| Dataset | Baseline | Best Entity | Worst Entity | |||
| Valid | Test | Valid | Test (Maj@10) | Valid | Test | |
| bAbI 12 | 95.1 | 95.5 | 98.5 [+3.4] | 99.0 [+3.5] | 92.4 [-2.7] | 93.0 [-2.5] |
| bAbI 14 | 76.1 | 77.5 | 78.4 [+2.3] | 79.5 [+2.0] | 75.2 [-0.9] | 76.6 [-0.9] |
| bAbI 16 | 53.5 | 53.3 | 61.7 [+8.2] | 61.3 [+8.0] | 55.5 [+2.0] | 55.9 [+2.6] |
| GSM8k | 14.3 | 10.5 | 14.6 [+0.3] | 11.3 [+0.8] | 12.7 [-1.6] | 10.1 [-0.4] |
| CLUTRR | 59.0 | 53.0 | 61.8 [+2.8] | 54.6 [+1.6] | 51.4 [-7.6] | 53.1 [+0.1] |
| Dataset | Index | Best performing entity | Worst performing entity | ||||
| entity | Acc change (valid) | Acc change (test) | entity | Acc change (valid) | Acc change (test) | ||
| bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5 | 1 | Penelope | 3.6 | 3.0 | Mihai | -4.0 | -2.3 |
| 2 | Loreen | 3.6 | 3.0 | Raisa | -4.0 | -3.6 | |
| 3 | Dorcas | 3.6 | 3.2 | Ethel | -3.7 | -3.5 | |
| 4 | Toby | 3.4 | 3.4 | Migdalia | -3.0 | -4.0 | |
| 5 | Flora | 3.3 | 2.9 | Azucena | -2.6 | -3.2 | |
| 6 | Nataliya | 3.3 | 3.4 | Bhavana | -2.6 | -3.6 | |
| 7 | Nathaniel | 3.3 | 3.4 | Surinder | -2.1 | -0.9 | |
| 8 | Lucretia | 3.3 | 3.3 | Sanjeev | -1.9 | -1.7 | |
| 9 | Keiko | 3.2 | 3.0 | Armida | -1.8 | -1.1 | |
| 10 | Bernardo | 3.2 | 3.1 | Jianwei | -1.7 | -1.4 | |
| Majority@10 | 3.5 | ||||||
| bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5 | 1 | Erich | 2.9 | 2.0 | Frederick | -1.4 | -1.2 |
| 2 | Jacqueline | 2.6 | 1.8 | Raisa | -1.3 | -1.3 | |
| 3 | Zoltan | 2.4 | 2.3 | Maryellen | -0.5 | -0.6 | |
| 4 | Chryystal | 2.3 | 1.5 | Meera | -0.5 | -0.5 | |
| 5 | Petra | 2.2 | 1.1 | — | — | — | |
| 6 | Pravin | 2.2 | 1.8 | — | — | — | |
| 7 | Claudio | 2.2 | 1.0 | — | — | — | |
| 8 | Sanjeev | 2.2 | 1.1 | — | — | — | |
| 9 | Soledad | 2.1 | 2.1 | — | — | — | |
| 10 | Prashant | 2.1 | 2.3 | — | — | — | |
| Majority@10 | 2.0 | ||||||
| bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3 | 1 | Loyd | 8.9 | 8.4 | — | — | — |
| 2 | Armida | 8.7 | 8.1 | — | — | — | |
| 3 | Brannon | 8.7 | 9.5 | — | — | — | |
| 4 | Pravin | 8.4 | 8.8 | — | — | — | |
| 5 | Bozena | 8.1 | 8.2 | — | — | — | |
| 6 | Surinder | 8.0 | 7.2 | — | — | — | |
| 7 | Bonita | 8.0 | 7.1 | — | — | — | |
| 8 | Trena | 8.0 | 6.9 | — | — | — | |
| 9 | Rosalia | 7.9 | 6.2 | — | — | — | |
| 10 | Leopoldo | 7.7 | 9.3 | — | — | — | |
| Majority@10 | 8.0 | ||||||
| GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5 | 1 | Douglas | 0.5 | 0.9 | Leopoldo | -2.5 | -0.8 |
| 2 | Juliann | 0.4 | 0.2 | Youssef | -1.9 | -0.3 | |
| 3 | Kathleen | 0.3 | 0.4 | Epifanioo | -1.7 | 0.5 | |
| 4 | Ernst | 0.3 | 0.4 | Sanjeev | -1.6 | -0.8 | |
| 5 | Tonja | 0.3 | 0.0 | Zoltan | -1.5 | -0.3 | |
| 6 | Angela | 0.2 | 0.2 | Jianwei | -1.4 | -0.4 | |
| 7 | Hina | 0.2 | -0.3 | Abelardo | -1.3 | -0.7 | |
| 8 | Diego | 0.2 | 0.4 | Alphonse | -1.3 | -0.7 | |
| 9 | Bradley | 0.1 | 0.9 | Dagoberto | -1.3 | -0.5 | |
| 10 | Jonathan | 0.1 | 0.8 | Nathaniel | -1.2 | -0.1 | |
| Majority@10 | 0.8 | ||||||
| CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0 | 1 | Ashlee | 4.4 | 1.2 | Rob | -8.2 | 0.6 |
| 2 | Penelope | 3.4 | 3.4 | Ahmad | -8.0 | -0.2 | |
| 3 | Allyson | 3.2 | 1.0 | Mihai | -7.6 | 0.6 | |
| 4 | Chryystal | 2.8 | 2.6 | Frederick | -7.6 | -0.4 | |
| 5 | Keiko | 2.8 | -4.4 | Martin | -7.4 | 0.6 | |
| 6 | Alisha | 2.6 | 3.0 | Jake | -7.4 | -0.4 | |
| 7 | Rosalba | 2.4 | 1.8 | Edward | -7.4 | 0.2 | |
| 8 | Jacqueline | 2.2 | 2.0 | Erik | -7.4 | 0.2 | |
| 9 | Zoltan | 2.2 | 1.2 | Giovanni | -7.4 | -0.4 | |
| 10 | Dillon | 2.0 | 1.4 | Douglas | -7.2 | -0.2 | |
| Majority@10 | 1.6 |
| Dataset | Baseline | Best Entity | Worst Entity | |||
| Valid | Test | Valid | Test (Maj@10) | Valid | Test | |
| bAbI 12 | 95.1 | 95.5 | 98.5 [+3.4] | 99.0 [+3.5] | 92.4 [-2.7] | 93.0 [-2.5] |
| bAbI 14 | 76.1 | 77.5 | 78.4 [+2.3] | 79.5 [+2.0] | 75.2 [-0.9] | 76.6 [-0.9] |
| bAbI 16 | 53.5 | 53.3 | 61.7 [+8.2] | 61.3 [+8.0] | 55.5 [+2.0] | 55.9 [+2.6] |
| GSM8k | 14.3 | 10.5 | 14.6 [+0.3] | 11.3 [+0.8] | 12.7 [-1.6] | 10.1 [-0.4] |
| CLUTRR | 59.0 | 53.0 | 61.8 [+2.8] | 54.6 [+1.6] | 51.4 [-7.6] | 53.1 [+0.1] |
| Dataset | Index | Best performing entity | Worst performing entity | ||||
| entity | Acc change (valid) | Acc change (test) | entity | Acc change (valid) | Acc change (test) | ||
| bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5 | 1 | Penelope | 3.6 | 3.0 | Mihai | -4.0 | -2.3 |
| 2 | Loreen | 3.6 | 3.0 | Raisa | -4.0 | -3.6 | |
| 3 | Dorcas | 3.6 | 3.2 | Ethel | -3.7 | -3.5 | |
| 4 | Toby | 3.4 | 3.4 | Migdalia | -3.0 | -4.0 | |
| 5 | Flora | 3.3 | 2.9 | Azucena | -2.6 | -3.2 | |
| 6 | Nataliya | 3.3 | 3.4 | Bhavana | -2.6 | -3.6 | |
| 7 | Nathaniel | 3.3 | 3.4 | Surinder | -2.1 | -0.9 | |
| 8 | Lucretia | 3.3 | 3.3 | Sanjeev | -1.9 | -1.7 | |
| 9 | Keiko | 3.2 | 3.0 | Armida | -1.8 | -1.1 | |
| 10 | Bernardo | 3.2 | 3.1 | Jianwei | -1.7 | -1.4 | |
| Majority@10 | 3.5 | ||||||
| bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5 | 1 | Erich | 2.9 | 2.0 | Frederick | -1.4 | -1.2 |
| 2 | Jacqueline | 2.6 | 1.8 | Raisa | -1.3 | -1.3 | |
| 3 | Zoltan | 2.4 | 2.3 | Maryellen | -0.5 | -0.6 | |
| 4 | Chryystal | 2.3 | 1.5 | Meera | -0.5 | -0.5 | |
| 5 | Petra | 2.2 | 1.1 | — | — | — | |
| 6 | Pravin | 2.2 | 1.8 | — | — | — | |
| 7 | Claudio | 2.2 | 1.0 | — | — | — | |
| 8 | Sanjeev | 2.2 | 1.1 | — | — | — | |
| 9 | Soledad | 2.1 | 2.1 | — | — | — | |
| 10 | Prashant | 2.1 | 2.3 | — | — | — | |
| Majority@10 | 2.0 | ||||||
| bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3 | 1 | Loyd | 8.9 | 8.4 | — | — | — |
| 2 | Armida | 8.7 | 8.1 | — | — | — | |
| 3 | Brannon | 8.7 | 9.5 | — | — | — | |
| 4 | Pravin | 8.4 | 8.8 | — | — | — | |
| 5 | Bozena | 8.1 | 8.2 | — | — | — | |
| 6 | Surinder | 8.0 | 7.2 | — | — | — | |
| 7 | Bonita | 8.0 | 7.1 | — | — | — | |
| 8 | Trena | 8.0 | 6.9 | — | — | — | |
| 9 | Rosalia | 7.9 | 6.2 | — | — | — | |
| 10 | Leopoldo | 7.7 | 9.3 | — | — | — | |
| Majority@10 | 8.0 | ||||||
| GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5 | 1 | Douglas | 0.5 | 0.9 | Leopoldo | -2.5 | -0.8 |
| 2 | Juliann | 0.4 | 0.2 | Youssef | -1.9 | -0.3 | |
| 3 | Kathleen | 0.3 | 0.4 | Epifanioo | -1.7 | 0.5 | |
| 4 | Ernst | 0.3 | 0.4 | Sanjeev | -1.6 | -0.8 | |
| 5 | Tonja | 0.3 | 0.0 | Zoltan | -1.5 | -0.3 | |
| 6 | Angela | 0.2 | 0.2 | Jianwei | -1.4 | -0.4 | |
| 7 | Hina | 0.2 | -0.3 | Abelardo | -1.3 | -0.7 | |
| 8 | Diego | 0.2 | 0.4 | Alphonse | -1.3 | -0.7 | |
| 9 | Bradley | 0.1 | 0.9 | Dagoberto | -1.3 | -0.5 | |
| 10 | Jonathan | 0.1 | 0.8 | Nathaniel | -1.2 | -0.1 | |
| Majority@10 | 0.8 | ||||||
| CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0 | 1 | Ashlee | 4.4 | 1.2 | Rob | -8.2 | 0.6 |
| 2 | Penelope | 3.4 | 3.4 | Ahmad | -8.0 | -0.2 | |
| 3 | Allyson | 3.2 | 1.0 | Mihai | -7.6 | 0.6 | |
| 4 | Chryystal | 2.8 | 2.6 | Frederick | -7.6 | -0.4 | |
| 5 | Keiko | 2.8 | -4.4 | Martin | -7.4 | 0.6 | |
| 6 | Alisha | 2.6 | 3.0 | Jake | -7.4 | -0.4 | |
| 7 | Rosalba | 2.4 | 1.8 | Edward | -7.4 | 0.2 | |
| 8 | Jacqueline | 2.2 | 2.0 | Erik | -7.4 | 0.2 | |
| 9 | Zoltan | 2.2 | 1.2 | Giovanni | -7.4 | -0.4 | |
| 10 | Dillon | 2.0 | 1.4 | Douglas | -7.2 | -0.2 | |
| Majority@10 | 1.6 |
| Dataset | Baseline | Best Entity | Worst Entity | |||
| Valid | Test | Valid | Test (Maj@10) | Valid | Test | |
| bAbI 12 | 95.1 | 95.5 | 98.5 [+3.4] | 99.0 [+3.5] | 92.4 [-2.7] | 93.0 [-2.5] |
| bAbI 14 | 76.1 | 77.5 | 78.4 [+2.3] | 79.5 [+2.0] | 75.2 [-0.9] | 76.6 [-0.9] |
| bAbI 16 | 53.5 | 53.3 | 61.7 [+8.2] | 61.3 [+8.0] | 55.5 [+2.0] | 55.9 [+2.6] |
| GSM8k | 14.3 | 10.5 | 14.6 [+0.3] | 11.3 [+0.8] | 12.7 [-1.6] | 10.1 [-0.4] |
| CLUTRR | 59.0 | 53.0 | 61.8 [+2.8] | 54.6 [+1.6] | 51.4 [-7.6] | 53.1 [+0.1] |
| Dataset | Index | Best performing entity | Worst performing entity | ||||
| entity | Acc change (valid) | Acc change (test) | entity | Acc change (valid) | Acc change (test) | ||
| bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5 | 1 | Penelope | 3.6 | 3.0 | Mihai | -4.0 | -2.3 |
| 2 | Loreen | 3.6 | 3.0 | Raisa | -4.0 | -3.6 | |
| 3 | Dorcas | 3.6 | 3.2 | Ethel | -3.7 | -3.5 | |
| 4 | Toby | 3.4 | 3.4 | Migdalia | -3.0 | -4.0 | |
| 5 | Flora | 3.3 | 2.9 | Azucena | -2.6 | -3.2 | |
| 6 | Nataliya | 3.3 | 3.4 | Bhavana | -2.6 | -3.6 | |
| 7 | Nathaniel | 3.3 | 3.4 | Surinder | -2.1 | -0.9 | |
| 8 | Lucretia | 3.3 | 3.3 | Sanjeev | -1.9 | -1.7 | |
| 9 | Keiko | 3.2 | 3.0 | Armida | -1.8 | -1.1 | |
| 10 | Bernardo | 3.2 | 3.1 | Jianwei | -1.7 | -1.4 | |
| Majority@10 | 3.5 | ||||||
| bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5 | 1 | Erich | 2.9 | 2.0 | Frederick | -1.4 | -1.2 |
| 2 | Jacqueline | 2.6 | 1.8 | Raisa | -1.3 | -1.3 | |
| 3 | Zoltan | 2.4 | 2.3 | Maryellen | -0.5 | -0.6 | |
| 4 | Chryystal | 2.3 | 1.5 | Meera | -0.5 | -0.5 | |
| 5 | Petra | 2.2 | 1.1 | — | — | — | |
| 6 | Pravin | 2.2 | 1.8 | — | — | — | |
| 7 | Claudio | 2.2 | 1.0 | — | — | — | |
| 8 | Sanjeev | 2.2 | 1.1 | — | — | — | |
| 9 | Soledad | 2.1 | 2.1 | — | — | — | |
| 10 | Prashant | 2.1 | 2.3 | — | — | — | |
| Majority@10 | 2.0 | ||||||
| bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3 | 1 | Loyd | 8.9 | 8.4 | — | — | — |
| 2 | Armida | 8.7 | 8.1 | — | — | — | |
| 3 | Brannon | 8.7 | 9.5 | — | — | — | |
| 4 | Pravin | 8.4 | 8.8 | — | — | — | |
| 5 | Bozena | 8.1 | 8.2 | — | — | — | |
| 6 | Surinder | 8.0 | 7.2 | — | — | — | |
| 7 | Bonita | 8.0 | 7.1 | — | — | — | |
| 8 | Trena | 8.0 | 6.9 | — | — | — | |
| 9 | Rosalia | 7.9 | 6.2 | — | — | — | |
| 10 | Leopoldo | 7.7 | 9.3 | — | — | — | |
| Majority@10 | 8.0 | ||||||
| GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5 | 1 | Douglas | 0.5 | 0.9 | Leopoldo | -2.5 | -0.8 |
| 2 | Juliann | 0.4 | 0.2 | Youssef | -1.9 | -0.3 | |
| 3 | Kathleen | 0.3 | 0.4 | Epifanioo | -1.7 | 0.5 | |
| 4 | Ernst | 0.3 | 0.4 | Sanjeev | -1.6 | -0.8 | |
| 5 | Tonja | 0.3 | 0.0 | Zoltan | -1.5 | -0.3 | |
| 6 | Angela | 0.2 | 0.2 | Jianwei | -1.4 | -0.4 | |
| 7 | Hina | 0.2 | -0.3 | Abelardo | -1.3 | -0.7 | |
| 8 | Diego | 0.2 | 0.4 | Alphonse | -1.3 | -0.7 | |
| 9 | Bradley | 0.1 | 0.9 | Dagoberto | -1.3 | -0.5 | |
| 10 | Jonathan | 0.1 | 0.8 | Nathaniel | -1.2 | -0.1 | |
| Majority@10 | 0.8 | ||||||
| CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0 | 1 | Ashlee | 4.4 | 1.2 | Rob | -8.2 | 0.6 |
| 2 | Penelope | 3.4 | 3.4 | Ahmad | -8.0 | -0.2 | |
| 3 | Allyson | 3.2 | 1.0 | Mihai | -7.6 | 0.6 | |
| 4 | Chryystal | 2.8 | 2.6 | Frederick | -7.6 | -0.4 | |
| 5 | Keiko | 2.8 | -4.4 | Martin | -7.4 | 0.6 | |
| 6 | Alisha | 2.6 | 3.0 | Jake | -7.4 | -0.4 | |
| 7 | Rosalba | 2.4 | 1.8 | Edward | -7.4 | 0.2 | |
| 8 | Jacqueline | 2.2 | 2.0 | Erik | -7.4 | 0.2 | |
| 9 | Zoltan | 2.2 | 1.2 | Giovanni | -7.4 | -0.4 | |
| 10 | Dillon | 2.0 | 1.4 | Douglas | -7.2 | -0.2 | |
| Majority@10 | 1.6 | ||||||
| Dataset | Baseline | Best Entity | Worst Entity | |||
| Valid | Test | Valid | Test (Maj@10) | Valid | Test | |
| bAbI 12 | 95.1 | 95.5 | 98.5 [+3.4] | 99.0 [+3.5] | 92.4 [-2.7] | 93.0 [-2.5] |
| bAbI 14 | 76.1 | 77.5 | 78.4 [+2.3] | 79.5 [+2.0] | 75.2 [-0.9] | 76.6 [-0.9] |
| bAbI 16 | 53.5 | 53.3 | 61.7 [+8.2] | 61.3 [+8.0] | 55.5 [+2.0] | 55.9 [+2.6] |
| GSM8k | 14.3 | 10.5 | 14.6 [+0.3] | 11.3 [+0.8] | 12.7 [-1.6] | 10.1 [-0.4] |
| CLUTRR | 59.0 | 53.0 | 61.8 [+2.8] | 54.6 [+1.6] | 51.4 [-7.6] | 53.1 [+0.1] |
| Dataset | Index | Best performing entity | Worst performing entity | ||||
| entity | Acc change (valid) | Acc change (test) | entity | Acc change (valid) | Acc change (test) | ||
| bAbI task 12 valid baseline Acc = 95.1 test baseline Acc = 95.5 | 1 | Penelope | 3.6 | 3.0 | Mihai | -4.0 | -2.3 |
| 2 | Loreen | 3.6 | 3.0 | Raisa | -4.0 | -3.6 | |
| 3 | Dorcas | 3.6 | 3.2 | Ethel | -3.7 | -3.5 | |
| 4 | Toby | 3.4 | 3.4 | Migdalia | -3.0 | -4.0 | |
| 5 | Flora | 3.3 | 2.9 | Azucena | -2.6 | -3.2 | |
| 6 | Nataliya | 3.3 | 3.4 | Bhavana | -2.6 | -3.6 | |
| 7 | Nathaniel | 3.3 | 3.4 | Surinder | -2.1 | -0.9 | |
| 8 | Lucretia | 3.3 | 3.3 | Sanjeev | -1.9 | -1.7 | |
| 9 | Keiko | 3.2 | 3.0 | Armida | -1.8 | -1.1 | |
| 10 | Bernardo | 3.2 | 3.1 | Jianwei | -1.7 | -1.4 | |
| Majority@10 | 3.5 | ||||||
| bAbI task 14 valid baseline Acc = 76.1 test baseline Acc = 77.5 | 1 | Erich | 2.9 | 2.0 | Frederick | -1.4 | -1.2 |
| 2 | Jacqueline | 2.6 | 1.8 | Raisa | -1.3 | -1.3 | |
| 3 | Zoltan | 2.4 | 2.3 | Maryellen | -0.5 | -0.6 | |
| 4 | Chryystal | 2.3 | 1.5 | Meera | -0.5 | -0.5 | |
| 5 | Petra | 2.2 | 1.1 | — | — | — | |
| 6 | Pravin | 2.2 | 1.8 | — | — | — | |
| 7 | Claudio | 2.2 | 1.0 | — | — | — | |
| 8 | Sanjeev | 2.2 | 1.1 | — | — | — | |
| 9 | Soledad | 2.1 | 2.1 | — | — | — | |
| 10 | Prashant | 2.1 | 2.3 | — | — | — | |
| Majority@10 | 2.0 | ||||||
| bAbI task 16 valid baseline Acc = 53.5 test baseline Acc = 53.3 | 1 | Loyd | 8.9 | 8.4 | — | — | — |
| 2 | Armida | 8.7 | 8.1 | — | — | — | |
| 3 | Brannon | 8.7 | 9.5 | — | — | — | |
| 4 | Pravin | 8.4 | 8.8 | — | — | — | |
| 5 | Bozena | 8.1 | 8.2 | — | — | — | |
| 6 | Surinder | 8.0 | 7.2 | — | — | — | |
| 7 | Bonita | 8.0 | 7.1 | — | — | — | |
| 8 | Trena | 8.0 | 6.9 | — | — | — | |
| 9 | Rosalia | 7.9 | 6.2 | — | — | — | |
| 10 | Leopoldo | 7.7 | 9.3 | — | — | — | |
| Majority@10 | 8.0 | ||||||
| GSM8k valid baseline Acc = 14.3 test baseline Acc = 10.5 | 1 | Douglas | 0.5 | 0.9 | Leopoldo | -2.5 | -0.8 |
| 2 | Juliann | 0.4 | 0.2 | Youssef | -1.9 | -0.3 | |
| 3 | Kathleen | 0.3 | 0.4 | Epifanioo | -1.7 | 0.5 | |
| 4 | Ernst | 0.3 | 0.4 | Sanjeev | -1.6 | -0.8 | |
| 5 | Tonja | 0.3 | 0.0 | Zoltan | -1.5 | -0.3 | |
| 6 | Angela | 0.2 | 0.2 | Jianwei | -1.4 | -0.4 | |
| 7 | Hina | 0.2 | -0.3 | Abelardo | -1.3 | -0.7 | |
| 8 | Diego | 0.2 | 0.4 | Alphonse | -1.3 | -0.7 | |
| 9 | Bradley | 0.1 | 0.9 | Dagoberto | -1.3 | -0.5 | |
| 10 | Jonathan | 0.1 | 0.8 | Nathaniel | -1.2 | -0.1 | |
| Majority@10 | 0.8 | ||||||
| CLUTRR valid baseline Acc = 59.0 test baseline Acc = 53.0 | 1 | Ashlee | 4.4 | 1.2 | Rob | -8.2 | 0.6 |
| 2 | Penelope | 3.4 | 3.4 | Ahmad | -8.0 | -0.2 | |
| 3 | Allyson | 3.2 | 1.0 | Mihai | -7.6 | 0.6 | |
| 4 | Chryystal | 2.8 | 2.6 | Frederick | -7.6 | -0.4 | |
| 5 | Keiko | 2.8 | -4.4 | Martin | -7.4 | 0.6 | |
| 6 | Alisha | 2.6 | 3.0 | Jake | -7.4 | -0.4 | |
| 7 | Rosalba | 2.4 | 1.8 | Edward | -7.4 | 0.2 | |
| 8 | Jacqueline | 2.2 | 2.0 | Erik | -7.4 | 0.2 | |
| 9 | Zoltan | 2.2 | 1.2 | Giovanni | -7.4 | -0.4 | |
| 10 | Dillon | 2.0 | 1.4 | Douglas | -7.2 | -0.2 | |
| Majority@10 | 1.6 |
| SD^E | But I, my grandparents were looking into it before. |
| SD | So I know what they’ve said. |
| B | Uh-huh. |
| ^H | Well, I’m trying to think. |
| Train | Dev | Test | |
| Utterances | 212,145 | 26,518 | 26,519 |
| Tokens | 1,833M | 227K | 227K |
| Tokens per att. | 8.64 | 8.56 | 8.32 |
| SD | 35.47% | 33.52% | 35.00% |
| %/Other | 27.45% | 25.77% | 25.60% |
| B | 18.52% | 21.54% | 23.02% |
| SV | 13.76% | 14.60% | 13.73% |
| AA | 4.81% | 4.57% | 2.65% |
| Context | 1 att. | 2 att. | ||
| Algo. | Dim. | Accuracy [%] | ||
| Maj. class | 35.00 | |||
| GLOVE | 6 | 51.62 | 50.37 | 50.14 |
| 12 | 52.82 | 52.60 | 52.55 | |
| 25 | 57.29 | 57.24 | 56.52 | |
| LDA | 6 | 51.10 | 51.16 | 51.05 |
| 12 | 53.32 | 52.99 | 52.73 | |
| 25 | 56.82 | 56.36 | 56.13 | |
| LSA | 6 | 51.07 | 50.62 | 50.58 |
| 12 | 51.94 | 51.44 | 51.04 | |
| 25 | 52.02 | 52.11 | 52.12 | |
| Random | 6 | 36.52 | 41.55 | 41.83 |
| 12 | 46.99 | 47.17 | 46.47 | |
| 25 | 48.46 | 47.69 | 47.59 | |
| Auto-enc. | 6 | 53.88 | 52.27 | 51.63 |
| 12 | 54.46 | 53.89 | 52.34 | |
| 25 | 52.43 | 53.18 | 52.57 | |
| Skip-Th. | 6 | 51.13 | 39.66 | |
| 12 | 53.33 | 52.38 | ||
| 25 | 55.21 | 57.54 |
| SD^E | But I, my grandparents were looking into it before. |
| SD | So I know what they’ve said. |
| B | Uh-huh. |
| ^H | Well, I’m trying to think. |
| Train | Dev | Test | |
| Utterances | 212,145 | 26,518 | 26,519 |
| Tokens | 1,833M | 227K | 227K |
| Tokens per att. | 8.64 | 8.56 | 8.32 |
| SD | 35.47% | 33.52% | 35.00% |
| %/Other | 27.45% | 25.77% | 25.60% |
| B | 18.52% | 21.54% | 23.02% |
| SV | 13.76% | 14.60% | 13.73% |
| AA | 4.81% | 4.57% | 2.65% |
| Context | 1 att. | 2 att. | ||
| Algo. | Dim. | Accuracy [%] | ||
| Maj. class | 35.00 | |||
| GLOVE | 6 | 51.62 | 50.37 | 50.14 |
| 12 | 52.82 | 52.60 | 52.55 | |
| 25 | 57.29 | 57.24 | 56.52 | |
| LDA | 6 | 51.10 | 51.16 | 51.05 |
| 12 | 53.32 | 52.99 | 52.73 | |
| 25 | 56.82 | 56.36 | 56.13 | |
| LSA | 6 | 51.07 | 50.62 | 50.58 |
| 12 | 51.94 | 51.44 | 51.04 | |
| 25 | 52.02 | 52.11 | 52.12 | |
| Random | 6 | 36.52 | 41.55 | 41.83 |
| 12 | 46.99 | 47.17 | 46.47 | |
| 25 | 48.46 | 47.69 | 47.59 | |
| Auto-enc. | 6 | 53.88 | 52.27 | 51.63 |
| 12 | 54.46 | 53.89 | 52.34 | |
| 25 | 52.43 | 53.18 | 52.57 | |
| Skip-Th. | 6 | 51.13 | 39.66 | |
| 12 | 53.33 | 52.38 | ||
| 25 | 55.21 | 57.54 |
| SD^E | But I, my grandparents were looking into it before. |
| SD | So I know what they’ve said. |
| B | Uh-huh. |
| ^H | Well, I’m trying to think. |
| Train | Dev | Test | |
| Utterances | 212,145 | 26,518 | 26,519 |
| Tokens | 1,833M | 227K | 227K |
| Tokens per att. | 8.64 | 8.56 | 8.32 |
| SD | 35.47% | 33.52% | 35.00% |
| %/Other | 27.45% | 25.77% | 25.60% |
| B | 18.52% | 21.54% | 23.02% |
| SV | 13.76% | 14.60% | 13.73% |
| AA | 4.81% | 4.57% | 2.65% |
| Context | 1 att. | 2 att. | ||
| Algo. | Dim. | Accuracy [%] | ||
| Maj. class | 35.00 | |||
| GLOVE | 6 | 51.62 | 50.37 | 50.14 |
| 12 | 52.82 | 52.60 | 52.55 | |
| 25 | 57.29 | 57.24 | 56.52 | |
| LDA | 6 | 51.10 | 51.16 | 51.05 |
| 12 | 53.32 | 52.99 | 52.73 | |
| 25 | 56.82 | 56.36 | 56.13 | |
| LSA | 6 | 51.07 | 50.62 | 50.58 |
| 12 | 51.94 | 51.44 | 51.04 | |
| 25 | 52.02 | 52.11 | 52.12 | |
| Random | 6 | 36.52 | 41.55 | 41.83 |
| 12 | 46.99 | 47.17 | 46.47 | |
| 25 | 48.46 | 47.69 | 47.59 | |
| Auto-enc. | 6 | 53.88 | 52.27 | 51.63 |
| 12 | 54.46 | 53.89 | 52.34 | |
| 25 | 52.43 | 53.18 | 52.57 | |
| Skip-Th. | 6 | 51.13 | 39.66 | |
| 12 | 53.33 | 52.38 | ||
| 25 | 55.21 | 57.54 | ||
| SD^E | But I, my grandparents were looking into it before. |
| SD | So I know what they’ve said. |
| B | Uh-huh. |
| ^H | Well, I’m trying to think. |
| Train | Dev | Test | |
| Utterances | 212,145 | 26,518 | 26,519 |
| Tokens | 1,833M | 227K | 227K |
| Tokens per att. | 8.64 | 8.56 | 8.32 |
| SD | 35.47% | 33.52% | 35.00% |
| %/Other | 27.45% | 25.77% | 25.60% |
| B | 18.52% | 21.54% | 23.02% |
| SV | 13.76% | 14.60% | 13.73% |
| AA | 4.81% | 4.57% | 2.65% |
| Context | 1 att. | 2 att. | ||
| Algo. | Dim. | Accuracy [%] | ||
| Maj. class | 35.00 | |||
| GLOVE | 6 | 51.62 | 50.37 | 50.14 |
| 12 | 52.82 | 52.60 | 52.55 | |
| 25 | 57.29 | 57.24 | 56.52 | |
| LDA | 6 | 51.10 | 51.16 | 51.05 |
| 12 | 53.32 | 52.99 | 52.73 | |
| 25 | 56.82 | 56.36 | 56.13 | |
| LSA | 6 | 51.07 | 50.62 | 50.58 |
| 12 | 51.94 | 51.44 | 51.04 | |
| 25 | 52.02 | 52.11 | 52.12 | |
| Random | 6 | 36.52 | 41.55 | 41.83 |
| 12 | 46.99 | 47.17 | 46.47 | |
| 25 | 48.46 | 47.69 | 47.59 | |
| Auto-enc. | 6 | 53.88 | 52.27 | 51.63 |
| 12 | 54.46 | 53.89 | 52.34 | |
| 25 | 52.43 | 53.18 | 52.57 | |
| Skip-Th. | 6 | 51.13 | 39.66 | |
| 12 | 53.33 | 52.38 | ||
| 25 | 55.21 | 57.54 |
| Size | Name | Loss | L1(%) | L2(%) |
| 774M | GPT-2 | 0.3264 | 34.99 | 40.03 |
| 7B | LLaMA | 0.2408 | 53.04 | 62.44 |
| Alpaca | 0.2646 | 52.51 | 61.71 | |
| Vicuna | 0.2593 | 51.90 | 61.19 | |
| 13B | LLaMA | 0.2406 | 55.66 | 64.20 |
| Alpaca | 0.2634 | 55.05 | 64.46 | |
| Vicuna | 0.2847 | 54.26 | 61.31 | |
| 30B | LLaMA | 0.2372 | 63.16 | 69.40 |
| 65B | LLaMA | 0.2375 | 64.05 | 70.07 |
| Model | Resemblance (%) | ||
| Name | Size | L1 | L2 |
| GPT-2 | 774M | 34.99 | 40.03 |
| 7B | 53.04 | 62.44 | |
| LLaMA | 13B | 55.56 | 64.20 |
| 30B | 63.16 | 69.40 | |
| 65B | 64.05 | 70.07 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 53.04 | 62.44 |
| Alpaca | 52.51 | 61.71 | |
| Vicuna | 51.90 | 61.19 | |
| 13B | LLaMA | 55.66 | 64.20 |
| Alpaca | 55.05 | 63.46 | |
| Vicuna | 54.26 | 61.31 |
| Model | Higher | Lower | |||
| Size | Name | L1 | L2 | L1 | L2 |
| 7B | Alpaca | 3 | 8 | 4 | 7 |
| Vicuna | 9 | 10 | 8 | 9 | |
| 13B | Alpaca | 4 | 7 | 16 | 25 |
| Vicuna | 5 | 8 | 14 | 16 |
| Group | Min | Max | Mean | SE |
| L1 | 0.0048 | 0.1525 | 0.0473 | 0.0037 |
| L2 | 0.0134 | 0.2165 | 0.0894 | 0.0060 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 2, 5, 7 | 12, 13, 17, 18, 21, 26, 28, 29, 30 |
| L2 | 2, 5, 6, 8, 12, 15, 21, 27 | 3, 13, 15, 18, 21, 23, 26, 28, 29, 30 | |
| Lower | L1 | 4, 10, 23, 31 | 0, 1, 4, 5, 9, 11, 20, 31 |
| L2 | 0, 10, 14, 19, 23, 28, 31 | 0, 1, 4, 5, 9, 11, 14, 20, 31 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 16, 32, 33, 36 | 4, 16, 27, 33, 36 |
| L2 | 1, 16, 19, 27, 32, 33, 36 | 4, 16, 19, 23, 27, 33, 36, 37 | |
| Lower | L1 | 0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39 | 0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39 |
| L2 | 0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39 | 0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 11.02 | 15.45 |
| Alpaca | 10.99 | 15.61 | |
| Vicuna | 11.35 | 15.44 | |
| 13B | LLaMA | 14.69 | 20.19 |
| Alpaca | 14.67 | 19.44 | |
| Vicuna | 14.02 | 18.57 |
| Size | Name | Loss | L1(%) | L2(%) |
| 774M | GPT-2 | 0.3264 | 34.99 | 40.03 |
| 7B | LLaMA | 0.2408 | 53.04 | 62.44 |
| Alpaca | 0.2646 | 52.51 | 61.71 | |
| Vicuna | 0.2593 | 51.90 | 61.19 | |
| 13B | LLaMA | 0.2406 | 55.66 | 64.20 |
| Alpaca | 0.2634 | 55.05 | 64.46 | |
| Vicuna | 0.2847 | 54.26 | 61.31 | |
| 30B | LLaMA | 0.2372 | 63.16 | 69.40 |
| 65B | LLaMA | 0.2375 | 64.05 | 70.07 |
| Model | Resemblance (%) | ||
| Name | Size | L1 | L2 |
| GPT-2 | 774M | 34.99 | 40.03 |
| 7B | 53.04 | 62.44 | |
| LLaMA | 13B | 55.56 | 64.20 |
| 30B | 63.16 | 69.40 | |
| 65B | 64.05 | 70.07 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 53.04 | 62.44 |
| Alpaca | 52.51 | 61.71 | |
| Vicuna | 51.90 | 61.19 | |
| 13B | LLaMA | 55.66 | 64.20 |
| Alpaca | 55.05 | 63.46 | |
| Vicuna | 54.26 | 61.31 |
| Model | Higher | Lower | |||
| Size | Name | L1 | L2 | L1 | L2 |
| 7B | Alpaca | 3 | 8 | 4 | 7 |
| Vicuna | 9 | 10 | 8 | 9 | |
| 13B | Alpaca | 4 | 7 | 16 | 25 |
| Vicuna | 5 | 8 | 14 | 16 |
| Group | Min | Max | Mean | SE |
| L1 | 0.0048 | 0.1525 | 0.0473 | 0.0037 |
| L2 | 0.0134 | 0.2165 | 0.0894 | 0.0060 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 2, 5, 7 | 12, 13, 17, 18, 21, 26, 28, 29, 30 |
| L2 | 2, 5, 6, 8, 12, 15, 21, 27 | 3, 13, 15, 18, 21, 23, 26, 28, 29, 30 | |
| Lower | L1 | 4, 10, 23, 31 | 0, 1, 4, 5, 9, 11, 20, 31 |
| L2 | 0, 10, 14, 19, 23, 28, 31 | 0, 1, 4, 5, 9, 11, 14, 20, 31 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 16, 32, 33, 36 | 4, 16, 27, 33, 36 |
| L2 | 1, 16, 19, 27, 32, 33, 36 | 4, 16, 19, 23, 27, 33, 36, 37 | |
| Lower | L1 | 0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39 | 0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39 |
| L2 | 0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39 | 0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 11.02 | 15.45 |
| Alpaca | 10.99 | 15.61 | |
| Vicuna | 11.35 | 15.44 | |
| 13B | LLaMA | 14.69 | 20.19 |
| Alpaca | 14.67 | 19.44 | |
| Vicuna | 14.02 | 18.57 |
| Size | Name | Loss | L1(%) | L2(%) |
| 774M | GPT-2 | 0.3264 | 34.99 | 40.03 |
| 7B | LLaMA | 0.2408 | 53.04 | 62.44 |
| Alpaca | 0.2646 | 52.51 | 61.71 | |
| Vicuna | 0.2593 | 51.90 | 61.19 | |
| 13B | LLaMA | 0.2406 | 55.66 | 64.20 |
| Alpaca | 0.2634 | 55.05 | 64.46 | |
| Vicuna | 0.2847 | 54.26 | 61.31 | |
| 30B | LLaMA | 0.2372 | 63.16 | 69.40 |
| 65B | LLaMA | 0.2375 | 64.05 | 70.07 |
| Model | Resemblance (%) | ||
| Name | Size | L1 | L2 |
| GPT-2 | 774M | 34.99 | 40.03 |
| 7B | 53.04 | 62.44 | |
| LLaMA | 13B | 55.56 | 64.20 |
| 30B | 63.16 | 69.40 | |
| 65B | 64.05 | 70.07 | |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 53.04 | 62.44 |
| Alpaca | 52.51 | 61.71 | |
| Vicuna | 51.90 | 61.19 | |
| 13B | LLaMA | 55.66 | 64.20 |
| Alpaca | 55.05 | 63.46 | |
| Vicuna | 54.26 | 61.31 | |
| Model | Higher | Lower | |||
| Size | Name | L1 | L2 | L1 | L2 |
| 7B | Alpaca | 3 | 8 | 4 | 7 |
| Vicuna | 9 | 10 | 8 | 9 | |
| 13B | Alpaca | 4 | 7 | 16 | 25 |
| Vicuna | 5 | 8 | 14 | 16 | |
| Group | Min | Max | Mean | SE |
| L1 | 0.0048 | 0.1525 | 0.0473 | 0.0037 |
| L2 | 0.0134 | 0.2165 | 0.0894 | 0.0060 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 2, 5, 7 | 12, 13, 17, 18, 21, 26, 28, 29, 30 |
| L2 | 2, 5, 6, 8, 12, 15, 21, 27 | 3, 13, 15, 18, 21, 23, 26, 28, 29, 30 | |
| Lower | L1 | 4, 10, 23, 31 | 0, 1, 4, 5, 9, 11, 20, 31 |
| L2 | 0, 10, 14, 19, 23, 28, 31 | 0, 1, 4, 5, 9, 11, 14, 20, 31 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 16, 32, 33, 36 | 4, 16, 27, 33, 36 |
| L2 | 1, 16, 19, 27, 32, 33, 36 | 4, 16, 19, 23, 27, 33, 36, 37 | |
| Lower | L1 | 0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39 | 0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39 |
| L2 | 0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39 | 0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 11.02 | 15.45 |
| Alpaca | 10.99 | 15.61 | |
| Vicuna | 11.35 | 15.44 | |
| 13B | LLaMA | 14.69 | 20.19 |
| Alpaca | 14.67 | 19.44 | |
| Vicuna | 14.02 | 18.57 | |
| Size | Name | Loss | L1(%) | L2(%) |
| 774M | GPT-2 | 0.3264 | 34.99 | 40.03 |
| 7B | LLaMA | 0.2408 | 53.04 | 62.44 |
| Alpaca | 0.2646 | 52.51 | 61.71 | |
| Vicuna | 0.2593 | 51.90 | 61.19 | |
| 13B | LLaMA | 0.2406 | 55.66 | 64.20 |
| Alpaca | 0.2634 | 55.05 | 64.46 | |
| Vicuna | 0.2847 | 54.26 | 61.31 | |
| 30B | LLaMA | 0.2372 | 63.16 | 69.40 |
| 65B | LLaMA | 0.2375 | 64.05 | 70.07 |
| Model | Resemblance (%) | ||
| Name | Size | L1 | L2 |
| GPT-2 | 774M | 34.99 | 40.03 |
| 7B | 53.04 | 62.44 | |
| LLaMA | 13B | 55.56 | 64.20 |
| 30B | 63.16 | 69.40 | |
| 65B | 64.05 | 70.07 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 53.04 | 62.44 |
| Alpaca | 52.51 | 61.71 | |
| Vicuna | 51.90 | 61.19 | |
| 13B | LLaMA | 55.66 | 64.20 |
| Alpaca | 55.05 | 63.46 | |
| Vicuna | 54.26 | 61.31 |
| Model | Higher | Lower | |||
| Size | Name | L1 | L2 | L1 | L2 |
| 7B | Alpaca | 3 | 8 | 4 | 7 |
| Vicuna | 9 | 10 | 8 | 9 | |
| 13B | Alpaca | 4 | 7 | 16 | 25 |
| Vicuna | 5 | 8 | 14 | 16 |
| Group | Min | Max | Mean | SE |
| L1 | 0.0048 | 0.1525 | 0.0473 | 0.0037 |
| L2 | 0.0134 | 0.2165 | 0.0894 | 0.0060 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 2, 5, 7 | 12, 13, 17, 18, 21, 26, 28, 29, 30 |
| L2 | 2, 5, 6, 8, 12, 15, 21, 27 | 3, 13, 15, 18, 21, 23, 26, 28, 29, 30 | |
| Lower | L1 | 4, 10, 23, 31 | 0, 1, 4, 5, 9, 11, 20, 31 |
| L2 | 0, 10, 14, 19, 23, 28, 31 | 0, 1, 4, 5, 9, 11, 14, 20, 31 |
| Relation | Group | Alpaca | Vicuna |
| Higher | L1 | 16, 32, 33, 36 | 4, 16, 27, 33, 36 |
| L2 | 1, 16, 19, 27, 32, 33, 36 | 4, 16, 19, 23, 27, 33, 36, 37 | |
| Lower | L1 | 0, 2, 3, 5, 6, 7, 8, 9, 14, 18, 24, 25, 28, 30, 35, 39 | 0, 1, 3, 5, 6, 9, 10, 14, 17, 18, 24, 30, 34, 39 |
| L2 | 0, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 15, 17, 18, 20, 24, 25, 28, 29, 30, 31, 35, 38, 39 | 0, 2, 3, 5, 6, 8, 9, 10, 11, 12, 14, 17, 18, 20, 21 30 |
| Model | Resemblance(%) | ||
| Size | Name | L1 | L2 |
| 7B | LLaMA | 11.02 | 15.45 |
| Alpaca | 10.99 | 15.61 | |
| Vicuna | 11.35 | 15.44 | |
| 13B | LLaMA | 14.69 | 20.19 |
| Alpaca | 14.67 | 19.44 | |
| Vicuna | 14.02 | 18.57 |
| Language | Original Text | Romanized Text | Translation |
| Bhojpuri | बाल्दियम् गार्वा | jorjiyan bhaasaas | Georgian language |
| Sinhala | सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था | ||
| Sindhi | सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था | grahalooka | planets |
| Khmer | khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄.\n<|im_start|>assistant | Bhojpuri | BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ |
| Language | Family | Script | # Sentences |
| Bhojpuri (bh) | Indo-European | Devanagari | 35,983 |
| Buryat (bxr) | Mongolic | Cyrillic | 41,692 |
| Erzya (myv) | Uralic | Cyrillic | 42,575 |
| Meadow Mari (mhr) | Uralic | Cyrillic | 144,529 |
| Min Dong (cdo) | Sino-Tibetan | Chinese | 33,978 |
| Mingrelian (xmf) | Kartvelian | Georgian | 63,032 |
| Sindhi (sd) | Indo-European | Arabic | 86,176 |
| Sorani Kurdish (ckb) | Indo-European | Arabic | 459,475 |
| Uyghur (ug) | Turkic | Arabic | 149,813 |
| Amháric (am) | Afro-Asiatic | Ge'ez | 88,320 |
| Divehi (dv) | Indo-European | Thaana | 34,779 |
| Khmer (km) | Austroasiatic | Khmer | 139,704 |
| Sinhala (si) | Indo-European | Sinhala | 219,866 |
| Tibetan (bo) | Sino-Tibetan | Tibetan | 131,362 |
| Task | Transliterator | am | ar | ka | ru | hi | sd | avg |
| NER (Macro F1) | UROMAN | 25.6 | 24.8 | 61.4 | 66.5 | 48.6 | 35.3 | 43.7 |
| Other | 25.5 | 23.7 | 57.3 | 63.9 | 56.7 | 35.9 | 43.8 | |
| UD (UAS / LAS) | UROMAN | 36.17 / 6.6 | 33.0 / 19.8 | - | 47.3 / 32.4 | 33.8 / 17.8 | - | 37.5 / 19.1 |
| Other | 29.9 / 5.4 | 32.6 / 19.9 | - | 45.0 / 19.9 | 33.2 / 17.9 | - | 35.2 / 15.8 |
| Method | Seen Script | Unseen Script | avg | |||||||||||
| bh | cdo | ckb | mhr | sd | ug | xmf | am | bo | dv | km | si | |||
| UROMAN | 32.59 | 27.34 | 67.73 | 64.68 | 35.33 | 28.10 | 52.58 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 | |
| BORROW | 53.42 (hi) | - | 12.46 (ar) | 45.86 (ru) | 16.79 (ar) | 12.85 (ar) | 24.77 (ru) | - | - | - | - | - | - | |
| RAND | 25.42 | 19.51 | 53.55 | 42.02 | 27.20 | 25.18 | 35.82 | 18.00 | 18.95 | 21.19 | 32.75 | 20.01 | 21.59 |
| Variant | Seen Script | Unseen Script | avg | ||||||||||
| bh | cdo | sd | xmf | mhr | ckb | ug | am | bo | dv | km | si | ||
| Non-TransLA+INV | 55.14 | 24.19 | 31.31 | 49.74 | 70.31 | 45.54 | 33.53 | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 | 18.39 |
| TransLA+INV | 32.59 | 27.34 | 35.33 | 52.58 | 64.68 | 67.73 | 28.10 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 |
| Non-TransLA+EmblEx | 60.00 | 28.91 | 42.47 | 51.99 | 61.05 | 79.12 | 50.42 | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 | 49.01 |
| TransLA + mPLMrl | 49.05 | 36.92 | 39.16 | 57.99 | 69.85 | 73.92 | 33.43 | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 | 47.44 |
| Variant | Seen Script | Unseen Script | ||||
| bh | myv | ug | bxr | am | avg | |
| Non-TransLA+INV | 28.46 / 11.53 | 45.28 / 26.27 | 33.44 / 15.28 | 39.75 / 19.77 | 19.08 / 1.85 | 33.20 / 10.81 |
| TransLA+INV | 25.12 / 10.17 | 45.74 / 26.64 | 32.30 / 15.10 | 37.92 / 17.23 | 36.07 / 7.58 | 35.43 / 12.41 |
| Non-TransLA+EmblEx | 26.68 / 10.10 | 48.34 / 25.34 | 41.20 / 22.81 | 39.51 / 16.02 | 36.4778.39 | 38.44 / 12.20 |
| TransLA + mPLMrl | 28.04 / 11.13 | 41.97 / 20.29 | 50.89 / 16.56 | 35.03 / 20.29 | 39.10 / 9.00 | 39.01 / 14.65 |
| Transliterator | Used for languages | Available at |
| UROMAN | All | github.com/isi-nlp/uroman |
| amseg | am | pypi.org/project/amseg/ |
| transliterate | ru, ka | pypi.org/project/transliterate/ |
| ai4bharat-transliteration | hi, sd | pypi.org/project/ai4bharat-transliteration/ |
| lang-trans | ar | pypi.org/project/lang-trans/ |
| Variant | Unseen Scripts | ||||
| am | bo | dv | km | si | |
| mBERT | 0.91 | 17.40 | 1.30 | 10.71 | 2.50 |
| Non-TransLA+INV | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 |
| TransLA + INV | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 |
| Non-TransLA+EmblEx | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 |
| TransLA + mPLMft | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 |
| Language | Original Text | Romanized Text | Translation |
| Bhojpuri | बाल्दियम् गार्वा | jorjiyan bhaasaas | Georgian language |
| Sinhala | सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था | ||
| Sindhi | सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था | grahalooka | planets |
| Khmer | khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄.\n<|im_start|>assistant | Bhojpuri | BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ |
| Language | Family | Script | # Sentences |
| Bhojpuri (bh) | Indo-European | Devanagari | 35,983 |
| Buryat (bxr) | Mongolic | Cyrillic | 41,692 |
| Erzya (myv) | Uralic | Cyrillic | 42,575 |
| Meadow Mari (mhr) | Uralic | Cyrillic | 144,529 |
| Min Dong (cdo) | Sino-Tibetan | Chinese | 33,978 |
| Mingrelian (xmf) | Kartvelian | Georgian | 63,032 |
| Sindhi (sd) | Indo-European | Arabic | 86,176 |
| Sorani Kurdish (ckb) | Indo-European | Arabic | 459,475 |
| Uyghur (ug) | Turkic | Arabic | 149,813 |
| Amháric (am) | Afro-Asiatic | Ge'ez | 88,320 |
| Divehi (dv) | Indo-European | Thaana | 34,779 |
| Khmer (km) | Austroasiatic | Khmer | 139,704 |
| Sinhala (si) | Indo-European | Sinhala | 219,866 |
| Tibetan (bo) | Sino-Tibetan | Tibetan | 131,362 |
| Task | Transliterator | am | ar | ka | ru | hi | sd | avg |
| NER (Macro F1) | UROMAN | 25.6 | 24.8 | 61.4 | 66.5 | 48.6 | 35.3 | 43.7 |
| Other | 25.5 | 23.7 | 57.3 | 63.9 | 56.7 | 35.9 | 43.8 | |
| UD (UAS / LAS) | UROMAN | 36.17 / 6.6 | 33.0 / 19.8 | - | 47.3 / 32.4 | 33.8 / 17.8 | - | 37.5 / 19.1 |
| Other | 29.9 / 5.4 | 32.6 / 19.9 | - | 45.0 / 19.9 | 33.2 / 17.9 | - | 35.2 / 15.8 |
| Method | Seen Script | Unseen Script | avg | |||||||||||
| bh | cdo | ckb | mhr | sd | ug | xmf | am | bo | dv | km | si | |||
| UROMAN | 32.59 | 27.34 | 67.73 | 64.68 | 35.33 | 28.10 | 52.58 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 | |
| BORROW | 53.42 (hi) | - | 12.46 (ar) | 45.86 (ru) | 16.79 (ar) | 12.85 (ar) | 24.77 (ru) | - | - | - | - | - | - | |
| RAND | 25.42 | 19.51 | 53.55 | 42.02 | 27.20 | 25.18 | 35.82 | 18.00 | 18.95 | 21.19 | 32.75 | 20.01 | 21.59 |
| Variant | Seen Script | Unseen Script | avg | ||||||||||
| bh | cdo | sd | xmf | mhr | ckb | ug | am | bo | dv | km | si | ||
| Non-TransLA+INV | 55.14 | 24.19 | 31.31 | 49.74 | 70.31 | 45.54 | 33.53 | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 | 18.39 |
| TransLA+INV | 32.59 | 27.34 | 35.33 | 52.58 | 64.68 | 67.73 | 28.10 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 |
| Non-TransLA+EmblEx | 60.00 | 28.91 | 42.47 | 51.99 | 61.05 | 79.12 | 50.42 | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 | 49.01 |
| TransLA + mPLMrl | 49.05 | 36.92 | 39.16 | 57.99 | 69.85 | 73.92 | 33.43 | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 | 47.44 |
| Variant | Seen Script | Unseen Script | ||||
| bh | myv | ug | bxr | am | avg | |
| Non-TransLA+INV | 28.46 / 11.53 | 45.28 / 26.27 | 33.44 / 15.28 | 39.75 / 19.77 | 19.08 / 1.85 | 33.20 / 10.81 |
| TransLA+INV | 25.12 / 10.17 | 45.74 / 26.64 | 32.30 / 15.10 | 37.92 / 17.23 | 36.07 / 7.58 | 35.43 / 12.41 |
| Non-TransLA+EmblEx | 26.68 / 10.10 | 48.34 / 25.34 | 41.20 / 22.81 | 39.51 / 16.02 | 36.4778.39 | 38.44 / 12.20 |
| TransLA + mPLMrl | 28.04 / 11.13 | 41.97 / 20.29 | 50.89 / 16.56 | 35.03 / 20.29 | 39.10 / 9.00 | 39.01 / 14.65 |
| Transliterator | Used for languages | Available at |
| UROMAN | All | github.com/isi-nlp/uroman |
| amseg | am | pypi.org/project/amseg/ |
| transliterate | ru, ka | pypi.org/project/transliterate/ |
| ai4bharat-transliteration | hi, sd | pypi.org/project/ai4bharat-transliteration/ |
| lang-trans | ar | pypi.org/project/lang-trans/ |
| Variant | Unseen Scripts | ||||
| am | bo | dv | km | si | |
| mBERT | 0.91 | 17.40 | 1.30 | 10.71 | 2.50 |
| Non-TransLA+INV | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 |
| TransLA + INV | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 |
| Non-TransLA+EmblEx | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 |
| TransLA + mPLMft | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 |
| Language | Original Text | Romanized Text | Translation |
| Bhojpuri | बाल्दियम् गार्वा | jorjiyan bhaasaas | Georgian language |
| Sinhala | सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था | ||
| Sindhi | सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था | grahalooka | planets |
| Khmer | khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄. +<|im_start|>assistant | Bhojpuri | BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ |
| Language | Family | Script | # Sentences |
| Bhojpuri (bh) | Indo-European | Devanagari | 35,983 |
| Buryat (bxr) | Mongolic | Cyrillic | 41,692 |
| Erzya (myv) | Uralic | Cyrillic | 42,575 |
| Meadow Mari (mhr) | Uralic | Cyrillic | 144,529 |
| Min Dong (cdo) | Sino-Tibetan | Chinese | 33,978 |
| Mingrelian (xmf) | Kartvelian | Georgian | 63,032 |
| Sindhi (sd) | Indo-European | Arabic | 86,176 |
| Sorani Kurdish (ckb) | Indo-European | Arabic | 459,475 |
| Uyghur (ug) | Turkic | Arabic | 149,813 |
| Amháric (am) | Afro-Asiatic | Ge'ez | 88,320 |
| Divehi (dv) | Indo-European | Thaana | 34,779 |
| Khmer (km) | Austroasiatic | Khmer | 139,704 |
| Sinhala (si) | Indo-European | Sinhala | 219,866 |
| Tibetan (bo) | Sino-Tibetan | Tibetan | 131,362 |
| Task | Transliterator | am | ar | ka | ru | hi | sd | avg |
| NER (Macro F1) | UROMAN | 25.6 | 24.8 | 61.4 | 66.5 | 48.6 | 35.3 | 43.7 |
| Other | 25.5 | 23.7 | 57.3 | 63.9 | 56.7 | 35.9 | 43.8 | |
| UD (UAS / LAS) | UROMAN | 36.17 / 6.6 | 33.0 / 19.8 | - | 47.3 / 32.4 | 33.8 / 17.8 | - | 37.5 / 19.1 |
| Other | 29.9 / 5.4 | 32.6 / 19.9 | - | 45.0 / 19.9 | 33.2 / 17.9 | - | 35.2 / 15.8 |
| Method | Seen Script | Unseen Script | avg | |||||||||||
| bh | cdo | ckb | mhr | sd | ug | xmf | am | bo | dv | km | si | |||
| UROMAN | 32.59 | 27.34 | 67.73 | 64.68 | 35.33 | 28.10 | 52.58 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 | |
| BORROW | 53.42 (hi) | - | 12.46 (ar) | 45.86 (ru) | 16.79 (ar) | 12.85 (ar) | 24.77 (ru) | - | - | - | - | - | - | |
| RAND | 25.42 | 19.51 | 53.55 | 42.02 | 27.20 | 25.18 | 35.82 | 18.00 | 18.95 | 21.19 | 32.75 | 20.01 | 21.59 | |
| Variant | Seen Script | Unseen Script | avg | ||||||||||
| bh | cdo | sd | xmf | mhr | ckb | ug | am | bo | dv | km | si | ||
| Non-TransLA+INV | 55.14 | 24.19 | 31.31 | 49.74 | 70.31 | 45.54 | 33.53 | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 | 18.39 |
| TransLA+INV | 32.59 | 27.34 | 35.33 | 52.58 | 64.68 | 67.73 | 28.10 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 |
| Non-TransLA+EmblEx | 60.00 | 28.91 | 42.47 | 51.99 | 61.05 | 79.12 | 50.42 | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 | 49.01 |
| TransLA + mPLMrl | 49.05 | 36.92 | 39.16 | 57.99 | 69.85 | 73.92 | 33.43 | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 | 47.44 |
| Variant | Seen Script | Unseen Script | ||||
| bh | myv | ug | bxr | am | avg | |
| Non-TransLA+INV | 28.46 / 11.53 | 45.28 / 26.27 | 33.44 / 15.28 | 39.75 / 19.77 | 19.08 / 1.85 | 33.20 / 10.81 |
| TransLA+INV | 25.12 / 10.17 | 45.74 / 26.64 | 32.30 / 15.10 | 37.92 / 17.23 | 36.07 / 7.58 | 35.43 / 12.41 |
| Non-TransLA+EmblEx | 26.68 / 10.10 | 48.34 / 25.34 | 41.20 / 22.81 | 39.51 / 16.02 | 36.4778.39 | 38.44 / 12.20 |
| TransLA + mPLMrl | 28.04 / 11.13 | 41.97 / 20.29 | 50.89 / 16.56 | 35.03 / 20.29 | 39.10 / 9.00 | 39.01 / 14.65 |
| Transliterator | Used for languages | Available at |
| UROMAN | All | github.com/isi-nlp/uroman |
| amseg | am | pypi.org/project/amseg/ |
| transliterate | ru, ka | pypi.org/project/transliterate/ |
| ai4bharat-transliteration | hi, sd | pypi.org/project/ai4bharat-transliteration/ |
| lang-trans | ar | pypi.org/project/lang-trans/ |
| Variant | Unseen Scripts | ||||
| am | bo | dv | km | si | |
| mBERT | 0.91 | 17.40 | 1.30 | 10.71 | 2.50 |
| Non-TransLA+INV | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 |
| TransLA + INV | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 |
| Non-TransLA+EmblEx | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 |
| TransLA + mPLMft | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 |
| Language | Original Text | Romanized Text | Translation |
| Bhojpuri | बाल्दियम् गार्वा | jorjiyan bhaasaas | Georgian language |
| Sinhala | सिकूल्तिूल्तिूल्तिूल्तिूल्तिूल्तिू�िूल्तिूल्तिूल्तिू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�िू�र्था | ||
| Sindhi | सिकूल्तिूल्तिूल्तिू�िू�िू�िू�र्था | grahalooka | planets |
| Khmer | khμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμmu μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμµ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμum μ μμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμμ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ µ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μ μµ μµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµüµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµúµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ µµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµµ₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₃₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₂₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₁₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀₀⁻⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁶⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁹⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁷⁰⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁵⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁴⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁽⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁻⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁼⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸ⁱ⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁅⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁸⁰⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₃⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄²⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄⁄₂⁄.\n<|im_start|>assistant | Bhojpuri | BJJUJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJJJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIajiJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJIJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJP JPJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJJPJJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJPJSJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ |
| Language | Family | Script | # Sentences |
| Bhojpuri (bh) | Indo-European | Devanagari | 35,983 |
| Buryat (bxr) | Mongolic | Cyrillic | 41,692 |
| Erzya (myv) | Uralic | Cyrillic | 42,575 |
| Meadow Mari (mhr) | Uralic | Cyrillic | 144,529 |
| Min Dong (cdo) | Sino-Tibetan | Chinese | 33,978 |
| Mingrelian (xmf) | Kartvelian | Georgian | 63,032 |
| Sindhi (sd) | Indo-European | Arabic | 86,176 |
| Sorani Kurdish (ckb) | Indo-European | Arabic | 459,475 |
| Uyghur (ug) | Turkic | Arabic | 149,813 |
| Amháric (am) | Afro-Asiatic | Ge'ez | 88,320 |
| Divehi (dv) | Indo-European | Thaana | 34,779 |
| Khmer (km) | Austroasiatic | Khmer | 139,704 |
| Sinhala (si) | Indo-European | Sinhala | 219,866 |
| Tibetan (bo) | Sino-Tibetan | Tibetan | 131,362 |
| Task | Transliterator | am | ar | ka | ru | hi | sd | avg |
| NER (Macro F1) | UROMAN | 25.6 | 24.8 | 61.4 | 66.5 | 48.6 | 35.3 | 43.7 |
| Other | 25.5 | 23.7 | 57.3 | 63.9 | 56.7 | 35.9 | 43.8 | |
| UD (UAS / LAS) | UROMAN | 36.17 / 6.6 | 33.0 / 19.8 | - | 47.3 / 32.4 | 33.8 / 17.8 | - | 37.5 / 19.1 |
| Other | 29.9 / 5.4 | 32.6 / 19.9 | - | 45.0 / 19.9 | 33.2 / 17.9 | - | 35.2 / 15.8 |
| Method | Seen Script | Unseen Script | avg | |||||||||||
| bh | cdo | ckb | mhr | sd | ug | xmf | am | bo | dv | km | si | |||
| UROMAN | 32.59 | 27.34 | 67.73 | 64.68 | 35.33 | 28.10 | 52.58 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 | |
| BORROW | 53.42 (hi) | - | 12.46 (ar) | 45.86 (ru) | 16.79 (ar) | 12.85 (ar) | 24.77 (ru) | - | - | - | - | - | - | |
| RAND | 25.42 | 19.51 | 53.55 | 42.02 | 27.20 | 25.18 | 35.82 | 18.00 | 18.95 | 21.19 | 32.75 | 20.01 | 21.59 |
| Variant | Seen Script | Unseen Script | avg | ||||||||||
| bh | cdo | sd | xmf | mhr | ckb | ug | am | bo | dv | km | si | ||
| Non-TransLA+INV | 55.14 | 24.19 | 31.31 | 49.74 | 70.31 | 45.54 | 33.53 | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 | 18.39 |
| TransLA+INV | 32.59 | 27.34 | 35.33 | 52.58 | 64.68 | 67.73 | 28.10 | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 | 26.89 |
| Non-TransLA+EmblEx | 60.00 | 28.91 | 42.47 | 51.99 | 61.05 | 79.12 | 50.42 | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 | 49.01 |
| TransLA + mPLMrl | 49.05 | 36.92 | 39.16 | 57.99 | 69.85 | 73.92 | 33.43 | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 | 47.44 |
| Variant | Seen Script | Unseen Script | ||||
| bh | myv | ug | bxr | am | avg | |
| Non-TransLA+INV | 28.46 / 11.53 | 45.28 / 26.27 | 33.44 / 15.28 | 39.75 / 19.77 | 19.08 / 1.85 | 33.20 / 10.81 |
| TransLA+INV | 25.12 / 10.17 | 45.74 / 26.64 | 32.30 / 15.10 | 37.92 / 17.23 | 36.07 / 7.58 | 35.43 / 12.41 |
| Non-TransLA+EmblEx | 26.68 / 10.10 | 48.34 / 25.34 | 41.20 / 22.81 | 39.51 / 16.02 | 36.4778.39 | 38.44 / 12.20 |
| TransLA + mPLMrl | 28.04 / 11.13 | 41.97 / 20.29 | 50.89 / 16.56 | 35.03 / 20.29 | 39.10 / 9.00 | 39.01 / 14.65 |
| Transliterator | Used for languages | Available at |
| UROMAN | All | github.com/isi-nlp/uroman |
| amseg | am | pypi.org/project/amseg/ |
| transliterate | ru, ka | pypi.org/project/transliterate/ |
| ai4bharat-transliteration | hi, sd | pypi.org/project/ai4bharat-transliteration/ |
| lang-trans | ar | pypi.org/project/lang-trans/ |
| Variant | Unseen Scripts | ||||
| am | bo | dv | km | si | |
| mBERT | 0.91 | 17.40 | 1.30 | 10.71 | 2.50 |
| Non-TransLA+INV | 3.26 | 19.86 | 18.72 | 13.81 | 23.14 |
| TransLA + INV | 25.69 | 35.95 | 29.99 | 41.76 | 31.83 |
| Non-TransLA+EmblEx | 47.60 | 40.96 | 31.21 | 53.94 | 45.89 |
| TransLA + mPLMft | 37.09 | 33.82 | 40.40 | 52.39 | 45.24 |
| Method | Input | Model | Output | Consistency | Factuality |
| Self-consistency | Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | ✓ | ✗ |
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | |||
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number. It can be divided evenly by 7 and 19." | |||
| Cross-Q consistency | Q1: "Is the number 3691 prime?" | Target | "Yes, the number 3691 is prime." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Target | "Yes, the only factors of 3691 are itself and one." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Target | "Yes, 3691 is a prime number and can only be divided by 1 and 3691." | |||
| Cross-M consistency | Q: "Is 3691 a prime number?" | Verifier | "Yes, 3691 is a prime number." | ✗ | ✓ |
| Cross-M & Q consistency | Q1: "Is the number 3691 prime?" | Verifier | "Yes." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Verifier | "Yes, the only factors of 3691 are 1 and itself." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Verifier | "Yes, 3691 can only be divided by 1 and 3691." |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 65.9 | 67.5 | 56.1 | 53.1 |
| SAC3-Q(gpt-3.5-turbo) | 99.4 | 99.7 | 99.7 | 99.7 |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 48.2 | 51.0 | 29.6 | 30.6 |
| SAC3-Q(gpt-3.5-turbo) | 93.2 | 96.4 | 97.0 | 97.4 |
| SAC3-QM (Falcon-7b) | 89.8 | 91.2 | 21.0 | 21.8 |
| SAC3-QM (Guanaco-33b) | 94.4 | 96.3 | 45.6 | 46.2 |
| SAC3-all (Falcon-7b) | 97.8 | 98.0 | 84.6 | 85.6 |
| SAC3-all (Guanaco-33b) | 99.4 | 99.4 | 85.6 | 86.7 |
| Method | HotpotQA-halu | NQ-open-halu | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 74.2 | 77.0 | 70.5 | 72.7 |
| SAC3-Q(gpt-3.5-turbo) | 81.3 | 81.4 | 77.2 | 78.5 |
| SAC3-QM (Falcon-7b) | 83.0 | 79.5 | 67.5 | 62.0 |
| SAC3-QM (Guanaco-33b) | 88.0 | 85.2 | 72.7 | 72.7 |
| SAC3-all (Falcon-7b) | 84.5 | 84.5 | 77.1 | 77.2 |
| SAC3-all (Guanaco-33b) | 87.0 | 86.8 | 77.2 | 77.8 |
| Method | Prime Number | Senator Search | HotpotQA -halu | NQ-open -halu |
| SC2(gpt-3.5-turbo) | 48.2 | 29.6 | 74.2 | 70.5 |
| SC2(gpt-4) | 38.3 | 18.4 | 79.7 | 76.3 |
| SC2(chat-bison) | 26.9 | 19.2 | 75.8 | 67.9 |
| SAC3-Q (gpt-3.5-turbo) | 93.2 | 97.0 | 81.3 | 77.2 |
| SAC3-Q (gpt-4) | 91.1 | 61.6 | 87.2 | 82.9 |
| SAC3-Q (chat-bison) | 90.3 | 66.3 | 82.8 | 72.7 |
| Objective | Template |
| Semantic question perturbation | For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \\nAre the following two inputs semantically equivalent? \\n[QUERYED INPUT] \\n[GENERATED INPUT] |
| Semantic equivalence check | Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \\nGuess: <most likely guess, as short as possible; not a complete sentence, just the guess!>\\nProbability: <the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \\nThe first QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} \\nThe second QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} |
| Method | Input | Model | Output | Consistency | Factuality |
| Self-consistency | Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | ✓ | ✗ |
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | |||
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number. It can be divided evenly by 7 and 19." | |||
| Cross-Q consistency | Q1: "Is the number 3691 prime?" | Target | "Yes, the number 3691 is prime." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Target | "Yes, the only factors of 3691 are itself and one." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Target | "Yes, 3691 is a prime number and can only be divided by 1 and 3691." | |||
| Cross-M consistency | Q: "Is 3691 a prime number?" | Verifier | "Yes, 3691 is a prime number." | ✗ | ✓ |
| Cross-M & Q consistency | Q1: "Is the number 3691 prime?" | Verifier | "Yes." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Verifier | "Yes, the only factors of 3691 are 1 and itself." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Verifier | "Yes, 3691 can only be divided by 1 and 3691." |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 65.9 | 67.5 | 56.1 | 53.1 |
| SAC3-Q(gpt-3.5-turbo) | 99.4 | 99.7 | 99.7 | 99.7 |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 48.2 | 51.0 | 29.6 | 30.6 |
| SAC3-Q(gpt-3.5-turbo) | 93.2 | 96.4 | 97.0 | 97.4 |
| SAC3-QM (Falcon-7b) | 89.8 | 91.2 | 21.0 | 21.8 |
| SAC3-QM (Guanaco-33b) | 94.4 | 96.3 | 45.6 | 46.2 |
| SAC3-all (Falcon-7b) | 97.8 | 98.0 | 84.6 | 85.6 |
| SAC3-all (Guanaco-33b) | 99.4 | 99.4 | 85.6 | 86.7 |
| Method | HotpotQA-halu | NQ-open-halu | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 74.2 | 77.0 | 70.5 | 72.7 |
| SAC3-Q(gpt-3.5-turbo) | 81.3 | 81.4 | 77.2 | 78.5 |
| SAC3-QM (Falcon-7b) | 83.0 | 79.5 | 67.5 | 62.0 |
| SAC3-QM (Guanaco-33b) | 88.0 | 85.2 | 72.7 | 72.7 |
| SAC3-all (Falcon-7b) | 84.5 | 84.5 | 77.1 | 77.2 |
| SAC3-all (Guanaco-33b) | 87.0 | 86.8 | 77.2 | 77.8 |
| Method | Prime Number | Senator Search | HotpotQA -halu | NQ-open -halu |
| SC2(gpt-3.5-turbo) | 48.2 | 29.6 | 74.2 | 70.5 |
| SC2(gpt-4) | 38.3 | 18.4 | 79.7 | 76.3 |
| SC2(chat-bison) | 26.9 | 19.2 | 75.8 | 67.9 |
| SAC3-Q (gpt-3.5-turbo) | 93.2 | 97.0 | 81.3 | 77.2 |
| SAC3-Q (gpt-4) | 91.1 | 61.6 | 87.2 | 82.9 |
| SAC3-Q (chat-bison) | 90.3 | 66.3 | 82.8 | 72.7 |
| Objective | Template |
| Semantic question perturbation | For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \\nAre the following two inputs semantically equivalent? \\n[QUERYED INPUT] \\n[GENERATED INPUT] |
| Semantic equivalence check | Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \\nGuess: <most likely guess, as short as possible; not a complete sentence, just the guess!>\\nProbability: <the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \\nThe first QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} \\nThe second QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} |
| Method | Input | Model | Output | Consistency | Factuality |
| Self-consistency | Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | ✓ | ✗ |
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | |||
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number. It can be divided evenly by 7 and 19." | |||
| Cross-Q consistency | Q1: "Is the number 3691 prime?" | Target | "Yes, the number 3691 is prime." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Target | "Yes, the only factors of 3691 are itself and one." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Target | "Yes, 3691 is a prime number and can only be divided by 1 and 3691." | |||
| Cross-M consistency | Q: "Is 3691 a prime number?" | Verifier | "Yes, 3691 is a prime number." | ✗ | ✓ |
| Cross-M & Q consistency | Q1: "Is the number 3691 prime?" | Verifier | "Yes." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Verifier | "Yes, the only factors of 3691 are 1 and itself." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Verifier | "Yes, 3691 can only be divided by 1 and 3691." |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 65.9 | 67.5 | 56.1 | 53.1 |
| SAC3-Q(gpt-3.5-turbo) | 99.4 | 99.7 | 99.7 | 99.7 |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 48.2 | 51.0 | 29.6 | 30.6 |
| SAC3-Q(gpt-3.5-turbo) | 93.2 | 96.4 | 97.0 | 97.4 |
| SAC3-QM (Falcon-7b) | 89.8 | 91.2 | 21.0 | 21.8 |
| SAC3-QM (Guanaco-33b) | 94.4 | 96.3 | 45.6 | 46.2 |
| SAC3-all (Falcon-7b) | 97.8 | 98.0 | 84.6 | 85.6 |
| SAC3-all (Guanaco-33b) | 99.4 | 99.4 | 85.6 | 86.7 |
| Method | HotpotQA-halu | NQ-open-halu | |||
| Score | Confidence | Score | Confidence | ||
| SC2(gpt-3.5-turbo) | 74.2 | 77.0 | 70.5 | 72.7 | |
| SAC3-Q(gpt-3.5-turbo) | 81.3 | 81.4 | 77.2 | 78.5 | |
| SAC3-QM (Falcon-7b) | 83.0 | 79.5 | 67.5 | 62.0 | |
| SAC3-QM (Guanaco-33b) | 88.0 | 85.2 | 72.7 | 72.7 | |
| SAC3-all (Falcon-7b) | 84.5 | 84.5 | 77.1 | 77.2 | |
| SAC3-all (Guanaco-33b) | 87.0 | 86.8 | 77.2 | 77.8 | |
| Method | Prime Number | Senator Search | HotpotQA -halu | NQ-open -halu |
| SC2(gpt-3.5-turbo) | 48.2 | 29.6 | 74.2 | 70.5 |
| SC2(gpt-4) | 38.3 | 18.4 | 79.7 | 76.3 |
| SC2(chat-bison) | 26.9 | 19.2 | 75.8 | 67.9 |
| SAC3-Q (gpt-3.5-turbo) | 93.2 | 97.0 | 81.3 | 77.2 |
| SAC3-Q (gpt-4) | 91.1 | 61.6 | 87.2 | 82.9 |
| SAC3-Q (chat-bison) | 90.3 | 66.3 | 82.8 | 72.7 |
| Objective | Template |
| Semantic question perturbation | For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \nAre the following two inputs semantically equivalent? \n[QUERYED INPUT] \n[GENERATED INPUT] |
| Semantic equivalence check | Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \nGuess: <most likely guess, as short as possible; not a complete sentence, just the guess!>\nProbability: <the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \nThe first QA pair is: \nQ: ${THE QUESTION} \nA: ${THE ANSWER} \nThe second QA pair is: \nQ: ${THE QUESTION} \nA: ${THE ANSWER} |
| Method | Input | Model | Output | Consistency | Factuality |
| Self-consistency | Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | ✓ | ✗ |
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number as it is divisible by 7 and 13." | |||
| Q: "Is 3691 a prime number?" | Target | "No, 3691 is not a prime number. It can be divided evenly by 7 and 19." | |||
| Cross-Q consistency | Q1: "Is the number 3691 prime?" | Target | "Yes, the number 3691 is prime." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Target | "Yes, the only factors of 3691 are itself and one." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Target | "Yes, 3691 is a prime number and can only be divided by 1 and 3691." | |||
| Cross-M consistency | Q: "Is 3691 a prime number?" | Verifier | "Yes, 3691 is a prime number." | ✗ | ✓ |
| Cross-M & Q consistency | Q1: "Is the number 3691 prime?" | Verifier | "Yes." | ✗ | ✓ |
| Q2: "Are the only factors of 3691 itself and one?" | Verifier | "Yes, the only factors of 3691 are 1 and itself." | |||
| Q3: "Can 3691 only be divided by 1 and 3691?" | Verifier | "Yes, 3691 can only be divided by 1 and 3691." |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 65.9 | 67.5 | 56.1 | 53.1 |
| SAC3-Q(gpt-3.5-turbo) | 99.4 | 99.7 | 99.7 | 99.7 |
| Method | Prime number | Senator search | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 48.2 | 51.0 | 29.6 | 30.6 |
| SAC3-Q(gpt-3.5-turbo) | 93.2 | 96.4 | 97.0 | 97.4 |
| SAC3-QM (Falcon-7b) | 89.8 | 91.2 | 21.0 | 21.8 |
| SAC3-QM (Guanaco-33b) | 94.4 | 96.3 | 45.6 | 46.2 |
| SAC3-all (Falcon-7b) | 97.8 | 98.0 | 84.6 | 85.6 |
| SAC3-all (Guanaco-33b) | 99.4 | 99.4 | 85.6 | 86.7 |
| Method | HotpotQA-halu | NQ-open-halu | ||
| Score | Confidence | Score | Confidence | |
| SC2(gpt-3.5-turbo) | 74.2 | 77.0 | 70.5 | 72.7 |
| SAC3-Q(gpt-3.5-turbo) | 81.3 | 81.4 | 77.2 | 78.5 |
| SAC3-QM (Falcon-7b) | 83.0 | 79.5 | 67.5 | 62.0 |
| SAC3-QM (Guanaco-33b) | 88.0 | 85.2 | 72.7 | 72.7 |
| SAC3-all (Falcon-7b) | 84.5 | 84.5 | 77.1 | 77.2 |
| SAC3-all (Guanaco-33b) | 87.0 | 86.8 | 77.2 | 77.8 |
| Method | Prime Number | Senator Search | HotpotQA -halu | NQ-open -halu |
| SC2(gpt-3.5-turbo) | 48.2 | 29.6 | 74.2 | 70.5 |
| SC2(gpt-4) | 38.3 | 18.4 | 79.7 | 76.3 |
| SC2(chat-bison) | 26.9 | 19.2 | 75.8 | 67.9 |
| SAC3-Q (gpt-3.5-turbo) | 93.2 | 97.0 | 81.3 | 77.2 |
| SAC3-Q (gpt-4) | 91.1 | 61.6 | 87.2 | 82.9 |
| SAC3-Q (chat-bison) | 90.3 | 66.3 | 82.8 | 72.7 |
| Objective | Template |
| Semantic question perturbation | For the question [QUERYED QUESTION], provide ${k} semantically equivalent questions \\nAre the following two inputs semantically equivalent? \\n[QUERYED INPUT] \\n[GENERATED INPUT] |
| Semantic equivalence check | Are the following two Question-Answering (QA) pairs semantically equivalent? Provide your best guess and the probability that it is correct (0.0 to 1.0). Given ONLY the guess (Yes or No) and probability, no other words or explanation. For example: \\nGuess: <most likely guess, as short as possible; not a complete sentence, just the guess!>\\nProbability: <the probability between 0.0 and 1.0 that your guess is correct, without any extra commentary whatsoever; just the probability! \\nThe first QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} \\nThe second QA pair is: \\nQ: ${THE QUESTION} \\nA: ${THE ANSWER} |
| Model | Precision | Recall | F1 |
| Sentence-T5 11B | .2826 | .523 | .3669 |
| GTR XXL | .3148 | .4885 | .3829 |
| SENTLI | .4438 | .8161 | .5749 |
| SENTLI, threshold tuned | .5721 | .7299 | .6414 |
| Flan-T5 XXL 0-shot | .6867 | .6552 | .6706 |
| Flan-T5 XXL 5-shot | .6297 | .6839 | .6556 |
| Flan-UL2 0-shot | .5922 | .7011 | .6421 |
| Flan-UL2 5-shot | .5536 | .7126 | .6231 |
| Flan-T5 XXL finetuned on ANLI | .356 | .8736 | .5058 |
| Flan-T5 XXL finetuned on SDOH-NLI | .9295 | .8333 | .8788 |
| Flan-T5 XXL finetuned on ANLI + SDOH-NLI | .8765 | .8563 | .8663 |
| Model | Size | Dataset | Test P | Test R | Test F1 | New factors P | New factors R | New factors F1 |
| T5 | small | ANLI + SDOH-NLI | .7987 | .7069 | .7500 | .6667 | .0833 | .1481 |
| T5 | base | ANLI + SDOH-NLI | .7396 | .8161 | .7760 | .75 | .375 | .5 |
| T5 | large | ANLI + SDOH-NLI | .8247 | .7299 | .7744 | .75 | .125 | .2143 |
| T5 | XL | ANLI + SDOH-NLI | .8391 | .8391 | .8391 | .9231 | .5 | .6486 |
| T5 | XXL | ANLI | .8391 | .8391 | .8391 | .75 | .1 | .8571 |
| T5 | XXL | SDOH-NLI | .8859 | .7586 | .8173 | .8571 | .25 | .3871 |
| T5 | XXL | ANLI + SDOH-NLI | .9024 | .8506 | .8757 | .8889 | 1. | .9412 |
| Flan-T5 | small | ANLI + SDOH-NLI | .8075 | .7471 | .7761 | .6667 | .0833 | .1481 |
| Flan-T5 | base | ANLI + SDOH-NLI | .7956 | .8276 | .8113 | .6216 | .9583 | .7541 |
| Flan-T5 | large | ANLI | .1777 | .8506 | .2939 | .2421 | .9583 | .3866 |
| Flan-T5 | large | SDOH-NLI | .8041 | .6839 | .7391 | .2308 | .125 | .1622 |
| Flan-T5 | large | ANLI + SDOH-NLI | .8212 | .8448 | .8329 | .8846 | .9583 | .92 |
| Flan-T5 | XL | ANLI | .2868 | .8851 | .4332 | .7667 | .9583 | .8519 |
| Flan-T5 | XL | SDOH | .9071 | .7299 | .8089 | 1. | .125 | .2222 |
| Flan-T5 | XL | ANLI + SDOH-NLI | .8647 | .8448 | .8547 | .8846 | .9583 | .92 |
| Flan-T5 | XXL | ANLI | .3560 | .8736 | .5058 | .7931 | .9583 | .8679 |
| Flan-T5 | XXL | SDOH-NLI | .9295 | .8333 | .8788 | 1. | .875 | .9333 |
| Flan-T5 | XXL | ANLI + SDOH-NLI | .8765 | .8563 | .8663 | .8846 | .9583 | .92 |
| Model | Precision | Recall | F1 |
| Sentence-T5 11B | .2826 | .523 | .3669 |
| GTR XXL | .3148 | .4885 | .3829 |
| SENTLI | .4438 | .8161 | .5749 |
| SENTLI, threshold tuned | .5721 | .7299 | .6414 |
| Flan-T5 XXL 0-shot | .6867 | .6552 | .6706 |
| Flan-T5 XXL 5-shot | .6297 | .6839 | .6556 |
| Flan-UL2 0-shot | .5922 | .7011 | .6421 |
| Flan-UL2 5-shot | .5536 | .7126 | .6231 |
| Flan-T5 XXL finetuned on ANLI | .356 | .8736 | .5058 |
| Flan-T5 XXL finetuned on SDOH-NLI | .9295 | .8333 | .8788 |
| Flan-T5 XXL finetuned on ANLI + SDOH-NLI | .8765 | .8563 | .8663 |
| Model | Size | Dataset | Test P | Test R | Test F1 | New factors P | New factors R | New factors F1 |
| T5 | small | ANLI + SDOH-NLI | .7987 | .7069 | .7500 | .6667 | .0833 | .1481 |
| T5 | base | ANLI + SDOH-NLI | .7396 | .8161 | .7760 | .75 | .375 | .5 |
| T5 | large | ANLI + SDOH-NLI | .8247 | .7299 | .7744 | .75 | .125 | .2143 |
| T5 | XL | ANLI + SDOH-NLI | .8391 | .8391 | .8391 | .9231 | .5 | .6486 |
| T5 | XXL | ANLI | .8391 | .8391 | .8391 | .75 | .1 | .8571 |
| T5 | XXL | SDOH-NLI | .8859 | .7586 | .8173 | .8571 | .25 | .3871 |
| T5 | XXL | ANLI + SDOH-NLI | .9024 | .8506 | .8757 | .8889 | 1. | .9412 |
| Flan-T5 | small | ANLI + SDOH-NLI | .8075 | .7471 | .7761 | .6667 | .0833 | .1481 |
| Flan-T5 | base | ANLI + SDOH-NLI | .7956 | .8276 | .8113 | .6216 | .9583 | .7541 |
| Flan-T5 | large | ANLI | .1777 | .8506 | .2939 | .2421 | .9583 | .3866 |
| Flan-T5 | large | SDOH-NLI | .8041 | .6839 | .7391 | .2308 | .125 | .1622 |
| Flan-T5 | large | ANLI + SDOH-NLI | .8212 | .8448 | .8329 | .8846 | .9583 | .92 |
| Flan-T5 | XL | ANLI | .2868 | .8851 | .4332 | .7667 | .9583 | .8519 |
| Flan-T5 | XL | SDOH | .9071 | .7299 | .8089 | 1. | .125 | .2222 |
| Flan-T5 | XL | ANLI + SDOH-NLI | .8647 | .8448 | .8547 | .8846 | .9583 | .92 |
| Flan-T5 | XXL | ANLI | .3560 | .8736 | .5058 | .7931 | .9583 | .8679 |
| Flan-T5 | XXL | SDOH-NLI | .9295 | .8333 | .8788 | 1. | .875 | .9333 |
| Flan-T5 | XXL | ANLI + SDOH-NLI | .8765 | .8563 | .8663 | .8846 | .9583 | .92 |
| Model | Precision | Recall | F1 |
| Sentence-T5 11B | .2826 | .523 | .3669 |
| GTR XXL | .3148 | .4885 | .3829 |
| SENTLI | .4438 | .8161 | .5749 |
| SENTLI, threshold tuned | .5721 | .7299 | .6414 |
| Flan-T5 XXL 0-shot | .6867 | .6552 | .6706 |
| Flan-T5 XXL 5-shot | .6297 | .6839 | .6556 |
| Flan-UL2 0-shot | .5922 | .7011 | .6421 |
| Flan-UL2 5-shot | .5536 | .7126 | .6231 |
| Flan-T5 XXL finetuned on ANLI | .356 | .8736 | .5058 |
| Flan-T5 XXL finetuned on SDOH-NLI | .9295 | .8333 | .8788 |
| Flan-T5 XXL finetuned on ANLI + SDOH-NLI | .8765 | .8563 | .8663 |
| Model | Size | Dataset | Test P | Test R | Test F1 | New factors P | New factors R | New factors F1 |
| T5 | small | ANLI + SDOH-NLI | .7987 | .7069 | .7500 | .6667 | .0833 | .1481 |
| T5 | base | ANLI + SDOH-NLI | .7396 | .8161 | .7760 | .75 | .375 | .5 |
| T5 | large | ANLI + SDOH-NLI | .8247 | .7299 | .7744 | .75 | .125 | .2143 |
| T5 | XL | ANLI + SDOH-NLI | .8391 | .8391 | .8391 | .9231 | .5 | .6486 |
| T5 | XXL | ANLI | .8391 | .8391 | .8391 | .75 | .1 | .8571 |
| T5 | XXL | SDOH-NLI | .8859 | .7586 | .8173 | .8571 | .25 | .3871 |
| T5 | XXL | ANLI + SDOH-NLI | .9024 | .8506 | .8757 | .8889 | 1. | .9412 |
| Flan-T5 | small | ANLI + SDOH-NLI | .8075 | .7471 | .7761 | .6667 | .0833 | .1481 |
| Flan-T5 | base | ANLI + SDOH-NLI | .7956 | .8276 | .8113 | .6216 | .9583 | .7541 |
| Flan-T5 | large | ANLI | .1777 | .8506 | .2939 | .2421 | .9583 | .3866 |
| Flan-T5 | large | SDOH-NLI | .8041 | .6839 | .7391 | .2308 | .125 | .1622 |
| Flan-T5 | large | ANLI + SDOH-NLI | .8212 | .8448 | .8329 | .8846 | .9583 | .92 |
| Flan-T5 | XL | ANLI | .2868 | .8851 | .4332 | .7667 | .9583 | .8519 |
| Flan-T5 | XL | SDOH | .9071 | .7299 | .8089 | 1. | .125 | .2222 |
| Flan-T5 | XL | ANLI + SDOH-NLI | .8647 | .8448 | .8547 | .8846 | .9583 | .92 |
| Flan-T5 | XXL | ANLI | .3560 | .8736 | .5058 | .7931 | .9583 | .8679 |
| Flan-T5 | XXL | SDOH-NLI | .9295 | .8333 | .8788 | 1. | .875 | .9333 |
| Flan-T5 | XXL | ANLI + SDOH-NLI | .8765 | .8563 | .8663 | .8846 | .9583 | .92 |
| Model | Precision | Recall | F1 |
| Sentence-T5 11B | .2826 | .523 | .3669 |
| GTR XXL | .3148 | .4885 | .3829 |
| SENTLI | .4438 | .8161 | .5749 |
| SENTLI, threshold tuned | .5721 | .7299 | .6414 |
| Flan-T5 XXL 0-shot | .6867 | .6552 | .6706 |
| Flan-T5 XXL 5-shot | .6297 | .6839 | .6556 |
| Flan-UL2 0-shot | .5922 | .7011 | .6421 |
| Flan-UL2 5-shot | .5536 | .7126 | .6231 |
| Flan-T5 XXL finetuned on ANLI | .356 | .8736 | .5058 |
| Flan-T5 XXL finetuned on SDOH-NLI | .9295 | .8333 | .8788 |
| Flan-T5 XXL finetuned on ANLI + SDOH-NLI | .8765 | .8563 | .8663 |
| Model | Size | Dataset | Test P | Test R | Test F1 | New factors P | New factors R | New factors F1 |
| T5 | small | ANLI + SDOH-NLI | .7987 | .7069 | .7500 | .6667 | .0833 | .1481 |
| T5 | base | ANLI + SDOH-NLI | .7396 | .8161 | .7760 | .75 | .375 | .5 |
| T5 | large | ANLI + SDOH-NLI | .8247 | .7299 | .7744 | .75 | .125 | .2143 |
| T5 | XL | ANLI + SDOH-NLI | .8391 | .8391 | .8391 | .9231 | .5 | .6486 |
| T5 | XXL | ANLI | .8391 | .8391 | .8391 | .75 | .1 | .8571 |
| T5 | XXL | SDOH-NLI | .8859 | .7586 | .8173 | .8571 | .25 | .3871 |
| T5 | XXL | ANLI + SDOH-NLI | .9024 | .8506 | .8757 | .8889 | 1. | .9412 |
| Flan-T5 | small | ANLI + SDOH-NLI | .8075 | .7471 | .7761 | .6667 | .0833 | .1481 |
| Flan-T5 | base | ANLI + SDOH-NLI | .7956 | .8276 | .8113 | .6216 | .9583 | .7541 |
| Flan-T5 | large | ANLI | .1777 | .8506 | .2939 | .2421 | .9583 | .3866 |
| Flan-T5 | large | SDOH-NLI | .8041 | .6839 | .7391 | .2308 | .125 | .1622 |
| Flan-T5 | large | ANLI + SDOH-NLI | .8212 | .8448 | .8329 | .8846 | .9583 | .92 |
| Flan-T5 | XL | ANLI | .2868 | .8851 | .4332 | .7667 | .9583 | .8519 |
| Flan-T5 | XL | SDOH | .9071 | .7299 | .8089 | 1. | .125 | .2222 |
| Flan-T5 | XL | ANLI + SDOH-NLI | .8647 | .8448 | .8547 | .8846 | .9583 | .92 |
| Flan-T5 | XXL | ANLI | .3560 | .8736 | .5058 | .7931 | .9583 | .8679 |
| Flan-T5 | XXL | SDOH-NLI | .9295 | .8333 | .8788 | 1. | .875 | .9333 |
| Flan-T5 | XXL | ANLI + SDOH-NLI | .8765 | .8563 | .8663 | .8846 | .9583 | .92 |
| Dataset | Domain | Criteria | #Docs | #labels |
| NYT | News | Topic | 13081 | 26 |
| Yelp | Reviews | Sentiment | 70000 | 5 |
| Clima | Climate | Question Type | 17175 | 8 |
| TREC | General | Question Type | 5952 | 6 |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| Yelp | NYT | LogitNorm | 82.0 | 33.2 | 69.4 | 57.9 | 84.4 | 33.9 | 66.6 | 55.0 |
| OneVsRest | 85.1 | 35.8 | 62.4 | 49.3 | 88.5 | 29.6 | 69.8 | 56.5 | ||
| RCL | 56.9 | 31.2 | 73.2 | 55.2 | 87.7 | 34.5 | 56.7 | 45.0 | ||
| Bayes Approx | 64.3 | 35.9 | 66.5 | 50.3 | 79.6 | 36.4 | 54.1 | 40.8 | ||
| CE Loss | 82.1 | 34.9 | 63.4 | 50.7 | 79.8 | 32.9 | 66.8 | 58.0 | ||
| CRL | 99.9 | 36.9 | 43.9 | 51.1 | 99.8 | 28.1 | 64.8 | 60.5 | ||
| SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 | 69.5 | 21.6 | 72.6 | 78.8 | ||
| Clima | LogitNorm | 79.2 | 25.5 | 72.5 | 54.8 | 82.6 | 25.4 | 70.7 | 54.0 | |
| OneVsRest | 83.0 | 25.1 | 69.5 | 53.2 | 86.7 | 28.0 | 53.8 | 33.7 | ||
| RCL | 59.9 | 27.8 | 67.4 | 42.1 | 73.9 | 22.5 | 70.7 | 56.3 | ||
| Bayes Approx | 74.7 | 28.7 | 56.7 | 33.6 | 63.0 | 28.0 | 66.8 | 40.9 | ||
| CE Loss | 83.2 | 27.8 | 48.6 | 37.0 | 76.4 | 22.0 | 78.2 | 68.4 | ||
| CRL | 99.6 | 29.0 | 49.1 | 38.8 | 99.9 | 20.6 | 63.0 | 56.9 | ||
| SELFOOD | 17.6 | 4.8 | 97.9 | 96.4 | 65.6 | 16.6 | 73.2 | 77.2 | ||
| TREC | LogitNorm | 68.2 | 27.0 | 79.8 | 82.1 | 60.8 | 23.0 | 83.9 | 84.4 | |
| OneVsRest | 84.8 | 40.0 | 63.9 | 69.7 | 75.5 | 33.0 | 57.9 | 61.2 | ||
| RCL | 89.8 | 43.2 | 40.1 | 51.5 | 77.4 | 32.4 | 65.0 | 68.6 | ||
| Bayes Approx | 92.5 | 44.7 | 32.1 | 44.7 | 51.8 | 25.9 | 72.7 | 69.4 | ||
| CE Loss | 74.2 | 33.7 | 52.8 | 62.7 | 58.5 | 27.5 | 75.3 | 78.3 | ||
| CRL | 33.3 | 15.5 | 66.7 | 77.2 | 100.0 | 37.3 | 56.7 | 73.9 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 55.1 | 20.6 | 80.2 | 90.3 | ||
| TREC | Yelp | LogitNorm | 32.3 | 0.3 | 95.5 | 78.1 | 15.2 | 0.5 | 95.7 | 44.7 |
| OneVsRest | 4.2 | 0.3 | 97.9 | 83.4 | 39.3 | 0.4 | 91.6 | 63.0 | ||
| RCL | 78.5 | 0.6 | 80.2 | 20.2 | 67.7 | 0.6 | 83.0 | 24.2 | ||
| Bayes Approx | 46.6 | 0.4 | 93.0 | 61.6 | 63.7 | 0.7 | 87.7 | 21.3 | ||
| CE Loss | 27.2 | 0.3 | 94.9 | 71.2 | 11.3 | 0.4 | 96.9 | 62.3 | ||
| CRL | 96.6 | 0.7 | 70.9 | 16.5 | 99.8 | 0.7 | 24.3 | 9.7 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| NYT | LogitNorm | 9.3 | 1.0 | 97.7 | 91.2 | 10.9 | 2.3 | 97.2 | 77.3 | |
| OneVsRest | 1.8 | 0.5 | 99.0 | 95.8 | 32.5 | 1.8 | 93.2 | 74.8 | ||
| RCL | 63.6 | 3.3 | 84.1 | 45.7 | 74.2 | 3.0 | 80.4 | 45.1 | ||
| Bayes Approx | 39.1 | 2.5 | 93.5 | 68.1 | 64.9 | 3.0 | 88.4 | 54.4 | ||
| CE Loss | 8.4 | 1.3 | 97.6 | 87.5 | 6.8 | 1.4 | 98.1 | 87.1 | ||
| CRL | 97.2 | 3.3 | 73.0 | 35.2 | 99.9 | 3.8 | 22.9 | 12.2 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Clima | LogitNorm | 11.3 | 0.8 | 97.6 | 88.5 | 10.4 | 1.6 | 97.3 | 75.9 | |
| OneVsRest | 3.4 | 0.6 | 98.7 | 93.2 | 36.3 | 1.8 | 91.7 | 60.1 | ||
| RCL | 73.5 | 2.4 | 81.2 | 37.1 | 60.7 | 2.0 | 86.4 | 48.2 | ||
| Bayes Approx | 51.7 | 2.0 | 90.5 | 54.4 | 61.1 | 2.4 | 87.6 | 40.4 | ||
| CE Loss | 15.1 | 1.7 | 96.2 | 69.5 | 12.2 | 1.8 | 96.4 | 67.8 | ||
| CRL | 86.3 | 2.2 | 80.7 | 39.5 | 98.9 | 2.7 | 30.1 | 13.0 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 |
| Method | FPR95 | ERR\nIn Dist: Clima\nOOD: NYT | AUROC | AUPR |
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 |
| + Grad Sub | 67.9 | 8.9 | 82.4 | 61.1 |
| + Grad Sub & Min | 98.4 | 10.7 | 49.2 | 32.3 |
| + Intra-Doc | 92.7 | 10.6 | 67.3 | 41.6 |
| - SILu | 14.6 | 3.9 | 96.7 | 89.3 |
| In Dist: NYT\nOOD: Clima | ||||
| SELFOOD | 10.5 | 1.7 | 98.4 | 90.5 |
| + Grad Sub only | 83.5 | 5.1 | 77.4 | 44.8 |
| + Grad Sub & Min | 96.5 | 5.2 | 60.3 | 36.0 |
| + Intra-Doc | 48.3 | 2.6 | 93.2 | 80.2 |
| - SiLU | 19.0 | 2.3 | 97.4 | 86.0 |
| OOD | Method | FPR95 | ERR | AUROC | AUPR |
| AGNews | LogitNorm | 45.9 | 0.9 | 86.7 | 15.8 |
| Bayes Approx | 57.5 | 0.9 | 87.7 | 23.9 | |
| CE-Loss | 45.5 | 0.9 | 86.5 | 16.6 | |
| CRL | 77.2 | 0.9 | 80.6 | 9.5 | |
| SELFOOD | 0.0 | 0.1 | 100.0 | 99.4 | |
| 20News | LogitNorm | 24.5 | 6.2 | 91.7 | 39.2 |
| Bayes Approx | 35.0 | 4.3 | 93.7 | 66.6 | |
| CE-Loss | 25.4 | 6.2 | 91.9 | 39.6 | |
| CRL | 59.9 | 6.4 | 87.4 | 36.9 | |
| SELFOOD | 28.2 | 4.9 | 94.2 | 61.5 |
| InD | OOD | Method | FPR95 | ERR | AUROC | AUPR |
| Clima | Yelp | SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 |
| SELFOOD + Maha | 0.8 | 0.2 | 99.2 | 97.2 | ||
| Yelp | NYT | SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 |
| SELFOOD + Maha | 12.4 | 4.4 | 97.4 | 97.5 |
| Text | CE-Loss | SELFOOD |
| Who said “What contemptible scoundrel stole the cork from my lunch?” | InD | OOD |
| Who said “Give me liberty or give me death”? | InD | OOD |
| Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next "place"... | OOD | InD |
| the pitch reads: "what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?" And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ... | OOD | InD |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| NYT | Clima | LogitNorm | 20.6 | 1.0 | 96.8 | 92.3 | 18.4 | 1.0 | 97.3 | 93.6 |
| OneVsRest | 84.7 | 1.7 | 87.3 | 82.6 | 92.7 | 1.9 | 91.1 | 85.1 | ||
| RCL | 46.9 | 1.5 | 94.4 | 87.5 | 86.7 | 2.3 | 86.7 | 76.6 | ||
| Bayes Approx | 40.6 | 1.2 | 95.4 | 90.0 | 80.0 | 2.4 | 88.4 | 76.5 | ||
| CE Loss | 22.1 | 1.0 | 96.9 | 92.5 | 15.7 | 1.0 | 97.3 | 92.8 | ||
| CRL | 32.3 | 1.7 | 95.4 | 86.5 | 79.3 | 3.2 | 79.8 | 67.0 | ||
| SELFOOD | 10.5 | 1.6 | 98.4 | 90.5 | 4.6 | 1.0 | 99.3 | 95.4 | ||
| TREC | LogitNorm | 13.3 | 5.5 | 97.2 | 87.9 | 17.5 | 6.4 | 96.1 | 84.1 | |
| OneVsRest | 96.0 | 6.2 | 84.0 | 79.4 | 97.2 | 6.3 | 90.5 | 82.8 | ||
| RCL | 62.9 | 6.1 | 92.2 | 85.8 | 73.9 | 8.8 | 88.4 | 73.1 | ||
| Bayes Approx | 38.8 | 4.5 | 94.8 | 89.9 | 67.5 | 8.3 | 88.7 | 74.7 | ||
| CE Loss | 14.5 | 5.8 | 97.0 | 87.1 | 21.9 | 6.7 | 95.5 | 83.5 | ||
| CRL | 67.9 | 6.5 | 89.9 | 79.1 | 85.6 | 9.6 | 71.4 | 61.4 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Yelp | LogitNorm | 1.7 | 0.6 | 99.3 | 87.9 | 1.2 | 0.6 | 98.5 | 87.9 | |
| OneVsRest | 15.1 | 0.7 | 97.4 | 78.1 | 8.7 | 0.6 | 97.6 | 78.8 | ||
| RCL | 45.1 | 0.5 | 94.9 | 84.7 | 76.7 | 1.0 | 90.2 | 60.0 | ||
| Bayes Approx | 28.7 | 0.4 | 96.4 | 86.8 | 83.0 | 1.1 | 87.7 | 56.6 | ||
| CE Loss | 3.4 | 0.7 | 99.2 | 82.3 | 1.8 | 0.6 | 99.3 | 85.4 | ||
| CRL | 69.5 | 0.5 | 89.3 | 77.8 | 73.3 | 1.0 | 83.8 | 58.3 | ||
| SELFOOD | 22.5 | 1.1 | 95.7 | 57.8 | 0.2 | 0.2 | 99.9 | 97.5 | ||
| Clima | TREC | LogitNorm | 37.8 | 5.5 | 94.6 | 91.9 | 60.5 | 7.0 | 91.6 | 88.0 |
| OneVsRest | 25.5 | 6.8 | 95.5 | 91.8 | 54.3 | 15.5 | 84.5 | 67.7 | ||
| RCL | 80.1 | 18.4 | 75.7 | 55.3 | 91.2 | 18.4 | 66.6 | 47.7 | ||
| Bayes Approx | 83.7 | 15.8 | 77.5 | 63.8 | 65.7 | 13.9 | 83.6 | 70.7 | ||
| CE Loss | 34.9 | 5.6 | 94.9 | 92.2 | 60.2 | 9.1 | 90.4 | 83.8 | ||
| CRL | 83.0 | 12.7 | 82.0 | 72.2 | 60.9 | 12.6 | 79.2 | 63.2 | ||
| SELFOOD | 0.3 | 1.1 | 99.4 | 99.4 | 38.9 | 8.8 | 90.1 | 84.7 | ||
| Yelp | LogitNorm | 57.7 | 0.8 | 93.0 | 81.3 | 54.7 | 1.2 | 92.6 | 72.3 | |
| OneVsRest | 43.7 | 1.0 | 92.7 | 72.5 | 58.8 | 1.8 | 73.5 | 34.7 | ||
| RCL | 91.0 | 2.3 | 60.7 | 12.2 | 98.0 | 2.3 | 56.4 | 12.7 | ||
| Bayes Approx | 81.4 | 2.2 | 76.5 | 22.7 | 96.8 | 2.2 | 52.3 | 16.3 | ||
| CE Loss | 46.3 | 0.9 | 93.8 | 80.1 | 79.5 | 1.9 | 86.5 | 47.2 | ||
| CRL | 98.7 | 1.4 | 67.3 | 53.3 | 72.9 | 2.1 | 50.9 | 19.0 | ||
| SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 | 37.6 | 1.5 | 72.3 | 54.9 | ||
| NYT | LogitNorm | 37.6 | 4.2 | 94.4 | 87.3 | 46.1 | 4.5 | 93.1 | 85.4 | |
| OneVsRest | 34.6 | 4.8 | 94.3 | 85.3 | 40.6 | 9.4 | 83.1 | 51.3 | ||
| RCL | 90.3 | 12.1 | 64.6 | 28.8 | 97.9 | 12.4 | 51.5 | 21.2 | ||
| Bayes Approx | 87.4 | 11.2 | 70.4 | 38.3 | 96.8 | 11.3 | 55.8 | 29.2 | ||
| CE Loss | 44.9 | 4.8 | 93.2 | 84.4 | 62.5 | 6.7 | 89.6 | 75.2 | ||
| CRL | 85.6 | 7.6 | 78.0 | 62.5 | 61.3 | 9.4 | 61.8 | 39.9 | ||
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 | 55.3 | 6.7 | 68.6 | 61.5 |
| Dataset | Domain | Criteria | #Docs | #labels |
| NYT | News | Topic | 13081 | 26 |
| Yelp | Reviews | Sentiment | 70000 | 5 |
| Clima | Climate | Question Type | 17175 | 8 |
| TREC | General | Question Type | 5952 | 6 |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| Yelp | NYT | LogitNorm | 82.0 | 33.2 | 69.4 | 57.9 | 84.4 | 33.9 | 66.6 | 55.0 |
| OneVsRest | 85.1 | 35.8 | 62.4 | 49.3 | 88.5 | 29.6 | 69.8 | 56.5 | ||
| RCL | 56.9 | 31.2 | 73.2 | 55.2 | 87.7 | 34.5 | 56.7 | 45.0 | ||
| Bayes Approx | 64.3 | 35.9 | 66.5 | 50.3 | 79.6 | 36.4 | 54.1 | 40.8 | ||
| CE Loss | 82.1 | 34.9 | 63.4 | 50.7 | 79.8 | 32.9 | 66.8 | 58.0 | ||
| CRL | 99.9 | 36.9 | 43.9 | 51.1 | 99.8 | 28.1 | 64.8 | 60.5 | ||
| SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 | 69.5 | 21.6 | 72.6 | 78.8 | ||
| Clima | LogitNorm | 79.2 | 25.5 | 72.5 | 54.8 | 82.6 | 25.4 | 70.7 | 54.0 | |
| OneVsRest | 83.0 | 25.1 | 69.5 | 53.2 | 86.7 | 28.0 | 53.8 | 33.7 | ||
| RCL | 59.9 | 27.8 | 67.4 | 42.1 | 73.9 | 22.5 | 70.7 | 56.3 | ||
| Bayes Approx | 74.7 | 28.7 | 56.7 | 33.6 | 63.0 | 28.0 | 66.8 | 40.9 | ||
| CE Loss | 83.2 | 27.8 | 48.6 | 37.0 | 76.4 | 22.0 | 78.2 | 68.4 | ||
| CRL | 99.6 | 29.0 | 49.1 | 38.8 | 99.9 | 20.6 | 63.0 | 56.9 | ||
| SELFOOD | 17.6 | 4.8 | 97.9 | 96.4 | 65.6 | 16.6 | 73.2 | 77.2 | ||
| TREC | LogitNorm | 68.2 | 27.0 | 79.8 | 82.1 | 60.8 | 23.0 | 83.9 | 84.4 | |
| OneVsRest | 84.8 | 40.0 | 63.9 | 69.7 | 75.5 | 33.0 | 57.9 | 61.2 | ||
| RCL | 89.8 | 43.2 | 40.1 | 51.5 | 77.4 | 32.4 | 65.0 | 68.6 | ||
| Bayes Approx | 92.5 | 44.7 | 32.1 | 44.7 | 51.8 | 25.9 | 72.7 | 69.4 | ||
| CE Loss | 74.2 | 33.7 | 52.8 | 62.7 | 58.5 | 27.5 | 75.3 | 78.3 | ||
| CRL | 33.3 | 15.5 | 66.7 | 77.2 | 100.0 | 37.3 | 56.7 | 73.9 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 55.1 | 20.6 | 80.2 | 90.3 | ||
| TREC | Yelp | LogitNorm | 32.3 | 0.3 | 95.5 | 78.1 | 15.2 | 0.5 | 95.7 | 44.7 |
| OneVsRest | 4.2 | 0.3 | 97.9 | 83.4 | 39.3 | 0.4 | 91.6 | 63.0 | ||
| RCL | 78.5 | 0.6 | 80.2 | 20.2 | 67.7 | 0.6 | 83.0 | 24.2 | ||
| Bayes Approx | 46.6 | 0.4 | 93.0 | 61.6 | 63.7 | 0.7 | 87.7 | 21.3 | ||
| CE Loss | 27.2 | 0.3 | 94.9 | 71.2 | 11.3 | 0.4 | 96.9 | 62.3 | ||
| CRL | 96.6 | 0.7 | 70.9 | 16.5 | 99.8 | 0.7 | 24.3 | 9.7 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| NYT | LogitNorm | 9.3 | 1.0 | 97.7 | 91.2 | 10.9 | 2.3 | 97.2 | 77.3 | |
| OneVsRest | 1.8 | 0.5 | 99.0 | 95.8 | 32.5 | 1.8 | 93.2 | 74.8 | ||
| RCL | 63.6 | 3.3 | 84.1 | 45.7 | 74.2 | 3.0 | 80.4 | 45.1 | ||
| Bayes Approx | 39.1 | 2.5 | 93.5 | 68.1 | 64.9 | 3.0 | 88.4 | 54.4 | ||
| CE Loss | 8.4 | 1.3 | 97.6 | 87.5 | 6.8 | 1.4 | 98.1 | 87.1 | ||
| CRL | 97.2 | 3.3 | 73.0 | 35.2 | 99.9 | 3.8 | 22.9 | 12.2 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Clima | LogitNorm | 11.3 | 0.8 | 97.6 | 88.5 | 10.4 | 1.6 | 97.3 | 75.9 | |
| OneVsRest | 3.4 | 0.6 | 98.7 | 93.2 | 36.3 | 1.8 | 91.7 | 60.1 | ||
| RCL | 73.5 | 2.4 | 81.2 | 37.1 | 60.7 | 2.0 | 86.4 | 48.2 | ||
| Bayes Approx | 51.7 | 2.0 | 90.5 | 54.4 | 61.1 | 2.4 | 87.6 | 40.4 | ||
| CE Loss | 15.1 | 1.7 | 96.2 | 69.5 | 12.2 | 1.8 | 96.4 | 67.8 | ||
| CRL | 86.3 | 2.2 | 80.7 | 39.5 | 98.9 | 2.7 | 30.1 | 13.0 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 |
| Method | FPR95 | ERR\nIn Dist: Clima\nOOD: NYT | AUROC | AUPR |
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 |
| + Grad Sub | 67.9 | 8.9 | 82.4 | 61.1 |
| + Grad Sub & Min | 98.4 | 10.7 | 49.2 | 32.3 |
| + Intra-Doc | 92.7 | 10.6 | 67.3 | 41.6 |
| - SILu | 14.6 | 3.9 | 96.7 | 89.3 |
| In Dist: NYT\nOOD: Clima | ||||
| SELFOOD | 10.5 | 1.7 | 98.4 | 90.5 |
| + Grad Sub only | 83.5 | 5.1 | 77.4 | 44.8 |
| + Grad Sub & Min | 96.5 | 5.2 | 60.3 | 36.0 |
| + Intra-Doc | 48.3 | 2.6 | 93.2 | 80.2 |
| - SiLU | 19.0 | 2.3 | 97.4 | 86.0 |
| OOD | Method | FPR95 | ERR | AUROC | AUPR |
| AGNews | LogitNorm | 45.9 | 0.9 | 86.7 | 15.8 |
| Bayes Approx | 57.5 | 0.9 | 87.7 | 23.9 | |
| CE-Loss | 45.5 | 0.9 | 86.5 | 16.6 | |
| CRL | 77.2 | 0.9 | 80.6 | 9.5 | |
| SELFOOD | 0.0 | 0.1 | 100.0 | 99.4 | |
| 20News | LogitNorm | 24.5 | 6.2 | 91.7 | 39.2 |
| Bayes Approx | 35.0 | 4.3 | 93.7 | 66.6 | |
| CE-Loss | 25.4 | 6.2 | 91.9 | 39.6 | |
| CRL | 59.9 | 6.4 | 87.4 | 36.9 | |
| SELFOOD | 28.2 | 4.9 | 94.2 | 61.5 |
| InD | OOD | Method | FPR95 | ERR | AUROC | AUPR |
| Clima | Yelp | SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 |
| SELFOOD + Maha | 0.8 | 0.2 | 99.2 | 97.2 | ||
| Yelp | NYT | SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 |
| SELFOOD + Maha | 12.4 | 4.4 | 97.4 | 97.5 |
| Text | CE-Loss | SELFOOD |
| Who said “What contemptible scoundrel stole the cork from my lunch?” | InD | OOD |
| Who said “Give me liberty or give me death”? | InD | OOD |
| Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next "place"... | OOD | InD |
| the pitch reads: "what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?" And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ... | OOD | InD |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| NYT | Clima | LogitNorm | 20.6 | 1.0 | 96.8 | 92.3 | 18.4 | 1.0 | 97.3 | 93.6 |
| OneVsRest | 84.7 | 1.7 | 87.3 | 82.6 | 92.7 | 1.9 | 91.1 | 85.1 | ||
| RCL | 46.9 | 1.5 | 94.4 | 87.5 | 86.7 | 2.3 | 86.7 | 76.6 | ||
| Bayes Approx | 40.6 | 1.2 | 95.4 | 90.0 | 80.0 | 2.4 | 88.4 | 76.5 | ||
| CE Loss | 22.1 | 1.0 | 96.9 | 92.5 | 15.7 | 1.0 | 97.3 | 92.8 | ||
| CRL | 32.3 | 1.7 | 95.4 | 86.5 | 79.3 | 3.2 | 79.8 | 67.0 | ||
| SELFOOD | 10.5 | 1.6 | 98.4 | 90.5 | 4.6 | 1.0 | 99.3 | 95.4 | ||
| TREC | LogitNorm | 13.3 | 5.5 | 97.2 | 87.9 | 17.5 | 6.4 | 96.1 | 84.1 | |
| OneVsRest | 96.0 | 6.2 | 84.0 | 79.4 | 97.2 | 6.3 | 90.5 | 82.8 | ||
| RCL | 62.9 | 6.1 | 92.2 | 85.8 | 73.9 | 8.8 | 88.4 | 73.1 | ||
| Bayes Approx | 38.8 | 4.5 | 94.8 | 89.9 | 67.5 | 8.3 | 88.7 | 74.7 | ||
| CE Loss | 14.5 | 5.8 | 97.0 | 87.1 | 21.9 | 6.7 | 95.5 | 83.5 | ||
| CRL | 67.9 | 6.5 | 89.9 | 79.1 | 85.6 | 9.6 | 71.4 | 61.4 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Yelp | LogitNorm | 1.7 | 0.6 | 99.3 | 87.9 | 1.2 | 0.6 | 98.5 | 87.9 | |
| OneVsRest | 15.1 | 0.7 | 97.4 | 78.1 | 8.7 | 0.6 | 97.6 | 78.8 | ||
| RCL | 45.1 | 0.5 | 94.9 | 84.7 | 76.7 | 1.0 | 90.2 | 60.0 | ||
| Bayes Approx | 28.7 | 0.4 | 96.4 | 86.8 | 83.0 | 1.1 | 87.7 | 56.6 | ||
| CE Loss | 3.4 | 0.7 | 99.2 | 82.3 | 1.8 | 0.6 | 99.3 | 85.4 | ||
| CRL | 69.5 | 0.5 | 89.3 | 77.8 | 73.3 | 1.0 | 83.8 | 58.3 | ||
| SELFOOD | 22.5 | 1.1 | 95.7 | 57.8 | 0.2 | 0.2 | 99.9 | 97.5 | ||
| Clima | TREC | LogitNorm | 37.8 | 5.5 | 94.6 | 91.9 | 60.5 | 7.0 | 91.6 | 88.0 |
| OneVsRest | 25.5 | 6.8 | 95.5 | 91.8 | 54.3 | 15.5 | 84.5 | 67.7 | ||
| RCL | 80.1 | 18.4 | 75.7 | 55.3 | 91.2 | 18.4 | 66.6 | 47.7 | ||
| Bayes Approx | 83.7 | 15.8 | 77.5 | 63.8 | 65.7 | 13.9 | 83.6 | 70.7 | ||
| CE Loss | 34.9 | 5.6 | 94.9 | 92.2 | 60.2 | 9.1 | 90.4 | 83.8 | ||
| CRL | 83.0 | 12.7 | 82.0 | 72.2 | 60.9 | 12.6 | 79.2 | 63.2 | ||
| SELFOOD | 0.3 | 1.1 | 99.4 | 99.4 | 38.9 | 8.8 | 90.1 | 84.7 | ||
| Yelp | LogitNorm | 57.7 | 0.8 | 93.0 | 81.3 | 54.7 | 1.2 | 92.6 | 72.3 | |
| OneVsRest | 43.7 | 1.0 | 92.7 | 72.5 | 58.8 | 1.8 | 73.5 | 34.7 | ||
| RCL | 91.0 | 2.3 | 60.7 | 12.2 | 98.0 | 2.3 | 56.4 | 12.7 | ||
| Bayes Approx | 81.4 | 2.2 | 76.5 | 22.7 | 96.8 | 2.2 | 52.3 | 16.3 | ||
| CE Loss | 46.3 | 0.9 | 93.8 | 80.1 | 79.5 | 1.9 | 86.5 | 47.2 | ||
| CRL | 98.7 | 1.4 | 67.3 | 53.3 | 72.9 | 2.1 | 50.9 | 19.0 | ||
| SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 | 37.6 | 1.5 | 72.3 | 54.9 | ||
| NYT | LogitNorm | 37.6 | 4.2 | 94.4 | 87.3 | 46.1 | 4.5 | 93.1 | 85.4 | |
| OneVsRest | 34.6 | 4.8 | 94.3 | 85.3 | 40.6 | 9.4 | 83.1 | 51.3 | ||
| RCL | 90.3 | 12.1 | 64.6 | 28.8 | 97.9 | 12.4 | 51.5 | 21.2 | ||
| Bayes Approx | 87.4 | 11.2 | 70.4 | 38.3 | 96.8 | 11.3 | 55.8 | 29.2 | ||
| CE Loss | 44.9 | 4.8 | 93.2 | 84.4 | 62.5 | 6.7 | 89.6 | 75.2 | ||
| CRL | 85.6 | 7.6 | 78.0 | 62.5 | 61.3 | 9.4 | 61.8 | 39.9 | ||
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 | 55.3 | 6.7 | 68.6 | 61.5 |
| Dataset | Domain | Criteria | #Docs | #labels |
| NYT | News | Topic | 13081 | 26 |
| Yelp | Reviews | Sentiment | 70000 | 5 |
| Clima | Climate | Question Type | 17175 | 8 |
| TREC | General | Question Type | 5952 | 6 |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| Yelp | NYT | LogitNorm | 82.0 | 33.2 | 69.4 | 57.9 | 84.4 | 33.9 | 66.6 | 55.0 |
| OneVsRest | 85.1 | 35.8 | 62.4 | 49.3 | 88.5 | 29.6 | 69.8 | 56.5 | ||
| RCL | 56.9 | 31.2 | 73.2 | 55.2 | 87.7 | 34.5 | 56.7 | 45.0 | ||
| Bayes Approx | 64.3 | 35.9 | 66.5 | 50.3 | 79.6 | 36.4 | 54.1 | 40.8 | ||
| CE Loss | 82.1 | 34.9 | 63.4 | 50.7 | 79.8 | 32.9 | 66.8 | 58.0 | ||
| CRL | 99.9 | 36.9 | 43.9 | 51.1 | 99.8 | 28.1 | 64.8 | 60.5 | ||
| SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 | 69.5 | 21.6 | 72.6 | 78.8 | ||
| Clima | LogitNorm | 79.2 | 25.5 | 72.5 | 54.8 | 82.6 | 25.4 | 70.7 | 54.0 | |
| OneVsRest | 83.0 | 25.1 | 69.5 | 53.2 | 86.7 | 28.0 | 53.8 | 33.7 | ||
| RCL | 59.9 | 27.8 | 67.4 | 42.1 | 73.9 | 22.5 | 70.7 | 56.3 | ||
| Bayes Approx | 74.7 | 28.7 | 56.7 | 33.6 | 63.0 | 28.0 | 66.8 | 40.9 | ||
| CE Loss | 83.2 | 27.8 | 48.6 | 37.0 | 76.4 | 22.0 | 78.2 | 68.4 | ||
| CRL | 99.6 | 29.0 | 49.1 | 38.8 | 99.9 | 20.6 | 63.0 | 56.9 | ||
| SELFOOD | 17.6 | 4.8 | 97.9 | 96.4 | 65.6 | 16.6 | 73.2 | 77.2 | ||
| TREC | LogitNorm | 68.2 | 27.0 | 79.8 | 82.1 | 60.8 | 23.0 | 83.9 | 84.4 | |
| OneVsRest | 84.8 | 40.0 | 63.9 | 69.7 | 75.5 | 33.0 | 57.9 | 61.2 | ||
| RCL | 89.8 | 43.2 | 40.1 | 51.5 | 77.4 | 32.4 | 65.0 | 68.6 | ||
| Bayes Approx | 92.5 | 44.7 | 32.1 | 44.7 | 51.8 | 25.9 | 72.7 | 69.4 | ||
| CE Loss | 74.2 | 33.7 | 52.8 | 62.7 | 58.5 | 27.5 | 75.3 | 78.3 | ||
| CRL | 33.3 | 15.5 | 66.7 | 77.2 | 100.0 | 37.3 | 56.7 | 73.9 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 55.1 | 20.6 | 80.2 | 90.3 | ||
| TREC | Yelp | LogitNorm | 32.3 | 0.3 | 95.5 | 78.1 | 15.2 | 0.5 | 95.7 | 44.7 |
| OneVsRest | 4.2 | 0.3 | 97.9 | 83.4 | 39.3 | 0.4 | 91.6 | 63.0 | ||
| RCL | 78.5 | 0.6 | 80.2 | 20.2 | 67.7 | 0.6 | 83.0 | 24.2 | ||
| Bayes Approx | 46.6 | 0.4 | 93.0 | 61.6 | 63.7 | 0.7 | 87.7 | 21.3 | ||
| CE Loss | 27.2 | 0.3 | 94.9 | 71.2 | 11.3 | 0.4 | 96.9 | 62.3 | ||
| CRL | 96.6 | 0.7 | 70.9 | 16.5 | 99.8 | 0.7 | 24.3 | 9.7 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| NYT | LogitNorm | 9.3 | 1.0 | 97.7 | 91.2 | 10.9 | 2.3 | 97.2 | 77.3 | |
| OneVsRest | 1.8 | 0.5 | 99.0 | 95.8 | 32.5 | 1.8 | 93.2 | 74.8 | ||
| RCL | 63.6 | 3.3 | 84.1 | 45.7 | 74.2 | 3.0 | 80.4 | 45.1 | ||
| Bayes Approx | 39.1 | 2.5 | 93.5 | 68.1 | 64.9 | 3.0 | 88.4 | 54.4 | ||
| CE Loss | 8.4 | 1.3 | 97.6 | 87.5 | 6.8 | 1.4 | 98.1 | 87.1 | ||
| CRL | 97.2 | 3.3 | 73.0 | 35.2 | 99.9 | 3.8 | 22.9 | 12.2 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Clima | LogitNorm | 11.3 | 0.8 | 97.6 | 88.5 | 10.4 | 1.6 | 97.3 | 75.9 | |
| OneVsRest | 3.4 | 0.6 | 98.7 | 93.2 | 36.3 | 1.8 | 91.7 | 60.1 | ||
| RCL | 73.5 | 2.4 | 81.2 | 37.1 | 60.7 | 2.0 | 86.4 | 48.2 | ||
| Bayes Approx | 51.7 | 2.0 | 90.5 | 54.4 | 61.1 | 2.4 | 87.6 | 40.4 | ||
| CE Loss | 15.1 | 1.7 | 96.2 | 69.5 | 12.2 | 1.8 | 96.4 | 67.8 | ||
| CRL | 86.3 | 2.2 | 80.7 | 39.5 | 98.9 | 2.7 | 30.1 | 13.0 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Method | FPR95 | ERR +In Dist: Clima +OOD: NYT | AUROC | AUPR |
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 |
| + Grad Sub | 67.9 | 8.9 | 82.4 | 61.1 |
| + Grad Sub & Min | 98.4 | 10.7 | 49.2 | 32.3 |
| + Intra-Doc | 92.7 | 10.6 | 67.3 | 41.6 |
| - SILu | 14.6 | 3.9 | 96.7 | 89.3 |
| In Dist: NYT +OOD: Clima | ||||
| SELFOOD | 10.5 | 1.7 | 98.4 | 90.5 |
| + Grad Sub only | 83.5 | 5.1 | 77.4 | 44.8 |
| + Grad Sub & Min | 96.5 | 5.2 | 60.3 | 36.0 |
| + Intra-Doc | 48.3 | 2.6 | 93.2 | 80.2 |
| - SiLU | 19.0 | 2.3 | 97.4 | 86.0 |
| OOD | Method | FPR95 | ERR | AUROC | AUPR |
| AGNews | LogitNorm | 45.9 | 0.9 | 86.7 | 15.8 |
| Bayes Approx | 57.5 | 0.9 | 87.7 | 23.9 | |
| CE-Loss | 45.5 | 0.9 | 86.5 | 16.6 | |
| CRL | 77.2 | 0.9 | 80.6 | 9.5 | |
| SELFOOD | 0.0 | 0.1 | 100.0 | 99.4 | |
| 20News | LogitNorm | 24.5 | 6.2 | 91.7 | 39.2 |
| Bayes Approx | 35.0 | 4.3 | 93.7 | 66.6 | |
| CE-Loss | 25.4 | 6.2 | 91.9 | 39.6 | |
| CRL | 59.9 | 6.4 | 87.4 | 36.9 | |
| SELFOOD | 28.2 | 4.9 | 94.2 | 61.5 |
| InD | OOD | Method | FPR95 | ERR | AUROC | AUPR |
| Clima | Yelp | SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 |
| SELFOOD + Maha | 0.8 | 0.2 | 99.2 | 97.2 | ||
| Yelp | NYT | SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 |
| SELFOOD + Maha | 12.4 | 4.4 | 97.4 | 97.5 |
| Text | CE-Loss | SELFOOD |
| Who said “What contemptible scoundrel stole the cork from my lunch?” | InD | OOD |
| Who said “Give me liberty or give me death”? | InD | OOD |
| Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next "place"... | OOD | InD |
| the pitch reads: "what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?" And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ... | OOD | InD |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| NYT | Clima | LogitNorm | 20.6 | 1.0 | 96.8 | 92.3 | 18.4 | 1.0 | 97.3 | 93.6 |
| OneVsRest | 84.7 | 1.7 | 87.3 | 82.6 | 92.7 | 1.9 | 91.1 | 85.1 | ||
| RCL | 46.9 | 1.5 | 94.4 | 87.5 | 86.7 | 2.3 | 86.7 | 76.6 | ||
| Bayes Approx | 40.6 | 1.2 | 95.4 | 90.0 | 80.0 | 2.4 | 88.4 | 76.5 | ||
| CE Loss | 22.1 | 1.0 | 96.9 | 92.5 | 15.7 | 1.0 | 97.3 | 92.8 | ||
| CRL | 32.3 | 1.7 | 95.4 | 86.5 | 79.3 | 3.2 | 79.8 | 67.0 | ||
| SELFOOD | 10.5 | 1.6 | 98.4 | 90.5 | 4.6 | 1.0 | 99.3 | 95.4 | ||
| TREC | LogitNorm | 13.3 | 5.5 | 97.2 | 87.9 | 17.5 | 6.4 | 96.1 | 84.1 | |
| OneVsRest | 96.0 | 6.2 | 84.0 | 79.4 | 97.2 | 6.3 | 90.5 | 82.8 | ||
| RCL | 62.9 | 6.1 | 92.2 | 85.8 | 73.9 | 8.8 | 88.4 | 73.1 | ||
| Bayes Approx | 38.8 | 4.5 | 94.8 | 89.9 | 67.5 | 8.3 | 88.7 | 74.7 | ||
| CE Loss | 14.5 | 5.8 | 97.0 | 87.1 | 21.9 | 6.7 | 95.5 | 83.5 | ||
| CRL | 67.9 | 6.5 | 89.9 | 79.1 | 85.6 | 9.6 | 71.4 | 61.4 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Yelp | LogitNorm | 1.7 | 0.6 | 99.3 | 87.9 | 1.2 | 0.6 | 98.5 | 87.9 | |
| OneVsRest | 15.1 | 0.7 | 97.4 | 78.1 | 8.7 | 0.6 | 97.6 | 78.8 | ||
| RCL | 45.1 | 0.5 | 94.9 | 84.7 | 76.7 | 1.0 | 90.2 | 60.0 | ||
| Bayes Approx | 28.7 | 0.4 | 96.4 | 86.8 | 83.0 | 1.1 | 87.7 | 56.6 | ||
| CE Loss | 3.4 | 0.7 | 99.2 | 82.3 | 1.8 | 0.6 | 99.3 | 85.4 | ||
| CRL | 69.5 | 0.5 | 89.3 | 77.8 | 73.3 | 1.0 | 83.8 | 58.3 | ||
| SELFOOD | 22.5 | 1.1 | 95.7 | 57.8 | 0.2 | 0.2 | 99.9 | 97.5 | ||
| Clima | TREC | LogitNorm | 37.8 | 5.5 | 94.6 | 91.9 | 60.5 | 7.0 | 91.6 | 88.0 |
| OneVsRest | 25.5 | 6.8 | 95.5 | 91.8 | 54.3 | 15.5 | 84.5 | 67.7 | ||
| RCL | 80.1 | 18.4 | 75.7 | 55.3 | 91.2 | 18.4 | 66.6 | 47.7 | ||
| Bayes Approx | 83.7 | 15.8 | 77.5 | 63.8 | 65.7 | 13.9 | 83.6 | 70.7 | ||
| CE Loss | 34.9 | 5.6 | 94.9 | 92.2 | 60.2 | 9.1 | 90.4 | 83.8 | ||
| CRL | 83.0 | 12.7 | 82.0 | 72.2 | 60.9 | 12.6 | 79.2 | 63.2 | ||
| SELFOOD | 0.3 | 1.1 | 99.4 | 99.4 | 38.9 | 8.8 | 90.1 | 84.7 | ||
| Yelp | LogitNorm | 57.7 | 0.8 | 93.0 | 81.3 | 54.7 | 1.2 | 92.6 | 72.3 | |
| OneVsRest | 43.7 | 1.0 | 92.7 | 72.5 | 58.8 | 1.8 | 73.5 | 34.7 | ||
| RCL | 91.0 | 2.3 | 60.7 | 12.2 | 98.0 | 2.3 | 56.4 | 12.7 | ||
| Bayes Approx | 81.4 | 2.2 | 76.5 | 22.7 | 96.8 | 2.2 | 52.3 | 16.3 | ||
| CE Loss | 46.3 | 0.9 | 93.8 | 80.1 | 79.5 | 1.9 | 86.5 | 47.2 | ||
| CRL | 98.7 | 1.4 | 67.3 | 53.3 | 72.9 | 2.1 | 50.9 | 19.0 | ||
| SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 | 37.6 | 1.5 | 72.3 | 54.9 | ||
| NYT | LogitNorm | 37.6 | 4.2 | 94.4 | 87.3 | 46.1 | 4.5 | 93.1 | 85.4 | |
| OneVsRest | 34.6 | 4.8 | 94.3 | 85.3 | 40.6 | 9.4 | 83.1 | 51.3 | ||
| RCL | 90.3 | 12.1 | 64.6 | 28.8 | 97.9 | 12.4 | 51.5 | 21.2 | ||
| Bayes Approx | 87.4 | 11.2 | 70.4 | 38.3 | 96.8 | 11.3 | 55.8 | 29.2 | ||
| CE Loss | 44.9 | 4.8 | 93.2 | 84.4 | 62.5 | 6.7 | 89.6 | 75.2 | ||
| CRL | 85.6 | 7.6 | 78.0 | 62.5 | 61.3 | 9.4 | 61.8 | 39.9 | ||
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 | 55.3 | 6.7 | 68.6 | 61.5 | ||
| Dataset | Domain | Criteria | #Docs | #labels |
| NYT | News | Topic | 13081 | 26 |
| Yelp | Reviews | Sentiment | 70000 | 5 |
| Clima | Climate | Question Type | 17175 | 8 |
| TREC | General | Question Type | 5952 | 6 |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| Yelp | NYT | LogitNorm | 82.0 | 33.2 | 69.4 | 57.9 | 84.4 | 33.9 | 66.6 | 55.0 |
| OneVsRest | 85.1 | 35.8 | 62.4 | 49.3 | 88.5 | 29.6 | 69.8 | 56.5 | ||
| RCL | 56.9 | 31.2 | 73.2 | 55.2 | 87.7 | 34.5 | 56.7 | 45.0 | ||
| Bayes Approx | 64.3 | 35.9 | 66.5 | 50.3 | 79.6 | 36.4 | 54.1 | 40.8 | ||
| CE Loss | 82.1 | 34.9 | 63.4 | 50.7 | 79.8 | 32.9 | 66.8 | 58.0 | ||
| CRL | 99.9 | 36.9 | 43.9 | 51.1 | 99.8 | 28.1 | 64.8 | 60.5 | ||
| SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 | 69.5 | 21.6 | 72.6 | 78.8 | ||
| Clima | LogitNorm | 79.2 | 25.5 | 72.5 | 54.8 | 82.6 | 25.4 | 70.7 | 54.0 | |
| OneVsRest | 83.0 | 25.1 | 69.5 | 53.2 | 86.7 | 28.0 | 53.8 | 33.7 | ||
| RCL | 59.9 | 27.8 | 67.4 | 42.1 | 73.9 | 22.5 | 70.7 | 56.3 | ||
| Bayes Approx | 74.7 | 28.7 | 56.7 | 33.6 | 63.0 | 28.0 | 66.8 | 40.9 | ||
| CE Loss | 83.2 | 27.8 | 48.6 | 37.0 | 76.4 | 22.0 | 78.2 | 68.4 | ||
| CRL | 99.6 | 29.0 | 49.1 | 38.8 | 99.9 | 20.6 | 63.0 | 56.9 | ||
| SELFOOD | 17.6 | 4.8 | 97.9 | 96.4 | 65.6 | 16.6 | 73.2 | 77.2 | ||
| TREC | LogitNorm | 68.2 | 27.0 | 79.8 | 82.1 | 60.8 | 23.0 | 83.9 | 84.4 | |
| OneVsRest | 84.8 | 40.0 | 63.9 | 69.7 | 75.5 | 33.0 | 57.9 | 61.2 | ||
| RCL | 89.8 | 43.2 | 40.1 | 51.5 | 77.4 | 32.4 | 65.0 | 68.6 | ||
| Bayes Approx | 92.5 | 44.7 | 32.1 | 44.7 | 51.8 | 25.9 | 72.7 | 69.4 | ||
| CE Loss | 74.2 | 33.7 | 52.8 | 62.7 | 58.5 | 27.5 | 75.3 | 78.3 | ||
| CRL | 33.3 | 15.5 | 66.7 | 77.2 | 100.0 | 37.3 | 56.7 | 73.9 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 55.1 | 20.6 | 80.2 | 90.3 | ||
| TREC | Yelp | LogitNorm | 32.3 | 0.3 | 95.5 | 78.1 | 15.2 | 0.5 | 95.7 | 44.7 |
| OneVsRest | 4.2 | 0.3 | 97.9 | 83.4 | 39.3 | 0.4 | 91.6 | 63.0 | ||
| RCL | 78.5 | 0.6 | 80.2 | 20.2 | 67.7 | 0.6 | 83.0 | 24.2 | ||
| Bayes Approx | 46.6 | 0.4 | 93.0 | 61.6 | 63.7 | 0.7 | 87.7 | 21.3 | ||
| CE Loss | 27.2 | 0.3 | 94.9 | 71.2 | 11.3 | 0.4 | 96.9 | 62.3 | ||
| CRL | 96.6 | 0.7 | 70.9 | 16.5 | 99.8 | 0.7 | 24.3 | 9.7 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| NYT | LogitNorm | 9.3 | 1.0 | 97.7 | 91.2 | 10.9 | 2.3 | 97.2 | 77.3 | |
| OneVsRest | 1.8 | 0.5 | 99.0 | 95.8 | 32.5 | 1.8 | 93.2 | 74.8 | ||
| RCL | 63.6 | 3.3 | 84.1 | 45.7 | 74.2 | 3.0 | 80.4 | 45.1 | ||
| Bayes Approx | 39.1 | 2.5 | 93.5 | 68.1 | 64.9 | 3.0 | 88.4 | 54.4 | ||
| CE Loss | 8.4 | 1.3 | 97.6 | 87.5 | 6.8 | 1.4 | 98.1 | 87.1 | ||
| CRL | 97.2 | 3.3 | 73.0 | 35.2 | 99.9 | 3.8 | 22.9 | 12.2 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Clima | LogitNorm | 11.3 | 0.8 | 97.6 | 88.5 | 10.4 | 1.6 | 97.3 | 75.9 | |
| OneVsRest | 3.4 | 0.6 | 98.7 | 93.2 | 36.3 | 1.8 | 91.7 | 60.1 | ||
| RCL | 73.5 | 2.4 | 81.2 | 37.1 | 60.7 | 2.0 | 86.4 | 48.2 | ||
| Bayes Approx | 51.7 | 2.0 | 90.5 | 54.4 | 61.1 | 2.4 | 87.6 | 40.4 | ||
| CE Loss | 15.1 | 1.7 | 96.2 | 69.5 | 12.2 | 1.8 | 96.4 | 67.8 | ||
| CRL | 86.3 | 2.2 | 80.7 | 39.5 | 98.9 | 2.7 | 30.1 | 13.0 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 |
| Method | FPR95 | ERR\nIn Dist: Clima\nOOD: NYT | AUROC | AUPR |
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 |
| + Grad Sub | 67.9 | 8.9 | 82.4 | 61.1 |
| + Grad Sub & Min | 98.4 | 10.7 | 49.2 | 32.3 |
| + Intra-Doc | 92.7 | 10.6 | 67.3 | 41.6 |
| - SILu | 14.6 | 3.9 | 96.7 | 89.3 |
| In Dist: NYT\nOOD: Clima | ||||
| SELFOOD | 10.5 | 1.7 | 98.4 | 90.5 |
| + Grad Sub only | 83.5 | 5.1 | 77.4 | 44.8 |
| + Grad Sub & Min | 96.5 | 5.2 | 60.3 | 36.0 |
| + Intra-Doc | 48.3 | 2.6 | 93.2 | 80.2 |
| - SiLU | 19.0 | 2.3 | 97.4 | 86.0 |
| OOD | Method | FPR95 | ERR | AUROC | AUPR |
| AGNews | LogitNorm | 45.9 | 0.9 | 86.7 | 15.8 |
| Bayes Approx | 57.5 | 0.9 | 87.7 | 23.9 | |
| CE-Loss | 45.5 | 0.9 | 86.5 | 16.6 | |
| CRL | 77.2 | 0.9 | 80.6 | 9.5 | |
| SELFOOD | 0.0 | 0.1 | 100.0 | 99.4 | |
| 20News | LogitNorm | 24.5 | 6.2 | 91.7 | 39.2 |
| Bayes Approx | 35.0 | 4.3 | 93.7 | 66.6 | |
| CE-Loss | 25.4 | 6.2 | 91.9 | 39.6 | |
| CRL | 59.9 | 6.4 | 87.4 | 36.9 | |
| SELFOOD | 28.2 | 4.9 | 94.2 | 61.5 |
| InD | OOD | Method | FPR95 | ERR | AUROC | AUPR |
| Clima | Yelp | SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 |
| SELFOOD + Maha | 0.8 | 0.2 | 99.2 | 97.2 | ||
| Yelp | NYT | SELFOOD | 63.2 | 19.6 | 79.4 | 82.7 |
| SELFOOD + Maha | 12.4 | 4.4 | 97.4 | 97.5 |
| Text | CE-Loss | SELFOOD |
| Who said “What contemptible scoundrel stole the cork from my lunch?” | InD | OOD |
| Who said “Give me liberty or give me death”? | InD | OOD |
| Its a tequila house - wouldn’t you want your staff to be knowledgeable about tequila? Call me silly. So Gregg and I are wondering Bourbon Street, looking for the next "place"... | OOD | InD |
| the pitch reads: "what’s worse than mean girl cheerleaders? how about resurrected mean-girl cheerleaders with supernatural powers?" And if this place could be farther away from my house, it’d be in Camden... Oh, the Vomitorium, every time i grace the hipster-stained ... | OOD | InD |
| In-dist | OOD | Method | BERT | RoBERTa | ||||||
| FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | FPR95(↓) | ERR(↓) | AUROC(↑) | AUPR(↑) | |||
| NYT | Clima | LogitNorm | 20.6 | 1.0 | 96.8 | 92.3 | 18.4 | 1.0 | 97.3 | 93.6 |
| OneVsRest | 84.7 | 1.7 | 87.3 | 82.6 | 92.7 | 1.9 | 91.1 | 85.1 | ||
| RCL | 46.9 | 1.5 | 94.4 | 87.5 | 86.7 | 2.3 | 86.7 | 76.6 | ||
| Bayes Approx | 40.6 | 1.2 | 95.4 | 90.0 | 80.0 | 2.4 | 88.4 | 76.5 | ||
| CE Loss | 22.1 | 1.0 | 96.9 | 92.5 | 15.7 | 1.0 | 97.3 | 92.8 | ||
| CRL | 32.3 | 1.7 | 95.4 | 86.5 | 79.3 | 3.2 | 79.8 | 67.0 | ||
| SELFOOD | 10.5 | 1.6 | 98.4 | 90.5 | 4.6 | 1.0 | 99.3 | 95.4 | ||
| TREC | LogitNorm | 13.3 | 5.5 | 97.2 | 87.9 | 17.5 | 6.4 | 96.1 | 84.1 | |
| OneVsRest | 96.0 | 6.2 | 84.0 | 79.4 | 97.2 | 6.3 | 90.5 | 82.8 | ||
| RCL | 62.9 | 6.1 | 92.2 | 85.8 | 73.9 | 8.8 | 88.4 | 73.1 | ||
| Bayes Approx | 38.8 | 4.5 | 94.8 | 89.9 | 67.5 | 8.3 | 88.7 | 74.7 | ||
| CE Loss | 14.5 | 5.8 | 97.0 | 87.1 | 21.9 | 6.7 | 95.5 | 83.5 | ||
| CRL | 67.9 | 6.5 | 89.9 | 79.1 | 85.6 | 9.6 | 71.4 | 61.4 | ||
| SELFOOD | 0.0 | 0.0 | 100.0 | 100.0 | 0.0 | 0.0 | 100.0 | 100.0 | ||
| Yelp | LogitNorm | 1.7 | 0.6 | 99.3 | 87.9 | 1.2 | 0.6 | 98.5 | 87.9 | |
| OneVsRest | 15.1 | 0.7 | 97.4 | 78.1 | 8.7 | 0.6 | 97.6 | 78.8 | ||
| RCL | 45.1 | 0.5 | 94.9 | 84.7 | 76.7 | 1.0 | 90.2 | 60.0 | ||
| Bayes Approx | 28.7 | 0.4 | 96.4 | 86.8 | 83.0 | 1.1 | 87.7 | 56.6 | ||
| CE Loss | 3.4 | 0.7 | 99.2 | 82.3 | 1.8 | 0.6 | 99.3 | 85.4 | ||
| CRL | 69.5 | 0.5 | 89.3 | 77.8 | 73.3 | 1.0 | 83.8 | 58.3 | ||
| SELFOOD | 22.5 | 1.1 | 95.7 | 57.8 | 0.2 | 0.2 | 99.9 | 97.5 | ||
| Clima | TREC | LogitNorm | 37.8 | 5.5 | 94.6 | 91.9 | 60.5 | 7.0 | 91.6 | 88.0 |
| OneVsRest | 25.5 | 6.8 | 95.5 | 91.8 | 54.3 | 15.5 | 84.5 | 67.7 | ||
| RCL | 80.1 | 18.4 | 75.7 | 55.3 | 91.2 | 18.4 | 66.6 | 47.7 | ||
| Bayes Approx | 83.7 | 15.8 | 77.5 | 63.8 | 65.7 | 13.9 | 83.6 | 70.7 | ||
| CE Loss | 34.9 | 5.6 | 94.9 | 92.2 | 60.2 | 9.1 | 90.4 | 83.8 | ||
| CRL | 83.0 | 12.7 | 82.0 | 72.2 | 60.9 | 12.6 | 79.2 | 63.2 | ||
| SELFOOD | 0.3 | 1.1 | 99.4 | 99.4 | 38.9 | 8.8 | 90.1 | 84.7 | ||
| Yelp | LogitNorm | 57.7 | 0.8 | 93.0 | 81.3 | 54.7 | 1.2 | 92.6 | 72.3 | |
| OneVsRest | 43.7 | 1.0 | 92.7 | 72.5 | 58.8 | 1.8 | 73.5 | 34.7 | ||
| RCL | 91.0 | 2.3 | 60.7 | 12.2 | 98.0 | 2.3 | 56.4 | 12.7 | ||
| Bayes Approx | 81.4 | 2.2 | 76.5 | 22.7 | 96.8 | 2.2 | 52.3 | 16.3 | ||
| CE Loss | 46.3 | 0.9 | 93.8 | 80.1 | 79.5 | 1.9 | 86.5 | 47.2 | ||
| CRL | 98.7 | 1.4 | 67.3 | 53.3 | 72.9 | 2.1 | 50.9 | 19.0 | ||
| SELFOOD | 14.7 | 0.6 | 98.0 | 90.8 | 37.6 | 1.5 | 72.3 | 54.9 | ||
| NYT | LogitNorm | 37.6 | 4.2 | 94.4 | 87.3 | 46.1 | 4.5 | 93.1 | 85.4 | |
| OneVsRest | 34.6 | 4.8 | 94.3 | 85.3 | 40.6 | 9.4 | 83.1 | 51.3 | ||
| RCL | 90.3 | 12.1 | 64.6 | 28.8 | 97.9 | 12.4 | 51.5 | 21.2 | ||
| Bayes Approx | 87.4 | 11.2 | 70.4 | 38.3 | 96.8 | 11.3 | 55.8 | 29.2 | ||
| CE Loss | 44.9 | 4.8 | 93.2 | 84.4 | 62.5 | 6.7 | 89.6 | 75.2 | ||
| CRL | 85.6 | 7.6 | 78.0 | 62.5 | 61.3 | 9.4 | 61.8 | 39.9 | ||
| SELFOOD | 17.2 | 2.6 | 97.8 | 94.5 | 55.3 | 6.7 | 68.6 | 61.5 |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| Full-shot fine-tuning (with 8.4k+ training dialogs): | ||||||||
| DAMD (Zhang et al., 2020) | 76.33 | 60.40 | 16.60 | 84.97 | - | - | - | - |
| SIMPLETOD (Hosseini-Asl et al., 2020) | 84.40 | 70.10 | 15.01 | 92.26 | - | - | - | - |
| SOLOIST (Peng et al., 2021a) | 85.50 | 72.90 | 16.54 | 95.74 | 81.70 | 67.10 | 13.60 | 88.00 |
| PPTOD (Su et al., 2022) | 89.20 | 79.40 | 18.62 | 102.92 | - | - | - | - |
| MARS (Sun et al., 2022) | 88.90 | 78.00 | 19.90 | 103.35 | 88.90 | 78.00 | 19.60 | 103.05 |
| Zero-shot transfer method (pre-trained on SGD): | ||||||||
| ANYTOD-XXL | - | - | - | - | 73.90 | 24.40 | 3.40 | 52.55 |
| Few-shot prompting: | ||||||||
| IG-TOD-CHATGPT-FS | - | - | - | - | - | 20.00 | 7.17 | - |
| FEW-SHOT-CHATGPT | 44.74 | 24.32 | 7.88 | 42.41 | 45.40 | 24.50 | 7.72 | 42.67 |
| Zero-shot prompting: | ||||||||
| IG-TOD-CHATGPT-ZS | - | - | - | - | - | 15.00 | 3.58 | - |
| SGP-TOD-CHATGPT (Ours) | 64.56 | 54.05 | 7.17 | 66.48 | 64.70 | 54.70 | 6.96 | 66.66 |
| SGP-TOD-CODEX (Ours) | 71.67 | 52.55 | 7.91 | 70.02 | 75.50 | 52.30 | 6.62 | 70.53 |
| SGP-TOD-GPT3.5 (Ours) | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| Model | Attraction | Train | Hotel | Restaurant | ||||||||||||
| Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | |
| Few-shot fine-tuning (with 50 training dialogs): | ||||||||||||||||
| SIMPLETOD | 65.66 | 46.97 | 5.85 | 62.17 | 59.00 | 44.00 | 7.07 | 58.57 | 62.50 | 40.00 | 7.70 | 58.95 | 75.50 | 44.50 | 11.00 | 71.00 |
| SOLOIST | 86.00 | 65.00 | 12.90 | 88.40 | 80.81 | 64.65 | 9.96 | 82.69 | 74.50 | 43.50 | 8.12 | 67.12 | 81.00 | 55.50 | 12.80 | 81.50 |
| Few-shot prompting: | ||||||||||||||||
| Few-SHOT-CHATGPT | 75.00 | 67.00 | 8.22 | 79.23 | 79.80 | 65.66 | 8.12 | 80.85 | 51.00 | 26.50 | 5.80 | 44.55 | 80.00 | 55.50 | 7.71 | 75.46 |
| Zero-shot prompting: | ||||||||||||||||
| SGP-TOD-CHATGPT | 95.00 | 94.00 | 7.13 | 101.63 | 76.77 | 74.24 | 6.75 | 82.26 | 76.50 | 57.00 | 5.16 | 71.91 | 90.00 | 82.50 | 6.72 | 92.97 |
| SGP-TOD-CODEX | 98.00 | 93.00 | 10.45 | 105.95 | 78.79 | 70.20 | 8.56 | 83.06 | 83.50 | 69.50 | 7.86 | 84.36 | 91.00 | 85.00 | 10.50 | 98.50 |
| SGP-TOD-GPT3.5 | 96.00 | 93.00 | 9.53 | 104.03 | 82.83 | 77.27 | 8.72 | 88.77 | 82.50 | 71.50 | 7.05 | 84.05 | 91.50 | 84.00 | 12.90 | 100.65 |
| Model | Task transfer | Domain transfer | ||
| F1 | Accuracy | F1 | Accuracy | |
| Zero-shot transfer (leave-one fine-tuning with 2.5k training dialogs): | ||||
| BERT+S | 24.25 | 24.89 | 25.70 | 28.56 |
| SAM | 49.82 | 51.30 | 55.91 | 57.92 |
| Zero-shot prompting: | ||||
| SGP-TOD-CODEX-INI | 45.18 | 47.99 | 47.21 | 49.97 |
| SGP-TOD-GPT3.5 | 47.67 | 48.27 | 49.76 | 50.39 |
| SGP-TOD-CODEX | 49.78 | 51.01 | 52.72 | 53.66 |
| SGP-TOD-GPT3.5-E2E | 50.84 | 50.74 | 53.50 | 53.21 |
| Model | FT/FS/ZS | Restaurant-Ext | |||
| Info. | Succ. | BLEU | BERTS. | ||
| Without domain-relevant knowledge | |||||
| ChatGPT | ZS | 44.00 | 6.00 | 4.31 | 85.96 |
| GPT-3.5 | ZS | 34.00 | 16.00 | 8.70 | 84.31 |
| With prior knowledge on Restaurant | |||||
| SOLOIST | FT | 78.00 | 0.00 | 10.62 | 87.24 |
| SGP-TOD-CHATGPT | ZS | 88.00 | 34.00 | 5.45 | 86.11 |
| SGP-TOD-GPT3.5 | ZS | 94.00 | 30.00 | 10.68 | 87.30 |
| With knowledge on Restaurant-Ext | |||||
| SOLOIST+TEACH | FT | 82.00 | 38.00 | 10.99 | 87.66 |
| FEW-SHOT-GPT3.5+TEACH | FS | 88.00 | 54.00 | 12.95 | 88.90 |
| SGP-TOD-CHATGPT-EXT | ZS | 88.00 | 78.00 | 6.25 | 86.15 |
| SGP-TOD-GPT3.5-EXT | ZS | 96.00 | 86.00 | 14.57 | 89.01 |
| Model | Restaurant | ||||
| Sw/o g ↑ | Sw/ g ↑ | Und. ↑ | App. ↑ | T. ↓ | |
| SOLOIST | 34.00 | 30.00 | 2.18 | 2.10 | 10.64 |
| FEW-SHOT-CHATGPT | 94.00 | 74.00 | 4.58 | 4.72 | 8.32 |
| SGP-TOD-CHATGPT | 100.00 | 92.00 | 4.86 | 4.88 | 7.28 |
| Model | Multiwoz 2.2 | |||
| Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.60 | 33.90 | 6.29 | 42.54 |
| Model | Task | Schema types | Training strategy | ||
| Fine-tuning | Pre-training | Prompting | |||
| ROBUSTSF (Shah et al., 2019) | SF | slot names/value examples | ✓ | ||
| TRADE (Wu et al., 2019) | DST | slot names/value examples | ✓ | ||
| ZSTL-SD (Campaigna et al., 2020) | DST | ontology, dialog templates | ✓(+synthesized data) | ||
| S-DST (Rastogi et al., 2020b) | DST | slot names/descriptionss+service, intent names/descriptionss | ✓ | ||
| T5DST (Lin et al., 2021b) | DST | slot names/descriptionss | ✓ | ||
| TRANSFERQA (Lin et al., 2021a) | DST | slot names/value examples | ✓(QA tasks) | ||
| IC-DST (Hu et al., 2022) | DST | slot names/value examples | ✓ | ||
| SDM-DST (Wang et al., 2022a) | DST | slot names/value examples | ✓ | ||
| BERT+S (Mosig et al., 2020) | E2E policy | system-side policy skeletons | ✓ | ||
| SAM (Mehri and Eskenazi, 2021) | E2E policy | user-aware policy skeletons | ✓ | ||
| ZSDG (Zhao and Eskenazi, 2018) | E2E dialog | ontology, response templates | ✓ | ||
| DAML (Qian and Yu, 2019) | E2E dialog | ontology, response templates | ✓ | ||
| ANYTOD (Zhao et al., 2022) | E2E dialog | policy programs+slot names/value examples+slot descriptions+user action names/states/descriptionss | ✓ | ✓(heterogeneous tasks) | |
| IG-TOD (Hudecek and Dusek, 2023) | E2E dialog | slot names+slot descriptions | ✓ | ||
| SGP-TOD (ours) | E2E dialog | user-aware policy skeletons(+slot names/value examples) | ✓ |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 82.28 | 55.65 | 6.51 | 75.48 | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.20 | 50.95 | 6.48 | 72.56 | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.74 | 33.13 | 6.18 | 42.12 | 38.60 | 33.90 | 6.29 | 42.54 |
| Slot | Value |
| "address" | "21 - 24 Northampton Street" |
| "area" | "west" |
| "food" | "british" |
| "id" | "14810" |
| "location" | [52.21031, 0.11381] |
| "name" | "saint johns chop house" |
| "phone" | "01223353110" |
| "postcode" | "cb30ad" |
| "pricerange" | "moderate" |
| "type" | "restaurant" |
| "delivery" | "Yes" |
| "delivery fee" | "6 pounds" |
| "dish" | "Beef Wellington" |
| "start_time" | "10:30 am" |
| "end_time" | "22:40 pm" |
| Turn | Utterance |
| 1 | User: I want a restaurant on the west end of town. |
| 2 | System: Ok. Are you looking for any particular type of food? |
| 3 | User: I don't care. |
| 4 | System: Are you looking for a particular price range? |
| 5 | User: A moderately priced restaurant would be good. |
| 6 | System: How about Saint Johns Chop House? They serve moderately priced British food on the west side of town. |
| 7 | User: Can i please have their address and phone number? |
| 8 | System: Saint Johns Chop House's address is 21-24 Northampton Street, their phone number is 01223353110. |
| 9 | User: Does the restaurant offers delivery service? If it does, what's the delivery fee? |
| 10 | System: Yes, it does, and the delivery charges 6 pounds. |
| 11 | User: Thank you, goodbye. |
| 12 | System: Thank you for using our system. Goodbye. |
| Turn | Interaction between a user and ChatGPT | Turn | Interaction between a user and SP-TOD-ChatGPT-Ext |
| 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. | 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. |
| 2 | System: Certainly, there are 3 Tuscan restaurants in town. What is your preferred price range? | 2 | System: I am sorry, there are no Tuscan restaurants. Here are some other options: Korean, Italian, Chinese. Which one would you like more information on? |
| 3 | User: I don't care about the price. Please recommend one for me, i also need the phone number please. | 3 | User: How about any Korean restaurants? I also need the phone number please. |
| 4 | System: The phone number for Pizza Hut City Centre is 01223323737. Is there anything else i can help you with today? | 4 | System: Little Seoul is a Korean restaurant. Their phone number is 01223308681. |
| 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? | 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? |
| 6 | System: I'm sorry, but I could not find any information on delivery fee in our database. You may want to contact the restaurant directly to inquire about this. Is there anything else I can assist you with today? | 6 | System: Yes, they offer delivery service and the delivery charge is 4 pounds. Can I help you with anything else? |
| 7 | User: No. Thank you, goodbye. | 7 | User: No. Thank you, goodbye. |
| 8 | System: Glad i could assist you. Have a good day. | 8 | System: You're welcome, have a great meal! Goodbye! |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| Full-shot fine-tuning (with 8.4k+ training dialogs): | ||||||||
| DAMD (Zhang et al., 2020) | 76.33 | 60.40 | 16.60 | 84.97 | - | - | - | - |
| SIMPLETOD (Hosseini-Asl et al., 2020) | 84.40 | 70.10 | 15.01 | 92.26 | - | - | - | - |
| SOLOIST (Peng et al., 2021a) | 85.50 | 72.90 | 16.54 | 95.74 | 81.70 | 67.10 | 13.60 | 88.00 |
| PPTOD (Su et al., 2022) | 89.20 | 79.40 | 18.62 | 102.92 | - | - | - | - |
| MARS (Sun et al., 2022) | 88.90 | 78.00 | 19.90 | 103.35 | 88.90 | 78.00 | 19.60 | 103.05 |
| Zero-shot transfer method (pre-trained on SGD): | ||||||||
| ANYTOD-XXL | - | - | - | - | 73.90 | 24.40 | 3.40 | 52.55 |
| Few-shot prompting: | ||||||||
| IG-TOD-CHATGPT-FS | - | - | - | - | - | 20.00 | 7.17 | - |
| FEW-SHOT-CHATGPT | 44.74 | 24.32 | 7.88 | 42.41 | 45.40 | 24.50 | 7.72 | 42.67 |
| Zero-shot prompting: | ||||||||
| IG-TOD-CHATGPT-ZS | - | - | - | - | - | 15.00 | 3.58 | - |
| SGP-TOD-CHATGPT (Ours) | 64.56 | 54.05 | 7.17 | 66.48 | 64.70 | 54.70 | 6.96 | 66.66 |
| SGP-TOD-CODEX (Ours) | 71.67 | 52.55 | 7.91 | 70.02 | 75.50 | 52.30 | 6.62 | 70.53 |
| SGP-TOD-GPT3.5 (Ours) | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| Model | Attraction | Train | Hotel | Restaurant | ||||||||||||
| Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | |
| Few-shot fine-tuning (with 50 training dialogs): | ||||||||||||||||
| SIMPLETOD | 65.66 | 46.97 | 5.85 | 62.17 | 59.00 | 44.00 | 7.07 | 58.57 | 62.50 | 40.00 | 7.70 | 58.95 | 75.50 | 44.50 | 11.00 | 71.00 |
| SOLOIST | 86.00 | 65.00 | 12.90 | 88.40 | 80.81 | 64.65 | 9.96 | 82.69 | 74.50 | 43.50 | 8.12 | 67.12 | 81.00 | 55.50 | 12.80 | 81.50 |
| Few-shot prompting: | ||||||||||||||||
| Few-SHOT-CHATGPT | 75.00 | 67.00 | 8.22 | 79.23 | 79.80 | 65.66 | 8.12 | 80.85 | 51.00 | 26.50 | 5.80 | 44.55 | 80.00 | 55.50 | 7.71 | 75.46 |
| Zero-shot prompting: | ||||||||||||||||
| SGP-TOD-CHATGPT | 95.00 | 94.00 | 7.13 | 101.63 | 76.77 | 74.24 | 6.75 | 82.26 | 76.50 | 57.00 | 5.16 | 71.91 | 90.00 | 82.50 | 6.72 | 92.97 |
| SGP-TOD-CODEX | 98.00 | 93.00 | 10.45 | 105.95 | 78.79 | 70.20 | 8.56 | 83.06 | 83.50 | 69.50 | 7.86 | 84.36 | 91.00 | 85.00 | 10.50 | 98.50 |
| SGP-TOD-GPT3.5 | 96.00 | 93.00 | 9.53 | 104.03 | 82.83 | 77.27 | 8.72 | 88.77 | 82.50 | 71.50 | 7.05 | 84.05 | 91.50 | 84.00 | 12.90 | 100.65 |
| Model | Task transfer | Domain transfer | ||
| F1 | Accuracy | F1 | Accuracy | |
| Zero-shot transfer (leave-one fine-tuning with 2.5k training dialogs): | ||||
| BERT+S | 24.25 | 24.89 | 25.70 | 28.56 |
| SAM | 49.82 | 51.30 | 55.91 | 57.92 |
| Zero-shot prompting: | ||||
| SGP-TOD-CODEX-INI | 45.18 | 47.99 | 47.21 | 49.97 |
| SGP-TOD-GPT3.5 | 47.67 | 48.27 | 49.76 | 50.39 |
| SGP-TOD-CODEX | 49.78 | 51.01 | 52.72 | 53.66 |
| SGP-TOD-GPT3.5-E2E | 50.84 | 50.74 | 53.50 | 53.21 |
| Model | FT/FS/ZS | Restaurant-Ext | |||
| Info. | Succ. | BLEU | BERTS. | ||
| Without domain-relevant knowledge | |||||
| ChatGPT | ZS | 44.00 | 6.00 | 4.31 | 85.96 |
| GPT-3.5 | ZS | 34.00 | 16.00 | 8.70 | 84.31 |
| With prior knowledge on Restaurant | |||||
| SOLOIST | FT | 78.00 | 0.00 | 10.62 | 87.24 |
| SGP-TOD-CHATGPT | ZS | 88.00 | 34.00 | 5.45 | 86.11 |
| SGP-TOD-GPT3.5 | ZS | 94.00 | 30.00 | 10.68 | 87.30 |
| With knowledge on Restaurant-Ext | |||||
| SOLOIST+TEACH | FT | 82.00 | 38.00 | 10.99 | 87.66 |
| FEW-SHOT-GPT3.5+TEACH | FS | 88.00 | 54.00 | 12.95 | 88.90 |
| SGP-TOD-CHATGPT-EXT | ZS | 88.00 | 78.00 | 6.25 | 86.15 |
| SGP-TOD-GPT3.5-EXT | ZS | 96.00 | 86.00 | 14.57 | 89.01 |
| Model | Restaurant | ||||
| Sw/o g ↑ | Sw/ g ↑ | Und. ↑ | App. ↑ | T. ↓ | |
| SOLOIST | 34.00 | 30.00 | 2.18 | 2.10 | 10.64 |
| FEW-SHOT-CHATGPT | 94.00 | 74.00 | 4.58 | 4.72 | 8.32 |
| SGP-TOD-CHATGPT | 100.00 | 92.00 | 4.86 | 4.88 | 7.28 |
| Model | Multiwoz 2.2 | |||
| Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.60 | 33.90 | 6.29 | 42.54 |
| Model | Task | Schema types | Training strategy | ||
| Fine-tuning | Pre-training | Prompting | |||
| ROBUSTSF (Shah et al., 2019) | SF | slot names/value examples | ✓ | ||
| TRADE (Wu et al., 2019) | DST | slot names/value examples | ✓ | ||
| ZSTL-SD (Campaigna et al., 2020) | DST | ontology, dialog templates | ✓(+synthesized data) | ||
| S-DST (Rastogi et al., 2020b) | DST | slot names/descriptionss+service, intent names/descriptionss | ✓ | ||
| T5DST (Lin et al., 2021b) | DST | slot names/descriptionss | ✓ | ||
| TRANSFERQA (Lin et al., 2021a) | DST | slot names/value examples | ✓(QA tasks) | ||
| IC-DST (Hu et al., 2022) | DST | slot names/value examples | ✓ | ||
| SDM-DST (Wang et al., 2022a) | DST | slot names/value examples | ✓ | ||
| BERT+S (Mosig et al., 2020) | E2E policy | system-side policy skeletons | ✓ | ||
| SAM (Mehri and Eskenazi, 2021) | E2E policy | user-aware policy skeletons | ✓ | ||
| ZSDG (Zhao and Eskenazi, 2018) | E2E dialog | ontology, response templates | ✓ | ||
| DAML (Qian and Yu, 2019) | E2E dialog | ontology, response templates | ✓ | ||
| ANYTOD (Zhao et al., 2022) | E2E dialog | policy programs+slot names/value examples+slot descriptions+user action names/states/descriptionss | ✓ | ✓(heterogeneous tasks) | |
| IG-TOD (Hudecek and Dusek, 2023) | E2E dialog | slot names+slot descriptions | ✓ | ||
| SGP-TOD (ours) | E2E dialog | user-aware policy skeletons(+slot names/value examples) | ✓ |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 82.28 | 55.65 | 6.51 | 75.48 | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.20 | 50.95 | 6.48 | 72.56 | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.74 | 33.13 | 6.18 | 42.12 | 38.60 | 33.90 | 6.29 | 42.54 |
| Slot | Value |
| "address" | "21 - 24 Northampton Street" |
| "area" | "west" |
| "food" | "british" |
| "id" | "14810" |
| "location" | [52.21031, 0.11381] |
| "name" | "saint johns chop house" |
| "phone" | "01223353110" |
| "postcode" | "cb30ad" |
| "pricerange" | "moderate" |
| "type" | "restaurant" |
| "delivery" | "Yes" |
| "delivery fee" | "6 pounds" |
| "dish" | "Beef Wellington" |
| "start_time" | "10:30 am" |
| "end_time" | "22:40 pm" |
| Turn | Utterance |
| 1 | User: I want a restaurant on the west end of town. |
| 2 | System: Ok. Are you looking for any particular type of food? |
| 3 | User: I don't care. |
| 4 | System: Are you looking for a particular price range? |
| 5 | User: A moderately priced restaurant would be good. |
| 6 | System: How about Saint Johns Chop House? They serve moderately priced British food on the west side of town. |
| 7 | User: Can i please have their address and phone number? |
| 8 | System: Saint Johns Chop House's address is 21-24 Northampton Street, their phone number is 01223353110. |
| 9 | User: Does the restaurant offers delivery service? If it does, what's the delivery fee? |
| 10 | System: Yes, it does, and the delivery charges 6 pounds. |
| 11 | User: Thank you, goodbye. |
| 12 | System: Thank you for using our system. Goodbye. |
| Turn | Interaction between a user and ChatGPT | Turn | Interaction between a user and SP-TOD-ChatGPT-Ext |
| 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. | 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. |
| 2 | System: Certainly, there are 3 Tuscan restaurants in town. What is your preferred price range? | 2 | System: I am sorry, there are no Tuscan restaurants. Here are some other options: Korean, Italian, Chinese. Which one would you like more information on? |
| 3 | User: I don't care about the price. Please recommend one for me, i also need the phone number please. | 3 | User: How about any Korean restaurants? I also need the phone number please. |
| 4 | System: The phone number for Pizza Hut City Centre is 01223323737. Is there anything else i can help you with today? | 4 | System: Little Seoul is a Korean restaurant. Their phone number is 01223308681. |
| 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? | 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? |
| 6 | System: I'm sorry, but I could not find any information on delivery fee in our database. You may want to contact the restaurant directly to inquire about this. Is there anything else I can assist you with today? | 6 | System: Yes, they offer delivery service and the delivery charge is 4 pounds. Can I help you with anything else? |
| 7 | User: No. Thank you, goodbye. | 7 | User: No. Thank you, goodbye. |
| 8 | System: Glad i could assist you. Have a good day. | 8 | System: You're welcome, have a great meal! Goodbye! |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| Full-shot fine-tuning (with 8.4k+ training dialogs): | ||||||||
| DAMD (Zhang et al., 2020) | 76.33 | 60.40 | 16.60 | 84.97 | - | - | - | - |
| SIMPLETOD (Hosseini-Asl et al., 2020) | 84.40 | 70.10 | 15.01 | 92.26 | - | - | - | - |
| SOLOIST (Peng et al., 2021a) | 85.50 | 72.90 | 16.54 | 95.74 | 81.70 | 67.10 | 13.60 | 88.00 |
| PPTOD (Su et al., 2022) | 89.20 | 79.40 | 18.62 | 102.92 | - | - | - | - |
| MARS (Sun et al., 2022) | 88.90 | 78.00 | 19.90 | 103.35 | 88.90 | 78.00 | 19.60 | 103.05 |
| Zero-shot transfer method (pre-trained on SGD): | ||||||||
| ANYTOD-XXL | - | - | - | - | 73.90 | 24.40 | 3.40 | 52.55 |
| Few-shot prompting: | ||||||||
| IG-TOD-CHATGPT-FS | - | - | - | - | - | 20.00 | 7.17 | - |
| FEW-SHOT-CHATGPT | 44.74 | 24.32 | 7.88 | 42.41 | 45.40 | 24.50 | 7.72 | 42.67 |
| Zero-shot prompting: | ||||||||
| IG-TOD-CHATGPT-ZS | - | - | - | - | - | 15.00 | 3.58 | - |
| SGP-TOD-CHATGPT (Ours) | 64.56 | 54.05 | 7.17 | 66.48 | 64.70 | 54.70 | 6.96 | 66.66 |
| SGP-TOD-CODEX (Ours) | 71.67 | 52.55 | 7.91 | 70.02 | 75.50 | 52.30 | 6.62 | 70.53 |
| SGP-TOD-GPT3.5 (Ours) | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| Model | Attraction | Train | Hotel | Restaurant | ||||||||||||
| Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | |
| Few-shot fine-tuning (with 50 training dialogs): | ||||||||||||||||
| SIMPLETOD | 65.66 | 46.97 | 5.85 | 62.17 | 59.00 | 44.00 | 7.07 | 58.57 | 62.50 | 40.00 | 7.70 | 58.95 | 75.50 | 44.50 | 11.00 | 71.00 |
| SOLOIST | 86.00 | 65.00 | 12.90 | 88.40 | 80.81 | 64.65 | 9.96 | 82.69 | 74.50 | 43.50 | 8.12 | 67.12 | 81.00 | 55.50 | 12.80 | 81.50 |
| Few-shot prompting: | ||||||||||||||||
| Few-SHOT-CHATGPT | 75.00 | 67.00 | 8.22 | 79.23 | 79.80 | 65.66 | 8.12 | 80.85 | 51.00 | 26.50 | 5.80 | 44.55 | 80.00 | 55.50 | 7.71 | 75.46 |
| Zero-shot prompting: | ||||||||||||||||
| SGP-TOD-CHATGPT | 95.00 | 94.00 | 7.13 | 101.63 | 76.77 | 74.24 | 6.75 | 82.26 | 76.50 | 57.00 | 5.16 | 71.91 | 90.00 | 82.50 | 6.72 | 92.97 |
| SGP-TOD-CODEX | 98.00 | 93.00 | 10.45 | 105.95 | 78.79 | 70.20 | 8.56 | 83.06 | 83.50 | 69.50 | 7.86 | 84.36 | 91.00 | 85.00 | 10.50 | 98.50 |
| SGP-TOD-GPT3.5 | 96.00 | 93.00 | 9.53 | 104.03 | 82.83 | 77.27 | 8.72 | 88.77 | 82.50 | 71.50 | 7.05 | 84.05 | 91.50 | 84.00 | 12.90 | 100.65 |
| Model | Task transfer | Domain transfer | ||
| F1 | Accuracy | F1 | Accuracy | |
| Zero-shot transfer (leave-one fine-tuning with 2.5k training dialogs): | ||||
| BERT+S | 24.25 | 24.89 | 25.70 | 28.56 |
| SAM | 49.82 | 51.30 | 55.91 | 57.92 |
| Zero-shot prompting: | ||||
| SGP-TOD-CODEX-INI | 45.18 | 47.99 | 47.21 | 49.97 |
| SGP-TOD-GPT3.5 | 47.67 | 48.27 | 49.76 | 50.39 |
| SGP-TOD-CODEX | 49.78 | 51.01 | 52.72 | 53.66 |
| SGP-TOD-GPT3.5-E2E | 50.84 | 50.74 | 53.50 | 53.21 |
| Model | FT/FS/ZS | Restaurant-Ext | |||
| Info. | Succ. | BLEU | BERTS. | ||
| Without domain-relevant knowledge | |||||
| ChatGPT | ZS | 44.00 | 6.00 | 4.31 | 85.96 |
| GPT-3.5 | ZS | 34.00 | 16.00 | 8.70 | 84.31 |
| With prior knowledge on Restaurant | |||||
| SOLOIST | FT | 78.00 | 0.00 | 10.62 | 87.24 |
| SGP-TOD-CHATGPT | ZS | 88.00 | 34.00 | 5.45 | 86.11 |
| SGP-TOD-GPT3.5 | ZS | 94.00 | 30.00 | 10.68 | 87.30 |
| With knowledge on Restaurant-Ext | |||||
| SOLOIST+TEACH | FT | 82.00 | 38.00 | 10.99 | 87.66 |
| FEW-SHOT-GPT3.5+TEACH | FS | 88.00 | 54.00 | 12.95 | 88.90 |
| SGP-TOD-CHATGPT-EXT | ZS | 88.00 | 78.00 | 6.25 | 86.15 |
| SGP-TOD-GPT3.5-EXT | ZS | 96.00 | 86.00 | 14.57 | 89.01 |
| Model | Restaurant | ||||
| Sw/o g ↑ | Sw/ g ↑ | Und. ↑ | App. ↑ | T. ↓ | |
| SOLOIST | 34.00 | 30.00 | 2.18 | 2.10 | 10.64 |
| FEW-SHOT-CHATGPT | 94.00 | 74.00 | 4.58 | 4.72 | 8.32 |
| SGP-TOD-CHATGPT | 100.00 | 92.00 | 4.86 | 4.88 | 7.28 |
| Model | Multiwoz 2.2 | |||
| Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.60 | 33.90 | 6.29 | 42.54 |
| Model | Task | Schema types | Training strategy | ||
| Fine-tuning | Pre-training | Prompting | |||
| ROBUSTSF (Shah et al., 2019) | SF | slot names/value examples | ✓ | ||
| TRADE (Wu et al., 2019) | DST | slot names/value examples | ✓ | ||
| ZSTL-SD (Campaigna et al., 2020) | DST | ontology, dialog templates | ✓(+synthesized data) | ||
| S-DST (Rastogi et al., 2020b) | DST | slot names/descriptionss+service, intent names/descriptionss | ✓ | ||
| T5DST (Lin et al., 2021b) | DST | slot names/descriptionss | ✓ | ||
| TRANSFERQA (Lin et al., 2021a) | DST | slot names/value examples | ✓(QA tasks) | ||
| IC-DST (Hu et al., 2022) | DST | slot names/value examples | ✓ | ||
| SDM-DST (Wang et al., 2022a) | DST | slot names/value examples | ✓ | ||
| BERT+S (Mosig et al., 2020) | E2E policy | system-side policy skeletons | ✓ | ||
| SAM (Mehri and Eskenazi, 2021) | E2E policy | user-aware policy skeletons | ✓ | ||
| ZSDG (Zhao and Eskenazi, 2018) | E2E dialog | ontology, response templates | ✓ | ||
| DAML (Qian and Yu, 2019) | E2E dialog | ontology, response templates | ✓ | ||
| ANYTOD (Zhao et al., 2022) | E2E dialog | policy programs+slot names/value examples+slot descriptions+user action names/states/descriptionss | ✓ | ✓(heterogeneous tasks) | |
| IG-TOD (Hudecek and Dusek, 2023) | E2E dialog | slot names+slot descriptions | ✓ | ||
| SGP-TOD (ours) | E2E dialog | user-aware policy skeletons(+slot names/value examples) | ✓ | ||
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 82.28 | 55.65 | 6.51 | 75.48 | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.20 | 50.95 | 6.48 | 72.56 | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.74 | 33.13 | 6.18 | 42.12 | 38.60 | 33.90 | 6.29 | 42.54 |
| Slot | Value |
| "address" | "21 - 24 Northampton Street" |
| "area" | "west" |
| "food" | "british" |
| "id" | "14810" |
| "location" | [52.21031, 0.11381] |
| "name" | "saint johns chop house" |
| "phone" | "01223353110" |
| "postcode" | "cb30ad" |
| "pricerange" | "moderate" |
| "type" | "restaurant" |
| "delivery" | "Yes" |
| "delivery fee" | "6 pounds" |
| "dish" | "Beef Wellington" |
| "start_time" | "10:30 am" |
| "end_time" | "22:40 pm" |
| Turn | Utterance |
| 1 | User: I want a restaurant on the west end of town. |
| 2 | System: Ok. Are you looking for any particular type of food? |
| 3 | User: I don't care. |
| 4 | System: Are you looking for a particular price range? |
| 5 | User: A moderately priced restaurant would be good. |
| 6 | System: How about Saint Johns Chop House? They serve moderately priced British food on the west side of town. |
| 7 | User: Can i please have their address and phone number? |
| 8 | System: Saint Johns Chop House's address is 21-24 Northampton Street, their phone number is 01223353110. |
| 9 | User: Does the restaurant offers delivery service? If it does, what's the delivery fee? |
| 10 | System: Yes, it does, and the delivery charges 6 pounds. |
| 11 | User: Thank you, goodbye. |
| 12 | System: Thank you for using our system. Goodbye. |
| Turn | Interaction between a user and ChatGPT | Turn | Interaction between a user and SP-TOD-ChatGPT-Ext |
| 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. | 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. |
| 2 | System: Certainly, there are 3 Tuscan restaurants in town. What is your preferred price range? | 2 | System: I am sorry, there are no Tuscan restaurants. Here are some other options: Korean, Italian, Chinese. Which one would you like more information on? |
| 3 | User: I don't care about the price. Please recommend one for me, i also need the phone number please. | 3 | User: How about any Korean restaurants? I also need the phone number please. |
| 4 | System: The phone number for Pizza Hut City Centre is 01223323737. Is there anything else i can help you with today? | 4 | System: Little Seoul is a Korean restaurant. Their phone number is 01223308681. |
| 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? | 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? |
| 6 | System: I'm sorry, but I could not find any information on delivery fee in our database. You may want to contact the restaurant directly to inquire about this. Is there anything else I can assist you with today? | 6 | System: Yes, they offer delivery service and the delivery charge is 4 pounds. Can I help you with anything else? |
| 7 | User: No. Thank you, goodbye. | 7 | User: No. Thank you, goodbye. |
| 8 | System: Glad i could assist you. Have a good day. | 8 | System: You're welcome, have a great meal! Goodbye! |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| Full-shot fine-tuning (with 8.4k+ training dialogs): | ||||||||
| DAMD (Zhang et al., 2020) | 76.33 | 60.40 | 16.60 | 84.97 | - | - | - | - |
| SIMPLETOD (Hosseini-Asl et al., 2020) | 84.40 | 70.10 | 15.01 | 92.26 | - | - | - | - |
| SOLOIST (Peng et al., 2021a) | 85.50 | 72.90 | 16.54 | 95.74 | 81.70 | 67.10 | 13.60 | 88.00 |
| PPTOD (Su et al., 2022) | 89.20 | 79.40 | 18.62 | 102.92 | - | - | - | - |
| MARS (Sun et al., 2022) | 88.90 | 78.00 | 19.90 | 103.35 | 88.90 | 78.00 | 19.60 | 103.05 |
| Zero-shot transfer method (pre-trained on SGD): | ||||||||
| ANYTOD-XXL | - | - | - | - | 73.90 | 24.40 | 3.40 | 52.55 |
| Few-shot prompting: | ||||||||
| IG-TOD-CHATGPT-FS | - | - | - | - | - | 20.00 | 7.17 | - |
| FEW-SHOT-CHATGPT | 44.74 | 24.32 | 7.88 | 42.41 | 45.40 | 24.50 | 7.72 | 42.67 |
| Zero-shot prompting: | ||||||||
| IG-TOD-CHATGPT-ZS | - | - | - | - | - | 15.00 | 3.58 | - |
| SGP-TOD-CHATGPT (Ours) | 64.56 | 54.05 | 7.17 | 66.48 | 64.70 | 54.70 | 6.96 | 66.66 |
| SGP-TOD-CODEX (Ours) | 71.67 | 52.55 | 7.91 | 70.02 | 75.50 | 52.30 | 6.62 | 70.53 |
| SGP-TOD-GPT3.5 (Ours) | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| Model | Attraction | Train | Hotel | Restaurant | ||||||||||||
| Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | Info. | Succ. | BLEU | Comb. | |
| Few-shot fine-tuning (with 50 training dialogs): | ||||||||||||||||
| SIMPLETOD | 65.66 | 46.97 | 5.85 | 62.17 | 59.00 | 44.00 | 7.07 | 58.57 | 62.50 | 40.00 | 7.70 | 58.95 | 75.50 | 44.50 | 11.00 | 71.00 |
| SOLOIST | 86.00 | 65.00 | 12.90 | 88.40 | 80.81 | 64.65 | 9.96 | 82.69 | 74.50 | 43.50 | 8.12 | 67.12 | 81.00 | 55.50 | 12.80 | 81.50 |
| Few-shot prompting: | ||||||||||||||||
| Few-SHOT-CHATGPT | 75.00 | 67.00 | 8.22 | 79.23 | 79.80 | 65.66 | 8.12 | 80.85 | 51.00 | 26.50 | 5.80 | 44.55 | 80.00 | 55.50 | 7.71 | 75.46 |
| Zero-shot prompting: | ||||||||||||||||
| SGP-TOD-CHATGPT | 95.00 | 94.00 | 7.13 | 101.63 | 76.77 | 74.24 | 6.75 | 82.26 | 76.50 | 57.00 | 5.16 | 71.91 | 90.00 | 82.50 | 6.72 | 92.97 |
| SGP-TOD-CODEX | 98.00 | 93.00 | 10.45 | 105.95 | 78.79 | 70.20 | 8.56 | 83.06 | 83.50 | 69.50 | 7.86 | 84.36 | 91.00 | 85.00 | 10.50 | 98.50 |
| SGP-TOD-GPT3.5 | 96.00 | 93.00 | 9.53 | 104.03 | 82.83 | 77.27 | 8.72 | 88.77 | 82.50 | 71.50 | 7.05 | 84.05 | 91.50 | 84.00 | 12.90 | 100.65 |
| Model | Task transfer | Domain transfer | ||
| F1 | Accuracy | F1 | Accuracy | |
| Zero-shot transfer (leave-one fine-tuning with 2.5k training dialogs): | ||||
| BERT+S | 24.25 | 24.89 | 25.70 | 28.56 |
| SAM | 49.82 | 51.30 | 55.91 | 57.92 |
| Zero-shot prompting: | ||||
| SGP-TOD-CODEX-INI | 45.18 | 47.99 | 47.21 | 49.97 |
| SGP-TOD-GPT3.5 | 47.67 | 48.27 | 49.76 | 50.39 |
| SGP-TOD-CODEX | 49.78 | 51.01 | 52.72 | 53.66 |
| SGP-TOD-GPT3.5-E2E | 50.84 | 50.74 | 53.50 | 53.21 |
| Model | FT/FS/ZS | Restaurant-Ext | |||
| Info. | Succ. | BLEU | BERTS. | ||
| Without domain-relevant knowledge | |||||
| ChatGPT | ZS | 44.00 | 6.00 | 4.31 | 85.96 |
| GPT-3.5 | ZS | 34.00 | 16.00 | 8.70 | 84.31 |
| With prior knowledge on Restaurant | |||||
| SOLOIST | FT | 78.00 | 0.00 | 10.62 | 87.24 |
| SGP-TOD-CHATGPT | ZS | 88.00 | 34.00 | 5.45 | 86.11 |
| SGP-TOD-GPT3.5 | ZS | 94.00 | 30.00 | 10.68 | 87.30 |
| With knowledge on Restaurant-Ext | |||||
| SOLOIST+TEACH | FT | 82.00 | 38.00 | 10.99 | 87.66 |
| FEW-SHOT-GPT3.5+TEACH | FS | 88.00 | 54.00 | 12.95 | 88.90 |
| SGP-TOD-CHATGPT-EXT | ZS | 88.00 | 78.00 | 6.25 | 86.15 |
| SGP-TOD-GPT3.5-EXT | ZS | 96.00 | 86.00 | 14.57 | 89.01 |
| Model | Restaurant | ||||
| Sw/o g ↑ | Sw/ g ↑ | Und. ↑ | App. ↑ | T. ↓ | |
| SOLOIST | 34.00 | 30.00 | 2.18 | 2.10 | 10.64 |
| FEW-SHOT-CHATGPT | 94.00 | 74.00 | 4.58 | 4.72 | 8.32 |
| SGP-TOD-CHATGPT | 100.00 | 92.00 | 4.86 | 4.88 | 7.28 |
| Model | Multiwoz 2.2 | |||
| Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.60 | 33.90 | 6.29 | 42.54 |
| Model | Task | Schema types | Training strategy | ||
| Fine-tuning | Pre-training | Prompting | |||
| ROBUSTSF (Shah et al., 2019) | SF | slot names/value examples | ✓ | ||
| TRADE (Wu et al., 2019) | DST | slot names/value examples | ✓ | ||
| ZSTL-SD (Campaigna et al., 2020) | DST | ontology, dialog templates | ✓(+synthesized data) | ||
| S-DST (Rastogi et al., 2020b) | DST | slot names/descriptionss+service, intent names/descriptionss | ✓ | ||
| T5DST (Lin et al., 2021b) | DST | slot names/descriptionss | ✓ | ||
| TRANSFERQA (Lin et al., 2021a) | DST | slot names/value examples | ✓(QA tasks) | ||
| IC-DST (Hu et al., 2022) | DST | slot names/value examples | ✓ | ||
| SDM-DST (Wang et al., 2022a) | DST | slot names/value examples | ✓ | ||
| BERT+S (Mosig et al., 2020) | E2E policy | system-side policy skeletons | ✓ | ||
| SAM (Mehri and Eskenazi, 2021) | E2E policy | user-aware policy skeletons | ✓ | ||
| ZSDG (Zhao and Eskenazi, 2018) | E2E dialog | ontology, response templates | ✓ | ||
| DAML (Qian and Yu, 2019) | E2E dialog | ontology, response templates | ✓ | ||
| ANYTOD (Zhao et al., 2022) | E2E dialog | policy programs+slot names/value examples+slot descriptions+user action names/states/descriptionss | ✓ | ✓(heterogeneous tasks) | |
| IG-TOD (Hudecek and Dusek, 2023) | E2E dialog | slot names+slot descriptions | ✓ | ||
| SGP-TOD (ours) | E2E dialog | user-aware policy skeletons(+slot names/value examples) | ✓ |
| Model | Multiwoz 2.0 | Multiwoz 2.2 | ||||||
| Inform | Success | BLEU | Combined | Inform | Success | BLEU | Combined | |
| SP-TOD-GPT3.5 | 83.88 | 69.87 | 9.09 | 85.97 | 82.00 | 72.50 | 9.22 | 86.47 |
| -policy | 82.28 | 55.65 | 6.51 | 75.48 | 81.80 | 56.20 | 6.63 | 75.63 |
| -policy -DB | 81.20 | 50.95 | 6.48 | 72.56 | 81.40 | 52.30 | 6.57 | 73.42 |
| -policy -DB -belief | 38.74 | 33.13 | 6.18 | 42.12 | 38.60 | 33.90 | 6.29 | 42.54 |
| Slot | Value |
| "address" | "21 - 24 Northampton Street" |
| "area" | "west" |
| "food" | "british" |
| "id" | "14810" |
| "location" | [52.21031, 0.11381] |
| "name" | "saint johns chop house" |
| "phone" | "01223353110" |
| "postcode" | "cb30ad" |
| "pricerange" | "moderate" |
| "type" | "restaurant" |
| "delivery" | "Yes" |
| "delivery fee" | "6 pounds" |
| "dish" | "Beef Wellington" |
| "start_time" | "10:30 am" |
| "end_time" | "22:40 pm" |
| Turn | Utterance |
| 1 | User: I want a restaurant on the west end of town. |
| 2 | System: Ok. Are you looking for any particular type of food? |
| 3 | User: I don't care. |
| 4 | System: Are you looking for a particular price range? |
| 5 | User: A moderately priced restaurant would be good. |
| 6 | System: How about Saint Johns Chop House? They serve moderately priced British food on the west side of town. |
| 7 | User: Can i please have their address and phone number? |
| 8 | System: Saint Johns Chop House's address is 21-24 Northampton Street, their phone number is 01223353110. |
| 9 | User: Does the restaurant offers delivery service? If it does, what's the delivery fee? |
| 10 | System: Yes, it does, and the delivery charges 6 pounds. |
| 11 | User: Thank you, goodbye. |
| 12 | System: Thank you for using our system. Goodbye. |
| Turn | Interaction between a user and ChatGPT | Turn | Interaction between a user and SP-TOD-ChatGPT-Ext |
| 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. | 1 | User: Hi, i'm looking for a place with Tuscan food in any price range. |
| 2 | System: Certainly, there are 3 Tuscan restaurants in town. What is your preferred price range? | 2 | System: I am sorry, there are no Tuscan restaurants. Here are some other options: Korean, Italian, Chinese. Which one would you like more information on? |
| 3 | User: I don't care about the price. Please recommend one for me, i also need the phone number please. | 3 | User: How about any Korean restaurants? I also need the phone number please. |
| 4 | System: The phone number for Pizza Hut City Centre is 01223323737. Is there anything else i can help you with today? | 4 | System: Little Seoul is a Korean restaurant. Their phone number is 01223308681. |
| 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? | 5 | User: Does the restaurant offer delivery service? How much does the delivery charge? |
| 6 | System: I'm sorry, but I could not find any information on delivery fee in our database. You may want to contact the restaurant directly to inquire about this. Is there anything else I can assist you with today? | 6 | System: Yes, they offer delivery service and the delivery charge is 4 pounds. Can I help you with anything else? |
| 7 | User: No. Thank you, goodbye. | 7 | User: No. Thank you, goodbye. |
| 8 | System: Glad i could assist you. Have a good day. | 8 | System: You're welcome, have a great meal! Goodbye! |
| FLOPs range (Tera FLOPs) | Best Baseline | SHARCS | ||
| Acc.(%) ↑ | FLOPs ↓ | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||
| 0-10 | 61.83 | 9.56 | 76.37 | 4.91 |
| 10-20 | 84.19 | 18 | 85.93 | 17.92 |
| 20-30 | 87.53 | 29.86 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||
| 0-10 | 48.87 | 8.76 | 72.7 | 8.94 |
| 10-20 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 83.27 | 27.99 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||
| 0-5 | 48.87 | 4.7 | 64.72 | 3.80 |
| 5-10 | 64.44 | 9.95 | 76.02 | 8.87 |
| 10-15 | 80.38 | 14.91 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||
| 0-4 | 65.61 | 3.78 | 78.92 | 3.99 |
| 4-6 | 76.01 | 5.42 | 81.48 | 5.21 |
| 6-8 | 83.86 | 7.92 | 83.40 | 7.33 |
| Accuracy (%)↑ | Speed up (×)↑ | |
| BERT | 90.90 | 1 |
| DistilBERT | 88.50 | 2 |
| BERxiT | 85.60 | 2.04 |
| FastBERT | 83.99 | 2.17 |
| DeeBERT | 83.82 | 2.14 |
| PABEE | 89.09 | 2.06 |
| RTJ | 84.44 | 2 |
| SHARCS (layer 2) | 88.43 | 2.57 |
| SHARCS (layer 4) | 90.05 | 2.05 |
| QNLI Acc. | TeraFLOPs | |
| RoBERTa Base | 92.8% | 24.5 |
| +Transkimmer | 89.45% | 14.08 |
| +Transkimmer + SHARCS | 88.83% (-0.62%) | 8.56 (~40%↓) |
| Router on RoBERTabase | AUC ↑ |
| SHARCS | 0.78 |
| BERxiT | 0.73 |
| FLOPs range (Tera Flops) | RTJ | DeeBERT | PABEE | FastBERT | BERxiT | SHARCS | ||||||
| Acc. ↑ | FLOPs ↓ | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||||||||||
| 0-10 | 54.27 | 9.59 | 61.83 | 9.56 | 53.98 | 8.53 | 45.39 | 2.84 | 53.69 | 8.41 | 76.37 | 4.91 |
| 10-20 | 82.29 | 19.77 | 77.18 | 19.80 | 84.19 | 18 | 59.79 | 15.06 | 83.39 | 19.91 | 85.93 | 17.92 |
| 20-30 | 86.92 | 25.76 | 81.35 | 23.67 | 87.53 | 29.86 | 77.24 | 29.95 | 87.23 | 27.92 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||||||||||
| 0-10 | 47.32 | 9.46 | 42.32 | 9.97 | 42.97 | 7.51 | 48.87 | 8.76 | 46.86 | 9.88 | 72.7 | 8.94 |
| 10-20 | 58.12 | 15.61 | 61.57 | 19.89 | 66.33 | 17.29 | 70.25 | 19.18 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 78.96 | 28.53 | 80.57 | 29.34 | 81.13 | 27.66 | 83.27 | 27.99 | 80.82 | 25.73 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||||||||||
| 0-5 | 41.92 | 2.76 | 43.98 | 2.76 | 41.45 | 2.76 | 48.87 | 4.7 | 41.47 | 4.07 | 64.72 | 3.80 |
| 5-10 | 52.72 | 9.56 | 59.87 | 9.45 | 64.44 | 9.95 | 64.24 | 9.76 | 55.10 | 9.76 | 76.02 | 8.87 |
| 10-15 | 72.93 | 13.90 | 80.38 | 14.91 | 79.29 | 14.12 | 79.72 | 14.84 | 60.25 | 11.68 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||||||||||
| 0-4 | 60.95 | 3.98 | 56.13 | 3.38 | 56.09 | 3.97 | 65.61 | 3.78 | 56.19 | 3.58 | 78.92 | 3.99 |
| 4-6 | 74.55 | 5.82 | 75.26 | 5.62 | 69.81 | 5.93 | 76.01 | 5.42 | - | - | 81.48 | 5.21 |
| 6-8 | 83.43 | 7.70 | 83.36 | 7.70 | 81.31 | 7.87 | 83.86 | 7.92 | 83.08 | 7.60 | 83.40 | 7.33 |
| FLOPs range (Tera FLOPs) | Best Baseline | SHARCS | ||
| Acc.(%) ↑ | FLOPs ↓ | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||
| 0-10 | 61.83 | 9.56 | 76.37 | 4.91 |
| 10-20 | 84.19 | 18 | 85.93 | 17.92 |
| 20-30 | 87.53 | 29.86 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||
| 0-10 | 48.87 | 8.76 | 72.7 | 8.94 |
| 10-20 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 83.27 | 27.99 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||
| 0-5 | 48.87 | 4.7 | 64.72 | 3.80 |
| 5-10 | 64.44 | 9.95 | 76.02 | 8.87 |
| 10-15 | 80.38 | 14.91 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||
| 0-4 | 65.61 | 3.78 | 78.92 | 3.99 |
| 4-6 | 76.01 | 5.42 | 81.48 | 5.21 |
| 6-8 | 83.86 | 7.92 | 83.40 | 7.33 |
| Accuracy (%)↑ | Speed up (×)↑ | |
| BERT | 90.90 | 1 |
| DistilBERT | 88.50 | 2 |
| BERxiT | 85.60 | 2.04 |
| FastBERT | 83.99 | 2.17 |
| DeeBERT | 83.82 | 2.14 |
| PABEE | 89.09 | 2.06 |
| RTJ | 84.44 | 2 |
| SHARCS (layer 2) | 88.43 | 2.57 |
| SHARCS (layer 4) | 90.05 | 2.05 |
| QNLI Acc. | TeraFLOPs | |
| RoBERTa Base | 92.8% | 24.5 |
| +Transkimmer | 89.45% | 14.08 |
| +Transkimmer + SHARCS | 88.83% (-0.62%) | 8.56 (~40%↓) |
| Router on RoBERTabase | AUC ↑ |
| SHARCS | 0.78 |
| BERxiT | 0.73 |
| FLOPs range (Tera Flops) | RTJ | DeeBERT | PABEE | FastBERT | BERxiT | SHARCS | ||||||
| Acc. ↑ | FLOPs ↓ | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||||||||||
| 0-10 | 54.27 | 9.59 | 61.83 | 9.56 | 53.98 | 8.53 | 45.39 | 2.84 | 53.69 | 8.41 | 76.37 | 4.91 |
| 10-20 | 82.29 | 19.77 | 77.18 | 19.80 | 84.19 | 18 | 59.79 | 15.06 | 83.39 | 19.91 | 85.93 | 17.92 |
| 20-30 | 86.92 | 25.76 | 81.35 | 23.67 | 87.53 | 29.86 | 77.24 | 29.95 | 87.23 | 27.92 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||||||||||
| 0-10 | 47.32 | 9.46 | 42.32 | 9.97 | 42.97 | 7.51 | 48.87 | 8.76 | 46.86 | 9.88 | 72.7 | 8.94 |
| 10-20 | 58.12 | 15.61 | 61.57 | 19.89 | 66.33 | 17.29 | 70.25 | 19.18 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 78.96 | 28.53 | 80.57 | 29.34 | 81.13 | 27.66 | 83.27 | 27.99 | 80.82 | 25.73 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||||||||||
| 0-5 | 41.92 | 2.76 | 43.98 | 2.76 | 41.45 | 2.76 | 48.87 | 4.7 | 41.47 | 4.07 | 64.72 | 3.80 |
| 5-10 | 52.72 | 9.56 | 59.87 | 9.45 | 64.44 | 9.95 | 64.24 | 9.76 | 55.10 | 9.76 | 76.02 | 8.87 |
| 10-15 | 72.93 | 13.90 | 80.38 | 14.91 | 79.29 | 14.12 | 79.72 | 14.84 | 60.25 | 11.68 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||||||||||
| 0-4 | 60.95 | 3.98 | 56.13 | 3.38 | 56.09 | 3.97 | 65.61 | 3.78 | 56.19 | 3.58 | 78.92 | 3.99 |
| 4-6 | 74.55 | 5.82 | 75.26 | 5.62 | 69.81 | 5.93 | 76.01 | 5.42 | - | - | 81.48 | 5.21 |
| 6-8 | 83.43 | 7.70 | 83.36 | 7.70 | 81.31 | 7.87 | 83.86 | 7.92 | 83.08 | 7.60 | 83.40 | 7.33 |
| FLOPs range (Tera FLOPs) | Best Baseline | SHARCS | ||
| Acc.(%) ↑ | FLOPs ↓ | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||
| 0-10 | 61.83 | 9.56 | 76.37 | 4.91 |
| 10-20 | 84.19 | 18 | 85.93 | 17.92 |
| 20-30 | 87.53 | 29.86 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||
| 0-10 | 48.87 | 8.76 | 72.7 | 8.94 |
| 10-20 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 83.27 | 27.99 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||
| 0-5 | 48.87 | 4.7 | 64.72 | 3.80 |
| 5-10 | 64.44 | 9.95 | 76.02 | 8.87 |
| 10-15 | 80.38 | 14.91 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||
| 0-4 | 65.61 | 3.78 | 78.92 | 3.99 |
| 4-6 | 76.01 | 5.42 | 81.48 | 5.21 |
| 6-8 | 83.86 | 7.92 | 83.40 | 7.33 |
| Accuracy (%)↑ | Speed up (×)↑ | |
| BERT | 90.90 | 1 |
| DistilBERT | 88.50 | 2 |
| BERxiT | 85.60 | 2.04 |
| FastBERT | 83.99 | 2.17 |
| DeeBERT | 83.82 | 2.14 |
| PABEE | 89.09 | 2.06 |
| RTJ | 84.44 | 2 |
| SHARCS (layer 2) | 88.43 | 2.57 |
| SHARCS (layer 4) | 90.05 | 2.05 |
| QNLI Acc. | TeraFLOPs | |
| RoBERTa Base | 92.8% | 24.5 |
| +Transkimmer | 89.45% | 14.08 |
| +Transkimmer + SHARCS | 88.83% (-0.62%) | 8.56 (~40%↓) |
| Router on RoBERTabase | AUC ↑ |
| SHARCS | 0.78 |
| BERxiT | 0.73 |
| FLOPs range (Tera Flops) | RTJ | DeeBERT | PABEE | FastBERT | BERxiT | SHARCS | ||||||
| Acc. ↑ | FLOPs ↓ | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||||||||||
| 0-10 | 54.27 | 9.59 | 61.83 | 9.56 | 53.98 | 8.53 | 45.39 | 2.84 | 53.69 | 8.41 | 76.37 | 4.91 |
| 10-20 | 82.29 | 19.77 | 77.18 | 19.80 | 84.19 | 18 | 59.79 | 15.06 | 83.39 | 19.91 | 85.93 | 17.92 |
| 20-30 | 86.92 | 25.76 | 81.35 | 23.67 | 87.53 | 29.86 | 77.24 | 29.95 | 87.23 | 27.92 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||||||||||
| 0-10 | 47.32 | 9.46 | 42.32 | 9.97 | 42.97 | 7.51 | 48.87 | 8.76 | 46.86 | 9.88 | 72.7 | 8.94 |
| 10-20 | 58.12 | 15.61 | 61.57 | 19.89 | 66.33 | 17.29 | 70.25 | 19.18 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 78.96 | 28.53 | 80.57 | 29.34 | 81.13 | 27.66 | 83.27 | 27.99 | 80.82 | 25.73 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||||||||||
| 0-5 | 41.92 | 2.76 | 43.98 | 2.76 | 41.45 | 2.76 | 48.87 | 4.7 | 41.47 | 4.07 | 64.72 | 3.80 |
| 5-10 | 52.72 | 9.56 | 59.87 | 9.45 | 64.44 | 9.95 | 64.24 | 9.76 | 55.10 | 9.76 | 76.02 | 8.87 |
| 10-15 | 72.93 | 13.90 | 80.38 | 14.91 | 79.29 | 14.12 | 79.72 | 14.84 | 60.25 | 11.68 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||||||||||
| 0-4 | 60.95 | 3.98 | 56.13 | 3.38 | 56.09 | 3.97 | 65.61 | 3.78 | 56.19 | 3.58 | 78.92 | 3.99 |
| 4-6 | 74.55 | 5.82 | 75.26 | 5.62 | 69.81 | 5.93 | 76.01 | 5.42 | - | - | 81.48 | 5.21 |
| 6-8 | 83.43 | 7.70 | 83.36 | 7.70 | 81.31 | 7.87 | 83.86 | 7.92 | 83.08 | 7.60 | 83.40 | 7.33 |
| FLOPs range (Tera FLOPs) | Best Baseline | SHARCS | ||
| Acc.(%) ↑ | FLOPs ↓ | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||
| 0-10 | 61.83 | 9.56 | 76.37 | 4.91 |
| 10-20 | 84.19 | 18 | 85.93 | 17.92 |
| 20-30 | 87.53 | 29.86 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||
| 0-10 | 48.87 | 8.76 | 72.7 | 8.94 |
| 10-20 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 83.27 | 27.99 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||
| 0-5 | 48.87 | 4.7 | 64.72 | 3.80 |
| 5-10 | 64.44 | 9.95 | 76.02 | 8.87 |
| 10-15 | 80.38 | 14.91 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||
| 0-4 | 65.61 | 3.78 | 78.92 | 3.99 |
| 4-6 | 76.01 | 5.42 | 81.48 | 5.21 |
| 6-8 | 83.86 | 7.92 | 83.40 | 7.33 |
| Accuracy (%)↑ | Speed up (×)↑ | |
| BERT | 90.90 | 1 |
| DistilBERT | 88.50 | 2 |
| BERxiT | 85.60 | 2.04 |
| FastBERT | 83.99 | 2.17 |
| DeeBERT | 83.82 | 2.14 |
| PABEE | 89.09 | 2.06 |
| RTJ | 84.44 | 2 |
| SHARCS (layer 2) | 88.43 | 2.57 |
| SHARCS (layer 4) | 90.05 | 2.05 |
| QNLI Acc. | TeraFLOPs | |
| RoBERTa Base | 92.8% | 24.5 |
| +Transkimmer | 89.45% | 14.08 |
| +Transkimmer + SHARCS | 88.83% (-0.62%) | 8.56 (~40%↓) |
| Router on RoBERTabase | AUC ↑ |
| SHARCS | 0.78 |
| BERxiT | 0.73 |
| FLOPs range (Tera Flops) | RTJ | DeeBERT | PABEE | FastBERT | BERxiT | SHARCS | ||||||
| Acc. ↑ | FLOPs ↓ | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | Acc. | FLOPs | |
| Roberta (Acc: 87.6, FLOPs: 33.5) | ||||||||||||
| 0-10 | 54.27 | 9.59 | 61.83 | 9.56 | 53.98 | 8.53 | 45.39 | 2.84 | 53.69 | 8.41 | 76.37 | 4.91 |
| 10-20 | 82.29 | 19.77 | 77.18 | 19.80 | 84.19 | 18 | 59.79 | 15.06 | 83.39 | 19.91 | 85.93 | 17.92 |
| 20-30 | 86.92 | 25.76 | 81.35 | 23.67 | 87.53 | 29.86 | 77.24 | 29.95 | 87.23 | 27.92 | 87.38 | 28.35 |
| BERT (Acc: 84.8, FLOPs: 33) | ||||||||||||
| 0-10 | 47.32 | 9.46 | 42.32 | 9.97 | 42.97 | 7.51 | 48.87 | 8.76 | 46.86 | 9.88 | 72.7 | 8.94 |
| 10-20 | 58.12 | 15.61 | 61.57 | 19.89 | 66.33 | 17.29 | 70.25 | 19.18 | 72.50 | 19.99 | 81.61 | 16.73 |
| 20-30 | 78.96 | 28.53 | 80.57 | 29.34 | 81.13 | 27.66 | 83.27 | 27.99 | 80.82 | 25.73 | 83.04 | 22.76 |
| DistilBERT (Acc: 82.2, FLOPs: 16.57) | ||||||||||||
| 0-5 | 41.92 | 2.76 | 43.98 | 2.76 | 41.45 | 2.76 | 48.87 | 4.7 | 41.47 | 4.07 | 64.72 | 3.80 |
| 5-10 | 52.72 | 9.56 | 59.87 | 9.45 | 64.44 | 9.95 | 64.24 | 9.76 | 55.10 | 9.76 | 76.02 | 8.87 |
| 10-15 | 72.93 | 13.90 | 80.38 | 14.91 | 79.29 | 14.12 | 79.72 | 14.84 | 60.25 | 11.68 | 81.61 | 14.39 |
| DynaBERT 0.25 width (Acc: 83.9, FLOPs: 8.26) | ||||||||||||
| 0-4 | 60.95 | 3.98 | 56.13 | 3.38 | 56.09 | 3.97 | 65.61 | 3.78 | 56.19 | 3.58 | 78.92 | 3.99 |
| 4-6 | 74.55 | 5.82 | 75.26 | 5.62 | 69.81 | 5.93 | 76.01 | 5.42 | - | - | 81.48 | 5.21 |
| 6-8 | 83.43 | 7.70 | 83.36 | 7.70 | 81.31 | 7.87 | 83.86 | 7.92 | 83.08 | 7.60 | 83.40 | 7.33 |
| Base PLM | Models | Syntax | External Resource | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | ||||
| BERT | DGEDT-BERT2 | Dep. Graph | x | 79.8 | 75.6 | 86.3 | 80.0 | 77.9 | 75.4 | - | - |
| kumaGCN | Dep. Graph | x | 82.0 | 78.8 | 86.4 | 80.3 | 77.9 | 77.0 | - | - | |
| dotGCN | x | x | 81.0 | 78.1 | 86.2 | 80.5 | 78.1 | 77.0 | 85.0 | 84.4 | |
| CHGMAN | Dep. Graph | x | 81.5 | 77.7 | 86.9 | 81.6 | - | - | 85.1 | 84.3 | |
| RGAT-BERT4 | Dep. Graph | x | 78.2 | 74.1 | 86.6 | 81.4 | 76.2 | 74.9 | - | - | |
| DGNN (BERT)4 | Dep. Graph | x | 81.4 | 79.0 | 87.2 | 81.7 | 76.2 | 75.0 | - | - | |
| WMM-GCN (BERT)3 | Dep. Graph | x | 82.8 | 79.5 | 88.5 | 82.6 | 78.9 | 77.4 | - | - | |
| Sentic GCN-BERT2 | Dep. Graph | ✓ (SenticNet) | 82.1 | 79.1 | 86.9 | 81.0 | - | - | - | - | |
| SGGCN-BERT | Dep. Graph | x | 82.8 | 80.2 | 87.2 | 82.5 | - | - | - | - | |
| Ours: S1R-ABSC (BERT)2 | Dep. Graph | x | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) | |
| RoBERTa | RoBERTa-RGAT4 | Dep. Graph | x | 83.4 | 80.3 | 87.4 | 80.6 | 74.4 | 72.9 | 84.5 | 83.7 |
| RoBERTa-PWCN2 | Dep. Graph | x | 84.2 | 81.2 | 87.4 | 81.1 | 76.6 | 75.6 | - | - | |
| AM-RoBERTa2 | x | x | 83.0 | 80.2 | 88.1 | 82.5 | - | - | 84.2 | 83.7 | |
| RoBERTa-DLGM2 | Dep. Graph | x | 84.4 | 82.0 | 88.6 | 83.6 | 75.5 | 74.6 | 84.8 | 84.3 | |
| KGAN | Dep. Graph | ✓ (WordNet) | 83.9 | 81.1 | 88.5 | 84.1 | 80.6 | 79.6 | - | - | |
| Ours: S1R-ABSC2 | Dep. Graph | x | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.0(±0.3) | 84.5(±0.3) |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMs (Val.) | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| RoBERTa-ASC | 82.1(±0.9) | 78.9(±1.1) | 87.6(±0.8) | 81.7(±0.8) | 75.6(±0.6) | 74.5(±0.7) | 84.4(±0.4) | 83.9(±0.5) |
| SIR-ABSC (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4(±0.7) | 80.4(±0.8) | 88.5(±0.6) | 83.2(±0.6) | 76.5(±0.8) | 75.3(±0.7) | 84.7(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.7(±0.7) | 80.7(±0.7) | 88.6(±0.6) | 83.3(±0.5) | 76.4(±0.7) | 75.4(±0.8) | 84.9(±0.4) | 84.4(±0.5) |
| • Increasing-VDC | 83.7(±0.9) | 80.5(±0.8) | 88.6(±0.7) | 83.2(±0.7) | 76.9(±0.9) | 76.0(±1.0) | 84.8(±0.3) | 84.4(±0.2) |
| SIR-ABSC (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.8(±0.6) | 80.8(±0.7) | 88.2(±0.6) | 82.9(±0.6) | 76.5(±0.7) | 75.4(±0.8) | 84.8(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.9(±0.6) | 80.7(±0.6) | 88.9(±0.7) | 83.7(±0.8) | 76.4(±0.8) | 75.2(±0.7) | 84.8(±0.4) | 84.5(±0.3) |
| • Increasing-VDC | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| SIR-ABSC | Lap14 | Rest14 | MAMs (Val) | ||||||
| Auto-VDC | DAA | Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 |
| ✓ | ✓ | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.4(±0.4) | 84.9(±0.4) |
| ✓ | x | 84.7(±0.6) | 81.9(±0.7) | 89.5(±0.5) | 84.4(±0.7) | 77.3(±0.6) | 76.2(±0.7) | 85.2(±0.3) | 84.8(±0.2) |
| x | ✓ | 84.6(±0.8) | 81.8(±0.8) | 89.4(±0.4) | 84.4(±0.8) | 77.2(±0.6) | 76.3(±0.6) | 85.2(±0.5) | 84.7(±0.5) |
| x | x | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| [g] token | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | 85.0(±0.8) | 82.1(±0.9) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.2(±0.4) | 84.8(±0.4) |
| Variants | ||||||||
| 1. w/o s ↔ g interaction | 84.2(±0.8) | 81.5(±0.9) | 89.3(±0.4) | 84.1(±0.6) | 77.2(±0.9) | 76.1(±0.9) | 85.0(±0.3) | 84.5(±0.3) |
| 2. [g] init. = aspect embed. | 84.7(±0.8) | 81.6(±0.7) | 89.3(±0.6) | 84.3(±0.6) | 77.2(±0.8) | 76.0(±0.9) | 85.0(±0.4) | 84.6(±0.4) |
| 3. SIR-ABSC-[g] (Appendix F) | 83.5(±0.6) | 80.5(±0.6) | 88.3(±0.4) | 82.9(±0.4) | 75.6(±0.8) | 74.3(±0.8) | 84.6(±0.3) | 84.2(±0.3) |
| Dataset | Train | Test | Val |
| Restaurant (SemEval-2014) | 3608 | 1120 | - |
| Laptop (SemEval-2014) | 2328 | 638 | - |
| 6248 | 692 | - | |
| MAMS-ATSA | 11186 | 1336 | 1332 |
| Datasets | Distribution | Train | Test | Val |
| Lap14 | % of multiple sent./item | 7.86 | 7.84 | - |
| Avg. sent./item | 1.09 | 1.09 | - | |
| Res14 | % of multiple sent./item | 4.02 | 4.38 | - |
| Avg. number of sent./item | 1.04 | 1.05 | - | |
| % of multiple sent./item | 59.44 | 60.55 | - | |
| Avg. number of sent./item | 1.99 | 1.96 | - | |
| MAMS-ATSA | % of multiple sent./item | 2.12 | 0.82 | 0.60 |
| Avg. number of sent./item | 1.02 | 1.01 | 1.01 |
| Pooler types | Lap14 | Rest14 | ||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | ||||||
| w/ max pooling | 83.2 | 80.0 | 88.7 | 83.3 | 74.8 | 73.7 |
| w/ avg pooling | 83.8 | 80.6 | 88.8 | 83.5 | 76.5 | 75.5 |
| w/ att pooling | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 |
| Sentiment: Neutral 0.01 1.84 0.39 0.08 0.05 0.04 0.40 X X 0.2 1.58 2.22 5.21 13.65 14.76 17.82 32.86 3.71 |
| Sentiment: Negative 0.22 6.14 0.39 5.56 1.50 5.07 2.44 X X 6.20 5.90 7.73 10.67 23.74 8.58 1.65 9.28 0.08 |
| Prediction: | Negative | 0.23 | 4.50 | 0.13 | 0.32 | 12.80 | 3.20 | 14.41 | 8.87 | 7.10 | 26.66 |
| Prediction: | Positive | 0.49 | 9.40 | 4.01 | 5.56 | 9.08 | 4.44 | 13.38 | 11.07 | 15.44 | 10.59 |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| GoBERTA (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4 | 80.4 | 88.5 | 83.2 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [222211110000] | 83.3 | 80.2 | 88.2 | 82.7 | 76.0 | 74.8 | 84.7 | 84.2 |
| • VDC = [333222111000] | 82.0 | 78.8 | 88.4 | 82.8 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [444422220000] | 83.4 | 80.4 | 88.4 | 83.1 | 75.6 | 74.4 | 84.7 | 84.2 |
| • VDC = [554433221100] | 83.1 | 79.9 | 88.5 | 83.2 | 75.2 | 73.7 | 84.6 | 84.2 |
| • VDC = [666444222000] | 83.2 | 80.0 | 87.8 | 82.0 | 76.0 | 74.8 | 84.7 | 84.2 |
| • Constant-VDC | 83.7 | 80.7 | 88.6 | 83.3 | 76.4 | 75.4 | 84.9 | 84.4 |
| • Increasing-VDC | 83.7 | 80.5 | 88.6 | 83.2 | 76.9 | 76.0 | 84.8 | 84.4 |
| • VDC = [000011112222] | 83.5 | 80.3 | 87.8 | 82.1 | 76.9 | 76.0 | 84.7 | 84.3 |
| • VDC = [000111222333] | 83.5 | 80.5 | 88.5 | 83.2 | 75.5 | 74.4 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.6 | 80.4 | 87.9 | 82.2 | 75.7 | 74.4 | 84.8 | 84.4 |
| • VDC = [001122334455] | 83.3 | 80.3 | 88.6 | 83.1 | 76.1 | 74.9 | 84.8 | 84.4 |
| • VDC = [000222444666] | 83.7 | 80.5 | 88.3 | 82.5 | 76.6 | 75.8 | 84.7 | 84.2 |
| GoBERTA (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.7 | 80.6 | 88.2 | 82.8 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [222211110000] | 83.5 | 80.4 | 88.1 | 82.7 | 75.4 | 74.2 | 84.7 | 84.3 |
| • VDC = [333222111000] | 82.6 | 79.6 | 87.1 | 81.1 | 76.1 | 75.1 | 84.5 | 84.0 |
| • VDC = [444422220000] | 83.4 | 80.5 | 87.9 | 82.2 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [554433221100] | 83.8 | 80.8 | 88.2 | 82.9 | 75.3 | 74.2 | 84.6 | 84.1 |
| • VDC = [666444222000] | 83.2 | 80.0 | 88.2 | 82.7 | 75.6 | 74.3 | 84.5 | 84.0 |
| • Constant-VDC | 83.7 | 80.4 | 88.9 | 83.7 | 76.4 | 75.2 | 85.0 | 84.5 |
| • Increasing-VDC | 83.8 | 80.8 | 89.1 | 83.8 | 77.1 | 75.9 | 85.0 | 84.6 |
| • VDC = [000011112222] | 84.1 | 81.1 | 88.3 | 82.8 | 77.2 | 76.3 | 84.6 | 84.2 |
| • VDC = [000111222333] | 83.8 | 80.8 | 89.1 | 83.8 | 76.7 | 75.5 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.5 | 80.5 | 88.9 | 83.5 | 75.7 | 74.6 | 84.8 | 84.3 |
| • VDC = [001122334455] | 83.2 | 80.2 | 89.3 | 84.1 | 74.7 | 75.9 | 85.0 | 84.6 |
| • VDC = [000222444666] | 82.5 | 79.4 | 88.9 | 83.8 | 76.9 | 75.9 | 84.8 | 84.5 |
| Ablations (Validation) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| 1. SIR-ABSC (RoBERTa) | ||||||||
| #1: RoBERTa baseline | 82.1 | 78.9 | 87.6 | 81.7 | 75.6 | 74.5 | 84.4 | 83.9 |
| #2: #1+[g] token+VDC | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 | 85.0 | 84.6 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 84.2 | 81.3 | 89.3 | 84.3 | 77.2 | 76.0 | 84.8 | 84.3 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 84.4 | 81.5 | 89.4 | 84.3 | 77.3 | 76.5 | 84.9 | 84.4 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 84.6 | 81.8 | 89.4 | 84.4 | 77.2 | 76.3 | 85.2 | 84.7 |
| #6 (SIR-ABSC): #5+Auto-VDC | 85.0 | 82.1 | 89.7 | 84.8 | 77.5 | 76.4 | 85.4 | 84.9 |
| 2. SIR-ABSC (BERT) | ||||||||
| #1: BERT baseline | 79.2 | 75.1 | 85.8 | 79.3 | 76.4 | 74.8 | 84.0 | 82.7 |
| #2: #1+[g] token+VDC | 80.8 | 77.1 | 86.7 | 80.6 | 77.1 | 75.9 | 84.3 | 83.1 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 81.1 | 77.5 | 86.9 | 80.9 | 77.2 | 75.9 | 84.3 | 83.2 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 81.6 | 78.2 | 87.0 | 81.1 | 77.4 | 76.1 | 84.4 | 83.5 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 81.8 | 78.6 | 87.0 | 81.2 | 77.6 | 76.4 | 84.5 | 83.5 |
| #6 (SIR-ABSC (BERT)): #5+Auto-VDC | 82.2 | 79.1 | 87.3 | 81.6 | 77.7 | 76.6 | 84.7 | 83.6 |
| SIR-ABSC (BERT) | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| BERT baseline | 79.2(±0.6) | 75.1(±0.5) | 85.8(±0.7) | 79.3(±0.7) | 76.4(±0.7) | 74.8(±07) | 84.0(±0.3) | 82.7(±0.4) |
| SIR-ABSC (BERT) | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) |
| Base PLM | Models | Syntax | External Resource | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | ||||
| BERT | DGEDT-BERT2 | Dep. Graph | x | 79.8 | 75.6 | 86.3 | 80.0 | 77.9 | 75.4 | - | - |
| kumaGCN | Dep. Graph | x | 82.0 | 78.8 | 86.4 | 80.3 | 77.9 | 77.0 | - | - | |
| dotGCN | x | x | 81.0 | 78.1 | 86.2 | 80.5 | 78.1 | 77.0 | 85.0 | 84.4 | |
| CHGMAN | Dep. Graph | x | 81.5 | 77.7 | 86.9 | 81.6 | - | - | 85.1 | 84.3 | |
| RGAT-BERT4 | Dep. Graph | x | 78.2 | 74.1 | 86.6 | 81.4 | 76.2 | 74.9 | - | - | |
| DGNN (BERT)4 | Dep. Graph | x | 81.4 | 79.0 | 87.2 | 81.7 | 76.2 | 75.0 | - | - | |
| WMM-GCN (BERT)3 | Dep. Graph | x | 82.8 | 79.5 | 88.5 | 82.6 | 78.9 | 77.4 | - | - | |
| Sentic GCN-BERT2 | Dep. Graph | ✓ (SenticNet) | 82.1 | 79.1 | 86.9 | 81.0 | - | - | - | - | |
| SGGCN-BERT | Dep. Graph | x | 82.8 | 80.2 | 87.2 | 82.5 | - | - | - | - | |
| Ours: S1R-ABSC (BERT)2 | Dep. Graph | x | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) | |
| RoBERTa | RoBERTa-RGAT4 | Dep. Graph | x | 83.4 | 80.3 | 87.4 | 80.6 | 74.4 | 72.9 | 84.5 | 83.7 |
| RoBERTa-PWCN2 | Dep. Graph | x | 84.2 | 81.2 | 87.4 | 81.1 | 76.6 | 75.6 | - | - | |
| AM-RoBERTa2 | x | x | 83.0 | 80.2 | 88.1 | 82.5 | - | - | 84.2 | 83.7 | |
| RoBERTa-DLGM2 | Dep. Graph | x | 84.4 | 82.0 | 88.6 | 83.6 | 75.5 | 74.6 | 84.8 | 84.3 | |
| KGAN | Dep. Graph | ✓ (WordNet) | 83.9 | 81.1 | 88.5 | 84.1 | 80.6 | 79.6 | - | - | |
| Ours: S1R-ABSC2 | Dep. Graph | x | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.0(±0.3) | 84.5(±0.3) |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMs (Val.) | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| RoBERTa-ASC | 82.1(±0.9) | 78.9(±1.1) | 87.6(±0.8) | 81.7(±0.8) | 75.6(±0.6) | 74.5(±0.7) | 84.4(±0.4) | 83.9(±0.5) |
| SIR-ABSC (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4(±0.7) | 80.4(±0.8) | 88.5(±0.6) | 83.2(±0.6) | 76.5(±0.8) | 75.3(±0.7) | 84.7(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.7(±0.7) | 80.7(±0.7) | 88.6(±0.6) | 83.3(±0.5) | 76.4(±0.7) | 75.4(±0.8) | 84.9(±0.4) | 84.4(±0.5) |
| • Increasing-VDC | 83.7(±0.9) | 80.5(±0.8) | 88.6(±0.7) | 83.2(±0.7) | 76.9(±0.9) | 76.0(±1.0) | 84.8(±0.3) | 84.4(±0.2) |
| SIR-ABSC (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.8(±0.6) | 80.8(±0.7) | 88.2(±0.6) | 82.9(±0.6) | 76.5(±0.7) | 75.4(±0.8) | 84.8(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.9(±0.6) | 80.7(±0.6) | 88.9(±0.7) | 83.7(±0.8) | 76.4(±0.8) | 75.2(±0.7) | 84.8(±0.4) | 84.5(±0.3) |
| • Increasing-VDC | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| SIR-ABSC | Lap14 | Rest14 | MAMs (Val) | ||||||
| Auto-VDC | DAA | Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 |
| ✓ | ✓ | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.4(±0.4) | 84.9(±0.4) |
| ✓ | x | 84.7(±0.6) | 81.9(±0.7) | 89.5(±0.5) | 84.4(±0.7) | 77.3(±0.6) | 76.2(±0.7) | 85.2(±0.3) | 84.8(±0.2) |
| x | ✓ | 84.6(±0.8) | 81.8(±0.8) | 89.4(±0.4) | 84.4(±0.8) | 77.2(±0.6) | 76.3(±0.6) | 85.2(±0.5) | 84.7(±0.5) |
| x | x | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| [g] token | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | 85.0(±0.8) | 82.1(±0.9) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.2(±0.4) | 84.8(±0.4) |
| Variants | ||||||||
| 1. w/o s ↔ g interaction | 84.2(±0.8) | 81.5(±0.9) | 89.3(±0.4) | 84.1(±0.6) | 77.2(±0.9) | 76.1(±0.9) | 85.0(±0.3) | 84.5(±0.3) |
| 2. [g] init. = aspect embed. | 84.7(±0.8) | 81.6(±0.7) | 89.3(±0.6) | 84.3(±0.6) | 77.2(±0.8) | 76.0(±0.9) | 85.0(±0.4) | 84.6(±0.4) |
| 3. SIR-ABSC-[g] (Appendix F) | 83.5(±0.6) | 80.5(±0.6) | 88.3(±0.4) | 82.9(±0.4) | 75.6(±0.8) | 74.3(±0.8) | 84.6(±0.3) | 84.2(±0.3) |
| Dataset | Train | Test | Val |
| Restaurant (SemEval-2014) | 3608 | 1120 | - |
| Laptop (SemEval-2014) | 2328 | 638 | - |
| 6248 | 692 | - | |
| MAMS-ATSA | 11186 | 1336 | 1332 |
| Datasets | Distribution | Train | Test | Val |
| Lap14 | % of multiple sent./item | 7.86 | 7.84 | - |
| Avg. sent./item | 1.09 | 1.09 | - | |
| Res14 | % of multiple sent./item | 4.02 | 4.38 | - |
| Avg. number of sent./item | 1.04 | 1.05 | - | |
| % of multiple sent./item | 59.44 | 60.55 | - | |
| Avg. number of sent./item | 1.99 | 1.96 | - | |
| MAMS-ATSA | % of multiple sent./item | 2.12 | 0.82 | 0.60 |
| Avg. number of sent./item | 1.02 | 1.01 | 1.01 |
| Pooler types | Lap14 | Rest14 | ||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | ||||||
| w/ max pooling | 83.2 | 80.0 | 88.7 | 83.3 | 74.8 | 73.7 |
| w/ avg pooling | 83.8 | 80.6 | 88.8 | 83.5 | 76.5 | 75.5 |
| w/ att pooling | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 |
| Sentiment: Neutral 0.01 1.84 0.39 0.08 0.05 0.04 0.40 X X 0.2 1.58 2.22 5.21 13.65 14.76 17.82 32.86 3.71 |
| Sentiment: Negative 0.22 6.14 0.39 5.56 1.50 5.07 2.44 X X 6.20 5.90 7.73 10.67 23.74 8.58 1.65 9.28 0.08 |
| Prediction: | Negative | 0.23 | 4.50 | 0.13 | 0.32 | 12.80 | 3.20 | 14.41 | 8.87 | 7.10 | 26.66 |
| Prediction: | Positive | 0.49 | 9.40 | 4.01 | 5.56 | 9.08 | 4.44 | 13.38 | 11.07 | 15.44 | 10.59 |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| GoBERTA (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4 | 80.4 | 88.5 | 83.2 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [222211110000] | 83.3 | 80.2 | 88.2 | 82.7 | 76.0 | 74.8 | 84.7 | 84.2 |
| • VDC = [333222111000] | 82.0 | 78.8 | 88.4 | 82.8 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [444422220000] | 83.4 | 80.4 | 88.4 | 83.1 | 75.6 | 74.4 | 84.7 | 84.2 |
| • VDC = [554433221100] | 83.1 | 79.9 | 88.5 | 83.2 | 75.2 | 73.7 | 84.6 | 84.2 |
| • VDC = [666444222000] | 83.2 | 80.0 | 87.8 | 82.0 | 76.0 | 74.8 | 84.7 | 84.2 |
| • Constant-VDC | 83.7 | 80.7 | 88.6 | 83.3 | 76.4 | 75.4 | 84.9 | 84.4 |
| • Increasing-VDC | 83.7 | 80.5 | 88.6 | 83.2 | 76.9 | 76.0 | 84.8 | 84.4 |
| • VDC = [000011112222] | 83.5 | 80.3 | 87.8 | 82.1 | 76.9 | 76.0 | 84.7 | 84.3 |
| • VDC = [000111222333] | 83.5 | 80.5 | 88.5 | 83.2 | 75.5 | 74.4 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.6 | 80.4 | 87.9 | 82.2 | 75.7 | 74.4 | 84.8 | 84.4 |
| • VDC = [001122334455] | 83.3 | 80.3 | 88.6 | 83.1 | 76.1 | 74.9 | 84.8 | 84.4 |
| • VDC = [000222444666] | 83.7 | 80.5 | 88.3 | 82.5 | 76.6 | 75.8 | 84.7 | 84.2 |
| GoBERTA (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.7 | 80.6 | 88.2 | 82.8 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [222211110000] | 83.5 | 80.4 | 88.1 | 82.7 | 75.4 | 74.2 | 84.7 | 84.3 |
| • VDC = [333222111000] | 82.6 | 79.6 | 87.1 | 81.1 | 76.1 | 75.1 | 84.5 | 84.0 |
| • VDC = [444422220000] | 83.4 | 80.5 | 87.9 | 82.2 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [554433221100] | 83.8 | 80.8 | 88.2 | 82.9 | 75.3 | 74.2 | 84.6 | 84.1 |
| • VDC = [666444222000] | 83.2 | 80.0 | 88.2 | 82.7 | 75.6 | 74.3 | 84.5 | 84.0 |
| • Constant-VDC | 83.7 | 80.4 | 88.9 | 83.7 | 76.4 | 75.2 | 85.0 | 84.5 |
| • Increasing-VDC | 83.8 | 80.8 | 89.1 | 83.8 | 77.1 | 75.9 | 85.0 | 84.6 |
| • VDC = [000011112222] | 84.1 | 81.1 | 88.3 | 82.8 | 77.2 | 76.3 | 84.6 | 84.2 |
| • VDC = [000111222333] | 83.8 | 80.8 | 89.1 | 83.8 | 76.7 | 75.5 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.5 | 80.5 | 88.9 | 83.5 | 75.7 | 74.6 | 84.8 | 84.3 |
| • VDC = [001122334455] | 83.2 | 80.2 | 89.3 | 84.1 | 74.7 | 75.9 | 85.0 | 84.6 |
| • VDC = [000222444666] | 82.5 | 79.4 | 88.9 | 83.8 | 76.9 | 75.9 | 84.8 | 84.5 |
| Ablations (Validation) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| 1. SIR-ABSC (RoBERTa) | ||||||||
| #1: RoBERTa baseline | 82.1 | 78.9 | 87.6 | 81.7 | 75.6 | 74.5 | 84.4 | 83.9 |
| #2: #1+[g] token+VDC | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 | 85.0 | 84.6 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 84.2 | 81.3 | 89.3 | 84.3 | 77.2 | 76.0 | 84.8 | 84.3 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 84.4 | 81.5 | 89.4 | 84.3 | 77.3 | 76.5 | 84.9 | 84.4 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 84.6 | 81.8 | 89.4 | 84.4 | 77.2 | 76.3 | 85.2 | 84.7 |
| #6 (SIR-ABSC): #5+Auto-VDC | 85.0 | 82.1 | 89.7 | 84.8 | 77.5 | 76.4 | 85.4 | 84.9 |
| 2. SIR-ABSC (BERT) | ||||||||
| #1: BERT baseline | 79.2 | 75.1 | 85.8 | 79.3 | 76.4 | 74.8 | 84.0 | 82.7 |
| #2: #1+[g] token+VDC | 80.8 | 77.1 | 86.7 | 80.6 | 77.1 | 75.9 | 84.3 | 83.1 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 81.1 | 77.5 | 86.9 | 80.9 | 77.2 | 75.9 | 84.3 | 83.2 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 81.6 | 78.2 | 87.0 | 81.1 | 77.4 | 76.1 | 84.4 | 83.5 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 81.8 | 78.6 | 87.0 | 81.2 | 77.6 | 76.4 | 84.5 | 83.5 |
| #6 (SIR-ABSC (BERT)): #5+Auto-VDC | 82.2 | 79.1 | 87.3 | 81.6 | 77.7 | 76.6 | 84.7 | 83.6 |
| SIR-ABSC (BERT) | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| BERT baseline | 79.2(±0.6) | 75.1(±0.5) | 85.8(±0.7) | 79.3(±0.7) | 76.4(±0.7) | 74.8(±07) | 84.0(±0.3) | 82.7(±0.4) |
| SIR-ABSC (BERT) | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) |
| Base PLM | Models | Syntax | External Resource | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | ||||
| BERT | DGEDT-BERT2 | Dep. Graph | x | 79.8 | 75.6 | 86.3 | 80.0 | 77.9 | 75.4 | - | - |
| kumaGCN | Dep. Graph | x | 82.0 | 78.8 | 86.4 | 80.3 | 77.9 | 77.0 | - | - | |
| dotGCN | x | x | 81.0 | 78.1 | 86.2 | 80.5 | 78.1 | 77.0 | 85.0 | 84.4 | |
| CHGMAN | Dep. Graph | x | 81.5 | 77.7 | 86.9 | 81.6 | - | - | 85.1 | 84.3 | |
| RGAT-BERT4 | Dep. Graph | x | 78.2 | 74.1 | 86.6 | 81.4 | 76.2 | 74.9 | - | - | |
| DGNN (BERT)4 | Dep. Graph | x | 81.4 | 79.0 | 87.2 | 81.7 | 76.2 | 75.0 | - | - | |
| WMM-GCN (BERT)3 | Dep. Graph | x | 82.8 | 79.5 | 88.5 | 82.6 | 78.9 | 77.4 | - | - | |
| Sentic GCN-BERT2 | Dep. Graph | ✓ (SenticNet) | 82.1 | 79.1 | 86.9 | 81.0 | - | - | - | - | |
| SGGCN-BERT | Dep. Graph | x | 82.8 | 80.2 | 87.2 | 82.5 | - | - | - | - | |
| Ours: S1R-ABSC (BERT)2 | Dep. Graph | x | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) | |
| RoBERTa | RoBERTa-RGAT4 | Dep. Graph | x | 83.4 | 80.3 | 87.4 | 80.6 | 74.4 | 72.9 | 84.5 | 83.7 |
| RoBERTa-PWCN2 | Dep. Graph | x | 84.2 | 81.2 | 87.4 | 81.1 | 76.6 | 75.6 | - | - | |
| AM-RoBERTa2 | x | x | 83.0 | 80.2 | 88.1 | 82.5 | - | - | 84.2 | 83.7 | |
| RoBERTa-DLGM2 | Dep. Graph | x | 84.4 | 82.0 | 88.6 | 83.6 | 75.5 | 74.6 | 84.8 | 84.3 | |
| KGAN | Dep. Graph | ✓ (WordNet) | 83.9 | 81.1 | 88.5 | 84.1 | 80.6 | 79.6 | - | - | |
| Ours: S1R-ABSC2 | Dep. Graph | x | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.0(±0.3) | 84.5(±0.3) | |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMs (Val.) | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| RoBERTa-ASC | 82.1(±0.9) | 78.9(±1.1) | 87.6(±0.8) | 81.7(±0.8) | 75.6(±0.6) | 74.5(±0.7) | 84.4(±0.4) | 83.9(±0.5) |
| SIR-ABSC (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4(±0.7) | 80.4(±0.8) | 88.5(±0.6) | 83.2(±0.6) | 76.5(±0.8) | 75.3(±0.7) | 84.7(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.7(±0.7) | 80.7(±0.7) | 88.6(±0.6) | 83.3(±0.5) | 76.4(±0.7) | 75.4(±0.8) | 84.9(±0.4) | 84.4(±0.5) |
| • Increasing-VDC | 83.7(±0.9) | 80.5(±0.8) | 88.6(±0.7) | 83.2(±0.7) | 76.9(±0.9) | 76.0(±1.0) | 84.8(±0.3) | 84.4(±0.2) |
| SIR-ABSC (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.8(±0.6) | 80.8(±0.7) | 88.2(±0.6) | 82.9(±0.6) | 76.5(±0.7) | 75.4(±0.8) | 84.8(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.9(±0.6) | 80.7(±0.6) | 88.9(±0.7) | 83.7(±0.8) | 76.4(±0.8) | 75.2(±0.7) | 84.8(±0.4) | 84.5(±0.3) |
| • Increasing-VDC | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| SIR-ABSC | Lap14 | Rest14 | MAMs (Val) | ||||||
| Auto-VDC | DAA | Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 |
| ✓ | ✓ | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.4(±0.4) | 84.9(±0.4) |
| ✓ | x | 84.7(±0.6) | 81.9(±0.7) | 89.5(±0.5) | 84.4(±0.7) | 77.3(±0.6) | 76.2(±0.7) | 85.2(±0.3) | 84.8(±0.2) |
| x | ✓ | 84.6(±0.8) | 81.8(±0.8) | 89.4(±0.4) | 84.4(±0.8) | 77.2(±0.6) | 76.3(±0.6) | 85.2(±0.5) | 84.7(±0.5) |
| x | x | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| [g] token | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | 85.0(±0.8) | 82.1(±0.9) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.2(±0.4) | 84.8(±0.4) |
| Variants | ||||||||
| 1. w/o s ↔ g interaction | 84.2(±0.8) | 81.5(±0.9) | 89.3(±0.4) | 84.1(±0.6) | 77.2(±0.9) | 76.1(±0.9) | 85.0(±0.3) | 84.5(±0.3) |
| 2. [g] init. = aspect embed. | 84.7(±0.8) | 81.6(±0.7) | 89.3(±0.6) | 84.3(±0.6) | 77.2(±0.8) | 76.0(±0.9) | 85.0(±0.4) | 84.6(±0.4) |
| 3. SIR-ABSC-[g] (Appendix F) | 83.5(±0.6) | 80.5(±0.6) | 88.3(±0.4) | 82.9(±0.4) | 75.6(±0.8) | 74.3(±0.8) | 84.6(±0.3) | 84.2(±0.3) |
| Dataset | Train | Test | Val |
| Restaurant (SemEval-2014) | 3608 | 1120 | - |
| Laptop (SemEval-2014) | 2328 | 638 | - |
| 6248 | 692 | - | |
| MAMS-ATSA | 11186 | 1336 | 1332 |
| Datasets | Distribution | Train | Test | Val |
| Lap14 | % of multiple sent./item | 7.86 | 7.84 | - |
| Avg. sent./item | 1.09 | 1.09 | - | |
| Res14 | % of multiple sent./item | 4.02 | 4.38 | - |
| Avg. number of sent./item | 1.04 | 1.05 | - | |
| % of multiple sent./item | 59.44 | 60.55 | - | |
| Avg. number of sent./item | 1.99 | 1.96 | - | |
| MAMS-ATSA | % of multiple sent./item | 2.12 | 0.82 | 0.60 |
| Avg. number of sent./item | 1.02 | 1.01 | 1.01 |
| Pooler types | Lap14 | Rest14 | ||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | ||||||
| w/ max pooling | 83.2 | 80.0 | 88.7 | 83.3 | 74.8 | 73.7 |
| w/ avg pooling | 83.8 | 80.6 | 88.8 | 83.5 | 76.5 | 75.5 |
| w/ att pooling | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 |
| Sentiment: Neutral 0.01 1.84 0.39 0.08 0.05 0.04 0.40 X X 0.2 1.58 2.22 5.21 13.65 14.76 17.82 32.86 3.71 | ||||||||||||||||
| Sentiment: Negative 0.22 6.14 0.39 5.56 1.50 5.07 2.44 X X 6.20 5.90 7.73 10.67 23.74 8.58 1.65 9.28 0.08 |
| Prediction: | Negative | 0.23 | 4.50 | 0.13 | 0.32 | 12.80 | 3.20 | 14.41 | 8.87 | 7.10 | 26.66 |
| Prediction: | Positive | 0.49 | 9.40 | 4.01 | 5.56 | 9.08 | 4.44 | 13.38 | 11.07 | 15.44 | 10.59 |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| GoBERTA (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4 | 80.4 | 88.5 | 83.2 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [222211110000] | 83.3 | 80.2 | 88.2 | 82.7 | 76.0 | 74.8 | 84.7 | 84.2 |
| • VDC = [333222111000] | 82.0 | 78.8 | 88.4 | 82.8 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [444422220000] | 83.4 | 80.4 | 88.4 | 83.1 | 75.6 | 74.4 | 84.7 | 84.2 |
| • VDC = [554433221100] | 83.1 | 79.9 | 88.5 | 83.2 | 75.2 | 73.7 | 84.6 | 84.2 |
| • VDC = [666444222000] | 83.2 | 80.0 | 87.8 | 82.0 | 76.0 | 74.8 | 84.7 | 84.2 |
| • Constant-VDC | 83.7 | 80.7 | 88.6 | 83.3 | 76.4 | 75.4 | 84.9 | 84.4 |
| • Increasing-VDC | 83.7 | 80.5 | 88.6 | 83.2 | 76.9 | 76.0 | 84.8 | 84.4 |
| • VDC = [000011112222] | 83.5 | 80.3 | 87.8 | 82.1 | 76.9 | 76.0 | 84.7 | 84.3 |
| • VDC = [000111222333] | 83.5 | 80.5 | 88.5 | 83.2 | 75.5 | 74.4 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.6 | 80.4 | 87.9 | 82.2 | 75.7 | 74.4 | 84.8 | 84.4 |
| • VDC = [001122334455] | 83.3 | 80.3 | 88.6 | 83.1 | 76.1 | 74.9 | 84.8 | 84.4 |
| • VDC = [000222444666] | 83.7 | 80.5 | 88.3 | 82.5 | 76.6 | 75.8 | 84.7 | 84.2 |
| GoBERTA (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.7 | 80.6 | 88.2 | 82.8 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [222211110000] | 83.5 | 80.4 | 88.1 | 82.7 | 75.4 | 74.2 | 84.7 | 84.3 |
| • VDC = [333222111000] | 82.6 | 79.6 | 87.1 | 81.1 | 76.1 | 75.1 | 84.5 | 84.0 |
| • VDC = [444422220000] | 83.4 | 80.5 | 87.9 | 82.2 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [554433221100] | 83.8 | 80.8 | 88.2 | 82.9 | 75.3 | 74.2 | 84.6 | 84.1 |
| • VDC = [666444222000] | 83.2 | 80.0 | 88.2 | 82.7 | 75.6 | 74.3 | 84.5 | 84.0 |
| • Constant-VDC | 83.7 | 80.4 | 88.9 | 83.7 | 76.4 | 75.2 | 85.0 | 84.5 |
| • Increasing-VDC | 83.8 | 80.8 | 89.1 | 83.8 | 77.1 | 75.9 | 85.0 | 84.6 |
| • VDC = [000011112222] | 84.1 | 81.1 | 88.3 | 82.8 | 77.2 | 76.3 | 84.6 | 84.2 |
| • VDC = [000111222333] | 83.8 | 80.8 | 89.1 | 83.8 | 76.7 | 75.5 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.5 | 80.5 | 88.9 | 83.5 | 75.7 | 74.6 | 84.8 | 84.3 |
| • VDC = [001122334455] | 83.2 | 80.2 | 89.3 | 84.1 | 74.7 | 75.9 | 85.0 | 84.6 |
| • VDC = [000222444666] | 82.5 | 79.4 | 88.9 | 83.8 | 76.9 | 75.9 | 84.8 | 84.5 |
| Ablations (Validation) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| 1. SIR-ABSC (RoBERTa) | ||||||||
| #1: RoBERTa baseline | 82.1 | 78.9 | 87.6 | 81.7 | 75.6 | 74.5 | 84.4 | 83.9 |
| #2: #1+[g] token+VDC | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 | 85.0 | 84.6 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 84.2 | 81.3 | 89.3 | 84.3 | 77.2 | 76.0 | 84.8 | 84.3 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 84.4 | 81.5 | 89.4 | 84.3 | 77.3 | 76.5 | 84.9 | 84.4 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 84.6 | 81.8 | 89.4 | 84.4 | 77.2 | 76.3 | 85.2 | 84.7 |
| #6 (SIR-ABSC): #5+Auto-VDC | 85.0 | 82.1 | 89.7 | 84.8 | 77.5 | 76.4 | 85.4 | 84.9 |
| 2. SIR-ABSC (BERT) | ||||||||
| #1: BERT baseline | 79.2 | 75.1 | 85.8 | 79.3 | 76.4 | 74.8 | 84.0 | 82.7 |
| #2: #1+[g] token+VDC | 80.8 | 77.1 | 86.7 | 80.6 | 77.1 | 75.9 | 84.3 | 83.1 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 81.1 | 77.5 | 86.9 | 80.9 | 77.2 | 75.9 | 84.3 | 83.2 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 81.6 | 78.2 | 87.0 | 81.1 | 77.4 | 76.1 | 84.4 | 83.5 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 81.8 | 78.6 | 87.0 | 81.2 | 77.6 | 76.4 | 84.5 | 83.5 |
| #6 (SIR-ABSC (BERT)): #5+Auto-VDC | 82.2 | 79.1 | 87.3 | 81.6 | 77.7 | 76.6 | 84.7 | 83.6 |
| SIR-ABSC (BERT) | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| BERT baseline | 79.2(±0.6) | 75.1(±0.5) | 85.8(±0.7) | 79.3(±0.7) | 76.4(±0.7) | 74.8(±07) | 84.0(±0.3) | 82.7(±0.4) |
| SIR-ABSC (BERT) | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) |
| Base PLM | Models | Syntax | External Resource | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | ||||
| BERT | DGEDT-BERT2 | Dep. Graph | x | 79.8 | 75.6 | 86.3 | 80.0 | 77.9 | 75.4 | - | - |
| kumaGCN | Dep. Graph | x | 82.0 | 78.8 | 86.4 | 80.3 | 77.9 | 77.0 | - | - | |
| dotGCN | x | x | 81.0 | 78.1 | 86.2 | 80.5 | 78.1 | 77.0 | 85.0 | 84.4 | |
| CHGMAN | Dep. Graph | x | 81.5 | 77.7 | 86.9 | 81.6 | - | - | 85.1 | 84.3 | |
| RGAT-BERT4 | Dep. Graph | x | 78.2 | 74.1 | 86.6 | 81.4 | 76.2 | 74.9 | - | - | |
| DGNN (BERT)4 | Dep. Graph | x | 81.4 | 79.0 | 87.2 | 81.7 | 76.2 | 75.0 | - | - | |
| WMM-GCN (BERT)3 | Dep. Graph | x | 82.8 | 79.5 | 88.5 | 82.6 | 78.9 | 77.4 | - | - | |
| Sentic GCN-BERT2 | Dep. Graph | ✓ (SenticNet) | 82.1 | 79.1 | 86.9 | 81.0 | - | - | - | - | |
| SGGCN-BERT | Dep. Graph | x | 82.8 | 80.2 | 87.2 | 82.5 | - | - | - | - | |
| Ours: S1R-ABSC (BERT)2 | Dep. Graph | x | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) | |
| RoBERTa | RoBERTa-RGAT4 | Dep. Graph | x | 83.4 | 80.3 | 87.4 | 80.6 | 74.4 | 72.9 | 84.5 | 83.7 |
| RoBERTa-PWCN2 | Dep. Graph | x | 84.2 | 81.2 | 87.4 | 81.1 | 76.6 | 75.6 | - | - | |
| AM-RoBERTa2 | x | x | 83.0 | 80.2 | 88.1 | 82.5 | - | - | 84.2 | 83.7 | |
| RoBERTa-DLGM2 | Dep. Graph | x | 84.4 | 82.0 | 88.6 | 83.6 | 75.5 | 74.6 | 84.8 | 84.3 | |
| KGAN | Dep. Graph | ✓ (WordNet) | 83.9 | 81.1 | 88.5 | 84.1 | 80.6 | 79.6 | - | - | |
| Ours: S1R-ABSC2 | Dep. Graph | x | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.0(±0.3) | 84.5(±0.3) |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMs (Val.) | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| RoBERTa-ASC | 82.1(±0.9) | 78.9(±1.1) | 87.6(±0.8) | 81.7(±0.8) | 75.6(±0.6) | 74.5(±0.7) | 84.4(±0.4) | 83.9(±0.5) |
| SIR-ABSC (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4(±0.7) | 80.4(±0.8) | 88.5(±0.6) | 83.2(±0.6) | 76.5(±0.8) | 75.3(±0.7) | 84.7(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.7(±0.7) | 80.7(±0.7) | 88.6(±0.6) | 83.3(±0.5) | 76.4(±0.7) | 75.4(±0.8) | 84.9(±0.4) | 84.4(±0.5) |
| • Increasing-VDC | 83.7(±0.9) | 80.5(±0.8) | 88.6(±0.7) | 83.2(±0.7) | 76.9(±0.9) | 76.0(±1.0) | 84.8(±0.3) | 84.4(±0.2) |
| SIR-ABSC (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.8(±0.6) | 80.8(±0.7) | 88.2(±0.6) | 82.9(±0.6) | 76.5(±0.7) | 75.4(±0.8) | 84.8(±0.4) | 84.3(±0.4) |
| • Constant-VDC | 83.9(±0.6) | 80.7(±0.6) | 88.9(±0.7) | 83.7(±0.8) | 76.4(±0.8) | 75.2(±0.7) | 84.8(±0.4) | 84.5(±0.3) |
| • Increasing-VDC | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| SIR-ABSC | Lap14 | Rest14 | MAMs (Val) | ||||||
| Auto-VDC | DAA | Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 |
| ✓ | ✓ | 85.0(±0.5) | 82.1(±0.6) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.4(±0.4) | 84.9(±0.4) |
| ✓ | x | 84.7(±0.6) | 81.9(±0.7) | 89.5(±0.5) | 84.4(±0.7) | 77.3(±0.6) | 76.2(±0.7) | 85.2(±0.3) | 84.8(±0.2) |
| x | ✓ | 84.6(±0.8) | 81.8(±0.8) | 89.4(±0.4) | 84.4(±0.8) | 77.2(±0.6) | 76.3(±0.6) | 85.2(±0.5) | 84.7(±0.5) |
| x | x | 84.1(±0.8) | 81.1(±0.8) | 89.3(±0.6) | 84.1(±0.8) | 77.2(±0.9) | 76.3(±0.9) | 85.0(±0.3) | 84.6(±0.3) |
| [g] token | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | 85.0(±0.8) | 82.1(±0.9) | 89.7(±0.6) | 84.8(±0.9) | 77.5(±0.6) | 76.4(±0.6) | 85.2(±0.4) | 84.8(±0.4) |
| Variants | ||||||||
| 1. w/o s ↔ g interaction | 84.2(±0.8) | 81.5(±0.9) | 89.3(±0.4) | 84.1(±0.6) | 77.2(±0.9) | 76.1(±0.9) | 85.0(±0.3) | 84.5(±0.3) |
| 2. [g] init. = aspect embed. | 84.7(±0.8) | 81.6(±0.7) | 89.3(±0.6) | 84.3(±0.6) | 77.2(±0.8) | 76.0(±0.9) | 85.0(±0.4) | 84.6(±0.4) |
| 3. SIR-ABSC-[g] (Appendix F) | 83.5(±0.6) | 80.5(±0.6) | 88.3(±0.4) | 82.9(±0.4) | 75.6(±0.8) | 74.3(±0.8) | 84.6(±0.3) | 84.2(±0.3) |
| Dataset | Train | Test | Val |
| Restaurant (SemEval-2014) | 3608 | 1120 | - |
| Laptop (SemEval-2014) | 2328 | 638 | - |
| 6248 | 692 | - | |
| MAMS-ATSA | 11186 | 1336 | 1332 |
| Datasets | Distribution | Train | Test | Val |
| Lap14 | % of multiple sent./item | 7.86 | 7.84 | - |
| Avg. sent./item | 1.09 | 1.09 | - | |
| Res14 | % of multiple sent./item | 4.02 | 4.38 | - |
| Avg. number of sent./item | 1.04 | 1.05 | - | |
| % of multiple sent./item | 59.44 | 60.55 | - | |
| Avg. number of sent./item | 1.99 | 1.96 | - | |
| MAMS-ATSA | % of multiple sent./item | 2.12 | 0.82 | 0.60 |
| Avg. number of sent./item | 1.02 | 1.01 | 1.01 |
| Pooler types | Lap14 | Rest14 | ||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| SIR-ABSC | ||||||
| w/ max pooling | 83.2 | 80.0 | 88.7 | 83.3 | 74.8 | 73.7 |
| w/ avg pooling | 83.8 | 80.6 | 88.8 | 83.5 | 76.5 | 75.5 |
| w/ att pooling | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 |
| Sentiment: Neutral 0.01 1.84 0.39 0.08 0.05 0.04 0.40 X X 0.2 1.58 2.22 5.21 13.65 14.76 17.82 32.86 3.71 |
| Sentiment: Negative 0.22 6.14 0.39 5.56 1.50 5.07 2.44 X X 6.20 5.90 7.73 10.67 23.74 8.58 1.65 9.28 0.08 |
| Prediction: | Negative | 0.23 | 4.50 | 0.13 | 0.32 | 12.80 | 3.20 | 14.41 | 8.87 | 7.10 | 26.66 |
| Prediction: | Positive | 0.49 | 9.40 | 4.01 | 5.56 | 9.08 | 4.44 | 13.38 | 11.07 | 15.44 | 10.59 |
| Variable Distance Control (VDC) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| GoBERTA (Position Distance) | ||||||||
| • Decreasing-VDC | 83.4 | 80.4 | 88.5 | 83.2 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [222211110000] | 83.3 | 80.2 | 88.2 | 82.7 | 76.0 | 74.8 | 84.7 | 84.2 |
| • VDC = [333222111000] | 82.0 | 78.8 | 88.4 | 82.8 | 76.5 | 75.3 | 84.7 | 84.3 |
| • VDC = [444422220000] | 83.4 | 80.4 | 88.4 | 83.1 | 75.6 | 74.4 | 84.7 | 84.2 |
| • VDC = [554433221100] | 83.1 | 79.9 | 88.5 | 83.2 | 75.2 | 73.7 | 84.6 | 84.2 |
| • VDC = [666444222000] | 83.2 | 80.0 | 87.8 | 82.0 | 76.0 | 74.8 | 84.7 | 84.2 |
| • Constant-VDC | 83.7 | 80.7 | 88.6 | 83.3 | 76.4 | 75.4 | 84.9 | 84.4 |
| • Increasing-VDC | 83.7 | 80.5 | 88.6 | 83.2 | 76.9 | 76.0 | 84.8 | 84.4 |
| • VDC = [000011112222] | 83.5 | 80.3 | 87.8 | 82.1 | 76.9 | 76.0 | 84.7 | 84.3 |
| • VDC = [000111222333] | 83.5 | 80.5 | 88.5 | 83.2 | 75.5 | 74.4 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.6 | 80.4 | 87.9 | 82.2 | 75.7 | 74.4 | 84.8 | 84.4 |
| • VDC = [001122334455] | 83.3 | 80.3 | 88.6 | 83.1 | 76.1 | 74.9 | 84.8 | 84.4 |
| • VDC = [000222444666] | 83.7 | 80.5 | 88.3 | 82.5 | 76.6 | 75.8 | 84.7 | 84.2 |
| GoBERTA (Dependency Graph) | ||||||||
| • Decreasing-VDC | 83.7 | 80.6 | 88.2 | 82.8 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [222211110000] | 83.5 | 80.4 | 88.1 | 82.7 | 75.4 | 74.2 | 84.7 | 84.3 |
| • VDC = [333222111000] | 82.6 | 79.6 | 87.1 | 81.1 | 76.1 | 75.1 | 84.5 | 84.0 |
| • VDC = [444422220000] | 83.4 | 80.5 | 87.9 | 82.2 | 76.5 | 75.4 | 84.8 | 84.3 |
| • VDC = [554433221100] | 83.8 | 80.8 | 88.2 | 82.9 | 75.3 | 74.2 | 84.6 | 84.1 |
| • VDC = [666444222000] | 83.2 | 80.0 | 88.2 | 82.7 | 75.6 | 74.3 | 84.5 | 84.0 |
| • Constant-VDC | 83.7 | 80.4 | 88.9 | 83.7 | 76.4 | 75.2 | 85.0 | 84.5 |
| • Increasing-VDC | 83.8 | 80.8 | 89.1 | 83.8 | 77.1 | 75.9 | 85.0 | 84.6 |
| • VDC = [000011112222] | 84.1 | 81.1 | 88.3 | 82.8 | 77.2 | 76.3 | 84.6 | 84.2 |
| • VDC = [000111222333] | 83.8 | 80.8 | 89.1 | 83.8 | 76.7 | 75.5 | 84.7 | 84.3 |
| • VDC = [000022224444] | 83.5 | 80.5 | 88.9 | 83.5 | 75.7 | 74.6 | 84.8 | 84.3 |
| • VDC = [001122334455] | 83.2 | 80.2 | 89.3 | 84.1 | 74.7 | 75.9 | 85.0 | 84.6 |
| • VDC = [000222444666] | 82.5 | 79.4 | 88.9 | 83.8 | 76.9 | 75.9 | 84.8 | 84.5 |
| Ablations (Validation) | Lap14 | Rest14 | MAMS | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| 1. SIR-ABSC (RoBERTa) | ||||||||
| #1: RoBERTa baseline | 82.1 | 78.9 | 87.6 | 81.7 | 75.6 | 74.5 | 84.4 | 83.9 |
| #2: #1+[g] token+VDC | 84.1 | 81.1 | 89.3 | 84.1 | 77.2 | 76.3 | 85.0 | 84.6 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 84.2 | 81.3 | 89.3 | 84.3 | 77.2 | 76.0 | 84.8 | 84.3 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 84.4 | 81.5 | 89.4 | 84.3 | 77.3 | 76.5 | 84.9 | 84.4 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 84.6 | 81.8 | 89.4 | 84.4 | 77.2 | 76.3 | 85.2 | 84.7 |
| #6 (SIR-ABSC): #5+Auto-VDC | 85.0 | 82.1 | 89.7 | 84.8 | 77.5 | 76.4 | 85.4 | 84.9 |
| 2. SIR-ABSC (BERT) | ||||||||
| #1: BERT baseline | 79.2 | 75.1 | 85.8 | 79.3 | 76.4 | 74.8 | 84.0 | 82.7 |
| #2: #1+[g] token+VDC | 80.8 | 77.1 | 86.7 | 80.6 | 77.1 | 75.9 | 84.3 | 83.1 |
| #3 (DAA Ablation #1): #2+DAA (w/o edge label) | 81.1 | 77.5 | 86.9 | 80.9 | 77.2 | 75.9 | 84.3 | 83.2 |
| #4 (DAA Ablation #2): #2+DAA (w/o edge direction) | 81.6 | 78.2 | 87.0 | 81.1 | 77.4 | 76.1 | 84.4 | 83.5 |
| #5 (DAA Ablation #3): #2+DAA (proposed) | 81.8 | 78.6 | 87.0 | 81.2 | 77.6 | 76.4 | 84.5 | 83.5 |
| #6 (SIR-ABSC (BERT)): #5+Auto-VDC | 82.2 | 79.1 | 87.3 | 81.6 | 77.7 | 76.6 | 84.7 | 83.6 |
| SIR-ABSC (BERT) | Lap14 | Rest14 | MAMs | |||||
| Acc. | F1 | Acc. | F1 | Acc. | F1 | Acc. | F1 | |
| BERT baseline | 79.2(±0.6) | 75.1(±0.5) | 85.8(±0.7) | 79.3(±0.7) | 76.4(±0.7) | 74.8(±07) | 84.0(±0.3) | 82.7(±0.4) |
| SIR-ABSC (BERT) | 82.2(±0.4) | 79.1(±0.4) | 87.3(±0.7) | 81.6(±0.8) | 77.7(±0.8) | 76.6(±0.7) | 84.6(±0.4) | 83.3(±0.5) |
| Methods | SPIDER | ||
| EX | TS | ||
| Fine-tuning | T5-3B + PICARD | 79.3 | 69.4 |
| RASAT + PICARD | 80.5 | 70.3 | |
| RESDSQL-3B + NatSQL | 84.1 | 73.5 | |
| In-context learning | GPT-3 ada (0-shot) | 2.3 | 0.3 |
| GPT-3 babbage (0-shot) | 5.7 | 3.9 | |
| GPT-3 curie (0-shot) | 12.6 | 8.3 | |
| GPT-3 davinci (0-shot) | 26.3 | 21.7 | |
| Codex couchman (0-shot) | 63.7 | 53.0 | |
| Codex davinci (0-shot) | 67.0 | 55.1 | |
| ChatGPT (0-shot) | 70.1 | 60.1 | |
| SQLPrompt (0-shot) | 76.6 | 68.0 | |
| SQLPrompt (4-shot) | 77.1 | 68.6 |
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 67.7 | 61.3 | 70.8 | 62.9 | 70.5 | 63.2 (↑ 0.3) |
| PaLM FLAN 540B qntz | 72.3 | 64.1 | 71.6 | 61.3 | 74.0 | 65.5 (↑ 1.4) |
| Table 3: Ablation Study: Few-shots | ||||||
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 65.9 | 59.6 | 71.8 | 63.8 | 74.7 | 66.6 (↑ 2.8) |
| PaLM FLAN 540B qntz | 71.2 | 63.2 | 70.7 | 61.1 | 74.7 | 65.2 (↑ 2.0) |
| EX | TS | |
| SQLPrompt (Prompt Design + Consistency + Execution Filtering +MixPrompt) | 70.5 | 63.2 |
| No MixPrompt | 67.7 | 61.3 (↓ 1.9) |
| Only Schema (No primary, No foreignkeys, no DB content) | 66.4 | 57.3 (↓ 5.9) |
| No Consistency | 55.9 | 49.6 (↓ 13.6) |
| No Execution Filtering | 55.2 | 48.7 (↓ 14.5) |
| Num of Mixture | Zero-shots | Few-shots | ||
| 2 | 4 | 6 | 16 | |
| EX | 74 | 76.6 | 77.3 | 77.1 |
| TS | 65.5 | 68.0 | 68.3 | 68.6 |
| Methods/Datasets | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B q | 68.1 | 61.7 | 71.9 | 64.3 | 74.5 | 66.7 |
| Question | What are the names of the singers whose birth years are either 1948 or 1949? |
| Ground Truth | SELECT Name FROM singer WHERE Birth Year = 1948 OR Birth Year = 1949 |
| LLM Generation | SELECT name from singer WHERE birth_year = 1948 OR birth_year = 1949 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer with the largest net worth? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer who is worth the most? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenship of singers and the number of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, count(*) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, how many singers are from that country? |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT Citizenship, count(*) FROM singer GROUP BY Citizenship |
| Execution Accuracy | 1 |
| Question | Please show the most common citizenship of singers. |
| Ground Truth | SELECT Citizenship FROM singer GROUP BY Citizenship ORDER BY COUNT(*) DESC LIMIT 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the most common citizenship? |
| Ground Truth | select citizenship from singer group by citizenship order by count(*) desc limit 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenships and the maximum net worth of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, what is the maximum net worth? |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | Show titles of songs and names of singers. |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1.title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | What are the song titles and singer names? |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1 title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | Show distinct names of singers that have songs with sales more than 300000. |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT DISTINCT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | what are the different names of the singers that have sales more than 300000? |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | Show the names of singers that have more than one song. |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Question | What are the names of the singers that have more than one songs? |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Methods | SPIDER | ||
| EX | TS | ||
| Fine-tuning | T5-3B + PICARD | 79.3 | 69.4 |
| RASAT + PICARD | 80.5 | 70.3 | |
| RESDSQL-3B + NatSQL | 84.1 | 73.5 | |
| In-context learning | GPT-3 ada (0-shot) | 2.3 | 0.3 |
| GPT-3 babbage (0-shot) | 5.7 | 3.9 | |
| GPT-3 curie (0-shot) | 12.6 | 8.3 | |
| GPT-3 davinci (0-shot) | 26.3 | 21.7 | |
| Codex couchman (0-shot) | 63.7 | 53.0 | |
| Codex davinci (0-shot) | 67.0 | 55.1 | |
| ChatGPT (0-shot) | 70.1 | 60.1 | |
| SQLPrompt (0-shot) | 76.6 | 68.0 | |
| SQLPrompt (4-shot) | 77.1 | 68.6 |
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 67.7 | 61.3 | 70.8 | 62.9 | 70.5 | 63.2 (↑ 0.3) |
| PaLM FLAN 540B qntz | 72.3 | 64.1 | 71.6 | 61.3 | 74.0 | 65.5 (↑ 1.4) |
| Table 3: Ablation Study: Few-shots | ||||||
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 65.9 | 59.6 | 71.8 | 63.8 | 74.7 | 66.6 (↑ 2.8) |
| PaLM FLAN 540B qntz | 71.2 | 63.2 | 70.7 | 61.1 | 74.7 | 65.2 (↑ 2.0) |
| EX | TS | |
| SQLPrompt (Prompt Design + Consistency + Execution Filtering +MixPrompt) | 70.5 | 63.2 |
| No MixPrompt | 67.7 | 61.3 (↓ 1.9) |
| Only Schema (No primary, No foreignkeys, no DB content) | 66.4 | 57.3 (↓ 5.9) |
| No Consistency | 55.9 | 49.6 (↓ 13.6) |
| No Execution Filtering | 55.2 | 48.7 (↓ 14.5) |
| Num of Mixture | Zero-shots | Few-shots | ||
| 2 | 4 | 6 | 16 | |
| EX | 74 | 76.6 | 77.3 | 77.1 |
| TS | 65.5 | 68.0 | 68.3 | 68.6 |
| Methods/Datasets | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B q | 68.1 | 61.7 | 71.9 | 64.3 | 74.5 | 66.7 |
| Question | What are the names of the singers whose birth years are either 1948 or 1949? |
| Ground Truth | SELECT Name FROM singer WHERE Birth Year = 1948 OR Birth Year = 1949 |
| LLM Generation | SELECT name from singer WHERE birth_year = 1948 OR birth_year = 1949 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer with the largest net worth? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer who is worth the most? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenship of singers and the number of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, count(*) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, how many singers are from that country? |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT Citizenship, count(*) FROM singer GROUP BY Citizenship |
| Execution Accuracy | 1 |
| Question | Please show the most common citizenship of singers. |
| Ground Truth | SELECT Citizenship FROM singer GROUP BY Citizenship ORDER BY COUNT(*) DESC LIMIT 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the most common citizenship? |
| Ground Truth | select citizenship from singer group by citizenship order by count(*) desc limit 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenships and the maximum net worth of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, what is the maximum net worth? |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | Show titles of songs and names of singers. |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1.title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | What are the song titles and singer names? |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1 title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | Show distinct names of singers that have songs with sales more than 300000. |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT DISTINCT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | what are the different names of the singers that have sales more than 300000? |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | Show the names of singers that have more than one song. |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Question | What are the names of the singers that have more than one songs? |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Methods | SPIDER | ||
| EX | TS | ||
| Fine-tuning | T5-3B + PICARD | 79.3 | 69.4 |
| RASAT + PICARD | 80.5 | 70.3 | |
| RESDSQL-3B + NatSQL | 84.1 | 73.5 | |
| In-context learning | GPT-3 ada (0-shot) | 2.3 | 0.3 |
| GPT-3 babbage (0-shot) | 5.7 | 3.9 | |
| GPT-3 curie (0-shot) | 12.6 | 8.3 | |
| GPT-3 davinci (0-shot) | 26.3 | 21.7 | |
| Codex couchman (0-shot) | 63.7 | 53.0 | |
| Codex davinci (0-shot) | 67.0 | 55.1 | |
| ChatGPT (0-shot) | 70.1 | 60.1 | |
| SQLPrompt (0-shot) | 76.6 | 68.0 | |
| SQLPrompt (4-shot) | 77.1 | 68.6 | |
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 67.7 | 61.3 | 70.8 | 62.9 | 70.5 | 63.2 (↑ 0.3) |
| PaLM FLAN 540B qntz | 72.3 | 64.1 | 71.6 | 61.3 | 74.0 | 65.5 (↑ 1.4) |
| Table 3: Ablation Study: Few-shots | ||||||
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 65.9 | 59.6 | 71.8 | 63.8 | 74.7 | 66.6 (↑ 2.8) |
| PaLM FLAN 540B qntz | 71.2 | 63.2 | 70.7 | 61.1 | 74.7 | 65.2 (↑ 2.0) |
| EX | TS | |
| SQLPrompt (Prompt Design + Consistency + Execution Filtering +MixPrompt) | 70.5 | 63.2 |
| No MixPrompt | 67.7 | 61.3 (↓ 1.9) |
| Only Schema (No primary, No foreignkeys, no DB content) | 66.4 | 57.3 (↓ 5.9) |
| No Consistency | 55.9 | 49.6 (↓ 13.6) |
| No Execution Filtering | 55.2 | 48.7 (↓ 14.5) |
| Num of Mixture | Zero-shots | Few-shots | ||
| 2 | 4 | 6 | 16 | |
| EX | 74 | 76.6 | 77.3 | 77.1 |
| TS | 65.5 | 68.0 | 68.3 | 68.6 |
| Methods/Datasets | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B q | 68.1 | 61.7 | 71.9 | 64.3 | 74.5 | 66.7 |
| Question | What are the names of the singers whose birth years are either 1948 or 1949? |
| Ground Truth | SELECT Name FROM singer WHERE Birth Year = 1948 OR Birth Year = 1949 |
| LLM Generation | SELECT name from singer WHERE birth_year = 1948 OR birth_year = 1949 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer with the largest net worth? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer who is worth the most? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenship of singers and the number of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, count(*) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, how many singers are from that country? |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT Citizenship, count(*) FROM singer GROUP BY Citizenship |
| Execution Accuracy | 1 |
| Question | Please show the most common citizenship of singers. |
| Ground Truth | SELECT Citizenship FROM singer GROUP BY Citizenship ORDER BY COUNT(*) DESC LIMIT 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the most common citizenship? |
| Ground Truth | select citizenship from singer group by citizenship order by count(*) desc limit 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenships and the maximum net worth of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, what is the maximum net worth? |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | Show titles of songs and names of singers. |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1.title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | What are the song titles and singer names? |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1 title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | Show distinct names of singers that have songs with sales more than 300000. |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT DISTINCT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | what are the different names of the singers that have sales more than 300000? |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | Show the names of singers that have more than one song. |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Question | What are the names of the singers that have more than one songs? |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Methods | SPIDER | ||
| EX | TS | ||
| Fine-tuning | T5-3B + PICARD | 79.3 | 69.4 |
| RASAT + PICARD | 80.5 | 70.3 | |
| RESDSQL-3B + NatSQL | 84.1 | 73.5 | |
| In-context learning | GPT-3 ada (0-shot) | 2.3 | 0.3 |
| GPT-3 babbage (0-shot) | 5.7 | 3.9 | |
| GPT-3 curie (0-shot) | 12.6 | 8.3 | |
| GPT-3 davinci (0-shot) | 26.3 | 21.7 | |
| Codex couchman (0-shot) | 63.7 | 53.0 | |
| Codex davinci (0-shot) | 67.0 | 55.1 | |
| ChatGPT (0-shot) | 70.1 | 60.1 | |
| SQLPrompt (0-shot) | 76.6 | 68.0 | |
| SQLPrompt (4-shot) | 77.1 | 68.6 |
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 67.7 | 61.3 | 70.8 | 62.9 | 70.5 | 63.2 (↑ 0.3) |
| PaLM FLAN 540B qntz | 72.3 | 64.1 | 71.6 | 61.3 | 74.0 | 65.5 (↑ 1.4) |
| Table 3: Ablation Study: Few-shots | ||||||
| Models | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B qntz | 65.9 | 59.6 | 71.8 | 63.8 | 74.7 | 66.6 (↑ 2.8) |
| PaLM FLAN 540B qntz | 71.2 | 63.2 | 70.7 | 61.1 | 74.7 | 65.2 (↑ 2.0) |
| EX | TS | |
| SQLPrompt (Prompt Design + Consistency + Execution Filtering +MixPrompt) | 70.5 | 63.2 |
| No MixPrompt | 67.7 | 61.3 (↓ 1.9) |
| Only Schema (No primary, No foreignkeys, no DB content) | 66.4 | 57.3 (↓ 5.9) |
| No Consistency | 55.9 | 49.6 (↓ 13.6) |
| No Execution Filtering | 55.2 | 48.7 (↓ 14.5) |
| Num of Mixture | Zero-shots | Few-shots | ||
| 2 | 4 | 6 | 16 | |
| EX | 74 | 76.6 | 77.3 | 77.1 |
| TS | 65.5 | 68.0 | 68.3 | 68.6 |
| Methods/Datasets | Concise | Verbose | MixPrompt | |||
| EX | TS | EX | TS | EX | TS | |
| PaLM FLAN 62B q | 68.1 | 61.7 | 71.9 | 64.3 | 74.5 | 66.7 |
| Question | What are the names of the singers whose birth years are either 1948 or 1949? |
| Ground Truth | SELECT Name FROM singer WHERE Birth Year = 1948 OR Birth Year = 1949 |
| LLM Generation | SELECT name from singer WHERE birth_year = 1948 OR birth_year = 1949 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer with the largest net worth? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the name of the singer who is worth the most? |
| Ground Truth | SELECT Name FROM singer ORDER BY Net_Worth_Millions DESC LIMIT 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id ORDER BY T1.net_worth_millions Desc LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenship of singers and the number of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, count(*) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, how many singers are from that country? |
| Ground Truth | SELECT Citizenship, COUNT(*) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT Citizenship, count(*) FROM singer GROUP BY Citizenship |
| Execution Accuracy | 1 |
| Question | Please show the most common citizenship of singers. |
| Ground Truth | SELECT Citizenship FROM singer GROUP BY Citizenship ORDER BY COUNT(*) DESC LIMIT 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | What is the most common citizenship? |
| Ground Truth | select citizenship from singer group by citizenship order by count(*) desc limit 1 |
| LLM Generation | SELECT citizenship FROM singer GROUP BY citizenship ORDER BY count(*) DESC LIMIT 1 |
| Execution Accuracy | 1 |
| Question | Show different citizenships and the maximum net worth of singers of each citizenship. |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | For each citizenship, what is the maximum net worth? |
| Ground Truth | SELECT Citizenship, max(Net_Worth_Millions) FROM singer GROUP BY Citizenship |
| LLM Generation | SELECT T1.citizenship, max(T1.net_worth_millions) FROM singer AS T1 GROUP BY T1.citizenship |
| Execution Accuracy | 1 |
| Question | Show titles of songs and names of singers. |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1.title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | What are the song titles and singer names? |
| Ground Truth | SELECT T2.Title, T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID |
| LLM Generation | SELECT T1 title, T2.name FROM song AS T1 JOIN singer AS T2 ON T1.singer_id = T2.singer_id |
| Execution Accuracy | 1 |
| Question | Show distinct names of singers that have songs with sales more than 300000. |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT DISTINCT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | what are the different names of the singers that have sales more than 300000? |
| Ground Truth | SELECT DISTINCT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID WHERE T2.Sales > 300000 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id WHERE T2.sales > 300000 |
| Execution Accuracy | 1 |
| Question | Show the names of singers that have more than one song. |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Question | What are the names of the singers that have more than one songs? |
| Ground Truth | SELECT T1.Name FROM singer AS T1 JOIN song AS T2 ON T1.Singer_ID = T2.Singer_ID GROUP BY T1.Name HAVING COUNT(*) > 1 |
| LLM Generation | SELECT T1.name FROM singer AS T1 JOIN song AS T2 ON T1.singer_id = T2.singer_id GROUP BY T1.singer_id HAVING COUNT(*) > 1 |
| Execution Accuracy | 1 |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 42.6 | 7.4 | 3.8 | 23.2 | 11.2 | 25.4 | 22.7 |
| Bible | 44.0 | 26.9 | 6.6 | 5.2 | 16.0 | 20.2 | 21.0 |
| 1810-1820s | 30.2 | 11.1 | 3.5 | 14.7 | 15.9 | 17.4 | 17.0 |
| 1890-1900s | 35.9 | 12.3 | 4.4 | 8.6 | 9.1 | 10.4 | 10.1 |
| 1990-2000s | 42.3 | 16.6 | 4.3 | 7.9 | 13.0 | 17.5 | 17.2 |
| English Twitter | 41.2 | 8.0 | 5.5 | 35.0 | 23.6 | 32.0 | 29.5 |
| James Joyce | 20.4 | 11.8 | 5.4 | 3.4 | 1.3 | 1.6 | 2.6 |
| Song Lyrics | 33.3 | 20.2 | 7.7 | 12.2 | 15.4 | 11.2 | 13.2 |
| Romantic Poetry | 20.4 | 15.7 | 2.8 | 1.1 | 3.4 | 6.2 | 4.9 |
| Shakespeare | 13.6 | 9.1 | 2.5 | 9.6 | 10.0 | 9.7 | 9.7 |
| Switchboard | 52.9 | 21.1 | 1.7 | 0.1 | 0.3 | 5.3 | 13.7 |
| Overall | 34.3 | 14.6 | 4.4 | 11.0 | 10.8 | 14.3 | 14.7 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 44.0 | 47.7 | 18.7 | 13.2 | 25.6 | 10.6 | 31.7 | 29.2 | 21.5 | 17.9 | 30 | 28.8 | 30.2 | 27.6 |
| Bible | 36.1 | 38.8 | 22 | 22.9 | 0.3 | 1.6 | 4.3 | 4.4 | 15.7 | 15.9 | 18.0 | 19.0 | 19.8 | 19.5 |
| 1810-1820s | 26.3 | 29.5 | 5.9 | 10.0 | 1.2 | 4.7 | 12.4 | 15.6 | 14.3 | 16.9 | 17.6 | 21.6 | 16.9 | 20.1 |
| 1890-1900s | 33.5 | 34.7 | 10.0 | 13.4 | 4.4 | 11.0 | 9.9 | 11.8 | 13.9 | 13.8 | 14.6 | 14.4 | 13.8 | 13.3 |
| 1990-200s | 50.2 | 56.2 | 22.6 | 32.1 | 11.8 | 31.4 | 16.7 | 20.7 | 28.5 | 32.5 | 31.5 | 34.7 | 28.4 | 32.8 |
| English Twitter | 46.1 | 54.1 | 20.1 | 22.1 | 32.4 | 33.5 | 37.4 | 41.8 | 30.1 | 29.5 | 34.9 | 36.4 | 32.5 | 35.0 |
| James Joyce | 22.3 | 22.8 | 10.9 | 13.2 | 3.2 | 7.9 | 2.9 | 3.3 | 2.7 | 2.3 | 3.1 | 2.5 | 3.3 | 2.8 |
| Song Lyrics | 42.6 | 40.5 | 22.1 | 23.2 | 10.3 | 12.4 | 19.3 | 12.9 | 22.3 | 18.4 | 19.3 | 16.2 | 24.2 | 20.1 |
| Romantic Poetry | 13.5 | 12.9 | 8.9 | 10.8 | 0.8 | 0.9 | 2.0 | 1.1 | 5.2 | 4.3 | 7.0 | 4.7 | 6.0 | 3.9 |
| Shakespeare | 11.8 | 11.6 | 11.1 | 10.4 | 1.3 | 4.1 | 12.9 | 15.1 | 15.3 | 14.7 | 13.4 | 15.2 | 13.8 | 15.2 |
| Switchboard | 54.6 | 59.3 | 29.7 | 35.1 | 5.2 | 6.1 | 0.1 | 0.1 | 0.3 | 0.1 | 9.7 | 13.4 | 15.6 | 23.0 |
| Overall | 34.6 | 37.1 | 16.5 | 18.8 | 8.8 | 11.3 | 13.6 | 14.2 | 15.4 | 15.1 | 18.1 | 18.8 | 18.6 | 19.4 |
| Input: Can’t sleep at all. Smh. Transfer: AAE → 1990s-2000s | |
| STEER | I mean, I can’t sleep at all. |
| GPT-3 | I am unable to obtain any rest; shaking my head in disbelief. |
| STRAP | I don’t want to sleep. |
| P-A-R | Can’t sleep at all, smh. |
| Input: Yeah one one way or another Transfer: switchboard → rom. poetry | |
| STEER | One way, or another, or both |
| GPT-3 | Aye, one mayhap will find the way. |
| STRAP | And one way or the other |
| P-A-R | Yeah one one way or another |
| Input: In his fear, he dare not face me Transfer: lyrics → bible | |
| STEER | And he will not dare to face me: for fear of me is in his eyes. |
| GPT-3 | And his fear was great, so that he could not stand before me. |
| STRAP | For he that is afraid of me is of me; but he that is of me is of him. |
| P-A-R | In fear he came and hid himself, because God was near to him |
| Reward Type | V | TSS | F | MS |
| Coarse | 19.9 | 42.8 | 79.0 | 62.6 |
| Fine-grained | 27.4 | 47.0 | 83.6 | 71.5 |
| MSTTR | n-gram entropy, n = | F | MS | |||
| 1 | 2 | 3 | ||||
| Df(STEER) | 0.912 | 8.8 | 15.0 | 19.3 | 86.8 | 67.2 |
| GYAFC | 0.931 | 9.1 | 15.0 | 17.8 | 86.2 | 78.4 |
| Yelp | 0.958 | 9.6 | 16.4 | 21.3 | 87.2 | - |
| CDS | 0.946 | 10.0 | 17.0 | 20.4 | 83.2 | - |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 1 |
| effective batch size | 32 |
| total steps | 50,000 |
| steps per evaluation | 1000 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| Hyperparameter | Tested |
| α | [0.2,0.4,0.6] |
| temperature | [0.7, 1.0, 1.3] |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 4 |
| effective batch size | 1024 |
| total steps | 30,000 |
| steps per evaluation | 2500 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| QUARK hyps | |
| sample steps | 2500 |
| KL penalty | 0.025 |
| entropy coef | 0.00 |
| policy temp. | 1.0 |
| vectorized reward token | True |
| number of buckets | 5 |
| Style | Train | Validation | Test |
| AAE | 720K | 7.3K | 7.3K |
| Bible | 31K | 1.7K | 1.7K |
| Coha 1810 | 205K | 5.3K | 5.3K |
| Coha 1890 | 1.2M | 32K | 32K |
| Coha 1990 | 1.9M | 49K | 49K |
| English Tweet | 5.2M | 40K | 40K |
| Joyce | 37K | 2.0K | 2.0K |
| Lyrics | 4.5M | 252K | 252K |
| Romantic Poetry | 27K | 1.5K | 1.5K |
| Shakespeare | 25K | 1.3K | 1.3K |
| Switchboard | 146K | 1.5K | 1.5K |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 38.2 | 7.8 | 1.6 | 18.1 | 9.4 | 20.7 | 17.6 |
| Bible | 34.4 | 22 | 2.0 | 4.0 | 12.3 | 14.4 | 13.6 |
| 1810-1820s | 23.1 | 9.2 | 1.0 | 12.8 | 11.5 | 12.3 | 12.6 |
| 1890-1900s | 34.3 | 13.6 | 2.2 | 8.0 | 8.0 | 10.3 | 10.3 |
| 1990-200s | 39.0 | 16.1 | 1.7 | 8.7 | 13.1 | 15.6 | 15.8 |
| English Twitter | 37.7 | 8.7 | 2.5 | 28.2 | 16.6 | 24.2 | 19.8 |
| James Joyce | 17.6 | 10.7 | 1.9 | 1.9 | 1.5 | 1.5 | 2.2 |
| Song Lyrics | 29.3 | 20.1 | 2.9 | 10.3 | 13.6 | 9.4 | 11.7 |
| Romantic Poetry | 16.8 | 14.8 | 0.9 | 0.5 | 2.6 | 4.2 | 3.9 |
| Shakespeare | 11.4 | 8.0 | 0.8 | 5.9 | 7.3 | 6.6 | 6.9 |
| Switchboard | 41.4 | 15.2 | 0.5 | 0.0 | 0.2 | 3.0 | 6.9 |
| Overall | 29.4 | 13.3 | 1.6 | 8.9 | 8.7 | 11.1 | 11.0 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 38.7 | 43.8 | 18.7 | 14.1 | 14.5 | 5.7 | 25.7 | 23.5 | 16.7 | 15 | 25.5 | 24.8 | 23.8 | 22.9 |
| Bible | 29.6 | 32.4 | 19.4 | 20.1 | 0.3 | 0.5 | 3.9 | 3.9 | 10.8 | 11.3 | 13.6 | 15.5 | 15.3 | 15.4 |
| 1810-1820s | 19.6 | 22.3 | 5.2 | 8.4 | 0.2 | 1.5 | 9.9 | 12.6 | 11.2 | 14.1 | 12.6 | 16.7 | 12.1 | 15.7 |
| 1890-1900s | 29.2 | 32.3 | 10.9 | 14.2 | 2.0 | 5.3 | 8.5 | 10.2 | 11.3 | 12.6 | 12.4 | 12.8 | 11.4 | 11.2 |
| 1990-200s | 44.9 | 51.9 | 22.1 | 31.6 | 6.7 | 17.9 | 15.3 | 18.9 | 26.3 | 30.6 | 30.1 | 32.6 | 26.8 | 31.0 |
| English Twitter | 44.4 | 51.2 | 20.6 | 22.1 | 23.6 | 22.5 | 28.0 | 32.3 | 20.3 | 19.8 | 24.7 | 26.0 | 22.8 | 24.5 |
| James Joyce | 19.0 | 21.5 | 11.7 | 13.2 | 1.5 | 3.5 | 2.0 | 2.6 | 2.9 | 3.1 | 2.9 | 2.3 | 3.0 | 3.1 |
| Song Lyrics | 36.4 | 36.1 | 21.3 | 23.6 | 6.0 | 4.3 | 15.3 | 10.0 | 19.0 | 14.6 | 15.7 | 13.3 | 19.2 | 16.1 |
| Romantic Poetry | 11.2 | 11.4 | 10.0 | 11.2 | 0.2 | 0.3 | 1.4 | 0.7 | 3.7 | 2.9 | 4.7 | 3.5 | 4.2 | 2.8 |
| Shakespeare | 10.1 | 9.7 | 9.9 | 9.4 | 0.3 | 1.6 | 10.2 | 12.4 | 11.3 | 12.5 | 11.1 | 12.5 | 11.5 | 12.8 |
| Switchboard | 42.5 | 47.3 | 22.6 | 27.7 | 2.1 | 1.7 | 0.1 | 0.0 | 0.3 | 0.0 | 4.8 | 6.6 | 8.1 | 12.4 |
| Overall | 29.6 | 32.7 | 15.7 | 17.9 | 5.2 | 5.9 | 10.9 | 11.6 | 12.2 | 12.4 | 14.4 | 15.1 | 14.4 | 15.3 |
| V | TSS | F | MS | |
| AAE Twitter | 42.6 | 68.2 | 87.0 | 72.5 |
| Bible | 44.0 | 80.9 | 85.9 | 63.9 |
| 1810-1820s | 30.2 | 49.8 | 81.9 | 74.7 |
| 1890-1900s | 35.9 | 56.6 | 85.4 | 74.6 |
| 1990-200s | 42.3 | 63.0 | 83.9 | 77.8 |
| English Twitter | 41.2 | 62.6 | 87.7 | 74.5 |
| James Joyce | 20.4 | 34.1 | 81.8 | 76.7 |
| Song Lyrics | 33.3 | 51.6 | 87.2 | 74.7 |
| Romantic Poetry | 20.4 | 35.3 | 80.5 | 74.7 |
| Shakespeare | 13.6 | 24.7 | 83.7 | 72.7 |
| Switchboard | 52.9 | 92.0 | 87.0 | 66.1 |
| Overall | 34.3 | 56.3 | 84.7 | 73.0 |
| v | TSS | F | MS | |
| AAE Twitter | 7.4 | 21.8 | 65.7 | 66.2 |
| Bible | 26.9 | 70.5 | 79.4 | 51.7 |
| 1810-1820s | 11.1 | 22.5 | 78.6 | 64.9 |
| 1890-1900s | 12.3 | 22.6 | 82.6 | 65.0 |
| 1990-2000s | 16.6 | 29.1 | 82.8 | 65.0 |
| English Twitter | 8.0 | 36.9 | 76.8 | 49.8 |
| James Joyce | 11.8 | 26.4 | 78.8 | 66.7 |
| Song Lyrics | 20.2 | 39.4 | 80.2 | 67.3 |
| Romantic Poetry | 15.7 | 46.5 | 64.5 | 62.1 |
| Shakespeare | 9.1 | 27.8 | 68.0 | 62.2 |
| Switchboard | 21.1 | 52.0 | 69.3 | 65.2 |
| Overall | 14.6 | 36.0 | 75.2 | 62.4 |
| V | TSS | F | MS | |
| AAE Twitter | 3.8 | 10.3 | 76.1 | 68.5 |
| Bible | 6.6 | 24 | 79.1 | 67.3 |
| 1810-1820s | 3.5 | 8.7 | 74.4 | 78.9 |
| 1890-1900s | 4.4 | 10 | 76 | 77.7 |
| 1990-2000s | 4.3 | 9.1 | 75.4 | 79.2 |
| English Twitter | 5.5 | 12.5 | 74.1 | 82.1 |
| James Joyce | 5.4 | 13.6 | 76.9 | 75.3 |
| Song Lyrics | 7.7 | 23.2 | 78.2 | 65.3 |
| Romantic Poetry | 2.8 | 7.6 | 77.1 | 76.1 |
| Shakespeare | 2.5 | 8.2 | 77 | 72 |
| Switchboard | 1.7 | 3.9 | 75.1 | 84.5 |
| Overall | 4.4 | 11.9 | 76.3 | 75.2 |
| V | TSS | F | MS | |
| AAE Twitter | 23.2 | 58.4 | 63.1 | 65.7 |
| Bible | 5.2 | 10.4 | 85 | 70.4 |
| 1810-1820s | 14.7 | 27.6 | 79.4 | 67 |
| 1890-1900s | 8.6 | 15 | 83.1 | 65 |
| 1990-2000s | 7.9 | 12.6 | 83.8 | 68.8 |
| English Twitter | 35 | 58.3 | 82 | 70 |
| James Joyce | 3.4 | 7.4 | 77.1 | 66 |
| Song Lyrics | 12.2 | 25.4 | 73.7 | 72.2 |
| Romantic Poetry | 1.1 | 3.5 | 61.7 | 61.5 |
| Shakespeare | 9.6 | 27.7 | 64.3 | 68.6 |
| Switchboard | 0.1 | 0.1 | 92.2 | 74.5 |
| Overall | 11 | 22.4 | 76.9 | 68.2 |
| V | TSS | F | MS | |
| AAE Twitter | 11.2 | 30.9 | 67.3 | 64.8 |
| Bible | 16 | 40.2 | 79 | 59.5 |
| 1810-1820s | 15.9 | 29.4 | 82 | 65.3 |
| 1890-1900s | 9.1 | 16 | 82.4 | 63.3 |
| 1990-2000s | 13 | 25.2 | 86.2 | 65.3 |
| English Twitter | 23.6 | 47 | 71.3 | 68 |
| James Joyce | 1.3 | 11 | 46.1 | 63.2 |
| SongLyrics | 15.4 | 40.1 | 71.8 | 62.4 |
| Romantic Poetry | 3.4 | 15.3 | 51.7 | 59.1 |
| Shakespeare | 10 | 29.5 | 61.9 | 69 |
| Switchboard | 0.3 | 0.6 | 90.3 | 75.4 |
| Overall | 10.8 | 25.9 | 71.8 | 65 |
| V | TSS | F | MS | |
| AAE Twitter | 25.4 | 79.1 | 58.4 | 59 |
| Bible | 20.2 | 43.6 | 78.8 | 65.6 |
| 1810-1820s | 17.4 | 35.6 | 80.4 | 62.4 |
| 1890-1900s | 10.4 | 17.2 | 84.6 | 66.8 |
| 1990-2000s | 17.5 | 26.9 | 89.1 | 71 |
| English Twitter | 32 | 55.9 | 80.2 | 68.4 |
| James Joyce | 1.6 | 10.9 | 48.6 | 63.2 |
| Song Lyrics | 11.2 | 25.8 | 74.7 | 66 |
| Romantic Poetry | 6.2 | 28.5 | 44.6 | 57.4 |
| Shakespeare | 9.7 | 28 | 62.8 | 68.4 |
| Switchboard | 5.3 | 10.7 | 85.4 | 71.2 |
| Overall | 14.3 | 32.9 | 71.6 | 65.4 |
| V | TSS | F | MS | |
| AAE Twitter | 22.7 | 65.4 | 57.9 | 64.9 |
| Bible | 21 | 43.8 | 81.4 | 65.7 |
| 1810-1820s | 17 | 33.1 | 81 | 66.2 |
| 1890-1900s | 10.1 | 16.9 | 83.1 | 68.8 |
| 1990-2000s | 17.2 | 25.7 | 88.2 | 72.1 |
| English Twitter | 29.5 | 51.4 | 77.8 | 69.9 |
| James Joyce | 2.6 | 20.9 | 33.9 | 62.4 |
| Song Lyrics | 13.2 | 30.8 | 69.5 | 69.5 |
| Romantic Poetry | 4.9 | 23.2 | 43.9 | 58.9 |
| Shakespeare | 9.7 | 28.7 | 60.8 | 69.5 |
| Switchboard | 13.7 | 24.8 | 82.2 | 73.5 |
| Overall | 14.7 | 33.2 | 69.1 | 67.4 |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 42.6 | 7.4 | 3.8 | 23.2 | 11.2 | 25.4 | 22.7 |
| Bible | 44.0 | 26.9 | 6.6 | 5.2 | 16.0 | 20.2 | 21.0 |
| 1810-1820s | 30.2 | 11.1 | 3.5 | 14.7 | 15.9 | 17.4 | 17.0 |
| 1890-1900s | 35.9 | 12.3 | 4.4 | 8.6 | 9.1 | 10.4 | 10.1 |
| 1990-2000s | 42.3 | 16.6 | 4.3 | 7.9 | 13.0 | 17.5 | 17.2 |
| English Twitter | 41.2 | 8.0 | 5.5 | 35.0 | 23.6 | 32.0 | 29.5 |
| James Joyce | 20.4 | 11.8 | 5.4 | 3.4 | 1.3 | 1.6 | 2.6 |
| Song Lyrics | 33.3 | 20.2 | 7.7 | 12.2 | 15.4 | 11.2 | 13.2 |
| Romantic Poetry | 20.4 | 15.7 | 2.8 | 1.1 | 3.4 | 6.2 | 4.9 |
| Shakespeare | 13.6 | 9.1 | 2.5 | 9.6 | 10.0 | 9.7 | 9.7 |
| Switchboard | 52.9 | 21.1 | 1.7 | 0.1 | 0.3 | 5.3 | 13.7 |
| Overall | 34.3 | 14.6 | 4.4 | 11.0 | 10.8 | 14.3 | 14.7 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 44.0 | 47.7 | 18.7 | 13.2 | 25.6 | 10.6 | 31.7 | 29.2 | 21.5 | 17.9 | 30 | 28.8 | 30.2 | 27.6 |
| Bible | 36.1 | 38.8 | 22 | 22.9 | 0.3 | 1.6 | 4.3 | 4.4 | 15.7 | 15.9 | 18.0 | 19.0 | 19.8 | 19.5 |
| 1810-1820s | 26.3 | 29.5 | 5.9 | 10.0 | 1.2 | 4.7 | 12.4 | 15.6 | 14.3 | 16.9 | 17.6 | 21.6 | 16.9 | 20.1 |
| 1890-1900s | 33.5 | 34.7 | 10.0 | 13.4 | 4.4 | 11.0 | 9.9 | 11.8 | 13.9 | 13.8 | 14.6 | 14.4 | 13.8 | 13.3 |
| 1990-200s | 50.2 | 56.2 | 22.6 | 32.1 | 11.8 | 31.4 | 16.7 | 20.7 | 28.5 | 32.5 | 31.5 | 34.7 | 28.4 | 32.8 |
| English Twitter | 46.1 | 54.1 | 20.1 | 22.1 | 32.4 | 33.5 | 37.4 | 41.8 | 30.1 | 29.5 | 34.9 | 36.4 | 32.5 | 35.0 |
| James Joyce | 22.3 | 22.8 | 10.9 | 13.2 | 3.2 | 7.9 | 2.9 | 3.3 | 2.7 | 2.3 | 3.1 | 2.5 | 3.3 | 2.8 |
| Song Lyrics | 42.6 | 40.5 | 22.1 | 23.2 | 10.3 | 12.4 | 19.3 | 12.9 | 22.3 | 18.4 | 19.3 | 16.2 | 24.2 | 20.1 |
| Romantic Poetry | 13.5 | 12.9 | 8.9 | 10.8 | 0.8 | 0.9 | 2.0 | 1.1 | 5.2 | 4.3 | 7.0 | 4.7 | 6.0 | 3.9 |
| Shakespeare | 11.8 | 11.6 | 11.1 | 10.4 | 1.3 | 4.1 | 12.9 | 15.1 | 15.3 | 14.7 | 13.4 | 15.2 | 13.8 | 15.2 |
| Switchboard | 54.6 | 59.3 | 29.7 | 35.1 | 5.2 | 6.1 | 0.1 | 0.1 | 0.3 | 0.1 | 9.7 | 13.4 | 15.6 | 23.0 |
| Overall | 34.6 | 37.1 | 16.5 | 18.8 | 8.8 | 11.3 | 13.6 | 14.2 | 15.4 | 15.1 | 18.1 | 18.8 | 18.6 | 19.4 |
| Input: Can’t sleep at all. Smh. Transfer: AAE → 1990s-2000s | |
| STEER | I mean, I can’t sleep at all. |
| GPT-3 | I am unable to obtain any rest; shaking my head in disbelief. |
| STRAP | I don’t want to sleep. |
| P-A-R | Can’t sleep at all, smh. |
| Input: Yeah one one way or another Transfer: switchboard → rom. poetry | |
| STEER | One way, or another, or both |
| GPT-3 | Aye, one mayhap will find the way. |
| STRAP | And one way or the other |
| P-A-R | Yeah one one way or another |
| Input: In his fear, he dare not face me Transfer: lyrics → bible | |
| STEER | And he will not dare to face me: for fear of me is in his eyes. |
| GPT-3 | And his fear was great, so that he could not stand before me. |
| STRAP | For he that is afraid of me is of me; but he that is of me is of him. |
| P-A-R | In fear he came and hid himself, because God was near to him |
| Reward Type | V | TSS | F | MS |
| Coarse | 19.9 | 42.8 | 79.0 | 62.6 |
| Fine-grained | 27.4 | 47.0 | 83.6 | 71.5 |
| MSTTR | n-gram entropy, n = | F | MS | |||
| 1 | 2 | 3 | ||||
| Df(STEER) | 0.912 | 8.8 | 15.0 | 19.3 | 86.8 | 67.2 |
| GYAFC | 0.931 | 9.1 | 15.0 | 17.8 | 86.2 | 78.4 |
| Yelp | 0.958 | 9.6 | 16.4 | 21.3 | 87.2 | - |
| CDS | 0.946 | 10.0 | 17.0 | 20.4 | 83.2 | - |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 1 |
| effective batch size | 32 |
| total steps | 50,000 |
| steps per evaluation | 1000 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| Hyperparameter | Tested |
| α | [0.2,0.4,0.6] |
| temperature | [0.7, 1.0, 1.3] |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 4 |
| effective batch size | 1024 |
| total steps | 30,000 |
| steps per evaluation | 2500 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| QUARK hyps | |
| sample steps | 2500 |
| KL penalty | 0.025 |
| entropy coef | 0.00 |
| policy temp. | 1.0 |
| vectorized reward token | True |
| number of buckets | 5 |
| Style | Train | Validation | Test |
| AAE | 720K | 7.3K | 7.3K |
| Bible | 31K | 1.7K | 1.7K |
| Coha 1810 | 205K | 5.3K | 5.3K |
| Coha 1890 | 1.2M | 32K | 32K |
| Coha 1990 | 1.9M | 49K | 49K |
| English Tweet | 5.2M | 40K | 40K |
| Joyce | 37K | 2.0K | 2.0K |
| Lyrics | 4.5M | 252K | 252K |
| Romantic Poetry | 27K | 1.5K | 1.5K |
| Shakespeare | 25K | 1.3K | 1.3K |
| Switchboard | 146K | 1.5K | 1.5K |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 38.2 | 7.8 | 1.6 | 18.1 | 9.4 | 20.7 | 17.6 |
| Bible | 34.4 | 22 | 2.0 | 4.0 | 12.3 | 14.4 | 13.6 |
| 1810-1820s | 23.1 | 9.2 | 1.0 | 12.8 | 11.5 | 12.3 | 12.6 |
| 1890-1900s | 34.3 | 13.6 | 2.2 | 8.0 | 8.0 | 10.3 | 10.3 |
| 1990-200s | 39.0 | 16.1 | 1.7 | 8.7 | 13.1 | 15.6 | 15.8 |
| English Twitter | 37.7 | 8.7 | 2.5 | 28.2 | 16.6 | 24.2 | 19.8 |
| James Joyce | 17.6 | 10.7 | 1.9 | 1.9 | 1.5 | 1.5 | 2.2 |
| Song Lyrics | 29.3 | 20.1 | 2.9 | 10.3 | 13.6 | 9.4 | 11.7 |
| Romantic Poetry | 16.8 | 14.8 | 0.9 | 0.5 | 2.6 | 4.2 | 3.9 |
| Shakespeare | 11.4 | 8.0 | 0.8 | 5.9 | 7.3 | 6.6 | 6.9 |
| Switchboard | 41.4 | 15.2 | 0.5 | 0.0 | 0.2 | 3.0 | 6.9 |
| Overall | 29.4 | 13.3 | 1.6 | 8.9 | 8.7 | 11.1 | 11.0 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 38.7 | 43.8 | 18.7 | 14.1 | 14.5 | 5.7 | 25.7 | 23.5 | 16.7 | 15 | 25.5 | 24.8 | 23.8 | 22.9 |
| Bible | 29.6 | 32.4 | 19.4 | 20.1 | 0.3 | 0.5 | 3.9 | 3.9 | 10.8 | 11.3 | 13.6 | 15.5 | 15.3 | 15.4 |
| 1810-1820s | 19.6 | 22.3 | 5.2 | 8.4 | 0.2 | 1.5 | 9.9 | 12.6 | 11.2 | 14.1 | 12.6 | 16.7 | 12.1 | 15.7 |
| 1890-1900s | 29.2 | 32.3 | 10.9 | 14.2 | 2.0 | 5.3 | 8.5 | 10.2 | 11.3 | 12.6 | 12.4 | 12.8 | 11.4 | 11.2 |
| 1990-200s | 44.9 | 51.9 | 22.1 | 31.6 | 6.7 | 17.9 | 15.3 | 18.9 | 26.3 | 30.6 | 30.1 | 32.6 | 26.8 | 31.0 |
| English Twitter | 44.4 | 51.2 | 20.6 | 22.1 | 23.6 | 22.5 | 28.0 | 32.3 | 20.3 | 19.8 | 24.7 | 26.0 | 22.8 | 24.5 |
| James Joyce | 19.0 | 21.5 | 11.7 | 13.2 | 1.5 | 3.5 | 2.0 | 2.6 | 2.9 | 3.1 | 2.9 | 2.3 | 3.0 | 3.1 |
| Song Lyrics | 36.4 | 36.1 | 21.3 | 23.6 | 6.0 | 4.3 | 15.3 | 10.0 | 19.0 | 14.6 | 15.7 | 13.3 | 19.2 | 16.1 |
| Romantic Poetry | 11.2 | 11.4 | 10.0 | 11.2 | 0.2 | 0.3 | 1.4 | 0.7 | 3.7 | 2.9 | 4.7 | 3.5 | 4.2 | 2.8 |
| Shakespeare | 10.1 | 9.7 | 9.9 | 9.4 | 0.3 | 1.6 | 10.2 | 12.4 | 11.3 | 12.5 | 11.1 | 12.5 | 11.5 | 12.8 |
| Switchboard | 42.5 | 47.3 | 22.6 | 27.7 | 2.1 | 1.7 | 0.1 | 0.0 | 0.3 | 0.0 | 4.8 | 6.6 | 8.1 | 12.4 |
| Overall | 29.6 | 32.7 | 15.7 | 17.9 | 5.2 | 5.9 | 10.9 | 11.6 | 12.2 | 12.4 | 14.4 | 15.1 | 14.4 | 15.3 |
| V | TSS | F | MS | |
| AAE Twitter | 42.6 | 68.2 | 87.0 | 72.5 |
| Bible | 44.0 | 80.9 | 85.9 | 63.9 |
| 1810-1820s | 30.2 | 49.8 | 81.9 | 74.7 |
| 1890-1900s | 35.9 | 56.6 | 85.4 | 74.6 |
| 1990-200s | 42.3 | 63.0 | 83.9 | 77.8 |
| English Twitter | 41.2 | 62.6 | 87.7 | 74.5 |
| James Joyce | 20.4 | 34.1 | 81.8 | 76.7 |
| Song Lyrics | 33.3 | 51.6 | 87.2 | 74.7 |
| Romantic Poetry | 20.4 | 35.3 | 80.5 | 74.7 |
| Shakespeare | 13.6 | 24.7 | 83.7 | 72.7 |
| Switchboard | 52.9 | 92.0 | 87.0 | 66.1 |
| Overall | 34.3 | 56.3 | 84.7 | 73.0 |
| v | TSS | F | MS | |
| AAE Twitter | 7.4 | 21.8 | 65.7 | 66.2 |
| Bible | 26.9 | 70.5 | 79.4 | 51.7 |
| 1810-1820s | 11.1 | 22.5 | 78.6 | 64.9 |
| 1890-1900s | 12.3 | 22.6 | 82.6 | 65.0 |
| 1990-2000s | 16.6 | 29.1 | 82.8 | 65.0 |
| English Twitter | 8.0 | 36.9 | 76.8 | 49.8 |
| James Joyce | 11.8 | 26.4 | 78.8 | 66.7 |
| Song Lyrics | 20.2 | 39.4 | 80.2 | 67.3 |
| Romantic Poetry | 15.7 | 46.5 | 64.5 | 62.1 |
| Shakespeare | 9.1 | 27.8 | 68.0 | 62.2 |
| Switchboard | 21.1 | 52.0 | 69.3 | 65.2 |
| Overall | 14.6 | 36.0 | 75.2 | 62.4 |
| V | TSS | F | MS | |
| AAE Twitter | 3.8 | 10.3 | 76.1 | 68.5 |
| Bible | 6.6 | 24 | 79.1 | 67.3 |
| 1810-1820s | 3.5 | 8.7 | 74.4 | 78.9 |
| 1890-1900s | 4.4 | 10 | 76 | 77.7 |
| 1990-2000s | 4.3 | 9.1 | 75.4 | 79.2 |
| English Twitter | 5.5 | 12.5 | 74.1 | 82.1 |
| James Joyce | 5.4 | 13.6 | 76.9 | 75.3 |
| Song Lyrics | 7.7 | 23.2 | 78.2 | 65.3 |
| Romantic Poetry | 2.8 | 7.6 | 77.1 | 76.1 |
| Shakespeare | 2.5 | 8.2 | 77 | 72 |
| Switchboard | 1.7 | 3.9 | 75.1 | 84.5 |
| Overall | 4.4 | 11.9 | 76.3 | 75.2 |
| V | TSS | F | MS | |
| AAE Twitter | 23.2 | 58.4 | 63.1 | 65.7 |
| Bible | 5.2 | 10.4 | 85 | 70.4 |
| 1810-1820s | 14.7 | 27.6 | 79.4 | 67 |
| 1890-1900s | 8.6 | 15 | 83.1 | 65 |
| 1990-2000s | 7.9 | 12.6 | 83.8 | 68.8 |
| English Twitter | 35 | 58.3 | 82 | 70 |
| James Joyce | 3.4 | 7.4 | 77.1 | 66 |
| Song Lyrics | 12.2 | 25.4 | 73.7 | 72.2 |
| Romantic Poetry | 1.1 | 3.5 | 61.7 | 61.5 |
| Shakespeare | 9.6 | 27.7 | 64.3 | 68.6 |
| Switchboard | 0.1 | 0.1 | 92.2 | 74.5 |
| Overall | 11 | 22.4 | 76.9 | 68.2 |
| V | TSS | F | MS | |
| AAE Twitter | 11.2 | 30.9 | 67.3 | 64.8 |
| Bible | 16 | 40.2 | 79 | 59.5 |
| 1810-1820s | 15.9 | 29.4 | 82 | 65.3 |
| 1890-1900s | 9.1 | 16 | 82.4 | 63.3 |
| 1990-2000s | 13 | 25.2 | 86.2 | 65.3 |
| English Twitter | 23.6 | 47 | 71.3 | 68 |
| James Joyce | 1.3 | 11 | 46.1 | 63.2 |
| SongLyrics | 15.4 | 40.1 | 71.8 | 62.4 |
| Romantic Poetry | 3.4 | 15.3 | 51.7 | 59.1 |
| Shakespeare | 10 | 29.5 | 61.9 | 69 |
| Switchboard | 0.3 | 0.6 | 90.3 | 75.4 |
| Overall | 10.8 | 25.9 | 71.8 | 65 |
| V | TSS | F | MS | |
| AAE Twitter | 25.4 | 79.1 | 58.4 | 59 |
| Bible | 20.2 | 43.6 | 78.8 | 65.6 |
| 1810-1820s | 17.4 | 35.6 | 80.4 | 62.4 |
| 1890-1900s | 10.4 | 17.2 | 84.6 | 66.8 |
| 1990-2000s | 17.5 | 26.9 | 89.1 | 71 |
| English Twitter | 32 | 55.9 | 80.2 | 68.4 |
| James Joyce | 1.6 | 10.9 | 48.6 | 63.2 |
| Song Lyrics | 11.2 | 25.8 | 74.7 | 66 |
| Romantic Poetry | 6.2 | 28.5 | 44.6 | 57.4 |
| Shakespeare | 9.7 | 28 | 62.8 | 68.4 |
| Switchboard | 5.3 | 10.7 | 85.4 | 71.2 |
| Overall | 14.3 | 32.9 | 71.6 | 65.4 |
| V | TSS | F | MS | |
| AAE Twitter | 22.7 | 65.4 | 57.9 | 64.9 |
| Bible | 21 | 43.8 | 81.4 | 65.7 |
| 1810-1820s | 17 | 33.1 | 81 | 66.2 |
| 1890-1900s | 10.1 | 16.9 | 83.1 | 68.8 |
| 1990-2000s | 17.2 | 25.7 | 88.2 | 72.1 |
| English Twitter | 29.5 | 51.4 | 77.8 | 69.9 |
| James Joyce | 2.6 | 20.9 | 33.9 | 62.4 |
| Song Lyrics | 13.2 | 30.8 | 69.5 | 69.5 |
| Romantic Poetry | 4.9 | 23.2 | 43.9 | 58.9 |
| Shakespeare | 9.7 | 28.7 | 60.8 | 69.5 |
| Switchboard | 13.7 | 24.8 | 82.2 | 73.5 |
| Overall | 14.7 | 33.2 | 69.1 | 67.4 |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 42.6 | 7.4 | 3.8 | 23.2 | 11.2 | 25.4 | 22.7 |
| Bible | 44.0 | 26.9 | 6.6 | 5.2 | 16.0 | 20.2 | 21.0 |
| 1810-1820s | 30.2 | 11.1 | 3.5 | 14.7 | 15.9 | 17.4 | 17.0 |
| 1890-1900s | 35.9 | 12.3 | 4.4 | 8.6 | 9.1 | 10.4 | 10.1 |
| 1990-2000s | 42.3 | 16.6 | 4.3 | 7.9 | 13.0 | 17.5 | 17.2 |
| English Twitter | 41.2 | 8.0 | 5.5 | 35.0 | 23.6 | 32.0 | 29.5 |
| James Joyce | 20.4 | 11.8 | 5.4 | 3.4 | 1.3 | 1.6 | 2.6 |
| Song Lyrics | 33.3 | 20.2 | 7.7 | 12.2 | 15.4 | 11.2 | 13.2 |
| Romantic Poetry | 20.4 | 15.7 | 2.8 | 1.1 | 3.4 | 6.2 | 4.9 |
| Shakespeare | 13.6 | 9.1 | 2.5 | 9.6 | 10.0 | 9.7 | 9.7 |
| Switchboard | 52.9 | 21.1 | 1.7 | 0.1 | 0.3 | 5.3 | 13.7 |
| Overall | 34.3 | 14.6 | 4.4 | 11.0 | 10.8 | 14.3 | 14.7 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 44.0 | 47.7 | 18.7 | 13.2 | 25.6 | 10.6 | 31.7 | 29.2 | 21.5 | 17.9 | 30 | 28.8 | 30.2 | 27.6 |
| Bible | 36.1 | 38.8 | 22 | 22.9 | 0.3 | 1.6 | 4.3 | 4.4 | 15.7 | 15.9 | 18.0 | 19.0 | 19.8 | 19.5 |
| 1810-1820s | 26.3 | 29.5 | 5.9 | 10.0 | 1.2 | 4.7 | 12.4 | 15.6 | 14.3 | 16.9 | 17.6 | 21.6 | 16.9 | 20.1 |
| 1890-1900s | 33.5 | 34.7 | 10.0 | 13.4 | 4.4 | 11.0 | 9.9 | 11.8 | 13.9 | 13.8 | 14.6 | 14.4 | 13.8 | 13.3 |
| 1990-200s | 50.2 | 56.2 | 22.6 | 32.1 | 11.8 | 31.4 | 16.7 | 20.7 | 28.5 | 32.5 | 31.5 | 34.7 | 28.4 | 32.8 |
| English Twitter | 46.1 | 54.1 | 20.1 | 22.1 | 32.4 | 33.5 | 37.4 | 41.8 | 30.1 | 29.5 | 34.9 | 36.4 | 32.5 | 35.0 |
| James Joyce | 22.3 | 22.8 | 10.9 | 13.2 | 3.2 | 7.9 | 2.9 | 3.3 | 2.7 | 2.3 | 3.1 | 2.5 | 3.3 | 2.8 |
| Song Lyrics | 42.6 | 40.5 | 22.1 | 23.2 | 10.3 | 12.4 | 19.3 | 12.9 | 22.3 | 18.4 | 19.3 | 16.2 | 24.2 | 20.1 |
| Romantic Poetry | 13.5 | 12.9 | 8.9 | 10.8 | 0.8 | 0.9 | 2.0 | 1.1 | 5.2 | 4.3 | 7.0 | 4.7 | 6.0 | 3.9 |
| Shakespeare | 11.8 | 11.6 | 11.1 | 10.4 | 1.3 | 4.1 | 12.9 | 15.1 | 15.3 | 14.7 | 13.4 | 15.2 | 13.8 | 15.2 |
| Switchboard | 54.6 | 59.3 | 29.7 | 35.1 | 5.2 | 6.1 | 0.1 | 0.1 | 0.3 | 0.1 | 9.7 | 13.4 | 15.6 | 23.0 |
| Overall | 34.6 | 37.1 | 16.5 | 18.8 | 8.8 | 11.3 | 13.6 | 14.2 | 15.4 | 15.1 | 18.1 | 18.8 | 18.6 | 19.4 |
| Input: Can’t sleep at all. Smh. Transfer: AAE → 1990s-2000s | |
| STEER | I mean, I can’t sleep at all. |
| GPT-3 | I am unable to obtain any rest; shaking my head in disbelief. |
| STRAP | I don’t want to sleep. |
| P-A-R | Can’t sleep at all, smh. |
| Input: Yeah one one way or another Transfer: switchboard → rom. poetry | |
| STEER | One way, or another, or both |
| GPT-3 | Aye, one mayhap will find the way. |
| STRAP | And one way or the other |
| P-A-R | Yeah one one way or another |
| Input: In his fear, he dare not face me Transfer: lyrics → bible | |
| STEER | And he will not dare to face me: for fear of me is in his eyes. |
| GPT-3 | And his fear was great, so that he could not stand before me. |
| STRAP | For he that is afraid of me is of me; but he that is of me is of him. |
| P-A-R | In fear he came and hid himself, because God was near to him |
| Reward Type | V | TSS | F | MS |
| Coarse | 19.9 | 42.8 | 79.0 | 62.6 |
| Fine-grained | 27.4 | 47.0 | 83.6 | 71.5 |
| MSTTR | n-gram entropy, n = | F | MS | |||
| 1 | 2 | 3 | ||||
| Df(STEER) | 0.912 | 8.8 | 15.0 | 19.3 | 86.8 | 67.2 |
| GYAFC | 0.931 | 9.1 | 15.0 | 17.8 | 86.2 | 78.4 |
| Yelp | 0.958 | 9.6 | 16.4 | 21.3 | 87.2 | - |
| CDS | 0.946 | 10.0 | 17.0 | 20.4 | 83.2 | - |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 1 |
| effective batch size | 32 |
| total steps | 50,000 |
| steps per evaluation | 1000 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| Hyperparameter | Tested |
| α | [0.2,0.4,0.6] |
| temperature | [0.7, 1.0, 1.3] |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 4 |
| effective batch size | 1024 |
| total steps | 30,000 |
| steps per evaluation | 2500 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| QUARK hyps | |
| sample steps | 2500 |
| KL penalty | 0.025 |
| entropy coef | 0.00 |
| policy temp. | 1.0 |
| vectorized reward token | True |
| number of buckets | 5 |
| Style | Train | Validation | Test |
| AAE | 720K | 7.3K | 7.3K |
| Bible | 31K | 1.7K | 1.7K |
| Coha 1810 | 205K | 5.3K | 5.3K |
| Coha 1890 | 1.2M | 32K | 32K |
| Coha 1990 | 1.9M | 49K | 49K |
| English Tweet | 5.2M | 40K | 40K |
| Joyce | 37K | 2.0K | 2.0K |
| Lyrics | 4.5M | 252K | 252K |
| Romantic Poetry | 27K | 1.5K | 1.5K |
| Shakespeare | 25K | 1.3K | 1.3K |
| Switchboard | 146K | 1.5K | 1.5K |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 38.2 | 7.8 | 1.6 | 18.1 | 9.4 | 20.7 | 17.6 |
| Bible | 34.4 | 22 | 2.0 | 4.0 | 12.3 | 14.4 | 13.6 |
| 1810-1820s | 23.1 | 9.2 | 1.0 | 12.8 | 11.5 | 12.3 | 12.6 |
| 1890-1900s | 34.3 | 13.6 | 2.2 | 8.0 | 8.0 | 10.3 | 10.3 |
| 1990-200s | 39.0 | 16.1 | 1.7 | 8.7 | 13.1 | 15.6 | 15.8 |
| English Twitter | 37.7 | 8.7 | 2.5 | 28.2 | 16.6 | 24.2 | 19.8 |
| James Joyce | 17.6 | 10.7 | 1.9 | 1.9 | 1.5 | 1.5 | 2.2 |
| Song Lyrics | 29.3 | 20.1 | 2.9 | 10.3 | 13.6 | 9.4 | 11.7 |
| Romantic Poetry | 16.8 | 14.8 | 0.9 | 0.5 | 2.6 | 4.2 | 3.9 |
| Shakespeare | 11.4 | 8.0 | 0.8 | 5.9 | 7.3 | 6.6 | 6.9 |
| Switchboard | 41.4 | 15.2 | 0.5 | 0.0 | 0.2 | 3.0 | 6.9 |
| Overall | 29.4 | 13.3 | 1.6 | 8.9 | 8.7 | 11.1 | 11.0 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 38.7 | 43.8 | 18.7 | 14.1 | 14.5 | 5.7 | 25.7 | 23.5 | 16.7 | 15 | 25.5 | 24.8 | 23.8 | 22.9 |
| Bible | 29.6 | 32.4 | 19.4 | 20.1 | 0.3 | 0.5 | 3.9 | 3.9 | 10.8 | 11.3 | 13.6 | 15.5 | 15.3 | 15.4 |
| 1810-1820s | 19.6 | 22.3 | 5.2 | 8.4 | 0.2 | 1.5 | 9.9 | 12.6 | 11.2 | 14.1 | 12.6 | 16.7 | 12.1 | 15.7 |
| 1890-1900s | 29.2 | 32.3 | 10.9 | 14.2 | 2.0 | 5.3 | 8.5 | 10.2 | 11.3 | 12.6 | 12.4 | 12.8 | 11.4 | 11.2 |
| 1990-200s | 44.9 | 51.9 | 22.1 | 31.6 | 6.7 | 17.9 | 15.3 | 18.9 | 26.3 | 30.6 | 30.1 | 32.6 | 26.8 | 31.0 |
| English Twitter | 44.4 | 51.2 | 20.6 | 22.1 | 23.6 | 22.5 | 28.0 | 32.3 | 20.3 | 19.8 | 24.7 | 26.0 | 22.8 | 24.5 |
| James Joyce | 19.0 | 21.5 | 11.7 | 13.2 | 1.5 | 3.5 | 2.0 | 2.6 | 2.9 | 3.1 | 2.9 | 2.3 | 3.0 | 3.1 |
| Song Lyrics | 36.4 | 36.1 | 21.3 | 23.6 | 6.0 | 4.3 | 15.3 | 10.0 | 19.0 | 14.6 | 15.7 | 13.3 | 19.2 | 16.1 |
| Romantic Poetry | 11.2 | 11.4 | 10.0 | 11.2 | 0.2 | 0.3 | 1.4 | 0.7 | 3.7 | 2.9 | 4.7 | 3.5 | 4.2 | 2.8 |
| Shakespeare | 10.1 | 9.7 | 9.9 | 9.4 | 0.3 | 1.6 | 10.2 | 12.4 | 11.3 | 12.5 | 11.1 | 12.5 | 11.5 | 12.8 |
| Switchboard | 42.5 | 47.3 | 22.6 | 27.7 | 2.1 | 1.7 | 0.1 | 0.0 | 0.3 | 0.0 | 4.8 | 6.6 | 8.1 | 12.4 |
| Overall | 29.6 | 32.7 | 15.7 | 17.9 | 5.2 | 5.9 | 10.9 | 11.6 | 12.2 | 12.4 | 14.4 | 15.1 | 14.4 | 15.3 |
| V | TSS | F | MS | |
| AAE Twitter | 42.6 | 68.2 | 87.0 | 72.5 |
| Bible | 44.0 | 80.9 | 85.9 | 63.9 |
| 1810-1820s | 30.2 | 49.8 | 81.9 | 74.7 |
| 1890-1900s | 35.9 | 56.6 | 85.4 | 74.6 |
| 1990-200s | 42.3 | 63.0 | 83.9 | 77.8 |
| English Twitter | 41.2 | 62.6 | 87.7 | 74.5 |
| James Joyce | 20.4 | 34.1 | 81.8 | 76.7 |
| Song Lyrics | 33.3 | 51.6 | 87.2 | 74.7 |
| Romantic Poetry | 20.4 | 35.3 | 80.5 | 74.7 |
| Shakespeare | 13.6 | 24.7 | 83.7 | 72.7 |
| Switchboard | 52.9 | 92.0 | 87.0 | 66.1 |
| Overall | 34.3 | 56.3 | 84.7 | 73.0 |
| v | TSS | F | MS | |
| AAE Twitter | 7.4 | 21.8 | 65.7 | 66.2 |
| Bible | 26.9 | 70.5 | 79.4 | 51.7 |
| 1810-1820s | 11.1 | 22.5 | 78.6 | 64.9 |
| 1890-1900s | 12.3 | 22.6 | 82.6 | 65.0 |
| 1990-2000s | 16.6 | 29.1 | 82.8 | 65.0 |
| English Twitter | 8.0 | 36.9 | 76.8 | 49.8 |
| James Joyce | 11.8 | 26.4 | 78.8 | 66.7 |
| Song Lyrics | 20.2 | 39.4 | 80.2 | 67.3 |
| Romantic Poetry | 15.7 | 46.5 | 64.5 | 62.1 |
| Shakespeare | 9.1 | 27.8 | 68.0 | 62.2 |
| Switchboard | 21.1 | 52.0 | 69.3 | 65.2 |
| Overall | 14.6 | 36.0 | 75.2 | 62.4 |
| V | TSS | F | MS | |
| AAE Twitter | 3.8 | 10.3 | 76.1 | 68.5 |
| Bible | 6.6 | 24 | 79.1 | 67.3 |
| 1810-1820s | 3.5 | 8.7 | 74.4 | 78.9 |
| 1890-1900s | 4.4 | 10 | 76 | 77.7 |
| 1990-2000s | 4.3 | 9.1 | 75.4 | 79.2 |
| English Twitter | 5.5 | 12.5 | 74.1 | 82.1 |
| James Joyce | 5.4 | 13.6 | 76.9 | 75.3 |
| Song Lyrics | 7.7 | 23.2 | 78.2 | 65.3 |
| Romantic Poetry | 2.8 | 7.6 | 77.1 | 76.1 |
| Shakespeare | 2.5 | 8.2 | 77 | 72 |
| Switchboard | 1.7 | 3.9 | 75.1 | 84.5 |
| Overall | 4.4 | 11.9 | 76.3 | 75.2 |
| V | TSS | F | MS | |
| AAE Twitter | 23.2 | 58.4 | 63.1 | 65.7 |
| Bible | 5.2 | 10.4 | 85 | 70.4 |
| 1810-1820s | 14.7 | 27.6 | 79.4 | 67 |
| 1890-1900s | 8.6 | 15 | 83.1 | 65 |
| 1990-2000s | 7.9 | 12.6 | 83.8 | 68.8 |
| English Twitter | 35 | 58.3 | 82 | 70 |
| James Joyce | 3.4 | 7.4 | 77.1 | 66 |
| Song Lyrics | 12.2 | 25.4 | 73.7 | 72.2 |
| Romantic Poetry | 1.1 | 3.5 | 61.7 | 61.5 |
| Shakespeare | 9.6 | 27.7 | 64.3 | 68.6 |
| Switchboard | 0.1 | 0.1 | 92.2 | 74.5 |
| Overall | 11 | 22.4 | 76.9 | 68.2 |
| V | TSS | F | MS | |
| AAE Twitter | 11.2 | 30.9 | 67.3 | 64.8 |
| Bible | 16 | 40.2 | 79 | 59.5 |
| 1810-1820s | 15.9 | 29.4 | 82 | 65.3 |
| 1890-1900s | 9.1 | 16 | 82.4 | 63.3 |
| 1990-2000s | 13 | 25.2 | 86.2 | 65.3 |
| English Twitter | 23.6 | 47 | 71.3 | 68 |
| James Joyce | 1.3 | 11 | 46.1 | 63.2 |
| SongLyrics | 15.4 | 40.1 | 71.8 | 62.4 |
| Romantic Poetry | 3.4 | 15.3 | 51.7 | 59.1 |
| Shakespeare | 10 | 29.5 | 61.9 | 69 |
| Switchboard | 0.3 | 0.6 | 90.3 | 75.4 |
| Overall | 10.8 | 25.9 | 71.8 | 65 |
| V | TSS | F | MS | |
| AAE Twitter | 25.4 | 79.1 | 58.4 | 59 |
| Bible | 20.2 | 43.6 | 78.8 | 65.6 |
| 1810-1820s | 17.4 | 35.6 | 80.4 | 62.4 |
| 1890-1900s | 10.4 | 17.2 | 84.6 | 66.8 |
| 1990-2000s | 17.5 | 26.9 | 89.1 | 71 |
| English Twitter | 32 | 55.9 | 80.2 | 68.4 |
| James Joyce | 1.6 | 10.9 | 48.6 | 63.2 |
| Song Lyrics | 11.2 | 25.8 | 74.7 | 66 |
| Romantic Poetry | 6.2 | 28.5 | 44.6 | 57.4 |
| Shakespeare | 9.7 | 28 | 62.8 | 68.4 |
| Switchboard | 5.3 | 10.7 | 85.4 | 71.2 |
| Overall | 14.3 | 32.9 | 71.6 | 65.4 |
| V | TSS | F | MS | |
| AAE Twitter | 22.7 | 65.4 | 57.9 | 64.9 |
| Bible | 21 | 43.8 | 81.4 | 65.7 |
| 1810-1820s | 17 | 33.1 | 81 | 66.2 |
| 1890-1900s | 10.1 | 16.9 | 83.1 | 68.8 |
| 1990-2000s | 17.2 | 25.7 | 88.2 | 72.1 |
| English Twitter | 29.5 | 51.4 | 77.8 | 69.9 |
| James Joyce | 2.6 | 20.9 | 33.9 | 62.4 |
| Song Lyrics | 13.2 | 30.8 | 69.5 | 69.5 |
| Romantic Poetry | 4.9 | 23.2 | 43.9 | 58.9 |
| Shakespeare | 9.7 | 28.7 | 60.8 | 69.5 |
| Switchboard | 13.7 | 24.8 | 82.2 | 73.5 |
| Overall | 14.7 | 33.2 | 69.1 | 67.4 |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 42.6 | 7.4 | 3.8 | 23.2 | 11.2 | 25.4 | 22.7 |
| Bible | 44.0 | 26.9 | 6.6 | 5.2 | 16.0 | 20.2 | 21.0 |
| 1810-1820s | 30.2 | 11.1 | 3.5 | 14.7 | 15.9 | 17.4 | 17.0 |
| 1890-1900s | 35.9 | 12.3 | 4.4 | 8.6 | 9.1 | 10.4 | 10.1 |
| 1990-2000s | 42.3 | 16.6 | 4.3 | 7.9 | 13.0 | 17.5 | 17.2 |
| English Twitter | 41.2 | 8.0 | 5.5 | 35.0 | 23.6 | 32.0 | 29.5 |
| James Joyce | 20.4 | 11.8 | 5.4 | 3.4 | 1.3 | 1.6 | 2.6 |
| Song Lyrics | 33.3 | 20.2 | 7.7 | 12.2 | 15.4 | 11.2 | 13.2 |
| Romantic Poetry | 20.4 | 15.7 | 2.8 | 1.1 | 3.4 | 6.2 | 4.9 |
| Shakespeare | 13.6 | 9.1 | 2.5 | 9.6 | 10.0 | 9.7 | 9.7 |
| Switchboard | 52.9 | 21.1 | 1.7 | 0.1 | 0.3 | 5.3 | 13.7 |
| Overall | 34.3 | 14.6 | 4.4 | 11.0 | 10.8 | 14.3 | 14.7 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 44.0 | 47.7 | 18.7 | 13.2 | 25.6 | 10.6 | 31.7 | 29.2 | 21.5 | 17.9 | 30 | 28.8 | 30.2 | 27.6 |
| Bible | 36.1 | 38.8 | 22 | 22.9 | 0.3 | 1.6 | 4.3 | 4.4 | 15.7 | 15.9 | 18.0 | 19.0 | 19.8 | 19.5 |
| 1810-1820s | 26.3 | 29.5 | 5.9 | 10.0 | 1.2 | 4.7 | 12.4 | 15.6 | 14.3 | 16.9 | 17.6 | 21.6 | 16.9 | 20.1 |
| 1890-1900s | 33.5 | 34.7 | 10.0 | 13.4 | 4.4 | 11.0 | 9.9 | 11.8 | 13.9 | 13.8 | 14.6 | 14.4 | 13.8 | 13.3 |
| 1990-200s | 50.2 | 56.2 | 22.6 | 32.1 | 11.8 | 31.4 | 16.7 | 20.7 | 28.5 | 32.5 | 31.5 | 34.7 | 28.4 | 32.8 |
| English Twitter | 46.1 | 54.1 | 20.1 | 22.1 | 32.4 | 33.5 | 37.4 | 41.8 | 30.1 | 29.5 | 34.9 | 36.4 | 32.5 | 35.0 |
| James Joyce | 22.3 | 22.8 | 10.9 | 13.2 | 3.2 | 7.9 | 2.9 | 3.3 | 2.7 | 2.3 | 3.1 | 2.5 | 3.3 | 2.8 |
| Song Lyrics | 42.6 | 40.5 | 22.1 | 23.2 | 10.3 | 12.4 | 19.3 | 12.9 | 22.3 | 18.4 | 19.3 | 16.2 | 24.2 | 20.1 |
| Romantic Poetry | 13.5 | 12.9 | 8.9 | 10.8 | 0.8 | 0.9 | 2.0 | 1.1 | 5.2 | 4.3 | 7.0 | 4.7 | 6.0 | 3.9 |
| Shakespeare | 11.8 | 11.6 | 11.1 | 10.4 | 1.3 | 4.1 | 12.9 | 15.1 | 15.3 | 14.7 | 13.4 | 15.2 | 13.8 | 15.2 |
| Switchboard | 54.6 | 59.3 | 29.7 | 35.1 | 5.2 | 6.1 | 0.1 | 0.1 | 0.3 | 0.1 | 9.7 | 13.4 | 15.6 | 23.0 |
| Overall | 34.6 | 37.1 | 16.5 | 18.8 | 8.8 | 11.3 | 13.6 | 14.2 | 15.4 | 15.1 | 18.1 | 18.8 | 18.6 | 19.4 |
| Input: Can’t sleep at all. Smh. Transfer: AAE → 1990s-2000s | |
| STEER | I mean, I can’t sleep at all. |
| GPT-3 | I am unable to obtain any rest; shaking my head in disbelief. |
| STRAP | I don’t want to sleep. |
| P-A-R | Can’t sleep at all, smh. |
| Input: Yeah one one way or another Transfer: switchboard → rom. poetry | |
| STEER | One way, or another, or both |
| GPT-3 | Aye, one mayhap will find the way. |
| STRAP | And one way or the other |
| P-A-R | Yeah one one way or another |
| Input: In his fear, he dare not face me Transfer: lyrics → bible | |
| STEER | And he will not dare to face me: for fear of me is in his eyes. |
| GPT-3 | And his fear was great, so that he could not stand before me. |
| STRAP | For he that is afraid of me is of me; but he that is of me is of him. |
| P-A-R | In fear he came and hid himself, because God was near to him |
| Reward Type | V | TSS | F | MS |
| Coarse | 19.9 | 42.8 | 79.0 | 62.6 |
| Fine-grained | 27.4 | 47.0 | 83.6 | 71.5 |
| MSTTR | n-gram entropy, n = | F | MS | |||
| 1 | 2 | 3 | ||||
| Df(STEER) | 0.912 | 8.8 | 15.0 | 19.3 | 86.8 | 67.2 |
| GYAFC | 0.931 | 9.1 | 15.0 | 17.8 | 86.2 | 78.4 |
| Yelp | 0.958 | 9.6 | 16.4 | 21.3 | 87.2 | - |
| CDS | 0.946 | 10.0 | 17.0 | 20.4 | 83.2 | - |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 1 |
| effective batch size | 32 |
| total steps | 50,000 |
| steps per evaluation | 1000 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| Hyperparameter | Tested |
| α | [0.2,0.4,0.6] |
| temperature | [0.7, 1.0, 1.3] |
| Hyperparameter | Assignment |
| model | GPT2-large |
| number of gpus | 4 |
| effective batch size | 1024 |
| total steps | 30,000 |
| steps per evaluation | 2500 |
| learning rate optimizer | AdamW |
| AdamW initial learning rate | 1e-05 |
| AdamW epsilon | 1e-05 |
| learning rate schedule | linear with no warmup |
| weight decay | 0.0 |
| max sequence length | 50 |
| max generation length | 100 |
| padding sequences | to max seq length |
| QUARK hyps | |
| sample steps | 2500 |
| KL penalty | 0.025 |
| entropy coef | 0.00 |
| policy temp. | 1.0 |
| vectorized reward token | True |
| number of buckets | 5 |
| Style | Train | Validation | Test |
| AAE | 720K | 7.3K | 7.3K |
| Bible | 31K | 1.7K | 1.7K |
| Coha 1810 | 205K | 5.3K | 5.3K |
| Coha 1890 | 1.2M | 32K | 32K |
| Coha 1990 | 1.9M | 49K | 49K |
| English Tweet | 5.2M | 40K | 40K |
| Joyce | 37K | 2.0K | 2.0K |
| Lyrics | 4.5M | 252K | 252K |
| Romantic Poetry | 27K | 1.5K | 1.5K |
| Shakespeare | 25K | 1.3K | 1.3K |
| Switchboard | 146K | 1.5K | 1.5K |
| Target Style | GPT-2 Large | GPT-3 (text-davinci-003) | |||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | |
| AAE Twitter | 38.2 | 7.8 | 1.6 | 18.1 | 9.4 | 20.7 | 17.6 |
| Bible | 34.4 | 22 | 2.0 | 4.0 | 12.3 | 14.4 | 13.6 |
| 1810-1820s | 23.1 | 9.2 | 1.0 | 12.8 | 11.5 | 12.3 | 12.6 |
| 1890-1900s | 34.3 | 13.6 | 2.2 | 8.0 | 8.0 | 10.3 | 10.3 |
| 1990-200s | 39.0 | 16.1 | 1.7 | 8.7 | 13.1 | 15.6 | 15.8 |
| English Twitter | 37.7 | 8.7 | 2.5 | 28.2 | 16.6 | 24.2 | 19.8 |
| James Joyce | 17.6 | 10.7 | 1.9 | 1.9 | 1.5 | 1.5 | 2.2 |
| Song Lyrics | 29.3 | 20.1 | 2.9 | 10.3 | 13.6 | 9.4 | 11.7 |
| Romantic Poetry | 16.8 | 14.8 | 0.9 | 0.5 | 2.6 | 4.2 | 3.9 |
| Shakespeare | 11.4 | 8.0 | 0.8 | 5.9 | 7.3 | 6.6 | 6.9 |
| Switchboard | 41.4 | 15.2 | 0.5 | 0.0 | 0.2 | 3.0 | 6.9 |
| Overall | 29.4 | 13.3 | 1.6 | 8.9 | 8.7 | 11.1 | 11.0 |
| Target Style | GPT2-Large | GPT-3 (text-davinci-003) | ||||||||||||
| STEER | STRAP | P-A-R | k=0 | k=1 | k=5 | k=10 | ||||||||
| Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | Inf. | For. | |
| AAE Twitter | 38.7 | 43.8 | 18.7 | 14.1 | 14.5 | 5.7 | 25.7 | 23.5 | 16.7 | 15 | 25.5 | 24.8 | 23.8 | 22.9 |
| Bible | 29.6 | 32.4 | 19.4 | 20.1 | 0.3 | 0.5 | 3.9 | 3.9 | 10.8 | 11.3 | 13.6 | 15.5 | 15.3 | 15.4 |
| 1810-1820s | 19.6 | 22.3 | 5.2 | 8.4 | 0.2 | 1.5 | 9.9 | 12.6 | 11.2 | 14.1 | 12.6 | 16.7 | 12.1 | 15.7 |
| 1890-1900s | 29.2 | 32.3 | 10.9 | 14.2 | 2.0 | 5.3 | 8.5 | 10.2 | 11.3 | 12.6 | 12.4 | 12.8 | 11.4 | 11.2 |
| 1990-200s | 44.9 | 51.9 | 22.1 | 31.6 | 6.7 | 17.9 | 15.3 | 18.9 | 26.3 | 30.6 | 30.1 | 32.6 | 26.8 | 31.0 |
| English Twitter | 44.4 | 51.2 | 20.6 | 22.1 | 23.6 | 22.5 | 28.0 | 32.3 | 20.3 | 19.8 | 24.7 | 26.0 | 22.8 | 24.5 |
| James Joyce | 19.0 | 21.5 | 11.7 | 13.2 | 1.5 | 3.5 | 2.0 | 2.6 | 2.9 | 3.1 | 2.9 | 2.3 | 3.0 | 3.1 |
| Song Lyrics | 36.4 | 36.1 | 21.3 | 23.6 | 6.0 | 4.3 | 15.3 | 10.0 | 19.0 | 14.6 | 15.7 | 13.3 | 19.2 | 16.1 |
| Romantic Poetry | 11.2 | 11.4 | 10.0 | 11.2 | 0.2 | 0.3 | 1.4 | 0.7 | 3.7 | 2.9 | 4.7 | 3.5 | 4.2 | 2.8 |
| Shakespeare | 10.1 | 9.7 | 9.9 | 9.4 | 0.3 | 1.6 | 10.2 | 12.4 | 11.3 | 12.5 | 11.1 | 12.5 | 11.5 | 12.8 |
| Switchboard | 42.5 | 47.3 | 22.6 | 27.7 | 2.1 | 1.7 | 0.1 | 0.0 | 0.3 | 0.0 | 4.8 | 6.6 | 8.1 | 12.4 |
| Overall | 29.6 | 32.7 | 15.7 | 17.9 | 5.2 | 5.9 | 10.9 | 11.6 | 12.2 | 12.4 | 14.4 | 15.1 | 14.4 | 15.3 |
| V | TSS | F | MS | |
| AAE Twitter | 42.6 | 68.2 | 87.0 | 72.5 |
| Bible | 44.0 | 80.9 | 85.9 | 63.9 |
| 1810-1820s | 30.2 | 49.8 | 81.9 | 74.7 |
| 1890-1900s | 35.9 | 56.6 | 85.4 | 74.6 |
| 1990-200s | 42.3 | 63.0 | 83.9 | 77.8 |
| English Twitter | 41.2 | 62.6 | 87.7 | 74.5 |
| James Joyce | 20.4 | 34.1 | 81.8 | 76.7 |
| Song Lyrics | 33.3 | 51.6 | 87.2 | 74.7 |
| Romantic Poetry | 20.4 | 35.3 | 80.5 | 74.7 |
| Shakespeare | 13.6 | 24.7 | 83.7 | 72.7 |
| Switchboard | 52.9 | 92.0 | 87.0 | 66.1 |
| Overall | 34.3 | 56.3 | 84.7 | 73.0 |
| v | TSS | F | MS | |
| AAE Twitter | 7.4 | 21.8 | 65.7 | 66.2 |
| Bible | 26.9 | 70.5 | 79.4 | 51.7 |
| 1810-1820s | 11.1 | 22.5 | 78.6 | 64.9 |
| 1890-1900s | 12.3 | 22.6 | 82.6 | 65.0 |
| 1990-2000s | 16.6 | 29.1 | 82.8 | 65.0 |
| English Twitter | 8.0 | 36.9 | 76.8 | 49.8 |
| James Joyce | 11.8 | 26.4 | 78.8 | 66.7 |
| Song Lyrics | 20.2 | 39.4 | 80.2 | 67.3 |
| Romantic Poetry | 15.7 | 46.5 | 64.5 | 62.1 |
| Shakespeare | 9.1 | 27.8 | 68.0 | 62.2 |
| Switchboard | 21.1 | 52.0 | 69.3 | 65.2 |
| Overall | 14.6 | 36.0 | 75.2 | 62.4 |
| V | TSS | F | MS | |
| AAE Twitter | 3.8 | 10.3 | 76.1 | 68.5 |
| Bible | 6.6 | 24 | 79.1 | 67.3 |
| 1810-1820s | 3.5 | 8.7 | 74.4 | 78.9 |
| 1890-1900s | 4.4 | 10 | 76 | 77.7 |
| 1990-2000s | 4.3 | 9.1 | 75.4 | 79.2 |
| English Twitter | 5.5 | 12.5 | 74.1 | 82.1 |
| James Joyce | 5.4 | 13.6 | 76.9 | 75.3 |
| Song Lyrics | 7.7 | 23.2 | 78.2 | 65.3 |
| Romantic Poetry | 2.8 | 7.6 | 77.1 | 76.1 |
| Shakespeare | 2.5 | 8.2 | 77 | 72 |
| Switchboard | 1.7 | 3.9 | 75.1 | 84.5 |
| Overall | 4.4 | 11.9 | 76.3 | 75.2 |
| V | TSS | F | MS | |
| AAE Twitter | 23.2 | 58.4 | 63.1 | 65.7 |
| Bible | 5.2 | 10.4 | 85 | 70.4 |
| 1810-1820s | 14.7 | 27.6 | 79.4 | 67 |
| 1890-1900s | 8.6 | 15 | 83.1 | 65 |
| 1990-2000s | 7.9 | 12.6 | 83.8 | 68.8 |
| English Twitter | 35 | 58.3 | 82 | 70 |
| James Joyce | 3.4 | 7.4 | 77.1 | 66 |
| Song Lyrics | 12.2 | 25.4 | 73.7 | 72.2 |
| Romantic Poetry | 1.1 | 3.5 | 61.7 | 61.5 |
| Shakespeare | 9.6 | 27.7 | 64.3 | 68.6 |
| Switchboard | 0.1 | 0.1 | 92.2 | 74.5 |
| Overall | 11 | 22.4 | 76.9 | 68.2 |
| V | TSS | F | MS | |
| AAE Twitter | 11.2 | 30.9 | 67.3 | 64.8 |
| Bible | 16 | 40.2 | 79 | 59.5 |
| 1810-1820s | 15.9 | 29.4 | 82 | 65.3 |
| 1890-1900s | 9.1 | 16 | 82.4 | 63.3 |
| 1990-2000s | 13 | 25.2 | 86.2 | 65.3 |
| English Twitter | 23.6 | 47 | 71.3 | 68 |
| James Joyce | 1.3 | 11 | 46.1 | 63.2 |
| SongLyrics | 15.4 | 40.1 | 71.8 | 62.4 |
| Romantic Poetry | 3.4 | 15.3 | 51.7 | 59.1 |
| Shakespeare | 10 | 29.5 | 61.9 | 69 |
| Switchboard | 0.3 | 0.6 | 90.3 | 75.4 |
| Overall | 10.8 | 25.9 | 71.8 | 65 |
| V | TSS | F | MS | |
| AAE Twitter | 25.4 | 79.1 | 58.4 | 59 |
| Bible | 20.2 | 43.6 | 78.8 | 65.6 |
| 1810-1820s | 17.4 | 35.6 | 80.4 | 62.4 |
| 1890-1900s | 10.4 | 17.2 | 84.6 | 66.8 |
| 1990-2000s | 17.5 | 26.9 | 89.1 | 71 |
| English Twitter | 32 | 55.9 | 80.2 | 68.4 |
| James Joyce | 1.6 | 10.9 | 48.6 | 63.2 |
| Song Lyrics | 11.2 | 25.8 | 74.7 | 66 |
| Romantic Poetry | 6.2 | 28.5 | 44.6 | 57.4 |
| Shakespeare | 9.7 | 28 | 62.8 | 68.4 |
| Switchboard | 5.3 | 10.7 | 85.4 | 71.2 |
| Overall | 14.3 | 32.9 | 71.6 | 65.4 |
| V | TSS | F | MS | |
| AAE Twitter | 22.7 | 65.4 | 57.9 | 64.9 |
| Bible | 21 | 43.8 | 81.4 | 65.7 |
| 1810-1820s | 17 | 33.1 | 81 | 66.2 |
| 1890-1900s | 10.1 | 16.9 | 83.1 | 68.8 |
| 1990-2000s | 17.2 | 25.7 | 88.2 | 72.1 |
| English Twitter | 29.5 | 51.4 | 77.8 | 69.9 |
| James Joyce | 2.6 | 20.9 | 33.9 | 62.4 |
| Song Lyrics | 13.2 | 30.8 | 69.5 | 69.5 |
| Romantic Poetry | 4.9 | 23.2 | 43.9 | 58.9 |
| Shakespeare | 9.7 | 28.7 | 60.8 | 69.5 |
| Switchboard | 13.7 | 24.8 | 82.2 | 73.5 |
| Overall | 14.7 | 33.2 | 69.1 | 67.4 |
| Dataset | Train Size | Test Size | Total Size | Batch Size |
| CoNLL2003 | 14041 | 3453 | 17494 | 128 |
| WNUT2017 | 3394 | 1287 | 4681 | 128 |
| FewNERD-L1 | 131767 | 18824 | 150591 | 32 |
| WikiNER-en | 129907 | 14435 | 144342 | 128 |
| BTC | - | 2002 | - | - |
| Tweebank | - | 1201 | - | - |
| HTNAME | - | 995 | 995 | - |
| HTUNSUP | 6160 | - | 6160 | 32 |
| HTGEN | 9424 | 818 | 10242 | 32 |
| F1 | Precision | Recall |
| 0.90 | 0.89 | 0.91 |
| Method | CoNLL2003 | WNUT2017 | BTC | Tweebank | ||||||||
| F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .64 ± .04 | .66 ± .04 | .55 ± .04 | .21 ± .07 | .14 ± .06 | .44 ± .06 | .15 ± .01 | .44 ± .01 | .09 ± .01 | .19 ± .04 | .24 ± .05 | .18 ± .05 |
| TwitterNER (Mishra and Diesner, 2016) | .68 ± .05 | .91 ± .05 | .55 ± .05 | .61 ± .09 | .84 ± .10 | .57 ± .10 | .32 ± .05 | .86 ± .04 | .26 ± .04 | .67 ± .04 | .87 ± .05 | .57 ± .05 |
| LUKE (Yamada et al., 2020) | .31 ± .11 | .89 ± .09 | .19 ± .09 | .55 ± .07 | .67 ± .05 | .44 ± .05 | .47 ± .01 | .92 ± .01 | .31 ± .01 | .37 ± .07 | .76 ± .06 | .33 ± .06 |
| ELMo (Peters et al., 2018) | .96 ± .02 | .95 ± .02 | .99 ± .02 | .59 ± .15 | .72 ± .18 | .37 ± .18 | .41 ± .06 | .83 ± .04 | .35 ± .18 | .71 ± .01 | .74 ± .01 | .68 ± .01 |
| Flair (Akbik et al., 2019) | .98 ± .02 | .97 ± .02 | 1.0 ± .02 | .60 ± .15 | .79 ± .18 | .34 ± .18 | .35 ± .06 | .86 ± .05 | .31 ± .05 | .63 ± .07 | .76 ± .09 | .62 ± .09 |
| NEAT (Original) (Li et al., 2022) | .17 ± .07 | .43 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| NEAT (Weakly Supervised) | .16 ± .07 | .42 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| Majority vote | .83 ± .06 | .74 ± .02 | .98 ± .02 | .65 ± .04 | .53 ± .06 | .90 ± .06 | .64 ± .03 | .76 ± .03 | .60 ± .03 | .53 ± .06 | .37 ± .04 | .88 ± .04 |
| SWEET - Domain Data | .86 ± .05 | .79 ± .02 | .97 ± .02 | .69 ± .06 | .61 ± .06 | .82 ± .06 | .63 ± .03 | .86 ± .04 | .56 ± .04 | .58 ± .06 | .43 ± .04 | .85 ± .04 |
| SWEET | .86 ± .05 | .79 ± .03 | .98 ± .03 | .68 ± .04 | .58 ± .07 | .83 ± .07 | .64 ± .03 | .84 ± .04 | .57 ± .04 | .58 ± .06 | .43 ± .04 | .86 ± .04 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .27 ± .03 | .18 ± .02 | .51± .02 | .47 ± .04 | .50 ± .03 | .43 ± .03 |
| TwitterNER (Mishra and Diesner, 2016) | .56 ± .04 | 75. ± .04 | 52.± .04 | .70 ± .02 | .70 ± .03 | .67 ± .03 |
| LUKE (Yamada et al., 2020) | .63 ± .03 | .85 ± .04 | .51± .04 | .68 ± .03 | .84 ± .02 | .56 ± .02 |
| ELMo (Peters et al., 2018) | .51 ± .02 | .56 ± .02 | .46± .02 | .69 ± .05 | .61 ± .06 | .74 ± .06 |
| Flair (Akbik et al., 2019) | .45 ± .02 | .73 ± .02 | .32± .02 | .63 ± .04 | .83 ± .04 | .49 ± .04 |
| NEAT (Original) (Li et al., 2022) | .78 ± .04 | .83 ± .05 | .74± .03 | .71 ± .01 | .63 ± .02 | .79 ± .03 |
| NEAT (Weakly Supervised) | .79 ± .02 | .80 ± .02 | .77± .02 | .71 ± .01 | .64 ± .02 | .78 ± .02 |
| Majority vote | .73 ± .02 | .59 ± .01 | .95 ± .01 | .74 ± .02 | .65 ± .03 | .85 ± .03 |
| SWEET -Domain Data | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| SWEET | .87 ± .01 | .83 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| Model | Fine-tuning Dataset | F1 | Precision | Recall |
| DeBERTa-v3-base | HTUNSUP | .67 ± .03 | .71 ± .02 | .62 ± .02 |
| HTGEN | .68 ± .01 | .71 ± .02 | .67 ± .02 | |
| CoNLL2003 | .67 ± .02 | .67 ± .03 | .69 ± .03 | |
| Few-NERD-L1 | .57 ± .03 | .80 ± .03 | .43 ± .03 | |
| WikiNER-en | .52 ± .01 | .48 ± .02 | .54 ± .02 | |
| WNUT2017 | .70 ± .02 | .71 ± .02 | .72 ± .02 | |
| RoBERTa-base | HTUNSUP | .82 ± .02 | .84 ± .03 | .83 ± .03 |
| HTGEN | .72 ± .02 | .81 ± .02 | .66 ± .02 | |
| CoNLL2003 | .72 ± .02 | .68 ± .03 | .77 ± .03 | |
| Few-NERD-L1 | .68 ± .02 | .81 ± .02 | .59 ± .02 | |
| WikiNER-en | .49 ± .03 | .43 ± .03 | .56 ± .03 | |
| WNUT2017 | .68 ± .02 | .73 ± .03 | .66 ± .03 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| All LFs - HTGEN LFs | .88 ± .01 | .85 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - HTUNSUP LFs | .86 ± .02 | .82 ± .02 | .91 ± .02 | .82 ± .02 | .77 ± .03 | .84 ± .03 |
| All LFs - HTGEN LFs- HTUNSUP LFs (Domain Data) | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| All LFs - Antirules | .85 ± .01 | .79 ± .01 | .92 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs | .85 ± .01 | .79 ± .01 | .93 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTUNSUP LFs | .83 ± .02 | .76 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs - HTUNSUP LFs | .82 ± .02 | .76 ± .02 | .92 ± .02 | .81 ± .01 | .76 ± .03 | .84 ± .03 |
| Only HTGEN LFs | .76 ± .01 | .76 ± .02 | .76 ± .02 | .81 ± .02 | .79 ± .03 | .81 ± .03 |
| Only HTGEN LFs + Antirules | .86 ± .01 | .82 ± .01 | .92 ± .01 | .75 ± .04 | .82 ± .05 | .64 ± .05 |
| Only HTUNSUP LFs | .87 ± .01 | .88 ± .01 | .88 ± .01 | .81 ± .03 | .77 ± .04 | .81 ± .04 |
| Only HTUNSUP LFs + Antirules | .87 ± .01 | .92 ± .02 | .85 ± .02 | .76 ± .04 | .83 ± .05 | .66 ± .05 |
| Only HTGEN LFs + HTUNSUP LFs | .85 ± .01 | .85 ± .02 | .87 ± .02 | .82 ± .03 | .78 ± .03 | .81 ± .03 |
| Only HTGEN LFs + HTUNSUP LFs + Antirules | .86 ± .02 | .86 ± .01 | .86 ± .01 | .82 ± .03 | .79 ± .04 | .81 ± .04 |
| Model | F1 | Precision | Recall |
| Label-Indep NEAT Rules | .43 ± .03 | .88 ± .02 | .28 ± .02 |
| Label-Indep NEAT Dictionary | .78 ± .02 | .83 ± .02 | .74 ± .02 |
| Label-Indep NEAT Disambiguation | .79 ± .02 | .80 ± .02 | .77 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .67 ± .11 | .82 ± .14 | .57 ± .14 |
| CoNLL2003 | .97 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .95 ± .01 | .91 ± .01 | |
| WikiNER-en | .96 ± .01 | .97 ± .02 | .97 ± .02 | |
| WNUT2017 | .58 ± .04 | .56 ± .04 | .65 ± .04 | |
| RoBERTa-base | HTUNSUP | .74 ± .06 | .84 ± .07 | .60 ± .07 |
| CoNLL2003 | .98 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .94 ± .02 | .92 ± .02 | |
| WikiNER-en | .93 ± .03 | .94 ± .05 | .97 ± .05 | |
| WNUT2017 | .89 ± .02 | .87 ± .02 | .88 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .41 ± .11 | .50 ± .14 | .52 ± .14 |
| CoNLL2003 | .72 ± .04 | .71 ± .08 | .73 ± .08 | |
| Few-NERD-L1 | .66 ± .12 | .82 ± .13 | .41 ± .13 | |
| WikiNER-en | .65 ± .11 | .69 ± .13 | .52 ± .13 | |
| WNUT2017 | .62 ± .09 | .71 ± .15 | .47 ± .15 | |
| RoBERTa-base | HTUNSUP | .53 ± .07 | .56 ± .13 | .63 ± .13 |
| CoNLL2003 | .71 ± .05 | .67 ± .08 | .71 ± .08 | |
| Few-NERD-L1 | .61 ± .10 | .67 ± .10 | .45 ± .10 | |
| WikiNER-en | .65 ± .10 | .63 ± .09 | .57 ± .09 | |
| WNUT2017 | .67 ± .09 | .81 ± .14 | .48 ± .14 |
| Dataset | Train Size | Test Size | Total Size | Batch Size |
| CoNLL2003 | 14041 | 3453 | 17494 | 128 |
| WNUT2017 | 3394 | 1287 | 4681 | 128 |
| FewNERD-L1 | 131767 | 18824 | 150591 | 32 |
| WikiNER-en | 129907 | 14435 | 144342 | 128 |
| BTC | - | 2002 | - | - |
| Tweebank | - | 1201 | - | - |
| HTNAME | - | 995 | 995 | - |
| HTUNSUP | 6160 | - | 6160 | 32 |
| HTGEN | 9424 | 818 | 10242 | 32 |
| F1 | Precision | Recall |
| 0.90 | 0.89 | 0.91 |
| Method | CoNLL2003 | WNUT2017 | BTC | Tweebank | ||||||||
| F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .64 ± .04 | .66 ± .04 | .55 ± .04 | .21 ± .07 | .14 ± .06 | .44 ± .06 | .15 ± .01 | .44 ± .01 | .09 ± .01 | .19 ± .04 | .24 ± .05 | .18 ± .05 |
| TwitterNER (Mishra and Diesner, 2016) | .68 ± .05 | .91 ± .05 | .55 ± .05 | .61 ± .09 | .84 ± .10 | .57 ± .10 | .32 ± .05 | .86 ± .04 | .26 ± .04 | .67 ± .04 | .87 ± .05 | .57 ± .05 |
| LUKE (Yamada et al., 2020) | .31 ± .11 | .89 ± .09 | .19 ± .09 | .55 ± .07 | .67 ± .05 | .44 ± .05 | .47 ± .01 | .92 ± .01 | .31 ± .01 | .37 ± .07 | .76 ± .06 | .33 ± .06 |
| ELMo (Peters et al., 2018) | .96 ± .02 | .95 ± .02 | .99 ± .02 | .59 ± .15 | .72 ± .18 | .37 ± .18 | .41 ± .06 | .83 ± .04 | .35 ± .18 | .71 ± .01 | .74 ± .01 | .68 ± .01 |
| Flair (Akbik et al., 2019) | .98 ± .02 | .97 ± .02 | 1.0 ± .02 | .60 ± .15 | .79 ± .18 | .34 ± .18 | .35 ± .06 | .86 ± .05 | .31 ± .05 | .63 ± .07 | .76 ± .09 | .62 ± .09 |
| NEAT (Original) (Li et al., 2022) | .17 ± .07 | .43 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| NEAT (Weakly Supervised) | .16 ± .07 | .42 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| Majority vote | .83 ± .06 | .74 ± .02 | .98 ± .02 | .65 ± .04 | .53 ± .06 | .90 ± .06 | .64 ± .03 | .76 ± .03 | .60 ± .03 | .53 ± .06 | .37 ± .04 | .88 ± .04 |
| SWEET - Domain Data | .86 ± .05 | .79 ± .02 | .97 ± .02 | .69 ± .06 | .61 ± .06 | .82 ± .06 | .63 ± .03 | .86 ± .04 | .56 ± .04 | .58 ± .06 | .43 ± .04 | .85 ± .04 |
| SWEET | .86 ± .05 | .79 ± .03 | .98 ± .03 | .68 ± .04 | .58 ± .07 | .83 ± .07 | .64 ± .03 | .84 ± .04 | .57 ± .04 | .58 ± .06 | .43 ± .04 | .86 ± .04 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .27 ± .03 | .18 ± .02 | .51± .02 | .47 ± .04 | .50 ± .03 | .43 ± .03 |
| TwitterNER (Mishra and Diesner, 2016) | .56 ± .04 | 75. ± .04 | 52.± .04 | .70 ± .02 | .70 ± .03 | .67 ± .03 |
| LUKE (Yamada et al., 2020) | .63 ± .03 | .85 ± .04 | .51± .04 | .68 ± .03 | .84 ± .02 | .56 ± .02 |
| ELMo (Peters et al., 2018) | .51 ± .02 | .56 ± .02 | .46± .02 | .69 ± .05 | .61 ± .06 | .74 ± .06 |
| Flair (Akbik et al., 2019) | .45 ± .02 | .73 ± .02 | .32± .02 | .63 ± .04 | .83 ± .04 | .49 ± .04 |
| NEAT (Original) (Li et al., 2022) | .78 ± .04 | .83 ± .05 | .74± .03 | .71 ± .01 | .63 ± .02 | .79 ± .03 |
| NEAT (Weakly Supervised) | .79 ± .02 | .80 ± .02 | .77± .02 | .71 ± .01 | .64 ± .02 | .78 ± .02 |
| Majority vote | .73 ± .02 | .59 ± .01 | .95 ± .01 | .74 ± .02 | .65 ± .03 | .85 ± .03 |
| SWEET -Domain Data | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| SWEET | .87 ± .01 | .83 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| Model | Fine-tuning Dataset | F1 | Precision | Recall |
| DeBERTa-v3-base | HTUNSUP | .67 ± .03 | .71 ± .02 | .62 ± .02 |
| HTGEN | .68 ± .01 | .71 ± .02 | .67 ± .02 | |
| CoNLL2003 | .67 ± .02 | .67 ± .03 | .69 ± .03 | |
| Few-NERD-L1 | .57 ± .03 | .80 ± .03 | .43 ± .03 | |
| WikiNER-en | .52 ± .01 | .48 ± .02 | .54 ± .02 | |
| WNUT2017 | .70 ± .02 | .71 ± .02 | .72 ± .02 | |
| RoBERTa-base | HTUNSUP | .82 ± .02 | .84 ± .03 | .83 ± .03 |
| HTGEN | .72 ± .02 | .81 ± .02 | .66 ± .02 | |
| CoNLL2003 | .72 ± .02 | .68 ± .03 | .77 ± .03 | |
| Few-NERD-L1 | .68 ± .02 | .81 ± .02 | .59 ± .02 | |
| WikiNER-en | .49 ± .03 | .43 ± .03 | .56 ± .03 | |
| WNUT2017 | .68 ± .02 | .73 ± .03 | .66 ± .03 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| All LFs - HTGEN LFs | .88 ± .01 | .85 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - HTUNSUP LFs | .86 ± .02 | .82 ± .02 | .91 ± .02 | .82 ± .02 | .77 ± .03 | .84 ± .03 |
| All LFs - HTGEN LFs- HTUNSUP LFs (Domain Data) | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| All LFs - Antirules | .85 ± .01 | .79 ± .01 | .92 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs | .85 ± .01 | .79 ± .01 | .93 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTUNSUP LFs | .83 ± .02 | .76 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs - HTUNSUP LFs | .82 ± .02 | .76 ± .02 | .92 ± .02 | .81 ± .01 | .76 ± .03 | .84 ± .03 |
| Only HTGEN LFs | .76 ± .01 | .76 ± .02 | .76 ± .02 | .81 ± .02 | .79 ± .03 | .81 ± .03 |
| Only HTGEN LFs + Antirules | .86 ± .01 | .82 ± .01 | .92 ± .01 | .75 ± .04 | .82 ± .05 | .64 ± .05 |
| Only HTUNSUP LFs | .87 ± .01 | .88 ± .01 | .88 ± .01 | .81 ± .03 | .77 ± .04 | .81 ± .04 |
| Only HTUNSUP LFs + Antirules | .87 ± .01 | .92 ± .02 | .85 ± .02 | .76 ± .04 | .83 ± .05 | .66 ± .05 |
| Only HTGEN LFs + HTUNSUP LFs | .85 ± .01 | .85 ± .02 | .87 ± .02 | .82 ± .03 | .78 ± .03 | .81 ± .03 |
| Only HTGEN LFs + HTUNSUP LFs + Antirules | .86 ± .02 | .86 ± .01 | .86 ± .01 | .82 ± .03 | .79 ± .04 | .81 ± .04 |
| Model | F1 | Precision | Recall |
| Label-Indep NEAT Rules | .43 ± .03 | .88 ± .02 | .28 ± .02 |
| Label-Indep NEAT Dictionary | .78 ± .02 | .83 ± .02 | .74 ± .02 |
| Label-Indep NEAT Disambiguation | .79 ± .02 | .80 ± .02 | .77 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .67 ± .11 | .82 ± .14 | .57 ± .14 |
| CoNLL2003 | .97 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .95 ± .01 | .91 ± .01 | |
| WikiNER-en | .96 ± .01 | .97 ± .02 | .97 ± .02 | |
| WNUT2017 | .58 ± .04 | .56 ± .04 | .65 ± .04 | |
| RoBERTa-base | HTUNSUP | .74 ± .06 | .84 ± .07 | .60 ± .07 |
| CoNLL2003 | .98 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .94 ± .02 | .92 ± .02 | |
| WikiNER-en | .93 ± .03 | .94 ± .05 | .97 ± .05 | |
| WNUT2017 | .89 ± .02 | .87 ± .02 | .88 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .41 ± .11 | .50 ± .14 | .52 ± .14 |
| CoNLL2003 | .72 ± .04 | .71 ± .08 | .73 ± .08 | |
| Few-NERD-L1 | .66 ± .12 | .82 ± .13 | .41 ± .13 | |
| WikiNER-en | .65 ± .11 | .69 ± .13 | .52 ± .13 | |
| WNUT2017 | .62 ± .09 | .71 ± .15 | .47 ± .15 | |
| RoBERTa-base | HTUNSUP | .53 ± .07 | .56 ± .13 | .63 ± .13 |
| CoNLL2003 | .71 ± .05 | .67 ± .08 | .71 ± .08 | |
| Few-NERD-L1 | .61 ± .10 | .67 ± .10 | .45 ± .10 | |
| WikiNER-en | .65 ± .10 | .63 ± .09 | .57 ± .09 | |
| WNUT2017 | .67 ± .09 | .81 ± .14 | .48 ± .14 |
| Dataset | Train Size | Test Size | Total Size | Batch Size |
| CoNLL2003 | 14041 | 3453 | 17494 | 128 |
| WNUT2017 | 3394 | 1287 | 4681 | 128 |
| FewNERD-L1 | 131767 | 18824 | 150591 | 32 |
| WikiNER-en | 129907 | 14435 | 144342 | 128 |
| BTC | - | 2002 | - | - |
| Tweebank | - | 1201 | - | - |
| HTNAME | - | 995 | 995 | - |
| HTUNSUP | 6160 | - | 6160 | 32 |
| HTGEN | 9424 | 818 | 10242 | 32 |
| F1 | Precision | Recall |
| 0.90 | 0.89 | 0.91 |
| Method | CoNLL2003 | WNUT2017 | BTC | Tweebank | ||||||||
| F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .64 ± .04 | .66 ± .04 | .55 ± .04 | .21 ± .07 | .14 ± .06 | .44 ± .06 | .15 ± .01 | .44 ± .01 | .09 ± .01 | .19 ± .04 | .24 ± .05 | .18 ± .05 |
| TwitterNER (Mishra and Diesner, 2016) | .68 ± .05 | .91 ± .05 | .55 ± .05 | .61 ± .09 | .84 ± .10 | .57 ± .10 | .32 ± .05 | .86 ± .04 | .26 ± .04 | .67 ± .04 | .87 ± .05 | .57 ± .05 |
| LUKE (Yamada et al., 2020) | .31 ± .11 | .89 ± .09 | .19 ± .09 | .55 ± .07 | .67 ± .05 | .44 ± .05 | .47 ± .01 | .92 ± .01 | .31 ± .01 | .37 ± .07 | .76 ± .06 | .33 ± .06 |
| ELMo (Peters et al., 2018) | .96 ± .02 | .95 ± .02 | .99 ± .02 | .59 ± .15 | .72 ± .18 | .37 ± .18 | .41 ± .06 | .83 ± .04 | .35 ± .18 | .71 ± .01 | .74 ± .01 | .68 ± .01 |
| Flair (Akbik et al., 2019) | .98 ± .02 | .97 ± .02 | 1.0 ± .02 | .60 ± .15 | .79 ± .18 | .34 ± .18 | .35 ± .06 | .86 ± .05 | .31 ± .05 | .63 ± .07 | .76 ± .09 | .62 ± .09 |
| NEAT (Original) (Li et al., 2022) | .17 ± .07 | .43 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| NEAT (Weakly Supervised) | .16 ± .07 | .42 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| Majority vote | .83 ± .06 | .74 ± .02 | .98 ± .02 | .65 ± .04 | .53 ± .06 | .90 ± .06 | .64 ± .03 | .76 ± .03 | .60 ± .03 | .53 ± .06 | .37 ± .04 | .88 ± .04 |
| SWEET - Domain Data | .86 ± .05 | .79 ± .02 | .97 ± .02 | .69 ± .06 | .61 ± .06 | .82 ± .06 | .63 ± .03 | .86 ± .04 | .56 ± .04 | .58 ± .06 | .43 ± .04 | .85 ± .04 |
| SWEET | .86 ± .05 | .79 ± .03 | .98 ± .03 | .68 ± .04 | .58 ± .07 | .83 ± .07 | .64 ± .03 | .84 ± .04 | .57 ± .04 | .58 ± .06 | .43 ± .04 | .86 ± .04 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .27 ± .03 | .18 ± .02 | .51± .02 | .47 ± .04 | .50 ± .03 | .43 ± .03 |
| TwitterNER (Mishra and Diesner, 2016) | .56 ± .04 | 75. ± .04 | 52.± .04 | .70 ± .02 | .70 ± .03 | .67 ± .03 |
| LUKE (Yamada et al., 2020) | .63 ± .03 | .85 ± .04 | .51± .04 | .68 ± .03 | .84 ± .02 | .56 ± .02 |
| ELMo (Peters et al., 2018) | .51 ± .02 | .56 ± .02 | .46± .02 | .69 ± .05 | .61 ± .06 | .74 ± .06 |
| Flair (Akbik et al., 2019) | .45 ± .02 | .73 ± .02 | .32± .02 | .63 ± .04 | .83 ± .04 | .49 ± .04 |
| NEAT (Original) (Li et al., 2022) | .78 ± .04 | .83 ± .05 | .74± .03 | .71 ± .01 | .63 ± .02 | .79 ± .03 |
| NEAT (Weakly Supervised) | .79 ± .02 | .80 ± .02 | .77± .02 | .71 ± .01 | .64 ± .02 | .78 ± .02 |
| Majority vote | .73 ± .02 | .59 ± .01 | .95 ± .01 | .74 ± .02 | .65 ± .03 | .85 ± .03 |
| SWEET -Domain Data | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| SWEET | .87 ± .01 | .83 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| Model | Fine-tuning Dataset | F1 | Precision | Recall |
| DeBERTa-v3-base | HTUNSUP | .67 ± .03 | .71 ± .02 | .62 ± .02 |
| HTGEN | .68 ± .01 | .71 ± .02 | .67 ± .02 | |
| CoNLL2003 | .67 ± .02 | .67 ± .03 | .69 ± .03 | |
| Few-NERD-L1 | .57 ± .03 | .80 ± .03 | .43 ± .03 | |
| WikiNER-en | .52 ± .01 | .48 ± .02 | .54 ± .02 | |
| WNUT2017 | .70 ± .02 | .71 ± .02 | .72 ± .02 | |
| RoBERTa-base | HTUNSUP | .82 ± .02 | .84 ± .03 | .83 ± .03 |
| HTGEN | .72 ± .02 | .81 ± .02 | .66 ± .02 | |
| CoNLL2003 | .72 ± .02 | .68 ± .03 | .77 ± .03 | |
| Few-NERD-L1 | .68 ± .02 | .81 ± .02 | .59 ± .02 | |
| WikiNER-en | .49 ± .03 | .43 ± .03 | .56 ± .03 | |
| WNUT2017 | .68 ± .02 | .73 ± .03 | .66 ± .03 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| All LFs - HTGEN LFs | .88 ± .01 | .85 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - HTUNSUP LFs | .86 ± .02 | .82 ± .02 | .91 ± .02 | .82 ± .02 | .77 ± .03 | .84 ± .03 |
| All LFs - HTGEN LFs- HTUNSUP LFs (Domain Data) | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| All LFs - Antirules | .85 ± .01 | .79 ± .01 | .92 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs | .85 ± .01 | .79 ± .01 | .93 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTUNSUP LFs | .83 ± .02 | .76 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs - HTUNSUP LFs | .82 ± .02 | .76 ± .02 | .92 ± .02 | .81 ± .01 | .76 ± .03 | .84 ± .03 |
| Only HTGEN LFs | .76 ± .01 | .76 ± .02 | .76 ± .02 | .81 ± .02 | .79 ± .03 | .81 ± .03 |
| Only HTGEN LFs + Antirules | .86 ± .01 | .82 ± .01 | .92 ± .01 | .75 ± .04 | .82 ± .05 | .64 ± .05 |
| Only HTUNSUP LFs | .87 ± .01 | .88 ± .01 | .88 ± .01 | .81 ± .03 | .77 ± .04 | .81 ± .04 |
| Only HTUNSUP LFs + Antirules | .87 ± .01 | .92 ± .02 | .85 ± .02 | .76 ± .04 | .83 ± .05 | .66 ± .05 |
| Only HTGEN LFs + HTUNSUP LFs | .85 ± .01 | .85 ± .02 | .87 ± .02 | .82 ± .03 | .78 ± .03 | .81 ± .03 |
| Only HTGEN LFs + HTUNSUP LFs + Antirules | .86 ± .02 | .86 ± .01 | .86 ± .01 | .82 ± .03 | .79 ± .04 | .81 ± .04 |
| Model | F1 | Precision | Recall |
| Label-Indep NEAT Rules | .43 ± .03 | .88 ± .02 | .28 ± .02 |
| Label-Indep NEAT Dictionary | .78 ± .02 | .83 ± .02 | .74 ± .02 |
| Label-Indep NEAT Disambiguation | .79 ± .02 | .80 ± .02 | .77 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .67 ± .11 | .82 ± .14 | .57 ± .14 |
| CoNLL2003 | .97 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .95 ± .01 | .91 ± .01 | |
| WikiNER-en | .96 ± .01 | .97 ± .02 | .97 ± .02 | |
| WNUT2017 | .58 ± .04 | .56 ± .04 | .65 ± .04 | |
| RoBERTa-base | HTUNSUP | .74 ± .06 | .84 ± .07 | .60 ± .07 |
| CoNLL2003 | .98 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .94 ± .02 | .92 ± .02 | |
| WikiNER-en | .93 ± .03 | .94 ± .05 | .97 ± .05 | |
| WNUT2017 | .89 ± .02 | .87 ± .02 | .88 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .41 ± .11 | .50 ± .14 | .52 ± .14 |
| CoNLL2003 | .72 ± .04 | .71 ± .08 | .73 ± .08 | |
| Few-NERD-L1 | .66 ± .12 | .82 ± .13 | .41 ± .13 | |
| WikiNER-en | .65 ± .11 | .69 ± .13 | .52 ± .13 | |
| WNUT2017 | .62 ± .09 | .71 ± .15 | .47 ± .15 | |
| RoBERTa-base | HTUNSUP | .53 ± .07 | .56 ± .13 | .63 ± .13 |
| CoNLL2003 | .71 ± .05 | .67 ± .08 | .71 ± .08 | |
| Few-NERD-L1 | .61 ± .10 | .67 ± .10 | .45 ± .10 | |
| WikiNER-en | .65 ± .10 | .63 ± .09 | .57 ± .09 | |
| WNUT2017 | .67 ± .09 | .81 ± .14 | .48 ± .14 |
| Dataset | Train Size | Test Size | Total Size | Batch Size |
| CoNLL2003 | 14041 | 3453 | 17494 | 128 |
| WNUT2017 | 3394 | 1287 | 4681 | 128 |
| FewNERD-L1 | 131767 | 18824 | 150591 | 32 |
| WikiNER-en | 129907 | 14435 | 144342 | 128 |
| BTC | - | 2002 | - | - |
| Tweebank | - | 1201 | - | - |
| HTNAME | - | 995 | 995 | - |
| HTUNSUP | 6160 | - | 6160 | 32 |
| HTGEN | 9424 | 818 | 10242 | 32 |
| F1 | Precision | Recall |
| 0.90 | 0.89 | 0.91 |
| Method | CoNLL2003 | WNUT2017 | BTC | Tweebank | ||||||||
| F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .64 ± .04 | .66 ± .04 | .55 ± .04 | .21 ± .07 | .14 ± .06 | .44 ± .06 | .15 ± .01 | .44 ± .01 | .09 ± .01 | .19 ± .04 | .24 ± .05 | .18 ± .05 |
| TwitterNER (Mishra and Diesner, 2016) | .68 ± .05 | .91 ± .05 | .55 ± .05 | .61 ± .09 | .84 ± .10 | .57 ± .10 | .32 ± .05 | .86 ± .04 | .26 ± .04 | .67 ± .04 | .87 ± .05 | .57 ± .05 |
| LUKE (Yamada et al., 2020) | .31 ± .11 | .89 ± .09 | .19 ± .09 | .55 ± .07 | .67 ± .05 | .44 ± .05 | .47 ± .01 | .92 ± .01 | .31 ± .01 | .37 ± .07 | .76 ± .06 | .33 ± .06 |
| ELMo (Peters et al., 2018) | .96 ± .02 | .95 ± .02 | .99 ± .02 | .59 ± .15 | .72 ± .18 | .37 ± .18 | .41 ± .06 | .83 ± .04 | .35 ± .18 | .71 ± .01 | .74 ± .01 | .68 ± .01 |
| Flair (Akbik et al., 2019) | .98 ± .02 | .97 ± .02 | 1.0 ± .02 | .60 ± .15 | .79 ± .18 | .34 ± .18 | .35 ± .06 | .86 ± .05 | .31 ± .05 | .63 ± .07 | .76 ± .09 | .62 ± .09 |
| NEAT (Original) (Li et al., 2022) | .17 ± .07 | .43 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| NEAT (Weakly Supervised) | .16 ± .07 | .42 ± .05 | .07 ± .05 | .22 ± .06 | .47 ± .04 | .16 ± .04 | .11 ± .02 | .46 ± .01 | .07 ± .01 | .24 ± .05 | .42 ± .04 | .17 ± .04 |
| Majority vote | .83 ± .06 | .74 ± .02 | .98 ± .02 | .65 ± .04 | .53 ± .06 | .90 ± .06 | .64 ± .03 | .76 ± .03 | .60 ± .03 | .53 ± .06 | .37 ± .04 | .88 ± .04 |
| SWEET - Domain Data | .86 ± .05 | .79 ± .02 | .97 ± .02 | .69 ± .06 | .61 ± .06 | .82 ± .06 | .63 ± .03 | .86 ± .04 | .56 ± .04 | .58 ± .06 | .43 ± .04 | .85 ± .04 |
| SWEET | .86 ± .05 | .79 ± .03 | .98 ± .03 | .68 ± .04 | .58 ± .07 | .83 ± .07 | .64 ± .03 | .84 ± .04 | .57 ± .04 | .58 ± .06 | .43 ± .04 | .86 ± .04 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| spaCy (Honnibal et al., 2020) | .27 ± .03 | .18 ± .02 | .51± .02 | .47 ± .04 | .50 ± .03 | .43 ± .03 |
| TwitterNER (Mishra and Diesner, 2016) | .56 ± .04 | 75. ± .04 | 52.± .04 | .70 ± .02 | .70 ± .03 | .67 ± .03 |
| LUKE (Yamada et al., 2020) | .63 ± .03 | .85 ± .04 | .51± .04 | .68 ± .03 | .84 ± .02 | .56 ± .02 |
| ELMo (Peters et al., 2018) | .51 ± .02 | .56 ± .02 | .46± .02 | .69 ± .05 | .61 ± .06 | .74 ± .06 |
| Flair (Akbik et al., 2019) | .45 ± .02 | .73 ± .02 | .32± .02 | .63 ± .04 | .83 ± .04 | .49 ± .04 |
| NEAT (Original) (Li et al., 2022) | .78 ± .04 | .83 ± .05 | .74± .03 | .71 ± .01 | .63 ± .02 | .79 ± .03 |
| NEAT (Weakly Supervised) | .79 ± .02 | .80 ± .02 | .77± .02 | .71 ± .01 | .64 ± .02 | .78 ± .02 |
| Majority vote | .73 ± .02 | .59 ± .01 | .95 ± .01 | .74 ± .02 | .65 ± .03 | .85 ± .03 |
| SWEET -Domain Data | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| SWEET | .87 ± .01 | .83 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| Model | Fine-tuning Dataset | F1 | Precision | Recall |
| DeBERTa-v3-base | HTUNSUP | .67 ± .03 | .71 ± .02 | .62 ± .02 |
| HTGEN | .68 ± .01 | .71 ± .02 | .67 ± .02 | |
| CoNLL2003 | .67 ± .02 | .67 ± .03 | .69 ± .03 | |
| Few-NERD-L1 | .57 ± .03 | .80 ± .03 | .43 ± .03 | |
| WikiNER-en | .52 ± .01 | .48 ± .02 | .54 ± .02 | |
| WNUT2017 | .70 ± .02 | .71 ± .02 | .72 ± .02 | |
| RoBERTa-base | HTUNSUP | .82 ± .02 | .84 ± .03 | .83 ± .03 |
| HTGEN | .72 ± .02 | .81 ± .02 | .66 ± .02 | |
| CoNLL2003 | .72 ± .02 | .68 ± .03 | .77 ± .03 | |
| Few-NERD-L1 | .68 ± .02 | .81 ± .02 | .59 ± .02 | |
| WikiNER-en | .49 ± .03 | .43 ± .03 | .56 ± .03 | |
| WNUT2017 | .68 ± .02 | .73 ± .03 | .66 ± .03 |
| Method | HTNAME | HTGEN | ||||
| F1 | Prec | Rec | F1 | Prec | Rec | |
| All LFs - HTGEN LFs | .88 ± .01 | .85 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - HTUNSUP LFs | .86 ± .02 | .82 ± .02 | .91 ± .02 | .82 ± .02 | .77 ± .03 | .84 ± .03 |
| All LFs - HTGEN LFs- HTUNSUP LFs (Domain Data) | .88 ± .01 | .85 ± .01 | .92 ± .01 | .75 ± .02 | .71 ± .03 | .78 ± .03 |
| All LFs - Antirules | .85 ± .01 | .79 ± .01 | .92 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs | .85 ± .01 | .79 ± .01 | .93 ± .01 | .81 ± .02 | .75 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTUNSUP LFs | .83 ± .02 | .76 ± .01 | .92 ± .01 | .81 ± .02 | .76 ± .03 | .84 ± .03 |
| All LFs - Antirules - HTGEN LFs - HTUNSUP LFs | .82 ± .02 | .76 ± .02 | .92 ± .02 | .81 ± .01 | .76 ± .03 | .84 ± .03 |
| Only HTGEN LFs | .76 ± .01 | .76 ± .02 | .76 ± .02 | .81 ± .02 | .79 ± .03 | .81 ± .03 |
| Only HTGEN LFs + Antirules | .86 ± .01 | .82 ± .01 | .92 ± .01 | .75 ± .04 | .82 ± .05 | .64 ± .05 |
| Only HTUNSUP LFs | .87 ± .01 | .88 ± .01 | .88 ± .01 | .81 ± .03 | .77 ± .04 | .81 ± .04 |
| Only HTUNSUP LFs + Antirules | .87 ± .01 | .92 ± .02 | .85 ± .02 | .76 ± .04 | .83 ± .05 | .66 ± .05 |
| Only HTGEN LFs + HTUNSUP LFs | .85 ± .01 | .85 ± .02 | .87 ± .02 | .82 ± .03 | .78 ± .03 | .81 ± .03 |
| Only HTGEN LFs + HTUNSUP LFs + Antirules | .86 ± .02 | .86 ± .01 | .86 ± .01 | .82 ± .03 | .79 ± .04 | .81 ± .04 |
| Model | F1 | Precision | Recall |
| Label-Indep NEAT Rules | .43 ± .03 | .88 ± .02 | .28 ± .02 |
| Label-Indep NEAT Dictionary | .78 ± .02 | .83 ± .02 | .74 ± .02 |
| Label-Indep NEAT Disambiguation | .79 ± .02 | .80 ± .02 | .77 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .67 ± .11 | .82 ± .14 | .57 ± .14 |
| CoNLL2003 | .97 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .95 ± .01 | .91 ± .01 | |
| WikiNER-en | .96 ± .01 | .97 ± .02 | .97 ± .02 | |
| WNUT2017 | .58 ± .04 | .56 ± .04 | .65 ± .04 | |
| RoBERTa-base | HTUNSUP | .74 ± .06 | .84 ± .07 | .60 ± .07 |
| CoNLL2003 | .98 ± .02 | .98 ± .02 | .99 ± .02 | |
| Few-NERD-L1 | .92 ± .01 | .94 ± .02 | .92 ± .02 | |
| WikiNER-en | .93 ± .03 | .94 ± .05 | .97 ± .05 | |
| WNUT2017 | .89 ± .02 | .87 ± .02 | .88 ± .02 |
| Model | Dataset | F1 | Precision | Recall |
| DeBERTaV3-base | HTUNSUP | .41 ± .11 | .50 ± .14 | .52 ± .14 |
| CoNLL2003 | .72 ± .04 | .71 ± .08 | .73 ± .08 | |
| Few-NERD-L1 | .66 ± .12 | .82 ± .13 | .41 ± .13 | |
| WikiNER-en | .65 ± .11 | .69 ± .13 | .52 ± .13 | |
| WNUT2017 | .62 ± .09 | .71 ± .15 | .47 ± .15 | |
| RoBERTa-base | HTUNSUP | .53 ± .07 | .56 ± .13 | .63 ± .13 |
| CoNLL2003 | .71 ± .05 | .67 ± .08 | .71 ± .08 | |
| Few-NERD-L1 | .61 ± .10 | .67 ± .10 | .45 ± .10 | |
| WikiNER-en | .65 ± .10 | .63 ± .09 | .57 ± .09 | |
| WNUT2017 | .67 ± .09 | .81 ± .14 | .48 ± .14 |
| Dataset | Source | # Symptoms | Size |
| SYMPTOMIFY (ours) | VAERS | 11,472 | 871,373 |
| COMETA (Basaldella et al., 2020) | 7,645 | 20,000 | |
| MedRed (Scepanovic et al., 2020) | 18 | 2,000 | |
| RedMed (Laverty and Altman, 2019) | 2,978 | n/a | |
| Twitter TwiMed (Alvaro et al., 2017) | 3,144 | 2,000 | |
| CADEC (Karimi et al., 2015) | AskaPatient | 6,754 | 1,253 |
| Twitter ADR (Nikfarjam et al., 2015) | 1,280 | 1,784 |
| Vaccine | Symptom | Explanation Fragment |
| FLU4 | Dizziness postural | This is evident from the following fragment of the report: "I became extremely dizzy when I tried to stand. |
| COVID19 | Lymphadenopathy | This is evident from the statement "My lymph nodes were swollen." |
| FLUA3 | Throat tightness | The following fragments from the report indicate the presence of the symptom: - "throat feeling like its about to close" - "swelling and lump in throat on right side" |
| HEPA | Thirst | The following fragments from the report support this explanation: "He then began drinking large amounts of water throughout the day and even waking at night to ask for a drink multiple times." |
| FLUA4 | Arthralgia | The report states "Shoulder pain" which is a common symptom of arthralgia. |
| FLUR4 | Dyspnoea | The symptom "Dyspnoea" is mentioned in the patient report because the patient is experiencing difficulty breathing, which is a common symptom of dyspnoea. |
| Symptom | LLM Definition | UMLS Definition |
| Melaena | Blood in the stool, typically caused by gastrointestinal bleeding. | The black, tarry, foul-smelling feces that contain degraded blood. |
| Ischaemia | Lack of blood flow to a tissue or organ. This may be due to obstruction or a problem with the blood vessels. | A decrease in blood supply caused by blockage of blood vessel. |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 |
| MetaMap | - | - | 3.74 | 10.65 |
| Symptom Embedding | 0.72 | 3.32 | 0.87 | 5.87 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| BERT-base ‡ | 1.34 | 15.81 | 1.45 | 18.80 |
| BART-base ‡ | 1.37 | 15.96 | 1.55 | 18.98 |
| ClinicalBERT ‡ | 1.53 | 16.32 | 1.76 | 19.46 |
| ClinicalBERT + CNN ‡ | 7.29 | 62.68 | 8.12 | 65.79 |
| GEN w/ BART | 31.15 | 78.90 | 31.22 | 81.27 |
| GEN w/ T5-base | 32.01 | 79.88 | 32.19 | 82.25 |
| Joint C + GEN | 32.89 | 79.93 | 32.70 | 82.97 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Model | Rare symptoms test set | |
| Macro F1 | Micro F1 | |
| String match † | 6.91 | 7.79 |
| GEN (T5-base) | 21.78 | 24.98 |
| GEN w/ definitions | 21.72 | 25.16 |
| GEN w/ synthetic reports | 23.31 | 25.73 |
| GPT-2 Prompting | 6.63 | 7.87 |
| ChatGPT Prompting (zero-shot) | 11.75 | 8.56 |
| ChatGPT Prompting (one-shot) | 16.06 | 10.49 |
| (C1) Symptom | (C2) Report | (C3) Zero-shot Falcon Explanation | (C4) Falcon with SYMPTOMIFY Explanation |
| (R1) Arthralgia | Patient having shoulder pain. Made appoint-ment with doctor. MRI completed. Possible SIRVA. | Arthralgia is mentioned in the patient report as it is a common symptom associated with SIRVA (Shoulder Impingement Syndrome with Radial Neuropathy) | The symptom "Arthralgia" is mentioned in the patient report because the patient reported shoulder pain. This is evident from the state-ment "Patient having shoulder pain". |
| (R2) Pyrexia | all over body rash and fever | Pyrexia is mentioned in the patient report as it is a common symptom of fever | The symptom "Pyrexia" is mentioned in the pa-tient report because the patient is experiencing a fever. |
| (R3) Muscle spasms | Patient stated he was getting muscle spasms in the neck and back after the flu shot. | Muscle spasms are a common side effect of the flu shot. The patient may have received the flu vaccine and experienced muscle spasms as a result. | The symptom "Muscle spasms" is mentioned in the patient report because the patient is ex-persiencing muscle spasms in the neck and back after receiving a flu shot. |
| (R4) Mood altered | Fever, cold chills, restless sleep, redness at site of injection, right leg soreness-unable to stand on that leg, loss of appetite, moodiness/crying. After 12 hours of injection | The patient reported mood altered as a result of the injection. This is mentioned in the patient report as it is a potential side effect of the med-ication. | The symptom "Mood altered" is mentioned in the patient report because the patient reported experiencing a change in their mood. This is evident from the statement "I was moody/cry-ing." |
| Yes | 1 |
| No | 2 |
| COVID-19 | FLU4 | HPV9 | TDAP | RV5 | |
| 1 | Headache | Pain | Syncope | Injection site pain | Diarrhoea |
| 2 | Pyrexia | Injection site pain | Dizziness | Pain | Haematochezia |
| 3 | Fatigue | Pain in extremity | Headache | Pain in extremity | Vomiting |
| 4 | Chills | Injection site erythema | Injection site pain | Injection site swelling | Intussusception |
| 5 | Pain | Pyrexia | Nausea | Injection site erythema | Pyrexia |
| Metamap | |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Multi-label Classification | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| batch size | 32 |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Generative Models | |
| Learning rate (BART-base) | 3e-5 {2e-5, 3e-5, 5e-5} |
| Learning rate (T5-base) | 1e-4 {3e-5, 5e-5, 1e-4, 2e-4} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 {64, 128, 256} |
| beam search size | 4 |
| batch size | 64 {32, 64} |
| Joint Learning (C + GEN) | |
| Learning rate | 3e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 |
| batch size | 64 {32, 64} |
| λj | 0.1 {0.05, 0.1, 0.2, 0.5} |
| LM Prompt (GPT-2) | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 512 |
| batch size | 64 {32, 64} |
| Symptom | |
| 1 | Headache |
| 2 | Pyrexia |
| 3 | Fatigue |
| 4 | Chills |
| 5 | Pain |
| 6 | Nausea |
| 7 | Pain in extremity |
| 8 | Dizziness |
| 9 | Injection site pain |
| 10 | Myalgia |
| step | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 0.12 | 0.02 | 0.02 | 0.04 | 0.07 | 0.07 | 0.02 | 0.02 | 0.08 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.05 | 0.06 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.05 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.0; 1/2 | - |
| Model | Full test | UMLS-mapped | |||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | ||
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 | |
| MetaMap | - | - | 3.74 | 10.65 | |
| Symptom Embeddings | 0.72 | 3.32 | 0.87 | 5.87 | |
| Pretrain-Finetune | BERT-base ‡ | 1.34 (±0.13) | 15.81 (±0.14) | 1.45 (±0.09) | 18.80 (±0.08) |
| BART-base ‡ | 1.37 (±0.18) | 15.96 (±0.17) | 1.55 (±0.12) | 18.98 (±0.13) | |
| ClinicalBERT ‡ | 1.53 (±0.12) | 16.32 (±0.12) | 1.76 (±0.11) | 19.46 (±0.11) | |
| ClinicalBERT + CNN ‡ | 7.29 (±0.35) | 62.68 (±0.25) | 8.12 (±0.47) | 65.79 (±0.25) | |
| GEN w/ BART † | 31.15 (±0.37) | 78.90 (±0.24) | 31.22 (±0.19) | 81.27 (±0.25) | |
| GEN w/ T5-base † | 32.01 (±0.31) | 79.88 (±0.27) | 32.19 (±0.22) | 82.25 (±0.25) | |
| Joint C + GEN | 32.89 (±0.21) | 79.93 (±0.19) | 32.70 (±0.43) | 82.97 (±0.31) | |
| Prompt | GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 | |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Type | Example |
| no [symptom name] | I feel cold, but do not have a fever, am not having chills, nausea, etc.Fever of 101, 102 for 2 Days; Chills; Nausea with no Vomiting.no visible signs of rash or irritation. |
| Medical history(not current symptom) | Medical history included Ankylosing, spondylitis/psoriasis, cardiac ablation from Jan2017Family medical history included her dad had rheumatic fever and high blood pressure.Past medical history included no adverse event. concurrent medical conditionsincluded Ulcer and Hypothyroidism. |
| Other negation expressions | Patient didn’t experience fever and rash, but feel muscle painAll but the pain and soreness at the injection site subsided by the next dayDenies any further fevers or chills. Denies any nausea vomiting. Denies loss stasis smell. |
| Dataset | Source | # Symptoms | Size |
| SYMPTOMIFY (ours) | VAERS | 11,472 | 871,373 |
| COMETA (Basaldella et al., 2020) | 7,645 | 20,000 | |
| MedRed (Scepanovic et al., 2020) | 18 | 2,000 | |
| RedMed (Laverty and Altman, 2019) | 2,978 | n/a | |
| Twitter TwiMed (Alvaro et al., 2017) | 3,144 | 2,000 | |
| CADEC (Karimi et al., 2015) | AskaPatient | 6,754 | 1,253 |
| Twitter ADR (Nikfarjam et al., 2015) | 1,280 | 1,784 |
| Vaccine | Symptom | Explanation Fragment |
| FLU4 | Dizziness postural | This is evident from the following fragment of the report: "I became extremely dizzy when I tried to stand. |
| COVID19 | Lymphadenopathy | This is evident from the statement "My lymph nodes were swollen." |
| FLUA3 | Throat tightness | The following fragments from the report indicate the presence of the symptom: - "throat feeling like its about to close" - "swelling and lump in throat on right side" |
| HEPA | Thirst | The following fragments from the report support this explanation: "He then began drinking large amounts of water throughout the day and even waking at night to ask for a drink multiple times." |
| FLUA4 | Arthralgia | The report states "Shoulder pain" which is a common symptom of arthralgia. |
| FLUR4 | Dyspnoea | The symptom "Dyspnoea" is mentioned in the patient report because the patient is experiencing difficulty breathing, which is a common symptom of dyspnoea. |
| Symptom | LLM Definition | UMLS Definition |
| Melaena | Blood in the stool, typically caused by gastrointestinal bleeding. | The black, tarry, foul-smelling feces that contain degraded blood. |
| Ischaemia | Lack of blood flow to a tissue or organ. This may be due to obstruction or a problem with the blood vessels. | A decrease in blood supply caused by blockage of blood vessel. |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 |
| MetaMap | - | - | 3.74 | 10.65 |
| Symptom Embedding | 0.72 | 3.32 | 0.87 | 5.87 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| BERT-base ‡ | 1.34 | 15.81 | 1.45 | 18.80 |
| BART-base ‡ | 1.37 | 15.96 | 1.55 | 18.98 |
| ClinicalBERT ‡ | 1.53 | 16.32 | 1.76 | 19.46 |
| ClinicalBERT + CNN ‡ | 7.29 | 62.68 | 8.12 | 65.79 |
| GEN w/ BART | 31.15 | 78.90 | 31.22 | 81.27 |
| GEN w/ T5-base | 32.01 | 79.88 | 32.19 | 82.25 |
| Joint C + GEN | 32.89 | 79.93 | 32.70 | 82.97 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Model | Rare symptoms test set | |
| Macro F1 | Micro F1 | |
| String match † | 6.91 | 7.79 |
| GEN (T5-base) | 21.78 | 24.98 |
| GEN w/ definitions | 21.72 | 25.16 |
| GEN w/ synthetic reports | 23.31 | 25.73 |
| GPT-2 Prompting | 6.63 | 7.87 |
| ChatGPT Prompting (zero-shot) | 11.75 | 8.56 |
| ChatGPT Prompting (one-shot) | 16.06 | 10.49 |
| (C1) Symptom | (C2) Report | (C3) Zero-shot Falcon Explanation | (C4) Falcon with SYMPTOMIFY Explanation |
| (R1) Arthralgia | Patient having shoulder pain. Made appoint-ment with doctor. MRI completed. Possible SIRVA. | Arthralgia is mentioned in the patient report as it is a common symptom associated with SIRVA (Shoulder Impingement Syndrome with Radial Neuropathy) | The symptom "Arthralgia" is mentioned in the patient report because the patient reported shoulder pain. This is evident from the state-ment "Patient having shoulder pain". |
| (R2) Pyrexia | all over body rash and fever | Pyrexia is mentioned in the patient report as it is a common symptom of fever | The symptom "Pyrexia" is mentioned in the pa-tient report because the patient is experiencing a fever. |
| (R3) Muscle spasms | Patient stated he was getting muscle spasms in the neck and back after the flu shot. | Muscle spasms are a common side effect of the flu shot. The patient may have received the flu vaccine and experienced muscle spasms as a result. | The symptom "Muscle spasms" is mentioned in the patient report because the patient is ex-persiencing muscle spasms in the neck and back after receiving a flu shot. |
| (R4) Mood altered | Fever, cold chills, restless sleep, redness at site of injection, right leg soreness-unable to stand on that leg, loss of appetite, moodiness/crying. After 12 hours of injection | The patient reported mood altered as a result of the injection. This is mentioned in the patient report as it is a potential side effect of the med-ication. | The symptom "Mood altered" is mentioned in the patient report because the patient reported experiencing a change in their mood. This is evident from the statement "I was moody/cry-ing." |
| Yes | 1 |
| No | 2 |
| COVID-19 | FLU4 | HPV9 | TDAP | RV5 | |
| 1 | Headache | Pain | Syncope | Injection site pain | Diarrhoea |
| 2 | Pyrexia | Injection site pain | Dizziness | Pain | Haematochezia |
| 3 | Fatigue | Pain in extremity | Headache | Pain in extremity | Vomiting |
| 4 | Chills | Injection site erythema | Injection site pain | Injection site swelling | Intussusception |
| 5 | Pain | Pyrexia | Nausea | Injection site erythema | Pyrexia |
| Metamap | |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Multi-label Classification | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| batch size | 32 |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Generative Models | |
| Learning rate (BART-base) | 3e-5 {2e-5, 3e-5, 5e-5} |
| Learning rate (T5-base) | 1e-4 {3e-5, 5e-5, 1e-4, 2e-4} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 {64, 128, 256} |
| beam search size | 4 |
| batch size | 64 {32, 64} |
| Joint Learning (C + GEN) | |
| Learning rate | 3e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 |
| batch size | 64 {32, 64} |
| λj | 0.1 {0.05, 0.1, 0.2, 0.5} |
| LM Prompt (GPT-2) | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 512 |
| batch size | 64 {32, 64} |
| Symptom | |
| 1 | Headache |
| 2 | Pyrexia |
| 3 | Fatigue |
| 4 | Chills |
| 5 | Pain |
| 6 | Nausea |
| 7 | Pain in extremity |
| 8 | Dizziness |
| 9 | Injection site pain |
| 10 | Myalgia |
| step | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 0.12 | 0.02 | 0.02 | 0.04 | 0.07 | 0.07 | 0.02 | 0.02 | 0.08 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.05 | 0.06 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.05 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.0; 1/2 | - |
| Model | Full test | UMLS-mapped | |||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | ||
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 | |
| MetaMap | - | - | 3.74 | 10.65 | |
| Symptom Embeddings | 0.72 | 3.32 | 0.87 | 5.87 | |
| Pretrain-Finetune | BERT-base ‡ | 1.34 (±0.13) | 15.81 (±0.14) | 1.45 (±0.09) | 18.80 (±0.08) |
| BART-base ‡ | 1.37 (±0.18) | 15.96 (±0.17) | 1.55 (±0.12) | 18.98 (±0.13) | |
| ClinicalBERT ‡ | 1.53 (±0.12) | 16.32 (±0.12) | 1.76 (±0.11) | 19.46 (±0.11) | |
| ClinicalBERT + CNN ‡ | 7.29 (±0.35) | 62.68 (±0.25) | 8.12 (±0.47) | 65.79 (±0.25) | |
| GEN w/ BART † | 31.15 (±0.37) | 78.90 (±0.24) | 31.22 (±0.19) | 81.27 (±0.25) | |
| GEN w/ T5-base † | 32.01 (±0.31) | 79.88 (±0.27) | 32.19 (±0.22) | 82.25 (±0.25) | |
| Joint C + GEN | 32.89 (±0.21) | 79.93 (±0.19) | 32.70 (±0.43) | 82.97 (±0.31) | |
| Prompt | GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 | |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Type | Example |
| no [symptom name] | I feel cold, but do not have a fever, am not having chills, nausea, etc.Fever of 101, 102 for 2 Days; Chills; Nausea with no Vomiting.no visible signs of rash or irritation. |
| Medical history(not current symptom) | Medical history included Ankylosing, spondylitis/psoriasis, cardiac ablation from Jan2017Family medical history included her dad had rheumatic fever and high blood pressure.Past medical history included no adverse event. concurrent medical conditionsincluded Ulcer and Hypothyroidism. |
| Other negation expressions | Patient didn’t experience fever and rash, but feel muscle painAll but the pain and soreness at the injection site subsided by the next dayDenies any further fevers or chills. Denies any nausea vomiting. Denies loss stasis smell. |
| Dataset | Source | # Symptoms | Size |
| SYMPTOMIFY (ours) | VAERS | 11,472 | 871,373 |
| COMETA (Basaldella et al., 2020) | 7,645 | 20,000 | |
| MedRed (Scepanovic et al., 2020) | 18 | 2,000 | |
| RedMed (Laverty and Altman, 2019) | 2,978 | n/a | |
| Twitter TwiMed (Alvaro et al., 2017) | 3,144 | 2,000 | |
| CADEC (Karimi et al., 2015) | AskaPatient | 6,754 | 1,253 |
| Twitter ADR (Nikfarjam et al., 2015) | 1,280 | 1,784 |
| Vaccine | Symptom | Explanation Fragment |
| FLU4 | Dizziness postural | This is evident from the following fragment of the report: "I became extremely dizzy when I tried to stand. |
| COVID19 | Lymphadenopathy | This is evident from the statement "My lymph nodes were swollen." |
| FLUA3 | Throat tightness | The following fragments from the report indicate the presence of the symptom: - "throat feeling like its about to close" - "swelling and lump in throat on right side" |
| HEPA | Thirst | The following fragments from the report support this explanation: "He then began drinking large amounts of water throughout the day and even waking at night to ask for a drink multiple times." |
| FLUA4 | Arthralgia | The report states "Shoulder pain" which is a common symptom of arthralgia. |
| FLUR4 | Dyspnoea | The symptom "Dyspnoea" is mentioned in the patient report because the patient is experiencing difficulty breathing, which is a common symptom of dyspnoea. |
| Symptom | LLM Definition | UMLS Definition |
| Melaena | Blood in the stool, typically caused by gastrointestinal bleeding. | The black, tarry, foul-smelling feces that contain degraded blood. |
| Ischaemia | Lack of blood flow to a tissue or organ. This may be due to obstruction or a problem with the blood vessels. | A decrease in blood supply caused by blockage of blood vessel. |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 |
| MetaMap | - | - | 3.74 | 10.65 |
| Symptom Embedding | 0.72 | 3.32 | 0.87 | 5.87 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| BERT-base ‡ | 1.34 | 15.81 | 1.45 | 18.80 |
| BART-base ‡ | 1.37 | 15.96 | 1.55 | 18.98 |
| ClinicalBERT ‡ | 1.53 | 16.32 | 1.76 | 19.46 |
| ClinicalBERT + CNN ‡ | 7.29 | 62.68 | 8.12 | 65.79 |
| GEN w/ BART | 31.15 | 78.90 | 31.22 | 81.27 |
| GEN w/ T5-base | 32.01 | 79.88 | 32.19 | 82.25 |
| Joint C + GEN | 32.89 | 79.93 | 32.70 | 82.97 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Model | Rare symptoms test set | |
| Macro F1 | Micro F1 | |
| String match † | 6.91 | 7.79 |
| GEN (T5-base) | 21.78 | 24.98 |
| GEN w/ definitions | 21.72 | 25.16 |
| GEN w/ synthetic reports | 23.31 | 25.73 |
| GPT-2 Prompting | 6.63 | 7.87 |
| ChatGPT Prompting (zero-shot) | 11.75 | 8.56 |
| ChatGPT Prompting (one-shot) | 16.06 | 10.49 |
| (C1) Symptom | (C2) Report | (C3) Zero-shot Falcon Explanation | (C4) Falcon with SYMPTOMIFY Explanation |
| (R1) Arthralgia | Patient having shoulder pain. Made appoint-ment with doctor. MRI completed. Possible SIRVA. | Arthralgia is mentioned in the patient report as it is a common symptom associated with SIRVA (Shoulder Impingement Syndrome with Radial Neuropathy) | The symptom "Arthralgia" is mentioned in the patient report because the patient reported shoulder pain. This is evident from the state-ment "Patient having shoulder pain". |
| (R2) Pyrexia | all over body rash and fever | Pyrexia is mentioned in the patient report as it is a common symptom of fever | The symptom "Pyrexia" is mentioned in the pa-tient report because the patient is experiencing a fever. |
| (R3) Muscle spasms | Patient stated he was getting muscle spasms in the neck and back after the flu shot. | Muscle spasms are a common side effect of the flu shot. The patient may have received the flu vaccine and experienced muscle spasms as a result. | The symptom "Muscle spasms" is mentioned in the patient report because the patient is ex-persiencing muscle spasms in the neck and back after receiving a flu shot. |
| (R4) Mood altered | Fever, cold chills, restless sleep, redness at site of injection, right leg soreness-unable to stand on that leg, loss of appetite, moodiness/crying. After 12 hours of injection | The patient reported mood altered as a result of the injection. This is mentioned in the patient report as it is a potential side effect of the med-ication. | The symptom "Mood altered" is mentioned in the patient report because the patient reported experiencing a change in their mood. This is evident from the statement "I was moody/cry-ing." |
| Yes | 1 |
| No | 2 |
| COVID-19 | FLU4 | HPV9 | TDAP | RV5 | |
| 1 | Headache | Pain | Syncope | Injection site pain | Diarrhoea |
| 2 | Pyrexia | Injection site pain | Dizziness | Pain | Haematochezia |
| 3 | Fatigue | Pain in extremity | Headache | Pain in extremity | Vomiting |
| 4 | Chills | Injection site erythema | Injection site pain | Injection site swelling | Intussusception |
| 5 | Pain | Pyrexia | Nausea | Injection site erythema | Pyrexia |
| Metamap | |
| threshold | 0.1 +{0.05, 0.1, 0.15, 0.2, 0.25} |
| Multi-label Classification | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| batch size | 32 |
| threshold | 0.1 +{0.05, 0.1, 0.15, 0.2, 0.25} |
| Generative Models | |
| Learning rate (BART-base) | 3e-5 {2e-5, 3e-5, 5e-5} |
| Learning rate (T5-base) | 1e-4 {3e-5, 5e-5, 1e-4, 2e-4} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 {64, 128, 256} |
| beam search size | 4 |
| batch size | 64 {32, 64} |
| Joint Learning (C + GEN) | |
| Learning rate | 3e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 |
| batch size | 64 {32, 64} |
| λj | 0.1 {0.05, 0.1, 0.2, 0.5} |
| LM Prompt (GPT-2) | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 512 |
| batch size | 64 {32, 64} |
| Symptom | |
| 1 | Headache |
| 2 | Pyrexia |
| 3 | Fatigue |
| 4 | Chills |
| 5 | Pain |
| 6 | Nausea |
| 7 | Pain in extremity |
| 8 | Dizziness |
| 9 | Injection site pain |
| 10 | Myalgia |
| step | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 0.12 | 0.02 | 0.02 | 0.04 | 0.07 | 0.07 | 0.02 | 0.02 | 0.08 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.05 | 0.06 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.05 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.0; 1/2 | - | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Model | Full test | UMLS-mapped | |||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | ||
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 | |
| MetaMap | - | - | 3.74 | 10.65 | |
| Symptom Embeddings | 0.72 | 3.32 | 0.87 | 5.87 | |
| Pretrain-Finetune | BERT-base ‡ | 1.34 (±0.13) | 15.81 (±0.14) | 1.45 (±0.09) | 18.80 (±0.08) |
| BART-base ‡ | 1.37 (±0.18) | 15.96 (±0.17) | 1.55 (±0.12) | 18.98 (±0.13) | |
| ClinicalBERT ‡ | 1.53 (±0.12) | 16.32 (±0.12) | 1.76 (±0.11) | 19.46 (±0.11) | |
| ClinicalBERT + CNN ‡ | 7.29 (±0.35) | 62.68 (±0.25) | 8.12 (±0.47) | 65.79 (±0.25) | |
| GEN w/ BART † | 31.15 (±0.37) | 78.90 (±0.24) | 31.22 (±0.19) | 81.27 (±0.25) | |
| GEN w/ T5-base † | 32.01 (±0.31) | 79.88 (±0.27) | 32.19 (±0.22) | 82.25 (±0.25) | |
| Joint C + GEN | 32.89 (±0.21) | 79.93 (±0.19) | 32.70 (±0.43) | 82.97 (±0.31) | |
| Prompt | GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 | |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 | |
| Type | Example |
| no [symptom name] | I feel cold, but do not have a fever, am not having chills, nausea, etc.Fever of 101, 102 for 2 Days; Chills; Nausea with no Vomiting.no visible signs of rash or irritation. |
| Medical history(not current symptom) | Medical history included Ankylosing, spondylitis/psoriasis, cardiac ablation from Jan2017Family medical history included her dad had rheumatic fever and high blood pressure.Past medical history included no adverse event. concurrent medical conditionsincluded Ulcer and Hypothyroidism. |
| Other negation expressions | Patient didn’t experience fever and rash, but feel muscle painAll but the pain and soreness at the injection site subsided by the next dayDenies any further fevers or chills. Denies any nausea vomiting. Denies loss stasis smell. |
| Dataset | Source | # Symptoms | Size |
| SYMPTOMIFY (ours) | VAERS | 11,472 | 871,373 |
| COMETA (Basaldella et al., 2020) | 7,645 | 20,000 | |
| MedRed (Scepanovic et al., 2020) | 18 | 2,000 | |
| RedMed (Laverty and Altman, 2019) | 2,978 | n/a | |
| Twitter TwiMed (Alvaro et al., 2017) | 3,144 | 2,000 | |
| CADEC (Karimi et al., 2015) | AskaPatient | 6,754 | 1,253 |
| Twitter ADR (Nikfarjam et al., 2015) | 1,280 | 1,784 |
| Vaccine | Symptom | Explanation Fragment |
| FLU4 | Dizziness postural | This is evident from the following fragment of the report: "I became extremely dizzy when I tried to stand. |
| COVID19 | Lymphadenopathy | This is evident from the statement "My lymph nodes were swollen." |
| FLUA3 | Throat tightness | The following fragments from the report indicate the presence of the symptom: - "throat feeling like its about to close" - "swelling and lump in throat on right side" |
| HEPA | Thirst | The following fragments from the report support this explanation: "He then began drinking large amounts of water throughout the day and even waking at night to ask for a drink multiple times." |
| FLUA4 | Arthralgia | The report states "Shoulder pain" which is a common symptom of arthralgia. |
| FLUR4 | Dyspnoea | The symptom "Dyspnoea" is mentioned in the patient report because the patient is experiencing difficulty breathing, which is a common symptom of dyspnoea. |
| Symptom | LLM Definition | UMLS Definition |
| Melaena | Blood in the stool, typically caused by gastrointestinal bleeding. | The black, tarry, foul-smelling feces that contain degraded blood. |
| Ischaemia | Lack of blood flow to a tissue or organ. This may be due to obstruction or a problem with the blood vessels. | A decrease in blood supply caused by blockage of blood vessel. |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 |
| MetaMap | - | - | 3.74 | 10.65 |
| Symptom Embedding | 0.72 | 3.32 | 0.87 | 5.87 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| BERT-base ‡ | 1.34 | 15.81 | 1.45 | 18.80 |
| BART-base ‡ | 1.37 | 15.96 | 1.55 | 18.98 |
| ClinicalBERT ‡ | 1.53 | 16.32 | 1.76 | 19.46 |
| ClinicalBERT + CNN ‡ | 7.29 | 62.68 | 8.12 | 65.79 |
| GEN w/ BART | 31.15 | 78.90 | 31.22 | 81.27 |
| GEN w/ T5-base | 32.01 | 79.88 | 32.19 | 82.25 |
| Joint C + GEN | 32.89 | 79.93 | 32.70 | 82.97 |
| Model | Full test | UMLS-mapped | ||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | |
| GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Model | Rare symptoms test set | |
| Macro F1 | Micro F1 | |
| String match † | 6.91 | 7.79 |
| GEN (T5-base) | 21.78 | 24.98 |
| GEN w/ definitions | 21.72 | 25.16 |
| GEN w/ synthetic reports | 23.31 | 25.73 |
| GPT-2 Prompting | 6.63 | 7.87 |
| ChatGPT Prompting (zero-shot) | 11.75 | 8.56 |
| ChatGPT Prompting (one-shot) | 16.06 | 10.49 |
| (C1) Symptom | (C2) Report | (C3) Zero-shot Falcon Explanation | (C4) Falcon with SYMPTOMIFY Explanation |
| (R1) Arthralgia | Patient having shoulder pain. Made appoint-ment with doctor. MRI completed. Possible SIRVA. | Arthralgia is mentioned in the patient report as it is a common symptom associated with SIRVA (Shoulder Impingement Syndrome with Radial Neuropathy) | The symptom "Arthralgia" is mentioned in the patient report because the patient reported shoulder pain. This is evident from the state-ment "Patient having shoulder pain". |
| (R2) Pyrexia | all over body rash and fever | Pyrexia is mentioned in the patient report as it is a common symptom of fever | The symptom "Pyrexia" is mentioned in the pa-tient report because the patient is experiencing a fever. |
| (R3) Muscle spasms | Patient stated he was getting muscle spasms in the neck and back after the flu shot. | Muscle spasms are a common side effect of the flu shot. The patient may have received the flu vaccine and experienced muscle spasms as a result. | The symptom "Muscle spasms" is mentioned in the patient report because the patient is ex-persiencing muscle spasms in the neck and back after receiving a flu shot. |
| (R4) Mood altered | Fever, cold chills, restless sleep, redness at site of injection, right leg soreness-unable to stand on that leg, loss of appetite, moodiness/crying. After 12 hours of injection | The patient reported mood altered as a result of the injection. This is mentioned in the patient report as it is a potential side effect of the med-ication. | The symptom "Mood altered" is mentioned in the patient report because the patient reported experiencing a change in their mood. This is evident from the statement "I was moody/cry-ing." |
| Yes | 1 |
| No | 2 |
| COVID-19 | FLU4 | HPV9 | TDAP | RV5 | |
| 1 | Headache | Pain | Syncope | Injection site pain | Diarrhoea |
| 2 | Pyrexia | Injection site pain | Dizziness | Pain | Haematochezia |
| 3 | Fatigue | Pain in extremity | Headache | Pain in extremity | Vomiting |
| 4 | Chills | Injection site erythema | Injection site pain | Injection site swelling | Intussusception |
| 5 | Pain | Pyrexia | Nausea | Injection site erythema | Pyrexia |
| Metamap | |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Multi-label Classification | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| batch size | 32 |
| threshold | 0.1\n{0.05, 0.1, 0.15, 0.2, 0.25} |
| Generative Models | |
| Learning rate (BART-base) | 3e-5 {2e-5, 3e-5, 5e-5} |
| Learning rate (T5-base) | 1e-4 {3e-5, 5e-5, 1e-4, 2e-4} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 {64, 128, 256} |
| beam search size | 4 |
| batch size | 64 {32, 64} |
| Joint Learning (C + GEN) | |
| Learning rate | 3e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 256 {128, 256} |
| max output length (decoder) | 128 |
| batch size | 64 {32, 64} |
| λj | 0.1 {0.05, 0.1, 0.2, 0.5} |
| LM Prompt (GPT-2) | |
| Learning rate | 5e-5 {2e-5, 3e-5, 5e-5} |
| max input length | 512 |
| batch size | 64 {32, 64} |
| Symptom | |
| 1 | Headache |
| 2 | Pyrexia |
| 3 | Fatigue |
| 4 | Chills |
| 5 | Pain |
| 6 | Nausea |
| 7 | Pain in extremity |
| 8 | Dizziness |
| 9 | Injection site pain |
| 10 | Myalgia |
| step | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 0.12 | 0.02 | 0.02 | 0.04 | 0.07 | 0.07 | 0.02 | 0.02 | 0.08 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.05 | 0.06 | 0.29 | 0.01 | 0.00 | 0.02 | 0.00 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.05 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.02 | 0.01 | 0.03 | 0.02 | 0.02 | 0.0; 1/2 | - |
| Model | Full test | UMLS-mapped | |||
| Macro F1 | Micro F1 | Macro F1 | Micro F1 | ||
| Exact String Matching | 8.87 | 30.38 | 10.41 | 32.98 | |
| MetaMap | - | - | 3.74 | 10.65 | |
| Symptom Embeddings | 0.72 | 3.32 | 0.87 | 5.87 | |
| Pretrain-Finetune | BERT-base ‡ | 1.34 (±0.13) | 15.81 (±0.14) | 1.45 (±0.09) | 18.80 (±0.08) |
| BART-base ‡ | 1.37 (±0.18) | 15.96 (±0.17) | 1.55 (±0.12) | 18.98 (±0.13) | |
| ClinicalBERT ‡ | 1.53 (±0.12) | 16.32 (±0.12) | 1.76 (±0.11) | 19.46 (±0.11) | |
| ClinicalBERT + CNN ‡ | 7.29 (±0.35) | 62.68 (±0.25) | 8.12 (±0.47) | 65.79 (±0.25) | |
| GEN w/ BART † | 31.15 (±0.37) | 78.90 (±0.24) | 31.22 (±0.19) | 81.27 (±0.25) | |
| GEN w/ T5-base † | 32.01 (±0.31) | 79.88 (±0.27) | 32.19 (±0.22) | 82.25 (±0.25) | |
| Joint C + GEN | 32.89 (±0.21) | 79.93 (±0.19) | 32.70 (±0.43) | 82.97 (±0.31) | |
| Prompt | GPT-2 (Supervised) | 13.93 | 59.41 | 14.27 | 61.90 |
| ChatGPT (Zero-shot) | 24.69 | 39.75 | 22.23 | 38.10 | |
| ChatGPT (One-shot) | 29.17 | 42.74 | 27.97 | 41.51 |
| Type | Example |
| no [symptom name] | I feel cold, but do not have a fever, am not having chills, nausea, etc.Fever of 101, 102 for 2 Days; Chills; Nausea with no Vomiting.no visible signs of rash or irritation. |
| Medical history(not current symptom) | Medical history included Ankylosing, spondylitis/psoriasis, cardiac ablation from Jan2017Family medical history included her dad had rheumatic fever and high blood pressure.Past medical history included no adverse event. concurrent medical conditionsincluded Ulcer and Hypothyroidism. |
| Other negation expressions | Patient didn’t experience fever and rash, but feel muscle painAll but the pain and soreness at the injection site subsided by the next dayDenies any further fevers or chills. Denies any nausea vomiting. Denies loss stasis smell. |
| Agent | Flu e ↑ | Flu i ↑ | Inf e ↑ | Rec ↑ |
| Ada-RG | 1.41 | 0.00 | 0.16 | 0.16 |
| Rule-AD | 4.82 | 0.59 | 0.86 | 0.40 |
| Key-QG | 4.97 | 0.72 | 0.71 | 0.36 |
| No-ReGen | 4.99 | 0.85 | 0.74 | 0.47 |
| SalesBot | 4.99 | 0.91 | 0.74 | 0.44 |
| Statistics (avg) | # | |
| Nb. words of Seller utterance | 62.5 | 35.3 |
| Nb. words of Shopper utterance | 20.8 | 20.7 |
| Nb. of turns | 11.9 | 12.9 |
| Nb. turns before the first rec. | 6.0 | 7.9 |
| Nb. of recommendations | 2.6 | 2.4 |
| Nb. of triggered revelations | 1.7 | 2.8 |
| % Correct recommendations (Rec) | 44 | 54 |
| Information Quiz Score (Infq) | 32.9 | 31.8 |
| Fluency Score (Fluε) | 4.4 | 4.2 |
| % Is Human (Flui) | 55 | 80 |
| name | Samsung 55" Class S95B OLED 4K Smart Tizen TV |
| price | $1,699.99 |
| description | Samsung OLED TV changes the game again with 8.3 million self lit pixels and ultra powerful 4K AI Neural Processing, all for a picture so real, it's surreal. Add on Dolby Atmos® sound built in, the latest Smart TV apps, and a LaserSlim design and get a viewing experience that's intensely cinematic. |
| features | ["55 inch", "OLED Technology", "Neural Quantum Processor with 4K Upscaling", "Smart Calibration", "Connectivity with Bluetooth, RF, Wi-Fi, USB, HDMI, Ethernet (LAN), Digital Audio Out x 1 (Optical)", "Supported internet services: Netflix, Google TV, Amazon Instant Video, YouTube, Browser"] |
| Product Category | Question | Answer |
| TV | The resolution of a TV refers to the number of vertical pixels or dots that make up the TV's picture. | T |
| A 4K TV with an OLED display technology and a refresh rate of 120Hz will generally have a better picture quality than a 1080p TV with an LCD display technology and a refresh rate of 60Hz. | T | |
| TVs with bigger screens always have better sound quality. | F | |
| Vacuum Cleaners | A canister vacuum is best suited for a large home with multiple floors and different types of flooring. | T |
| If you have pets, it's important to choose a vacuum cleaner with a HEPA filter to capture pet hair and dander. | T | |
| Stick vacuums are not suitable for cleaning homes with stairs. | F | |
| Mattress | Memory foam mattresses mold to the shape of your body. | T |
| Gel memory foam mattresses are ideal for hot sleepers who want to stay cool at night. | T | |
| A mattress with more coils is always better than a mattress with fewer coils. | F | |
| Guitar | The acoustic guitar is the most common guitar for first-time buyers. | T |
| Acoustic guitars are self-amplified instruments that come in a variety of sizes, and the most common shape is the Parlor. | F | |
| Acoustic-electric guitars are the easiest way to perform with an acoustic guitar, and they always come with an integrated pickup system. | F | |
| Laptop | A laptop with a larger screen size is always better for gaming. | F |
| A laptop with a dedicated graphics card is unnecessary for basic office tasks. | T | |
| An SSD provides faster storage than an HDD. | T | |
| Coffee Makers | Drip coffee makers are ideal for making coffee for one person. | F |
| Single-serve coffee makers are convenient because they require no measurement of coffee grinds or water. | T | |
| Espresso machines can only make espresso shots. | F |
| Agent | Flu e ↑ | Flu i ↑ | Inf e ↑ | Rec ↑ |
| Ada-RG | 1.41 | 0.00 | 0.16 | 0.16 |
| Rule-AD | 4.82 | 0.59 | 0.86 | 0.40 |
| Key-QG | 4.97 | 0.72 | 0.71 | 0.36 |
| No-ReGen | 4.99 | 0.85 | 0.74 | 0.47 |
| SalesBot | 4.99 | 0.91 | 0.74 | 0.44 |
| Statistics (avg) | # | |
| Nb. words of Seller utterance | 62.5 | 35.3 |
| Nb. words of Shopper utterance | 20.8 | 20.7 |
| Nb. of turns | 11.9 | 12.9 |
| Nb. turns before the first rec. | 6.0 | 7.9 |
| Nb. of recommendations | 2.6 | 2.4 |
| Nb. of triggered revelations | 1.7 | 2.8 |
| % Correct recommendations (Rec) | 44 | 54 |
| Information Quiz Score (Infq) | 32.9 | 31.8 |
| Fluency Score (Fluε) | 4.4 | 4.2 |
| % Is Human (Flui) | 55 | 80 |
| name | Samsung 55" Class S95B OLED 4K Smart Tizen TV |
| price | $1,699.99 |
| description | Samsung OLED TV changes the game again with 8.3 million self lit pixels and ultra powerful 4K AI Neural Processing, all for a picture so real, it's surreal. Add on Dolby Atmos® sound built in, the latest Smart TV apps, and a LaserSlim design and get a viewing experience that's intensely cinematic. |
| features | ["55 inch", "OLED Technology", "Neural Quantum Processor with 4K Upscaling", "Smart Calibration", "Connectivity with Bluetooth, RF, Wi-Fi, USB, HDMI, Ethernet (LAN), Digital Audio Out x 1 (Optical)", "Supported internet services: Netflix, Google TV, Amazon Instant Video, YouTube, Browser"] |
| Product Category | Question | Answer |
| TV | The resolution of a TV refers to the number of vertical pixels or dots that make up the TV's picture. | T |
| A 4K TV with an OLED display technology and a refresh rate of 120Hz will generally have a better picture quality than a 1080p TV with an LCD display technology and a refresh rate of 60Hz. | T | |
| TVs with bigger screens always have better sound quality. | F | |
| Vacuum Cleaners | A canister vacuum is best suited for a large home with multiple floors and different types of flooring. | T |
| If you have pets, it's important to choose a vacuum cleaner with a HEPA filter to capture pet hair and dander. | T | |
| Stick vacuums are not suitable for cleaning homes with stairs. | F | |
| Mattress | Memory foam mattresses mold to the shape of your body. | T |
| Gel memory foam mattresses are ideal for hot sleepers who want to stay cool at night. | T | |
| A mattress with more coils is always better than a mattress with fewer coils. | F | |
| Guitar | The acoustic guitar is the most common guitar for first-time buyers. | T |
| Acoustic guitars are self-amplified instruments that come in a variety of sizes, and the most common shape is the Parlor. | F | |
| Acoustic-electric guitars are the easiest way to perform with an acoustic guitar, and they always come with an integrated pickup system. | F | |
| Laptop | A laptop with a larger screen size is always better for gaming. | F |
| A laptop with a dedicated graphics card is unnecessary for basic office tasks. | T | |
| An SSD provides faster storage than an HDD. | T | |
| Coffee Makers | Drip coffee makers are ideal for making coffee for one person. | F |
| Single-serve coffee makers are convenient because they require no measurement of coffee grinds or water. | T | |
| Espresso machines can only make espresso shots. | F |
| Agent | Flu e ↑ | Flu i ↑ | Inf e ↑ | Rec ↑ |
| Ada-RG | 1.41 | 0.00 | 0.16 | 0.16 |
| Rule-AD | 4.82 | 0.59 | 0.86 | 0.40 |
| Key-QG | 4.97 | 0.72 | 0.71 | 0.36 |
| No-ReGen | 4.99 | 0.85 | 0.74 | 0.47 |
| SalesBot | 4.99 | 0.91 | 0.74 | 0.44 |
| Statistics (avg) | # | |
| Nb. words of Seller utterance | 62.5 | 35.3 |
| Nb. words of Shopper utterance | 20.8 | 20.7 |
| Nb. of turns | 11.9 | 12.9 |
| Nb. turns before the first rec. | 6.0 | 7.9 |
| Nb. of recommendations | 2.6 | 2.4 |
| Nb. of triggered revelations | 1.7 | 2.8 |
| % Correct recommendations (Rec) | 44 | 54 |
| Information Quiz Score (Infq) | 32.9 | 31.8 |
| Fluency Score (Fluε) | 4.4 | 4.2 |
| % Is Human (Flui) | 55 | 80 |
| name | Samsung 55" Class S95B OLED 4K Smart Tizen TV |
| price | $1,699.99 |
| description | Samsung OLED TV changes the game again with 8.3 million self lit pixels and ultra powerful 4K AI Neural Processing, all for a picture so real, it's surreal. Add on Dolby Atmos® sound built in, the latest Smart TV apps, and a LaserSlim design and get a viewing experience that's intensely cinematic. |
| features | ["55 inch", "OLED Technology", "Neural Quantum Processor with 4K Upscaling", "Smart Calibration", "Connectivity with Bluetooth, RF, Wi-Fi, USB, HDMI, Ethernet (LAN), Digital Audio Out x 1 (Optical)", "Supported internet services: Netflix, Google TV, Amazon Instant Video, YouTube, Browser"] |
| Product Category | Question | Answer |
| TV | The resolution of a TV refers to the number of vertical pixels or dots that make up the TV's picture. | T |
| A 4K TV with an OLED display technology and a refresh rate of 120Hz will generally have a better picture quality than a 1080p TV with an LCD display technology and a refresh rate of 60Hz. | T | |
| TVs with bigger screens always have better sound quality. | F | |
| Vacuum Cleaners | A canister vacuum is best suited for a large home with multiple floors and different types of flooring. | T |
| If you have pets, it's important to choose a vacuum cleaner with a HEPA filter to capture pet hair and dander. | T | |
| Stick vacuums are not suitable for cleaning homes with stairs. | F | |
| Mattress | Memory foam mattresses mold to the shape of your body. | T |
| Gel memory foam mattresses are ideal for hot sleepers who want to stay cool at night. | T | |
| A mattress with more coils is always better than a mattress with fewer coils. | F | |
| Guitar | The acoustic guitar is the most common guitar for first-time buyers. | T |
| Acoustic guitars are self-amplified instruments that come in a variety of sizes, and the most common shape is the Parlor. | F | |
| Acoustic-electric guitars are the easiest way to perform with an acoustic guitar, and they always come with an integrated pickup system. | F | |
| Laptop | A laptop with a larger screen size is always better for gaming. | F |
| A laptop with a dedicated graphics card is unnecessary for basic office tasks. | T | |
| An SSD provides faster storage than an HDD. | T | |
| Coffee Makers | Drip coffee makers are ideal for making coffee for one person. | F |
| Single-serve coffee makers are convenient because they require no measurement of coffee grinds or water. | T | |
| Espresso machines can only make espresso shots. | F |
| Agent | Flu e ↑ | Flu i ↑ | Inf e ↑ | Rec ↑ |
| Ada-RG | 1.41 | 0.00 | 0.16 | 0.16 |
| Rule-AD | 4.82 | 0.59 | 0.86 | 0.40 |
| Key-QG | 4.97 | 0.72 | 0.71 | 0.36 |
| No-ReGen | 4.99 | 0.85 | 0.74 | 0.47 |
| SalesBot | 4.99 | 0.91 | 0.74 | 0.44 |
| Statistics (avg) | # | |
| Nb. words of Seller utterance | 62.5 | 35.3 |
| Nb. words of Shopper utterance | 20.8 | 20.7 |
| Nb. of turns | 11.9 | 12.9 |
| Nb. turns before the first rec. | 6.0 | 7.9 |
| Nb. of recommendations | 2.6 | 2.4 |
| Nb. of triggered revelations | 1.7 | 2.8 |
| % Correct recommendations (Rec) | 44 | 54 |
| Information Quiz Score (Infq) | 32.9 | 31.8 |
| Fluency Score (Fluε) | 4.4 | 4.2 |
| % Is Human (Flui) | 55 | 80 |
| name | Samsung 55" Class S95B OLED 4K Smart Tizen TV |
| price | $1,699.99 |
| description | Samsung OLED TV changes the game again with 8.3 million self lit pixels and ultra powerful 4K AI Neural Processing, all for a picture so real, it's surreal. Add on Dolby Atmos® sound built in, the latest Smart TV apps, and a LaserSlim design and get a viewing experience that's intensely cinematic. |
| features | ["55 inch", "OLED Technology", "Neural Quantum Processor with 4K Upscaling", "Smart Calibration", "Connectivity with Bluetooth, RF, Wi-Fi, USB, HDMI, Ethernet (LAN), Digital Audio Out x 1 (Optical)", "Supported internet services: Netflix, Google TV, Amazon Instant Video, YouTube, Browser"] |
| Product Category | Question | Answer |
| TV | The resolution of a TV refers to the number of vertical pixels or dots that make up the TV's picture. | T |
| A 4K TV with an OLED display technology and a refresh rate of 120Hz will generally have a better picture quality than a 1080p TV with an LCD display technology and a refresh rate of 60Hz. | T | |
| TVs with bigger screens always have better sound quality. | F | |
| Vacuum Cleaners | A canister vacuum is best suited for a large home with multiple floors and different types of flooring. | T |
| If you have pets, it's important to choose a vacuum cleaner with a HEPA filter to capture pet hair and dander. | T | |
| Stick vacuums are not suitable for cleaning homes with stairs. | F | |
| Mattress | Memory foam mattresses mold to the shape of your body. | T |
| Gel memory foam mattresses are ideal for hot sleepers who want to stay cool at night. | T | |
| A mattress with more coils is always better than a mattress with fewer coils. | F | |
| Guitar | The acoustic guitar is the most common guitar for first-time buyers. | T |
| Acoustic guitars are self-amplified instruments that come in a variety of sizes, and the most common shape is the Parlor. | F | |
| Acoustic-electric guitars are the easiest way to perform with an acoustic guitar, and they always come with an integrated pickup system. | F | |
| Laptop | A laptop with a larger screen size is always better for gaming. | F |
| A laptop with a dedicated graphics card is unnecessary for basic office tasks. | T | |
| An SSD provides faster storage than an HDD. | T | |
| Coffee Makers | Drip coffee makers are ideal for making coffee for one person. | F |
| Single-serve coffee makers are convenient because they require no measurement of coffee grinds or water. | T | |
| Espresso machines can only make espresso shots. | F |
| Architecture | Params | Params w/o Embedding | MACs / Token | MACs / Token |
| Transformer-small | 48.32M | 31.54M | 48.7M | 67.2M |
| Transformer-base | 60.92M | 44.14M | 61.5M | 79.8M |
| Transformer-big | 209.92M | 176.37M | 210.9M | 247.5M |
| Transformer-base-3 | 38.85M | 22.07M | 39.0M | 48.2M |
| Transformer-base-9 | 82.99M | 66.21M | 83.7M | 111.2M |
| Transformer-base-12 | 105.06M | 88.28M | 105.9M | 142.6M |
| Subset | #Words | #Sents | #Docs |
| EUbookshop | 172M | 9.3M | 14k |
| OpenSubtitles | 136M | 23M | 29k |
| Europarl | 42M | 1.8M | 10k |
| NewsCommentary | 8.8M | 0.4M | 9.7k |
| TildeMODEL | 23M | 1.6M | 52k |
| Combined | 382M | 36M | 115k |
| Context Length | Accuracy |
| 0 | 50.06 |
| 60 | 67.75 |
| 120 | 81.89 |
| 250 | 82.26 |
| 500 | 82.38 |
| 750 | 82.49 |
| 1000 | 81.96 |
| Subset Number | #Lines | #Tokens | ||||
| 1 | 2 | 3 | 1 | 2 | 3 | |
| 1 | 0.28 | 0.29 | 0.27 | 4.29 | 4.27 | 4.08 |
| 2 | 0.52 | 0.49 | 0.48 | 8.13 | 8.05 | 7.62 |
| 3 | 0.82 | 0.86 | 0.85 | 13.12 | 13.92 | 12.85 |
| 4 | 1.39 | 1.42 | 1.40 | 21.77 | 22.28 | 21.48 |
| 5 | 2.37 | 2.41 | 2.30 | 37.70 | 39.21 | 36.30 |
| 6 | 4.00 | 4.06 | 3.99 | 62.64 | 64.93 | 62.83 |
| 7 | 6.81 | 6.88 | 6.87 | 109.17 | 111.99 | 109.54 |
| 8 | 11.47 | 11.39 | 11.51 | 183.84 | 181.83 | 182.84 |
| 9 | 19.53 | 19.66 | 19.58 | 312.61 | 313.62 | 310.31 |
| Module | Computational Cost | Number of the Module | Corresponding Operation |
| Embedding | n*512 | 4 | The embedding and positional encoding layers in both encoder and decoder |
| Output Projection | n*512*32768 | 1 | The output projection layer for generating the final output probabilities |
| Feed Forward | n*512*2048*2 | 12 | The feed forward layers in both encoder and decoder layers |
| Attention Project | n*512*64*8*4 | 18 | The query, key, value, and output projection in the multi-head attention |
| Dot Product | n2*64*8*2 | 18 | The dot product between key and query as well as attention weight and value |
| Total | n2*18432+n*60819456 | 1 | Total computation cost of the Transformer base |
| n | Computational Cost | Computational Cost / token |
| 32 | 1965096960 | 61409280 |
| 1024 | 81606475776 | 79693824 |
| Subset | #Words | #Sents | #Docs |
| UNPC | 417.2M | 17.4M | 94.7k |
| NewsCommentary | 5.3M | 0.1M | 7.8k |
| Combined | 422.5M | 17.5M | 102.5k |
| Architecture | Params | Params w/o Embedding | MACs / Token | MACs / Token |
| Transformer-small | 48.32M | 31.54M | 48.7M | 67.2M |
| Transformer-base | 60.92M | 44.14M | 61.5M | 79.8M |
| Transformer-big | 209.92M | 176.37M | 210.9M | 247.5M |
| Transformer-base-3 | 38.85M | 22.07M | 39.0M | 48.2M |
| Transformer-base-9 | 82.99M | 66.21M | 83.7M | 111.2M |
| Transformer-base-12 | 105.06M | 88.28M | 105.9M | 142.6M |
| Subset | #Words | #Sents | #Docs |
| EUbookshop | 172M | 9.3M | 14k |
| OpenSubtitles | 136M | 23M | 29k |
| Europarl | 42M | 1.8M | 10k |
| NewsCommentary | 8.8M | 0.4M | 9.7k |
| TildeMODEL | 23M | 1.6M | 52k |
| Combined | 382M | 36M | 115k |
| Context Length | Accuracy |
| 0 | 50.06 |
| 60 | 67.75 |
| 120 | 81.89 |
| 250 | 82.26 |
| 500 | 82.38 |
| 750 | 82.49 |
| 1000 | 81.96 |
| Subset Number | #Lines | #Tokens | ||||
| 1 | 2 | 3 | 1 | 2 | 3 | |
| 1 | 0.28 | 0.29 | 0.27 | 4.29 | 4.27 | 4.08 |
| 2 | 0.52 | 0.49 | 0.48 | 8.13 | 8.05 | 7.62 |
| 3 | 0.82 | 0.86 | 0.85 | 13.12 | 13.92 | 12.85 |
| 4 | 1.39 | 1.42 | 1.40 | 21.77 | 22.28 | 21.48 |
| 5 | 2.37 | 2.41 | 2.30 | 37.70 | 39.21 | 36.30 |
| 6 | 4.00 | 4.06 | 3.99 | 62.64 | 64.93 | 62.83 |
| 7 | 6.81 | 6.88 | 6.87 | 109.17 | 111.99 | 109.54 |
| 8 | 11.47 | 11.39 | 11.51 | 183.84 | 181.83 | 182.84 |
| 9 | 19.53 | 19.66 | 19.58 | 312.61 | 313.62 | 310.31 |
| Module | Computational Cost | Number of the Module | Corresponding Operation |
| Embedding | n*512 | 4 | The embedding and positional encoding layers in both encoder and decoder |
| Output Projection | n*512*32768 | 1 | The output projection layer for generating the final output probabilities |
| Feed Forward | n*512*2048*2 | 12 | The feed forward layers in both encoder and decoder layers |
| Attention Project | n*512*64*8*4 | 18 | The query, key, value, and output projection in the multi-head attention |
| Dot Product | n2*64*8*2 | 18 | The dot product between key and query as well as attention weight and value |
| Total | n2*18432+n*60819456 | 1 | Total computation cost of the Transformer base |
| n | Computational Cost | Computational Cost / token |
| 32 | 1965096960 | 61409280 |
| 1024 | 81606475776 | 79693824 |
| Subset | #Words | #Sents | #Docs |
| UNPC | 417.2M | 17.4M | 94.7k |
| NewsCommentary | 5.3M | 0.1M | 7.8k |
| Combined | 422.5M | 17.5M | 102.5k |
| Architecture | Params | Params w/o Embedding | MACs / Token | MACs / Token |
| Transformer-small | 48.32M | 31.54M | 48.7M | 67.2M |
| Transformer-base | 60.92M | 44.14M | 61.5M | 79.8M |
| Transformer-big | 209.92M | 176.37M | 210.9M | 247.5M |
| Transformer-base-3 | 38.85M | 22.07M | 39.0M | 48.2M |
| Transformer-base-9 | 82.99M | 66.21M | 83.7M | 111.2M |
| Transformer-base-12 | 105.06M | 88.28M | 105.9M | 142.6M |
| Subset | #Words | #Sents | #Docs |
| EUbookshop | 172M | 9.3M | 14k |
| OpenSubtitles | 136M | 23M | 29k |
| Europarl | 42M | 1.8M | 10k |
| NewsCommentary | 8.8M | 0.4M | 9.7k |
| TildeMODEL | 23M | 1.6M | 52k |
| Combined | 382M | 36M | 115k |
| Context Length | Accuracy |
| 0 | 50.06 |
| 60 | 67.75 |
| 120 | 81.89 |
| 250 | 82.26 |
| 500 | 82.38 |
| 750 | 82.49 |
| 1000 | 81.96 |
| Subset Number | #Lines | #Tokens | ||||
| 1 | 2 | 3 | 1 | 2 | 3 | |
| 1 | 0.28 | 0.29 | 0.27 | 4.29 | 4.27 | 4.08 |
| 2 | 0.52 | 0.49 | 0.48 | 8.13 | 8.05 | 7.62 |
| 3 | 0.82 | 0.86 | 0.85 | 13.12 | 13.92 | 12.85 |
| 4 | 1.39 | 1.42 | 1.40 | 21.77 | 22.28 | 21.48 |
| 5 | 2.37 | 2.41 | 2.30 | 37.70 | 39.21 | 36.30 |
| 6 | 4.00 | 4.06 | 3.99 | 62.64 | 64.93 | 62.83 |
| 7 | 6.81 | 6.88 | 6.87 | 109.17 | 111.99 | 109.54 |
| 8 | 11.47 | 11.39 | 11.51 | 183.84 | 181.83 | 182.84 |
| 9 | 19.53 | 19.66 | 19.58 | 312.61 | 313.62 | 310.31 |
| Module | Computational Cost | Number of the Module | Corresponding Operation |
| Embedding | n*512 | 4 | The embedding and positional encoding layers in both encoder and decoder |
| Output Projection | n*512*32768 | 1 | The output projection layer for generating the final output probabilities |
| Feed Forward | n*512*2048*2 | 12 | The feed forward layers in both encoder and decoder layers |
| Attention Project | n*512*64*8*4 | 18 | The query, key, value, and output projection in the multi-head attention |
| Dot Product | n2*64*8*2 | 18 | The dot product between key and query as well as attention weight and value |
| Total | n2*18432+n*60819456 | 1 | Total computation cost of the Transformer base |
| n | Computational Cost | Computational Cost / token |
| 32 | 1965096960 | 61409280 |
| 1024 | 81606475776 | 79693824 |
| Subset | #Words | #Sents | #Docs |
| UNPC | 417.2M | 17.4M | 94.7k |
| NewsCommentary | 5.3M | 0.1M | 7.8k |
| Combined | 422.5M | 17.5M | 102.5k |
| Architecture | Params | Params w/o Embedding | MACs / Token | MACs / Token |
| Transformer-small | 48.32M | 31.54M | 48.7M | 67.2M |
| Transformer-base | 60.92M | 44.14M | 61.5M | 79.8M |
| Transformer-big | 209.92M | 176.37M | 210.9M | 247.5M |
| Transformer-base-3 | 38.85M | 22.07M | 39.0M | 48.2M |
| Transformer-base-9 | 82.99M | 66.21M | 83.7M | 111.2M |
| Transformer-base-12 | 105.06M | 88.28M | 105.9M | 142.6M |
| Subset | #Words | #Sents | #Docs |
| EUbookshop | 172M | 9.3M | 14k |
| OpenSubtitles | 136M | 23M | 29k |
| Europarl | 42M | 1.8M | 10k |
| NewsCommentary | 8.8M | 0.4M | 9.7k |
| TildeMODEL | 23M | 1.6M | 52k |
| Combined | 382M | 36M | 115k |
| Context Length | Accuracy |
| 0 | 50.06 |
| 60 | 67.75 |
| 120 | 81.89 |
| 250 | 82.26 |
| 500 | 82.38 |
| 750 | 82.49 |
| 1000 | 81.96 |
| Subset Number | #Lines | #Tokens | ||||
| 1 | 2 | 3 | 1 | 2 | 3 | |
| 1 | 0.28 | 0.29 | 0.27 | 4.29 | 4.27 | 4.08 |
| 2 | 0.52 | 0.49 | 0.48 | 8.13 | 8.05 | 7.62 |
| 3 | 0.82 | 0.86 | 0.85 | 13.12 | 13.92 | 12.85 |
| 4 | 1.39 | 1.42 | 1.40 | 21.77 | 22.28 | 21.48 |
| 5 | 2.37 | 2.41 | 2.30 | 37.70 | 39.21 | 36.30 |
| 6 | 4.00 | 4.06 | 3.99 | 62.64 | 64.93 | 62.83 |
| 7 | 6.81 | 6.88 | 6.87 | 109.17 | 111.99 | 109.54 |
| 8 | 11.47 | 11.39 | 11.51 | 183.84 | 181.83 | 182.84 |
| 9 | 19.53 | 19.66 | 19.58 | 312.61 | 313.62 | 310.31 |
| Module | Computational Cost | Number of the Module | Corresponding Operation |
| Embedding | n*512 | 4 | The embedding and positional encoding layers in both encoder and decoder |
| Output Projection | n*512*32768 | 1 | The output projection layer for generating the final output probabilities |
| Feed Forward | n*512*2048*2 | 12 | The feed forward layers in both encoder and decoder layers |
| Attention Project | n*512*64*8*4 | 18 | The query, key, value, and output projection in the multi-head attention |
| Dot Product | n2*64*8*2 | 18 | The dot product between key and query as well as attention weight and value |
| Total | n2*18432+n*60819456 | 1 | Total computation cost of the Transformer base |
| n | Computational Cost | Computational Cost / token |
| 32 | 1965096960 | 61409280 |
| 1024 | 81606475776 | 79693824 |
| Subset | #Words | #Sents | #Docs |
| UNPC | 417.2M | 17.4M | 94.7k |
| NewsCommentary | 5.3M | 0.1M | 7.8k |
| Combined | 422.5M | 17.5M | 102.5k |
| Model | #Params | FLOPs | Speed | Neg Log Ppl | GLUE | SGLUE | SQuAD |
| Transformer Tiny | 16M | 1.21 | 38.4 | -2.47 | 69.3 | 56.9 | 73.6 |
| Transformer Small | 60M | 3.70 | 22.7 | -2.02 | 78.1 | 65.3 | 81.9 |
| Transformer Base | 223M | 11.4 | 9.3 | -1.75 | 83.8 | 74.0 | 86.3 |
| Transformer Large | 738M | 34.3 | 3.6 | -1.61 | 86.4 | 78.3 | 88.6 |
| Transformer XL | 2.9B | 63.8 | 1.3 | -1.49 | 87.8 | 81.5 | 89.5 |
| Evolved Transformer Tiny | 19M | 1.31 | 39.7 | -2.45 | 69.6 | 57.1 | 69.6 |
| Evolved Transformer Small | 79M | 4.23 | 23.7 | -2.04 | 75.7 | 66.2 | 80.2 |
| Evolved Transformer Base | 218M | 10.2 | 8.9 | -1.79 | 83.0 | 70.5 | 84.8 |
| Evolved Transformer Large | 1.0B | 49.3 | 2.1 | -1.62 | 86.2 | 77.1 | 88.0 |
| Evolved Transformer XL | 2.2B | 71.3 | 0.8 | -1.55 | 87.0 | 78.3 | 88.2 |
| Universal Transformer Tiny | 11M | 1.77 | 38.1 | -2.73 | 69.8 | 56.1 | 62.3 |
| Universal Transformer Small | 52M | 7.30 | 18.3 | -2.12 | 76.8 | 64.2 | 75.4 |
| Universal Transformer Base | 127M | 20.3 | 8.4 | -1.91 | 80.0 | 67.9 | 80.1 |
| Universal Transformer Large | 283M | 27.6 | 1.6 | -1.67 | 84.0 | 73.4 | 85.4 |
| Switch Transformer Tiny | 174M | 3.25 | 29.7 | -2.01 | 78.2 | 63.8 | 80.7 |
| Switch Transformer Small | 460M | 4.63 | 22.3 | -1.85 | 80.3 | 68.0 | 82.9 |
| Switch Transformer Base | 2.0B | 12.7 | 8.4 | -1.66 | 84.2 | 74.1 | 86.5 |
| Switch Transformer Large | 3.9B | 23.0 | 4.1 | -1.56 | 84.6 | 75.8 | 87.9 |
| Switch Transformer XL | 29.6B | 43.3 | 0.8 | -1.62 | 84.0 | 75.2 | 87.5 |
| Performer Tiny | 16M | 1.14 | 42.0 | -2.88 | 50.5 | 48.8 | 15.0 |
| Performer Small | 61M | 3.50 | 39.0 | -2.44 | 57.8 | 51.1 | 31.1 |
| Performer Base | 224M | 10.8 | 11.7 | -2.23 | 61.4 | 53.4 | 37.8 |
| Performer Large | 739M | 32.8 | 4.4 | -2.16 | 62.4 | 52.4 | 30.8 |
| Funnel Transformer Tiny | 16M | 1.10 | 39.9 | -2.58 | 63.4 | 49.4 | 54.6 |
| Funnel Transformer Small | 61M | 2.96 | 32.7 | -2.11 | 70.0 | 58.5 | 75.1 |
| Funnel Transformer Base | 223M | 8.10 | 11.9 | -1.83 | 76.3 | 62.9 | 81.6 |
| Funnel Transformer Large | 739M | 22.6 | 5.0 | -1.69 | 79.8 | 67.1 | 83.8 |
| Funnel Transformer XL | 2.9B | 40.3 | 1.89 | -1.61 | 79.8 | 68.0 | 83.7 |
| ALBERT Small | 15M | 3.57 | 42.0 | -2.36 | 73.7 | 62.0 | 77.1 |
| ALBERT Base | 21M | 9.40 | 16.4 | -2.28 | 69.0 | 57.2 | 64.3 |
| ALBERT Large | 34M | 31.6 | 5.1 | -2.20 | 62.9 | 54.1 | 27.3 |
| MoS-Transformer Tiny | 27M | 1.29 | 39.7 | -2.37 | 70.6 | 57.9 | 74.1 |
| MoS-Transformer Small | 81M | 3.70 | 26.3 | -1.98 | 79.7 | 67.1 | 83.1 |
| MoS-Transformer Base | 257M | 11.4 | 8.6 | -1.70 | 84.5 | 73.9 | 86.8 |
| MoS-Transformer Large | 800M | 35.0 | 3.4 | -1.56 | 86.5 | 79.7 | 89.1 |
| MoS-Transformer XL | 2.9B | 112 | 1.2 | -1.45 | 88.2 | 81.4 | 90.0 |
| GLU-Transformer Tiny | 26M | 1.29 | 31.7 | -2.35 | 70.5 | 57.0 | 74.2 |
| GLU-Transformer Small | 77M | 3.70 | 26.4 | -1.97 | 79.1 | 67.4 | 83.0 |
| GLU-Transformer Base | 248M | 11.4 | 8.6 | -1.71 | 84.6 | 74.5 | 87.2 |
| GLU-Transformer Large | 748M | 35.0 | 3.4 | -1.56 | 84.2 | 74.3 | 86.2 |
| GLU-Transformer XL | 2.85B | 61.3 | 1.0 | -1.49 | 87.6 | 82.9 | 89.4 |
| LConv Tiny | 17M | 1.20 | 31.2 | -2.50 | 51.1 | 51.3 | 49.5 |
| LConv Small | 67M | 3.80 | 12.8 | -2.10 | 71.8 | 59.9 | 64.7 |
| LConv Base | 210M | 10.6 | 12.8 | -1.95 | 73.8 | 63.6 | 70.3 |
| LConv Large | 741M | 41.0 | 3.0 | -1.76 | 76.8 | 65.6 | 76.3 |
| LConv XL | 2.3B | 77.0 | 1.0 | -1.75 | 73.3 | 64.1 | 72.9 |
| DConv Tiny | 22M | 1.39 | 27.3 | -2.46 | 51.1 | 48.9 | 30.2 |
| DConv Small | 96M | 4.97 | 19.8 | -2.08 | 68.6 | 57.4 | 64.3 |
| DConv Base | 324M | 15.3 | 7.6 | -1.90 | 72.9 | 60.1 | 63.7 |
| DConv Large | 1.2B | 78.0 | 1.1 | -1.82 | 70.8 | 58.5 | 58.2 |
| MLP-Mixer Small | 67M | 3.83 | 22.3 | -2.15 | 65.4 | 55.1 | 58.7 |
| MLP-Mixer Base | 233M | 12.4 | 10.7 | -1.90 | 64.4 | 58.1 | 60.5 |
| MLP-Mixer Large | 739M | 38.3 | 3.9 | -1.73 | 52.2 | 47.8 | 60.9 |
| MLP-Mixer XL | 2.86B | 48.3 | 1.2 | -1.61 | 57.3 | 58.9 | 65.7 |
| Model | αF,U | αF,D | αP,U | αP,D | αU,D |
| Transformer | 0.54 | 0.28 | 0.47 | 0.24 | 0.49 |
| GLU-Trans. | 0.49 | 0.24 | 0.42 | 0.22 | 0.46 |
| LConv | 0.32 | 0.13 | 0.29 | 0.11 | 0.48 |
| Funnel | 0.47 | 0.22 | 0.38 | 0.18 | 0.46 |
| Switch | 0.23 | 0.14 | 0.13 | 0.08 | 0.58 |
| Universal | 0.50 | 0.20 | 0.56 | 0.22 | 0.35 |
| ALBERT | 0.08 | -0.12 | 0.13 | -0.21 | -1.67 |
| Evolved | 0.44 | 0.22 | 0.42 | 0.21 | 0.47 |
| Performer | 0.25 | 0.05 | 0.24 | 0.05 | 0.24 |
| MoS-Trans. | 0.43 | 0.21 | 0.43 | 0.20 | 0.47 |
| MLP-Mixer | 0.32 | -0.03 | 0.26 | 0.65 | -0.02 |
| Model | NL | dfff | dmodel | dkv | NH | #Params |
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M |
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M |
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M |
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M |
| XL | 24/24 | 16384 | 1024 | 128 | 32 | 3B |
| Model | NL | dff | dmodel | dvk | NH | NE | #Params |
| Tiny | 4 | 1024 | 512 | 64 | 12 | 32 | 173M |
| Small | 6 | 2048 | 512 | 64 | 12 | 32 | 460M |
| Base | 12 | 3072 | 768 | 64 | 12 | 32 | 2B |
| Large | 24 | 3072 | 768 | 64 | 12 | 32 | 8B |
| XL | 48 | 3072 | 768 | 64 | 12 | 128 | 30B |
| Model | NR | dFF | dmodel | dkv | NH | #Params |
| UT Tiny | 3/3 | 1024 | 128 | 32 | 8 | 11M |
| UT Small | 3/3 | 2048 | 512 | 32 | 8 | 52M |
| UT Base | 3/3 | 3072 | 768 | 64 | 12 | 127M |
| UT Large | 3/3 | 32768 | 1024 | 64 | 16 | 283M |
| Model | #Params | FLOPs | Speed | Neg Log Ppl | GLUE | SGLUE | SQuAD |
| Transformer Tiny | 16M | 1.21 | 38.4 | -2.47 | 69.3 | 56.9 | 73.6 |
| Transformer Small | 60M | 3.70 | 22.7 | -2.02 | 78.1 | 65.3 | 81.9 |
| Transformer Base | 223M | 11.4 | 9.3 | -1.75 | 83.8 | 74.0 | 86.3 |
| Transformer Large | 738M | 34.3 | 3.6 | -1.61 | 86.4 | 78.3 | 88.6 |
| Transformer XL | 2.9B | 63.8 | 1.3 | -1.49 | 87.8 | 81.5 | 89.5 |
| Evolved Transformer Tiny | 19M | 1.31 | 39.7 | -2.45 | 69.6 | 57.1 | 69.6 |
| Evolved Transformer Small | 79M | 4.23 | 23.7 | -2.04 | 75.7 | 66.2 | 80.2 |
| Evolved Transformer Base | 218M | 10.2 | 8.9 | -1.79 | 83.0 | 70.5 | 84.8 |
| Evolved Transformer Large | 1.0B | 49.3 | 2.1 | -1.62 | 86.2 | 77.1 | 88.0 |
| Evolved Transformer XL | 2.2B | 71.3 | 0.8 | -1.55 | 87.0 | 78.3 | 88.2 |
| Universal Transformer Tiny | 11M | 1.77 | 38.1 | -2.73 | 69.8 | 56.1 | 62.3 |
| Universal Transformer Small | 52M | 7.30 | 18.3 | -2.12 | 76.8 | 64.2 | 75.4 |
| Universal Transformer Base | 127M | 20.3 | 8.4 | -1.91 | 80.0 | 67.9 | 80.1 |
| Universal Transformer Large | 283M | 27.6 | 1.6 | -1.67 | 84.0 | 73.4 | 85.4 |
| Switch Transformer Tiny | 174M | 3.25 | 29.7 | -2.01 | 78.2 | 63.8 | 80.7 |
| Switch Transformer Small | 460M | 4.63 | 22.3 | -1.85 | 80.3 | 68.0 | 82.9 |
| Switch Transformer Base | 2.0B | 12.7 | 8.4 | -1.66 | 84.2 | 74.1 | 86.5 |
| Switch Transformer Large | 3.9B | 23.0 | 4.1 | -1.56 | 84.6 | 75.8 | 87.9 |
| Switch Transformer XL | 29.6B | 43.3 | 0.8 | -1.62 | 84.0 | 75.2 | 87.5 |
| Performer Tiny | 16M | 1.14 | 42.0 | -2.88 | 50.5 | 48.8 | 15.0 |
| Performer Small | 61M | 3.50 | 39.0 | -2.44 | 57.8 | 51.1 | 31.1 |
| Performer Base | 224M | 10.8 | 11.7 | -2.23 | 61.4 | 53.4 | 37.8 |
| Performer Large | 739M | 32.8 | 4.4 | -2.16 | 62.4 | 52.4 | 30.8 |
| Funnel Transformer Tiny | 16M | 1.10 | 39.9 | -2.58 | 63.4 | 49.4 | 54.6 |
| Funnel Transformer Small | 61M | 2.96 | 32.7 | -2.11 | 70.0 | 58.5 | 75.1 |
| Funnel Transformer Base | 223M | 8.10 | 11.9 | -1.83 | 76.3 | 62.9 | 81.6 |
| Funnel Transformer Large | 739M | 22.6 | 5.0 | -1.69 | 79.8 | 67.1 | 83.8 |
| Funnel Transformer XL | 2.9B | 40.3 | 1.89 | -1.61 | 79.8 | 68.0 | 83.7 |
| ALBERT Small | 15M | 3.57 | 42.0 | -2.36 | 73.7 | 62.0 | 77.1 |
| ALBERT Base | 21M | 9.40 | 16.4 | -2.28 | 69.0 | 57.2 | 64.3 |
| ALBERT Large | 34M | 31.6 | 5.1 | -2.20 | 62.9 | 54.1 | 27.3 |
| MoS-Transformer Tiny | 27M | 1.29 | 39.7 | -2.37 | 70.6 | 57.9 | 74.1 |
| MoS-Transformer Small | 81M | 3.70 | 26.3 | -1.98 | 79.7 | 67.1 | 83.1 |
| MoS-Transformer Base | 257M | 11.4 | 8.6 | -1.70 | 84.5 | 73.9 | 86.8 |
| MoS-Transformer Large | 800M | 35.0 | 3.4 | -1.56 | 86.5 | 79.7 | 89.1 |
| MoS-Transformer XL | 2.9B | 112 | 1.2 | -1.45 | 88.2 | 81.4 | 90.0 |
| GLU-Transformer Tiny | 26M | 1.29 | 31.7 | -2.35 | 70.5 | 57.0 | 74.2 |
| GLU-Transformer Small | 77M | 3.70 | 26.4 | -1.97 | 79.1 | 67.4 | 83.0 |
| GLU-Transformer Base | 248M | 11.4 | 8.6 | -1.71 | 84.6 | 74.5 | 87.2 |
| GLU-Transformer Large | 748M | 35.0 | 3.4 | -1.56 | 84.2 | 74.3 | 86.2 |
| GLU-Transformer XL | 2.85B | 61.3 | 1.0 | -1.49 | 87.6 | 82.9 | 89.4 |
| LConv Tiny | 17M | 1.20 | 31.2 | -2.50 | 51.1 | 51.3 | 49.5 |
| LConv Small | 67M | 3.80 | 12.8 | -2.10 | 71.8 | 59.9 | 64.7 |
| LConv Base | 210M | 10.6 | 12.8 | -1.95 | 73.8 | 63.6 | 70.3 |
| LConv Large | 741M | 41.0 | 3.0 | -1.76 | 76.8 | 65.6 | 76.3 |
| LConv XL | 2.3B | 77.0 | 1.0 | -1.75 | 73.3 | 64.1 | 72.9 |
| DConv Tiny | 22M | 1.39 | 27.3 | -2.46 | 51.1 | 48.9 | 30.2 |
| DConv Small | 96M | 4.97 | 19.8 | -2.08 | 68.6 | 57.4 | 64.3 |
| DConv Base | 324M | 15.3 | 7.6 | -1.90 | 72.9 | 60.1 | 63.7 |
| DConv Large | 1.2B | 78.0 | 1.1 | -1.82 | 70.8 | 58.5 | 58.2 |
| MLP-Mixer Small | 67M | 3.83 | 22.3 | -2.15 | 65.4 | 55.1 | 58.7 |
| MLP-Mixer Base | 233M | 12.4 | 10.7 | -1.90 | 64.4 | 58.1 | 60.5 |
| MLP-Mixer Large | 739M | 38.3 | 3.9 | -1.73 | 52.2 | 47.8 | 60.9 |
| MLP-Mixer XL | 2.86B | 48.3 | 1.2 | -1.61 | 57.3 | 58.9 | 65.7 |
| Model | αF,U | αF,D | αP,U | αP,D | αU,D |
| Transformer | 0.54 | 0.28 | 0.47 | 0.24 | 0.49 |
| GLU-Trans. | 0.49 | 0.24 | 0.42 | 0.22 | 0.46 |
| LConv | 0.32 | 0.13 | 0.29 | 0.11 | 0.48 |
| Funnel | 0.47 | 0.22 | 0.38 | 0.18 | 0.46 |
| Switch | 0.23 | 0.14 | 0.13 | 0.08 | 0.58 |
| Universal | 0.50 | 0.20 | 0.56 | 0.22 | 0.35 |
| ALBERT | 0.08 | -0.12 | 0.13 | -0.21 | -1.67 |
| Evolved | 0.44 | 0.22 | 0.42 | 0.21 | 0.47 |
| Performer | 0.25 | 0.05 | 0.24 | 0.05 | 0.24 |
| MoS-Trans. | 0.43 | 0.21 | 0.43 | 0.20 | 0.47 |
| MLP-Mixer | 0.32 | -0.03 | 0.26 | 0.65 | -0.02 |
| Model | NL | dfff | dmodel | dkv | NH | #Params |
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M |
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M |
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M |
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M |
| XL | 24/24 | 16384 | 1024 | 128 | 32 | 3B |
| Model | NL | dff | dmodel | dvk | NH | NE | #Params |
| Tiny | 4 | 1024 | 512 | 64 | 12 | 32 | 173M |
| Small | 6 | 2048 | 512 | 64 | 12 | 32 | 460M |
| Base | 12 | 3072 | 768 | 64 | 12 | 32 | 2B |
| Large | 24 | 3072 | 768 | 64 | 12 | 32 | 8B |
| XL | 48 | 3072 | 768 | 64 | 12 | 128 | 30B |
| Model | NR | dFF | dmodel | dkv | NH | #Params |
| UT Tiny | 3/3 | 1024 | 128 | 32 | 8 | 11M |
| UT Small | 3/3 | 2048 | 512 | 32 | 8 | 52M |
| UT Base | 3/3 | 3072 | 768 | 64 | 12 | 127M |
| UT Large | 3/3 | 32768 | 1024 | 64 | 16 | 283M |
| Model | #Params | FLOPs | Speed | Neg Log Ppl | GLUE | SGLUE | SQuAD |
| Transformer Tiny | 16M | 1.21 | 38.4 | -2.47 | 69.3 | 56.9 | 73.6 |
| Transformer Small | 60M | 3.70 | 22.7 | -2.02 | 78.1 | 65.3 | 81.9 |
| Transformer Base | 223M | 11.4 | 9.3 | -1.75 | 83.8 | 74.0 | 86.3 |
| Transformer Large | 738M | 34.3 | 3.6 | -1.61 | 86.4 | 78.3 | 88.6 |
| Transformer XL | 2.9B | 63.8 | 1.3 | -1.49 | 87.8 | 81.5 | 89.5 |
| Evolved Transformer Tiny | 19M | 1.31 | 39.7 | -2.45 | 69.6 | 57.1 | 69.6 |
| Evolved Transformer Small | 79M | 4.23 | 23.7 | -2.04 | 75.7 | 66.2 | 80.2 |
| Evolved Transformer Base | 218M | 10.2 | 8.9 | -1.79 | 83.0 | 70.5 | 84.8 |
| Evolved Transformer Large | 1.0B | 49.3 | 2.1 | -1.62 | 86.2 | 77.1 | 88.0 |
| Evolved Transformer XL | 2.2B | 71.3 | 0.8 | -1.55 | 87.0 | 78.3 | 88.2 |
| Universal Transformer Tiny | 11M | 1.77 | 38.1 | -2.73 | 69.8 | 56.1 | 62.3 |
| Universal Transformer Small | 52M | 7.30 | 18.3 | -2.12 | 76.8 | 64.2 | 75.4 |
| Universal Transformer Base | 127M | 20.3 | 8.4 | -1.91 | 80.0 | 67.9 | 80.1 |
| Universal Transformer Large | 283M | 27.6 | 1.6 | -1.67 | 84.0 | 73.4 | 85.4 |
| Switch Transformer Tiny | 174M | 3.25 | 29.7 | -2.01 | 78.2 | 63.8 | 80.7 |
| Switch Transformer Small | 460M | 4.63 | 22.3 | -1.85 | 80.3 | 68.0 | 82.9 |
| Switch Transformer Base | 2.0B | 12.7 | 8.4 | -1.66 | 84.2 | 74.1 | 86.5 |
| Switch Transformer Large | 3.9B | 23.0 | 4.1 | -1.56 | 84.6 | 75.8 | 87.9 |
| Switch Transformer XL | 29.6B | 43.3 | 0.8 | -1.62 | 84.0 | 75.2 | 87.5 |
| Performer Tiny | 16M | 1.14 | 42.0 | -2.88 | 50.5 | 48.8 | 15.0 |
| Performer Small | 61M | 3.50 | 39.0 | -2.44 | 57.8 | 51.1 | 31.1 |
| Performer Base | 224M | 10.8 | 11.7 | -2.23 | 61.4 | 53.4 | 37.8 |
| Performer Large | 739M | 32.8 | 4.4 | -2.16 | 62.4 | 52.4 | 30.8 |
| Funnel Transformer Tiny | 16M | 1.10 | 39.9 | -2.58 | 63.4 | 49.4 | 54.6 |
| Funnel Transformer Small | 61M | 2.96 | 32.7 | -2.11 | 70.0 | 58.5 | 75.1 |
| Funnel Transformer Base | 223M | 8.10 | 11.9 | -1.83 | 76.3 | 62.9 | 81.6 |
| Funnel Transformer Large | 739M | 22.6 | 5.0 | -1.69 | 79.8 | 67.1 | 83.8 |
| Funnel Transformer XL | 2.9B | 40.3 | 1.89 | -1.61 | 79.8 | 68.0 | 83.7 |
| ALBERT Small | 15M | 3.57 | 42.0 | -2.36 | 73.7 | 62.0 | 77.1 |
| ALBERT Base | 21M | 9.40 | 16.4 | -2.28 | 69.0 | 57.2 | 64.3 |
| ALBERT Large | 34M | 31.6 | 5.1 | -2.20 | 62.9 | 54.1 | 27.3 |
| MoS-Transformer Tiny | 27M | 1.29 | 39.7 | -2.37 | 70.6 | 57.9 | 74.1 |
| MoS-Transformer Small | 81M | 3.70 | 26.3 | -1.98 | 79.7 | 67.1 | 83.1 |
| MoS-Transformer Base | 257M | 11.4 | 8.6 | -1.70 | 84.5 | 73.9 | 86.8 |
| MoS-Transformer Large | 800M | 35.0 | 3.4 | -1.56 | 86.5 | 79.7 | 89.1 |
| MoS-Transformer XL | 2.9B | 112 | 1.2 | -1.45 | 88.2 | 81.4 | 90.0 |
| GLU-Transformer Tiny | 26M | 1.29 | 31.7 | -2.35 | 70.5 | 57.0 | 74.2 |
| GLU-Transformer Small | 77M | 3.70 | 26.4 | -1.97 | 79.1 | 67.4 | 83.0 |
| GLU-Transformer Base | 248M | 11.4 | 8.6 | -1.71 | 84.6 | 74.5 | 87.2 |
| GLU-Transformer Large | 748M | 35.0 | 3.4 | -1.56 | 84.2 | 74.3 | 86.2 |
| GLU-Transformer XL | 2.85B | 61.3 | 1.0 | -1.49 | 87.6 | 82.9 | 89.4 |
| LConv Tiny | 17M | 1.20 | 31.2 | -2.50 | 51.1 | 51.3 | 49.5 |
| LConv Small | 67M | 3.80 | 12.8 | -2.10 | 71.8 | 59.9 | 64.7 |
| LConv Base | 210M | 10.6 | 12.8 | -1.95 | 73.8 | 63.6 | 70.3 |
| LConv Large | 741M | 41.0 | 3.0 | -1.76 | 76.8 | 65.6 | 76.3 |
| LConv XL | 2.3B | 77.0 | 1.0 | -1.75 | 73.3 | 64.1 | 72.9 |
| DConv Tiny | 22M | 1.39 | 27.3 | -2.46 | 51.1 | 48.9 | 30.2 |
| DConv Small | 96M | 4.97 | 19.8 | -2.08 | 68.6 | 57.4 | 64.3 |
| DConv Base | 324M | 15.3 | 7.6 | -1.90 | 72.9 | 60.1 | 63.7 |
| DConv Large | 1.2B | 78.0 | 1.1 | -1.82 | 70.8 | 58.5 | 58.2 |
| MLP-Mixer Small | 67M | 3.83 | 22.3 | -2.15 | 65.4 | 55.1 | 58.7 |
| MLP-Mixer Base | 233M | 12.4 | 10.7 | -1.90 | 64.4 | 58.1 | 60.5 |
| MLP-Mixer Large | 739M | 38.3 | 3.9 | -1.73 | 52.2 | 47.8 | 60.9 |
| MLP-Mixer XL | 2.86B | 48.3 | 1.2 | -1.61 | 57.3 | 58.9 | 65.7 |
| Model | αF,U | αF,D | αP,U | αP,D | αU,D |
| Transformer | 0.54 | 0.28 | 0.47 | 0.24 | 0.49 |
| GLU-Trans. | 0.49 | 0.24 | 0.42 | 0.22 | 0.46 |
| LConv | 0.32 | 0.13 | 0.29 | 0.11 | 0.48 |
| Funnel | 0.47 | 0.22 | 0.38 | 0.18 | 0.46 |
| Switch | 0.23 | 0.14 | 0.13 | 0.08 | 0.58 |
| Universal | 0.50 | 0.20 | 0.56 | 0.22 | 0.35 |
| ALBERT | 0.08 | -0.12 | 0.13 | -0.21 | -1.67 |
| Evolved | 0.44 | 0.22 | 0.42 | 0.21 | 0.47 |
| Performer | 0.25 | 0.05 | 0.24 | 0.05 | 0.24 |
| MoS-Trans. | 0.43 | 0.21 | 0.43 | 0.20 | 0.47 |
| MLP-Mixer | 0.32 | -0.03 | 0.26 | 0.65 | -0.02 |
| Model | NL | dfff | dmodel | dkv | NH | #Params |
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M |
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M |
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M |
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M |
| XL | 24/24 | 16384 | 1024 | 128 | 32 | 3B |
| Model | NL | dff | dmodel | dvk | NH | NE | #Params |
| Tiny | 4 | 1024 | 512 | 64 | 12 | 32 | 173M |
| Small | 6 | 2048 | 512 | 64 | 12 | 32 | 460M |
| Base | 12 | 3072 | 768 | 64 | 12 | 32 | 2B |
| Large | 24 | 3072 | 768 | 64 | 12 | 32 | 8B |
| XL | 48 | 3072 | 768 | 64 | 12 | 128 | 30B |
| Model | NR | dFF | dmodel | dkv | NH | #Params |
| UT Tiny | 3/3 | 1024 | 128 | 32 | 8 | 11M |
| UT Small | 3/3 | 2048 | 512 | 32 | 8 | 52M |
| UT Base | 3/3 | 3072 | 768 | 64 | 12 | 127M |
| UT Large | 3/3 | 32768 | 1024 | 64 | 16 | 283M |
| Model | #Params | FLOPs | Speed | Neg Log Ppl | GLUE | SGLUE | SQuAD |
| Transformer Tiny | 16M | 1.21 | 38.4 | -2.47 | 69.3 | 56.9 | 73.6 |
| Transformer Small | 60M | 3.70 | 22.7 | -2.02 | 78.1 | 65.3 | 81.9 |
| Transformer Base | 223M | 11.4 | 9.3 | -1.75 | 83.8 | 74.0 | 86.3 |
| Transformer Large | 738M | 34.3 | 3.6 | -1.61 | 86.4 | 78.3 | 88.6 |
| Transformer XL | 2.9B | 63.8 | 1.3 | -1.49 | 87.8 | 81.5 | 89.5 |
| Evolved Transformer Tiny | 19M | 1.31 | 39.7 | -2.45 | 69.6 | 57.1 | 69.6 |
| Evolved Transformer Small | 79M | 4.23 | 23.7 | -2.04 | 75.7 | 66.2 | 80.2 |
| Evolved Transformer Base | 218M | 10.2 | 8.9 | -1.79 | 83.0 | 70.5 | 84.8 |
| Evolved Transformer Large | 1.0B | 49.3 | 2.1 | -1.62 | 86.2 | 77.1 | 88.0 |
| Evolved Transformer XL | 2.2B | 71.3 | 0.8 | -1.55 | 87.0 | 78.3 | 88.2 |
| Universal Transformer Tiny | 11M | 1.77 | 38.1 | -2.73 | 69.8 | 56.1 | 62.3 |
| Universal Transformer Small | 52M | 7.30 | 18.3 | -2.12 | 76.8 | 64.2 | 75.4 |
| Universal Transformer Base | 127M | 20.3 | 8.4 | -1.91 | 80.0 | 67.9 | 80.1 |
| Universal Transformer Large | 283M | 27.6 | 1.6 | -1.67 | 84.0 | 73.4 | 85.4 |
| Switch Transformer Tiny | 174M | 3.25 | 29.7 | -2.01 | 78.2 | 63.8 | 80.7 |
| Switch Transformer Small | 460M | 4.63 | 22.3 | -1.85 | 80.3 | 68.0 | 82.9 |
| Switch Transformer Base | 2.0B | 12.7 | 8.4 | -1.66 | 84.2 | 74.1 | 86.5 |
| Switch Transformer Large | 3.9B | 23.0 | 4.1 | -1.56 | 84.6 | 75.8 | 87.9 |
| Switch Transformer XL | 29.6B | 43.3 | 0.8 | -1.62 | 84.0 | 75.2 | 87.5 |
| Performer Tiny | 16M | 1.14 | 42.0 | -2.88 | 50.5 | 48.8 | 15.0 |
| Performer Small | 61M | 3.50 | 39.0 | -2.44 | 57.8 | 51.1 | 31.1 |
| Performer Base | 224M | 10.8 | 11.7 | -2.23 | 61.4 | 53.4 | 37.8 |
| Performer Large | 739M | 32.8 | 4.4 | -2.16 | 62.4 | 52.4 | 30.8 |
| Funnel Transformer Tiny | 16M | 1.10 | 39.9 | -2.58 | 63.4 | 49.4 | 54.6 |
| Funnel Transformer Small | 61M | 2.96 | 32.7 | -2.11 | 70.0 | 58.5 | 75.1 |
| Funnel Transformer Base | 223M | 8.10 | 11.9 | -1.83 | 76.3 | 62.9 | 81.6 |
| Funnel Transformer Large | 739M | 22.6 | 5.0 | -1.69 | 79.8 | 67.1 | 83.8 |
| Funnel Transformer XL | 2.9B | 40.3 | 1.89 | -1.61 | 79.8 | 68.0 | 83.7 |
| ALBERT Small | 15M | 3.57 | 42.0 | -2.36 | 73.7 | 62.0 | 77.1 |
| ALBERT Base | 21M | 9.40 | 16.4 | -2.28 | 69.0 | 57.2 | 64.3 |
| ALBERT Large | 34M | 31.6 | 5.1 | -2.20 | 62.9 | 54.1 | 27.3 |
| MoS-Transformer Tiny | 27M | 1.29 | 39.7 | -2.37 | 70.6 | 57.9 | 74.1 |
| MoS-Transformer Small | 81M | 3.70 | 26.3 | -1.98 | 79.7 | 67.1 | 83.1 |
| MoS-Transformer Base | 257M | 11.4 | 8.6 | -1.70 | 84.5 | 73.9 | 86.8 |
| MoS-Transformer Large | 800M | 35.0 | 3.4 | -1.56 | 86.5 | 79.7 | 89.1 |
| MoS-Transformer XL | 2.9B | 112 | 1.2 | -1.45 | 88.2 | 81.4 | 90.0 |
| GLU-Transformer Tiny | 26M | 1.29 | 31.7 | -2.35 | 70.5 | 57.0 | 74.2 |
| GLU-Transformer Small | 77M | 3.70 | 26.4 | -1.97 | 79.1 | 67.4 | 83.0 |
| GLU-Transformer Base | 248M | 11.4 | 8.6 | -1.71 | 84.6 | 74.5 | 87.2 |
| GLU-Transformer Large | 748M | 35.0 | 3.4 | -1.56 | 84.2 | 74.3 | 86.2 |
| GLU-Transformer XL | 2.85B | 61.3 | 1.0 | -1.49 | 87.6 | 82.9 | 89.4 |
| LConv Tiny | 17M | 1.20 | 31.2 | -2.50 | 51.1 | 51.3 | 49.5 |
| LConv Small | 67M | 3.80 | 12.8 | -2.10 | 71.8 | 59.9 | 64.7 |
| LConv Base | 210M | 10.6 | 12.8 | -1.95 | 73.8 | 63.6 | 70.3 |
| LConv Large | 741M | 41.0 | 3.0 | -1.76 | 76.8 | 65.6 | 76.3 |
| LConv XL | 2.3B | 77.0 | 1.0 | -1.75 | 73.3 | 64.1 | 72.9 |
| DConv Tiny | 22M | 1.39 | 27.3 | -2.46 | 51.1 | 48.9 | 30.2 |
| DConv Small | 96M | 4.97 | 19.8 | -2.08 | 68.6 | 57.4 | 64.3 |
| DConv Base | 324M | 15.3 | 7.6 | -1.90 | 72.9 | 60.1 | 63.7 |
| DConv Large | 1.2B | 78.0 | 1.1 | -1.82 | 70.8 | 58.5 | 58.2 |
| MLP-Mixer Small | 67M | 3.83 | 22.3 | -2.15 | 65.4 | 55.1 | 58.7 |
| MLP-Mixer Base | 233M | 12.4 | 10.7 | -1.90 | 64.4 | 58.1 | 60.5 |
| MLP-Mixer Large | 739M | 38.3 | 3.9 | -1.73 | 52.2 | 47.8 | 60.9 |
| MLP-Mixer XL | 2.86B | 48.3 | 1.2 | -1.61 | 57.3 | 58.9 | 65.7 |
| Model | αF,U | αF,D | αP,U | αP,D | αU,D |
| Transformer | 0.54 | 0.28 | 0.47 | 0.24 | 0.49 |
| GLU-Trans. | 0.49 | 0.24 | 0.42 | 0.22 | 0.46 |
| LConv | 0.32 | 0.13 | 0.29 | 0.11 | 0.48 |
| Funnel | 0.47 | 0.22 | 0.38 | 0.18 | 0.46 |
| Switch | 0.23 | 0.14 | 0.13 | 0.08 | 0.58 |
| Universal | 0.50 | 0.20 | 0.56 | 0.22 | 0.35 |
| ALBERT | 0.08 | -0.12 | 0.13 | -0.21 | -1.67 |
| Evolved | 0.44 | 0.22 | 0.42 | 0.21 | 0.47 |
| Performer | 0.25 | 0.05 | 0.24 | 0.05 | 0.24 |
| MoS-Trans. | 0.43 | 0.21 | 0.43 | 0.20 | 0.47 |
| MLP-Mixer | 0.32 | -0.03 | 0.26 | 0.65 | -0.02 |
| Model | NL | dfff | dmodel | dkv | NH | #Params |
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M |
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M |
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M |
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M |
| XL | 24/24 | 16384 | 1024 | 128 | 32 | 3B |
| Model | NL | dff | dmodel | dvk | NH | NE | #Params |
| Tiny | 4 | 1024 | 512 | 64 | 12 | 32 | 173M |
| Small | 6 | 2048 | 512 | 64 | 12 | 32 | 460M |
| Base | 12 | 3072 | 768 | 64 | 12 | 32 | 2B |
| Large | 24 | 3072 | 768 | 64 | 12 | 32 | 8B |
| XL | 48 | 3072 | 768 | 64 | 12 | 128 | 30B |
| Model | NR | dFF | dmodel | dkv | NH | #Params |
| UT Tiny | 3/3 | 1024 | 128 | 32 | 8 | 11M |
| UT Small | 3/3 | 2048 | 512 | 32 | 8 | 52M |
| UT Base | 3/3 | 3072 | 768 | 64 | 12 | 127M |
| UT Large | 3/3 | 32768 | 1024 | 64 | 16 | 283M |
| Model | # Pretrained images | # Pretrained Steps | # Params per token | VQA | NLVR2 | COCO | Flickr30K | ||||
| test-dev | test-standard | dev | test-P | TR | IR | TR | IR | ||||
| Base-size models pretrained in the similar settings | |||||||||||
| UNITERBASE (Chen et al., 2020) | 4M | 200k | 86M | 72.70 | 72.91 | 77.18 | 77.85 | 64.4 | 50.3 | 85.9 | 72.5 |
| VILLABASE (Gan et al., 2020) | 4M | 200k | 86M | 73.59 | 73.67 | 78.39 | 79.30 | - | - | 86.6 | 74.7 |
| UNIMOBASE (Li et al., 2021b) | 4M | 500K | 120M | 73.79 | 74.02 | - | - | - | - | 89.7 | 74.7 |
| ViLT (Kim et al., 2021a) | 4M | 200k | 120M | 71.26 | - | 75.70 | 76.13 | 61.5 | 42.7 | 83.5 | 64.4 |
| ALBEFBASE (Li et al., 2021a) | 4M | 240k | 210M | 74.54 | 74.70 | 80.24 | 80.50 | 73.1 | 56.8 | 94.3 | 82.8 |
| VLMOBASE (Bao et al., 2022b) | 4M | 200k | 180M | 76.64 | 76.89 | 82.77 | 83.34 | 74.8 | 57.2 | 92.3 | 79.3 |
| BEiT-3BASE* | 4M | 200k | 180M | 76.21 | 76.75 | 84.93 | 85.76 | 78.7 | 60.3 | 95.3 | 83.8 |
| VL-MoEBASE/16E | 4M | 200k | 180M | 78.21 | 78.63 | 85.52 | 86.77 | 79.4 | 61.2 | 96.1 | 84.9 |
| Pretained with more aggressive cost, including compute / data / model | |||||||||||
| VLMOLARGE (Bao et al., 2022b) | 4M | 200k | 560M | 79.94 | 79.98 | 85.64 | 86.86 | 78.2 | 60.6 | 95.3 | 84.5 |
| ALBEFBASE (Li et al., 2021a) | 14M | 800k | 210M | 75.84 | 76.04 | 82.55 | 83.14 | 77.6 | 60.7 | 95.9 | 85.6 |
| BLIPLARGE (Li et al., 2022) | 129M | 1.26M | 427M | 78.24 | 78.17 | 82.48 | 83.08 | 81.9 | 64.3 | 97.3 | 87.3 |
| SIMVLMBASE (Wang et al., 2022c) | 1.8B | 1M | 230M | 77.87 | 78.14 | 81.72 | 81.77 | - | - | - | - |
| SIMVLMHUGE (Wang et al., 2022c) | 1.8B | 1M | 1.7B | 80.03 | 80.34 | 84.53 | 85.15 | - | - | - | - |
| BEiT-3HUGE (Wang et al., 2022b) | 21M | 1M | 1.9B | 84.19 | 84.03 | 91.51 | 92.58 | 84.8 | 67.2 | 98.0 | 90.3 |
| PALIHUGE (Wang et al., 2022b) | 1.6B | 1M | 17B | 84.30 | 84.30 | - | - | - | - | - | - |
| BLIP2XL (Li et al., 2023b) | 129M | 250k | 4.1B | 81.55 | 81.66 | - | - | 85.4 | 68.3 | 97.6 | 89.7 |
| BEiT-3LARGE* | 4M | 200k | 560M | 78.14 | 78.23 | 85.23 | 86.15 | 79.2 | 61.4 | 95.7 | 84.1 |
| VL-MoELARGE/16E | 4M | 200k | 560M | 79.91 | 79.95 | 86.28 | 87.14 | 79.9 | 62.3 | 96.5 | 85.3 |
| Models | Pretraining | Tasks | ||
| # Images | # Steps | ImageNet | MNLI-m | |
| Vision Pretraining | ||||
| V1TB/16 | 300M | 500k | 83.6 | - |
| BE1TB/16 | 1.2M | 500k | 85.2 | - |
| V-MoEB/16-16E | 300M | 500k | 85.3 | - |
| Vision-Language Pretraining | ||||
| SIMVLMBASE | 1.8B | 1M | 80.6 | 64.4 |
| BEiT-3*BASE | 4M | 200k | 83.2 | 67.0 |
| VL-MoEBASE/16E | 4M | 200k | 84.5 | 68.1 |
| Scaling Strategy | NLVR2 | Flickr30k | ImageNet Acc@1 | MNLI-m Acc | Avg. | ||||
| T-MoE | V-MoE | dev | test-P | TR R@1 | IR R@1 | ||||
| [1] | X | X | 67.42 | 68.21 | 80.4 | 61.7 | 67.2 | 54.3 | 66.5 |
| [2] | ✓ | X | 72.42 | 72.73 | 83.2 | 64.7 | 67.8 | 58.3 | 69.9 |
| [3] | X | ✓ | 71.19 | 72.23 | 82.9 | 64.5 | 69.2 | 55.2 | 69.2 |
| [4] | ✓ | ✓ | 72.98 | 73.34 | 84.7 | 65.3 | 69.0 | 58.1 | 70.6 |
| Models | Size | Methods | Efficiency | Val | |||
| #E | Param/Mem | EP | KN | TPS | Speedup | Loss | |
| BEiT | - | 180M/0.3G | - | - | 1002.3 | - | 4.51 |
| VL-MoE | 16 | 180M/1.6G | X | X | 450.5 | ×0.9 | 4.49 |
| VL-MoE | 16 | 180M/0.3G | ✓ | X | 685.0 | ×1.4 | 4.50 |
| VL-MoE | 16 | 180M/0.3G | ✓ | ✓ | 887.5 | ×1.8 | 4.48 |
| VL-MoE | 8 | 180M/0.3G | ✓ | ✓ | 911.5 | ×1.4 | 4.51 |
| VL-MoE | 16 | 105M/0.3G | ✓ | ✓ | 1211.3 | ×1.3 | 4.50 |
| Models | Size | IN0shot | |||
| # Param | # E | 50k | 75k | 100k | |
| Contrastive Pretraining | |||||
| DENSE | 105M | - | 50.3 | 63.2 | 67.5 |
| LIMOE | 105M | 8 | 53.7 | 62.9 | 62.0 |
| LIMOE | 105M | 16 | 54.6 | 63.1 | 62.1 |
| VL-MoE | 105M | 8 | 55.2 | 64.2 | 68.3 |
| VL-MoE | 105M | 16 | 57.2 | 65.1 | 69.0 |
| DENSE | 180M | - | 60.1 | 70.3 | 78.2 |
| LIMOE | 180M | 8 | 61.5 | 70.4 | 68.2 |
| LIMOE | 180M | 16 | 61.2 | 69.3 | 67.5 |
| VL-MoE | 180M | 8 | 62.5 | 71.7 | 78.9 |
| VL-MoE | 180M | 16 | 63.2 | 72.4 | 79.5 |
| Model | # Pretrained images | # Pretrained Steps | # Params per token | VQA | NLVR2 | COCO | Flickr30K | ||||
| test-dev | test-standard | dev | test-P | TR | IR | TR | IR | ||||
| Base-size models pretrained in the similar settings | |||||||||||
| UNITERBASE (Chen et al., 2020) | 4M | 200k | 86M | 72.70 | 72.91 | 77.18 | 77.85 | 64.4 | 50.3 | 85.9 | 72.5 |
| VILLABASE (Gan et al., 2020) | 4M | 200k | 86M | 73.59 | 73.67 | 78.39 | 79.30 | - | - | 86.6 | 74.7 |
| UNIMOBASE (Li et al., 2021b) | 4M | 500K | 120M | 73.79 | 74.02 | - | - | - | - | 89.7 | 74.7 |
| ViLT (Kim et al., 2021a) | 4M | 200k | 120M | 71.26 | - | 75.70 | 76.13 | 61.5 | 42.7 | 83.5 | 64.4 |
| ALBEFBASE (Li et al., 2021a) | 4M | 240k | 210M | 74.54 | 74.70 | 80.24 | 80.50 | 73.1 | 56.8 | 94.3 | 82.8 |
| VLMOBASE (Bao et al., 2022b) | 4M | 200k | 180M | 76.64 | 76.89 | 82.77 | 83.34 | 74.8 | 57.2 | 92.3 | 79.3 |
| BEiT-3BASE* | 4M | 200k | 180M | 76.21 | 76.75 | 84.93 | 85.76 | 78.7 | 60.3 | 95.3 | 83.8 |
| VL-MoEBASE/16E | 4M | 200k | 180M | 78.21 | 78.63 | 85.52 | 86.77 | 79.4 | 61.2 | 96.1 | 84.9 |
| Pretained with more aggressive cost, including compute / data / model | |||||||||||
| VLMOLARGE (Bao et al., 2022b) | 4M | 200k | 560M | 79.94 | 79.98 | 85.64 | 86.86 | 78.2 | 60.6 | 95.3 | 84.5 |
| ALBEFBASE (Li et al., 2021a) | 14M | 800k | 210M | 75.84 | 76.04 | 82.55 | 83.14 | 77.6 | 60.7 | 95.9 | 85.6 |
| BLIPLARGE (Li et al., 2022) | 129M | 1.26M | 427M | 78.24 | 78.17 | 82.48 | 83.08 | 81.9 | 64.3 | 97.3 | 87.3 |
| SIMVLMBASE (Wang et al., 2022c) | 1.8B | 1M | 230M | 77.87 | 78.14 | 81.72 | 81.77 | - | - | - | - |
| SIMVLMHUGE (Wang et al., 2022c) | 1.8B | 1M | 1.7B | 80.03 | 80.34 | 84.53 | 85.15 | - | - | - | - |
| BEiT-3HUGE (Wang et al., 2022b) | 21M | 1M | 1.9B | 84.19 | 84.03 | 91.51 | 92.58 | 84.8 | 67.2 | 98.0 | 90.3 |
| PALIHUGE (Wang et al., 2022b) | 1.6B | 1M | 17B | 84.30 | 84.30 | - | - | - | - | - | - |
| BLIP2XL (Li et al., 2023b) | 129M | 250k | 4.1B | 81.55 | 81.66 | - | - | 85.4 | 68.3 | 97.6 | 89.7 |
| BEiT-3LARGE* | 4M | 200k | 560M | 78.14 | 78.23 | 85.23 | 86.15 | 79.2 | 61.4 | 95.7 | 84.1 |
| VL-MoELARGE/16E | 4M | 200k | 560M | 79.91 | 79.95 | 86.28 | 87.14 | 79.9 | 62.3 | 96.5 | 85.3 |
| Models | Pretraining | Tasks | ||
| # Images | # Steps | ImageNet | MNLI-m | |
| Vision Pretraining | ||||
| V1TB/16 | 300M | 500k | 83.6 | - |
| BE1TB/16 | 1.2M | 500k | 85.2 | - |
| V-MoEB/16-16E | 300M | 500k | 85.3 | - |
| Vision-Language Pretraining | ||||
| SIMVLMBASE | 1.8B | 1M | 80.6 | 64.4 |
| BEiT-3*BASE | 4M | 200k | 83.2 | 67.0 |
| VL-MoEBASE/16E | 4M | 200k | 84.5 | 68.1 |
| Scaling Strategy | NLVR2 | Flickr30k | ImageNet Acc@1 | MNLI-m Acc | Avg. | ||||
| T-MoE | V-MoE | dev | test-P | TR R@1 | IR R@1 | ||||
| [1] | X | X | 67.42 | 68.21 | 80.4 | 61.7 | 67.2 | 54.3 | 66.5 |
| [2] | ✓ | X | 72.42 | 72.73 | 83.2 | 64.7 | 67.8 | 58.3 | 69.9 |
| [3] | X | ✓ | 71.19 | 72.23 | 82.9 | 64.5 | 69.2 | 55.2 | 69.2 |
| [4] | ✓ | ✓ | 72.98 | 73.34 | 84.7 | 65.3 | 69.0 | 58.1 | 70.6 |
| Models | Size | Methods | Efficiency | Val | |||
| #E | Param/Mem | EP | KN | TPS | Speedup | Loss | |
| BEiT | - | 180M/0.3G | - | - | 1002.3 | - | 4.51 |
| VL-MoE | 16 | 180M/1.6G | X | X | 450.5 | ×0.9 | 4.49 |
| VL-MoE | 16 | 180M/0.3G | ✓ | X | 685.0 | ×1.4 | 4.50 |
| VL-MoE | 16 | 180M/0.3G | ✓ | ✓ | 887.5 | ×1.8 | 4.48 |
| VL-MoE | 8 | 180M/0.3G | ✓ | ✓ | 911.5 | ×1.4 | 4.51 |
| VL-MoE | 16 | 105M/0.3G | ✓ | ✓ | 1211.3 | ×1.3 | 4.50 |
| Models | Size | IN0shot | |||
| # Param | # E | 50k | 75k | 100k | |
| Contrastive Pretraining | |||||
| DENSE | 105M | - | 50.3 | 63.2 | 67.5 |
| LIMOE | 105M | 8 | 53.7 | 62.9 | 62.0 |
| LIMOE | 105M | 16 | 54.6 | 63.1 | 62.1 |
| VL-MoE | 105M | 8 | 55.2 | 64.2 | 68.3 |
| VL-MoE | 105M | 16 | 57.2 | 65.1 | 69.0 |
| DENSE | 180M | - | 60.1 | 70.3 | 78.2 |
| LIMOE | 180M | 8 | 61.5 | 70.4 | 68.2 |
| LIMOE | 180M | 16 | 61.2 | 69.3 | 67.5 |
| VL-MoE | 180M | 8 | 62.5 | 71.7 | 78.9 |
| VL-MoE | 180M | 16 | 63.2 | 72.4 | 79.5 |
| Model | # Pretrained images | # Pretrained Steps | # Params per token | VQA | NLVR2 | COCO | Flickr30K | ||||
| test-dev | test-standard | dev | test-P | TR | IR | TR | IR | ||||
| Base-size models pretrained in the similar settings | |||||||||||
| UNITERBASE (Chen et al., 2020) | 4M | 200k | 86M | 72.70 | 72.91 | 77.18 | 77.85 | 64.4 | 50.3 | 85.9 | 72.5 |
| VILLABASE (Gan et al., 2020) | 4M | 200k | 86M | 73.59 | 73.67 | 78.39 | 79.30 | - | - | 86.6 | 74.7 |
| UNIMOBASE (Li et al., 2021b) | 4M | 500K | 120M | 73.79 | 74.02 | - | - | - | - | 89.7 | 74.7 |
| ViLT (Kim et al., 2021a) | 4M | 200k | 120M | 71.26 | - | 75.70 | 76.13 | 61.5 | 42.7 | 83.5 | 64.4 |
| ALBEFBASE (Li et al., 2021a) | 4M | 240k | 210M | 74.54 | 74.70 | 80.24 | 80.50 | 73.1 | 56.8 | 94.3 | 82.8 |
| VLMOBASE (Bao et al., 2022b) | 4M | 200k | 180M | 76.64 | 76.89 | 82.77 | 83.34 | 74.8 | 57.2 | 92.3 | 79.3 |
| BEiT-3BASE* | 4M | 200k | 180M | 76.21 | 76.75 | 84.93 | 85.76 | 78.7 | 60.3 | 95.3 | 83.8 |
| VL-MoEBASE/16E | 4M | 200k | 180M | 78.21 | 78.63 | 85.52 | 86.77 | 79.4 | 61.2 | 96.1 | 84.9 |
| Pretained with more aggressive cost, including compute / data / model | |||||||||||
| VLMOLARGE (Bao et al., 2022b) | 4M | 200k | 560M | 79.94 | 79.98 | 85.64 | 86.86 | 78.2 | 60.6 | 95.3 | 84.5 |
| ALBEFBASE (Li et al., 2021a) | 14M | 800k | 210M | 75.84 | 76.04 | 82.55 | 83.14 | 77.6 | 60.7 | 95.9 | 85.6 |
| BLIPLARGE (Li et al., 2022) | 129M | 1.26M | 427M | 78.24 | 78.17 | 82.48 | 83.08 | 81.9 | 64.3 | 97.3 | 87.3 |
| SIMVLMBASE (Wang et al., 2022c) | 1.8B | 1M | 230M | 77.87 | 78.14 | 81.72 | 81.77 | - | - | - | - |
| SIMVLMHUGE (Wang et al., 2022c) | 1.8B | 1M | 1.7B | 80.03 | 80.34 | 84.53 | 85.15 | - | - | - | - |
| BEiT-3HUGE (Wang et al., 2022b) | 21M | 1M | 1.9B | 84.19 | 84.03 | 91.51 | 92.58 | 84.8 | 67.2 | 98.0 | 90.3 |
| PALIHUGE (Wang et al., 2022b) | 1.6B | 1M | 17B | 84.30 | 84.30 | - | - | - | - | - | - |
| BLIP2XL (Li et al., 2023b) | 129M | 250k | 4.1B | 81.55 | 81.66 | - | - | 85.4 | 68.3 | 97.6 | 89.7 |
| BEiT-3LARGE* | 4M | 200k | 560M | 78.14 | 78.23 | 85.23 | 86.15 | 79.2 | 61.4 | 95.7 | 84.1 |
| VL-MoELARGE/16E | 4M | 200k | 560M | 79.91 | 79.95 | 86.28 | 87.14 | 79.9 | 62.3 | 96.5 | 85.3 |
| Models | Pretraining | Tasks | ||
| # Images | # Steps | ImageNet | MNLI-m | |
| Vision Pretraining | ||||
| V1TB/16 | 300M | 500k | 83.6 | - |
| BE1TB/16 | 1.2M | 500k | 85.2 | - |
| V-MoEB/16-16E | 300M | 500k | 85.3 | - |
| Vision-Language Pretraining | ||||
| SIMVLMBASE | 1.8B | 1M | 80.6 | 64.4 |
| BEiT-3*BASE | 4M | 200k | 83.2 | 67.0 |
| VL-MoEBASE/16E | 4M | 200k | 84.5 | 68.1 |
| Scaling Strategy | NLVR2 | Flickr30k | ImageNet Acc@1 | MNLI-m Acc | Avg. | ||||
| T-MoE | V-MoE | dev | test-P | TR R@1 | IR R@1 | ||||
| [1] | X | X | 67.42 | 68.21 | 80.4 | 61.7 | 67.2 | 54.3 | 66.5 |
| [2] | ✓ | X | 72.42 | 72.73 | 83.2 | 64.7 | 67.8 | 58.3 | 69.9 |
| [3] | X | ✓ | 71.19 | 72.23 | 82.9 | 64.5 | 69.2 | 55.2 | 69.2 |
| [4] | ✓ | ✓ | 72.98 | 73.34 | 84.7 | 65.3 | 69.0 | 58.1 | 70.6 |
| Models | Size | Methods | Efficiency | Val | |||
| #E | Param/Mem | EP | KN | TPS | Speedup | Loss | |
| BEiT | - | 180M/0.3G | - | - | 1002.3 | - | 4.51 |
| VL-MoE | 16 | 180M/1.6G | X | X | 450.5 | ×0.9 | 4.49 |
| VL-MoE | 16 | 180M/0.3G | ✓ | X | 685.0 | ×1.4 | 4.50 |
| VL-MoE | 16 | 180M/0.3G | ✓ | ✓ | 887.5 | ×1.8 | 4.48 |
| VL-MoE | 8 | 180M/0.3G | ✓ | ✓ | 911.5 | ×1.4 | 4.51 |
| VL-MoE | 16 | 105M/0.3G | ✓ | ✓ | 1211.3 | ×1.3 | 4.50 |
| Models | Size | IN0shot | |||
| # Param | # E | 50k | 75k | 100k | |
| Contrastive Pretraining | |||||
| DENSE | 105M | - | 50.3 | 63.2 | 67.5 |
| LIMOE | 105M | 8 | 53.7 | 62.9 | 62.0 |
| LIMOE | 105M | 16 | 54.6 | 63.1 | 62.1 |
| VL-MoE | 105M | 8 | 55.2 | 64.2 | 68.3 |
| VL-MoE | 105M | 16 | 57.2 | 65.1 | 69.0 |
| DENSE | 180M | - | 60.1 | 70.3 | 78.2 |
| LIMOE | 180M | 8 | 61.5 | 70.4 | 68.2 |
| LIMOE | 180M | 16 | 61.2 | 69.3 | 67.5 |
| VL-MoE | 180M | 8 | 62.5 | 71.7 | 78.9 |
| VL-MoE | 180M | 16 | 63.2 | 72.4 | 79.5 |
| Model | # Pretrained images | # Pretrained Steps | # Params per token | VQA | NLVR2 | COCO | Flickr30K | ||||
| test-dev | test-standard | dev | test-P | TR | IR | TR | IR | ||||
| Base-size models pretrained in the similar settings | |||||||||||
| UNITERBASE (Chen et al., 2020) | 4M | 200k | 86M | 72.70 | 72.91 | 77.18 | 77.85 | 64.4 | 50.3 | 85.9 | 72.5 |
| VILLABASE (Gan et al., 2020) | 4M | 200k | 86M | 73.59 | 73.67 | 78.39 | 79.30 | - | - | 86.6 | 74.7 |
| UNIMOBASE (Li et al., 2021b) | 4M | 500K | 120M | 73.79 | 74.02 | - | - | - | - | 89.7 | 74.7 |
| ViLT (Kim et al., 2021a) | 4M | 200k | 120M | 71.26 | - | 75.70 | 76.13 | 61.5 | 42.7 | 83.5 | 64.4 |
| ALBEFBASE (Li et al., 2021a) | 4M | 240k | 210M | 74.54 | 74.70 | 80.24 | 80.50 | 73.1 | 56.8 | 94.3 | 82.8 |
| VLMOBASE (Bao et al., 2022b) | 4M | 200k | 180M | 76.64 | 76.89 | 82.77 | 83.34 | 74.8 | 57.2 | 92.3 | 79.3 |
| BEiT-3BASE* | 4M | 200k | 180M | 76.21 | 76.75 | 84.93 | 85.76 | 78.7 | 60.3 | 95.3 | 83.8 |
| VL-MoEBASE/16E | 4M | 200k | 180M | 78.21 | 78.63 | 85.52 | 86.77 | 79.4 | 61.2 | 96.1 | 84.9 |
| Pretained with more aggressive cost, including compute / data / model | |||||||||||
| VLMOLARGE (Bao et al., 2022b) | 4M | 200k | 560M | 79.94 | 79.98 | 85.64 | 86.86 | 78.2 | 60.6 | 95.3 | 84.5 |
| ALBEFBASE (Li et al., 2021a) | 14M | 800k | 210M | 75.84 | 76.04 | 82.55 | 83.14 | 77.6 | 60.7 | 95.9 | 85.6 |
| BLIPLARGE (Li et al., 2022) | 129M | 1.26M | 427M | 78.24 | 78.17 | 82.48 | 83.08 | 81.9 | 64.3 | 97.3 | 87.3 |
| SIMVLMBASE (Wang et al., 2022c) | 1.8B | 1M | 230M | 77.87 | 78.14 | 81.72 | 81.77 | - | - | - | - |
| SIMVLMHUGE (Wang et al., 2022c) | 1.8B | 1M | 1.7B | 80.03 | 80.34 | 84.53 | 85.15 | - | - | - | - |
| BEiT-3HUGE (Wang et al., 2022b) | 21M | 1M | 1.9B | 84.19 | 84.03 | 91.51 | 92.58 | 84.8 | 67.2 | 98.0 | 90.3 |
| PALIHUGE (Wang et al., 2022b) | 1.6B | 1M | 17B | 84.30 | 84.30 | - | - | - | - | - | - |
| BLIP2XL (Li et al., 2023b) | 129M | 250k | 4.1B | 81.55 | 81.66 | - | - | 85.4 | 68.3 | 97.6 | 89.7 |
| BEiT-3LARGE* | 4M | 200k | 560M | 78.14 | 78.23 | 85.23 | 86.15 | 79.2 | 61.4 | 95.7 | 84.1 |
| VL-MoELARGE/16E | 4M | 200k | 560M | 79.91 | 79.95 | 86.28 | 87.14 | 79.9 | 62.3 | 96.5 | 85.3 |
| Models | Pretraining | Tasks | ||
| # Images | # Steps | ImageNet | MNLI-m | |
| Vision Pretraining | ||||
| V1TB/16 | 300M | 500k | 83.6 | - |
| BE1TB/16 | 1.2M | 500k | 85.2 | - |
| V-MoEB/16-16E | 300M | 500k | 85.3 | - |
| Vision-Language Pretraining | ||||
| SIMVLMBASE | 1.8B | 1M | 80.6 | 64.4 |
| BEiT-3*BASE | 4M | 200k | 83.2 | 67.0 |
| VL-MoEBASE/16E | 4M | 200k | 84.5 | 68.1 |
| Scaling Strategy | NLVR2 | Flickr30k | ImageNet Acc@1 | MNLI-m Acc | Avg. | ||||
| T-MoE | V-MoE | dev | test-P | TR R@1 | IR R@1 | ||||
| [1] | X | X | 67.42 | 68.21 | 80.4 | 61.7 | 67.2 | 54.3 | 66.5 |
| [2] | ✓ | X | 72.42 | 72.73 | 83.2 | 64.7 | 67.8 | 58.3 | 69.9 |
| [3] | X | ✓ | 71.19 | 72.23 | 82.9 | 64.5 | 69.2 | 55.2 | 69.2 |
| [4] | ✓ | ✓ | 72.98 | 73.34 | 84.7 | 65.3 | 69.0 | 58.1 | 70.6 |
| Models | Size | Methods | Efficiency | Val | |||
| #E | Param/Mem | EP | KN | TPS | Speedup | Loss | |
| BEiT | - | 180M/0.3G | - | - | 1002.3 | - | 4.51 |
| VL-MoE | 16 | 180M/1.6G | X | X | 450.5 | ×0.9 | 4.49 |
| VL-MoE | 16 | 180M/0.3G | ✓ | X | 685.0 | ×1.4 | 4.50 |
| VL-MoE | 16 | 180M/0.3G | ✓ | ✓ | 887.5 | ×1.8 | 4.48 |
| VL-MoE | 8 | 180M/0.3G | ✓ | ✓ | 911.5 | ×1.4 | 4.51 |
| VL-MoE | 16 | 105M/0.3G | ✓ | ✓ | 1211.3 | ×1.3 | 4.50 |
| Models | Size | IN0shot | |||
| # Param | # E | 50k | 75k | 100k | |
| Contrastive Pretraining | |||||
| DENSE | 105M | - | 50.3 | 63.2 | 67.5 |
| LIMOE | 105M | 8 | 53.7 | 62.9 | 62.0 |
| LIMOE | 105M | 16 | 54.6 | 63.1 | 62.1 |
| VL-MoE | 105M | 8 | 55.2 | 64.2 | 68.3 |
| VL-MoE | 105M | 16 | 57.2 | 65.1 | 69.0 |
| DENSE | 180M | - | 60.1 | 70.3 | 78.2 |
| LIMOE | 180M | 8 | 61.5 | 70.4 | 68.2 |
| LIMOE | 180M | 16 | 61.2 | 69.3 | 67.5 |
| VL-MoE | 180M | 8 | 62.5 | 71.7 | 78.9 |
| VL-MoE | 180M | 16 | 63.2 | 72.4 | 79.5 |
| method | Sup | RefCOCO | RefCOCO+ | RefCOCOg | |||||
| val | testA | testB | val | testA | testB | val | test | ||
| Random | - | 15.73 | 13.51 | 19.20 | 16.29 | 13.57 | 19.60 | 18.12 | 19.10 |
| Supervised SOTA | Full | 87.51 | 90.40 | 82.67 | 81.13 | 85.52 | 72.96 | 83.35 | 81.64 |
| VC(Zhang et al., 2018) | Weak | - | 33.29 | 30.13 | - | 34.60 | 31.58 | - | - |
| ARN(Liu et al., 2019a) | 34.26 | 36.43 | 33.07 | 34.53 | 36.01 | 33.75 | - | - | |
| KPRN(Liu et al., 2019b) | 35.04 | 34.74 | 36.98 | 35.96 | 35.24 | 36.96 | - | - | |
| DTWREG(Sun et al., 2021) | 39.21 | 41.14 | 37.72 | 39.18 | 40.10 | 38.08 | - | - | |
| CPT(Yao et al., 2021) | No | ||||||||
| CPT-Blk w/ VinVL | 26.9 | 27.5 | 27.4 | 25.4 | 25.0 | 27.0 | 32.1 | 32.3 | |
| CPT-Seg w/ VinVL | 32.2 | 36.1 | 30.3 | 31.9 | 35.2 | 28.8 | 36.7 | 36.5 | |
| CLIP | No | ||||||||
| GradCAM | 44.65 | 53.49 | 36.19 | 49.41 | 59.66 | 38.62 | 52.29 | 51.28 | |
| ReCLIP | 54.04 | 58.60 | 49.54 | 55.07 | 60.47 | 47.41 | 60.85 | 61.05 | |
| Pseudo-labeling | No | ||||||||
| Pseudo-Q(Jiang et al., 2022) | 56.02 | 58.25 | 54.13 | 38.88 | 45.06 | 32.13 | 46.25 | 47.44 | |
| Pseudo-Q-adapted | 59.44 | 61.87 | 56.47 | 42.76 | 48.27 | 36.90 | 50.35 | 49.68 | |
| SGEPG(Ours) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 | 61.33 | 60.61 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| w/o SG | 63.45 | 63.9 | 63.47 | 43.8 | 45.81 | 44.14 |
| w/o ARS | 65.24 | 63.81 | 64.65 | 48.5 | 50.47 | 46.02 |
| w/o DDRQ | 68.82 | 72.46 | 64.99 | 53.78 | 58.26 | 46.14 |
| SGEPG | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| replace rate (α) | 0% | 25% | 37.5% | 50% | 75% |
| val | 53.78 | 54.29 | 54.94 | 55.21 | 55.33 |
| testA | 58.26 | 60.15 | 58.99 | 60.67 | 60.9 |
| testB | 46.14 | 46.41 | 46.7 | 46.76 | 46.8 |
| Expression number per object | Object number per image | ||||||||
| 1 | 2 | 3 | 4 | 1 | 2 | 3 | 4 | 5 | |
| SGEPG | 66.17 | 68.21 | 68.6 | 68.81 | 62.3 | 67.61 | 68.82 | 68.6 | 68.02 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| preliminary scene graph | 36.23 | 45.31 | 25.32 | 37.21 | 46.11 | 26.55 |
| Core scene graph (SGEPG) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| method | Sup | RefCOCO | RefCOCO+ | RefCOCOg | |||||
| val | testA | testB | val | testA | testB | val | test | ||
| Random | - | 15.73 | 13.51 | 19.20 | 16.29 | 13.57 | 19.60 | 18.12 | 19.10 |
| Supervised SOTA | Full | 87.51 | 90.40 | 82.67 | 81.13 | 85.52 | 72.96 | 83.35 | 81.64 |
| VC(Zhang et al., 2018) | Weak | - | 33.29 | 30.13 | - | 34.60 | 31.58 | - | - |
| ARN(Liu et al., 2019a) | 34.26 | 36.43 | 33.07 | 34.53 | 36.01 | 33.75 | - | - | |
| KPRN(Liu et al., 2019b) | 35.04 | 34.74 | 36.98 | 35.96 | 35.24 | 36.96 | - | - | |
| DTWREG(Sun et al., 2021) | 39.21 | 41.14 | 37.72 | 39.18 | 40.10 | 38.08 | - | - | |
| CPT(Yao et al., 2021) | No | ||||||||
| CPT-Blk w/ VinVL | 26.9 | 27.5 | 27.4 | 25.4 | 25.0 | 27.0 | 32.1 | 32.3 | |
| CPT-Seg w/ VinVL | 32.2 | 36.1 | 30.3 | 31.9 | 35.2 | 28.8 | 36.7 | 36.5 | |
| CLIP | No | ||||||||
| GradCAM | 44.65 | 53.49 | 36.19 | 49.41 | 59.66 | 38.62 | 52.29 | 51.28 | |
| ReCLIP | 54.04 | 58.60 | 49.54 | 55.07 | 60.47 | 47.41 | 60.85 | 61.05 | |
| Pseudo-labeling | No | ||||||||
| Pseudo-Q(Jiang et al., 2022) | 56.02 | 58.25 | 54.13 | 38.88 | 45.06 | 32.13 | 46.25 | 47.44 | |
| Pseudo-Q-adapted | 59.44 | 61.87 | 56.47 | 42.76 | 48.27 | 36.90 | 50.35 | 49.68 | |
| SGEPG(Ours) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 | 61.33 | 60.61 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| w/o SG | 63.45 | 63.9 | 63.47 | 43.8 | 45.81 | 44.14 |
| w/o ARS | 65.24 | 63.81 | 64.65 | 48.5 | 50.47 | 46.02 |
| w/o DDRQ | 68.82 | 72.46 | 64.99 | 53.78 | 58.26 | 46.14 |
| SGEPG | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| replace rate (α) | 0% | 25% | 37.5% | 50% | 75% |
| val | 53.78 | 54.29 | 54.94 | 55.21 | 55.33 |
| testA | 58.26 | 60.15 | 58.99 | 60.67 | 60.9 |
| testB | 46.14 | 46.41 | 46.7 | 46.76 | 46.8 |
| Expression number per object | Object number per image | ||||||||
| 1 | 2 | 3 | 4 | 1 | 2 | 3 | 4 | 5 | |
| SGEPG | 66.17 | 68.21 | 68.6 | 68.81 | 62.3 | 67.61 | 68.82 | 68.6 | 68.02 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| preliminary scene graph | 36.23 | 45.31 | 25.32 | 37.21 | 46.11 | 26.55 |
| Core scene graph (SGEPG) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| method | Sup | RefCOCO | RefCOCO+ | RefCOCOg | |||||
| val | testA | testB | val | testA | testB | val | test | ||
| Random | - | 15.73 | 13.51 | 19.20 | 16.29 | 13.57 | 19.60 | 18.12 | 19.10 |
| Supervised SOTA | Full | 87.51 | 90.40 | 82.67 | 81.13 | 85.52 | 72.96 | 83.35 | 81.64 |
| VC(Zhang et al., 2018) | Weak | - | 33.29 | 30.13 | - | 34.60 | 31.58 | - | - |
| ARN(Liu et al., 2019a) | 34.26 | 36.43 | 33.07 | 34.53 | 36.01 | 33.75 | - | - | |
| KPRN(Liu et al., 2019b) | 35.04 | 34.74 | 36.98 | 35.96 | 35.24 | 36.96 | - | - | |
| DTWREG(Sun et al., 2021) | 39.21 | 41.14 | 37.72 | 39.18 | 40.10 | 38.08 | - | - | |
| CPT(Yao et al., 2021) | No | ||||||||
| CPT-Blk w/ VinVL | 26.9 | 27.5 | 27.4 | 25.4 | 25.0 | 27.0 | 32.1 | 32.3 | |
| CPT-Seg w/ VinVL | 32.2 | 36.1 | 30.3 | 31.9 | 35.2 | 28.8 | 36.7 | 36.5 | |
| CLIP | No | ||||||||
| GradCAM | 44.65 | 53.49 | 36.19 | 49.41 | 59.66 | 38.62 | 52.29 | 51.28 | |
| ReCLIP | 54.04 | 58.60 | 49.54 | 55.07 | 60.47 | 47.41 | 60.85 | 61.05 | |
| Pseudo-labeling | No | ||||||||
| Pseudo-Q(Jiang et al., 2022) | 56.02 | 58.25 | 54.13 | 38.88 | 45.06 | 32.13 | 46.25 | 47.44 | |
| Pseudo-Q-adapted | 59.44 | 61.87 | 56.47 | 42.76 | 48.27 | 36.90 | 50.35 | 49.68 | |
| SGEPG(Ours) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 | 61.33 | 60.61 | |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| w/o SG | 63.45 | 63.9 | 63.47 | 43.8 | 45.81 | 44.14 |
| w/o ARS | 65.24 | 63.81 | 64.65 | 48.5 | 50.47 | 46.02 |
| w/o DDRQ | 68.82 | 72.46 | 64.99 | 53.78 | 58.26 | 46.14 |
| SGEPG | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| replace rate (α) | 0% | 25% | 37.5% | 50% | 75% |
| val | 53.78 | 54.29 | 54.94 | 55.21 | 55.33 |
| testA | 58.26 | 60.15 | 58.99 | 60.67 | 60.9 |
| testB | 46.14 | 46.41 | 46.7 | 46.76 | 46.8 |
| Expression number per object | Object number per image | ||||||||
| 1 | 2 | 3 | 4 | 1 | 2 | 3 | 4 | 5 | |
| SGEPG | 66.17 | 68.21 | 68.6 | 68.81 | 62.3 | 67.61 | 68.82 | 68.6 | 68.02 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| preliminary scene graph | 36.23 | 45.31 | 25.32 | 37.21 | 46.11 | 26.55 |
| Core scene graph (SGEPG) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| method | Sup | RefCOCO | RefCOCO+ | RefCOCOg | |||||
| val | testA | testB | val | testA | testB | val | test | ||
| Random | - | 15.73 | 13.51 | 19.20 | 16.29 | 13.57 | 19.60 | 18.12 | 19.10 |
| Supervised SOTA | Full | 87.51 | 90.40 | 82.67 | 81.13 | 85.52 | 72.96 | 83.35 | 81.64 |
| VC(Zhang et al., 2018) | Weak | - | 33.29 | 30.13 | - | 34.60 | 31.58 | - | - |
| ARN(Liu et al., 2019a) | 34.26 | 36.43 | 33.07 | 34.53 | 36.01 | 33.75 | - | - | |
| KPRN(Liu et al., 2019b) | 35.04 | 34.74 | 36.98 | 35.96 | 35.24 | 36.96 | - | - | |
| DTWREG(Sun et al., 2021) | 39.21 | 41.14 | 37.72 | 39.18 | 40.10 | 38.08 | - | - | |
| CPT(Yao et al., 2021) | No | ||||||||
| CPT-Blk w/ VinVL | 26.9 | 27.5 | 27.4 | 25.4 | 25.0 | 27.0 | 32.1 | 32.3 | |
| CPT-Seg w/ VinVL | 32.2 | 36.1 | 30.3 | 31.9 | 35.2 | 28.8 | 36.7 | 36.5 | |
| CLIP | No | ||||||||
| GradCAM | 44.65 | 53.49 | 36.19 | 49.41 | 59.66 | 38.62 | 52.29 | 51.28 | |
| ReCLIP | 54.04 | 58.60 | 49.54 | 55.07 | 60.47 | 47.41 | 60.85 | 61.05 | |
| Pseudo-labeling | No | ||||||||
| Pseudo-Q(Jiang et al., 2022) | 56.02 | 58.25 | 54.13 | 38.88 | 45.06 | 32.13 | 46.25 | 47.44 | |
| Pseudo-Q-adapted | 59.44 | 61.87 | 56.47 | 42.76 | 48.27 | 36.90 | 50.35 | 49.68 | |
| SGEPG(Ours) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 | 61.33 | 60.61 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| w/o SG | 63.45 | 63.9 | 63.47 | 43.8 | 45.81 | 44.14 |
| w/o ARS | 65.24 | 63.81 | 64.65 | 48.5 | 50.47 | 46.02 |
| w/o DDRQ | 68.82 | 72.46 | 64.99 | 53.78 | 58.26 | 46.14 |
| SGEPG | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| replace rate (α) | 0% | 25% | 37.5% | 50% | 75% |
| val | 53.78 | 54.29 | 54.94 | 55.21 | 55.33 |
| testA | 58.26 | 60.15 | 58.99 | 60.67 | 60.9 |
| testB | 46.14 | 46.41 | 46.7 | 46.76 | 46.8 |
| Expression number per object | Object number per image | ||||||||
| 1 | 2 | 3 | 4 | 1 | 2 | 3 | 4 | 5 | |
| SGEPG | 66.17 | 68.21 | 68.6 | 68.81 | 62.3 | 67.61 | 68.82 | 68.6 | 68.02 |
| RefCOCO | RefCOCO+ | |||||
| val | testA | testB | val | testA | testB | |
| preliminary scene graph | 36.23 | 45.31 | 25.32 | 37.21 | 46.11 | 26.55 |
| Core scene graph (SGEPG) | 69.61 | 72.34 | 65.67 | 55.21 | 60.67 | 46.76 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 73.49 | 64.83 | 57.05 | 51.84 | 47.16 | 42.06 | 35.96 | 53.20 |
| UIE | 74.45 | 66.09 | 58.24 | 53.01 | 48.24 | 45.77 | 48.41 | 56.32 | |
| Ent-F1 | AdaKGC | 74.45 | 66.05 | 58.26 | 53.06 | 48.39 | 45.92 | 48.57 | 56.39 |
| Relation (NYT) | TANL | 90.13 | 83.11 | 76.63 | 73.03 | 67.03 | 63.40 | 47.13 | 71.49 |
| UIE | 90.38 | 81.76 | 74.27 | 71.61 | 65.30 | 62.95 | 59.81 | 72.30 | |
| Rel-S F1 | AdaKGC | 90.18 | 81.65 | 74.13 | 71.48 | 65.10 | 62.75 | 59.77 | 72.15 |
| Event Trigger (ACE2005) | TEXT2EVENT | 67.10 | 55.81 | 50.38 | 52.99 | 45.82 | 41.39 | 41.65 | 50.73 |
| UIE | 70.94 | 60.00 | 57.01 | 62.52 | 60.39 | 54.90 | 53.62 | 59.91 | |
| Evt Tri F1 | AdaKGC | 70.57 | 59.75 | 56.50 | 62.31 | 60.57 | 55.49 | 54.32 | 59.93 |
| Event Argument (ACE2005) | TEXT2EVENT | 49.32 | 37.83 | 33.43 | 35.49 | 30.67 | 27.49 | 27.79 | 34.57 |
| UIE | 50.70 | 41.45 | 39.66 | 44.14 | 43.65 | 38.86 | 37.77 | 42.32 | |
| Evt Arg F1 | AdaKGC | 51.87 | 42.74 | 40.68 | 45.16 | 43.97 | 40.10 | 39.05 | 43.37 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 68.37 | 58.50 | 54.51 | 52.46 | 48.67 | 44.13 | 41.51 | 52.59 |
| UIE | 69.26 | 59.80 | 55.71 | 53.66 | 49.55 | 46.97 | 48.11 | 54.72 | |
| Ent-F1 | AdaKGC | 69.48 | 59.97 | 55.94 | 53.89 | 49.92 | 47.44 | 48.43 | 55.01 |
| Relation (NYT) | TANL | 88.67 | 81.59 | 76.48 | 72.22 | 72.35 | 61.76 | 57.87 | 72.99 |
| UIE | 90.17 | 83.14 | 77.81 | 71.45 | 72.97 | 65.57 | 63.12 | 74.89 | |
| Rel-S F1 | AdaKGC | 90.07 | 83.06 | 77.68 | 71.38 | 72.94 | 65.73 | 63.32 | 74.88 |
| Event Trigger (ACE2005) | TEXT2EVENT | 69.26 | 56.99 | 53.32 | 46.03 | 40.44 | 56.86 | 48.84 | 53.11 |
| UIE | 74.69 | 66.34 | 63.12 | 63.21 | 59.86 | 53.18 | 53.26 | 61.95 | |
| Evt Tri F1 | AdaKGC | 74.84 | 66.99 | 63.28 | 63.07 | 60.94 | 54.72 | 54.80 | 62.66 |
| Event Argument (ACE2005) | TEXT2EVENT | 50.32 | 38.14 | 35.79 | 31.94 | 28.79 | 37.75 | 33.54 | 36.61 |
| UIE | 51.94 | 45.22 | 42.97 | 43.12 | 40.66 | 36.66 | 36.17 | 42.39 | |
| Evt Arg F1 | AdaKGC | 55.08 | 48.58 | 46.19 | 46.27 | 45.22 | 40.52 | 40.41 | 46.04 |
| Error Analysis | Input Example | Gold Type(s) | Predicted Type(s) | Proportion |
| Weak Transfer | Kelly, who declined to talk to reporters here, travel to Tokyo Sunday for talks with Japanese officials. | Meet[travel] | Transport[travel] | 38% |
| Inheritance Deficiency | He started his entertainment career at ABC, where he is credited with creating the ‘movie of the week’ concept. | Start-Position[started] | Personnel[started] | 24% |
| Relevance Neglect | Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Transport[arrived]Meet[brief] | Transport[arrived] | 21% |
| Class Imbalance | Within weeks he was arrested and charged with sodomising an official driver several years previously and with abusing his powers to cover up the offence. | Arrest-Jail[arrested]Charge-Indict[charged] | Transport[arrested] | 13% |
| Potential Annotation | Anne-Marie will get the couple’s 19-room home in New York state, which was on the market last year for 21.5 million dollars, as well as their fine art collection. | None Event | Transfer-ownership[get] | 4% |
| #Maj | #Sub | #Train | #Val | #Test | |
| NERD | 8 | 66 | 131,767 | 18,824 | 37,648 |
| NYT | 4 | 24 | 56,196 | 5,000 | 5,000 |
| ACE-2005 | 8 | 33 | 19,216 | 901 | 676 |
| Taxonomy | Horizontal | Vertical | Hybrid | Analogous | ||||||||
| #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | |
| NERD | 30 | 6 | 7 | 30 | 6 | 7 | 30 | 6 | 7 | 66 | - | 7 |
| NYT | 10 | 2 | 8 | 10 | 2 | 8 | 10 | 2 | 8 | 24 | - | 7 |
| ACE-2005 | 15 | 3 | 7 | 15 | 3 | 7 | 15 | 3 | 7 | 33 | - | 7 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity | TANL | 68.34 | 62.70 | 57.22 | 52.94 | 47.73 | 42.24 | 36.09 | 52.47 |
| (NERD) | UIE | 67.58 | 63.39 | 60.37 | 58.00 | 58.17 | 54.26 | 49.53 | 58.76 |
| Ent-F1 | AdaKGC | 68.22 | 63.87 | 61.01 | 58.18 | 58.76 | 54.58 | 49.81 | 59.20 |
| Relation | TANL | 89.80 | 86.82 | 83.51 | 77.99 | 73.61 | 73.32 | 66.37 | 78.77 |
| (NYT) | UIE | 89.66 | 86.56 | 83.27 | 77.80 | 73.30 | 73.27 | 66.33 | 78.60 |
| Rel-S F1 | AdaKGC | 89.17 | 86.12 | 82.76 | 77.33 | 72.85 | 72.91 | 65.84 | 78.14 |
| Event Trigger | TEXT2EVENT | 64.40 | 60.39 | 53.96 | 52.64 | 40.83 | 32.45 | 29.58 | 47.75 |
| (ACE2005) | UIE | 69.69 | 64.53 | 62.10 | 58.47 | 49.76 | 42.51 | 39.31 | 55.20 |
| Evt Tri F1 | AdaKGC | 69.63 | 64.77 | 61.86 | 58.31 | 50.08 | 43.07 | 40.40 | 55.45 |
| Event Argument | TEXT2EVENT | 45.88 | 40.96 | 34.21 | 33.40 | 24.49 | 19.62 | 16.77 | 30.76 |
| (ACE2005) | UIE | 49.96 | 42.67 | 40.50 | 36.56 | 28.70 | 24.06 | 20.47 | 34.70 |
| Evt Arg F1 | AdaKGC | 51.74 | 44.56 | 42.48 | 32.61 | 30.52 | 25.41 | 22.10 | 35.63 |
| Input: The Belgrade district court said that Markovic will be tried along with 10 other Milosevic-era officials who face similar charges of ‘inappropriate use of state property’ that carry a sentence of up to five years in jail. | Labels |
| Iteration 1 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "phone write", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 2 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 3 schema: "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite" | Sentence[sentence] |
| Iteration 4 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite", "pardon" | Sentence[sentence] |
| Iteration 5 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "exe-cute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] |
| Iteration 6 schema: "end position", "attack", "start position", "charge indirect", "trans-fer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "par-don" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indirect", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Input: Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Labels |
| Iteration 1 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "contact", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[arrived] Contact[brief] |
| Iteration 2 schema: "personnel", "attack", "justice", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 3 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 4 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 5 schema: "personnel", "attack", "charge indict", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organi-zation", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demon-strate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Input: The charismatic leader of Turkey's governing party was named prime minister Tuesday, a step that probably boosts chances that the United States will get permission to deploy troops in the country along Iraq's northern border. | Labels |
| Iteration 1 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 2 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 3 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 4 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 5 schema: "attack", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Personnel[named] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Elect[named] |
| Input: Webb also said details of the breakdowns of the Welches' previous marriages were likely to come up, and cited reports of alleged extramarital affairs by both. | Labels |
| Iteration 1 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Divorce[breakdowns] \nMarry[marriages] |
| Iteration 2 schema: "end position", "attack", "begin", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 3 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradITE", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 4 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 5 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "condemn", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 6 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 7 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "hurt", "communication", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "wed", "extradite", "pardon" | Separate[breakdowns] \nWed[marriages] |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 73.49 | 64.83 | 57.05 | 51.84 | 47.16 | 42.06 | 35.96 | 53.20 |
| UIE | 74.45 | 66.09 | 58.24 | 53.01 | 48.24 | 45.77 | 48.41 | 56.32 | |
| Ent-F1 | AdaKGC | 74.45 | 66.05 | 58.26 | 53.06 | 48.39 | 45.92 | 48.57 | 56.39 |
| Relation (NYT) | TANL | 90.13 | 83.11 | 76.63 | 73.03 | 67.03 | 63.40 | 47.13 | 71.49 |
| UIE | 90.38 | 81.76 | 74.27 | 71.61 | 65.30 | 62.95 | 59.81 | 72.30 | |
| Rel-S F1 | AdaKGC | 90.18 | 81.65 | 74.13 | 71.48 | 65.10 | 62.75 | 59.77 | 72.15 |
| Event Trigger (ACE2005) | TEXT2EVENT | 67.10 | 55.81 | 50.38 | 52.99 | 45.82 | 41.39 | 41.65 | 50.73 |
| UIE | 70.94 | 60.00 | 57.01 | 62.52 | 60.39 | 54.90 | 53.62 | 59.91 | |
| Evt Tri F1 | AdaKGC | 70.57 | 59.75 | 56.50 | 62.31 | 60.57 | 55.49 | 54.32 | 59.93 |
| Event Argument (ACE2005) | TEXT2EVENT | 49.32 | 37.83 | 33.43 | 35.49 | 30.67 | 27.49 | 27.79 | 34.57 |
| UIE | 50.70 | 41.45 | 39.66 | 44.14 | 43.65 | 38.86 | 37.77 | 42.32 | |
| Evt Arg F1 | AdaKGC | 51.87 | 42.74 | 40.68 | 45.16 | 43.97 | 40.10 | 39.05 | 43.37 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 68.37 | 58.50 | 54.51 | 52.46 | 48.67 | 44.13 | 41.51 | 52.59 |
| UIE | 69.26 | 59.80 | 55.71 | 53.66 | 49.55 | 46.97 | 48.11 | 54.72 | |
| Ent-F1 | AdaKGC | 69.48 | 59.97 | 55.94 | 53.89 | 49.92 | 47.44 | 48.43 | 55.01 |
| Relation (NYT) | TANL | 88.67 | 81.59 | 76.48 | 72.22 | 72.35 | 61.76 | 57.87 | 72.99 |
| UIE | 90.17 | 83.14 | 77.81 | 71.45 | 72.97 | 65.57 | 63.12 | 74.89 | |
| Rel-S F1 | AdaKGC | 90.07 | 83.06 | 77.68 | 71.38 | 72.94 | 65.73 | 63.32 | 74.88 |
| Event Trigger (ACE2005) | TEXT2EVENT | 69.26 | 56.99 | 53.32 | 46.03 | 40.44 | 56.86 | 48.84 | 53.11 |
| UIE | 74.69 | 66.34 | 63.12 | 63.21 | 59.86 | 53.18 | 53.26 | 61.95 | |
| Evt Tri F1 | AdaKGC | 74.84 | 66.99 | 63.28 | 63.07 | 60.94 | 54.72 | 54.80 | 62.66 |
| Event Argument (ACE2005) | TEXT2EVENT | 50.32 | 38.14 | 35.79 | 31.94 | 28.79 | 37.75 | 33.54 | 36.61 |
| UIE | 51.94 | 45.22 | 42.97 | 43.12 | 40.66 | 36.66 | 36.17 | 42.39 | |
| Evt Arg F1 | AdaKGC | 55.08 | 48.58 | 46.19 | 46.27 | 45.22 | 40.52 | 40.41 | 46.04 |
| Error Analysis | Input Example | Gold Type(s) | Predicted Type(s) | Proportion |
| Weak Transfer | Kelly, who declined to talk to reporters here, travel to Tokyo Sunday for talks with Japanese officials. | Meet[travel] | Transport[travel] | 38% |
| Inheritance Deficiency | He started his entertainment career at ABC, where he is credited with creating the ‘movie of the week’ concept. | Start-Position[started] | Personnel[started] | 24% |
| Relevance Neglect | Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Transport[arrived]Meet[brief] | Transport[arrived] | 21% |
| Class Imbalance | Within weeks he was arrested and charged with sodomising an official driver several years previously and with abusing his powers to cover up the offence. | Arrest-Jail[arrested]Charge-Indict[charged] | Transport[arrested] | 13% |
| Potential Annotation | Anne-Marie will get the couple’s 19-room home in New York state, which was on the market last year for 21.5 million dollars, as well as their fine art collection. | None Event | Transfer-ownership[get] | 4% |
| #Maj | #Sub | #Train | #Val | #Test | |
| NERD | 8 | 66 | 131,767 | 18,824 | 37,648 |
| NYT | 4 | 24 | 56,196 | 5,000 | 5,000 |
| ACE-2005 | 8 | 33 | 19,216 | 901 | 676 |
| Taxonomy | Horizontal | Vertical | Hybrid | Analogous | ||||||||
| #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | |
| NERD | 30 | 6 | 7 | 30 | 6 | 7 | 30 | 6 | 7 | 66 | - | 7 |
| NYT | 10 | 2 | 8 | 10 | 2 | 8 | 10 | 2 | 8 | 24 | - | 7 |
| ACE-2005 | 15 | 3 | 7 | 15 | 3 | 7 | 15 | 3 | 7 | 33 | - | 7 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity | TANL | 68.34 | 62.70 | 57.22 | 52.94 | 47.73 | 42.24 | 36.09 | 52.47 |
| (NERD) | UIE | 67.58 | 63.39 | 60.37 | 58.00 | 58.17 | 54.26 | 49.53 | 58.76 |
| Ent-F1 | AdaKGC | 68.22 | 63.87 | 61.01 | 58.18 | 58.76 | 54.58 | 49.81 | 59.20 |
| Relation | TANL | 89.80 | 86.82 | 83.51 | 77.99 | 73.61 | 73.32 | 66.37 | 78.77 |
| (NYT) | UIE | 89.66 | 86.56 | 83.27 | 77.80 | 73.30 | 73.27 | 66.33 | 78.60 |
| Rel-S F1 | AdaKGC | 89.17 | 86.12 | 82.76 | 77.33 | 72.85 | 72.91 | 65.84 | 78.14 |
| Event Trigger | TEXT2EVENT | 64.40 | 60.39 | 53.96 | 52.64 | 40.83 | 32.45 | 29.58 | 47.75 |
| (ACE2005) | UIE | 69.69 | 64.53 | 62.10 | 58.47 | 49.76 | 42.51 | 39.31 | 55.20 |
| Evt Tri F1 | AdaKGC | 69.63 | 64.77 | 61.86 | 58.31 | 50.08 | 43.07 | 40.40 | 55.45 |
| Event Argument | TEXT2EVENT | 45.88 | 40.96 | 34.21 | 33.40 | 24.49 | 19.62 | 16.77 | 30.76 |
| (ACE2005) | UIE | 49.96 | 42.67 | 40.50 | 36.56 | 28.70 | 24.06 | 20.47 | 34.70 |
| Evt Arg F1 | AdaKGC | 51.74 | 44.56 | 42.48 | 32.61 | 30.52 | 25.41 | 22.10 | 35.63 |
| Input: The Belgrade district court said that Markovic will be tried along with 10 other Milosevic-era officials who face similar charges of ‘inappropriate use of state property’ that carry a sentence of up to five years in jail. | Labels |
| Iteration 1 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "phone write", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 2 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 3 schema: "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite" | Sentence[sentence] |
| Iteration 4 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite", "pardon" | Sentence[sentence] |
| Iteration 5 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "exe-cute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] |
| Iteration 6 schema: "end position", "attack", "start position", "charge indirect", "trans-fer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "par-don" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indirect", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Input: Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Labels |
| Iteration 1 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "contact", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[arrived] Contact[brief] |
| Iteration 2 schema: "personnel", "attack", "justice", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 3 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 4 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 5 schema: "personnel", "attack", "charge indict", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organi-zation", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demon-strate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Input: The charismatic leader of Turkey's governing party was named prime minister Tuesday, a step that probably boosts chances that the United States will get permission to deploy troops in the country along Iraq's northern border. | Labels |
| Iteration 1 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 2 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 3 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 4 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 5 schema: "attack", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Personnel[named] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Elect[named] |
| Input: Webb also said details of the breakdowns of the Welches' previous marriages were likely to come up, and cited reports of alleged extramarital affairs by both. | Labels |
| Iteration 1 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Divorce[breakdowns] \nMarry[marriages] |
| Iteration 2 schema: "end position", "attack", "begin", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 3 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradITE", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 4 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 5 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "condemn", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 6 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 7 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "hurt", "communication", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "wed", "extradite", "pardon" | Separate[breakdowns] \nWed[marriages] |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 73.49 | 64.83 | 57.05 | 51.84 | 47.16 | 42.06 | 35.96 | 53.20 |
| UIE | 74.45 | 66.09 | 58.24 | 53.01 | 48.24 | 45.77 | 48.41 | 56.32 | |
| Ent-F1 | AdaKGC | 74.45 | 66.05 | 58.26 | 53.06 | 48.39 | 45.92 | 48.57 | 56.39 |
| Relation (NYT) | TANL | 90.13 | 83.11 | 76.63 | 73.03 | 67.03 | 63.40 | 47.13 | 71.49 |
| UIE | 90.38 | 81.76 | 74.27 | 71.61 | 65.30 | 62.95 | 59.81 | 72.30 | |
| Rel-S F1 | AdaKGC | 90.18 | 81.65 | 74.13 | 71.48 | 65.10 | 62.75 | 59.77 | 72.15 |
| Event Trigger (ACE2005) | TEXT2EVENT | 67.10 | 55.81 | 50.38 | 52.99 | 45.82 | 41.39 | 41.65 | 50.73 |
| UIE | 70.94 | 60.00 | 57.01 | 62.52 | 60.39 | 54.90 | 53.62 | 59.91 | |
| Evt Tri F1 | AdaKGC | 70.57 | 59.75 | 56.50 | 62.31 | 60.57 | 55.49 | 54.32 | 59.93 |
| Event Argument (ACE2005) | TEXT2EVENT | 49.32 | 37.83 | 33.43 | 35.49 | 30.67 | 27.49 | 27.79 | 34.57 |
| UIE | 50.70 | 41.45 | 39.66 | 44.14 | 43.65 | 38.86 | 37.77 | 42.32 | |
| Evt Arg F1 | AdaKGC | 51.87 | 42.74 | 40.68 | 45.16 | 43.97 | 40.10 | 39.05 | 43.37 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 68.37 | 58.50 | 54.51 | 52.46 | 48.67 | 44.13 | 41.51 | 52.59 |
| UIE | 69.26 | 59.80 | 55.71 | 53.66 | 49.55 | 46.97 | 48.11 | 54.72 | |
| Ent-F1 | AdaKGC | 69.48 | 59.97 | 55.94 | 53.89 | 49.92 | 47.44 | 48.43 | 55.01 |
| Relation (NYT) | TANL | 88.67 | 81.59 | 76.48 | 72.22 | 72.35 | 61.76 | 57.87 | 72.99 |
| UIE | 90.17 | 83.14 | 77.81 | 71.45 | 72.97 | 65.57 | 63.12 | 74.89 | |
| Rel-S F1 | AdaKGC | 90.07 | 83.06 | 77.68 | 71.38 | 72.94 | 65.73 | 63.32 | 74.88 |
| Event Trigger (ACE2005) | TEXT2EVENT | 69.26 | 56.99 | 53.32 | 46.03 | 40.44 | 56.86 | 48.84 | 53.11 |
| UIE | 74.69 | 66.34 | 63.12 | 63.21 | 59.86 | 53.18 | 53.26 | 61.95 | |
| Evt Tri F1 | AdaKGC | 74.84 | 66.99 | 63.28 | 63.07 | 60.94 | 54.72 | 54.80 | 62.66 |
| Event Argument (ACE2005) | TEXT2EVENT | 50.32 | 38.14 | 35.79 | 31.94 | 28.79 | 37.75 | 33.54 | 36.61 |
| UIE | 51.94 | 45.22 | 42.97 | 43.12 | 40.66 | 36.66 | 36.17 | 42.39 | |
| Evt Arg F1 | AdaKGC | 55.08 | 48.58 | 46.19 | 46.27 | 45.22 | 40.52 | 40.41 | 46.04 |
| Error Analysis | Input Example | Gold Type(s) | Predicted Type(s) | Proportion |
| Weak Transfer | Kelly, who declined to talk to reporters here, travel to Tokyo Sunday for talks with Japanese officials. | Meet[travel] | Transport[travel] | 38% |
| Inheritance Deficiency | He started his entertainment career at ABC, where he is credited with creating the ‘movie of the week’ concept. | Start-Position[started] | Personnel[started] | 24% |
| Relevance Neglect | Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Transport[arrived]Meet[brief] | Transport[arrived] | 21% |
| Class Imbalance | Within weeks he was arrested and charged with sodomising an official driver several years previously and with abusing his powers to cover up the offence. | Arrest-Jail[arrested]Charge-Indict[charged] | Transport[arrested] | 13% |
| Potential Annotation | Anne-Marie will get the couple’s 19-room home in New York state, which was on the market last year for 21.5 million dollars, as well as their fine art collection. | None Event | Transfer-ownership[get] | 4% |
| #Maj | #Sub | #Train | #Val | #Test | |
| NERD | 8 | 66 | 131,767 | 18,824 | 37,648 |
| NYT | 4 | 24 | 56,196 | 5,000 | 5,000 |
| ACE-2005 | 8 | 33 | 19,216 | 901 | 676 |
| Taxonomy | Horizontal | Vertical | Hybrid | Analogous | ||||||||
| #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | |
| NERD | 30 | 6 | 7 | 30 | 6 | 7 | 30 | 6 | 7 | 66 | - | 7 |
| NYT | 10 | 2 | 8 | 10 | 2 | 8 | 10 | 2 | 8 | 24 | - | 7 |
| ACE-2005 | 15 | 3 | 7 | 15 | 3 | 7 | 15 | 3 | 7 | 33 | - | 7 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity | TANL | 68.34 | 62.70 | 57.22 | 52.94 | 47.73 | 42.24 | 36.09 | 52.47 |
| (NERD) | UIE | 67.58 | 63.39 | 60.37 | 58.00 | 58.17 | 54.26 | 49.53 | 58.76 |
| Ent-F1 | AdaKGC | 68.22 | 63.87 | 61.01 | 58.18 | 58.76 | 54.58 | 49.81 | 59.20 |
| Relation | TANL | 89.80 | 86.82 | 83.51 | 77.99 | 73.61 | 73.32 | 66.37 | 78.77 |
| (NYT) | UIE | 89.66 | 86.56 | 83.27 | 77.80 | 73.30 | 73.27 | 66.33 | 78.60 |
| Rel-S F1 | AdaKGC | 89.17 | 86.12 | 82.76 | 77.33 | 72.85 | 72.91 | 65.84 | 78.14 |
| Event Trigger | TEXT2EVENT | 64.40 | 60.39 | 53.96 | 52.64 | 40.83 | 32.45 | 29.58 | 47.75 |
| (ACE2005) | UIE | 69.69 | 64.53 | 62.10 | 58.47 | 49.76 | 42.51 | 39.31 | 55.20 |
| Evt Tri F1 | AdaKGC | 69.63 | 64.77 | 61.86 | 58.31 | 50.08 | 43.07 | 40.40 | 55.45 |
| Event Argument | TEXT2EVENT | 45.88 | 40.96 | 34.21 | 33.40 | 24.49 | 19.62 | 16.77 | 30.76 |
| (ACE2005) | UIE | 49.96 | 42.67 | 40.50 | 36.56 | 28.70 | 24.06 | 20.47 | 34.70 |
| Evt Arg F1 | AdaKGC | 51.74 | 44.56 | 42.48 | 32.61 | 30.52 | 25.41 | 22.10 | 35.63 |
| Input: The Belgrade district court said that Markovic will be tried along with 10 other Milosevic-era officials who face similar charges of ‘inappropriate use of state property’ that carry a sentence of up to five years in jail. | Labels |
| Iteration 1 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "phone write", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 2 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 3 schema: "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite" | Sentence[sentence] |
| Iteration 4 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite", "pardon" | Sentence[sentence] |
| Iteration 5 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "exe-cute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] |
| Iteration 6 schema: "end position", "attack", "start position", "charge indirect", "trans-fer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "par-don" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indirect", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Input: Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Labels |
| Iteration 1 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "contact", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[arrived] Contact[brief] |
| Iteration 2 schema: "personnel", "attack", "justice", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 3 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 4 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 5 schema: "personnel", "attack", "charge indict", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organi-zation", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demon-strate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Input: The charismatic leader of Turkey's governing party was named prime minister Tuesday, a step that probably boosts chances that the United States will get permission to deploy troops in the country along Iraq's northern border. | Labels |
| Iteration 1 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 2 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 3 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 4 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 5 schema: "attack", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Personnel[named] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Elect[named] |
| Input: Webb also said details of the breakdowns of the Welches' previous marriages were likely to come up, and cited reports of alleged extramarital affairs by both. | Labels |
| Iteration 1 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Divorce[breakdowns] +Marry[marriages] |
| Iteration 2 schema: "end position", "attack", "begin", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Separate[breakdowns] +Marry[marriages] |
| Iteration 3 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradITE", "pardon" | Separate[breakdowns] +Marry[marriages] |
| Iteration 4 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] +Marry[marriages] |
| Iteration 5 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "condemn", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] +Marry[marriages] |
| Iteration 6 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] +Marry[marriages] |
| Iteration 7 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "hurt", "communication", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "wed", "extradite", "pardon" | Separate[breakdowns] +Wed[marriages] |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 73.49 | 64.83 | 57.05 | 51.84 | 47.16 | 42.06 | 35.96 | 53.20 |
| UIE | 74.45 | 66.09 | 58.24 | 53.01 | 48.24 | 45.77 | 48.41 | 56.32 | |
| Ent-F1 | AdaKGC | 74.45 | 66.05 | 58.26 | 53.06 | 48.39 | 45.92 | 48.57 | 56.39 |
| Relation (NYT) | TANL | 90.13 | 83.11 | 76.63 | 73.03 | 67.03 | 63.40 | 47.13 | 71.49 |
| UIE | 90.38 | 81.76 | 74.27 | 71.61 | 65.30 | 62.95 | 59.81 | 72.30 | |
| Rel-S F1 | AdaKGC | 90.18 | 81.65 | 74.13 | 71.48 | 65.10 | 62.75 | 59.77 | 72.15 |
| Event Trigger (ACE2005) | TEXT2EVENT | 67.10 | 55.81 | 50.38 | 52.99 | 45.82 | 41.39 | 41.65 | 50.73 |
| UIE | 70.94 | 60.00 | 57.01 | 62.52 | 60.39 | 54.90 | 53.62 | 59.91 | |
| Evt Tri F1 | AdaKGC | 70.57 | 59.75 | 56.50 | 62.31 | 60.57 | 55.49 | 54.32 | 59.93 |
| Event Argument (ACE2005) | TEXT2EVENT | 49.32 | 37.83 | 33.43 | 35.49 | 30.67 | 27.49 | 27.79 | 34.57 |
| UIE | 50.70 | 41.45 | 39.66 | 44.14 | 43.65 | 38.86 | 37.77 | 42.32 | |
| Evt Arg F1 | AdaKGC | 51.87 | 42.74 | 40.68 | 45.16 | 43.97 | 40.10 | 39.05 | 43.37 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity (NERD) | TANL | 68.37 | 58.50 | 54.51 | 52.46 | 48.67 | 44.13 | 41.51 | 52.59 |
| UIE | 69.26 | 59.80 | 55.71 | 53.66 | 49.55 | 46.97 | 48.11 | 54.72 | |
| Ent-F1 | AdaKGC | 69.48 | 59.97 | 55.94 | 53.89 | 49.92 | 47.44 | 48.43 | 55.01 |
| Relation (NYT) | TANL | 88.67 | 81.59 | 76.48 | 72.22 | 72.35 | 61.76 | 57.87 | 72.99 |
| UIE | 90.17 | 83.14 | 77.81 | 71.45 | 72.97 | 65.57 | 63.12 | 74.89 | |
| Rel-S F1 | AdaKGC | 90.07 | 83.06 | 77.68 | 71.38 | 72.94 | 65.73 | 63.32 | 74.88 |
| Event Trigger (ACE2005) | TEXT2EVENT | 69.26 | 56.99 | 53.32 | 46.03 | 40.44 | 56.86 | 48.84 | 53.11 |
| UIE | 74.69 | 66.34 | 63.12 | 63.21 | 59.86 | 53.18 | 53.26 | 61.95 | |
| Evt Tri F1 | AdaKGC | 74.84 | 66.99 | 63.28 | 63.07 | 60.94 | 54.72 | 54.80 | 62.66 |
| Event Argument (ACE2005) | TEXT2EVENT | 50.32 | 38.14 | 35.79 | 31.94 | 28.79 | 37.75 | 33.54 | 36.61 |
| UIE | 51.94 | 45.22 | 42.97 | 43.12 | 40.66 | 36.66 | 36.17 | 42.39 | |
| Evt Arg F1 | AdaKGC | 55.08 | 48.58 | 46.19 | 46.27 | 45.22 | 40.52 | 40.41 | 46.04 |
| Error Analysis | Input Example | Gold Type(s) | Predicted Type(s) | Proportion |
| Weak Transfer | Kelly, who declined to talk to reporters here, travel to Tokyo Sunday for talks with Japanese officials. | Meet[travel] | Transport[travel] | 38% |
| Inheritance Deficiency | He started his entertainment career at ABC, where he is credited with creating the ‘movie of the week’ concept. | Start-Position[started] | Personnel[started] | 24% |
| Relevance Neglect | Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Transport[arrived]Meet[brief] | Transport[arrived] | 21% |
| Class Imbalance | Within weeks he was arrested and charged with sodomising an official driver several years previously and with abusing his powers to cover up the offence. | Arrest-Jail[arrested]Charge-Indict[charged] | Transport[arrested] | 13% |
| Potential Annotation | Anne-Marie will get the couple’s 19-room home in New York state, which was on the market last year for 21.5 million dollars, as well as their fine art collection. | None Event | Transfer-ownership[get] | 4% |
| #Maj | #Sub | #Train | #Val | #Test | |
| NERD | 8 | 66 | 131,767 | 18,824 | 37,648 |
| NYT | 4 | 24 | 56,196 | 5,000 | 5,000 |
| ACE-2005 | 8 | 33 | 19,216 | 901 | 676 |
| Taxonomy | Horizontal | Vertical | Hybrid | Analogous | ||||||||
| #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | #Init | #Add | #N | |
| NERD | 30 | 6 | 7 | 30 | 6 | 7 | 30 | 6 | 7 | 66 | - | 7 |
| NYT | 10 | 2 | 8 | 10 | 2 | 8 | 10 | 2 | 8 | 24 | - | 7 |
| ACE-2005 | 15 | 3 | 7 | 15 | 3 | 7 | 15 | 3 | 7 | 33 | - | 7 |
| Model | Iter 1 | Iter 2 | Iter 3 | Iter 4 | Iter 5 | Iter 6 | Iter 7 | AVE | |
| Entity | TANL | 68.34 | 62.70 | 57.22 | 52.94 | 47.73 | 42.24 | 36.09 | 52.47 |
| (NERD) | UIE | 67.58 | 63.39 | 60.37 | 58.00 | 58.17 | 54.26 | 49.53 | 58.76 |
| Ent-F1 | AdaKGC | 68.22 | 63.87 | 61.01 | 58.18 | 58.76 | 54.58 | 49.81 | 59.20 |
| Relation | TANL | 89.80 | 86.82 | 83.51 | 77.99 | 73.61 | 73.32 | 66.37 | 78.77 |
| (NYT) | UIE | 89.66 | 86.56 | 83.27 | 77.80 | 73.30 | 73.27 | 66.33 | 78.60 |
| Rel-S F1 | AdaKGC | 89.17 | 86.12 | 82.76 | 77.33 | 72.85 | 72.91 | 65.84 | 78.14 |
| Event Trigger | TEXT2EVENT | 64.40 | 60.39 | 53.96 | 52.64 | 40.83 | 32.45 | 29.58 | 47.75 |
| (ACE2005) | UIE | 69.69 | 64.53 | 62.10 | 58.47 | 49.76 | 42.51 | 39.31 | 55.20 |
| Evt Tri F1 | AdaKGC | 69.63 | 64.77 | 61.86 | 58.31 | 50.08 | 43.07 | 40.40 | 55.45 |
| Event Argument | TEXT2EVENT | 45.88 | 40.96 | 34.21 | 33.40 | 24.49 | 19.62 | 16.77 | 30.76 |
| (ACE2005) | UIE | 49.96 | 42.67 | 40.50 | 36.56 | 28.70 | 24.06 | 20.47 | 34.70 |
| Evt Arg F1 | AdaKGC | 51.74 | 44.56 | 42.48 | 32.61 | 30.52 | 25.41 | 22.10 | 35.63 |
| Input: The Belgrade district court said that Markovic will be tried along with 10 other Milosevic-era officials who face similar charges of ‘inappropriate use of state property’ that carry a sentence of up to five years in jail. | Labels |
| Iteration 1 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "phone write", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 2 schema: "attack", "start position", "transfer ownership", "be born", "sentence", "die", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "marry", "extradite" | Sentence[sentence] |
| Iteration 3 schema: "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite" | Sentence[sentence] |
| Iteration 4 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "fine", "convict", "end organization", "sue", "acquit", "execute", "marry", "extradite", "pardon" | Sentence[sentence] |
| Iteration 5 schema: "end position", "attack", "start position", "transfer money", "transfer ownership", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "exe-cute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] |
| Iteration 6 schema: "end position", "attack", "start position", "charge indirect", "trans-fer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "par-don" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indirect", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Sentence[sentence] Trial hearing[tried] Charge-Indict[charges] |
| Input: Kelly, the US assistant secretary for East Asia and Pacific Affairs, arrived in Seoul from Beijing Friday to brief Yoon, the foreign minister. | Labels |
| Iteration 1 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "contact", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[arrived] Contact[brief] |
| Iteration 2 schema: "personnel", "attack", "justice", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "business", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 3 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 4 schema: "personnel", "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 5 schema: "personnel", "attack", "charge indict", "transfer money", "trans-fer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organi-zation", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demon-strate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sent-ence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[arrived] Meet[brief] |
| Input: The charismatic leader of Turkey's governing party was named prime minister Tuesday, a step that probably boosts chances that the United States will get permission to deploy troops in the country along Iraq's northern border. | Labels |
| Iteration 1 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "life", "fine", "sue", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 2 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "meet", "life", "contact", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 3 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "contact", "merge organization", "business", "trial hearing", "fine", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 4 schema: "attack", "justice", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 5 schema: "attack", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "life", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy] |
| Iteration 6 schema: "end position", "attack", "personnel", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Personnel[named] |
| Iteration 7 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Transport[deploy]Elect[named] |
| Input: Webb also said details of the breakdowns of the Welches' previous marriages were likely to come up, and cited reports of alleged extramarital affairs by both. | Labels |
| Iteration 1 schema: "end position", "attack", "start position", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "transport", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "divorce", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Divorce[breakdowns] \nMarry[marriages] |
| Iteration 2 schema: "end position", "attack", "begin", "nominate", "charge indict", "transfer money", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 3 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "die", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "trial hearing", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "execute", "marry", "extradITE", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 4 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "sentence", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "meet", "injure", "phone write", "merge organization", "declare bankruptcy", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 5 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "transfer ownership", "release parole", "be born", "condemn", "pass away", "demonstrate", "arrest jail", "carry", "elect", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 6 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "injure", "phone write", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "marry", "extradite", "pardon" | Separate[breakdowns] \nMarry[marriages] |
| Iteration 7 schema: "end", "attack", "begin", "nominate", "prosecute", "remittance", "giveaway", "release parole", "be born", "condemn", "pass away", "parade", "arrest jail", "carry", "vote", "start organization", "encounter", "hurt", "communication", "merge organization", "go out of business", "attend the trial", "fine", "convict", "end organization", "sue", "separate", "acquit", "appeal", "perform", "wed", "extradite", "pardon" | Separate[breakdowns] \nWed[marriages] |
| Model | LLaMA-GPT4-7B | Vicuna-7B | Vicuna-13B | SAIL-7B | ||||||||
| Search | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG |
| CSQA | 48.4 | 47.7 | 49.6 | 44.9 | 45.6 | 47.6 | 50.6 | 51.1 | 50.9 | 51.5 | 51.0 | 51.8 |
| OBQA | 42.2 | 44.4 | 44.6 | 37.2 | 39.4 | 42.6 | 49.0 | 47.2 | 49.4 | 49.2 | 50.2 | 52.0 |
| ARC-C | 43.0 | 45.2 | 47.3 | 40.5 | 44.5 | 46.3 | 53.2 | 51.6 | 51.8 | 47.7 | 48.1 | 48.4 |
| Avg. | 44.5 | 45.8 | 47.2 | 40.9 | 43.3 | 45.5 | 51.0 | 50.0 | 50.7 | 49.5 | 49.8 | 50.7 |
| Search Effect | none | +1.3 | +2.7 | none | +2.4 | +4.6 | none | -1.0 | -0.3 | none | +0.3 | +1.2 |
| Model | Size | True | True * Info |
| No Search Augmentation | |||
| GPT-3 | 175B | 0.28 | 0.25 |
| LLaMA | 65B | 0.57 | 0.53 |
| LLaMA | 7B | 0.33 | 0.29 |
| Alpaca | 7B | 0.33 | 0.33 |
| Vicuna | 7B | 0.56 | 0.52 |
| W/ Search Augmentation | |||
| Vicuna | 13B | 0.71 | 0.69 |
| Vicuna | 7B | 0.68 | 0.65 |
| SAIL | 7B | 0.73 | 0.73 |
| Model | Metric | Climate | PubHealth | Fact Avg. | HSD | SBIC | Fairness Avg. | All Avg. |
| Vicuna-7B | Acc | 57.9 | 60.6 | 59.2 | 55.9 | 74.5 | 65.2 | 62.2 |
| F1 | 38.8 | 56.63 | 47.7 | 68.5 | 84.3 | 76.4 | 62.04 | |
| Vicuna-13B | Acc | 51.4 | 54.4 | 52.9 | 57.7 | 72.3 | 65.0 | 59.0 |
| F1 | 42.5 | 57.7 | 50.1 | 69.6 | 82.9 | 76.3 | 63.2 | |
| LLaMA-7B | Acc | 58.8 | 59.9 | 59.3 | 62.3 | 74.8 | 68.6 | 63.9 |
| F1 | 46.6 | 57.5 | 52.0 | 72.3 | 84.4 | 78.4 | 65.2 | |
| SAIL-7B | Acc | 63.5 | 69.2 | 66.4 | 70.1 | 76.4 | 73.2 | 69.8 |
| F1 | 51.0 | 63.6 | 57.3 | 75.1 | 83.9 | 79.5 | 68.4 |
| Model | Metric | Climate | PubHealth | Avg. |
| Vicuna-7B | Acc | 57.7 | 60.1 | 58.9 |
| Acc Diff | -0.2 | -0.5 | -0.3 | |
| F1 | 49.5 | 57.6 | 53.6 | |
| F1 Diff | +10.7 | +1.0 | +5.9 | |
| Vicuna-13B | Acc | 53.5 | 50.3 | 51.9 |
| Acc Diff | +2.1 | -4.1 | -1.0 | |
| F1 | 46.6 | 56.8 | 51.7 | |
| F1 Diff | +4.1 | -0.9 | +1.6 | |
| LLaMA-7B | Acc | 55.8 | 62.8 | 59.3 |
| Acc Diff | -3.0 | +2.9 | -0.1 | |
| F1 | 50.2 | 59.7 | 54.9 | |
| F1 Diff | +3.6 | +2.2 | +2.9 | |
| SAIL-7B | Acc | 65.8 | 70.7 | 68.3 |
| Acc Diff | +2.3 | +1.5 | +1.9 | |
| F1 | 55.2 | 64.5 | 59.9 | |
| F1 Diff | +4.2 | +0.9 | +2.5 |
| Models | Vicuna-7B-v1.1 | SAIL-7B |
| Novel \nVerbs | Include \nConsider | Calculate \nMatch \nRevolutionize \nCheck \nInclude \nIncrease |
| Count | 2 | 6 |
| Models | Avg. | Std. | Diversity |
| GPT-4 | 303.8 | 121.5 | 0.48 |
| ChatGPT | 135.1 | 63.6 | 0.56 |
| Vicuna-13B | 204.1 | 82.9 | 0.45 |
| Vicuna-7B | 196.5 | 90.3 | 0.45 |
| SAIL-7B + Search | 246.2 | 87.7 | 0.44 |
| SAIL-7B | 206.6 | 86.9 | 0.47 |
| Model | LLaMA-GPT4-7B | Vicuna-7B | Vicuna-13B | SAIL-7B | ||||||||
| Search | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG |
| CSQA | 48.4 | 47.7 | 49.6 | 44.9 | 45.6 | 47.6 | 50.6 | 51.1 | 50.9 | 51.5 | 51.0 | 51.8 |
| OBQA | 42.2 | 44.4 | 44.6 | 37.2 | 39.4 | 42.6 | 49.0 | 47.2 | 49.4 | 49.2 | 50.2 | 52.0 |
| ARC-C | 43.0 | 45.2 | 47.3 | 40.5 | 44.5 | 46.3 | 53.2 | 51.6 | 51.8 | 47.7 | 48.1 | 48.4 |
| Avg. | 44.5 | 45.8 | 47.2 | 40.9 | 43.3 | 45.5 | 51.0 | 50.0 | 50.7 | 49.5 | 49.8 | 50.7 |
| Search Effect | none | +1.3 | +2.7 | none | +2.4 | +4.6 | none | -1.0 | -0.3 | none | +0.3 | +1.2 |
| Model | Size | True | True * Info |
| No Search Augmentation | |||
| GPT-3 | 175B | 0.28 | 0.25 |
| LLaMA | 65B | 0.57 | 0.53 |
| LLaMA | 7B | 0.33 | 0.29 |
| Alpaca | 7B | 0.33 | 0.33 |
| Vicuna | 7B | 0.56 | 0.52 |
| W/ Search Augmentation | |||
| Vicuna | 13B | 0.71 | 0.69 |
| Vicuna | 7B | 0.68 | 0.65 |
| SAIL | 7B | 0.73 | 0.73 |
| Model | Metric | Climate | PubHealth | Fact Avg. | HSD | SBIC | Fairness Avg. | All Avg. |
| Vicuna-7B | Acc | 57.9 | 60.6 | 59.2 | 55.9 | 74.5 | 65.2 | 62.2 |
| F1 | 38.8 | 56.63 | 47.7 | 68.5 | 84.3 | 76.4 | 62.04 | |
| Vicuna-13B | Acc | 51.4 | 54.4 | 52.9 | 57.7 | 72.3 | 65.0 | 59.0 |
| F1 | 42.5 | 57.7 | 50.1 | 69.6 | 82.9 | 76.3 | 63.2 | |
| LLaMA-7B | Acc | 58.8 | 59.9 | 59.3 | 62.3 | 74.8 | 68.6 | 63.9 |
| F1 | 46.6 | 57.5 | 52.0 | 72.3 | 84.4 | 78.4 | 65.2 | |
| SAIL-7B | Acc | 63.5 | 69.2 | 66.4 | 70.1 | 76.4 | 73.2 | 69.8 |
| F1 | 51.0 | 63.6 | 57.3 | 75.1 | 83.9 | 79.5 | 68.4 |
| Model | Metric | Climate | PubHealth | Avg. |
| Vicuna-7B | Acc | 57.7 | 60.1 | 58.9 |
| Acc Diff | -0.2 | -0.5 | -0.3 | |
| F1 | 49.5 | 57.6 | 53.6 | |
| F1 Diff | +10.7 | +1.0 | +5.9 | |
| Vicuna-13B | Acc | 53.5 | 50.3 | 51.9 |
| Acc Diff | +2.1 | -4.1 | -1.0 | |
| F1 | 46.6 | 56.8 | 51.7 | |
| F1 Diff | +4.1 | -0.9 | +1.6 | |
| LLaMA-7B | Acc | 55.8 | 62.8 | 59.3 |
| Acc Diff | -3.0 | +2.9 | -0.1 | |
| F1 | 50.2 | 59.7 | 54.9 | |
| F1 Diff | +3.6 | +2.2 | +2.9 | |
| SAIL-7B | Acc | 65.8 | 70.7 | 68.3 |
| Acc Diff | +2.3 | +1.5 | +1.9 | |
| F1 | 55.2 | 64.5 | 59.9 | |
| F1 Diff | +4.2 | +0.9 | +2.5 |
| Models | Vicuna-7B-v1.1 | SAIL-7B |
| Novel \nVerbs | Include \nConsider | Calculate \nMatch \nRevolutionize \nCheck \nInclude \nIncrease |
| Count | 2 | 6 |
| Models | Avg. | Std. | Diversity |
| GPT-4 | 303.8 | 121.5 | 0.48 |
| ChatGPT | 135.1 | 63.6 | 0.56 |
| Vicuna-13B | 204.1 | 82.9 | 0.45 |
| Vicuna-7B | 196.5 | 90.3 | 0.45 |
| SAIL-7B + Search | 246.2 | 87.7 | 0.44 |
| SAIL-7B | 206.6 | 86.9 | 0.47 |
| Model | LLaMA-GPT4-7B | Vicuna-7B | Vicuna-13B | SAIL-7B | ||||||||
| Search | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG |
| CSQA | 48.4 | 47.7 | 49.6 | 44.9 | 45.6 | 47.6 | 50.6 | 51.1 | 50.9 | 51.5 | 51.0 | 51.8 |
| OBQA | 42.2 | 44.4 | 44.6 | 37.2 | 39.4 | 42.6 | 49.0 | 47.2 | 49.4 | 49.2 | 50.2 | 52.0 |
| ARC-C | 43.0 | 45.2 | 47.3 | 40.5 | 44.5 | 46.3 | 53.2 | 51.6 | 51.8 | 47.7 | 48.1 | 48.4 |
| Avg. | 44.5 | 45.8 | 47.2 | 40.9 | 43.3 | 45.5 | 51.0 | 50.0 | 50.7 | 49.5 | 49.8 | 50.7 |
| Search Effect | none | +1.3 | +2.7 | none | +2.4 | +4.6 | none | -1.0 | -0.3 | none | +0.3 | +1.2 |
| Model | Size | True | True * Info |
| No Search Augmentation | |||
| GPT-3 | 175B | 0.28 | 0.25 |
| LLaMA | 65B | 0.57 | 0.53 |
| LLaMA | 7B | 0.33 | 0.29 |
| Alpaca | 7B | 0.33 | 0.33 |
| Vicuna | 7B | 0.56 | 0.52 |
| W/ Search Augmentation | |||
| Vicuna | 13B | 0.71 | 0.69 |
| Vicuna | 7B | 0.68 | 0.65 |
| SAIL | 7B | 0.73 | 0.73 |
| Model | Metric | Climate | PubHealth | Fact Avg. | HSD | SBIC | Fairness Avg. | All Avg. |
| Vicuna-7B | Acc | 57.9 | 60.6 | 59.2 | 55.9 | 74.5 | 65.2 | 62.2 |
| F1 | 38.8 | 56.63 | 47.7 | 68.5 | 84.3 | 76.4 | 62.04 | |
| Vicuna-13B | Acc | 51.4 | 54.4 | 52.9 | 57.7 | 72.3 | 65.0 | 59.0 |
| F1 | 42.5 | 57.7 | 50.1 | 69.6 | 82.9 | 76.3 | 63.2 | |
| LLaMA-7B | Acc | 58.8 | 59.9 | 59.3 | 62.3 | 74.8 | 68.6 | 63.9 |
| F1 | 46.6 | 57.5 | 52.0 | 72.3 | 84.4 | 78.4 | 65.2 | |
| SAIL-7B | Acc | 63.5 | 69.2 | 66.4 | 70.1 | 76.4 | 73.2 | 69.8 |
| F1 | 51.0 | 63.6 | 57.3 | 75.1 | 83.9 | 79.5 | 68.4 |
| Model | Metric | Climate | PubHealth | Avg. |
| Vicuna-7B | Acc | 57.7 | 60.1 | 58.9 |
| Acc Diff | -0.2 | -0.5 | -0.3 | |
| F1 | 49.5 | 57.6 | 53.6 | |
| F1 Diff | +10.7 | +1.0 | +5.9 | |
| Vicuna-13B | Acc | 53.5 | 50.3 | 51.9 |
| Acc Diff | +2.1 | -4.1 | -1.0 | |
| F1 | 46.6 | 56.8 | 51.7 | |
| F1 Diff | +4.1 | -0.9 | +1.6 | |
| LLaMA-7B | Acc | 55.8 | 62.8 | 59.3 |
| Acc Diff | -3.0 | +2.9 | -0.1 | |
| F1 | 50.2 | 59.7 | 54.9 | |
| F1 Diff | +3.6 | +2.2 | +2.9 | |
| SAIL-7B | Acc | 65.8 | 70.7 | 68.3 |
| Acc Diff | +2.3 | +1.5 | +1.9 | |
| F1 | 55.2 | 64.5 | 59.9 | |
| F1 Diff | +4.2 | +0.9 | +2.5 |
| Models | Vicuna-7B-v1.1 | SAIL-7B |
| Novel +Verbs | Include +Consider | Calculate +Match +Revolutionize +Check +Include +Increase |
| Count | 2 | 6 |
| Models | Avg. | Std. | Diversity |
| GPT-4 | 303.8 | 121.5 | 0.48 |
| ChatGPT | 135.1 | 63.6 | 0.56 |
| Vicuna-13B | 204.1 | 82.9 | 0.45 |
| Vicuna-7B | 196.5 | 90.3 | 0.45 |
| SAIL-7B + Search | 246.2 | 87.7 | 0.44 |
| SAIL-7B | 206.6 | 86.9 | 0.47 |
| Model | LLaMA-GPT4-7B | Vicuna-7B | Vicuna-13B | SAIL-7B | ||||||||
| Search | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG | None | Wiki | DDG |
| CSQA | 48.4 | 47.7 | 49.6 | 44.9 | 45.6 | 47.6 | 50.6 | 51.1 | 50.9 | 51.5 | 51.0 | 51.8 |
| OBQA | 42.2 | 44.4 | 44.6 | 37.2 | 39.4 | 42.6 | 49.0 | 47.2 | 49.4 | 49.2 | 50.2 | 52.0 |
| ARC-C | 43.0 | 45.2 | 47.3 | 40.5 | 44.5 | 46.3 | 53.2 | 51.6 | 51.8 | 47.7 | 48.1 | 48.4 |
| Avg. | 44.5 | 45.8 | 47.2 | 40.9 | 43.3 | 45.5 | 51.0 | 50.0 | 50.7 | 49.5 | 49.8 | 50.7 |
| Search Effect | none | +1.3 | +2.7 | none | +2.4 | +4.6 | none | -1.0 | -0.3 | none | +0.3 | +1.2 |
| Model | Size | True | True * Info |
| No Search Augmentation | |||
| GPT-3 | 175B | 0.28 | 0.25 |
| LLaMA | 65B | 0.57 | 0.53 |
| LLaMA | 7B | 0.33 | 0.29 |
| Alpaca | 7B | 0.33 | 0.33 |
| Vicuna | 7B | 0.56 | 0.52 |
| W/ Search Augmentation | |||
| Vicuna | 13B | 0.71 | 0.69 |
| Vicuna | 7B | 0.68 | 0.65 |
| SAIL | 7B | 0.73 | 0.73 |
| Model | Metric | Climate | PubHealth | Fact Avg. | HSD | SBIC | Fairness Avg. | All Avg. |
| Vicuna-7B | Acc | 57.9 | 60.6 | 59.2 | 55.9 | 74.5 | 65.2 | 62.2 |
| F1 | 38.8 | 56.63 | 47.7 | 68.5 | 84.3 | 76.4 | 62.04 | |
| Vicuna-13B | Acc | 51.4 | 54.4 | 52.9 | 57.7 | 72.3 | 65.0 | 59.0 |
| F1 | 42.5 | 57.7 | 50.1 | 69.6 | 82.9 | 76.3 | 63.2 | |
| LLaMA-7B | Acc | 58.8 | 59.9 | 59.3 | 62.3 | 74.8 | 68.6 | 63.9 |
| F1 | 46.6 | 57.5 | 52.0 | 72.3 | 84.4 | 78.4 | 65.2 | |
| SAIL-7B | Acc | 63.5 | 69.2 | 66.4 | 70.1 | 76.4 | 73.2 | 69.8 |
| F1 | 51.0 | 63.6 | 57.3 | 75.1 | 83.9 | 79.5 | 68.4 |
| Model | Metric | Climate | PubHealth | Avg. |
| Vicuna-7B | Acc | 57.7 | 60.1 | 58.9 |
| Acc Diff | -0.2 | -0.5 | -0.3 | |
| F1 | 49.5 | 57.6 | 53.6 | |
| F1 Diff | +10.7 | +1.0 | +5.9 | |
| Vicuna-13B | Acc | 53.5 | 50.3 | 51.9 |
| Acc Diff | +2.1 | -4.1 | -1.0 | |
| F1 | 46.6 | 56.8 | 51.7 | |
| F1 Diff | +4.1 | -0.9 | +1.6 | |
| LLaMA-7B | Acc | 55.8 | 62.8 | 59.3 |
| Acc Diff | -3.0 | +2.9 | -0.1 | |
| F1 | 50.2 | 59.7 | 54.9 | |
| F1 Diff | +3.6 | +2.2 | +2.9 | |
| SAIL-7B | Acc | 65.8 | 70.7 | 68.3 |
| Acc Diff | +2.3 | +1.5 | +1.9 | |
| F1 | 55.2 | 64.5 | 59.9 | |
| F1 Diff | +4.2 | +0.9 | +2.5 |
| Models | Vicuna-7B-v1.1 | SAIL-7B |
| Novel \nVerbs | Include \nConsider | Calculate \nMatch \nRevolutionize \nCheck \nInclude \nIncrease |
| Count | 2 | 6 |
| Models | Avg. | Std. | Diversity |
| GPT-4 | 303.8 | 121.5 | 0.48 |
| ChatGPT | 135.1 | 63.6 | 0.56 |
| Vicuna-13B | 204.1 | 82.9 | 0.45 |
| Vicuna-7B | 196.5 | 90.3 | 0.45 |
| SAIL-7B + Search | 246.2 | 87.7 | 0.44 |
| SAIL-7B | 206.6 | 86.9 | 0.47 |
| Dataset | v | Lang. Pair | #Docs | #Sents | #Hours |
| MuST-C | 1.0 | En-De | 2,093 | 234K | 408 |
| En-Es | 2,514 | 266K | 496 | ||
| En-Fr | 2,460 | 275K | 485 | ||
| En-It | 2,324 | 254K | 457 | ||
| En-NI | 2,219 | 248K | 434 | ||
| En-Pt | 2,001 | 206K | 377 | ||
| En-Ro | 2,166 | 236K | 424 | ||
| En-Ru | 2,448 | 265K | 482 | ||
| 2.0 | En-De | 2,537 | 251K | 450 | |
| mTEDx | Es-Es | 988 | 102K | 178 | |
| Pt-Pt | 812 | 90K | 153 | ||
| Es-En | 378 | 36K | 64 | ||
| Pt-En | 279 | 31K | 53 | ||
| Es-Fr | 43 | 4K | 6 | ||
| CoVoST2 | En-De | — | 231K | 362 |
| Model | En-De | En-Es | En-Fr | En-It | En-Nl | En-Pt | En-Ro | En-Ru | Average |
| Fairseq ST | 22.7 / — | 27.2 / — | 32.9 / — | 22.7 / — | 27.3 / — | 28.1 / — | 21.9 / — | 15.3 / — | 24.8 / — |
| Baseline | 23.2 / 50.5 | 27.5 / 55.0 | 33.1 / 58.3 | 22.9 / 50.4 | 27.3 / 53.9 | 28.7 / 54.9 | 22.2 / 49.0 | 15.3 / 40.4 | 25.0 / 51.6 |
| + SEGAUGMENT | 25.0 / 52.6 | 29.3 / 56.8 | 35.7 / 60.7 | 25.6 / 52.9 | 30.3 / 57.0 | 31.8 / 57.8 | 25.0 / 51.8 | 16.8 / 42.5 | 27.4 / 54.0 |
| + ASR SEGAUGMENT | 25.5 / 52.8 | 29.5 / 56.9 | 35.8 / 60.7 | 25.7 / 52.9 | 30.0 / 56.8 | 31.6 / 57.6 | 25.0 / 52.0 | 16.7 / 42.4 | 27.5 / 54.0 |
| Model | En-Es | En-Fr |
| Chimera (Han et al., 2021) | 30.6 / — | 35.6 / — |
| STEMM (Fang et al., 2022) | 31.0 / — | 37.4 / — |
| ConST (Ye et al., 2022) | 32.0 / — | 38.3 / — |
| STPT (Tang et al., 2022) | 33.1 / — | 39.7 / — |
| SpeechUT (Zhang et al., 2022b) | 33.6 / — | 41.4 / — |
| w2v-mBART | 33.3 / 60.2 | 40.8 / 64.5 |
| + SEGAUGMENT | 33.7 / 60.7 | 41.5 / 65.1 |
| Model | Es-En | Pt-En | Es-Fr |
| Bilingual E2E-ST | 7.0 / — | 8.1 / — | 1.7 / — |
| Baseline | 15.6 / 40.5 | 15.3 / 38.6 | 2.7 / 23.4 |
| + SEGAUGMENT | 18.5 / 43.5 | 17.8 / 41.1 | 3.8 / 23.9 |
| + ASR SEGAUGMENT | 19.1 / 43.9 | 20.3 / 43.7 | 5.3 / 27.2 |
| Model | En-De |
| Bi-AST (Wang et al., 2021) | 16.3 / — |
| STR (Lam et al., 2022b) | 18.8 / — |
| Baseline | 17.4 / 43.2 |
| + SEGAUGMENT | 19.0 / 44.9 |
| + ASR SEGAUGMENT | 19.3 / 45.4 |
| Lang. Pair | Model | test set Segmentation | |||
| Manual | SHAS-long | ||||
| En-De | Baseline | 23.2 | 21.2 | ||
| + SEGAUGMENT | 25.0 | 25.1 | 24.5 | 25.2 | |
| ← special tok. | 25.1 | 25.2 | |||
| En-Es | Baseline | 27.5 | 26.1 | ||
| + SEGAUGMENT | 29.3 | 29.3 | 29.0 | 29.4 | |
| ← special tok. | 29.2 | 29.4 | |||
| En-Fr | Baseline | 33.1 | 30.6 | ||
| + SEGAUGMENT | 35.7 | 35.7 | 34.9 | 35.3 | |
| ← special tok. | 35.6 | 35.3 | |||
| En-It | Baseline | 22.9 | 21.5 | ||
| + SEGAUGMENT | 25.6 | 25.6 | 24.4 | 25.1 | |
| ← special tok. | 25.0 | 25.1 | |||
| En-Nl | Baseline | 27.3 | 25.7 | ||
| + SEGAUGMENT | 30.3 | 30.3 | 29.7 | 29.8 | |
| ← special tok. | 29.4 | 29.8 | |||
| En-Pt | Baseline | 28.7 | 26.9 | ||
| + SEGAUGMENT | 31.8 | 31.8 | 31.4 | 31.8 | |
| ← special tok. | 31.5 | 31.8 | |||
| En-Ro | Baseline | 22.2 | 20.5 | ||
| + SEGAUGMENT | 25.0 | 25.0 | 24.1 | 24.6 | |
| ← special tok. | 24.8 | 24.6 | |||
| En-Ru | Baseline | 15.3 | 14.6 | ||
| + SEGAUGMENT | 16.8 | 16.8 | 16.8 | 17.0 | |
| ← special tok. | 16.4 | 17.0 | |||
| Average | Baseline | 25.0 | 23.4 | ||
| + SEGAUGMENT | 27.4 | 27.4 | 26.9 | 27.3 | |
| ← special tok. | 27.1 | 27.3 |
| Model | #p | ASR | En-Es | En-Fr |
| Baseline | 31M | ✓ | 27.5 / 55.0 | 33.1 / 58.3 |
| + SEGAUGMENT | 31M | ✓ | 29.3 / 56.8 | 35.7 / 60.7 |
| Revisit-ST | 48M | ✗ | 28.1 / — | 33.4 / — |
| Baseline | 31M | ✗ | 25.3 / 52.3 | 30.1 / 55.4 |
| + SEGAUGMENT | 31M | ✗ | 28.5 / 55.7 | 34.6 / 59.6 |
| Model | En-De |
| Baseline | 24.3 / 51.9 |
| Baseline + in-data-KD | 25.2 / 52.7 |
| Baseline + SEGAUGMENT | 26.2 / 53.7 |
| Model | MuST-C | mTEDx | CoVoST | |
| En-De | Es-Es (xs) | Pt-Pt | En-De | |
| Baseline | 11.1 | 18 (19.7) | 28.3 | 26.6 |
| + SEGAUGMENT | 10.4 | 16.6 (18.1) | 23.8 | 23.8 |
| Lang. Pair | Model | test set Segmentation | ||||||
| Manual | SHAS | |||||||
| — | Best | s | m | l | xl | Best | ||
| En-De | Baseline | 23.2 | 20.6 | 21.7 | 21.2 | 20.2 | 25.2 | |
| + SEGAUGMENT | 25.0 | 25.1 | 23.1 | 24.3 | 24.5 | 24.7 | ||
| ← special tok. | 25.1 | 23.1 | 24.5 | 25.2 | 25.1 | |||
| En-Es | Baseline | 27.5 | 25.0 | 26.4 | 26.1 | 24.8 | 29.4 | |
| + SEGAUGMENT | 29.3 | 29.3 | 27.3 | 28.7 | 29.0 | 28.9 | ||
| ← special tok. | 29.2 | 26.6 | 28.6 | 29.4 | 29.1 | |||
| En-Fr | Baseline | 33.1 | 30.0 | 31.0 | 30.6 | 29.0 | 35.3 | |
| + SEGAUGMENT | 35.7 | 35.7 | 32.3 | 34.1 | 34.9 | 34.8 | ||
| ← special tok. | 35.6 | 31.4 | 33.8 | 35.3 | 35.1 | |||
| En-It | Baseline | 22.9 | 20.0 | 21.6 | 21.5 | 20.3 | 25.1 | |
| + SEGAUGMENT | 25.6 | 25.6 | 22.7 | 24.2 | 24.4 | 24.8 | ||
| ← special tok. | 25.0 | 22.3 | 24.3 | 25.1 | 24.5 | |||
| En-NI | Baseline | 27.3 | 24.9 | 26.5 | 25.7 | 24.3 | 29.8 | |
| + SEGAUGMENT | 30.3 | 30.3 | 27.5 | 29.2 | 29.7 | 29.7 | ||
| ← special tok. | 29.4 | 26.9 | 26.9 | 29.8 | 29.3 | |||
| En-Pt | Baseline | 28.7 | 25.8 | 27.1 | 26.9 | 25.4 | 31.8 | |
| + SEGAUGMENT | 31.8 | 31.8 | 29.2 | 30.7 | 31.4 | 31.1 | ||
| ← special tok. | 31.5 | 28.8 | 30.8 | 31.8 | 31.3 | |||
| En-Ro | Baseline | 22.2 | 19.9 | 20.8 | 20.5 | 18.7 | 24.6 | |
| + SEGAUGMENT | 25.0 | 25.0 | 22.5 | 23.4 | 24.1 | 24.3 | ||
| ← special tok. | 24.8 | 22.6 | 24.1 | 24.6 | 24.0 | |||
| En-Ru | Baseline | 15.3 | 13.4 | 14.7 | 14.6 | 13.8 | 17.0 | |
| + SEGAUGMENT | 16.8 | 16.8 | 15.0 | 16.4 | 16.8 | 16.7 | ||
| ← special tok. | 16.4 | 14.9 | 16.4 | 17.0 | 16.8 | |||
| Avg | Baseline | 25.0 | 22.4 | 23.7 | 23.4 | 22.1 | 27.3 | |
| + SEGAUGMENT | 27.4 | 27.4 | 24.9 | 26.4 | 26.9 | 26.9 | ||
| ← special tok. | 27.1 | 24.6 | 26.2 | 27.3 | 26.9 |
| Original Data | SEGAUGMENT | BLEU | ||||
| s | m | l | xl | |||
| (1) | ✓ | 24.3 | ||||
| (2) | ✓ | 22.6 | ||||
| (3) | ✓ | 23.4 | ||||
| (4) | ✓ | 22.4 | ||||
| (5) | ✓ | 22.7 | ||||
| (6) | ✓ | ✓ | ✓ | ✓ | 25.4 | |
| (7) | ✓ | ✓ | 24.8 | |||
| (8) | ✓ | ✓ | 25.1 | |||
| (9) | ✓ | ✓ | 25.1 | |||
| (10) | ✓ | ✓ | 25.1 | |||
| (11) | ✓ | ✓ | ✓ | 25.6 | ||
| (12) | ✓ | ✓ | ✓ | 25.6 | ||
| (13) | ✓ | ✓ | ✓ | 25.8 | ||
| (14) | ✓ | ✓ | ✓ | 25.8 | ||
| (15) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (16) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (17) | ✓ | ✓ | ✓ | ✓ | ✓ | 26.2 |
| Original Data | SEGAUGMENT | WER(↓) | BLEU(↑) | |||
| s | m | l | xl | |||
| (1) | ✓ | 11.1 | 31.0 | |||
| (2) | ✓ | ✓ | 10.2 | 31.3 | ||
| (3) | ✓ | ✓ | 10.3 | 31.6 | ||
| (4) | ✓ | ✓ | 11.0 | 31.7 | ||
| (5) | ✓ | 10.9 | 31.7 | |||
| (6) | ✓ | ✓ | ✓ | ✓ | 10.4 | 32.1 |
| MT model Ml | Original train | Original development | Synthetic train (Sl) | |||
| BLEU | doc-BLEU | BLEU | doc-BLEU | BLEU | doc-BLEU | |
| s | 95.1 | 79.2 | 24.5 | 26.2 | - | 68.3 |
| m | 90.0 | 74.4 | 26.0 | 27.5 | - | 70.4 |
| l | 72.9 | 61.3 | 25.5 | 27.0 | - | 64.6 |
| xl | 59.6 | 51.8 | 25.9 | 27.4 | - | 56.8 |
| SEGAUGMENT | Expanded | Isolated | Mixed | Equal |
| s | 12.3% | 56.9% | 22.3% | 8.5% |
| m | 31.1% | 33.3% | 24.3% | 11.3% |
| l | 63.2% | 7.9% | 24.6% | 4.2% |
| xl | 55.7% | 1.2% | 43.0% | 0.1% |
| 28.3% | 38.7% | 25.0% | 8.0% |
| Dataset | v | Lang. Pair | Sentence-level version | |||||
| Original | SEGAUGMENT | |||||||
| s | m | l | xl | Concat | ||||
| MuST-C | v1.0 | En-De | 225K | 433K | 253K | 120K | 76K | 882K |
| En-Es | 260K | 430K | 253K | 120K | 76K | 879K | ||
| En-Fr | 269K | 421K | 253K | 120K | 76K | 871K | ||
| En-It | 248K | 399K | 253K | 120K | 76K | 848K | ||
| En-NI | 244K | 433K | 253K | 120K | 76K | 882K | ||
| En-Pt | 201K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ro | 231K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ru | 260K | 419K | 253K | 120K | 76K | 868K | ||
| v2.0 | En-De | 249K | 401K | 231K | 109K | 70K | 812K | |
| mTEDx | Es-Es | 101k | 150K | 88K | 43K | 27K | 308K | |
| Pt-Pt | 90K | 136K | 78K | 37K | 24K | 274K | ||
| Es-En | 36K | 53K | 32K | 15K | 10K | 110K | ||
| Pt-En | 30K | 47K | 27K | 13K | 8K | 94K | ||
| Es-Fr | 3.5K | 5.5K | 3K | 1.5K | 1K | 11K | ||
| CoVoST2 | En-De | 231K | 504K | 255K | 0 | 0 | 759K |
| Dataset | v | Lang. Pair | Sentence-level version | ||||
| Original | SEGAUGMENT | ||||||
| s | m | l | xl | ||||
| MuST-C | v1.0 | En-De | 6.27 | 3.83 | 6.87 | 15.21 | 23.96 |
| En-Es | 6.63 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Fr | 6.28 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-It | 6.41 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| En-Nl | 6.23 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Pt | 6.50 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ro | 6.37 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ru | 6.47 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| v2.0 | En-De | 6.27 | 3.75 | 6.83 | 15.16 | 23.94 | |
| mTEDx | Es-Es | 5.90 | 3.79 | 6.91 | 15.24 | 23.86 | |
| Pt-Pt | 5.81 | 3.71 | 6.86 | 15.17 | 24.03 | ||
| Es-En | 6.06 | 3.87 | 6.91 | 15.27 | 23.78 | ||
| Pt-En | 5.86 | 3.68 | 6.83 | 15.16 | 24.02 | ||
| Es-Fr | 6.08 | 3.80 | 6.89 | 15.10 | 23.84 | ||
| CoVoST2 | En-De | 5.61 | 1.82 | 3.68 | — | — |
| Dataset | v | Lang. Pair | #Docs | #Sents | #Hours |
| MuST-C | 1.0 | En-De | 2,093 | 234K | 408 |
| En-Es | 2,514 | 266K | 496 | ||
| En-Fr | 2,460 | 275K | 485 | ||
| En-It | 2,324 | 254K | 457 | ||
| En-NI | 2,219 | 248K | 434 | ||
| En-Pt | 2,001 | 206K | 377 | ||
| En-Ro | 2,166 | 236K | 424 | ||
| En-Ru | 2,448 | 265K | 482 | ||
| 2.0 | En-De | 2,537 | 251K | 450 | |
| mTEDx | Es-Es | 988 | 102K | 178 | |
| Pt-Pt | 812 | 90K | 153 | ||
| Es-En | 378 | 36K | 64 | ||
| Pt-En | 279 | 31K | 53 | ||
| Es-Fr | 43 | 4K | 6 | ||
| CoVoST2 | En-De | — | 231K | 362 |
| Model | En-De | En-Es | En-Fr | En-It | En-Nl | En-Pt | En-Ro | En-Ru | Average |
| Fairseq ST | 22.7 / — | 27.2 / — | 32.9 / — | 22.7 / — | 27.3 / — | 28.1 / — | 21.9 / — | 15.3 / — | 24.8 / — |
| Baseline | 23.2 / 50.5 | 27.5 / 55.0 | 33.1 / 58.3 | 22.9 / 50.4 | 27.3 / 53.9 | 28.7 / 54.9 | 22.2 / 49.0 | 15.3 / 40.4 | 25.0 / 51.6 |
| + SEGAUGMENT | 25.0 / 52.6 | 29.3 / 56.8 | 35.7 / 60.7 | 25.6 / 52.9 | 30.3 / 57.0 | 31.8 / 57.8 | 25.0 / 51.8 | 16.8 / 42.5 | 27.4 / 54.0 |
| + ASR SEGAUGMENT | 25.5 / 52.8 | 29.5 / 56.9 | 35.8 / 60.7 | 25.7 / 52.9 | 30.0 / 56.8 | 31.6 / 57.6 | 25.0 / 52.0 | 16.7 / 42.4 | 27.5 / 54.0 |
| Model | En-Es | En-Fr |
| Chimera (Han et al., 2021) | 30.6 / — | 35.6 / — |
| STEMM (Fang et al., 2022) | 31.0 / — | 37.4 / — |
| ConST (Ye et al., 2022) | 32.0 / — | 38.3 / — |
| STPT (Tang et al., 2022) | 33.1 / — | 39.7 / — |
| SpeechUT (Zhang et al., 2022b) | 33.6 / — | 41.4 / — |
| w2v-mBART | 33.3 / 60.2 | 40.8 / 64.5 |
| + SEGAUGMENT | 33.7 / 60.7 | 41.5 / 65.1 |
| Model | Es-En | Pt-En | Es-Fr |
| Bilingual E2E-ST | 7.0 / — | 8.1 / — | 1.7 / — |
| Baseline | 15.6 / 40.5 | 15.3 / 38.6 | 2.7 / 23.4 |
| + SEGAUGMENT | 18.5 / 43.5 | 17.8 / 41.1 | 3.8 / 23.9 |
| + ASR SEGAUGMENT | 19.1 / 43.9 | 20.3 / 43.7 | 5.3 / 27.2 |
| Model | En-De |
| Bi-AST (Wang et al., 2021) | 16.3 / — |
| STR (Lam et al., 2022b) | 18.8 / — |
| Baseline | 17.4 / 43.2 |
| + SEGAUGMENT | 19.0 / 44.9 |
| + ASR SEGAUGMENT | 19.3 / 45.4 |
| Lang. Pair | Model | test set Segmentation | |||
| Manual | SHAS-long | ||||
| En-De | Baseline | 23.2 | 21.2 | ||
| + SEGAUGMENT | 25.0 | 25.1 | 24.5 | 25.2 | |
| ← special tok. | 25.1 | 25.2 | |||
| En-Es | Baseline | 27.5 | 26.1 | ||
| + SEGAUGMENT | 29.3 | 29.3 | 29.0 | 29.4 | |
| ← special tok. | 29.2 | 29.4 | |||
| En-Fr | Baseline | 33.1 | 30.6 | ||
| + SEGAUGMENT | 35.7 | 35.7 | 34.9 | 35.3 | |
| ← special tok. | 35.6 | 35.3 | |||
| En-It | Baseline | 22.9 | 21.5 | ||
| + SEGAUGMENT | 25.6 | 25.6 | 24.4 | 25.1 | |
| ← special tok. | 25.0 | 25.1 | |||
| En-Nl | Baseline | 27.3 | 25.7 | ||
| + SEGAUGMENT | 30.3 | 30.3 | 29.7 | 29.8 | |
| ← special tok. | 29.4 | 29.8 | |||
| En-Pt | Baseline | 28.7 | 26.9 | ||
| + SEGAUGMENT | 31.8 | 31.8 | 31.4 | 31.8 | |
| ← special tok. | 31.5 | 31.8 | |||
| En-Ro | Baseline | 22.2 | 20.5 | ||
| + SEGAUGMENT | 25.0 | 25.0 | 24.1 | 24.6 | |
| ← special tok. | 24.8 | 24.6 | |||
| En-Ru | Baseline | 15.3 | 14.6 | ||
| + SEGAUGMENT | 16.8 | 16.8 | 16.8 | 17.0 | |
| ← special tok. | 16.4 | 17.0 | |||
| Average | Baseline | 25.0 | 23.4 | ||
| + SEGAUGMENT | 27.4 | 27.4 | 26.9 | 27.3 | |
| ← special tok. | 27.1 | 27.3 |
| Model | #p | ASR | En-Es | En-Fr |
| Baseline | 31M | ✓ | 27.5 / 55.0 | 33.1 / 58.3 |
| + SEGAUGMENT | 31M | ✓ | 29.3 / 56.8 | 35.7 / 60.7 |
| Revisit-ST | 48M | ✗ | 28.1 / — | 33.4 / — |
| Baseline | 31M | ✗ | 25.3 / 52.3 | 30.1 / 55.4 |
| + SEGAUGMENT | 31M | ✗ | 28.5 / 55.7 | 34.6 / 59.6 |
| Model | En-De |
| Baseline | 24.3 / 51.9 |
| Baseline + in-data-KD | 25.2 / 52.7 |
| Baseline + SEGAUGMENT | 26.2 / 53.7 |
| Model | MuST-C | mTEDx | CoVoST | |
| En-De | Es-Es (xs) | Pt-Pt | En-De | |
| Baseline | 11.1 | 18 (19.7) | 28.3 | 26.6 |
| + SEGAUGMENT | 10.4 | 16.6 (18.1) | 23.8 | 23.8 |
| Lang. Pair | Model | test set Segmentation | ||||||
| Manual | SHAS | |||||||
| — | Best | s | m | l | xl | Best | ||
| En-De | Baseline | 23.2 | 20.6 | 21.7 | 21.2 | 20.2 | 25.2 | |
| + SEGAUGMENT | 25.0 | 25.1 | 23.1 | 24.3 | 24.5 | 24.7 | ||
| ← special tok. | 25.1 | 23.1 | 24.5 | 25.2 | 25.1 | |||
| En-Es | Baseline | 27.5 | 25.0 | 26.4 | 26.1 | 24.8 | 29.4 | |
| + SEGAUGMENT | 29.3 | 29.3 | 27.3 | 28.7 | 29.0 | 28.9 | ||
| ← special tok. | 29.2 | 26.6 | 28.6 | 29.4 | 29.1 | |||
| En-Fr | Baseline | 33.1 | 30.0 | 31.0 | 30.6 | 29.0 | 35.3 | |
| + SEGAUGMENT | 35.7 | 35.7 | 32.3 | 34.1 | 34.9 | 34.8 | ||
| ← special tok. | 35.6 | 31.4 | 33.8 | 35.3 | 35.1 | |||
| En-It | Baseline | 22.9 | 20.0 | 21.6 | 21.5 | 20.3 | 25.1 | |
| + SEGAUGMENT | 25.6 | 25.6 | 22.7 | 24.2 | 24.4 | 24.8 | ||
| ← special tok. | 25.0 | 22.3 | 24.3 | 25.1 | 24.5 | |||
| En-NI | Baseline | 27.3 | 24.9 | 26.5 | 25.7 | 24.3 | 29.8 | |
| + SEGAUGMENT | 30.3 | 30.3 | 27.5 | 29.2 | 29.7 | 29.7 | ||
| ← special tok. | 29.4 | 26.9 | 26.9 | 29.8 | 29.3 | |||
| En-Pt | Baseline | 28.7 | 25.8 | 27.1 | 26.9 | 25.4 | 31.8 | |
| + SEGAUGMENT | 31.8 | 31.8 | 29.2 | 30.7 | 31.4 | 31.1 | ||
| ← special tok. | 31.5 | 28.8 | 30.8 | 31.8 | 31.3 | |||
| En-Ro | Baseline | 22.2 | 19.9 | 20.8 | 20.5 | 18.7 | 24.6 | |
| + SEGAUGMENT | 25.0 | 25.0 | 22.5 | 23.4 | 24.1 | 24.3 | ||
| ← special tok. | 24.8 | 22.6 | 24.1 | 24.6 | 24.0 | |||
| En-Ru | Baseline | 15.3 | 13.4 | 14.7 | 14.6 | 13.8 | 17.0 | |
| + SEGAUGMENT | 16.8 | 16.8 | 15.0 | 16.4 | 16.8 | 16.7 | ||
| ← special tok. | 16.4 | 14.9 | 16.4 | 17.0 | 16.8 | |||
| Avg | Baseline | 25.0 | 22.4 | 23.7 | 23.4 | 22.1 | 27.3 | |
| + SEGAUGMENT | 27.4 | 27.4 | 24.9 | 26.4 | 26.9 | 26.9 | ||
| ← special tok. | 27.1 | 24.6 | 26.2 | 27.3 | 26.9 |
| Original Data | SEGAUGMENT | BLEU | ||||
| s | m | l | xl | |||
| (1) | ✓ | 24.3 | ||||
| (2) | ✓ | 22.6 | ||||
| (3) | ✓ | 23.4 | ||||
| (4) | ✓ | 22.4 | ||||
| (5) | ✓ | 22.7 | ||||
| (6) | ✓ | ✓ | ✓ | ✓ | 25.4 | |
| (7) | ✓ | ✓ | 24.8 | |||
| (8) | ✓ | ✓ | 25.1 | |||
| (9) | ✓ | ✓ | 25.1 | |||
| (10) | ✓ | ✓ | 25.1 | |||
| (11) | ✓ | ✓ | ✓ | 25.6 | ||
| (12) | ✓ | ✓ | ✓ | 25.6 | ||
| (13) | ✓ | ✓ | ✓ | 25.8 | ||
| (14) | ✓ | ✓ | ✓ | 25.8 | ||
| (15) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (16) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (17) | ✓ | ✓ | ✓ | ✓ | ✓ | 26.2 |
| Original Data | SEGAUGMENT | WER(↓) | BLEU(↑) | |||
| s | m | l | xl | |||
| (1) | ✓ | 11.1 | 31.0 | |||
| (2) | ✓ | ✓ | 10.2 | 31.3 | ||
| (3) | ✓ | ✓ | 10.3 | 31.6 | ||
| (4) | ✓ | ✓ | 11.0 | 31.7 | ||
| (5) | ✓ | 10.9 | 31.7 | |||
| (6) | ✓ | ✓ | ✓ | ✓ | 10.4 | 32.1 |
| MT model Ml | Original train | Original development | Synthetic train (Sl) | |||
| BLEU | doc-BLEU | BLEU | doc-BLEU | BLEU | doc-BLEU | |
| s | 95.1 | 79.2 | 24.5 | 26.2 | - | 68.3 |
| m | 90.0 | 74.4 | 26.0 | 27.5 | - | 70.4 |
| l | 72.9 | 61.3 | 25.5 | 27.0 | - | 64.6 |
| xl | 59.6 | 51.8 | 25.9 | 27.4 | - | 56.8 |
| SEGAUGMENT | Expanded | Isolated | Mixed | Equal |
| s | 12.3% | 56.9% | 22.3% | 8.5% |
| m | 31.1% | 33.3% | 24.3% | 11.3% |
| l | 63.2% | 7.9% | 24.6% | 4.2% |
| xl | 55.7% | 1.2% | 43.0% | 0.1% |
| 28.3% | 38.7% | 25.0% | 8.0% |
| Dataset | v | Lang. Pair | Sentence-level version | |||||
| Original | SEGAUGMENT | |||||||
| s | m | l | xl | Concat | ||||
| MuST-C | v1.0 | En-De | 225K | 433K | 253K | 120K | 76K | 882K |
| En-Es | 260K | 430K | 253K | 120K | 76K | 879K | ||
| En-Fr | 269K | 421K | 253K | 120K | 76K | 871K | ||
| En-It | 248K | 399K | 253K | 120K | 76K | 848K | ||
| En-NI | 244K | 433K | 253K | 120K | 76K | 882K | ||
| En-Pt | 201K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ro | 231K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ru | 260K | 419K | 253K | 120K | 76K | 868K | ||
| v2.0 | En-De | 249K | 401K | 231K | 109K | 70K | 812K | |
| mTEDx | Es-Es | 101k | 150K | 88K | 43K | 27K | 308K | |
| Pt-Pt | 90K | 136K | 78K | 37K | 24K | 274K | ||
| Es-En | 36K | 53K | 32K | 15K | 10K | 110K | ||
| Pt-En | 30K | 47K | 27K | 13K | 8K | 94K | ||
| Es-Fr | 3.5K | 5.5K | 3K | 1.5K | 1K | 11K | ||
| CoVoST2 | En-De | 231K | 504K | 255K | 0 | 0 | 759K |
| Dataset | v | Lang. Pair | Sentence-level version | ||||
| Original | SEGAUGMENT | ||||||
| s | m | l | xl | ||||
| MuST-C | v1.0 | En-De | 6.27 | 3.83 | 6.87 | 15.21 | 23.96 |
| En-Es | 6.63 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Fr | 6.28 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-It | 6.41 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| En-Nl | 6.23 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Pt | 6.50 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ro | 6.37 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ru | 6.47 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| v2.0 | En-De | 6.27 | 3.75 | 6.83 | 15.16 | 23.94 | |
| mTEDx | Es-Es | 5.90 | 3.79 | 6.91 | 15.24 | 23.86 | |
| Pt-Pt | 5.81 | 3.71 | 6.86 | 15.17 | 24.03 | ||
| Es-En | 6.06 | 3.87 | 6.91 | 15.27 | 23.78 | ||
| Pt-En | 5.86 | 3.68 | 6.83 | 15.16 | 24.02 | ||
| Es-Fr | 6.08 | 3.80 | 6.89 | 15.10 | 23.84 | ||
| CoVoST2 | En-De | 5.61 | 1.82 | 3.68 | — | — |
| Dataset | v | Lang. Pair | #Docs | #Sents | #Hours |
| MuST-C | 1.0 | En-De | 2,093 | 234K | 408 |
| En-Es | 2,514 | 266K | 496 | ||
| En-Fr | 2,460 | 275K | 485 | ||
| En-It | 2,324 | 254K | 457 | ||
| En-NI | 2,219 | 248K | 434 | ||
| En-Pt | 2,001 | 206K | 377 | ||
| En-Ro | 2,166 | 236K | 424 | ||
| En-Ru | 2,448 | 265K | 482 | ||
| 2.0 | En-De | 2,537 | 251K | 450 | |
| mTEDx | Es-Es | 988 | 102K | 178 | |
| Pt-Pt | 812 | 90K | 153 | ||
| Es-En | 378 | 36K | 64 | ||
| Pt-En | 279 | 31K | 53 | ||
| Es-Fr | 43 | 4K | 6 | ||
| CoVoST2 | En-De | — | 231K | 362 | |
| Model | En-De | En-Es | En-Fr | En-It | En-Nl | En-Pt | En-Ro | En-Ru | Average |
| Fairseq ST | 22.7 / — | 27.2 / — | 32.9 / — | 22.7 / — | 27.3 / — | 28.1 / — | 21.9 / — | 15.3 / — | 24.8 / — |
| Baseline | 23.2 / 50.5 | 27.5 / 55.0 | 33.1 / 58.3 | 22.9 / 50.4 | 27.3 / 53.9 | 28.7 / 54.9 | 22.2 / 49.0 | 15.3 / 40.4 | 25.0 / 51.6 |
| + SEGAUGMENT | 25.0 / 52.6 | 29.3 / 56.8 | 35.7 / 60.7 | 25.6 / 52.9 | 30.3 / 57.0 | 31.8 / 57.8 | 25.0 / 51.8 | 16.8 / 42.5 | 27.4 / 54.0 |
| + ASR SEGAUGMENT | 25.5 / 52.8 | 29.5 / 56.9 | 35.8 / 60.7 | 25.7 / 52.9 | 30.0 / 56.8 | 31.6 / 57.6 | 25.0 / 52.0 | 16.7 / 42.4 | 27.5 / 54.0 |
| Model | En-Es | En-Fr |
| Chimera (Han et al., 2021) | 30.6 / — | 35.6 / — |
| STEMM (Fang et al., 2022) | 31.0 / — | 37.4 / — |
| ConST (Ye et al., 2022) | 32.0 / — | 38.3 / — |
| STPT (Tang et al., 2022) | 33.1 / — | 39.7 / — |
| SpeechUT (Zhang et al., 2022b) | 33.6 / — | 41.4 / — |
| w2v-mBART | 33.3 / 60.2 | 40.8 / 64.5 |
| + SEGAUGMENT | 33.7 / 60.7 | 41.5 / 65.1 |
| Model | Es-En | Pt-En | Es-Fr |
| Bilingual E2E-ST | 7.0 / — | 8.1 / — | 1.7 / — |
| Baseline | 15.6 / 40.5 | 15.3 / 38.6 | 2.7 / 23.4 |
| + SEGAUGMENT | 18.5 / 43.5 | 17.8 / 41.1 | 3.8 / 23.9 |
| + ASR SEGAUGMENT | 19.1 / 43.9 | 20.3 / 43.7 | 5.3 / 27.2 |
| Model | En-De |
| Bi-AST (Wang et al., 2021) | 16.3 / — |
| STR (Lam et al., 2022b) | 18.8 / — |
| Baseline | 17.4 / 43.2 |
| + SEGAUGMENT | 19.0 / 44.9 |
| + ASR SEGAUGMENT | 19.3 / 45.4 |
| Lang. Pair | Model | test set Segmentation | |||
| Manual | SHAS-long | ||||
| En-De | Baseline | 23.2 | 21.2 | ||
| + SEGAUGMENT | 25.0 | 25.1 | 24.5 | 25.2 | |
| ← special tok. | 25.1 | 25.2 | |||
| En-Es | Baseline | 27.5 | 26.1 | ||
| + SEGAUGMENT | 29.3 | 29.3 | 29.0 | 29.4 | |
| ← special tok. | 29.2 | 29.4 | |||
| En-Fr | Baseline | 33.1 | 30.6 | ||
| + SEGAUGMENT | 35.7 | 35.7 | 34.9 | 35.3 | |
| ← special tok. | 35.6 | 35.3 | |||
| En-It | Baseline | 22.9 | 21.5 | ||
| + SEGAUGMENT | 25.6 | 25.6 | 24.4 | 25.1 | |
| ← special tok. | 25.0 | 25.1 | |||
| En-Nl | Baseline | 27.3 | 25.7 | ||
| + SEGAUGMENT | 30.3 | 30.3 | 29.7 | 29.8 | |
| ← special tok. | 29.4 | 29.8 | |||
| En-Pt | Baseline | 28.7 | 26.9 | ||
| + SEGAUGMENT | 31.8 | 31.8 | 31.4 | 31.8 | |
| ← special tok. | 31.5 | 31.8 | |||
| En-Ro | Baseline | 22.2 | 20.5 | ||
| + SEGAUGMENT | 25.0 | 25.0 | 24.1 | 24.6 | |
| ← special tok. | 24.8 | 24.6 | |||
| En-Ru | Baseline | 15.3 | 14.6 | ||
| + SEGAUGMENT | 16.8 | 16.8 | 16.8 | 17.0 | |
| ← special tok. | 16.4 | 17.0 | |||
| Average | Baseline | 25.0 | 23.4 | ||
| + SEGAUGMENT | 27.4 | 27.4 | 26.9 | 27.3 | |
| ← special tok. | 27.1 | 27.3 | |||
| Model | #p | ASR | En-Es | En-Fr |
| Baseline | 31M | ✓ | 27.5 / 55.0 | 33.1 / 58.3 |
| + SEGAUGMENT | 31M | ✓ | 29.3 / 56.8 | 35.7 / 60.7 |
| Revisit-ST | 48M | ✗ | 28.1 / — | 33.4 / — |
| Baseline | 31M | ✗ | 25.3 / 52.3 | 30.1 / 55.4 |
| + SEGAUGMENT | 31M | ✗ | 28.5 / 55.7 | 34.6 / 59.6 |
| Model | En-De |
| Baseline | 24.3 / 51.9 |
| Baseline + in-data-KD | 25.2 / 52.7 |
| Baseline + SEGAUGMENT | 26.2 / 53.7 |
| Model | MuST-C | mTEDx | CoVoST | |
| En-De | Es-Es (xs) | Pt-Pt | En-De | |
| Baseline | 11.1 | 18 (19.7) | 28.3 | 26.6 |
| + SEGAUGMENT | 10.4 | 16.6 (18.1) | 23.8 | 23.8 |
| Lang. Pair | Model | test set Segmentation | ||||||
| Manual | SHAS | |||||||
| — | Best | s | m | l | xl | Best | ||
| En-De | Baseline | 23.2 | 20.6 | 21.7 | 21.2 | 20.2 | 25.2 | |
| + SEGAUGMENT | 25.0 | 25.1 | 23.1 | 24.3 | 24.5 | 24.7 | ||
| ← special tok. | 25.1 | 23.1 | 24.5 | 25.2 | 25.1 | |||
| En-Es | Baseline | 27.5 | 25.0 | 26.4 | 26.1 | 24.8 | 29.4 | |
| + SEGAUGMENT | 29.3 | 29.3 | 27.3 | 28.7 | 29.0 | 28.9 | ||
| ← special tok. | 29.2 | 26.6 | 28.6 | 29.4 | 29.1 | |||
| En-Fr | Baseline | 33.1 | 30.0 | 31.0 | 30.6 | 29.0 | 35.3 | |
| + SEGAUGMENT | 35.7 | 35.7 | 32.3 | 34.1 | 34.9 | 34.8 | ||
| ← special tok. | 35.6 | 31.4 | 33.8 | 35.3 | 35.1 | |||
| En-It | Baseline | 22.9 | 20.0 | 21.6 | 21.5 | 20.3 | 25.1 | |
| + SEGAUGMENT | 25.6 | 25.6 | 22.7 | 24.2 | 24.4 | 24.8 | ||
| ← special tok. | 25.0 | 22.3 | 24.3 | 25.1 | 24.5 | |||
| En-NI | Baseline | 27.3 | 24.9 | 26.5 | 25.7 | 24.3 | 29.8 | |
| + SEGAUGMENT | 30.3 | 30.3 | 27.5 | 29.2 | 29.7 | 29.7 | ||
| ← special tok. | 29.4 | 26.9 | 26.9 | 29.8 | 29.3 | |||
| En-Pt | Baseline | 28.7 | 25.8 | 27.1 | 26.9 | 25.4 | 31.8 | |
| + SEGAUGMENT | 31.8 | 31.8 | 29.2 | 30.7 | 31.4 | 31.1 | ||
| ← special tok. | 31.5 | 28.8 | 30.8 | 31.8 | 31.3 | |||
| En-Ro | Baseline | 22.2 | 19.9 | 20.8 | 20.5 | 18.7 | 24.6 | |
| + SEGAUGMENT | 25.0 | 25.0 | 22.5 | 23.4 | 24.1 | 24.3 | ||
| ← special tok. | 24.8 | 22.6 | 24.1 | 24.6 | 24.0 | |||
| En-Ru | Baseline | 15.3 | 13.4 | 14.7 | 14.6 | 13.8 | 17.0 | |
| + SEGAUGMENT | 16.8 | 16.8 | 15.0 | 16.4 | 16.8 | 16.7 | ||
| ← special tok. | 16.4 | 14.9 | 16.4 | 17.0 | 16.8 | |||
| Avg | Baseline | 25.0 | 22.4 | 23.7 | 23.4 | 22.1 | 27.3 | |
| + SEGAUGMENT | 27.4 | 27.4 | 24.9 | 26.4 | 26.9 | 26.9 | ||
| ← special tok. | 27.1 | 24.6 | 26.2 | 27.3 | 26.9 | |||
| Original Data | SEGAUGMENT | BLEU | ||||
| s | m | l | xl | |||
| (1) | ✓ | 24.3 | ||||
| (2) | ✓ | 22.6 | ||||
| (3) | ✓ | 23.4 | ||||
| (4) | ✓ | 22.4 | ||||
| (5) | ✓ | 22.7 | ||||
| (6) | ✓ | ✓ | ✓ | ✓ | 25.4 | |
| (7) | ✓ | ✓ | 24.8 | |||
| (8) | ✓ | ✓ | 25.1 | |||
| (9) | ✓ | ✓ | 25.1 | |||
| (10) | ✓ | ✓ | 25.1 | |||
| (11) | ✓ | ✓ | ✓ | 25.6 | ||
| (12) | ✓ | ✓ | ✓ | 25.6 | ||
| (13) | ✓ | ✓ | ✓ | 25.8 | ||
| (14) | ✓ | ✓ | ✓ | 25.8 | ||
| (15) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (16) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (17) | ✓ | ✓ | ✓ | ✓ | ✓ | 26.2 |
| Original Data | SEGAUGMENT | WER(↓) | BLEU(↑) | |||
| s | m | l | xl | |||
| (1) | ✓ | 11.1 | 31.0 | |||
| (2) | ✓ | ✓ | 10.2 | 31.3 | ||
| (3) | ✓ | ✓ | 10.3 | 31.6 | ||
| (4) | ✓ | ✓ | 11.0 | 31.7 | ||
| (5) | ✓ | 10.9 | 31.7 | |||
| (6) | ✓ | ✓ | ✓ | ✓ | 10.4 | 32.1 |
| MT model Ml | Original train | Original development | Synthetic train (Sl) | |||
| BLEU | doc-BLEU | BLEU | doc-BLEU | BLEU | doc-BLEU | |
| s | 95.1 | 79.2 | 24.5 | 26.2 | - | 68.3 |
| m | 90.0 | 74.4 | 26.0 | 27.5 | - | 70.4 |
| l | 72.9 | 61.3 | 25.5 | 27.0 | - | 64.6 |
| xl | 59.6 | 51.8 | 25.9 | 27.4 | - | 56.8 |
| SEGAUGMENT | Expanded | Isolated | Mixed | Equal |
| s | 12.3% | 56.9% | 22.3% | 8.5% |
| m | 31.1% | 33.3% | 24.3% | 11.3% |
| l | 63.2% | 7.9% | 24.6% | 4.2% |
| xl | 55.7% | 1.2% | 43.0% | 0.1% |
| 28.3% | 38.7% | 25.0% | 8.0% |
| Dataset | v | Lang. Pair | Sentence-level version | |||||
| Original | SEGAUGMENT | |||||||
| s | m | l | xl | Concat | ||||
| MuST-C | v1.0 | En-De | 225K | 433K | 253K | 120K | 76K | 882K |
| En-Es | 260K | 430K | 253K | 120K | 76K | 879K | ||
| En-Fr | 269K | 421K | 253K | 120K | 76K | 871K | ||
| En-It | 248K | 399K | 253K | 120K | 76K | 848K | ||
| En-NI | 244K | 433K | 253K | 120K | 76K | 882K | ||
| En-Pt | 201K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ro | 231K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ru | 260K | 419K | 253K | 120K | 76K | 868K | ||
| v2.0 | En-De | 249K | 401K | 231K | 109K | 70K | 812K | |
| mTEDx | Es-Es | 101k | 150K | 88K | 43K | 27K | 308K | |
| Pt-Pt | 90K | 136K | 78K | 37K | 24K | 274K | ||
| Es-En | 36K | 53K | 32K | 15K | 10K | 110K | ||
| Pt-En | 30K | 47K | 27K | 13K | 8K | 94K | ||
| Es-Fr | 3.5K | 5.5K | 3K | 1.5K | 1K | 11K | ||
| CoVoST2 | En-De | 231K | 504K | 255K | 0 | 0 | 759K | |
| Dataset | v | Lang. Pair | Sentence-level version | ||||
| Original | SEGAUGMENT | ||||||
| s | m | l | xl | ||||
| MuST-C | v1.0 | En-De | 6.27 | 3.83 | 6.87 | 15.21 | 23.96 |
| En-Es | 6.63 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Fr | 6.28 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-It | 6.41 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| En-Nl | 6.23 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Pt | 6.50 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ro | 6.37 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ru | 6.47 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| v2.0 | En-De | 6.27 | 3.75 | 6.83 | 15.16 | 23.94 | |
| mTEDx | Es-Es | 5.90 | 3.79 | 6.91 | 15.24 | 23.86 | |
| Pt-Pt | 5.81 | 3.71 | 6.86 | 15.17 | 24.03 | ||
| Es-En | 6.06 | 3.87 | 6.91 | 15.27 | 23.78 | ||
| Pt-En | 5.86 | 3.68 | 6.83 | 15.16 | 24.02 | ||
| Es-Fr | 6.08 | 3.80 | 6.89 | 15.10 | 23.84 | ||
| CoVoST2 | En-De | 5.61 | 1.82 | 3.68 | — | — | |
| Dataset | v | Lang. Pair | #Docs | #Sents | #Hours |
| MuST-C | 1.0 | En-De | 2,093 | 234K | 408 |
| En-Es | 2,514 | 266K | 496 | ||
| En-Fr | 2,460 | 275K | 485 | ||
| En-It | 2,324 | 254K | 457 | ||
| En-NI | 2,219 | 248K | 434 | ||
| En-Pt | 2,001 | 206K | 377 | ||
| En-Ro | 2,166 | 236K | 424 | ||
| En-Ru | 2,448 | 265K | 482 | ||
| 2.0 | En-De | 2,537 | 251K | 450 | |
| mTEDx | Es-Es | 988 | 102K | 178 | |
| Pt-Pt | 812 | 90K | 153 | ||
| Es-En | 378 | 36K | 64 | ||
| Pt-En | 279 | 31K | 53 | ||
| Es-Fr | 43 | 4K | 6 | ||
| CoVoST2 | En-De | — | 231K | 362 |
| Model | En-De | En-Es | En-Fr | En-It | En-Nl | En-Pt | En-Ro | En-Ru | Average |
| Fairseq ST | 22.7 / — | 27.2 / — | 32.9 / — | 22.7 / — | 27.3 / — | 28.1 / — | 21.9 / — | 15.3 / — | 24.8 / — |
| Baseline | 23.2 / 50.5 | 27.5 / 55.0 | 33.1 / 58.3 | 22.9 / 50.4 | 27.3 / 53.9 | 28.7 / 54.9 | 22.2 / 49.0 | 15.3 / 40.4 | 25.0 / 51.6 |
| + SEGAUGMENT | 25.0 / 52.6 | 29.3 / 56.8 | 35.7 / 60.7 | 25.6 / 52.9 | 30.3 / 57.0 | 31.8 / 57.8 | 25.0 / 51.8 | 16.8 / 42.5 | 27.4 / 54.0 |
| + ASR SEGAUGMENT | 25.5 / 52.8 | 29.5 / 56.9 | 35.8 / 60.7 | 25.7 / 52.9 | 30.0 / 56.8 | 31.6 / 57.6 | 25.0 / 52.0 | 16.7 / 42.4 | 27.5 / 54.0 |
| Model | En-Es | En-Fr |
| Chimera (Han et al., 2021) | 30.6 / — | 35.6 / — |
| STEMM (Fang et al., 2022) | 31.0 / — | 37.4 / — |
| ConST (Ye et al., 2022) | 32.0 / — | 38.3 / — |
| STPT (Tang et al., 2022) | 33.1 / — | 39.7 / — |
| SpeechUT (Zhang et al., 2022b) | 33.6 / — | 41.4 / — |
| w2v-mBART | 33.3 / 60.2 | 40.8 / 64.5 |
| + SEGAUGMENT | 33.7 / 60.7 | 41.5 / 65.1 |
| Model | Es-En | Pt-En | Es-Fr |
| Bilingual E2E-ST | 7.0 / — | 8.1 / — | 1.7 / — |
| Baseline | 15.6 / 40.5 | 15.3 / 38.6 | 2.7 / 23.4 |
| + SEGAUGMENT | 18.5 / 43.5 | 17.8 / 41.1 | 3.8 / 23.9 |
| + ASR SEGAUGMENT | 19.1 / 43.9 | 20.3 / 43.7 | 5.3 / 27.2 |
| Model | En-De |
| Bi-AST (Wang et al., 2021) | 16.3 / — |
| STR (Lam et al., 2022b) | 18.8 / — |
| Baseline | 17.4 / 43.2 |
| + SEGAUGMENT | 19.0 / 44.9 |
| + ASR SEGAUGMENT | 19.3 / 45.4 |
| Lang. Pair | Model | test set Segmentation | |||
| Manual | SHAS-long | ||||
| En-De | Baseline | 23.2 | 21.2 | ||
| + SEGAUGMENT | 25.0 | 25.1 | 24.5 | 25.2 | |
| ← special tok. | 25.1 | 25.2 | |||
| En-Es | Baseline | 27.5 | 26.1 | ||
| + SEGAUGMENT | 29.3 | 29.3 | 29.0 | 29.4 | |
| ← special tok. | 29.2 | 29.4 | |||
| En-Fr | Baseline | 33.1 | 30.6 | ||
| + SEGAUGMENT | 35.7 | 35.7 | 34.9 | 35.3 | |
| ← special tok. | 35.6 | 35.3 | |||
| En-It | Baseline | 22.9 | 21.5 | ||
| + SEGAUGMENT | 25.6 | 25.6 | 24.4 | 25.1 | |
| ← special tok. | 25.0 | 25.1 | |||
| En-Nl | Baseline | 27.3 | 25.7 | ||
| + SEGAUGMENT | 30.3 | 30.3 | 29.7 | 29.8 | |
| ← special tok. | 29.4 | 29.8 | |||
| En-Pt | Baseline | 28.7 | 26.9 | ||
| + SEGAUGMENT | 31.8 | 31.8 | 31.4 | 31.8 | |
| ← special tok. | 31.5 | 31.8 | |||
| En-Ro | Baseline | 22.2 | 20.5 | ||
| + SEGAUGMENT | 25.0 | 25.0 | 24.1 | 24.6 | |
| ← special tok. | 24.8 | 24.6 | |||
| En-Ru | Baseline | 15.3 | 14.6 | ||
| + SEGAUGMENT | 16.8 | 16.8 | 16.8 | 17.0 | |
| ← special tok. | 16.4 | 17.0 | |||
| Average | Baseline | 25.0 | 23.4 | ||
| + SEGAUGMENT | 27.4 | 27.4 | 26.9 | 27.3 | |
| ← special tok. | 27.1 | 27.3 |
| Model | #p | ASR | En-Es | En-Fr |
| Baseline | 31M | ✓ | 27.5 / 55.0 | 33.1 / 58.3 |
| + SEGAUGMENT | 31M | ✓ | 29.3 / 56.8 | 35.7 / 60.7 |
| Revisit-ST | 48M | ✗ | 28.1 / — | 33.4 / — |
| Baseline | 31M | ✗ | 25.3 / 52.3 | 30.1 / 55.4 |
| + SEGAUGMENT | 31M | ✗ | 28.5 / 55.7 | 34.6 / 59.6 |
| Model | En-De |
| Baseline | 24.3 / 51.9 |
| Baseline + in-data-KD | 25.2 / 52.7 |
| Baseline + SEGAUGMENT | 26.2 / 53.7 |
| Model | MuST-C | mTEDx | CoVoST | |
| En-De | Es-Es (xs) | Pt-Pt | En-De | |
| Baseline | 11.1 | 18 (19.7) | 28.3 | 26.6 |
| + SEGAUGMENT | 10.4 | 16.6 (18.1) | 23.8 | 23.8 |
| Lang. Pair | Model | test set Segmentation | ||||||
| Manual | SHAS | |||||||
| — | Best | s | m | l | xl | Best | ||
| En-De | Baseline | 23.2 | 20.6 | 21.7 | 21.2 | 20.2 | 25.2 | |
| + SEGAUGMENT | 25.0 | 25.1 | 23.1 | 24.3 | 24.5 | 24.7 | ||
| ← special tok. | 25.1 | 23.1 | 24.5 | 25.2 | 25.1 | |||
| En-Es | Baseline | 27.5 | 25.0 | 26.4 | 26.1 | 24.8 | 29.4 | |
| + SEGAUGMENT | 29.3 | 29.3 | 27.3 | 28.7 | 29.0 | 28.9 | ||
| ← special tok. | 29.2 | 26.6 | 28.6 | 29.4 | 29.1 | |||
| En-Fr | Baseline | 33.1 | 30.0 | 31.0 | 30.6 | 29.0 | 35.3 | |
| + SEGAUGMENT | 35.7 | 35.7 | 32.3 | 34.1 | 34.9 | 34.8 | ||
| ← special tok. | 35.6 | 31.4 | 33.8 | 35.3 | 35.1 | |||
| En-It | Baseline | 22.9 | 20.0 | 21.6 | 21.5 | 20.3 | 25.1 | |
| + SEGAUGMENT | 25.6 | 25.6 | 22.7 | 24.2 | 24.4 | 24.8 | ||
| ← special tok. | 25.0 | 22.3 | 24.3 | 25.1 | 24.5 | |||
| En-NI | Baseline | 27.3 | 24.9 | 26.5 | 25.7 | 24.3 | 29.8 | |
| + SEGAUGMENT | 30.3 | 30.3 | 27.5 | 29.2 | 29.7 | 29.7 | ||
| ← special tok. | 29.4 | 26.9 | 26.9 | 29.8 | 29.3 | |||
| En-Pt | Baseline | 28.7 | 25.8 | 27.1 | 26.9 | 25.4 | 31.8 | |
| + SEGAUGMENT | 31.8 | 31.8 | 29.2 | 30.7 | 31.4 | 31.1 | ||
| ← special tok. | 31.5 | 28.8 | 30.8 | 31.8 | 31.3 | |||
| En-Ro | Baseline | 22.2 | 19.9 | 20.8 | 20.5 | 18.7 | 24.6 | |
| + SEGAUGMENT | 25.0 | 25.0 | 22.5 | 23.4 | 24.1 | 24.3 | ||
| ← special tok. | 24.8 | 22.6 | 24.1 | 24.6 | 24.0 | |||
| En-Ru | Baseline | 15.3 | 13.4 | 14.7 | 14.6 | 13.8 | 17.0 | |
| + SEGAUGMENT | 16.8 | 16.8 | 15.0 | 16.4 | 16.8 | 16.7 | ||
| ← special tok. | 16.4 | 14.9 | 16.4 | 17.0 | 16.8 | |||
| Avg | Baseline | 25.0 | 22.4 | 23.7 | 23.4 | 22.1 | 27.3 | |
| + SEGAUGMENT | 27.4 | 27.4 | 24.9 | 26.4 | 26.9 | 26.9 | ||
| ← special tok. | 27.1 | 24.6 | 26.2 | 27.3 | 26.9 |
| Original Data | SEGAUGMENT | BLEU | ||||
| s | m | l | xl | |||
| (1) | ✓ | 24.3 | ||||
| (2) | ✓ | 22.6 | ||||
| (3) | ✓ | 23.4 | ||||
| (4) | ✓ | 22.4 | ||||
| (5) | ✓ | 22.7 | ||||
| (6) | ✓ | ✓ | ✓ | ✓ | 25.4 | |
| (7) | ✓ | ✓ | 24.8 | |||
| (8) | ✓ | ✓ | 25.1 | |||
| (9) | ✓ | ✓ | 25.1 | |||
| (10) | ✓ | ✓ | 25.1 | |||
| (11) | ✓ | ✓ | ✓ | 25.6 | ||
| (12) | ✓ | ✓ | ✓ | 25.6 | ||
| (13) | ✓ | ✓ | ✓ | 25.8 | ||
| (14) | ✓ | ✓ | ✓ | 25.8 | ||
| (15) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (16) | ✓ | ✓ | ✓ | ✓ | 26.0 | |
| (17) | ✓ | ✓ | ✓ | ✓ | ✓ | 26.2 |
| Original Data | SEGAUGMENT | WER(↓) | BLEU(↑) | |||
| s | m | l | xl | |||
| (1) | ✓ | 11.1 | 31.0 | |||
| (2) | ✓ | ✓ | 10.2 | 31.3 | ||
| (3) | ✓ | ✓ | 10.3 | 31.6 | ||
| (4) | ✓ | ✓ | 11.0 | 31.7 | ||
| (5) | ✓ | 10.9 | 31.7 | |||
| (6) | ✓ | ✓ | ✓ | ✓ | 10.4 | 32.1 |
| MT model Ml | Original train | Original development | Synthetic train (Sl) | |||
| BLEU | doc-BLEU | BLEU | doc-BLEU | BLEU | doc-BLEU | |
| s | 95.1 | 79.2 | 24.5 | 26.2 | - | 68.3 |
| m | 90.0 | 74.4 | 26.0 | 27.5 | - | 70.4 |
| l | 72.9 | 61.3 | 25.5 | 27.0 | - | 64.6 |
| xl | 59.6 | 51.8 | 25.9 | 27.4 | - | 56.8 |
| SEGAUGMENT | Expanded | Isolated | Mixed | Equal |
| s | 12.3% | 56.9% | 22.3% | 8.5% |
| m | 31.1% | 33.3% | 24.3% | 11.3% |
| l | 63.2% | 7.9% | 24.6% | 4.2% |
| xl | 55.7% | 1.2% | 43.0% | 0.1% |
| 28.3% | 38.7% | 25.0% | 8.0% |
| Dataset | v | Lang. Pair | Sentence-level version | |||||
| Original | SEGAUGMENT | |||||||
| s | m | l | xl | Concat | ||||
| MuST-C | v1.0 | En-De | 225K | 433K | 253K | 120K | 76K | 882K |
| En-Es | 260K | 430K | 253K | 120K | 76K | 879K | ||
| En-Fr | 269K | 421K | 253K | 120K | 76K | 871K | ||
| En-It | 248K | 399K | 253K | 120K | 76K | 848K | ||
| En-NI | 244K | 433K | 253K | 120K | 76K | 882K | ||
| En-Pt | 201K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ro | 231K | 433K | 253K | 120K | 76K | 882K | ||
| En-Ru | 260K | 419K | 253K | 120K | 76K | 868K | ||
| v2.0 | En-De | 249K | 401K | 231K | 109K | 70K | 812K | |
| mTEDx | Es-Es | 101k | 150K | 88K | 43K | 27K | 308K | |
| Pt-Pt | 90K | 136K | 78K | 37K | 24K | 274K | ||
| Es-En | 36K | 53K | 32K | 15K | 10K | 110K | ||
| Pt-En | 30K | 47K | 27K | 13K | 8K | 94K | ||
| Es-Fr | 3.5K | 5.5K | 3K | 1.5K | 1K | 11K | ||
| CoVoST2 | En-De | 231K | 504K | 255K | 0 | 0 | 759K |
| Dataset | v | Lang. Pair | Sentence-level version | ||||
| Original | SEGAUGMENT | ||||||
| s | m | l | xl | ||||
| MuST-C | v1.0 | En-De | 6.27 | 3.83 | 6.87 | 15.21 | 23.96 |
| En-Es | 6.63 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Fr | 6.28 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-It | 6.41 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| En-Nl | 6.23 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Pt | 6.50 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ro | 6.37 | 3.84 | 6.87 | 15.21 | 23.96 | ||
| En-Ru | 6.47 | 3.83 | 6.87 | 15.21 | 23.96 | ||
| v2.0 | En-De | 6.27 | 3.75 | 6.83 | 15.16 | 23.94 | |
| mTEDx | Es-Es | 5.90 | 3.79 | 6.91 | 15.24 | 23.86 | |
| Pt-Pt | 5.81 | 3.71 | 6.86 | 15.17 | 24.03 | ||
| Es-En | 6.06 | 3.87 | 6.91 | 15.27 | 23.78 | ||
| Pt-En | 5.86 | 3.68 | 6.83 | 15.16 | 24.02 | ||
| Es-Fr | 6.08 | 3.80 | 6.89 | 15.10 | 23.84 | ||
| CoVoST2 | En-De | 5.61 | 1.82 | 3.68 | — | — |
| Comp. complexity | Example (K) | Memory complexity | Example (K) | |
| Transformer | O(qkd) | qkd = 8389 | O(qk + kd) | 197 |
| Segmented | O(qsd) | qsd = 524 | O(qs + kd) | 74 |
| SRformer | O(qsd + kd2) | qsd + kd2 = 4718 | O(qs + kd + md2) | 139 |
| SRformer | Segmented | Baseline | |||||
| Dataset | Model | ROUGE2 | ROUGEL | ROUGELsum | ROUGE1 | ROUGE1 | ROUGE1 |
| CNN-DM | T5 | 16.79 | 27.29 | 36.81 | 39.39 | 33.31 | 41.60 |
| BART | 19.80 | 29.44 | 40.40 | 43.19 | 33.87 | 44.54 | |
| XSUM | T5 | 11.72 | 26.88 | 26.88 | 34.48 | 27.03 | 35.35 |
| BART | 16.58 | 31.54 | 31.54 | 39.03 | 36.22 | 41.33 | |
| ArXiv | T5 | 11.50 | 24.53 | 31.86 | 36.04 | 32.64 | 37.79 |
| BART | 14.97 | 25.29 | 37.86 | 42.99 | 36.43 | 43.95 | |
| MediaSum | T5 | 12.89 | 25.72 | 26.29 | 28.59 | 24.43 | 29.25 |
| BART | 14.70 | 24.75 | 28.46 | 32.36 | 28.40 | 33.88 |
| Setting | ROUGE1 | Setting | ROUGE1 |
| SRformer | 39.39 | - | - |
| (I) | 36.86 | (II) | 36.77 |
| (III) | 35.84 | (IV) | 29.72 |
| Model | Mem | Time | Power |
| BART | 49 | 91 | 99 |
| TransformerXL | 82 | 70 | 89 |
| LongT5 | 73 | 80 | 90 |
| SRformer | 60 | 60 | 80 |
| Comp. complexity | Example (K) | Memory complexity | Example (K) | |
| Transformer | O(qkd) | qkd = 8389 | O(qk + kd) | 197 |
| Segmented | O(qsd) | qsd = 524 | O(qs + kd) | 74 |
| SRformer | O(qsd + kd2) | qsd + kd2 = 4718 | O(qs + kd + md2) | 139 |
| SRformer | Segmented | Baseline | |||||
| Dataset | Model | ROUGE2 | ROUGEL | ROUGELsum | ROUGE1 | ROUGE1 | ROUGE1 |
| CNN-DM | T5 | 16.79 | 27.29 | 36.81 | 39.39 | 33.31 | 41.60 |
| BART | 19.80 | 29.44 | 40.40 | 43.19 | 33.87 | 44.54 | |
| XSUM | T5 | 11.72 | 26.88 | 26.88 | 34.48 | 27.03 | 35.35 |
| BART | 16.58 | 31.54 | 31.54 | 39.03 | 36.22 | 41.33 | |
| ArXiv | T5 | 11.50 | 24.53 | 31.86 | 36.04 | 32.64 | 37.79 |
| BART | 14.97 | 25.29 | 37.86 | 42.99 | 36.43 | 43.95 | |
| MediaSum | T5 | 12.89 | 25.72 | 26.29 | 28.59 | 24.43 | 29.25 |
| BART | 14.70 | 24.75 | 28.46 | 32.36 | 28.40 | 33.88 |
| Setting | ROUGE1 | Setting | ROUGE1 |
| SRformer | 39.39 | - | - |
| (I) | 36.86 | (II) | 36.77 |
| (III) | 35.84 | (IV) | 29.72 |
| Model | Mem | Time | Power |
| BART | 49 | 91 | 99 |
| TransformerXL | 82 | 70 | 89 |
| LongT5 | 73 | 80 | 90 |
| SRformer | 60 | 60 | 80 |
| Comp. complexity | Example (K) | Memory complexity | Example (K) | |
| Transformer | O(qkd) | qkd = 8389 | O(qk + kd) | 197 |
| Segmented | O(qsd) | qsd = 524 | O(qs + kd) | 74 |
| SRformer | O(qsd + kd2) | qsd + kd2 = 4718 | O(qs + kd + md2) | 139 |
| SRformer | Segmented | Baseline | |||||
| Dataset | Model | ROUGE2 | ROUGEL | ROUGELsum | ROUGE1 | ROUGE1 | ROUGE1 |
| CNN-DM | T5 | 16.79 | 27.29 | 36.81 | 39.39 | 33.31 | 41.60 |
| BART | 19.80 | 29.44 | 40.40 | 43.19 | 33.87 | 44.54 | |
| XSUM | T5 | 11.72 | 26.88 | 26.88 | 34.48 | 27.03 | 35.35 |
| BART | 16.58 | 31.54 | 31.54 | 39.03 | 36.22 | 41.33 | |
| ArXiv | T5 | 11.50 | 24.53 | 31.86 | 36.04 | 32.64 | 37.79 |
| BART | 14.97 | 25.29 | 37.86 | 42.99 | 36.43 | 43.95 | |
| MediaSum | T5 | 12.89 | 25.72 | 26.29 | 28.59 | 24.43 | 29.25 |
| BART | 14.70 | 24.75 | 28.46 | 32.36 | 28.40 | 33.88 | |
| Setting | ROUGE1 | Setting | ROUGE1 |
| SRformer | 39.39 | - | - |
| (I) | 36.86 | (II) | 36.77 |
| (III) | 35.84 | (IV) | 29.72 |
| Model | Mem | Time | Power |
| BART | 49 | 91 | 99 |
| TransformerXL | 82 | 70 | 89 |
| LongT5 | 73 | 80 | 90 |
| SRformer | 60 | 60 | 80 |
| Comp. complexity | Example (K) | Memory complexity | Example (K) | |
| Transformer | O(qkd) | qkd = 8389 | O(qk + kd) | 197 |
| Segmented | O(qsd) | qsd = 524 | O(qs + kd) | 74 |
| SRformer | O(qsd + kd2) | qsd + kd2 = 4718 | O(qs + kd + md2) | 139 |
| SRformer | Segmented | Baseline | |||||
| Dataset | Model | ROUGE2 | ROUGEL | ROUGELsum | ROUGE1 | ROUGE1 | ROUGE1 |
| CNN-DM | T5 | 16.79 | 27.29 | 36.81 | 39.39 | 33.31 | 41.60 |
| BART | 19.80 | 29.44 | 40.40 | 43.19 | 33.87 | 44.54 | |
| XSUM | T5 | 11.72 | 26.88 | 26.88 | 34.48 | 27.03 | 35.35 |
| BART | 16.58 | 31.54 | 31.54 | 39.03 | 36.22 | 41.33 | |
| ArXiv | T5 | 11.50 | 24.53 | 31.86 | 36.04 | 32.64 | 37.79 |
| BART | 14.97 | 25.29 | 37.86 | 42.99 | 36.43 | 43.95 | |
| MediaSum | T5 | 12.89 | 25.72 | 26.29 | 28.59 | 24.43 | 29.25 |
| BART | 14.70 | 24.75 | 28.46 | 32.36 | 28.40 | 33.88 |
| Setting | ROUGE1 | Setting | ROUGE1 |
| SRformer | 39.39 | - | - |
| (I) | 36.86 | (II) | 36.77 |
| (III) | 35.84 | (IV) | 29.72 |
| Model | Mem | Time | Power |
| BART | 49 | 91 | 99 |
| TransformerXL | 82 | 70 | 89 |
| LongT5 | 73 | 80 | 90 |
| SRformer | 60 | 60 | 80 |
| Models | Indo-Aryan | Kas | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | cat | glg | ast | oci | ||||
| Vanilla NMT | 40.3 | 46.8 | 22.3 | 40.0 | 49.3 | 47.6 | 29.6 | 21.3 | 33.0 | 41.0 | 40.7 | 33.0 | 37.08 | |
| Word-drop | 39.5 | 47.2 | 21.8 | 40.6 | 49.0 | 47.6 | 28.6 | 20.6 | 37.6 | 43.6 | 43.4 | 36.0 | 37.96 | |
| BPE-drop | 39.1 | 46.8 | 22.6 | 40.4 | 48.7 | 46.7 | 29.2 | 21.1 | 33.8 | 41.7 | 41.5 | 33.0 | 37.05 | |
| SwitchOut | 36.1 | 43.2 | 20.1 | 38.2 | 45.6 | 42.7 | 28.3 | 18.8 | 29.0 | 34.9 | 34.9 | 29.1 | 33.41 | |
| OBPE | 41.3 | 47.5 | 23.4 | 41.8 | 50.4 | 49.7 | 30.5 | 21.1 | 34.1 | 41.2 | 41.3 | 33.8 | 38.00 | |
| BPE-Dropout | 39.8 | 47.4 | 22.5 | 39.9 | 49.6 | 47.7 | 29.3 | 21.2 | 33.2 | 40.8 | 41.4 | 33.0 | 37.15 | |
| Random Char Noise | 40.9 | 48.4 | 23.8 | 40.8 | 50.0 | 47.5 | 31.2 | 21.9 | 40.9 | 46.1 | 46.4 | 38.2 | 39.68 |
| SELECTNOISE + Greedy | 42.1 | 51.0 | 25.2 | 43.4 | 51.7 | 49.9 | 33.4 | 23.7 | 42.0 | 47.1 | 47.4 | 38.5 | 41.28 |
| SELECTNOISE + Top-k | 42.4 | 49.9 | 26.0 | 43.0 | 51.0 | 48.8 | 33.4 | 23.3 | 41.5 | 47.1 | 47.8 | 38.5 | 41.06 |
| SELECTNOISE + Top-p | 42.0 | 49.6 | 24.1 | 42.4 | 50.6 | 48.8 | 33.6 | 23.3 | 41.6 | 47.1 | 47.5 | 38.8 | 40.78 |
| Supervised Noise Injection Model | |||||||||||||
| Selective noise + Greedy | 41.4 | 49.1 | 25.4 | 42.2 | 50.1 | 48.7 | 32.9 | 22.2 | 41.6 | 47.2 | 47.7 | 38.7 | 40.60 |
| Selective noise + Top-k | 41.7 | 49.3 | 26.3 | 43.3 | 50.8 | 48.7 | 34.2 | 23.6 | 41.9 | 46.8 | 47.5 | 38.7 | 41.10 |
| Selective noise + Top-p | 41.4 | 49.9 | 27.3 | 43.3 | 51.6 | 48.9 | 33.9 | 23.4 | 41.6 | 47.7 | 48.2 | 39.0 | 41.35 |
| Models | bho | hne | san | Indo-Aryan | awa | npi | kas | cat | Romance | Average | |||
| mai | mag | glg | ast | oci | |||||||||
| Vanilla NMT | 11.1 | 17.2 | 2.7 | 10.1 | 18.5 | 18.3 | 5.1 | 2.6 | 5.3 | 10.1 | 12.3 | 5.2 | 9.86 |
| Word-drop | 8.7 | 13.7 | 1.9 | 7.7 | 15.2 | 16.1 | 3.0 | 1.6 | 6.9 | 10.7 | 13.3 | 6.5 | 8.76 |
| BPE-drop | 10.8 | 16.1 | 2.7 | 10.0 | 17.2 | 17.8 | 4.0 | 2.1 | 5.1 | 9.1 | 11.2 | 4.7 | 9.23 |
| SwitchOut | 4.3 | 7.7 | 1.4 | 4.9 | 8.4 | 7.9 | 2.9 | 1.2 | 3.5 | 6.3 | 8.2 | 3.8 | 5.04 |
| OBPE | 11.1 | 16.6 | 2.9 | 10.4 | 18.7 | 19.7 | 4.8 | 1.9 | 6.2 | 10.7 | 12.9 | 6.1 | 10.16 |
| BPE-Dropout | 11.6 | 17.5 | 3.1 | 10.1 | 19.3 | 18.3 | 5.4 | 2.5 | 5.4 | 10.1 | 13.0 | 5.4 | 10.14 |
| Random Char Noise | 12.8 | 18.8 | 3.1 | 10.2 | 19.4 | 18.6 | 6.3 | 2.9 | 10.9 | 14.3 | 17.2 | 8.7 | 11.93 |
| SELECTNOISE + Greedy | 12.5 | 20.1 | 3.7 | 11.9 | 21.2 | 20.2 | 7.1 | 3.0 | 10.8 | 15.0 | 17.4 | 9.0 | 12.66 |
| SELECTNOISE + Top-k | 12.3 | 19.7 | 3.8 | 12.0 | 20.2 | 19.5 | 7.2 | 2.8 | 10.5 | 15.0 | 17.5 | 8.8 | 12.44 |
| SELECTNOISE + Top-p | 12.7 | 19.5 | 3.8 | 11.9 | 20.3 | 19.6 | 6.7 | 3.2 | 10.7 | 14.8 | 17.1 | 8.9 | 12.43 |
| Selective noise + Greedy | 13.1 | 19.5 | 4.0 | 11.8 | 19.6 | 19.3 | 6.8 | 2.4 | 10.5 | 15.0 | 17.9 | 8.9 | 12.4 |
| Selective noise + Top-k | 12.7 | 19.1 | 3.9 | 12.2 | 20.1 | 19.3 | 7.0 | 2.9 | 10.8 | 15.0 | 17.4 | 8.9 | 12.44 |
| Selective noise + Top-p | 12.7 | 20.0 | 4.1 | 12.6 | 21.2 | 19.7 | 7.0 | 2.7 | 10.5 | 15.4 | 18.1 | 9.1 | 12.76 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.500 | 0.531 | 0.368 | 0.500 | 0.559 | 0.576 | 0.435 | 0.377 | 0.295 | 0.390 | 0.406 | 0.232 | 0.431 |
| Word-drop | 0.497 | 0.533 | 0.357 | 0.498 | 0.551 | 0.563 | 0.417 | 0.353 | 0.361 | 0.440 | 0.454 | 0.312 | 0.445 |
| BPE-drop | 0.506 | 0.537 | 0.367 | 0.509 | 0.557 | 0.572 | 0.422 | 0.363 | 0.316 | 0.415 | 0.432 | 0.283 | 0.440 |
| SwitchOut | 0.411 | 0.446 | 0.318 | 0.415 | 0.467 | 0.466 | 0.38 | 0.335 | 0.278 | 0.337 | 0.347 | 0.262 | 0.372 |
| OBPE | 0.502 | 0.525 | 0.371 | 0.502 | 0.561 | 0.583 | 0.436 | 0.381 | 0.306 | 0.404 | 0.416 | 0.266 | 0.438 |
| BPE-Dropout | 0.501 | 0.526 | 0.371 | 0.497 | 0.558 | 0.574 | 0.439 | 0.393 | 0.300 | 0.389 | 0.410 | 0.231 | 0.432 |
| Random Char Noise | 0.521 | 0.547 | 0.371 | 0.501 | 0.569 | 0.584 | 0.441 | 0.380 | 0.391 | 0.487 | 0.491 | 0.319 | 0.467 |
| SELECTNOISE + Greedy | 0.525 | 0.563 | 0.386 | 0.511 | 0.578 | 0.606 | 0.458 | 0.394 | 0.392 | 0.499 | 0.511 | 0.319 | 0.478 |
| SELECTNOISE + Top-k | 0.524 | 0.558 | 0.386 | 0.507 | 0.576 | 0.599 | 0.454 | 0.388 | 0.400 | 0.497 | 0.516 | 0.321 | 0.477 |
| SELECTNOISE + Top-p | 0.527 | 0.599 | 0.372 | 0.505 | 0.573 | 0.599 | 0.457 | 0.391 | 0.399 | 0.501 | 0.509 | 0.321 | 0.479 |
| Selective noise + Greedy | 0.527 | 0.560 | 0.389 | 0.507 | 0.572 | 0.600 | 0.451 | 0.381 | 0.392 | 0.499 | 0.511 | 0.319 | 0.476 |
| Selective noise + Top-k | 0.526 | 0.549 | 0.401 | 0.509 | 0.573 | 0.463 | 0.463 | 0.390 | 0.400 | 0.494 | 0.506 | 0.326 | 0.467 |
| Selective noise + Top-p | 0.524 | 0.558 | 0.400 | 0.510 | 0.584 | 0.455 | 0.455 | 0.386 | 0.391 | 0.501 | 0.512 | 0.321 | 0.466 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.642 | 0.676 | 0.471 | 0.621 | 0.711 | 0.736 | 0.542 | 0.387 | 0.499 | 0.534 | 0.497 | 0.408 | 0.560 |
| Word-drop | 0.659 | 0.702 | 0.494 | 0.650 | 0.725 | 0.747 | 0.564 | 0.409 | 0.484 | 0.551 | 0.538 | 0.421 | 0.579 |
| BPE-drop | 0.653 | 0.687 | 0.497 | 0.645 | 0.711 | 0.732 | 0.554 | 0.400 | 0.438 | 0.515 | 0.505 | 0.389 | 0.560 |
| SwitchOut | 0.565 | 0.605 | 0.462 | 0.564 | 0.626 | 0.632 | 0.533 | 0.394 | 0.405 | 0.461 | 0.445 | 0.362 | 0.504 |
| OBPE | 0.664 | 0.676 | 0.452 | 0.630 | 0.707 | 0.740 | 0.544 | 0.392 | 0.456 | 0.524 | 0.501 | 0.400 | 0.557 |
| BPE-Dropout | 0.644 | 0.672 | 0.471 | 0.616 | 0.710 | 0.733 | 0.537 | 0.381 | 0.503 | 0.534 | 0.500 | 0.411 | 0.559 |
| Random Char Noise | 0.673 | 0.700 | 0.492 | 0.641 | 0.725 | 0.746 | 0.559 | 0.401 | 0.522 | 0.610 | 0.584 | 0.441 | 0.591 |
| SELECTNOISE + Greedy | 0.672 | 0.714 | 0.493 | 0.647 | 0.735 | 0.765 | 0.575 | 0.412 | 0.523 | 0.620 | 0.598 | 0.434 | 0.599 |
| SELECTNOISE + Top-k | 0.678 | 0.708 | 0.504 | 0.649 | 0.730 | 0.758 | 0.585 | 0.419 | 0.524 | 0.621 | 0.603 | 0.438 | 0.601 |
| SELECTNOISE + Top-p | 0.677 | 0.559 | 0.502 | 0.643 | 0.730 | 0.758 | 0.586 | 0.411 | 0.526 | 0.625 | 0.600 | 0.442 | 0.588 |
| Selective noise + Greedy | 0.681 | 0.711 | 0.505 | 0.649 | 0.728 | 0.761 | 0.582 | 0.411 | 0.522 | 0.618 | 0.603 | 0.441 | 0.601 |
| Selective noise + Top-k | 0.677 | 0.700 | 0.506 | 0.655 | 0.703 | 0.757 | 0.581 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.598 |
| Selective noise + Top-p | 0.680 | 0.708 | 0.511 | 0.655 | 0.738 | 0.756 | 0.589 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.603 |
| Models | Languages | ||
| bho | san | npi | |
| Annotator set-1 | |||
| Vanilla NMT | 3.54 | 2.42 | 2.21 |
| BPE Dropout | 3.29 | 2.37 | 1.83 |
| SELECTNOISE Model | 4.17 | 2.83 | 2.50 |
| Annotator set-2 | |||
| Vanilla NMT | 3.42 | 1.96 | 2.17 |
| BPE Dropout | 2.79 | 1.83 | 1.96 |
| SELECTNOISE Model | 3.54 | 2.17 | 2.21 |
| Language | Data size | BLEU | chrF |
| hne | 997 | 19.5 | 49.6 |
| 6000 | 20.3 | 50.3 | |
| mai | 997 | 11.9 | 42.4 |
| 6000 | 12.4 | 43.2 | |
| npi | 997 | 6.7 | 33.6 |
| 6000 | 7.2 | 33.8 |
| Language | Model | BLEU | chrF |
| Bodo | Vanilla NMT | 2.4 | 18.2 |
| SELECTNOISE | 2.7 | 18.7 | |
| Tamil | Vanilla NMT | 0.6 | 11.7 |
| SELECTNOISE | 0.9 | 13.3 |
| Models | Languages | |||
| BLEU | chrF | BLEURT | COMET | |
| Hindi (hi) | ||||
| Vanilla NMT | 33.4 | 60.2 | 0.724 | 0.868 |
| Random Char Noise | 33.0 | 59.5 | 0.722 | 0.865 |
| SELECTNOISE | 34.2 | 60.5 | 0.726 | 0.869 |
| Spanish (es) | ||||
| Vanilla NMT | 21.5 | 53.5 | 0.695 | 0.810 |
| Random Char Noise | 21.7 | 53.0 | 0.689 | 0.806 |
| SELECTNOISE | 21.3 | 53.1 | 0.689 | 0.869 |
| Candidate Pool (Sc) Template |
| C1: {I': f1, 'D': f3, 'S': {E1: f4, E2: f5, ...}}, |
| C2: {I': f4, 'D': f3, 'S': {E3: f6, E2: f7, ...}}, |
| Ci: {I': f2, 'D': f4, 'S': {E1: f2, E3: f1, ...}} |
| Cn: {I': f1, 'D': f4, 'S': {E1: f4, E3: f5, ...}} |
| Few Sample Elements of Candidate Pool (Sc) |
| 'O': {I': 62, 'D': 1561, 'S': {'‘ह’: 1482, ...}} |
| '瓦': {I': 92, 'D': 97, 'S': {'‘ो’': 1482, ...}} |
| 'O': {I': 1552, 'D': 15, 'S': {'‘य’: 397, ...}} |
| 'S': {I': 0, 'D': 33, 'S': {} |
| } |
| Languages | #Dev | #Test | HRL | Source |
| bho, mag, mai, npi, awa, san, kas, hne | 997 | 1012 | hi | FLORES-200 |
| ast, cat, glg, oci | 997 | 1012 | es | FLORES-200 |
| HR Lang Pair | #Train | # Dev | #Test | Source |
| hi-en | 10.1M | 997 | 1012 | AI4Bharat |
| es-en | 6.6M | 997 | 1012 | Rapp (2021) |
| BHO: रूप्व केहोँ पाथस्मा,तुका केरागपिकोयपिक रेशेयपि रागपिक रहोँ प्रधान सेलन्दीपय रागपिक रूप्व के रागपिक रूप्व के रागपिक रूप्व के रागपिक रू�विक रूप्व के रागपिक रू�विक रू�विक रू�विक रू�विक रू��िक रूप्व के रागपिक रू�विक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक वस्थिके \nENG: The announcement was made after Trump had a phone conversation with Turkish President Recep Tayyip Erdoğan. |
| RCN: Trump opened up about the phone call with Turkish President Recep Tayyip Erdoğan. |
| UCN: After a phone call with Turkish President Recep Tayyip Erdoğan, Trump announced his decision. |
| SCN: Trump's announcement came after a phone call with Turkish President Recep Tayyip Erdoğan. |
| HNE: पूललमह बलाटयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके \nENG: Police said that the body appeared to have been there for about a day. |
| RCN: The police said the body had been lying on the ground for the past few days. |
| UCN: The police said the bodies had been lying on the road for the past few days. |
| SCN: The police said the body had been lying for the past several days. |
| GLG: A ONU samén ten previsto create un fondo para axudar aes paises afectados polo quentamento global a afrontar o seu impacto. |
| ENG: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| RCN: The UN is also planning to set up a fund to tax years of countries affected by global pole to cope with or have an impact. |
| UCN: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| SCN: The UN also plans to set up a fund to relieve the countries affected by the global warming to face up to their impact. |
| AST: Munchos cudadianos de Bishkek acusaron a los manifestantes sureños del desorde. |
| ENG: Several Bishkek residents blamed protesters from the south for the lawlessness. |
| RCN: Cubans of Bishkek accused southern protestors of the deorde. |
| UCN: Many people in Bishkek accused the Swiss demonstrators of disorder. |
| SCN: Many people in Bishkek accused southern demonstrators of disorder. |
| Models | Indo-Aryan | Kas | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | cat | glg | ast | oci | ||||
| Vanilla NMT | 40.3 | 46.8 | 22.3 | 40.0 | 49.3 | 47.6 | 29.6 | 21.3 | 33.0 | 41.0 | 40.7 | 33.0 | 37.08 | |
| Word-drop | 39.5 | 47.2 | 21.8 | 40.6 | 49.0 | 47.6 | 28.6 | 20.6 | 37.6 | 43.6 | 43.4 | 36.0 | 37.96 | |
| BPE-drop | 39.1 | 46.8 | 22.6 | 40.4 | 48.7 | 46.7 | 29.2 | 21.1 | 33.8 | 41.7 | 41.5 | 33.0 | 37.05 | |
| SwitchOut | 36.1 | 43.2 | 20.1 | 38.2 | 45.6 | 42.7 | 28.3 | 18.8 | 29.0 | 34.9 | 34.9 | 29.1 | 33.41 | |
| OBPE | 41.3 | 47.5 | 23.4 | 41.8 | 50.4 | 49.7 | 30.5 | 21.1 | 34.1 | 41.2 | 41.3 | 33.8 | 38.00 | |
| BPE-Dropout | 39.8 | 47.4 | 22.5 | 39.9 | 49.6 | 47.7 | 29.3 | 21.2 | 33.2 | 40.8 | 41.4 | 33.0 | 37.15 | |
| Random Char Noise | 40.9 | 48.4 | 23.8 | 40.8 | 50.0 | 47.5 | 31.2 | 21.9 | 40.9 | 46.1 | 46.4 | 38.2 | 39.68 |
| SELECTNOISE + Greedy | 42.1 | 51.0 | 25.2 | 43.4 | 51.7 | 49.9 | 33.4 | 23.7 | 42.0 | 47.1 | 47.4 | 38.5 | 41.28 |
| SELECTNOISE + Top-k | 42.4 | 49.9 | 26.0 | 43.0 | 51.0 | 48.8 | 33.4 | 23.3 | 41.5 | 47.1 | 47.8 | 38.5 | 41.06 |
| SELECTNOISE + Top-p | 42.0 | 49.6 | 24.1 | 42.4 | 50.6 | 48.8 | 33.6 | 23.3 | 41.6 | 47.1 | 47.5 | 38.8 | 40.78 |
| Supervised Noise Injection Model | |||||||||||||
| Selective noise + Greedy | 41.4 | 49.1 | 25.4 | 42.2 | 50.1 | 48.7 | 32.9 | 22.2 | 41.6 | 47.2 | 47.7 | 38.7 | 40.60 |
| Selective noise + Top-k | 41.7 | 49.3 | 26.3 | 43.3 | 50.8 | 48.7 | 34.2 | 23.6 | 41.9 | 46.8 | 47.5 | 38.7 | 41.10 |
| Selective noise + Top-p | 41.4 | 49.9 | 27.3 | 43.3 | 51.6 | 48.9 | 33.9 | 23.4 | 41.6 | 47.7 | 48.2 | 39.0 | 41.35 |
| Models | bho | hne | san | Indo-Aryan | awa | npi | kas | cat | Romance | Average | |||
| mai | mag | glg | ast | oci | |||||||||
| Vanilla NMT | 11.1 | 17.2 | 2.7 | 10.1 | 18.5 | 18.3 | 5.1 | 2.6 | 5.3 | 10.1 | 12.3 | 5.2 | 9.86 |
| Word-drop | 8.7 | 13.7 | 1.9 | 7.7 | 15.2 | 16.1 | 3.0 | 1.6 | 6.9 | 10.7 | 13.3 | 6.5 | 8.76 |
| BPE-drop | 10.8 | 16.1 | 2.7 | 10.0 | 17.2 | 17.8 | 4.0 | 2.1 | 5.1 | 9.1 | 11.2 | 4.7 | 9.23 |
| SwitchOut | 4.3 | 7.7 | 1.4 | 4.9 | 8.4 | 7.9 | 2.9 | 1.2 | 3.5 | 6.3 | 8.2 | 3.8 | 5.04 |
| OBPE | 11.1 | 16.6 | 2.9 | 10.4 | 18.7 | 19.7 | 4.8 | 1.9 | 6.2 | 10.7 | 12.9 | 6.1 | 10.16 |
| BPE-Dropout | 11.6 | 17.5 | 3.1 | 10.1 | 19.3 | 18.3 | 5.4 | 2.5 | 5.4 | 10.1 | 13.0 | 5.4 | 10.14 |
| Random Char Noise | 12.8 | 18.8 | 3.1 | 10.2 | 19.4 | 18.6 | 6.3 | 2.9 | 10.9 | 14.3 | 17.2 | 8.7 | 11.93 |
| SELECTNOISE + Greedy | 12.5 | 20.1 | 3.7 | 11.9 | 21.2 | 20.2 | 7.1 | 3.0 | 10.8 | 15.0 | 17.4 | 9.0 | 12.66 |
| SELECTNOISE + Top-k | 12.3 | 19.7 | 3.8 | 12.0 | 20.2 | 19.5 | 7.2 | 2.8 | 10.5 | 15.0 | 17.5 | 8.8 | 12.44 |
| SELECTNOISE + Top-p | 12.7 | 19.5 | 3.8 | 11.9 | 20.3 | 19.6 | 6.7 | 3.2 | 10.7 | 14.8 | 17.1 | 8.9 | 12.43 |
| Selective noise + Greedy | 13.1 | 19.5 | 4.0 | 11.8 | 19.6 | 19.3 | 6.8 | 2.4 | 10.5 | 15.0 | 17.9 | 8.9 | 12.4 |
| Selective noise + Top-k | 12.7 | 19.1 | 3.9 | 12.2 | 20.1 | 19.3 | 7.0 | 2.9 | 10.8 | 15.0 | 17.4 | 8.9 | 12.44 |
| Selective noise + Top-p | 12.7 | 20.0 | 4.1 | 12.6 | 21.2 | 19.7 | 7.0 | 2.7 | 10.5 | 15.4 | 18.1 | 9.1 | 12.76 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.500 | 0.531 | 0.368 | 0.500 | 0.559 | 0.576 | 0.435 | 0.377 | 0.295 | 0.390 | 0.406 | 0.232 | 0.431 |
| Word-drop | 0.497 | 0.533 | 0.357 | 0.498 | 0.551 | 0.563 | 0.417 | 0.353 | 0.361 | 0.440 | 0.454 | 0.312 | 0.445 |
| BPE-drop | 0.506 | 0.537 | 0.367 | 0.509 | 0.557 | 0.572 | 0.422 | 0.363 | 0.316 | 0.415 | 0.432 | 0.283 | 0.440 |
| SwitchOut | 0.411 | 0.446 | 0.318 | 0.415 | 0.467 | 0.466 | 0.38 | 0.335 | 0.278 | 0.337 | 0.347 | 0.262 | 0.372 |
| OBPE | 0.502 | 0.525 | 0.371 | 0.502 | 0.561 | 0.583 | 0.436 | 0.381 | 0.306 | 0.404 | 0.416 | 0.266 | 0.438 |
| BPE-Dropout | 0.501 | 0.526 | 0.371 | 0.497 | 0.558 | 0.574 | 0.439 | 0.393 | 0.300 | 0.389 | 0.410 | 0.231 | 0.432 |
| Random Char Noise | 0.521 | 0.547 | 0.371 | 0.501 | 0.569 | 0.584 | 0.441 | 0.380 | 0.391 | 0.487 | 0.491 | 0.319 | 0.467 |
| SELECTNOISE + Greedy | 0.525 | 0.563 | 0.386 | 0.511 | 0.578 | 0.606 | 0.458 | 0.394 | 0.392 | 0.499 | 0.511 | 0.319 | 0.478 |
| SELECTNOISE + Top-k | 0.524 | 0.558 | 0.386 | 0.507 | 0.576 | 0.599 | 0.454 | 0.388 | 0.400 | 0.497 | 0.516 | 0.321 | 0.477 |
| SELECTNOISE + Top-p | 0.527 | 0.599 | 0.372 | 0.505 | 0.573 | 0.599 | 0.457 | 0.391 | 0.399 | 0.501 | 0.509 | 0.321 | 0.479 |
| Selective noise + Greedy | 0.527 | 0.560 | 0.389 | 0.507 | 0.572 | 0.600 | 0.451 | 0.381 | 0.392 | 0.499 | 0.511 | 0.319 | 0.476 |
| Selective noise + Top-k | 0.526 | 0.549 | 0.401 | 0.509 | 0.573 | 0.463 | 0.463 | 0.390 | 0.400 | 0.494 | 0.506 | 0.326 | 0.467 |
| Selective noise + Top-p | 0.524 | 0.558 | 0.400 | 0.510 | 0.584 | 0.455 | 0.455 | 0.386 | 0.391 | 0.501 | 0.512 | 0.321 | 0.466 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.642 | 0.676 | 0.471 | 0.621 | 0.711 | 0.736 | 0.542 | 0.387 | 0.499 | 0.534 | 0.497 | 0.408 | 0.560 |
| Word-drop | 0.659 | 0.702 | 0.494 | 0.650 | 0.725 | 0.747 | 0.564 | 0.409 | 0.484 | 0.551 | 0.538 | 0.421 | 0.579 |
| BPE-drop | 0.653 | 0.687 | 0.497 | 0.645 | 0.711 | 0.732 | 0.554 | 0.400 | 0.438 | 0.515 | 0.505 | 0.389 | 0.560 |
| SwitchOut | 0.565 | 0.605 | 0.462 | 0.564 | 0.626 | 0.632 | 0.533 | 0.394 | 0.405 | 0.461 | 0.445 | 0.362 | 0.504 |
| OBPE | 0.664 | 0.676 | 0.452 | 0.630 | 0.707 | 0.740 | 0.544 | 0.392 | 0.456 | 0.524 | 0.501 | 0.400 | 0.557 |
| BPE-Dropout | 0.644 | 0.672 | 0.471 | 0.616 | 0.710 | 0.733 | 0.537 | 0.381 | 0.503 | 0.534 | 0.500 | 0.411 | 0.559 |
| Random Char Noise | 0.673 | 0.700 | 0.492 | 0.641 | 0.725 | 0.746 | 0.559 | 0.401 | 0.522 | 0.610 | 0.584 | 0.441 | 0.591 |
| SELECTNOISE + Greedy | 0.672 | 0.714 | 0.493 | 0.647 | 0.735 | 0.765 | 0.575 | 0.412 | 0.523 | 0.620 | 0.598 | 0.434 | 0.599 |
| SELECTNOISE + Top-k | 0.678 | 0.708 | 0.504 | 0.649 | 0.730 | 0.758 | 0.585 | 0.419 | 0.524 | 0.621 | 0.603 | 0.438 | 0.601 |
| SELECTNOISE + Top-p | 0.677 | 0.559 | 0.502 | 0.643 | 0.730 | 0.758 | 0.586 | 0.411 | 0.526 | 0.625 | 0.600 | 0.442 | 0.588 |
| Selective noise + Greedy | 0.681 | 0.711 | 0.505 | 0.649 | 0.728 | 0.761 | 0.582 | 0.411 | 0.522 | 0.618 | 0.603 | 0.441 | 0.601 |
| Selective noise + Top-k | 0.677 | 0.700 | 0.506 | 0.655 | 0.703 | 0.757 | 0.581 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.598 |
| Selective noise + Top-p | 0.680 | 0.708 | 0.511 | 0.655 | 0.738 | 0.756 | 0.589 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.603 |
| Models | Languages | ||
| bho | san | npi | |
| Annotator set-1 | |||
| Vanilla NMT | 3.54 | 2.42 | 2.21 |
| BPE Dropout | 3.29 | 2.37 | 1.83 |
| SELECTNOISE Model | 4.17 | 2.83 | 2.50 |
| Annotator set-2 | |||
| Vanilla NMT | 3.42 | 1.96 | 2.17 |
| BPE Dropout | 2.79 | 1.83 | 1.96 |
| SELECTNOISE Model | 3.54 | 2.17 | 2.21 |
| Language | Data size | BLEU | chrF |
| hne | 997 | 19.5 | 49.6 |
| 6000 | 20.3 | 50.3 | |
| mai | 997 | 11.9 | 42.4 |
| 6000 | 12.4 | 43.2 | |
| npi | 997 | 6.7 | 33.6 |
| 6000 | 7.2 | 33.8 |
| Language | Model | BLEU | chrF |
| Bodo | Vanilla NMT | 2.4 | 18.2 |
| SELECTNOISE | 2.7 | 18.7 | |
| Tamil | Vanilla NMT | 0.6 | 11.7 |
| SELECTNOISE | 0.9 | 13.3 |
| Models | Languages | |||
| BLEU | chrF | BLEURT | COMET | |
| Hindi (hi) | ||||
| Vanilla NMT | 33.4 | 60.2 | 0.724 | 0.868 |
| Random Char Noise | 33.0 | 59.5 | 0.722 | 0.865 |
| SELECTNOISE | 34.2 | 60.5 | 0.726 | 0.869 |
| Spanish (es) | ||||
| Vanilla NMT | 21.5 | 53.5 | 0.695 | 0.810 |
| Random Char Noise | 21.7 | 53.0 | 0.689 | 0.806 |
| SELECTNOISE | 21.3 | 53.1 | 0.689 | 0.869 |
| Candidate Pool (Sc) Template |
| C1: {I': f1, 'D': f3, 'S': {E1: f4, E2: f5, ...}}, |
| C2: {I': f4, 'D': f3, 'S': {E3: f6, E2: f7, ...}}, |
| Ci: {I': f2, 'D': f4, 'S': {E1: f2, E3: f1, ...}} |
| Cn: {I': f1, 'D': f4, 'S': {E1: f4, E3: f5, ...}} |
| Few Sample Elements of Candidate Pool (Sc) |
| 'O': {I': 62, 'D': 1561, 'S': {'‘ह’: 1482, ...}} |
| '瓦': {I': 92, 'D': 97, 'S': {'‘ो’': 1482, ...}} |
| 'O': {I': 1552, 'D': 15, 'S': {'‘य’: 397, ...}} |
| 'S': {I': 0, 'D': 33, 'S': {} |
| } |
| Languages | #Dev | #Test | HRL | Source |
| bho, mag, mai, npi, awa, san, kas, hne | 997 | 1012 | hi | FLORES-200 |
| ast, cat, glg, oci | 997 | 1012 | es | FLORES-200 |
| HR Lang Pair | #Train | # Dev | #Test | Source |
| hi-en | 10.1M | 997 | 1012 | AI4Bharat |
| es-en | 6.6M | 997 | 1012 | Rapp (2021) |
| BHO: रूप्व केहोँ पाथस्मा,तुका केरागपिकोयपिक रेशेयपि रागपिक रहोँ प्रधान सेलन्दीपय रागपिक रूप्व के रागपिक रूप्व के रागपिक रूप्व के रागपिक रू�विक रूप्व के रागपिक रू�विक रू�विक रू�विक रू�विक रू��िक रूप्व के रागपिक रू�विक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक वस्थिके \nENG: The announcement was made after Trump had a phone conversation with Turkish President Recep Tayyip Erdoğan. |
| RCN: Trump opened up about the phone call with Turkish President Recep Tayyip Erdoğan. |
| UCN: After a phone call with Turkish President Recep Tayyip Erdoğan, Trump announced his decision. |
| SCN: Trump's announcement came after a phone call with Turkish President Recep Tayyip Erdoğan. |
| HNE: पूललमह बलाटयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके \nENG: Police said that the body appeared to have been there for about a day. |
| RCN: The police said the body had been lying on the ground for the past few days. |
| UCN: The police said the bodies had been lying on the road for the past few days. |
| SCN: The police said the body had been lying for the past several days. |
| GLG: A ONU samén ten previsto create un fondo para axudar aes paises afectados polo quentamento global a afrontar o seu impacto. |
| ENG: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| RCN: The UN is also planning to set up a fund to tax years of countries affected by global pole to cope with or have an impact. |
| UCN: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| SCN: The UN also plans to set up a fund to relieve the countries affected by the global warming to face up to their impact. |
| AST: Munchos cudadianos de Bishkek acusaron a los manifestantes sureños del desorde. |
| ENG: Several Bishkek residents blamed protesters from the south for the lawlessness. |
| RCN: Cubans of Bishkek accused southern protestors of the deorde. |
| UCN: Many people in Bishkek accused the Swiss demonstrators of disorder. |
| SCN: Many people in Bishkek accused southern demonstrators of disorder. |
| Models | Indo-Aryan | Kas | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | cat | glg | ast | oci | ||||
| Vanilla NMT | 40.3 | 46.8 | 22.3 | 40.0 | 49.3 | 47.6 | 29.6 | 21.3 | 33.0 | 41.0 | 40.7 | 33.0 | 37.08 | |
| Word-drop | 39.5 | 47.2 | 21.8 | 40.6 | 49.0 | 47.6 | 28.6 | 20.6 | 37.6 | 43.6 | 43.4 | 36.0 | 37.96 | |
| BPE-drop | 39.1 | 46.8 | 22.6 | 40.4 | 48.7 | 46.7 | 29.2 | 21.1 | 33.8 | 41.7 | 41.5 | 33.0 | 37.05 | |
| SwitchOut | 36.1 | 43.2 | 20.1 | 38.2 | 45.6 | 42.7 | 28.3 | 18.8 | 29.0 | 34.9 | 34.9 | 29.1 | 33.41 | |
| OBPE | 41.3 | 47.5 | 23.4 | 41.8 | 50.4 | 49.7 | 30.5 | 21.1 | 34.1 | 41.2 | 41.3 | 33.8 | 38.00 | |
| BPE-Dropout | 39.8 | 47.4 | 22.5 | 39.9 | 49.6 | 47.7 | 29.3 | 21.2 | 33.2 | 40.8 | 41.4 | 33.0 | 37.15 | |
| Random Char Noise | 40.9 | 48.4 | 23.8 | 40.8 | 50.0 | 47.5 | 31.2 | 21.9 | 40.9 | 46.1 | 46.4 | 38.2 | 39.68 | |
| SELECTNOISE + Greedy | 42.1 | 51.0 | 25.2 | 43.4 | 51.7 | 49.9 | 33.4 | 23.7 | 42.0 | 47.1 | 47.4 | 38.5 | 41.28 |
| SELECTNOISE + Top-k | 42.4 | 49.9 | 26.0 | 43.0 | 51.0 | 48.8 | 33.4 | 23.3 | 41.5 | 47.1 | 47.8 | 38.5 | 41.06 |
| SELECTNOISE + Top-p | 42.0 | 49.6 | 24.1 | 42.4 | 50.6 | 48.8 | 33.6 | 23.3 | 41.6 | 47.1 | 47.5 | 38.8 | 40.78 |
| Supervised Noise Injection Model | |||||||||||||
| Selective noise + Greedy | 41.4 | 49.1 | 25.4 | 42.2 | 50.1 | 48.7 | 32.9 | 22.2 | 41.6 | 47.2 | 47.7 | 38.7 | 40.60 |
| Selective noise + Top-k | 41.7 | 49.3 | 26.3 | 43.3 | 50.8 | 48.7 | 34.2 | 23.6 | 41.9 | 46.8 | 47.5 | 38.7 | 41.10 |
| Selective noise + Top-p | 41.4 | 49.9 | 27.3 | 43.3 | 51.6 | 48.9 | 33.9 | 23.4 | 41.6 | 47.7 | 48.2 | 39.0 | 41.35 |
| Models | bho | hne | san | Indo-Aryan | awa | npi | kas | cat | Romance | Average | |||
| mai | mag | glg | ast | oci | |||||||||
| Vanilla NMT | 11.1 | 17.2 | 2.7 | 10.1 | 18.5 | 18.3 | 5.1 | 2.6 | 5.3 | 10.1 | 12.3 | 5.2 | 9.86 |
| Word-drop | 8.7 | 13.7 | 1.9 | 7.7 | 15.2 | 16.1 | 3.0 | 1.6 | 6.9 | 10.7 | 13.3 | 6.5 | 8.76 |
| BPE-drop | 10.8 | 16.1 | 2.7 | 10.0 | 17.2 | 17.8 | 4.0 | 2.1 | 5.1 | 9.1 | 11.2 | 4.7 | 9.23 |
| SwitchOut | 4.3 | 7.7 | 1.4 | 4.9 | 8.4 | 7.9 | 2.9 | 1.2 | 3.5 | 6.3 | 8.2 | 3.8 | 5.04 |
| OBPE | 11.1 | 16.6 | 2.9 | 10.4 | 18.7 | 19.7 | 4.8 | 1.9 | 6.2 | 10.7 | 12.9 | 6.1 | 10.16 |
| BPE-Dropout | 11.6 | 17.5 | 3.1 | 10.1 | 19.3 | 18.3 | 5.4 | 2.5 | 5.4 | 10.1 | 13.0 | 5.4 | 10.14 |
| Random Char Noise | 12.8 | 18.8 | 3.1 | 10.2 | 19.4 | 18.6 | 6.3 | 2.9 | 10.9 | 14.3 | 17.2 | 8.7 | 11.93 |
| SELECTNOISE + Greedy | 12.5 | 20.1 | 3.7 | 11.9 | 21.2 | 20.2 | 7.1 | 3.0 | 10.8 | 15.0 | 17.4 | 9.0 | 12.66 |
| SELECTNOISE + Top-k | 12.3 | 19.7 | 3.8 | 12.0 | 20.2 | 19.5 | 7.2 | 2.8 | 10.5 | 15.0 | 17.5 | 8.8 | 12.44 |
| SELECTNOISE + Top-p | 12.7 | 19.5 | 3.8 | 11.9 | 20.3 | 19.6 | 6.7 | 3.2 | 10.7 | 14.8 | 17.1 | 8.9 | 12.43 |
| Selective noise + Greedy | 13.1 | 19.5 | 4.0 | 11.8 | 19.6 | 19.3 | 6.8 | 2.4 | 10.5 | 15.0 | 17.9 | 8.9 | 12.4 |
| Selective noise + Top-k | 12.7 | 19.1 | 3.9 | 12.2 | 20.1 | 19.3 | 7.0 | 2.9 | 10.8 | 15.0 | 17.4 | 8.9 | 12.44 |
| Selective noise + Top-p | 12.7 | 20.0 | 4.1 | 12.6 | 21.2 | 19.7 | 7.0 | 2.7 | 10.5 | 15.4 | 18.1 | 9.1 | 12.76 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.500 | 0.531 | 0.368 | 0.500 | 0.559 | 0.576 | 0.435 | 0.377 | 0.295 | 0.390 | 0.406 | 0.232 | 0.431 |
| Word-drop | 0.497 | 0.533 | 0.357 | 0.498 | 0.551 | 0.563 | 0.417 | 0.353 | 0.361 | 0.440 | 0.454 | 0.312 | 0.445 |
| BPE-drop | 0.506 | 0.537 | 0.367 | 0.509 | 0.557 | 0.572 | 0.422 | 0.363 | 0.316 | 0.415 | 0.432 | 0.283 | 0.440 |
| SwitchOut | 0.411 | 0.446 | 0.318 | 0.415 | 0.467 | 0.466 | 0.38 | 0.335 | 0.278 | 0.337 | 0.347 | 0.262 | 0.372 |
| OBPE | 0.502 | 0.525 | 0.371 | 0.502 | 0.561 | 0.583 | 0.436 | 0.381 | 0.306 | 0.404 | 0.416 | 0.266 | 0.438 |
| BPE-Dropout | 0.501 | 0.526 | 0.371 | 0.497 | 0.558 | 0.574 | 0.439 | 0.393 | 0.300 | 0.389 | 0.410 | 0.231 | 0.432 |
| Random Char Noise | 0.521 | 0.547 | 0.371 | 0.501 | 0.569 | 0.584 | 0.441 | 0.380 | 0.391 | 0.487 | 0.491 | 0.319 | 0.467 |
| SELECTNOISE + Greedy | 0.525 | 0.563 | 0.386 | 0.511 | 0.578 | 0.606 | 0.458 | 0.394 | 0.392 | 0.499 | 0.511 | 0.319 | 0.478 |
| SELECTNOISE + Top-k | 0.524 | 0.558 | 0.386 | 0.507 | 0.576 | 0.599 | 0.454 | 0.388 | 0.400 | 0.497 | 0.516 | 0.321 | 0.477 |
| SELECTNOISE + Top-p | 0.527 | 0.599 | 0.372 | 0.505 | 0.573 | 0.599 | 0.457 | 0.391 | 0.399 | 0.501 | 0.509 | 0.321 | 0.479 |
| Selective noise + Greedy | 0.527 | 0.560 | 0.389 | 0.507 | 0.572 | 0.600 | 0.451 | 0.381 | 0.392 | 0.499 | 0.511 | 0.319 | 0.476 |
| Selective noise + Top-k | 0.526 | 0.549 | 0.401 | 0.509 | 0.573 | 0.463 | 0.463 | 0.390 | 0.400 | 0.494 | 0.506 | 0.326 | 0.467 |
| Selective noise + Top-p | 0.524 | 0.558 | 0.400 | 0.510 | 0.584 | 0.455 | 0.455 | 0.386 | 0.391 | 0.501 | 0.512 | 0.321 | 0.466 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.642 | 0.676 | 0.471 | 0.621 | 0.711 | 0.736 | 0.542 | 0.387 | 0.499 | 0.534 | 0.497 | 0.408 | 0.560 |
| Word-drop | 0.659 | 0.702 | 0.494 | 0.650 | 0.725 | 0.747 | 0.564 | 0.409 | 0.484 | 0.551 | 0.538 | 0.421 | 0.579 |
| BPE-drop | 0.653 | 0.687 | 0.497 | 0.645 | 0.711 | 0.732 | 0.554 | 0.400 | 0.438 | 0.515 | 0.505 | 0.389 | 0.560 |
| SwitchOut | 0.565 | 0.605 | 0.462 | 0.564 | 0.626 | 0.632 | 0.533 | 0.394 | 0.405 | 0.461 | 0.445 | 0.362 | 0.504 |
| OBPE | 0.664 | 0.676 | 0.452 | 0.630 | 0.707 | 0.740 | 0.544 | 0.392 | 0.456 | 0.524 | 0.501 | 0.400 | 0.557 |
| BPE-Dropout | 0.644 | 0.672 | 0.471 | 0.616 | 0.710 | 0.733 | 0.537 | 0.381 | 0.503 | 0.534 | 0.500 | 0.411 | 0.559 |
| Random Char Noise | 0.673 | 0.700 | 0.492 | 0.641 | 0.725 | 0.746 | 0.559 | 0.401 | 0.522 | 0.610 | 0.584 | 0.441 | 0.591 |
| SELECTNOISE + Greedy | 0.672 | 0.714 | 0.493 | 0.647 | 0.735 | 0.765 | 0.575 | 0.412 | 0.523 | 0.620 | 0.598 | 0.434 | 0.599 |
| SELECTNOISE + Top-k | 0.678 | 0.708 | 0.504 | 0.649 | 0.730 | 0.758 | 0.585 | 0.419 | 0.524 | 0.621 | 0.603 | 0.438 | 0.601 |
| SELECTNOISE + Top-p | 0.677 | 0.559 | 0.502 | 0.643 | 0.730 | 0.758 | 0.586 | 0.411 | 0.526 | 0.625 | 0.600 | 0.442 | 0.588 |
| Selective noise + Greedy | 0.681 | 0.711 | 0.505 | 0.649 | 0.728 | 0.761 | 0.582 | 0.411 | 0.522 | 0.618 | 0.603 | 0.441 | 0.601 |
| Selective noise + Top-k | 0.677 | 0.700 | 0.506 | 0.655 | 0.703 | 0.757 | 0.581 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.598 |
| Selective noise + Top-p | 0.680 | 0.708 | 0.511 | 0.655 | 0.738 | 0.756 | 0.589 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.603 |
| Models | Languages | ||
| bho | san | npi | |
| Annotator set-1 | |||
| Vanilla NMT | 3.54 | 2.42 | 2.21 |
| BPE Dropout | 3.29 | 2.37 | 1.83 |
| SELECTNOISE Model | 4.17 | 2.83 | 2.50 |
| Annotator set-2 | |||
| Vanilla NMT | 3.42 | 1.96 | 2.17 |
| BPE Dropout | 2.79 | 1.83 | 1.96 |
| SELECTNOISE Model | 3.54 | 2.17 | 2.21 |
| Language | Data size | BLEU | chrF |
| hne | 997 | 19.5 | 49.6 |
| 6000 | 20.3 | 50.3 | |
| mai | 997 | 11.9 | 42.4 |
| 6000 | 12.4 | 43.2 | |
| npi | 997 | 6.7 | 33.6 |
| 6000 | 7.2 | 33.8 |
| Language | Model | BLEU | chrF |
| Bodo | Vanilla NMT | 2.4 | 18.2 |
| SELECTNOISE | 2.7 | 18.7 | |
| Tamil | Vanilla NMT | 0.6 | 11.7 |
| SELECTNOISE | 0.9 | 13.3 |
| Models | Languages | |||
| BLEU | chrF | BLEURT | COMET | |
| Hindi (hi) | ||||
| Vanilla NMT | 33.4 | 60.2 | 0.724 | 0.868 |
| Random Char Noise | 33.0 | 59.5 | 0.722 | 0.865 |
| SELECTNOISE | 34.2 | 60.5 | 0.726 | 0.869 |
| Spanish (es) | ||||
| Vanilla NMT | 21.5 | 53.5 | 0.695 | 0.810 |
| Random Char Noise | 21.7 | 53.0 | 0.689 | 0.806 |
| SELECTNOISE | 21.3 | 53.1 | 0.689 | 0.869 |
| Candidate Pool (Sc) Template |
| C1: {I': f1, 'D': f3, 'S': {E1: f4, E2: f5, ...}}, |
| C2: {I': f4, 'D': f3, 'S': {E3: f6, E2: f7, ...}}, |
| Ci: {I': f2, 'D': f4, 'S': {E1: f2, E3: f1, ...}} |
| Cn: {I': f1, 'D': f4, 'S': {E1: f4, E3: f5, ...}} |
| Few Sample Elements of Candidate Pool (Sc) |
| 'O': {I': 62, 'D': 1561, 'S': {'‘ह’: 1482, ...}} |
| '瓦': {I': 92, 'D': 97, 'S': {'‘ो’': 1482, ...}} |
| 'O': {I': 1552, 'D': 15, 'S': {'‘य’: 397, ...}} |
| 'S': {I': 0, 'D': 33, 'S': {} |
| } |
| Languages | #Dev | #Test | HRL | Source |
| bho, mag, mai, npi, awa, san, kas, hne | 997 | 1012 | hi | FLORES-200 |
| ast, cat, glg, oci | 997 | 1012 | es | FLORES-200 |
| HR Lang Pair | #Train | # Dev | #Test | Source |
| hi-en | 10.1M | 997 | 1012 | AI4Bharat |
| es-en | 6.6M | 997 | 1012 | Rapp (2021) |
| BHO: रूप्व केहोँ पाथस्मा,तुका केरागपिकोयपिक रेशेयपि रागपिक रहोँ प्रधान सेलन्दीपय रागपिक रूप्व के रागपिक रूप्व के रागपिक रूप्व के रागपिक रू�विक रूप्व के रागपिक रू�विक रू�विक रू�विक रू�विक रू��िक रूप्व के रागपिक रू�विक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक वस्थिके +ENG: The announcement was made after Trump had a phone conversation with Turkish President Recep Tayyip Erdoğan. |
| RCN: Trump opened up about the phone call with Turkish President Recep Tayyip Erdoğan. |
| UCN: After a phone call with Turkish President Recep Tayyip Erdoğan, Trump announced his decision. |
| SCN: Trump's announcement came after a phone call with Turkish President Recep Tayyip Erdoğan. |
| HNE: पूललमह बलाटयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके +ENG: Police said that the body appeared to have been there for about a day. |
| RCN: The police said the body had been lying on the ground for the past few days. |
| UCN: The police said the bodies had been lying on the road for the past few days. |
| SCN: The police said the body had been lying for the past several days. |
| GLG: A ONU samén ten previsto create un fondo para axudar aes paises afectados polo quentamento global a afrontar o seu impacto. |
| ENG: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| RCN: The UN is also planning to set up a fund to tax years of countries affected by global pole to cope with or have an impact. |
| UCN: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| SCN: The UN also plans to set up a fund to relieve the countries affected by the global warming to face up to their impact. |
| AST: Munchos cudadianos de Bishkek acusaron a los manifestantes sureños del desorde. |
| ENG: Several Bishkek residents blamed protesters from the south for the lawlessness. |
| RCN: Cubans of Bishkek accused southern protestors of the deorde. |
| UCN: Many people in Bishkek accused the Swiss demonstrators of disorder. |
| SCN: Many people in Bishkek accused southern demonstrators of disorder. |
| Models | Indo-Aryan | Kas | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | cat | glg | ast | oci | ||||
| Vanilla NMT | 40.3 | 46.8 | 22.3 | 40.0 | 49.3 | 47.6 | 29.6 | 21.3 | 33.0 | 41.0 | 40.7 | 33.0 | 37.08 | |
| Word-drop | 39.5 | 47.2 | 21.8 | 40.6 | 49.0 | 47.6 | 28.6 | 20.6 | 37.6 | 43.6 | 43.4 | 36.0 | 37.96 | |
| BPE-drop | 39.1 | 46.8 | 22.6 | 40.4 | 48.7 | 46.7 | 29.2 | 21.1 | 33.8 | 41.7 | 41.5 | 33.0 | 37.05 | |
| SwitchOut | 36.1 | 43.2 | 20.1 | 38.2 | 45.6 | 42.7 | 28.3 | 18.8 | 29.0 | 34.9 | 34.9 | 29.1 | 33.41 | |
| OBPE | 41.3 | 47.5 | 23.4 | 41.8 | 50.4 | 49.7 | 30.5 | 21.1 | 34.1 | 41.2 | 41.3 | 33.8 | 38.00 | |
| BPE-Dropout | 39.8 | 47.4 | 22.5 | 39.9 | 49.6 | 47.7 | 29.3 | 21.2 | 33.2 | 40.8 | 41.4 | 33.0 | 37.15 | |
| Random Char Noise | 40.9 | 48.4 | 23.8 | 40.8 | 50.0 | 47.5 | 31.2 | 21.9 | 40.9 | 46.1 | 46.4 | 38.2 | 39.68 |
| SELECTNOISE + Greedy | 42.1 | 51.0 | 25.2 | 43.4 | 51.7 | 49.9 | 33.4 | 23.7 | 42.0 | 47.1 | 47.4 | 38.5 | 41.28 |
| SELECTNOISE + Top-k | 42.4 | 49.9 | 26.0 | 43.0 | 51.0 | 48.8 | 33.4 | 23.3 | 41.5 | 47.1 | 47.8 | 38.5 | 41.06 |
| SELECTNOISE + Top-p | 42.0 | 49.6 | 24.1 | 42.4 | 50.6 | 48.8 | 33.6 | 23.3 | 41.6 | 47.1 | 47.5 | 38.8 | 40.78 |
| Supervised Noise Injection Model | |||||||||||||
| Selective noise + Greedy | 41.4 | 49.1 | 25.4 | 42.2 | 50.1 | 48.7 | 32.9 | 22.2 | 41.6 | 47.2 | 47.7 | 38.7 | 40.60 |
| Selective noise + Top-k | 41.7 | 49.3 | 26.3 | 43.3 | 50.8 | 48.7 | 34.2 | 23.6 | 41.9 | 46.8 | 47.5 | 38.7 | 41.10 |
| Selective noise + Top-p | 41.4 | 49.9 | 27.3 | 43.3 | 51.6 | 48.9 | 33.9 | 23.4 | 41.6 | 47.7 | 48.2 | 39.0 | 41.35 |
| Models | bho | hne | san | Indo-Aryan | awa | npi | kas | cat | Romance | Average | |||
| mai | mag | glg | ast | oci | |||||||||
| Vanilla NMT | 11.1 | 17.2 | 2.7 | 10.1 | 18.5 | 18.3 | 5.1 | 2.6 | 5.3 | 10.1 | 12.3 | 5.2 | 9.86 |
| Word-drop | 8.7 | 13.7 | 1.9 | 7.7 | 15.2 | 16.1 | 3.0 | 1.6 | 6.9 | 10.7 | 13.3 | 6.5 | 8.76 |
| BPE-drop | 10.8 | 16.1 | 2.7 | 10.0 | 17.2 | 17.8 | 4.0 | 2.1 | 5.1 | 9.1 | 11.2 | 4.7 | 9.23 |
| SwitchOut | 4.3 | 7.7 | 1.4 | 4.9 | 8.4 | 7.9 | 2.9 | 1.2 | 3.5 | 6.3 | 8.2 | 3.8 | 5.04 |
| OBPE | 11.1 | 16.6 | 2.9 | 10.4 | 18.7 | 19.7 | 4.8 | 1.9 | 6.2 | 10.7 | 12.9 | 6.1 | 10.16 |
| BPE-Dropout | 11.6 | 17.5 | 3.1 | 10.1 | 19.3 | 18.3 | 5.4 | 2.5 | 5.4 | 10.1 | 13.0 | 5.4 | 10.14 |
| Random Char Noise | 12.8 | 18.8 | 3.1 | 10.2 | 19.4 | 18.6 | 6.3 | 2.9 | 10.9 | 14.3 | 17.2 | 8.7 | 11.93 |
| SELECTNOISE + Greedy | 12.5 | 20.1 | 3.7 | 11.9 | 21.2 | 20.2 | 7.1 | 3.0 | 10.8 | 15.0 | 17.4 | 9.0 | 12.66 |
| SELECTNOISE + Top-k | 12.3 | 19.7 | 3.8 | 12.0 | 20.2 | 19.5 | 7.2 | 2.8 | 10.5 | 15.0 | 17.5 | 8.8 | 12.44 |
| SELECTNOISE + Top-p | 12.7 | 19.5 | 3.8 | 11.9 | 20.3 | 19.6 | 6.7 | 3.2 | 10.7 | 14.8 | 17.1 | 8.9 | 12.43 |
| Selective noise + Greedy | 13.1 | 19.5 | 4.0 | 11.8 | 19.6 | 19.3 | 6.8 | 2.4 | 10.5 | 15.0 | 17.9 | 8.9 | 12.4 |
| Selective noise + Top-k | 12.7 | 19.1 | 3.9 | 12.2 | 20.1 | 19.3 | 7.0 | 2.9 | 10.8 | 15.0 | 17.4 | 8.9 | 12.44 |
| Selective noise + Top-p | 12.7 | 20.0 | 4.1 | 12.6 | 21.2 | 19.7 | 7.0 | 2.7 | 10.5 | 15.4 | 18.1 | 9.1 | 12.76 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.500 | 0.531 | 0.368 | 0.500 | 0.559 | 0.576 | 0.435 | 0.377 | 0.295 | 0.390 | 0.406 | 0.232 | 0.431 |
| Word-drop | 0.497 | 0.533 | 0.357 | 0.498 | 0.551 | 0.563 | 0.417 | 0.353 | 0.361 | 0.440 | 0.454 | 0.312 | 0.445 |
| BPE-drop | 0.506 | 0.537 | 0.367 | 0.509 | 0.557 | 0.572 | 0.422 | 0.363 | 0.316 | 0.415 | 0.432 | 0.283 | 0.440 |
| SwitchOut | 0.411 | 0.446 | 0.318 | 0.415 | 0.467 | 0.466 | 0.38 | 0.335 | 0.278 | 0.337 | 0.347 | 0.262 | 0.372 |
| OBPE | 0.502 | 0.525 | 0.371 | 0.502 | 0.561 | 0.583 | 0.436 | 0.381 | 0.306 | 0.404 | 0.416 | 0.266 | 0.438 |
| BPE-Dropout | 0.501 | 0.526 | 0.371 | 0.497 | 0.558 | 0.574 | 0.439 | 0.393 | 0.300 | 0.389 | 0.410 | 0.231 | 0.432 |
| Random Char Noise | 0.521 | 0.547 | 0.371 | 0.501 | 0.569 | 0.584 | 0.441 | 0.380 | 0.391 | 0.487 | 0.491 | 0.319 | 0.467 |
| SELECTNOISE + Greedy | 0.525 | 0.563 | 0.386 | 0.511 | 0.578 | 0.606 | 0.458 | 0.394 | 0.392 | 0.499 | 0.511 | 0.319 | 0.478 |
| SELECTNOISE + Top-k | 0.524 | 0.558 | 0.386 | 0.507 | 0.576 | 0.599 | 0.454 | 0.388 | 0.400 | 0.497 | 0.516 | 0.321 | 0.477 |
| SELECTNOISE + Top-p | 0.527 | 0.599 | 0.372 | 0.505 | 0.573 | 0.599 | 0.457 | 0.391 | 0.399 | 0.501 | 0.509 | 0.321 | 0.479 |
| Selective noise + Greedy | 0.527 | 0.560 | 0.389 | 0.507 | 0.572 | 0.600 | 0.451 | 0.381 | 0.392 | 0.499 | 0.511 | 0.319 | 0.476 |
| Selective noise + Top-k | 0.526 | 0.549 | 0.401 | 0.509 | 0.573 | 0.463 | 0.463 | 0.390 | 0.400 | 0.494 | 0.506 | 0.326 | 0.467 |
| Selective noise + Top-p | 0.524 | 0.558 | 0.400 | 0.510 | 0.584 | 0.455 | 0.455 | 0.386 | 0.391 | 0.501 | 0.512 | 0.321 | 0.466 |
| Models | Indo-Aryan | Romance | Average | ||||||||||
| bho | hne | san | mai | mag | awa | npi | kas | cat | glg | ast | oci | ||
| Vanilla NMT | 0.642 | 0.676 | 0.471 | 0.621 | 0.711 | 0.736 | 0.542 | 0.387 | 0.499 | 0.534 | 0.497 | 0.408 | 0.560 |
| Word-drop | 0.659 | 0.702 | 0.494 | 0.650 | 0.725 | 0.747 | 0.564 | 0.409 | 0.484 | 0.551 | 0.538 | 0.421 | 0.579 |
| BPE-drop | 0.653 | 0.687 | 0.497 | 0.645 | 0.711 | 0.732 | 0.554 | 0.400 | 0.438 | 0.515 | 0.505 | 0.389 | 0.560 |
| SwitchOut | 0.565 | 0.605 | 0.462 | 0.564 | 0.626 | 0.632 | 0.533 | 0.394 | 0.405 | 0.461 | 0.445 | 0.362 | 0.504 |
| OBPE | 0.664 | 0.676 | 0.452 | 0.630 | 0.707 | 0.740 | 0.544 | 0.392 | 0.456 | 0.524 | 0.501 | 0.400 | 0.557 |
| BPE-Dropout | 0.644 | 0.672 | 0.471 | 0.616 | 0.710 | 0.733 | 0.537 | 0.381 | 0.503 | 0.534 | 0.500 | 0.411 | 0.559 |
| Random Char Noise | 0.673 | 0.700 | 0.492 | 0.641 | 0.725 | 0.746 | 0.559 | 0.401 | 0.522 | 0.610 | 0.584 | 0.441 | 0.591 |
| SELECTNOISE + Greedy | 0.672 | 0.714 | 0.493 | 0.647 | 0.735 | 0.765 | 0.575 | 0.412 | 0.523 | 0.620 | 0.598 | 0.434 | 0.599 |
| SELECTNOISE + Top-k | 0.678 | 0.708 | 0.504 | 0.649 | 0.730 | 0.758 | 0.585 | 0.419 | 0.524 | 0.621 | 0.603 | 0.438 | 0.601 |
| SELECTNOISE + Top-p | 0.677 | 0.559 | 0.502 | 0.643 | 0.730 | 0.758 | 0.586 | 0.411 | 0.526 | 0.625 | 0.600 | 0.442 | 0.588 |
| Selective noise + Greedy | 0.681 | 0.711 | 0.505 | 0.649 | 0.728 | 0.761 | 0.582 | 0.411 | 0.522 | 0.618 | 0.603 | 0.441 | 0.601 |
| Selective noise + Top-k | 0.677 | 0.700 | 0.506 | 0.655 | 0.703 | 0.757 | 0.581 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.598 |
| Selective noise + Top-p | 0.680 | 0.708 | 0.511 | 0.655 | 0.738 | 0.756 | 0.589 | 0.414 | 0.522 | 0.623 | 0.605 | 0.439 | 0.603 |
| Models | Languages | ||
| bho | san | npi | |
| Annotator set-1 | |||
| Vanilla NMT | 3.54 | 2.42 | 2.21 |
| BPE Dropout | 3.29 | 2.37 | 1.83 |
| SELECTNOISE Model | 4.17 | 2.83 | 2.50 |
| Annotator set-2 | |||
| Vanilla NMT | 3.42 | 1.96 | 2.17 |
| BPE Dropout | 2.79 | 1.83 | 1.96 |
| SELECTNOISE Model | 3.54 | 2.17 | 2.21 |
| Language | Data size | BLEU | chrF |
| hne | 997 | 19.5 | 49.6 |
| 6000 | 20.3 | 50.3 | |
| mai | 997 | 11.9 | 42.4 |
| 6000 | 12.4 | 43.2 | |
| npi | 997 | 6.7 | 33.6 |
| 6000 | 7.2 | 33.8 |
| Language | Model | BLEU | chrF |
| Bodo | Vanilla NMT | 2.4 | 18.2 |
| SELECTNOISE | 2.7 | 18.7 | |
| Tamil | Vanilla NMT | 0.6 | 11.7 |
| SELECTNOISE | 0.9 | 13.3 |
| Models | Languages | |||
| BLEU | chrF | BLEURT | COMET | |
| Hindi (hi) | ||||
| Vanilla NMT | 33.4 | 60.2 | 0.724 | 0.868 |
| Random Char Noise | 33.0 | 59.5 | 0.722 | 0.865 |
| SELECTNOISE | 34.2 | 60.5 | 0.726 | 0.869 |
| Spanish (es) | ||||
| Vanilla NMT | 21.5 | 53.5 | 0.695 | 0.810 |
| Random Char Noise | 21.7 | 53.0 | 0.689 | 0.806 |
| SELECTNOISE | 21.3 | 53.1 | 0.689 | 0.869 |
| Candidate Pool (Sc) Template |
| C1: {I': f1, 'D': f3, 'S': {E1: f4, E2: f5, ...}}, |
| C2: {I': f4, 'D': f3, 'S': {E3: f6, E2: f7, ...}}, |
| Ci: {I': f2, 'D': f4, 'S': {E1: f2, E3: f1, ...}} |
| Cn: {I': f1, 'D': f4, 'S': {E1: f4, E3: f5, ...}} |
| Few Sample Elements of Candidate Pool (Sc) |
| 'O': {I': 62, 'D': 1561, 'S': {'‘ह’: 1482, ...}} |
| '瓦': {I': 92, 'D': 97, 'S': {'‘ो’': 1482, ...}} |
| 'O': {I': 1552, 'D': 15, 'S': {'‘य’: 397, ...}} |
| 'S': {I': 0, 'D': 33, 'S': {} |
| } |
| Languages | #Dev | #Test | HRL | Source |
| bho, mag, mai, npi, awa, san, kas, hne | 997 | 1012 | hi | FLORES-200 |
| ast, cat, glg, oci | 997 | 1012 | es | FLORES-200 |
| HR Lang Pair | #Train | # Dev | #Test | Source |
| hi-en | 10.1M | 997 | 1012 | AI4Bharat |
| es-en | 6.6M | 997 | 1012 | Rapp (2021) |
| BHO: रूप्व केहोँ पाथस्मा,तुका केरागपिकोयपिक रेशेयपि रागपिक रहोँ प्रधान सेलन्दीपय रागपिक रूप्व के रागपिक रूप्व के रागपिक रूप्व के रागपिक रू�विक रूप्व के रागपिक रू�विक रू�विक रू�विक रू�विक रू��िक रूप्व के रागपिक रू�विक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक रू��िक वस्थिके \nENG: The announcement was made after Trump had a phone conversation with Turkish President Recep Tayyip Erdoğan. |
| RCN: Trump opened up about the phone call with Turkish President Recep Tayyip Erdoğan. |
| UCN: After a phone call with Turkish President Recep Tayyip Erdoğan, Trump announced his decision. |
| SCN: Trump's announcement came after a phone call with Turkish President Recep Tayyip Erdoğan. |
| HNE: पूललमह बलाटयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके रागल कर्वील्लयके \nENG: Police said that the body appeared to have been there for about a day. |
| RCN: The police said the body had been lying on the ground for the past few days. |
| UCN: The police said the bodies had been lying on the road for the past few days. |
| SCN: The police said the body had been lying for the past several days. |
| GLG: A ONU samén ten previsto create un fondo para axudar aes paises afectados polo quentamento global a afrontar o seu impacto. |
| ENG: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| RCN: The UN is also planning to set up a fund to tax years of countries affected by global pole to cope with or have an impact. |
| UCN: The UN also intends to set up a fund to relieve the countries affected by the global fear of facing or having a serious impact. |
| SCN: The UN also plans to set up a fund to relieve the countries affected by the global warming to face up to their impact. |
| AST: Munchos cudadianos de Bishkek acusaron a los manifestantes sureños del desorde. |
| ENG: Several Bishkek residents blamed protesters from the south for the lawlessness. |
| RCN: Cubans of Bishkek accused southern protestors of the deorde. |
| UCN: Many people in Bishkek accused the Swiss demonstrators of disorder. |
| SCN: Many people in Bishkek accused southern demonstrators of disorder. |
| Algorithm 1: The key views selection process of our KVZEL. | |
| Input: The mention with context m, candidate entity ep, candidate view list V of ep, number of iterations Q. Output: The selected view list V. | |
| 1 | Construct initial clues by Eq. 1. |
| 2 | for each iteration q = 1 to Q do |
| 3 | Calculate LSelection by Eq. 2 and train our key views selection module. |
| 4 | Calculate the most key view Vk by Eq. 3 in the inference stage. |
| 5 | Add Vk to existing clues and the selected view list V. |
| 6 | Remove Vk from the candidate view list V. |
| 7 | end |
| 8 | return V |
| Domains | Length | Entities | Mentions |
| Training | |||
| American Football | 665.06 | 31929 | 3898 |
| Doctor Who | 264.14 | 40281 | 8334 |
| Fallout | 229.68 | 16992 | 3286 |
| Final Fantasy | 497.35 | 14044 | 6041 |
| Military | 870.55 | 104520 | 13063 |
| Pro Wrestling | 639.53 | 10133 | 1392 |
| StarWars | 379.76 | 87056 | 11824 |
| World of Warcraft | 242.16 | 27677 | 1437 |
| Validation | |||
| Coronation Street | 264.55 | 17809 | 1464 |
| Muppets | 161.50 | 21344 | 2028 |
| Ice Hockey | 282.62 | 28684 | 2233 |
| Elder Scrolls | 269.03 | 21712 | 4275 |
| Test | |||
| Forgotten Realms | 257.29 | 15603 | 1200 |
| Lego | 223.86 | 10076 | 1199 |
| Star Trek | 393.21 | 34430 | 4227 |
| YuGiOh | 643.68 | 10031 | 3374 |
| Methods | Forgotten Realms | Lego | Star Trek | YuGiOh | Macro Acc. | Micro Acc. |
| Cross-encoder | 85.60 | 76.90 | 75.80 | 67.22 | 76.38 | 74.21 |
| Repeat | - | - | - | - | 77.58 | - |
| Uni-MPR | 85.55 | 77.42 | 78.23 | 68.29 | 77.62 | 75.78 |
| Bi-MPR | 89.09 | 77.18 | 79.20 | 69.98 | 78.61 | 76.70 |
| KVZEL w/o iteration | 87.10 | 76.90 | 78.42 | 71.36 | 78.45 | 77.35 |
| KVZEL w/o V0 | 88.30 | 76.08 | 78.96 | 73.41 | 79.19 | 78.24 |
| KVZEL (Ours) | 88.80 | 77.93 | 80.65 | 73.50 | 80.22 | 79.30 |
| Length | Number | Proportion |
| (0,200) | 2752 | 40.40% |
| [200,500) | 2025 | 29.73% |
| [500,1000) | 954 | 14.00% |
| [1000,2000) | 554 | 8.13% |
| [2000,+∞) | 527 | 7.74% |
| Mention with context | Description input of baselines | Description input of KVZEL |
| Lucifer reuses the cat single-stud mould introduced for the Friends theme. It features Lucifer standing up with bent back, elevated head looking straight ahead and long tail lying flat curled... | The Cat is a Friends animal figure introduced in 2013. The black one, released in the first series of Friends Animals, became available even in December 2012... | The first selected view: The cat is a single-stud mould, which features it standing up with bent back, elevated head looking straight ahead and long tail lying flat curled around the left side... |
| Having just defeated Sherry, Akiza walks over to her. Sherry protests that Akiza ran away from the Academy. Akiza admits that she ran away, and acknowledges that Duel Academy Sanctuary may foster elite... She asks Akiza where she learned the trick of riding her Duel Runner with her eyes shut... | Sherry LeBlanc is a Psychic Duelist with an ability referred to as Hand Scan, in which she can view her opponent's current hand. Design. Appearance. Sherry has long, waist length, blond hair with bangs that curl away from her face in a horizontal direction and has emerald color eyes... | The first selected view: Akiza chose to leave the academy due to its eli-tist ways, which made Sherry feel betrayed and harbor a hatred for Akiza... The second selected view: After the Duel, Sherry and Akiza become friends again after Sherry questions Akiza on where she learned to drive a Duel Runner with her eyes closed... |
| Algorithm 1: The key views selection process of our KVZEL. | |
| Input: The mention with context m, candidate entity ep, candidate view list V of ep, number of iterations Q. Output: The selected view list V. | |
| 1 | Construct initial clues by Eq. 1. |
| 2 | for each iteration q = 1 to Q do |
| 3 | Calculate LSelection by Eq. 2 and train our key views selection module. |
| 4 | Calculate the most key view Vk by Eq. 3 in the inference stage. |
| 5 | Add Vk to existing clues and the selected view list V. |
| 6 | Remove Vk from the candidate view list V. |
| 7 | end |
| 8 | return V |
| Domains | Length | Entities | Mentions |
| Training | |||
| American Football | 665.06 | 31929 | 3898 |
| Doctor Who | 264.14 | 40281 | 8334 |
| Fallout | 229.68 | 16992 | 3286 |
| Final Fantasy | 497.35 | 14044 | 6041 |
| Military | 870.55 | 104520 | 13063 |
| Pro Wrestling | 639.53 | 10133 | 1392 |
| StarWars | 379.76 | 87056 | 11824 |
| World of Warcraft | 242.16 | 27677 | 1437 |
| Validation | |||
| Coronation Street | 264.55 | 17809 | 1464 |
| Muppets | 161.50 | 21344 | 2028 |
| Ice Hockey | 282.62 | 28684 | 2233 |
| Elder Scrolls | 269.03 | 21712 | 4275 |
| Test | |||
| Forgotten Realms | 257.29 | 15603 | 1200 |
| Lego | 223.86 | 10076 | 1199 |
| Star Trek | 393.21 | 34430 | 4227 |
| YuGiOh | 643.68 | 10031 | 3374 |
| Methods | Forgotten Realms | Lego | Star Trek | YuGiOh | Macro Acc. | Micro Acc. |
| Cross-encoder | 85.60 | 76.90 | 75.80 | 67.22 | 76.38 | 74.21 |
| Repeat | - | - | - | - | 77.58 | - |
| Uni-MPR | 85.55 | 77.42 | 78.23 | 68.29 | 77.62 | 75.78 |
| Bi-MPR | 89.09 | 77.18 | 79.20 | 69.98 | 78.61 | 76.70 |
| KVZEL w/o iteration | 87.10 | 76.90 | 78.42 | 71.36 | 78.45 | 77.35 |
| KVZEL w/o V0 | 88.30 | 76.08 | 78.96 | 73.41 | 79.19 | 78.24 |
| KVZEL (Ours) | 88.80 | 77.93 | 80.65 | 73.50 | 80.22 | 79.30 |
| Length | Number | Proportion |
| (0,200) | 2752 | 40.40% |
| [200,500) | 2025 | 29.73% |
| [500,1000) | 954 | 14.00% |
| [1000,2000) | 554 | 8.13% |
| [2000,+∞) | 527 | 7.74% |
| Mention with context | Description input of baselines | Description input of KVZEL |
| Lucifer reuses the cat single-stud mould introduced for the Friends theme. It features Lucifer standing up with bent back, elevated head looking straight ahead and long tail lying flat curled... | The Cat is a Friends animal figure introduced in 2013. The black one, released in the first series of Friends Animals, became available even in December 2012... | The first selected view: The cat is a single-stud mould, which features it standing up with bent back, elevated head looking straight ahead and long tail lying flat curled around the left side... |
| Having just defeated Sherry, Akiza walks over to her. Sherry protests that Akiza ran away from the Academy. Akiza admits that she ran away, and acknowledges that Duel Academy Sanctuary may foster elite... She asks Akiza where she learned the trick of riding her Duel Runner with her eyes shut... | Sherry LeBlanc is a Psychic Duelist with an ability referred to as Hand Scan, in which she can view her opponent's current hand. Design. Appearance. Sherry has long, waist length, blond hair with bangs that curl away from her face in a horizontal direction and has emerald color eyes... | The first selected view: Akiza chose to leave the academy due to its eli-tist ways, which made Sherry feel betrayed and harbor a hatred for Akiza... The second selected view: After the Duel, Sherry and Akiza become friends again after Sherry questions Akiza on where she learned to drive a Duel Runner with her eyes closed... |
| Algorithm 1: The key views selection process of our KVZEL. | |
| Input: The mention with context m, candidate entity ep, candidate view list V of ep, number of iterations Q. Output: The selected view list V. | |
| 1 | Construct initial clues by Eq. 1. |
| 2 | for each iteration q = 1 to Q do |
| 3 | Calculate LSelection by Eq. 2 and train our key views selection module. |
| 4 | Calculate the most key view Vk by Eq. 3 in the inference stage. |
| 5 | Add Vk to existing clues and the selected view list V. |
| 6 | Remove Vk from the candidate view list V. |
| 7 | end |
| 8 | return V |
| Domains | Length | Entities | Mentions |
| Training | |||
| American Football | 665.06 | 31929 | 3898 |
| Doctor Who | 264.14 | 40281 | 8334 |
| Fallout | 229.68 | 16992 | 3286 |
| Final Fantasy | 497.35 | 14044 | 6041 |
| Military | 870.55 | 104520 | 13063 |
| Pro Wrestling | 639.53 | 10133 | 1392 |
| StarWars | 379.76 | 87056 | 11824 |
| World of Warcraft | 242.16 | 27677 | 1437 |
| Validation | |||
| Coronation Street | 264.55 | 17809 | 1464 |
| Muppets | 161.50 | 21344 | 2028 |
| Ice Hockey | 282.62 | 28684 | 2233 |
| Elder Scrolls | 269.03 | 21712 | 4275 |
| Test | |||
| Forgotten Realms | 257.29 | 15603 | 1200 |
| Lego | 223.86 | 10076 | 1199 |
| Star Trek | 393.21 | 34430 | 4227 |
| YuGiOh | 643.68 | 10031 | 3374 |
| Methods | Forgotten Realms | Lego | Star Trek | YuGiOh | Macro Acc. | Micro Acc. |
| Cross-encoder | 85.60 | 76.90 | 75.80 | 67.22 | 76.38 | 74.21 |
| Repeat | - | - | - | - | 77.58 | - |
| Uni-MPR | 85.55 | 77.42 | 78.23 | 68.29 | 77.62 | 75.78 |
| Bi-MPR | 89.09 | 77.18 | 79.20 | 69.98 | 78.61 | 76.70 |
| KVZEL w/o iteration | 87.10 | 76.90 | 78.42 | 71.36 | 78.45 | 77.35 |
| KVZEL w/o V0 | 88.30 | 76.08 | 78.96 | 73.41 | 79.19 | 78.24 |
| KVZEL (Ours) | 88.80 | 77.93 | 80.65 | 73.50 | 80.22 | 79.30 |
| Length | Number | Proportion |
| (0,200) | 2752 | 40.40% |
| [200,500) | 2025 | 29.73% |
| [500,1000) | 954 | 14.00% |
| [1000,2000) | 554 | 8.13% |
| [2000,+∞) | 527 | 7.74% |
| Mention with context | Description input of baselines | Description input of KVZEL |
| Lucifer reuses the cat single-stud mould introduced for the Friends theme. It features Lucifer standing up with bent back, elevated head looking straight ahead and long tail lying flat curled... | The Cat is a Friends animal figure introduced in 2013. The black one, released in the first series of Friends Animals, became available even in December 2012... | The first selected view: The cat is a single-stud mould, which features it standing up with bent back, elevated head looking straight ahead and long tail lying flat curled around the left side... |
| Having just defeated Sherry, Akiza walks over to her. Sherry protests that Akiza ran away from the Academy. Akiza admits that she ran away, and acknowledges that Duel Academy Sanctuary may foster elite... She asks Akiza where she learned the trick of riding her Duel Runner with her eyes shut... | Sherry LeBlanc is a Psychic Duelist with an ability referred to as Hand Scan, in which she can view her opponent's current hand. Design. Appearance. Sherry has long, waist length, blond hair with bangs that curl away from her face in a horizontal direction and has emerald color eyes... | The first selected view: Akiza chose to leave the academy due to its eli-tist ways, which made Sherry feel betrayed and harbor a hatred for Akiza... The second selected view: After the Duel, Sherry and Akiza become friends again after Sherry questions Akiza on where she learned to drive a Duel Runner with her eyes closed... |
| Algorithm 1: The key views selection process of our KVZEL. | |
| Input: The mention with context m, candidate entity ep, candidate view list V of ep, number of iterations Q. Output: The selected view list V. | |
| 1 | Construct initial clues by Eq. 1. |
| 2 | for each iteration q = 1 to Q do |
| 3 | Calculate LSelection by Eq. 2 and train our key views selection module. |
| 4 | Calculate the most key view Vk by Eq. 3 in the inference stage. |
| 5 | Add Vk to existing clues and the selected view list V. |
| 6 | Remove Vk from the candidate view list V. |
| 7 | end |
| 8 | return V |
| Domains | Length | Entities | Mentions |
| Training | |||
| American Football | 665.06 | 31929 | 3898 |
| Doctor Who | 264.14 | 40281 | 8334 |
| Fallout | 229.68 | 16992 | 3286 |
| Final Fantasy | 497.35 | 14044 | 6041 |
| Military | 870.55 | 104520 | 13063 |
| Pro Wrestling | 639.53 | 10133 | 1392 |
| StarWars | 379.76 | 87056 | 11824 |
| World of Warcraft | 242.16 | 27677 | 1437 |
| Validation | |||
| Coronation Street | 264.55 | 17809 | 1464 |
| Muppets | 161.50 | 21344 | 2028 |
| Ice Hockey | 282.62 | 28684 | 2233 |
| Elder Scrolls | 269.03 | 21712 | 4275 |
| Test | |||
| Forgotten Realms | 257.29 | 15603 | 1200 |
| Lego | 223.86 | 10076 | 1199 |
| Star Trek | 393.21 | 34430 | 4227 |
| YuGiOh | 643.68 | 10031 | 3374 |
| Methods | Forgotten Realms | Lego | Star Trek | YuGiOh | Macro Acc. | Micro Acc. |
| Cross-encoder | 85.60 | 76.90 | 75.80 | 67.22 | 76.38 | 74.21 |
| Repeat | - | - | - | - | 77.58 | - |
| Uni-MPR | 85.55 | 77.42 | 78.23 | 68.29 | 77.62 | 75.78 |
| Bi-MPR | 89.09 | 77.18 | 79.20 | 69.98 | 78.61 | 76.70 |
| KVZEL w/o iteration | 87.10 | 76.90 | 78.42 | 71.36 | 78.45 | 77.35 |
| KVZEL w/o V0 | 88.30 | 76.08 | 78.96 | 73.41 | 79.19 | 78.24 |
| KVZEL (Ours) | 88.80 | 77.93 | 80.65 | 73.50 | 80.22 | 79.30 |
| Length | Number | Proportion |
| (0,200) | 2752 | 40.40% |
| [200,500) | 2025 | 29.73% |
| [500,1000) | 954 | 14.00% |
| [1000,2000) | 554 | 8.13% |
| [2000,+∞) | 527 | 7.74% |
| Mention with context | Description input of baselines | Description input of KVZEL |
| Lucifer reuses the cat single-stud mould introduced for the Friends theme. It features Lucifer standing up with bent back, elevated head looking straight ahead and long tail lying flat curled... | The Cat is a Friends animal figure introduced in 2013. The black one, released in the first series of Friends Animals, became available even in December 2012... | The first selected view: The cat is a single-stud mould, which features it standing up with bent back, elevated head looking straight ahead and long tail lying flat curled around the left side... |
| Having just defeated Sherry, Akiza walks over to her. Sherry protests that Akiza ran away from the Academy. Akiza admits that she ran away, and acknowledges that Duel Academy Sanctuary may foster elite... She asks Akiza where she learned the trick of riding her Duel Runner with her eyes shut... | Sherry LeBlanc is a Psychic Duelist with an ability referred to as Hand Scan, in which she can view her opponent's current hand. Design. Appearance. Sherry has long, waist length, blond hair with bangs that curl away from her face in a horizontal direction and has emerald color eyes... | The first selected view: Akiza chose to leave the academy due to its eli-tist ways, which made Sherry feel betrayed and harbor a hatred for Akiza... The second selected view: After the Duel, Sherry and Akiza become friends again after Sherry questions Akiza on where she learned to drive a Duel Runner with her eyes closed... |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 78.0 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 79.1 |
| ShiP+Picard (Zhao et al., 2022) | T5-3B | 81.4 |
| RESDSQL+NatSQL (Li et al., 2023a) | T5-3B | 84.1 |
| Previous Work with In-context Learning | ||
| Rajkumar et al. (2022) | Codex | 67.0 |
| Chang and Fosler-Lussier (2023) | Codex | 76.8 |
| LEVER (Ni et al., 2023) | Codex | 81.9 |
| DIN-SQL (Pourreza and Rafiei, 2023) | Codex | 75.6 |
| DIN-SQL (Pourreza and Rafiei, 2023) | GPT-4 | 82.8 |
| Self-Debugging (Chen et al., 2023) | Codex | 84.1 |
| This work | ||
| Codex | 75.7 | |
| Zero-shot | ChatGPT | 75.7 |
| CodeLlama | 70.7 | |
| Codex | 82.1 | |
| Out-of-domain Demo Only | ChatGPT | 80.7 |
| CodeLlama | 75.6 | |
| Codex | 81.5 | |
| Synthetic In-domain Demo Only | ChatGPT | 78.5 |
| CodeLlama | 78.3 | |
| Codex | 85.2 | |
| ODIS | ChatGPT | 81.5 |
| CodeLlama | 80.0 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 27.2 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 29.8 |
| Previous Work with In-context Learning | ||
| Zero-shot (Lan et al., 2023) | Codex | 23.9 |
| Few-shot (Lan et al., 2023) | Codex | 40.4 |
| SKILL-KNN (An et al., 2023) | Codex | 43.0 |
| This work | ||
| Codex | 26.8 | |
| Zero-shot | ChatGPT | 25.7 |
| CodeLlama | 18.8 | |
| Codex | 53.3 | |
| Out-of-domain Demo Only | ChatGPT | 45.6 |
| CodeLlama | 37.9 | |
| Codex | 45.2 | |
| Synthetic In-domain Demo Only | ChatGPT | 33.1 |
| CodeLlama | 34.9 | |
| Codex | 54.8 | |
| ODIS | ChatGPT | 52.9 |
| CodeLlama | 42.3 |
| Method | Spider | Kaggle |
| Random | 78.7 | 38.0 |
| SimNLQ | 81.2 | 42.6 |
| SimSQL | 82.1 | 53.3 |
| SimSQL (Oracle) | 82.9 | 56.2 |
| Method | Spider | Kaggle |
| Random | 77.0 | 38.2 |
| SimNLQ | 79.2 | 41.5 |
| SimSQL | 79.7 | 45.2 |
| CovSQL | 81.5 | 45.2 |
| CovSQL (Oracle) | 81.8 | 48.9 |
| CovSQL (In-domain Annotation) | 82.4 | 52.9 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 58.1 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 65.0 |
| This work | ||
| Zero-shot | Codex | 65.3 |
| Out-of-domain Demo Only | Codex | 70.9 |
| Synthetic In-domain Demo Only | Codex | 70.9 |
| ODIS | Codex | 73.4 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 78.0 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 79.1 |
| ShiP+Picard (Zhao et al., 2022) | T5-3B | 81.4 |
| RESDSQL+NatSQL (Li et al., 2023a) | T5-3B | 84.1 |
| Previous Work with In-context Learning | ||
| Rajkumar et al. (2022) | Codex | 67.0 |
| Chang and Fosler-Lussier (2023) | Codex | 76.8 |
| LEVER (Ni et al., 2023) | Codex | 81.9 |
| DIN-SQL (Pourreza and Rafiei, 2023) | Codex | 75.6 |
| DIN-SQL (Pourreza and Rafiei, 2023) | GPT-4 | 82.8 |
| Self-Debugging (Chen et al., 2023) | Codex | 84.1 |
| This work | ||
| Codex | 75.7 | |
| Zero-shot | ChatGPT | 75.7 |
| CodeLlama | 70.7 | |
| Codex | 82.1 | |
| Out-of-domain Demo Only | ChatGPT | 80.7 |
| CodeLlama | 75.6 | |
| Codex | 81.5 | |
| Synthetic In-domain Demo Only | ChatGPT | 78.5 |
| CodeLlama | 78.3 | |
| Codex | 85.2 | |
| ODIS | ChatGPT | 81.5 |
| CodeLlama | 80.0 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 27.2 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 29.8 |
| Previous Work with In-context Learning | ||
| Zero-shot (Lan et al., 2023) | Codex | 23.9 |
| Few-shot (Lan et al., 2023) | Codex | 40.4 |
| SKILL-KNN (An et al., 2023) | Codex | 43.0 |
| This work | ||
| Codex | 26.8 | |
| Zero-shot | ChatGPT | 25.7 |
| CodeLlama | 18.8 | |
| Codex | 53.3 | |
| Out-of-domain Demo Only | ChatGPT | 45.6 |
| CodeLlama | 37.9 | |
| Codex | 45.2 | |
| Synthetic In-domain Demo Only | ChatGPT | 33.1 |
| CodeLlama | 34.9 | |
| Codex | 54.8 | |
| ODIS | ChatGPT | 52.9 |
| CodeLlama | 42.3 |
| Method | Spider | Kaggle |
| Random | 78.7 | 38.0 |
| SimNLQ | 81.2 | 42.6 |
| SimSQL | 82.1 | 53.3 |
| SimSQL (Oracle) | 82.9 | 56.2 |
| Method | Spider | Kaggle |
| Random | 77.0 | 38.2 |
| SimNLQ | 79.2 | 41.5 |
| SimSQL | 79.7 | 45.2 |
| CovSQL | 81.5 | 45.2 |
| CovSQL (Oracle) | 81.8 | 48.9 |
| CovSQL (In-domain Annotation) | 82.4 | 52.9 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 58.1 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 65.0 |
| This work | ||
| Zero-shot | Codex | 65.3 |
| Out-of-domain Demo Only | Codex | 70.9 |
| Synthetic In-domain Demo Only | Codex | 70.9 |
| ODIS | Codex | 73.4 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 78.0 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 79.1 |
| ShiP+Picard (Zhao et al., 2022) | T5-3B | 81.4 |
| RESDSQL+NatSQL (Li et al., 2023a) | T5-3B | 84.1 |
| Previous Work with In-context Learning | ||
| Rajkumar et al. (2022) | Codex | 67.0 |
| Chang and Fosler-Lussier (2023) | Codex | 76.8 |
| LEVER (Ni et al., 2023) | Codex | 81.9 |
| DIN-SQL (Pourreza and Rafiei, 2023) | Codex | 75.6 |
| DIN-SQL (Pourreza and Rafiei, 2023) | GPT-4 | 82.8 |
| Self-Debugging (Chen et al., 2023) | Codex | 84.1 |
| This work | ||
| Codex | 75.7 | |
| Zero-shot | ChatGPT | 75.7 |
| CodeLlama | 70.7 | |
| Codex | 82.1 | |
| Out-of-domain Demo Only | ChatGPT | 80.7 |
| CodeLlama | 75.6 | |
| Codex | 81.5 | |
| Synthetic In-domain Demo Only | ChatGPT | 78.5 |
| CodeLlama | 78.3 | |
| Codex | 85.2 | |
| ODIS | ChatGPT | 81.5 |
| CodeLlama | 80.0 | |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 27.2 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 29.8 |
| Previous Work with In-context Learning | ||
| Zero-shot (Lan et al., 2023) | Codex | 23.9 |
| Few-shot (Lan et al., 2023) | Codex | 40.4 |
| SKILL-KNN (An et al., 2023) | Codex | 43.0 |
| This work | ||
| Codex | 26.8 | |
| Zero-shot | ChatGPT | 25.7 |
| CodeLlama | 18.8 | |
| Codex | 53.3 | |
| Out-of-domain Demo Only | ChatGPT | 45.6 |
| CodeLlama | 37.9 | |
| Codex | 45.2 | |
| Synthetic In-domain Demo Only | ChatGPT | 33.1 |
| CodeLlama | 34.9 | |
| Codex | 54.8 | |
| ODIS | ChatGPT | 52.9 |
| CodeLlama | 42.3 | |
| Method | Spider | Kaggle |
| Random | 78.7 | 38.0 |
| SimNLQ | 81.2 | 42.6 |
| SimSQL | 82.1 | 53.3 |
| SimSQL (Oracle) | 82.9 | 56.2 |
| Method | Spider | Kaggle |
| Random | 77.0 | 38.2 |
| SimNLQ | 79.2 | 41.5 |
| SimSQL | 79.7 | 45.2 |
| CovSQL | 81.5 | 45.2 |
| CovSQL (Oracle) | 81.8 | 48.9 |
| CovSQL (In-domain Annotation) | 82.4 | 52.9 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 58.1 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 65.0 |
| This work | ||
| Zero-shot | Codex | 65.3 |
| Out-of-domain Demo Only | Codex | 70.9 |
| Synthetic In-domain Demo Only | Codex | 70.9 |
| ODIS | Codex | 73.4 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 78.0 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 79.1 |
| ShiP+Picard (Zhao et al., 2022) | T5-3B | 81.4 |
| RESDSQL+NatSQL (Li et al., 2023a) | T5-3B | 84.1 |
| Previous Work with In-context Learning | ||
| Rajkumar et al. (2022) | Codex | 67.0 |
| Chang and Fosler-Lussier (2023) | Codex | 76.8 |
| LEVER (Ni et al., 2023) | Codex | 81.9 |
| DIN-SQL (Pourreza and Rafiei, 2023) | Codex | 75.6 |
| DIN-SQL (Pourreza and Rafiei, 2023) | GPT-4 | 82.8 |
| Self-Debugging (Chen et al., 2023) | Codex | 84.1 |
| This work | ||
| Codex | 75.7 | |
| Zero-shot | ChatGPT | 75.7 |
| CodeLlama | 70.7 | |
| Codex | 82.1 | |
| Out-of-domain Demo Only | ChatGPT | 80.7 |
| CodeLlama | 75.6 | |
| Codex | 81.5 | |
| Synthetic In-domain Demo Only | ChatGPT | 78.5 |
| CodeLlama | 78.3 | |
| Codex | 85.2 | |
| ODIS | ChatGPT | 81.5 |
| CodeLlama | 80.0 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 27.2 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 29.8 |
| Previous Work with In-context Learning | ||
| Zero-shot (Lan et al., 2023) | Codex | 23.9 |
| Few-shot (Lan et al., 2023) | Codex | 40.4 |
| SKILL-KNN (An et al., 2023) | Codex | 43.0 |
| This work | ||
| Codex | 26.8 | |
| Zero-shot | ChatGPT | 25.7 |
| CodeLlama | 18.8 | |
| Codex | 53.3 | |
| Out-of-domain Demo Only | ChatGPT | 45.6 |
| CodeLlama | 37.9 | |
| Codex | 45.2 | |
| Synthetic In-domain Demo Only | ChatGPT | 33.1 |
| CodeLlama | 34.9 | |
| Codex | 54.8 | |
| ODIS | ChatGPT | 52.9 |
| CodeLlama | 42.3 |
| Method | Spider | Kaggle |
| Random | 78.7 | 38.0 |
| SimNLQ | 81.2 | 42.6 |
| SimSQL | 82.1 | 53.3 |
| SimSQL (Oracle) | 82.9 | 56.2 |
| Method | Spider | Kaggle |
| Random | 77.0 | 38.2 |
| SimNLQ | 79.2 | 41.5 |
| SimSQL | 79.7 | 45.2 |
| CovSQL | 81.5 | 45.2 |
| CovSQL (Oracle) | 81.8 | 48.9 |
| CovSQL (In-domain Annotation) | 82.4 | 52.9 |
| Method | LLM | EX |
| Previous Work with Finetuning | ||
| SmBoP (Rubin and Berant, 2021) | RB | 58.1 |
| T5+Picard (Scholak et al., 2021) | T5-3B | 65.0 |
| This work | ||
| Zero-shot | Codex | 65.3 |
| Out-of-domain Demo Only | Codex | 70.9 |
| Synthetic In-domain Demo Only | Codex | 70.9 |
| ODIS | Codex | 73.4 |
| Method | Temporal (EM/F1) | Commonsense (Acc.) | Tabular (Acc.) | Strategy (Acc.) | Truthful (Acc.) | Avg. | |
| (text-davinci-003) | |||||||
| (w/o retrieval) | Zero-Shot | 40.57/44.94 | 65.52 | 66.08 | 63.32 | 61.06 | 56.91 |
| Zero-Shot-CoT | 41.71/45.33 | 63.31 | 60.50 | 58.52 | 53.10 | 53.75 | |
| Few-Shot | 45.14/49.59 | 80.34 | 63.50 | 66.37 | 65.49 | 61.73 | |
| Manual-CoT | 44.57/54.22 | 75.42 | 74.92 | 71.18 | 72.57 | 65.48 | |
| Auto-CoT (Similarity) | 44.00/54.13 | 74.44 | 77.25 | 70.61 | 72.57 | 65.50 | |
| Auto-CoT (Diversity) | 46.28/54.68 | 73.38 | 74.50 | 70.74 | 71.68 | 65.21 | |
| (w/ retrieval) | Manual-CoT-IR | 47.42/57.37 | 75.67 | 79.25 | 69.43 | 69.03 | 66.36 |
| IRCoT | 47.42/56.28 | 75.27 | 78.00 | 67.68 | 71.68 | 66.06 | |
| \\( SKR_{prompt} \\) | 45.71/56.31 | 75.02 | 77.33 | 69.43 | 70.80 | 65.77 | |
| \\( SKR_{icl} \\) | 47.42/57.74 | 75.51 | 77.75 | 71.18 | 73.45 | 67.17 | |
| \\( SKR_{cls} \\) | 46.28/56.54 | 75.83 | 79.25 | 70.30 | 72.57 | 66.80 | |
| \\( SKR_{knn} \\) | 48.00/58.47 | 76.66 | 79.83 | 71.62 | 74.34 | 68.15 | |
| \\( CoT-RR^† \\) | 44.97/56.58 | 76.98 | 82.08 | 74.67 | 69.91 | 67.53 | |
| (gpt-3.5-turbo-0301) | |||||||
| (w/o retrieval) | Zero-Shot | 51.42/54.94 | 73.30 | 66.50 | 54.14 | 74.33 | 62.43 |
| Zero-Shot-CoT | 56.57/58.92 | 61.58 | 63.08 | 44.10 | 65.49 | 58.29 | |
| Few-Shot | 52.57/55.77 | 78.86 | 68.75 | 61.13 | 69.91 | 64.50 | |
| Manual-CoT | 54.85/61.72 | 74.77 | 73.25 | 61.36 | 81.41 | 67.89 | |
| Auto-CoT (Similarity) | 54.28/57.66 | 74.44 | 71.58 | 61.57 | 79.64 | 66.52 | |
| Auto-CoT (Diversity) | 55.43/58.21 | 73.30 | 71.50 | 58.52 | 80.54 | 66.25 | |
| (w/ retrieval) | Manual-CoT-IR | 59.41/63.08 | 73.38 | 76.58 | 57.21 | 76.99 | 67.77 |
| IRCoT | 57.14/61.67 | 72.73 | 76.25 | 55.46 | 79.64 | 67.15 | |
| \\( SKR_{prompt} \\) | 54.86/61.64 | 75.02 | 73.91 | 61.57 | 80.53 | 67.92 | |
| \\( SKR_{icl} \\) | 58.29/63.81 | 75.10 | 74.42 | 62.01 | 82.30 | 69.32 | |
| \\( SKR_{cls} \\) | 59.43/64.04 | 75.10 | 76.42 | 62.01 | 84.95 | 70.33 | |
| \\( SKR_{knn} \\) | 61.14/66.13 | 75.43 | 76.75 | 62.01 | 82.30 | 70.62 | |
| \\( CoT-RR^† \\) | 60.35/65.59 | 75.84 | 76.08 | 62.88 | 83.18 | 70.65 |
| Template | Positive Response (LLM known) | Negative Response (LLM unknown) | Acc. |
| Do you need additional information to answer this question? | No, additional information is not needed to answer this... | Yes, additional information is needed to answer this... | 73.17 |
| Would you like any extra prompts to help you? | No, I do not need any extra... | Yes, please. | 72.32 |
| Would you like any additional clues? | No, the answer is... | Yes, please provide... | 72.32 |
| Can you answer this question based on what you know? | Yes, the correct answer to this question is... | No, I cannot answer it based on what I know. | 72.07 |
| Can you solve this question now? | Yes, the correct answer is... | No, this is not a solvable... | 71.58 |
| Manual-CoT | Manual-CoT-IR |
| A1: Cleopatra and Caesar are a 17th-century Italian paint-ing by Francesco Barberini. So the answer is false. (X) | A1: Cleopatra and Caesar is an oil on canvas painting by the French Academic artist Jean-Léon Gérôme, completed in 1866. So the answer is true. (✓) |
| A2: Boulevard des Capucines is a street in Paris, France. [...] Thus, Boulevard des Capucines is not a 19th-century European work of art. So the answer is false. (X) | A2: Boulevard des Capucines is an oil painting [...] by French Impressionist artist Claude Monet created in 1873. So the answer is true. (✓) |
| A3: Bulb Fields is a 1990 painting by Mark Rothko. So the answer is false. (X) | A3: Bulb Fields is an oil painting created by Vincent van Gogh in early 1883. So the answer is true. (✓) |
| A: An osteria is an Italian type of eating establishment. It is not a 19th-century artwork. So the answer is false. (X) | A: In a Roman Osteria is a painting by the Danish painter Carl Bloch. It was painted in 1866. So the answer is true. (✓) |
| Method | Temporal | Commonsense | Tabular | Strategy | Truthful |
| Self-Ask | 58.91 | 74.45 | 72.91 | 59.38 | 78.77 |
| Recite-and-Answer | 62.45 | 74.61 | 75.00 | 58.95 | 79.64 |
| DSP | 63.95 | 73.21 | 75.75 | 58.73 | 77.88 |
| \\( \\text{SKR}_{\\text{knn}} \\) | 66.13 | 75.43 | 76.75 | 62.01 | 82.30 |
| Retriever | Retrieval Quality | Manual-CoT-IR | Our \\( \\mathbf{S}\\mathbf{K}\\mathbf{R}_{\\mathbf{k}\\mathbf{n}\\mathbf{n}} \\) |
| DPR | high | 76.99 | 82.30 |
| SimCSE | semantically matched | 73.45 (↓ 3.54) | 81.42 (↓ 0.88) |
| BERT | relatively poor | 69.91 (↓ 7.08) | 80.53 (↓ 1.67) |
| Corpora | Selected Candidates | Manual-CoT-IR | Our SKRknn |
| Full C | rank 1~3 | 76.99 | 82.30 |
| Full C | rank 8~10 | 76.11 (↓0.88) | 82.30 (-) |
| Half of C | rank 1~3 | 74.37 (↓2.62) | 81.41 (↓0.89) |
| Method | Temporal (EM/F1) | Commonsense (Acc.) | Tabular (Acc.) | Strategy (Acc.) | Truthful (Acc.) | Avg. | |
| (text-davinci-003) | |||||||
| (w/o retrieval) | Zero-Shot | 40.57/44.94 | 65.52 | 66.08 | 63.32 | 61.06 | 56.91 |
| Zero-Shot-CoT | 41.71/45.33 | 63.31 | 60.50 | 58.52 | 53.10 | 53.75 | |
| Few-Shot | 45.14/49.59 | 80.34 | 63.50 | 66.37 | 65.49 | 61.73 | |
| Manual-CoT | 44.57/54.22 | 75.42 | 74.92 | 71.18 | 72.57 | 65.48 | |
| Auto-CoT (Similarity) | 44.00/54.13 | 74.44 | 77.25 | 70.61 | 72.57 | 65.50 | |
| Auto-CoT (Diversity) | 46.28/54.68 | 73.38 | 74.50 | 70.74 | 71.68 | 65.21 | |
| (w/ retrieval) | Manual-CoT-IR | 47.42/57.37 | 75.67 | 79.25 | 69.43 | 69.03 | 66.36 |
| IRCoT | 47.42/56.28 | 75.27 | 78.00 | 67.68 | 71.68 | 66.06 | |
| \\( SKR_{prompt} \\) | 45.71/56.31 | 75.02 | 77.33 | 69.43 | 70.80 | 65.77 | |
| \\( SKR_{icl} \\) | 47.42/57.74 | 75.51 | 77.75 | 71.18 | 73.45 | 67.17 | |
| \\( SKR_{cls} \\) | 46.28/56.54 | 75.83 | 79.25 | 70.30 | 72.57 | 66.80 | |
| \\( SKR_{knn} \\) | 48.00/58.47 | 76.66 | 79.83 | 71.62 | 74.34 | 68.15 | |
| \\( CoT-RR^† \\) | 44.97/56.58 | 76.98 | 82.08 | 74.67 | 69.91 | 67.53 | |
| (gpt-3.5-turbo-0301) | |||||||
| (w/o retrieval) | Zero-Shot | 51.42/54.94 | 73.30 | 66.50 | 54.14 | 74.33 | 62.43 |
| Zero-Shot-CoT | 56.57/58.92 | 61.58 | 63.08 | 44.10 | 65.49 | 58.29 | |
| Few-Shot | 52.57/55.77 | 78.86 | 68.75 | 61.13 | 69.91 | 64.50 | |
| Manual-CoT | 54.85/61.72 | 74.77 | 73.25 | 61.36 | 81.41 | 67.89 | |
| Auto-CoT (Similarity) | 54.28/57.66 | 74.44 | 71.58 | 61.57 | 79.64 | 66.52 | |
| Auto-CoT (Diversity) | 55.43/58.21 | 73.30 | 71.50 | 58.52 | 80.54 | 66.25 | |
| (w/ retrieval) | Manual-CoT-IR | 59.41/63.08 | 73.38 | 76.58 | 57.21 | 76.99 | 67.77 |
| IRCoT | 57.14/61.67 | 72.73 | 76.25 | 55.46 | 79.64 | 67.15 | |
| \\( SKR_{prompt} \\) | 54.86/61.64 | 75.02 | 73.91 | 61.57 | 80.53 | 67.92 | |
| \\( SKR_{icl} \\) | 58.29/63.81 | 75.10 | 74.42 | 62.01 | 82.30 | 69.32 | |
| \\( SKR_{cls} \\) | 59.43/64.04 | 75.10 | 76.42 | 62.01 | 84.95 | 70.33 | |
| \\( SKR_{knn} \\) | 61.14/66.13 | 75.43 | 76.75 | 62.01 | 82.30 | 70.62 | |
| \\( CoT-RR^† \\) | 60.35/65.59 | 75.84 | 76.08 | 62.88 | 83.18 | 70.65 |
| Template | Positive Response (LLM known) | Negative Response (LLM unknown) | Acc. |
| Do you need additional information to answer this question? | No, additional information is not needed to answer this... | Yes, additional information is needed to answer this... | 73.17 |
| Would you like any extra prompts to help you? | No, I do not need any extra... | Yes, please. | 72.32 |
| Would you like any additional clues? | No, the answer is... | Yes, please provide... | 72.32 |
| Can you answer this question based on what you know? | Yes, the correct answer to this question is... | No, I cannot answer it based on what I know. | 72.07 |
| Can you solve this question now? | Yes, the correct answer is... | No, this is not a solvable... | 71.58 |
| Manual-CoT | Manual-CoT-IR |
| A1: Cleopatra and Caesar are a 17th-century Italian paint-ing by Francesco Barberini. So the answer is false. (X) | A1: Cleopatra and Caesar is an oil on canvas painting by the French Academic artist Jean-Léon Gérôme, completed in 1866. So the answer is true. (✓) |
| A2: Boulevard des Capucines is a street in Paris, France. [...] Thus, Boulevard des Capucines is not a 19th-century European work of art. So the answer is false. (X) | A2: Boulevard des Capucines is an oil painting [...] by French Impressionist artist Claude Monet created in 1873. So the answer is true. (✓) |
| A3: Bulb Fields is a 1990 painting by Mark Rothko. So the answer is false. (X) | A3: Bulb Fields is an oil painting created by Vincent van Gogh in early 1883. So the answer is true. (✓) |
| A: An osteria is an Italian type of eating establishment. It is not a 19th-century artwork. So the answer is false. (X) | A: In a Roman Osteria is a painting by the Danish painter Carl Bloch. It was painted in 1866. So the answer is true. (✓) |
| Method | Temporal | Commonsense | Tabular | Strategy | Truthful |
| Self-Ask | 58.91 | 74.45 | 72.91 | 59.38 | 78.77 |
| Recite-and-Answer | 62.45 | 74.61 | 75.00 | 58.95 | 79.64 |
| DSP | 63.95 | 73.21 | 75.75 | 58.73 | 77.88 |
| \\( \\text{SKR}_{\\text{knn}} \\) | 66.13 | 75.43 | 76.75 | 62.01 | 82.30 |
| Retriever | Retrieval Quality | Manual-CoT-IR | Our \\( \\mathbf{S}\\mathbf{K}\\mathbf{R}_{\\mathbf{k}\\mathbf{n}\\mathbf{n}} \\) |
| DPR | high | 76.99 | 82.30 |
| SimCSE | semantically matched | 73.45 (↓ 3.54) | 81.42 (↓ 0.88) |
| BERT | relatively poor | 69.91 (↓ 7.08) | 80.53 (↓ 1.67) |
| Corpora | Selected Candidates | Manual-CoT-IR | Our SKRknn |
| Full C | rank 1~3 | 76.99 | 82.30 |
| Full C | rank 8~10 | 76.11 (↓0.88) | 82.30 (-) |
| Half of C | rank 1~3 | 74.37 (↓2.62) | 81.41 (↓0.89) |
| Method | Temporal (EM/F1) | Commonsense (Acc.) | Tabular (Acc.) | Strategy (Acc.) | Truthful (Acc.) | Avg. | |
| (text-davinci-003) | |||||||
| (w/o retrieval) | Zero-Shot | 40.57/44.94 | 65.52 | 66.08 | 63.32 | 61.06 | 56.91 |
| Zero-Shot-CoT | 41.71/45.33 | 63.31 | 60.50 | 58.52 | 53.10 | 53.75 | |
| Few-Shot | 45.14/49.59 | 80.34 | 63.50 | 66.37 | 65.49 | 61.73 | |
| Manual-CoT | 44.57/54.22 | 75.42 | 74.92 | 71.18 | 72.57 | 65.48 | |
| Auto-CoT (Similarity) | 44.00/54.13 | 74.44 | 77.25 | 70.61 | 72.57 | 65.50 | |
| Auto-CoT (Diversity) | 46.28/54.68 | 73.38 | 74.50 | 70.74 | 71.68 | 65.21 | |
| (w/ retrieval) | Manual-CoT-IR | 47.42/57.37 | 75.67 | 79.25 | 69.43 | 69.03 | 66.36 |
| IRCoT | 47.42/56.28 | 75.27 | 78.00 | 67.68 | 71.68 | 66.06 | |
| \( SKR_{prompt} \) | 45.71/56.31 | 75.02 | 77.33 | 69.43 | 70.80 | 65.77 | |
| \( SKR_{icl} \) | 47.42/57.74 | 75.51 | 77.75 | 71.18 | 73.45 | 67.17 | |
| \( SKR_{cls} \) | 46.28/56.54 | 75.83 | 79.25 | 70.30 | 72.57 | 66.80 | |
| \( SKR_{knn} \) | 48.00/58.47 | 76.66 | 79.83 | 71.62 | 74.34 | 68.15 | |
| \( CoT-RR^† \) | 44.97/56.58 | 76.98 | 82.08 | 74.67 | 69.91 | 67.53 | |
| (gpt-3.5-turbo-0301) | |||||||
| (w/o retrieval) | Zero-Shot | 51.42/54.94 | 73.30 | 66.50 | 54.14 | 74.33 | 62.43 |
| Zero-Shot-CoT | 56.57/58.92 | 61.58 | 63.08 | 44.10 | 65.49 | 58.29 | |
| Few-Shot | 52.57/55.77 | 78.86 | 68.75 | 61.13 | 69.91 | 64.50 | |
| Manual-CoT | 54.85/61.72 | 74.77 | 73.25 | 61.36 | 81.41 | 67.89 | |
| Auto-CoT (Similarity) | 54.28/57.66 | 74.44 | 71.58 | 61.57 | 79.64 | 66.52 | |
| Auto-CoT (Diversity) | 55.43/58.21 | 73.30 | 71.50 | 58.52 | 80.54 | 66.25 | |
| (w/ retrieval) | Manual-CoT-IR | 59.41/63.08 | 73.38 | 76.58 | 57.21 | 76.99 | 67.77 |
| IRCoT | 57.14/61.67 | 72.73 | 76.25 | 55.46 | 79.64 | 67.15 | |
| \( SKR_{prompt} \) | 54.86/61.64 | 75.02 | 73.91 | 61.57 | 80.53 | 67.92 | |
| \( SKR_{icl} \) | 58.29/63.81 | 75.10 | 74.42 | 62.01 | 82.30 | 69.32 | |
| \( SKR_{cls} \) | 59.43/64.04 | 75.10 | 76.42 | 62.01 | 84.95 | 70.33 | |
| \( SKR_{knn} \) | 61.14/66.13 | 75.43 | 76.75 | 62.01 | 82.30 | 70.62 | |
| \( CoT-RR^† \) | 60.35/65.59 | 75.84 | 76.08 | 62.88 | 83.18 | 70.65 | |
| Template | Positive Response (LLM known) | Negative Response (LLM unknown) | Acc. |
| Do you need additional information to answer this question? | No, additional information is not needed to answer this... | Yes, additional information is needed to answer this... | 73.17 |
| Would you like any extra prompts to help you? | No, I do not need any extra... | Yes, please. | 72.32 |
| Would you like any additional clues? | No, the answer is... | Yes, please provide... | 72.32 |
| Can you answer this question based on what you know? | Yes, the correct answer to this question is... | No, I cannot answer it based on what I know. | 72.07 |
| Can you solve this question now? | Yes, the correct answer is... | No, this is not a solvable... | 71.58 |
| Manual-CoT | Manual-CoT-IR |
| A1: Cleopatra and Caesar are a 17th-century Italian paint-ing by Francesco Barberini. So the answer is false. (X) | A1: Cleopatra and Caesar is an oil on canvas painting by the French Academic artist Jean-Léon Gérôme, completed in 1866. So the answer is true. (✓) |
| A2: Boulevard des Capucines is a street in Paris, France. [...] Thus, Boulevard des Capucines is not a 19th-century European work of art. So the answer is false. (X) | A2: Boulevard des Capucines is an oil painting [...] by French Impressionist artist Claude Monet created in 1873. So the answer is true. (✓) |
| A3: Bulb Fields is a 1990 painting by Mark Rothko. So the answer is false. (X) | A3: Bulb Fields is an oil painting created by Vincent van Gogh in early 1883. So the answer is true. (✓) |
| A: An osteria is an Italian type of eating establishment. It is not a 19th-century artwork. So the answer is false. (X) | A: In a Roman Osteria is a painting by the Danish painter Carl Bloch. It was painted in 1866. So the answer is true. (✓) |
| Method | Temporal | Commonsense | Tabular | Strategy | Truthful |
| Self-Ask | 58.91 | 74.45 | 72.91 | 59.38 | 78.77 |
| Recite-and-Answer | 62.45 | 74.61 | 75.00 | 58.95 | 79.64 |
| DSP | 63.95 | 73.21 | 75.75 | 58.73 | 77.88 |
| \( \text{SKR}_{\text{knn}} \) | 66.13 | 75.43 | 76.75 | 62.01 | 82.30 |
| Retriever | Retrieval Quality | Manual-CoT-IR | Our \( \mathbf{S}\mathbf{K}\mathbf{R}_{\mathbf{k}\mathbf{n}\mathbf{n}} \) |
| DPR | high | 76.99 | 82.30 |
| SimCSE | semantically matched | 73.45 (↓ 3.54) | 81.42 (↓ 0.88) |
| BERT | relatively poor | 69.91 (↓ 7.08) | 80.53 (↓ 1.67) |
| Corpora | Selected Candidates | Manual-CoT-IR | Our SKRknn |
| Full C | rank 1~3 | 76.99 | 82.30 |
| Full C | rank 8~10 | 76.11 (↓0.88) | 82.30 (-) |
| Half of C | rank 1~3 | 74.37 (↓2.62) | 81.41 (↓0.89) |
| Method | Temporal (EM/F1) | Commonsense (Acc.) | Tabular (Acc.) | Strategy (Acc.) | Truthful (Acc.) | Avg. | |
| (text-davinci-003) | |||||||
| (w/o retrieval) | Zero-Shot | 40.57/44.94 | 65.52 | 66.08 | 63.32 | 61.06 | 56.91 |
| Zero-Shot-CoT | 41.71/45.33 | 63.31 | 60.50 | 58.52 | 53.10 | 53.75 | |
| Few-Shot | 45.14/49.59 | 80.34 | 63.50 | 66.37 | 65.49 | 61.73 | |
| Manual-CoT | 44.57/54.22 | 75.42 | 74.92 | 71.18 | 72.57 | 65.48 | |
| Auto-CoT (Similarity) | 44.00/54.13 | 74.44 | 77.25 | 70.61 | 72.57 | 65.50 | |
| Auto-CoT (Diversity) | 46.28/54.68 | 73.38 | 74.50 | 70.74 | 71.68 | 65.21 | |
| (w/ retrieval) | Manual-CoT-IR | 47.42/57.37 | 75.67 | 79.25 | 69.43 | 69.03 | 66.36 |
| IRCoT | 47.42/56.28 | 75.27 | 78.00 | 67.68 | 71.68 | 66.06 | |
| \\( SKR_{prompt} \\) | 45.71/56.31 | 75.02 | 77.33 | 69.43 | 70.80 | 65.77 | |
| \\( SKR_{icl} \\) | 47.42/57.74 | 75.51 | 77.75 | 71.18 | 73.45 | 67.17 | |
| \\( SKR_{cls} \\) | 46.28/56.54 | 75.83 | 79.25 | 70.30 | 72.57 | 66.80 | |
| \\( SKR_{knn} \\) | 48.00/58.47 | 76.66 | 79.83 | 71.62 | 74.34 | 68.15 | |
| \\( CoT-RR^† \\) | 44.97/56.58 | 76.98 | 82.08 | 74.67 | 69.91 | 67.53 | |
| (gpt-3.5-turbo-0301) | |||||||
| (w/o retrieval) | Zero-Shot | 51.42/54.94 | 73.30 | 66.50 | 54.14 | 74.33 | 62.43 |
| Zero-Shot-CoT | 56.57/58.92 | 61.58 | 63.08 | 44.10 | 65.49 | 58.29 | |
| Few-Shot | 52.57/55.77 | 78.86 | 68.75 | 61.13 | 69.91 | 64.50 | |
| Manual-CoT | 54.85/61.72 | 74.77 | 73.25 | 61.36 | 81.41 | 67.89 | |
| Auto-CoT (Similarity) | 54.28/57.66 | 74.44 | 71.58 | 61.57 | 79.64 | 66.52 | |
| Auto-CoT (Diversity) | 55.43/58.21 | 73.30 | 71.50 | 58.52 | 80.54 | 66.25 | |
| (w/ retrieval) | Manual-CoT-IR | 59.41/63.08 | 73.38 | 76.58 | 57.21 | 76.99 | 67.77 |
| IRCoT | 57.14/61.67 | 72.73 | 76.25 | 55.46 | 79.64 | 67.15 | |
| \\( SKR_{prompt} \\) | 54.86/61.64 | 75.02 | 73.91 | 61.57 | 80.53 | 67.92 | |
| \\( SKR_{icl} \\) | 58.29/63.81 | 75.10 | 74.42 | 62.01 | 82.30 | 69.32 | |
| \\( SKR_{cls} \\) | 59.43/64.04 | 75.10 | 76.42 | 62.01 | 84.95 | 70.33 | |
| \\( SKR_{knn} \\) | 61.14/66.13 | 75.43 | 76.75 | 62.01 | 82.30 | 70.62 | |
| \\( CoT-RR^† \\) | 60.35/65.59 | 75.84 | 76.08 | 62.88 | 83.18 | 70.65 |
| Template | Positive Response (LLM known) | Negative Response (LLM unknown) | Acc. |
| Do you need additional information to answer this question? | No, additional information is not needed to answer this... | Yes, additional information is needed to answer this... | 73.17 |
| Would you like any extra prompts to help you? | No, I do not need any extra... | Yes, please. | 72.32 |
| Would you like any additional clues? | No, the answer is... | Yes, please provide... | 72.32 |
| Can you answer this question based on what you know? | Yes, the correct answer to this question is... | No, I cannot answer it based on what I know. | 72.07 |
| Can you solve this question now? | Yes, the correct answer is... | No, this is not a solvable... | 71.58 |
| Manual-CoT | Manual-CoT-IR |
| A1: Cleopatra and Caesar are a 17th-century Italian paint-ing by Francesco Barberini. So the answer is false. (X) | A1: Cleopatra and Caesar is an oil on canvas painting by the French Academic artist Jean-Léon Gérôme, completed in 1866. So the answer is true. (✓) |
| A2: Boulevard des Capucines is a street in Paris, France. [...] Thus, Boulevard des Capucines is not a 19th-century European work of art. So the answer is false. (X) | A2: Boulevard des Capucines is an oil painting [...] by French Impressionist artist Claude Monet created in 1873. So the answer is true. (✓) |
| A3: Bulb Fields is a 1990 painting by Mark Rothko. So the answer is false. (X) | A3: Bulb Fields is an oil painting created by Vincent van Gogh in early 1883. So the answer is true. (✓) |
| A: An osteria is an Italian type of eating establishment. It is not a 19th-century artwork. So the answer is false. (X) | A: In a Roman Osteria is a painting by the Danish painter Carl Bloch. It was painted in 1866. So the answer is true. (✓) |
| Method | Temporal | Commonsense | Tabular | Strategy | Truthful |
| Self-Ask | 58.91 | 74.45 | 72.91 | 59.38 | 78.77 |
| Recite-and-Answer | 62.45 | 74.61 | 75.00 | 58.95 | 79.64 |
| DSP | 63.95 | 73.21 | 75.75 | 58.73 | 77.88 |
| \\( \\text{SKR}_{\\text{knn}} \\) | 66.13 | 75.43 | 76.75 | 62.01 | 82.30 |
| Retriever | Retrieval Quality | Manual-CoT-IR | Our \\( \\mathbf{S}\\mathbf{K}\\mathbf{R}_{\\mathbf{k}\\mathbf{n}\\mathbf{n}} \\) |
| DPR | high | 76.99 | 82.30 |
| SimCSE | semantically matched | 73.45 (↓ 3.54) | 81.42 (↓ 0.88) |
| BERT | relatively poor | 69.91 (↓ 7.08) | 80.53 (↓ 1.67) |
| Corpora | Selected Candidates | Manual-CoT-IR | Our SKRknn |
| Full C | rank 1~3 | 76.99 | 82.30 |
| Full C | rank 8~10 | 76.11 (↓0.88) | 82.30 (-) |
| Half of C | rank 1~3 | 74.37 (↓2.62) | 81.41 (↓0.89) |
| METHOD | Progressively | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Standard | X | 15.8 | 28.3 | 72.9 | 35.1 | 28.2 | 36.1 |
| +Zero-shot SP | X | 22.4(↑ 6.6) | 28.3(0) | 73.2(↑ 0.3) | 43.1(↑ 8.0) | 25.4(↓ 2.8) | 38.5(↑ 2.4) |
| ✓ | 24.0(↑ 8.2) | 28.7(↑ 0.4) | 72.2(↓ 0.7) | 51.7(↑ 16.6) | 26.8(↓ 1.4) | 40.7(↑ 4.6) | |
| +In-context SP | X | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 73.9(↑ 1.0) | 50.6(↑ 15.5) | 29.4(↑ 0.8) | 41.7(↑ 5.6) |
| ✓ | 25.3(↑ 9.5) | 29.5(↑ 1.2) | 73.9(↑ 1.0) | 52.9(↑ 17.8) | 28.6(↑ 0.4) | 42.0(↑ 5.9) | |
| +Auto-SP | X | 24.3(↑ 8.5) | 29.9(↑ 1.6) | 72.6(↓ 0.3) | 54.0(↑ 18.9) | 27.6(↓ 0.6) | 41.7(↑ 5.6) |
| ✓ | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 72.9(0) | 56.7(↑ 21.6) | 28.2(0) | 42.5(↑ 6.4) | |
| +Complex-SP | X | 23.0(↑ 7.2) | 29.9(↑ 1.6) | 73.2(↑ 0.3) | 52.3(↑ 17.2) | 29.6(↑ 1.4) | 41.6(↑ 5.5) |
| ✓ | 24.6(↑ 8.8) | 28.7(↑ 0.4) | 72.9(0) | 55.7(↑ 20.6) | 30.0(↑ 1.8) | 42.4(↑ 6.3) |
| ANSWER SIDE | PROBLEM SIDE | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Chain-of-Thought | No Refinement | 56.1 | 44.9 | 80.3 | 90.8 | 41.0 | 62.6 |
| In-context SP | 56.7(↑ 0.6) | 48.8(↑ 3.9) | 81.6(↑ 1.3) | 88.5(↓ 2.3) | 43.0(↑ 2.0) | 63.7(↑ 1.1) | |
| Auto-SP | 56.9(↑ 0.8) | 49.2(↑ 4.3) | 78.3(↓ 2.0) | 89.7(↓ 1.1) | 42.8(↑ 1.8) | 63.4(↑ 0.8) | |
| Complex-SP | 58.1(↑ 2.0) | 47.3(↑ 2.4) | 81.3(↑ 1.0) | 90.2(↓ 0.6) | 43.2(↑ 2.2) | 64.0(↑ 1.4) | |
| Least-to-Most | No Refinement | 59.3 | 42.1 | 82.9 | 83.9 | 39.0 | 61.4 |
| In-context SP | 61.2(↑ 1.9) | 44.1(↑ 2.0) | 84.9(↑ 2.0) | 85.1(↑ 1.2) | 41.2(↑ 2.2) | 63.3(↑ 1.9) | |
| Auto-SP | 61.6(↑ 2.3) | 44.9(↑ 2.8) | 82.9(0) | 83.9(0) | 41.2(↑ 2.2) | 62.9(↑ 1.5) | |
| Complex-SP | 62.9(↑ 3.6) | 47.6(↑ 5.5) | 84.2(↑ 1.3) | 86.2(↑ 2.3) | 40.6(↑ 1.6) | 64.3 (↑ 2.9) | |
| Auto-CoT | No Refinement | 59.4 | 46.5 | 75.6 | 92.5 | 41.4 | 63.1 |
| In-context SP | 59.4(0) | 47.2(↑ 0.7) | 77.6(↑ 2.0) | 90.6(↓ 1.9) | 45.8(↑ 4.4) | 64.1(↑ 1.0) | |
| Auto-SP | 60.5(↑ 1.1) | 48.0(↑ 1.5) | 75.6(0) | 89.7(↓ 2.8) | 44.4(↑ 3.0) | 63.6(↑ 0.5) | |
| Complex-SP | 60.1(↑ 0.7) | 50.8(↑ 4.3) | 78.2(↑ 2.6) | 93.1(↑ 0.6) | 43.0(↑ 1.6) | 65.0(↑ 1.9) | |
| Complex-CoT | No Refinement | 67.5 | 47.6 | 78.3 | 94.3 | 43.6 | 66.3 |
| In-context SP | 66.2(↓ 1.3) | 50.8(↑ 3.1) | 80.9(↑ 2.6) | 90.8(↓ 3.5) | 44.6(↑ 1.3) | 66.7(↑ 0.4) | |
| Auto-SP | 68.7(↑ 1.2) | 45.7(↓ 1.9) | 78.3(0) | 92.0(↓ 2.3) | 42.4(↓ 1.2) | 65.4(↓ 0.9) | |
| Complex-SP | 68.5(↑ 1.0) | 50.4(↑ 2.8) | 78.6(↑ 0.3) | 93.1(↓ 1.2) | 44.8(↑ 1.2) | 67.1(↑ 0.8) |
| METHOD | SC Path | Self-Polish | DATASET | |
| AQuA | MathQA | |||
| Auto-CoT | 5 | X | 61.8 | 65.2 |
| 5 | ✓ | 66.1 | 68.4 | |
| 10 | X | 61.4 | 67.4 | |
| 10 | ✓ | 65.0 | 71.0 | |
| 20 | X | 63.0 | 69.0 | |
| 20 | ✓ | 68.1 | 72.0 | |
| Complex-CoT | 5 | X | 61.4 | 61.6 |
| 5 | ✓ | 65.0 | 64.4 | |
| 10 | X | 65.4 | 62.4 | |
| 10 | ✓ | 66.1 | 64.8 | |
| 20 | X | 65.4 | 64.0 | |
| 20 | ✓ | 67.0 | 65.6 |
| METHOD | SHOTS | Mean | Order Deviation |
| Std+SP | 2 | 21.3 | 0.8 |
| 3 | 22.5 | 0.6 | |
| 4 | 24.1 | 0.9 | |
| 5 | 25.0 | 1.5 | |
| 6 | 26.3 | 0.9 | |
| CoT+SP | 2 | 59.0 | 2.0 |
| 3 | 61.3 | 1.5 | |
| 4 | 61.5 | 1.1 | |
| 5 | 62.1 | 1.8 | |
| 6 | 61.3 | 1.6 |
| METHOD | MATH |
| No Refinement | 21 |
| Zero-shot SP | 23.5 |
| In-context SP | 24.5 |
| METHOD | Progressively | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Standard | X | 15.8 | 28.3 | 72.9 | 35.1 | 28.2 | 36.1 |
| +Zero-shot SP | X | 22.4(↑ 6.6) | 28.3(0) | 73.2(↑ 0.3) | 43.1(↑ 8.0) | 25.4(↓ 2.8) | 38.5(↑ 2.4) |
| ✓ | 24.0(↑ 8.2) | 28.7(↑ 0.4) | 72.2(↓ 0.7) | 51.7(↑ 16.6) | 26.8(↓ 1.4) | 40.7(↑ 4.6) | |
| +In-context SP | X | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 73.9(↑ 1.0) | 50.6(↑ 15.5) | 29.4(↑ 0.8) | 41.7(↑ 5.6) |
| ✓ | 25.3(↑ 9.5) | 29.5(↑ 1.2) | 73.9(↑ 1.0) | 52.9(↑ 17.8) | 28.6(↑ 0.4) | 42.0(↑ 5.9) | |
| +Auto-SP | X | 24.3(↑ 8.5) | 29.9(↑ 1.6) | 72.6(↓ 0.3) | 54.0(↑ 18.9) | 27.6(↓ 0.6) | 41.7(↑ 5.6) |
| ✓ | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 72.9(0) | 56.7(↑ 21.6) | 28.2(0) | 42.5(↑ 6.4) | |
| +Complex-SP | X | 23.0(↑ 7.2) | 29.9(↑ 1.6) | 73.2(↑ 0.3) | 52.3(↑ 17.2) | 29.6(↑ 1.4) | 41.6(↑ 5.5) |
| ✓ | 24.6(↑ 8.8) | 28.7(↑ 0.4) | 72.9(0) | 55.7(↑ 20.6) | 30.0(↑ 1.8) | 42.4(↑ 6.3) |
| ANSWER SIDE | PROBLEM SIDE | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Chain-of-Thought | No Refinement | 56.1 | 44.9 | 80.3 | 90.8 | 41.0 | 62.6 |
| In-context SP | 56.7(↑ 0.6) | 48.8(↑ 3.9) | 81.6(↑ 1.3) | 88.5(↓ 2.3) | 43.0(↑ 2.0) | 63.7(↑ 1.1) | |
| Auto-SP | 56.9(↑ 0.8) | 49.2(↑ 4.3) | 78.3(↓ 2.0) | 89.7(↓ 1.1) | 42.8(↑ 1.8) | 63.4(↑ 0.8) | |
| Complex-SP | 58.1(↑ 2.0) | 47.3(↑ 2.4) | 81.3(↑ 1.0) | 90.2(↓ 0.6) | 43.2(↑ 2.2) | 64.0(↑ 1.4) | |
| Least-to-Most | No Refinement | 59.3 | 42.1 | 82.9 | 83.9 | 39.0 | 61.4 |
| In-context SP | 61.2(↑ 1.9) | 44.1(↑ 2.0) | 84.9(↑ 2.0) | 85.1(↑ 1.2) | 41.2(↑ 2.2) | 63.3(↑ 1.9) | |
| Auto-SP | 61.6(↑ 2.3) | 44.9(↑ 2.8) | 82.9(0) | 83.9(0) | 41.2(↑ 2.2) | 62.9(↑ 1.5) | |
| Complex-SP | 62.9(↑ 3.6) | 47.6(↑ 5.5) | 84.2(↑ 1.3) | 86.2(↑ 2.3) | 40.6(↑ 1.6) | 64.3 (↑ 2.9) | |
| Auto-CoT | No Refinement | 59.4 | 46.5 | 75.6 | 92.5 | 41.4 | 63.1 |
| In-context SP | 59.4(0) | 47.2(↑ 0.7) | 77.6(↑ 2.0) | 90.6(↓ 1.9) | 45.8(↑ 4.4) | 64.1(↑ 1.0) | |
| Auto-SP | 60.5(↑ 1.1) | 48.0(↑ 1.5) | 75.6(0) | 89.7(↓ 2.8) | 44.4(↑ 3.0) | 63.6(↑ 0.5) | |
| Complex-SP | 60.1(↑ 0.7) | 50.8(↑ 4.3) | 78.2(↑ 2.6) | 93.1(↑ 0.6) | 43.0(↑ 1.6) | 65.0(↑ 1.9) | |
| Complex-CoT | No Refinement | 67.5 | 47.6 | 78.3 | 94.3 | 43.6 | 66.3 |
| In-context SP | 66.2(↓ 1.3) | 50.8(↑ 3.1) | 80.9(↑ 2.6) | 90.8(↓ 3.5) | 44.6(↑ 1.3) | 66.7(↑ 0.4) | |
| Auto-SP | 68.7(↑ 1.2) | 45.7(↓ 1.9) | 78.3(0) | 92.0(↓ 2.3) | 42.4(↓ 1.2) | 65.4(↓ 0.9) | |
| Complex-SP | 68.5(↑ 1.0) | 50.4(↑ 2.8) | 78.6(↑ 0.3) | 93.1(↓ 1.2) | 44.8(↑ 1.2) | 67.1(↑ 0.8) |
| METHOD | SC Path | Self-Polish | DATASET | |
| AQuA | MathQA | |||
| Auto-CoT | 5 | X | 61.8 | 65.2 |
| 5 | ✓ | 66.1 | 68.4 | |
| 10 | X | 61.4 | 67.4 | |
| 10 | ✓ | 65.0 | 71.0 | |
| 20 | X | 63.0 | 69.0 | |
| 20 | ✓ | 68.1 | 72.0 | |
| Complex-CoT | 5 | X | 61.4 | 61.6 |
| 5 | ✓ | 65.0 | 64.4 | |
| 10 | X | 65.4 | 62.4 | |
| 10 | ✓ | 66.1 | 64.8 | |
| 20 | X | 65.4 | 64.0 | |
| 20 | ✓ | 67.0 | 65.6 |
| METHOD | SHOTS | Mean | Order Deviation |
| Std+SP | 2 | 21.3 | 0.8 |
| 3 | 22.5 | 0.6 | |
| 4 | 24.1 | 0.9 | |
| 5 | 25.0 | 1.5 | |
| 6 | 26.3 | 0.9 | |
| CoT+SP | 2 | 59.0 | 2.0 |
| 3 | 61.3 | 1.5 | |
| 4 | 61.5 | 1.1 | |
| 5 | 62.1 | 1.8 | |
| 6 | 61.3 | 1.6 |
| METHOD | MATH |
| No Refinement | 21 |
| Zero-shot SP | 23.5 |
| In-context SP | 24.5 |
| METHOD | Progressively | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Standard | X | 15.8 | 28.3 | 72.9 | 35.1 | 28.2 | 36.1 |
| +Zero-shot SP | X | 22.4(↑ 6.6) | 28.3(0) | 73.2(↑ 0.3) | 43.1(↑ 8.0) | 25.4(↓ 2.8) | 38.5(↑ 2.4) |
| ✓ | 24.0(↑ 8.2) | 28.7(↑ 0.4) | 72.2(↓ 0.7) | 51.7(↑ 16.6) | 26.8(↓ 1.4) | 40.7(↑ 4.6) | |
| +In-context SP | X | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 73.9(↑ 1.0) | 50.6(↑ 15.5) | 29.4(↑ 0.8) | 41.7(↑ 5.6) |
| ✓ | 25.3(↑ 9.5) | 29.5(↑ 1.2) | 73.9(↑ 1.0) | 52.9(↑ 17.8) | 28.6(↑ 0.4) | 42.0(↑ 5.9) | |
| +Auto-SP | X | 24.3(↑ 8.5) | 29.9(↑ 1.6) | 72.6(↓ 0.3) | 54.0(↑ 18.9) | 27.6(↓ 0.6) | 41.7(↑ 5.6) |
| ✓ | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 72.9(0) | 56.7(↑ 21.6) | 28.2(0) | 42.5(↑ 6.4) | |
| +Complex-SP | X | 23.0(↑ 7.2) | 29.9(↑ 1.6) | 73.2(↑ 0.3) | 52.3(↑ 17.2) | 29.6(↑ 1.4) | 41.6(↑ 5.5) |
| ✓ | 24.6(↑ 8.8) | 28.7(↑ 0.4) | 72.9(0) | 55.7(↑ 20.6) | 30.0(↑ 1.8) | 42.4(↑ 6.3) | |
| ANSWER SIDE | PROBLEM SIDE | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Chain-of-Thought | No Refinement | 56.1 | 44.9 | 80.3 | 90.8 | 41.0 | 62.6 |
| In-context SP | 56.7(↑ 0.6) | 48.8(↑ 3.9) | 81.6(↑ 1.3) | 88.5(↓ 2.3) | 43.0(↑ 2.0) | 63.7(↑ 1.1) | |
| Auto-SP | 56.9(↑ 0.8) | 49.2(↑ 4.3) | 78.3(↓ 2.0) | 89.7(↓ 1.1) | 42.8(↑ 1.8) | 63.4(↑ 0.8) | |
| Complex-SP | 58.1(↑ 2.0) | 47.3(↑ 2.4) | 81.3(↑ 1.0) | 90.2(↓ 0.6) | 43.2(↑ 2.2) | 64.0(↑ 1.4) | |
| Least-to-Most | No Refinement | 59.3 | 42.1 | 82.9 | 83.9 | 39.0 | 61.4 |
| In-context SP | 61.2(↑ 1.9) | 44.1(↑ 2.0) | 84.9(↑ 2.0) | 85.1(↑ 1.2) | 41.2(↑ 2.2) | 63.3(↑ 1.9) | |
| Auto-SP | 61.6(↑ 2.3) | 44.9(↑ 2.8) | 82.9(0) | 83.9(0) | 41.2(↑ 2.2) | 62.9(↑ 1.5) | |
| Complex-SP | 62.9(↑ 3.6) | 47.6(↑ 5.5) | 84.2(↑ 1.3) | 86.2(↑ 2.3) | 40.6(↑ 1.6) | 64.3 (↑ 2.9) | |
| Auto-CoT | No Refinement | 59.4 | 46.5 | 75.6 | 92.5 | 41.4 | 63.1 |
| In-context SP | 59.4(0) | 47.2(↑ 0.7) | 77.6(↑ 2.0) | 90.6(↓ 1.9) | 45.8(↑ 4.4) | 64.1(↑ 1.0) | |
| Auto-SP | 60.5(↑ 1.1) | 48.0(↑ 1.5) | 75.6(0) | 89.7(↓ 2.8) | 44.4(↑ 3.0) | 63.6(↑ 0.5) | |
| Complex-SP | 60.1(↑ 0.7) | 50.8(↑ 4.3) | 78.2(↑ 2.6) | 93.1(↑ 0.6) | 43.0(↑ 1.6) | 65.0(↑ 1.9) | |
| Complex-CoT | No Refinement | 67.5 | 47.6 | 78.3 | 94.3 | 43.6 | 66.3 |
| In-context SP | 66.2(↓ 1.3) | 50.8(↑ 3.1) | 80.9(↑ 2.6) | 90.8(↓ 3.5) | 44.6(↑ 1.3) | 66.7(↑ 0.4) | |
| Auto-SP | 68.7(↑ 1.2) | 45.7(↓ 1.9) | 78.3(0) | 92.0(↓ 2.3) | 42.4(↓ 1.2) | 65.4(↓ 0.9) | |
| Complex-SP | 68.5(↑ 1.0) | 50.4(↑ 2.8) | 78.6(↑ 0.3) | 93.1(↓ 1.2) | 44.8(↑ 1.2) | 67.1(↑ 0.8) | |
| METHOD | SC Path | Self-Polish | DATASET | |
| AQuA | MathQA | |||
| Auto-CoT | 5 | X | 61.8 | 65.2 |
| 5 | ✓ | 66.1 | 68.4 | |
| 10 | X | 61.4 | 67.4 | |
| 10 | ✓ | 65.0 | 71.0 | |
| 20 | X | 63.0 | 69.0 | |
| 20 | ✓ | 68.1 | 72.0 | |
| Complex-CoT | 5 | X | 61.4 | 61.6 |
| 5 | ✓ | 65.0 | 64.4 | |
| 10 | X | 65.4 | 62.4 | |
| 10 | ✓ | 66.1 | 64.8 | |
| 20 | X | 65.4 | 64.0 | |
| 20 | ✓ | 67.0 | 65.6 | |
| METHOD | SHOTS | Mean | Order Deviation |
| Std+SP | 2 | 21.3 | 0.8 |
| 3 | 22.5 | 0.6 | |
| 4 | 24.1 | 0.9 | |
| 5 | 25.0 | 1.5 | |
| 6 | 26.3 | 0.9 | |
| CoT+SP | 2 | 59.0 | 2.0 |
| 3 | 61.3 | 1.5 | |
| 4 | 61.5 | 1.1 | |
| 5 | 62.1 | 1.8 | |
| 6 | 61.3 | 1.6 |
| METHOD | MATH |
| No Refinement | 21 |
| Zero-shot SP | 23.5 |
| In-context SP | 24.5 |
| METHOD | Progressively | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Standard | X | 15.8 | 28.3 | 72.9 | 35.1 | 28.2 | 36.1 |
| +Zero-shot SP | X | 22.4(↑ 6.6) | 28.3(0) | 73.2(↑ 0.3) | 43.1(↑ 8.0) | 25.4(↓ 2.8) | 38.5(↑ 2.4) |
| ✓ | 24.0(↑ 8.2) | 28.7(↑ 0.4) | 72.2(↓ 0.7) | 51.7(↑ 16.6) | 26.8(↓ 1.4) | 40.7(↑ 4.6) | |
| +In-context SP | X | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 73.9(↑ 1.0) | 50.6(↑ 15.5) | 29.4(↑ 0.8) | 41.7(↑ 5.6) |
| ✓ | 25.3(↑ 9.5) | 29.5(↑ 1.2) | 73.9(↑ 1.0) | 52.9(↑ 17.8) | 28.6(↑ 0.4) | 42.0(↑ 5.9) | |
| +Auto-SP | X | 24.3(↑ 8.5) | 29.9(↑ 1.6) | 72.6(↓ 0.3) | 54.0(↑ 18.9) | 27.6(↓ 0.6) | 41.7(↑ 5.6) |
| ✓ | 24.3(↑ 8.5) | 30.3(↑ 2.0) | 72.9(0) | 56.7(↑ 21.6) | 28.2(0) | 42.5(↑ 6.4) | |
| +Complex-SP | X | 23.0(↑ 7.2) | 29.9(↑ 1.6) | 73.2(↑ 0.3) | 52.3(↑ 17.2) | 29.6(↑ 1.4) | 41.6(↑ 5.5) |
| ✓ | 24.6(↑ 8.8) | 28.7(↑ 0.4) | 72.9(0) | 55.7(↑ 20.6) | 30.0(↑ 1.8) | 42.4(↑ 6.3) |
| ANSWER SIDE | PROBLEM SIDE | DATASET | AVERAGE | ||||
| GSM8K | AQuA | SVAMP | MultiArith | MathQA | |||
| Chain-of-Thought | No Refinement | 56.1 | 44.9 | 80.3 | 90.8 | 41.0 | 62.6 |
| In-context SP | 56.7(↑ 0.6) | 48.8(↑ 3.9) | 81.6(↑ 1.3) | 88.5(↓ 2.3) | 43.0(↑ 2.0) | 63.7(↑ 1.1) | |
| Auto-SP | 56.9(↑ 0.8) | 49.2(↑ 4.3) | 78.3(↓ 2.0) | 89.7(↓ 1.1) | 42.8(↑ 1.8) | 63.4(↑ 0.8) | |
| Complex-SP | 58.1(↑ 2.0) | 47.3(↑ 2.4) | 81.3(↑ 1.0) | 90.2(↓ 0.6) | 43.2(↑ 2.2) | 64.0(↑ 1.4) | |
| Least-to-Most | No Refinement | 59.3 | 42.1 | 82.9 | 83.9 | 39.0 | 61.4 |
| In-context SP | 61.2(↑ 1.9) | 44.1(↑ 2.0) | 84.9(↑ 2.0) | 85.1(↑ 1.2) | 41.2(↑ 2.2) | 63.3(↑ 1.9) | |
| Auto-SP | 61.6(↑ 2.3) | 44.9(↑ 2.8) | 82.9(0) | 83.9(0) | 41.2(↑ 2.2) | 62.9(↑ 1.5) | |
| Complex-SP | 62.9(↑ 3.6) | 47.6(↑ 5.5) | 84.2(↑ 1.3) | 86.2(↑ 2.3) | 40.6(↑ 1.6) | 64.3 (↑ 2.9) | |
| Auto-CoT | No Refinement | 59.4 | 46.5 | 75.6 | 92.5 | 41.4 | 63.1 |
| In-context SP | 59.4(0) | 47.2(↑ 0.7) | 77.6(↑ 2.0) | 90.6(↓ 1.9) | 45.8(↑ 4.4) | 64.1(↑ 1.0) | |
| Auto-SP | 60.5(↑ 1.1) | 48.0(↑ 1.5) | 75.6(0) | 89.7(↓ 2.8) | 44.4(↑ 3.0) | 63.6(↑ 0.5) | |
| Complex-SP | 60.1(↑ 0.7) | 50.8(↑ 4.3) | 78.2(↑ 2.6) | 93.1(↑ 0.6) | 43.0(↑ 1.6) | 65.0(↑ 1.9) | |
| Complex-CoT | No Refinement | 67.5 | 47.6 | 78.3 | 94.3 | 43.6 | 66.3 |
| In-context SP | 66.2(↓ 1.3) | 50.8(↑ 3.1) | 80.9(↑ 2.6) | 90.8(↓ 3.5) | 44.6(↑ 1.3) | 66.7(↑ 0.4) | |
| Auto-SP | 68.7(↑ 1.2) | 45.7(↓ 1.9) | 78.3(0) | 92.0(↓ 2.3) | 42.4(↓ 1.2) | 65.4(↓ 0.9) | |
| Complex-SP | 68.5(↑ 1.0) | 50.4(↑ 2.8) | 78.6(↑ 0.3) | 93.1(↓ 1.2) | 44.8(↑ 1.2) | 67.1(↑ 0.8) |
| METHOD | SC Path | Self-Polish | DATASET | |
| AQuA | MathQA | |||
| Auto-CoT | 5 | X | 61.8 | 65.2 |
| 5 | ✓ | 66.1 | 68.4 | |
| 10 | X | 61.4 | 67.4 | |
| 10 | ✓ | 65.0 | 71.0 | |
| 20 | X | 63.0 | 69.0 | |
| 20 | ✓ | 68.1 | 72.0 | |
| Complex-CoT | 5 | X | 61.4 | 61.6 |
| 5 | ✓ | 65.0 | 64.4 | |
| 10 | X | 65.4 | 62.4 | |
| 10 | ✓ | 66.1 | 64.8 | |
| 20 | X | 65.4 | 64.0 | |
| 20 | ✓ | 67.0 | 65.6 |
| METHOD | SHOTS | Mean | Order Deviation |
| Std+SP | 2 | 21.3 | 0.8 |
| 3 | 22.5 | 0.6 | |
| 4 | 24.1 | 0.9 | |
| 5 | 25.0 | 1.5 | |
| 6 | 26.3 | 0.9 | |
| CoT+SP | 2 | 59.0 | 2.0 |
| 3 | 61.3 | 1.5 | |
| 4 | 61.5 | 1.1 | |
| 5 | 62.1 | 1.8 | |
| 6 | 61.3 | 1.6 |
| METHOD | MATH |
| No Refinement | 21 |
| Zero-shot SP | 23.5 |
| In-context SP | 24.5 |
| Benchmark | DRRN \nBaseline | GPT-4 \nBaseline | Behavior Cloned \nSupervised | Transformer \nSelf-Supervised | ||||
| Score | Steps | Score | Steps | Score | Steps | Score | Steps | |
| Arithmetic | 0.17 | 10 | 0.76 | 15 | 0.54 | 5 | 0.44 | 5 |
| Sorting | 0.03 | 21 | 0.04 | 7 | 0.72 | 7 | 0.59 | 6 |
| TWC | 0.57 | 27 | 0.90 | 6 | 0.90 | 6 | 0.89 | 15 |
| Average | 0.26 | 19 | 0.57 | 9 | 0.72 | 6 | 0.64 | 9 |
| Arithmetic | |
| SS | TAKE(X), READ(X), LOOK-AROUND, TAKE(Y), PUT(Y, Z) |
| Gold | TAKE(X), READ(X), TAKE(Y), PUT(Y, Z) |
| Sorting | |
| SS | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Gold | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Text World Common Sense | |
| SS | (TAKE(X), LOOK-AROUND, PUT(X, Y)) |
| (TAKE(X), INVENTORY, OPEN(Y), PUT(X, Y)) | |
| Gold | (TAKE(X), PUT(X, Y)) |
| (TAKE(X), OPEN(Y), PUT(X, Y)) |
| Benchmark | DRRN \nBaseline | GPT-4 \nBaseline | Behavior Cloned \nSupervised | Transformer \nSelf-Supervised | ||||
| Score | Steps | Score | Steps | Score | Steps | Score | Steps | |
| Arithmetic | 0.17 | 10 | 0.76 | 15 | 0.54 | 5 | 0.44 | 5 |
| Sorting | 0.03 | 21 | 0.04 | 7 | 0.72 | 7 | 0.59 | 6 |
| TWC | 0.57 | 27 | 0.90 | 6 | 0.90 | 6 | 0.89 | 15 |
| Average | 0.26 | 19 | 0.57 | 9 | 0.72 | 6 | 0.64 | 9 |
| Arithmetic | |
| SS | TAKE(X), READ(X), LOOK-AROUND, TAKE(Y), PUT(Y, Z) |
| Gold | TAKE(X), READ(X), TAKE(Y), PUT(Y, Z) |
| Sorting | |
| SS | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Gold | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Text World Common Sense | |
| SS | (TAKE(X), LOOK-AROUND, PUT(X, Y)) |
| (TAKE(X), INVENTORY, OPEN(Y), PUT(X, Y)) | |
| Gold | (TAKE(X), PUT(X, Y)) |
| (TAKE(X), OPEN(Y), PUT(X, Y)) |
| Benchmark | DRRN +Baseline | GPT-4 +Baseline | Behavior Cloned +Supervised | Transformer +Self-Supervised | ||||
| Score | Steps | Score | Steps | Score | Steps | Score | Steps | |
| Arithmetic | 0.17 | 10 | 0.76 | 15 | 0.54 | 5 | 0.44 | 5 |
| Sorting | 0.03 | 21 | 0.04 | 7 | 0.72 | 7 | 0.59 | 6 |
| TWC | 0.57 | 27 | 0.90 | 6 | 0.90 | 6 | 0.89 | 15 |
| Average | 0.26 | 19 | 0.57 | 9 | 0.72 | 6 | 0.64 | 9 |
| Arithmetic | |
| SS | TAKE(X), READ(X), LOOK-AROUND, TAKE(Y), PUT(Y, Z) |
| Gold | TAKE(X), READ(X), TAKE(Y), PUT(Y, Z) |
| Sorting | |
| SS | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Gold | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Text World Common Sense | |
| SS | (TAKE(X), LOOK-AROUND, PUT(X, Y)) |
| (TAKE(X), INVENTORY, OPEN(Y), PUT(X, Y)) | |
| Gold | (TAKE(X), PUT(X, Y)) |
| (TAKE(X), OPEN(Y), PUT(X, Y)) | |
| Benchmark | DRRN \nBaseline | GPT-4 \nBaseline | Behavior Cloned \nSupervised | Transformer \nSelf-Supervised | ||||
| Score | Steps | Score | Steps | Score | Steps | Score | Steps | |
| Arithmetic | 0.17 | 10 | 0.76 | 15 | 0.54 | 5 | 0.44 | 5 |
| Sorting | 0.03 | 21 | 0.04 | 7 | 0.72 | 7 | 0.59 | 6 |
| TWC | 0.57 | 27 | 0.90 | 6 | 0.90 | 6 | 0.89 | 15 |
| Average | 0.26 | 19 | 0.57 | 9 | 0.72 | 6 | 0.64 | 9 |
| Arithmetic | |
| SS | TAKE(X), READ(X), LOOK-AROUND, TAKE(Y), PUT(Y, Z) |
| Gold | TAKE(X), READ(X), TAKE(Y), PUT(Y, Z) |
| Sorting | |
| SS | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Gold | TAKE(X), PUT(X, Y), TAKE(Z), PUT(Z, Y), TAKE(A), ... |
| Text World Common Sense | |
| SS | (TAKE(X), LOOK-AROUND, PUT(X, Y)) |
| (TAKE(X), INVENTORY, OPEN(Y), PUT(X, Y)) | |
| Gold | (TAKE(X), PUT(X, Y)) |
| (TAKE(X), OPEN(Y), PUT(X, Y)) |
| Training | Testing | GAT | Our approach |
| Clean | Clean | 1.0 | 1.0 |
| Supporting | 0.24 | 1.0 | |
| Irrelevant | 0.51 | 1.0 | |
| Disconnected | 0.8 | 1.0 | |
| Supporting | Clean | 0.92 | 0.89 |
| Supporting | 0.98 | 0.97 | |
| Irrelevant | 0.5 | 0.94 | |
| Disconnected | 0.92 | 0.91 | |
| Irrelevant | Clean | 0.92 | 0.94 |
| Supporting | 0.77 | 0.94 | |
| Irrelevant | 0.93 | 0.96 | |
| Disconnected | 0.85 | 0.88 | |
| Disconnected | Clean | 0.75 | 0.91 |
| Supporting | 0.78 | 0.93 | |
| Irrelevant | 0.56 | 0.91 | |
| Disconnected | 0.96 | 0.93 | |
| Average | 0.77 | 0.94 |
| Method | Recall |
| SYGMA | 55.0 |
| Our approach | 53.7 |
| SYGMA (with GT entities) | 68.0 |
| Our approach (with GT entities) | 70.1 |
| Training | Testing | GAT | Our approach |
| Clean | Clean | 1.0 | 1.0 |
| Supporting | 0.24 | 1.0 | |
| Irrelevant | 0.51 | 1.0 | |
| Disconnected | 0.8 | 1.0 | |
| Supporting | Clean | 0.92 | 0.89 |
| Supporting | 0.98 | 0.97 | |
| Irrelevant | 0.5 | 0.94 | |
| Disconnected | 0.92 | 0.91 | |
| Irrelevant | Clean | 0.92 | 0.94 |
| Supporting | 0.77 | 0.94 | |
| Irrelevant | 0.93 | 0.96 | |
| Disconnected | 0.85 | 0.88 | |
| Disconnected | Clean | 0.75 | 0.91 |
| Supporting | 0.78 | 0.93 | |
| Irrelevant | 0.56 | 0.91 | |
| Disconnected | 0.96 | 0.93 | |
| Average | 0.77 | 0.94 |
| Method | Recall |
| SYGMA | 55.0 |
| Our approach | 53.7 |
| SYGMA (with GT entities) | 68.0 |
| Our approach (with GT entities) | 70.1 |
| Training | Testing | GAT | Our approach |
| Clean | Clean | 1.0 | 1.0 |
| Supporting | 0.24 | 1.0 | |
| Irrelevant | 0.51 | 1.0 | |
| Disconnected | 0.8 | 1.0 | |
| Supporting | Clean | 0.92 | 0.89 |
| Supporting | 0.98 | 0.97 | |
| Irrelevant | 0.5 | 0.94 | |
| Disconnected | 0.92 | 0.91 | |
| Irrelevant | Clean | 0.92 | 0.94 |
| Supporting | 0.77 | 0.94 | |
| Irrelevant | 0.93 | 0.96 | |
| Disconnected | 0.85 | 0.88 | |
| Disconnected | Clean | 0.75 | 0.91 |
| Supporting | 0.78 | 0.93 | |
| Irrelevant | 0.56 | 0.91 | |
| Disconnected | 0.96 | 0.93 | |
| Average | 0.77 | 0.94 |
| Method | Recall |
| SYGMA | 55.0 |
| Our approach | 53.7 |
| SYGMA (with GT entities) | 68.0 |
| Our approach (with GT entities) | 70.1 |
| Training | Testing | GAT | Our approach |
| Clean | Clean | 1.0 | 1.0 |
| Supporting | 0.24 | 1.0 | |
| Irrelevant | 0.51 | 1.0 | |
| Disconnected | 0.8 | 1.0 | |
| Supporting | Clean | 0.92 | 0.89 |
| Supporting | 0.98 | 0.97 | |
| Irrelevant | 0.5 | 0.94 | |
| Disconnected | 0.92 | 0.91 | |
| Irrelevant | Clean | 0.92 | 0.94 |
| Supporting | 0.77 | 0.94 | |
| Irrelevant | 0.93 | 0.96 | |
| Disconnected | 0.85 | 0.88 | |
| Disconnected | Clean | 0.75 | 0.91 |
| Supporting | 0.78 | 0.93 | |
| Irrelevant | 0.56 | 0.91 | |
| Disconnected | 0.96 | 0.93 | |
| Average | 0.77 | 0.94 |
| Method | Recall |
| SYGMA | 55.0 |
| Our approach | 53.7 |
| SYGMA (with GT entities) | 68.0 |
| Our approach (with GT entities) | 70.1 |
| Dataset | Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | ||
| held-out | 522.6 | 18.4 | 172.4 | 2.0 | 53 |
| manual | 417.9 | 17.1 | 9.7 | 3.9 | 25 |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 33.8 | 40.7 | 75.0 | 72.5 | 65.7 | 71.1 |
| RESIDE | 41.5 | 45.7 | 84.0 | 78.5 | 75.6 | 79.4 |
| DISTRE | 42.2 | 48.6 | 68.0 | 67.0 | 65.3 | 66.8 |
| Intra+inter | 42.3 | 46.5 | 91.8 | 84.0 | 78.7 | 84.8 |
| CIL | 50.8 | 52.2 | 90.1 | 86.1 | 81.8 | 86.0 |
| HiCLRE | 45.3 | 50.5 | 82.0 | 78.5 | 74.0 | 78.2 |
| PARE | 53.4 | 54.4 | 85.0 | 85.0 | 82.7 | 84.8 |
| Teacher | 50.6 | 52.2 | 90.0 | 82.0 | 78.7 | 83.6 |
| Last Student | 53.9 | 55.3 | 88.0 | 84.0 | 82.7 | 84.9 |
| Model | AUC | Micro-F1 | Macro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 57.7 | 57.0 | 21.2 | 91.0 | 88.5 | 88.0 | 89.2 |
| Intra+inter | 53.6 | 53.5 | 2.2 | 93.0 | 92.5 | 90.0 | 91.8 |
| CIL | 60.2 | 58.8 | 32.3 | 94.1 | 91.5 | 89.4 | 91.7 |
| HiCLRE | 61.8 | 62.8 | 34.7 | 85.0 | 84.5 | 83.7 | 84.4 |
| PARE | 62.7 | 60.7 | 36.1 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 34.7 | 85.0 | 83.5 | 84.3 | 84.3 |
| Last Student | 63.9 | 63.8 | 35.2 | 94.0 | 90.5 | 88.0 | 90.8 |
| Model | AUC | F1 | P@M |
| Last Student | 53.9 | 55.3 | 84.9 |
| a: - Un | 52.5 | 53.2 | 86.1 |
| b: - STS | 53.2 | 54.5 | 83.3 |
| c: - PW | 51.9 | 52.5 | 84.8 |
| d: - FNF | 53.3 | 54.9 | 82.5 |
| e: - TIW | 52.1 | 52.6 | 84.6 |
| f: Probe | 50.6 | 52.5 | 80.0 |
| g: JS div | 53.6 | 55.0 | 83.9 |
| Setting | AUC | F1 | P@M |
| l = 11 | 53.4 | 55.1 | 82.8 |
| l = 10 | 53.5 | 54.9 | 83.6 |
| l = 9 | 53.6 | 55.0 | 84.0 |
| l = 8 | 53.7 | 55.1 | 84.7 |
| l = 7 | 53.9 | 55.3 | 84.9 |
| l = 6 | 53.8 | 55.3 | 84.8 |
| l = 5 | 53.7 | 55.1 | 84.6 |
| l = 3 | 53.5 | 55.0 | 84.7 |
| l = 2 | 53.5 | 54.9 | 84.6 |
| l = 1 | 53.4 | 54.9 | 84.5 |
| Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | |
| 698.7 | 157.7 | 138.0 | 56.0 | 81 |
| Model | AUC | Macro-F1 |
| Intra+inter | 88.7 | 81.1 |
| CIL | 89.7 | 82.6 |
| HiCLRE | 87.9 | 80.3 |
| IRN (Song et al., 2023) | 90.9 | 82.5 |
| PARE | 91.4 | 83.9 |
| Teacher | 89.7 | 82.8 |
| Last Student | 90.9 | 84.1 |
| Class | Num. of items | Percentage (%) |
| BC | 3,044 | 78.07 |
| BI | 742 | 19.03 |
| TISC | 94 | 2.41 |
| TCSI | 19 | 0.49 |
| Sentence | Teacher | Student |
| Carl Friedrich von Weizsacker was born in Kiel, Germany, on June 28, 1912. | /people/person/place_of_birth | /people/person/place_lived |
| Presented by Brooklyn College and the office of Borough President Marty Markowitz. | /business/person/company | /people/person/place_lived |
| Furthermore, the relationship between the central government, dominated by three small Arab tribes living along the Nile, and Darfur's Arabs, who claim a heritage going back to the e Prophet Muhammad, is often antagonistic. | /people/person/ethnicity | /people/person/place_of_birth |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| CIL | 60.2 | 58.8 | 94.1 | 91.5 | 89.4 | 91.7 |
| PARE | 62.7 | 60.7 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 85.0 | 83.5 | 84.3 | 84.3 |
| Teacher† | 61.7 | 60.6 | 89.0 | 90.0 | 88.7 | 89.2 |
| Last Student | 63.9 | 63.8 | 94.0 | 90.5 | 88.0 | 90.8 |
| Last Student† | 63.4 | 60.7 | 96.0 | 96.0 | 95.3 | 95.8 |
| Dataset | Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | ||
| held-out | 522.6 | 18.4 | 172.4 | 2.0 | 53 |
| manual | 417.9 | 17.1 | 9.7 | 3.9 | 25 |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 33.8 | 40.7 | 75.0 | 72.5 | 65.7 | 71.1 |
| RESIDE | 41.5 | 45.7 | 84.0 | 78.5 | 75.6 | 79.4 |
| DISTRE | 42.2 | 48.6 | 68.0 | 67.0 | 65.3 | 66.8 |
| Intra+inter | 42.3 | 46.5 | 91.8 | 84.0 | 78.7 | 84.8 |
| CIL | 50.8 | 52.2 | 90.1 | 86.1 | 81.8 | 86.0 |
| HiCLRE | 45.3 | 50.5 | 82.0 | 78.5 | 74.0 | 78.2 |
| PARE | 53.4 | 54.4 | 85.0 | 85.0 | 82.7 | 84.8 |
| Teacher | 50.6 | 52.2 | 90.0 | 82.0 | 78.7 | 83.6 |
| Last Student | 53.9 | 55.3 | 88.0 | 84.0 | 82.7 | 84.9 |
| Model | AUC | Micro-F1 | Macro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 57.7 | 57.0 | 21.2 | 91.0 | 88.5 | 88.0 | 89.2 |
| Intra+inter | 53.6 | 53.5 | 2.2 | 93.0 | 92.5 | 90.0 | 91.8 |
| CIL | 60.2 | 58.8 | 32.3 | 94.1 | 91.5 | 89.4 | 91.7 |
| HiCLRE | 61.8 | 62.8 | 34.7 | 85.0 | 84.5 | 83.7 | 84.4 |
| PARE | 62.7 | 60.7 | 36.1 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 34.7 | 85.0 | 83.5 | 84.3 | 84.3 |
| Last Student | 63.9 | 63.8 | 35.2 | 94.0 | 90.5 | 88.0 | 90.8 |
| Model | AUC | F1 | P@M |
| Last Student | 53.9 | 55.3 | 84.9 |
| a: - Un | 52.5 | 53.2 | 86.1 |
| b: - STS | 53.2 | 54.5 | 83.3 |
| c: - PW | 51.9 | 52.5 | 84.8 |
| d: - FNF | 53.3 | 54.9 | 82.5 |
| e: - TIW | 52.1 | 52.6 | 84.6 |
| f: Probe | 50.6 | 52.5 | 80.0 |
| g: JS div | 53.6 | 55.0 | 83.9 |
| Setting | AUC | F1 | P@M |
| l = 11 | 53.4 | 55.1 | 82.8 |
| l = 10 | 53.5 | 54.9 | 83.6 |
| l = 9 | 53.6 | 55.0 | 84.0 |
| l = 8 | 53.7 | 55.1 | 84.7 |
| l = 7 | 53.9 | 55.3 | 84.9 |
| l = 6 | 53.8 | 55.3 | 84.8 |
| l = 5 | 53.7 | 55.1 | 84.6 |
| l = 3 | 53.5 | 55.0 | 84.7 |
| l = 2 | 53.5 | 54.9 | 84.6 |
| l = 1 | 53.4 | 54.9 | 84.5 |
| Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | |
| 698.7 | 157.7 | 138.0 | 56.0 | 81 |
| Model | AUC | Macro-F1 |
| Intra+inter | 88.7 | 81.1 |
| CIL | 89.7 | 82.6 |
| HiCLRE | 87.9 | 80.3 |
| IRN (Song et al., 2023) | 90.9 | 82.5 |
| PARE | 91.4 | 83.9 |
| Teacher | 89.7 | 82.8 |
| Last Student | 90.9 | 84.1 |
| Class | Num. of items | Percentage (%) |
| BC | 3,044 | 78.07 |
| BI | 742 | 19.03 |
| TISC | 94 | 2.41 |
| TCSI | 19 | 0.49 |
| Sentence | Teacher | Student |
| Carl Friedrich von Weizsacker was born in Kiel, Germany, on June 28, 1912. | /people/person/place_of_birth | /people/person/place_lived |
| Presented by Brooklyn College and the office of Borough President Marty Markowitz. | /business/person/company | /people/person/place_lived |
| Furthermore, the relationship between the central government, dominated by three small Arab tribes living along the Nile, and Darfur's Arabs, who claim a heritage going back to the e Prophet Muhammad, is often antagonistic. | /people/person/ethnicity | /people/person/place_of_birth |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| CIL | 60.2 | 58.8 | 94.1 | 91.5 | 89.4 | 91.7 |
| PARE | 62.7 | 60.7 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 85.0 | 83.5 | 84.3 | 84.3 |
| Teacher† | 61.7 | 60.6 | 89.0 | 90.0 | 88.7 | 89.2 |
| Last Student | 63.9 | 63.8 | 94.0 | 90.5 | 88.0 | 90.8 |
| Last Student† | 63.4 | 60.7 | 96.0 | 96.0 | 95.3 | 95.8 |
| Dataset | Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | ||
| held-out | 522.6 | 18.4 | 172.4 | 2.0 | 53 |
| manual | 417.9 | 17.1 | 9.7 | 3.9 | 25 |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 33.8 | 40.7 | 75.0 | 72.5 | 65.7 | 71.1 |
| RESIDE | 41.5 | 45.7 | 84.0 | 78.5 | 75.6 | 79.4 |
| DISTRE | 42.2 | 48.6 | 68.0 | 67.0 | 65.3 | 66.8 |
| Intra+inter | 42.3 | 46.5 | 91.8 | 84.0 | 78.7 | 84.8 |
| CIL | 50.8 | 52.2 | 90.1 | 86.1 | 81.8 | 86.0 |
| HiCLRE | 45.3 | 50.5 | 82.0 | 78.5 | 74.0 | 78.2 |
| PARE | 53.4 | 54.4 | 85.0 | 85.0 | 82.7 | 84.8 |
| Teacher | 50.6 | 52.2 | 90.0 | 82.0 | 78.7 | 83.6 |
| Last Student | 53.9 | 55.3 | 88.0 | 84.0 | 82.7 | 84.9 |
| Model | AUC | Micro-F1 | Macro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 57.7 | 57.0 | 21.2 | 91.0 | 88.5 | 88.0 | 89.2 |
| Intra+inter | 53.6 | 53.5 | 2.2 | 93.0 | 92.5 | 90.0 | 91.8 |
| CIL | 60.2 | 58.8 | 32.3 | 94.1 | 91.5 | 89.4 | 91.7 |
| HiCLRE | 61.8 | 62.8 | 34.7 | 85.0 | 84.5 | 83.7 | 84.4 |
| PARE | 62.7 | 60.7 | 36.1 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 34.7 | 85.0 | 83.5 | 84.3 | 84.3 |
| Last Student | 63.9 | 63.8 | 35.2 | 94.0 | 90.5 | 88.0 | 90.8 |
| Model | AUC | F1 | P@M |
| Last Student | 53.9 | 55.3 | 84.9 |
| a: - Un | 52.5 | 53.2 | 86.1 |
| b: - STS | 53.2 | 54.5 | 83.3 |
| c: - PW | 51.9 | 52.5 | 84.8 |
| d: - FNF | 53.3 | 54.9 | 82.5 |
| e: - TIW | 52.1 | 52.6 | 84.6 |
| f: Probe | 50.6 | 52.5 | 80.0 |
| g: JS div | 53.6 | 55.0 | 83.9 |
| Setting | AUC | F1 | P@M |
| l = 11 | 53.4 | 55.1 | 82.8 |
| l = 10 | 53.5 | 54.9 | 83.6 |
| l = 9 | 53.6 | 55.0 | 84.0 |
| l = 8 | 53.7 | 55.1 | 84.7 |
| l = 7 | 53.9 | 55.3 | 84.9 |
| l = 6 | 53.8 | 55.3 | 84.8 |
| l = 5 | 53.7 | 55.1 | 84.6 |
| l = 3 | 53.5 | 55.0 | 84.7 |
| l = 2 | 53.5 | 54.9 | 84.6 |
| l = 1 | 53.4 | 54.9 | 84.5 |
| Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | |
| 698.7 | 157.7 | 138.0 | 56.0 | 81 |
| Model | AUC | Macro-F1 |
| Intra+inter | 88.7 | 81.1 |
| CIL | 89.7 | 82.6 |
| HiCLRE | 87.9 | 80.3 |
| IRN (Song et al., 2023) | 90.9 | 82.5 |
| PARE | 91.4 | 83.9 |
| Teacher | 89.7 | 82.8 |
| Last Student | 90.9 | 84.1 |
| Class | Num. of items | Percentage (%) |
| BC | 3,044 | 78.07 |
| BI | 742 | 19.03 |
| TISC | 94 | 2.41 |
| TCSI | 19 | 0.49 |
| Sentence | Teacher | Student |
| Carl Friedrich von Weizsacker was born in Kiel, Germany, on June 28, 1912. | /people/person/place_of_birth | /people/person/place_lived |
| Presented by Brooklyn College and the office of Borough President Marty Markowitz. | /business/person/company | /people/person/place_lived |
| Furthermore, the relationship between the central government, dominated by three small Arab tribes living along the Nile, and Darfur's Arabs, who claim a heritage going back to the e Prophet Muhammad, is often antagonistic. | /people/person/ethnicity | /people/person/place_of_birth |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| CIL | 60.2 | 58.8 | 94.1 | 91.5 | 89.4 | 91.7 |
| PARE | 62.7 | 60.7 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 85.0 | 83.5 | 84.3 | 84.3 |
| Teacher† | 61.7 | 60.6 | 89.0 | 90.0 | 88.7 | 89.2 |
| Last Student | 63.9 | 63.8 | 94.0 | 90.5 | 88.0 | 90.8 |
| Last Student† | 63.4 | 60.7 | 96.0 | 96.0 | 95.3 | 95.8 |
| Dataset | Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | ||
| held-out | 522.6 | 18.4 | 172.4 | 2.0 | 53 |
| manual | 417.9 | 17.1 | 9.7 | 3.9 | 25 |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 33.8 | 40.7 | 75.0 | 72.5 | 65.7 | 71.1 |
| RESIDE | 41.5 | 45.7 | 84.0 | 78.5 | 75.6 | 79.4 |
| DISTRE | 42.2 | 48.6 | 68.0 | 67.0 | 65.3 | 66.8 |
| Intra+inter | 42.3 | 46.5 | 91.8 | 84.0 | 78.7 | 84.8 |
| CIL | 50.8 | 52.2 | 90.1 | 86.1 | 81.8 | 86.0 |
| HiCLRE | 45.3 | 50.5 | 82.0 | 78.5 | 74.0 | 78.2 |
| PARE | 53.4 | 54.4 | 85.0 | 85.0 | 82.7 | 84.8 |
| Teacher | 50.6 | 52.2 | 90.0 | 82.0 | 78.7 | 83.6 |
| Last Student | 53.9 | 55.3 | 88.0 | 84.0 | 82.7 | 84.9 |
| Model | AUC | Micro-F1 | Macro-F1 | P@100 | P@200 | P@300 | P@M |
| PCNN+ATT | 57.7 | 57.0 | 21.2 | 91.0 | 88.5 | 88.0 | 89.2 |
| Intra+inter | 53.6 | 53.5 | 2.2 | 93.0 | 92.5 | 90.0 | 91.8 |
| CIL | 60.2 | 58.8 | 32.3 | 94.1 | 91.5 | 89.4 | 91.7 |
| HiCLRE | 61.8 | 62.8 | 34.7 | 85.0 | 84.5 | 83.7 | 84.4 |
| PARE | 62.7 | 60.7 | 36.1 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 34.7 | 85.0 | 83.5 | 84.3 | 84.3 |
| Last Student | 63.9 | 63.8 | 35.2 | 94.0 | 90.5 | 88.0 | 90.8 |
| Model | AUC | F1 | P@M |
| Last Student | 53.9 | 55.3 | 84.9 |
| a: - Un | 52.5 | 53.2 | 86.1 |
| b: - STS | 53.2 | 54.5 | 83.3 |
| c: - PW | 51.9 | 52.5 | 84.8 |
| d: - FNF | 53.3 | 54.9 | 82.5 |
| e: - TIW | 52.1 | 52.6 | 84.6 |
| f: Probe | 50.6 | 52.5 | 80.0 |
| g: JS div | 53.6 | 55.0 | 83.9 |
| Setting | AUC | F1 | P@M |
| l = 11 | 53.4 | 55.1 | 82.8 |
| l = 10 | 53.5 | 54.9 | 83.6 |
| l = 9 | 53.6 | 55.0 | 84.0 |
| l = 8 | 53.7 | 55.1 | 84.7 |
| l = 7 | 53.9 | 55.3 | 84.9 |
| l = 6 | 53.8 | 55.3 | 84.8 |
| l = 5 | 53.7 | 55.1 | 84.6 |
| l = 3 | 53.5 | 55.0 | 84.7 |
| l = 2 | 53.5 | 54.9 | 84.6 |
| l = 1 | 53.4 | 54.9 | 84.5 |
| Train (k) | Test (k) | Rel. | ||
| Sen. | Fac. | Sen. | Fac. | |
| 698.7 | 157.7 | 138.0 | 56.0 | 81 |
| Model | AUC | Macro-F1 |
| Intra+inter | 88.7 | 81.1 |
| CIL | 89.7 | 82.6 |
| HiCLRE | 87.9 | 80.3 |
| IRN (Song et al., 2023) | 90.9 | 82.5 |
| PARE | 91.4 | 83.9 |
| Teacher | 89.7 | 82.8 |
| Last Student | 90.9 | 84.1 |
| Class | Num. of items | Percentage (%) |
| BC | 3,044 | 78.07 |
| BI | 742 | 19.03 |
| TISC | 94 | 2.41 |
| TCSI | 19 | 0.49 |
| Sentence | Teacher | Student |
| Carl Friedrich von Weizsacker was born in Kiel, Germany, on June 28, 1912. | /people/person/place_of_birth | /people/person/place_lived |
| Presented by Brooklyn College and the office of Borough President Marty Markowitz. | /business/person/company | /people/person/place_lived |
| Furthermore, the relationship between the central government, dominated by three small Arab tribes living along the Nile, and Darfur's Arabs, who claim a heritage going back to the e Prophet Muhammad, is often antagonistic. | /people/person/ethnicity | /people/person/place_of_birth |
| Model | AUC | Micro-F1 | P@100 | P@200 | P@300 | P@M |
| CIL | 60.2 | 58.8 | 94.1 | 91.5 | 89.4 | 91.7 |
| PARE | 62.7 | 60.7 | 97.0 | 95.0 | 95.0 | 95.7 |
| Teacher | 61.3 | 62.4 | 85.0 | 83.5 | 84.3 | 84.3 |
| Teacher† | 61.7 | 60.6 | 89.0 | 90.0 | 88.7 | 89.2 |
| Last Student | 63.9 | 63.8 | 94.0 | 90.5 | 88.0 | 90.8 |
| Last Student† | 63.4 | 60.7 | 96.0 | 96.0 | 95.3 | 95.8 |
| Methods | MSQ | HotpotQA | 2Wiki | CWebQ | Average | |||||
| EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | |
| Fine-tuning methods with extra corpus | ||||||||||
| DPR† (Karpukhin et al., 2020) | 7.5 | 16.6 | 14.7 | 23.0 | 6.5 | 14.6 | 21.3 | 30.2 | 12.5 | 21.1 |
| RAG† (Lewis et al., 2020) | 3.9 | 8.2 | 9.6 | 15.5 | 15.7 | 18.6 | 13.3 | 15.8 | 10.6 | 14.5 |
| REALM† (Guu et al., 2020) | 3.7 | 8.5 | 15.3 | 22.0 | 5.1 | 9.6 | 21.1 | 27.3 | 11.3 | 16.9 |
| T5-11B-SSM (Roberts et al., 2020) | 10.6 | 16.6 | 15.4 | 22.4 | 15.6 | 20.5 | 28.5 | 35.5 | 17.5 | 23.8 |
| Retrieval-based methods with LLMs | ||||||||||
| DPR† + ChatGPT | 1.7 | 4.1 | 15.8 | 21.4 | 10.9 | 18.1 | 13.7 | 18.7 | 10.5 | 15.6 |
| DPR† + Alpaca-13B (Taori et al., 2023) | 2.2 | 8.4 | 12.0 | 21.5 | 12.6 | 21.2 | 15.9 | 27.1 | 10.7 | 19.5 |
| DPR† + Vicuna-13B (Chiang et al., 2023) | 2.2 | 7.4 | 18.6 | 25.8 | 23.9 | 27.6 | 20.3 | 27.9 | 16.2 | 22.1 |
| DPR† + WizardLM-13B (Xu et al., 2023) | 3.5 | 10.0 | 19.9 | 28.4 | 22.8 | 27.5 | 24.2 | 32.2 | 17.6 | 24.5 |
| DPR† + InstructGPT (Ouyang et al., 2022) | 4.8 | 11.6 | 26.3 | 34.8 | 23.3 | 27.1 | 34.4 | 41.6 | 22.2 | 28.8 |
| LLM-only methods on ChatGPT | ||||||||||
| Zero-Shot | 3.1 | 7.3 | 22.4 | 30.0 | 18.7 | 21.7 | 31.6 | 37.5 | 19.0 | 24.1 |
| Self-Prompting (Li et al., 2022) | 2.9 | 6.2 | 23.8 | 31.2 | 18.9 | 23.5 | 26.8 | 32.6 | 18.1 | 23.4 |
| GENREAD (Yu et al., 2023) | 8.6 | 14.6 | 33.2 | 42.6 | 30.4 | 35.3 | 33.7 | 40.1 | 26.5 | 33.2 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0 | 8.8 | 22.6 | 29.6 | 24.3 | 27.1 | 30.3 | 36.2 | 20.6 | 25.4 |
| Auto-CoT (Zhang et al., 2022) | 8.1 | 13.6 | 26.1 | 36.3 | 26.2 | 30.2 | 29.9 | 38.4 | 22.6 | 29.6 |
| Manual-CoT (random) (Wei et al., 2023) | 12.3 | 19.2 | 32.4 | 43.7 | 27.7 | 34.6 | 36.6 | 43.0 | 27.3 | 35.1 |
| SP-CoT (Ours) | 14.5 | 22.6 | 33.2 | 42.9 | 30.1 | 34.7 | 37.5 | 43.6 | 28.8 | 36.0 |
| Model | Size | Method | MSQ | HotpotQA | 2Wiki | CWebQ | Mean | Boost |
| Alpaca (Taori et al., 2023) | 13B | Zero-shot | 0.9 | 9.5 | 12.5 | 12.6 | 8.9 | - |
| Alpaca (Taori et al., 2023) | 13B | SP-CoT | 5.3 | 19.8 | 17.8 | 26.5 | 17.4 | 8.5↑ |
| Vicuna (Chiang et al., 2023) | 13B | Zero-shot | 2.3 | 11.5 | 18.1 | 18.3 | 12.6 | - |
| Vicuna (Chiang et al., 2023) | 13B | SP-CoT | 8.5 | 23.3 | 22.0 | 32.2 | 21.5 | 8.9↑ |
| WizardLM (Xu et al., 2023) | 13B | Zero-shot | 2.0 | 10.3 | 13.6 | 17.5 | 10.9 | - |
| WizardLM (Xu et al., 2023) | 13B | SP-CoT | 7.3 | 24.0 | 27.2 | 31.3 | 22.5 | 11.6↑ |
| InstructGPT (Ouyang et al., 2022) | 175B | Zero-shot | 4.4 | 25.6 | 22.8 | 38.6 | 22.9 | - |
| InstructGPT (Ouyang et al., 2022) | 175B | SP-CoT | 14.8 | 37.4 | 33.7 | 46.6 | 33.1 | 10.2↑ |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Random | 12.0±0.8 | 29.5±0.6 | 25.6±1.2 | 34.3±0.9 | 25.4±0.3 |
| Retrieve | 10.5 | 27.9 | 24.3 | 33.5 | 24.1 |
| ClusterCenter | 10.4 | 26.2 | 22.8 | 33.0 | 23.1 |
| RetrieveInTypeCluster | 11.6 | 28.7 | 23.5 | 35.3 | 24.8 |
| RetrieveInCluster | 11.5 | 30.9 | 27.8 | 34.0 | 26.1 |
| MSQ | HotpotQA | CWebQ | 2Wiki | |
| Q len. | 18.11 | 15.83 | 13.37 | 11.98 |
| A len. | 2.8 | 2.46 | 2.42 | 2.41 |
| Steps | 2.65 | 2.68 | - | 2.47 |
| Size | 2417 | 7405 | 3519 | 12576 |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Zero-shot | 3.1/7.3 | 22.4/30.0 | 18.7/21.7 | 31.6/37.5 | 19.0/24.1 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0/8.8 | 22.6/29.6 | 24.3/27.1 | 30.3/36.2 | 20.6/25.4 |
| Manual-CoT (Wei et al., 2023) (Random) | 12.3/19.2 | 32.4/43.7 | 27.7/34.6 | 36.6/43.0 | 27.3/35.1 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 13.7/21.9 | 33.9/44.7 | 31.5/38.6 | 37.2/44.0 | 29.1/37.3 |
| Auto-CoT (Zhang et al., 2022) | 8.1/13.6 | 26.1/36.3 | 26.2/30.2 | 29.9/38.3 | 22.6/29.6 |
| SP-CoT (Ours) | 14.5/22.6 | 33.2/42.9 | 30.1/34.7 | 37.5/43.6 | 28.8/36.0 |
| SP-CoT + Self-Consistency (Wang et al., 2023) | 18.3/28.3 | -/- | -/- | 47.1/54.0 | -/- |
| Method | ChatGPT | InstructGPT | Alpaca-13B | Vicuna-13B | Wizard-13B |
| Zero-shot | 20.8 | 22.9 | 8.9 | 12.6 | 10.9 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 28.7 | 33.1 | 18.0 | 23.7 | 24.7 |
| Manual-CoT (Wei et al., 2023) (Random) | 26.3 | 32.4 | 17.7 | 22.0 | 24.0 |
| Auto-CoT (Zhang et al., 2022) | 21.2 | 31.4 | 16.9 | 21.3 | 23.3 |
| SP-CoT (Ours) | 26.1 | 33.1 | 17.4 | 21.5 | 22.5 |
| ChatGPT | Alpaca-13B | Vicuna-13B | WizardLM-13B | |
| Number of Samples | 3550 | 1354 | 1562 | 1604 |
| Avg. Number of Hops | 3.04 | 2.82 | 2.92 | 3.16 |
| Avg. Question tokens | 34.61 | 25.35 | 31.05 | 34.30 |
| Avg. Answer tokens | 1.86 | 1.99 | 2.02 | 1.91 |
| Method | EM | F1 |
| Zero-shot | 12.9 | 24.3 |
| Zero-shot-CoT (Kojima et al., 2023) | 13.5 | 25.2 |
| Manual-CoT (Wei et al., 2023) (Random) | 17.7 | 30.7 |
| SP-CoT (Ours) | 16.0 | 28.2 |
| Graph | QA | Decomposition |
| ○→○ | Q: What is another common name for some members of the genus to which daffodils belong, besides daffodil? \nA: Jonquil | 1. Q: What is the genus to which daffodils belong? \nA: Narcissus \n2. Q: What is another common name for some members of the Narcissus genus besides daffodil? \nA: Jonquil |
| ○→○→○ | Q: Who held the titles of King of the French and The Duke of Orleans during the year when John Lindley first described the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Louis Philippe | 1. Q: What is the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Scilloideae \n2. Q: In what year was the subfamily Scilloideae first described by John Lindley? \nA: 1830 \n3. Q: Who became King of the French in 1830 and also held the title of The Duke of Orleans? \nA: Louis Philippe |
| ○ | Q: Which country, commonly visited by ships traveling from Europe and where most species of roses are native to, is located south of the remote island where Napoleon Bonaparte was exiled and eventually died? \nA: South Africa | 1. Q: Where are most species of roses native to? \nA: Asia \n2. Q: To which remote island was Napoleon Bonaparte exiled and where did he eventually die? \nA: Saint Helena \n3. Q: What country is located south of Saint Helena and was a common destination for ships traveling from Asia and Europe? \nA: South Africa |
| ○→○→○ | Q: What work did the person who shared credit with the father of the biblical figure referenced in the title of the song "Stairway to Heaven" defend his optimism in, while also famously contributing to the development of infinitesimal calculus with Newton? \nA: Theodicy | 1. Q: Who is the biblical figure that the title of the song "Stairway to Heaven" references? \nA: Jacob \n2. Q: Who was the father of Jacob? \nA: Isaac \n3. Q: Who shares credit with Isaac Newton for developing the infinitesimal calculus? \nA: Gottfried Wilhelm Leibniz \n4. Q: What work did Gottfried Wilhelm Leibniz famously defend his optimism in? \nA: Theodicy |
| ○→○ | Q: What body of water separates the country that is a key ally of the country that Iraq launched Scud missiles at during the Gulf War, from Canada? \nA: the Atlantic Ocean | 1. Q: What country did Iraq launch Scud missiles at during the Gulf War, and why is this relevant to the conflict? \nA: Israel \n2. Q: What is the native region of Carrot? \nA: the Middle East \n3. Q: What is the name of the country that is a key ally of Israel in the Middle East? \nA: the United States \n4. Q: What body of water separates the United States from Canada? \nA: the Atlantic Ocean |
| ○→○ | Q: What is the name of the town in the country that is the largest in South America after the country where Lionel Messi represents in international football competitions where Tom Brady's wife was born? \nA: Très de Maio | 1. Q: Which country does Lionel Messi represent in international football competitions? \nA: Argentina \n2. Q: What is the largest country in South America after Argentina? \nA: Brazil \n3. Q: What is the name of the model that Tom Brady is married to? \nA: Gisele Bündchen \n4. Q: What is the name of the town in Brazil where Gisele Bündchen was born? \nA: Très de Maio |
| Methods | MSQ | HotpotQA | 2Wiki | CWebQ | Average | |||||
| EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | |
| Fine-tuning methods with extra corpus | ||||||||||
| DPR† (Karpukhin et al., 2020) | 7.5 | 16.6 | 14.7 | 23.0 | 6.5 | 14.6 | 21.3 | 30.2 | 12.5 | 21.1 |
| RAG† (Lewis et al., 2020) | 3.9 | 8.2 | 9.6 | 15.5 | 15.7 | 18.6 | 13.3 | 15.8 | 10.6 | 14.5 |
| REALM† (Guu et al., 2020) | 3.7 | 8.5 | 15.3 | 22.0 | 5.1 | 9.6 | 21.1 | 27.3 | 11.3 | 16.9 |
| T5-11B-SSM (Roberts et al., 2020) | 10.6 | 16.6 | 15.4 | 22.4 | 15.6 | 20.5 | 28.5 | 35.5 | 17.5 | 23.8 |
| Retrieval-based methods with LLMs | ||||||||||
| DPR† + ChatGPT | 1.7 | 4.1 | 15.8 | 21.4 | 10.9 | 18.1 | 13.7 | 18.7 | 10.5 | 15.6 |
| DPR† + Alpaca-13B (Taori et al., 2023) | 2.2 | 8.4 | 12.0 | 21.5 | 12.6 | 21.2 | 15.9 | 27.1 | 10.7 | 19.5 |
| DPR† + Vicuna-13B (Chiang et al., 2023) | 2.2 | 7.4 | 18.6 | 25.8 | 23.9 | 27.6 | 20.3 | 27.9 | 16.2 | 22.1 |
| DPR† + WizardLM-13B (Xu et al., 2023) | 3.5 | 10.0 | 19.9 | 28.4 | 22.8 | 27.5 | 24.2 | 32.2 | 17.6 | 24.5 |
| DPR† + InstructGPT (Ouyang et al., 2022) | 4.8 | 11.6 | 26.3 | 34.8 | 23.3 | 27.1 | 34.4 | 41.6 | 22.2 | 28.8 |
| LLM-only methods on ChatGPT | ||||||||||
| Zero-Shot | 3.1 | 7.3 | 22.4 | 30.0 | 18.7 | 21.7 | 31.6 | 37.5 | 19.0 | 24.1 |
| Self-Prompting (Li et al., 2022) | 2.9 | 6.2 | 23.8 | 31.2 | 18.9 | 23.5 | 26.8 | 32.6 | 18.1 | 23.4 |
| GENREAD (Yu et al., 2023) | 8.6 | 14.6 | 33.2 | 42.6 | 30.4 | 35.3 | 33.7 | 40.1 | 26.5 | 33.2 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0 | 8.8 | 22.6 | 29.6 | 24.3 | 27.1 | 30.3 | 36.2 | 20.6 | 25.4 |
| Auto-CoT (Zhang et al., 2022) | 8.1 | 13.6 | 26.1 | 36.3 | 26.2 | 30.2 | 29.9 | 38.4 | 22.6 | 29.6 |
| Manual-CoT (random) (Wei et al., 2023) | 12.3 | 19.2 | 32.4 | 43.7 | 27.7 | 34.6 | 36.6 | 43.0 | 27.3 | 35.1 |
| SP-CoT (Ours) | 14.5 | 22.6 | 33.2 | 42.9 | 30.1 | 34.7 | 37.5 | 43.6 | 28.8 | 36.0 |
| Model | Size | Method | MSQ | HotpotQA | 2Wiki | CWebQ | Mean | Boost |
| Alpaca (Taori et al., 2023) | 13B | Zero-shot | 0.9 | 9.5 | 12.5 | 12.6 | 8.9 | - |
| Alpaca (Taori et al., 2023) | 13B | SP-CoT | 5.3 | 19.8 | 17.8 | 26.5 | 17.4 | 8.5↑ |
| Vicuna (Chiang et al., 2023) | 13B | Zero-shot | 2.3 | 11.5 | 18.1 | 18.3 | 12.6 | - |
| Vicuna (Chiang et al., 2023) | 13B | SP-CoT | 8.5 | 23.3 | 22.0 | 32.2 | 21.5 | 8.9↑ |
| WizardLM (Xu et al., 2023) | 13B | Zero-shot | 2.0 | 10.3 | 13.6 | 17.5 | 10.9 | - |
| WizardLM (Xu et al., 2023) | 13B | SP-CoT | 7.3 | 24.0 | 27.2 | 31.3 | 22.5 | 11.6↑ |
| InstructGPT (Ouyang et al., 2022) | 175B | Zero-shot | 4.4 | 25.6 | 22.8 | 38.6 | 22.9 | - |
| InstructGPT (Ouyang et al., 2022) | 175B | SP-CoT | 14.8 | 37.4 | 33.7 | 46.6 | 33.1 | 10.2↑ |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Random | 12.0±0.8 | 29.5±0.6 | 25.6±1.2 | 34.3±0.9 | 25.4±0.3 |
| Retrieve | 10.5 | 27.9 | 24.3 | 33.5 | 24.1 |
| ClusterCenter | 10.4 | 26.2 | 22.8 | 33.0 | 23.1 |
| RetrieveInTypeCluster | 11.6 | 28.7 | 23.5 | 35.3 | 24.8 |
| RetrieveInCluster | 11.5 | 30.9 | 27.8 | 34.0 | 26.1 |
| MSQ | HotpotQA | CWebQ | 2Wiki | |
| Q len. | 18.11 | 15.83 | 13.37 | 11.98 |
| A len. | 2.8 | 2.46 | 2.42 | 2.41 |
| Steps | 2.65 | 2.68 | - | 2.47 |
| Size | 2417 | 7405 | 3519 | 12576 |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Zero-shot | 3.1/7.3 | 22.4/30.0 | 18.7/21.7 | 31.6/37.5 | 19.0/24.1 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0/8.8 | 22.6/29.6 | 24.3/27.1 | 30.3/36.2 | 20.6/25.4 |
| Manual-CoT (Wei et al., 2023) (Random) | 12.3/19.2 | 32.4/43.7 | 27.7/34.6 | 36.6/43.0 | 27.3/35.1 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 13.7/21.9 | 33.9/44.7 | 31.5/38.6 | 37.2/44.0 | 29.1/37.3 |
| Auto-CoT (Zhang et al., 2022) | 8.1/13.6 | 26.1/36.3 | 26.2/30.2 | 29.9/38.3 | 22.6/29.6 |
| SP-CoT (Ours) | 14.5/22.6 | 33.2/42.9 | 30.1/34.7 | 37.5/43.6 | 28.8/36.0 |
| SP-CoT + Self-Consistency (Wang et al., 2023) | 18.3/28.3 | -/- | -/- | 47.1/54.0 | -/- |
| Method | ChatGPT | InstructGPT | Alpaca-13B | Vicuna-13B | Wizard-13B |
| Zero-shot | 20.8 | 22.9 | 8.9 | 12.6 | 10.9 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 28.7 | 33.1 | 18.0 | 23.7 | 24.7 |
| Manual-CoT (Wei et al., 2023) (Random) | 26.3 | 32.4 | 17.7 | 22.0 | 24.0 |
| Auto-CoT (Zhang et al., 2022) | 21.2 | 31.4 | 16.9 | 21.3 | 23.3 |
| SP-CoT (Ours) | 26.1 | 33.1 | 17.4 | 21.5 | 22.5 |
| ChatGPT | Alpaca-13B | Vicuna-13B | WizardLM-13B | |
| Number of Samples | 3550 | 1354 | 1562 | 1604 |
| Avg. Number of Hops | 3.04 | 2.82 | 2.92 | 3.16 |
| Avg. Question tokens | 34.61 | 25.35 | 31.05 | 34.30 |
| Avg. Answer tokens | 1.86 | 1.99 | 2.02 | 1.91 |
| Method | EM | F1 |
| Zero-shot | 12.9 | 24.3 |
| Zero-shot-CoT (Kojima et al., 2023) | 13.5 | 25.2 |
| Manual-CoT (Wei et al., 2023) (Random) | 17.7 | 30.7 |
| SP-CoT (Ours) | 16.0 | 28.2 |
| Graph | QA | Decomposition |
| ○→○ | Q: What is another common name for some members of the genus to which daffodils belong, besides daffodil? \nA: Jonquil | 1. Q: What is the genus to which daffodils belong? \nA: Narcissus \n2. Q: What is another common name for some members of the Narcissus genus besides daffodil? \nA: Jonquil |
| ○→○→○ | Q: Who held the titles of King of the French and The Duke of Orleans during the year when John Lindley first described the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Louis Philippe | 1. Q: What is the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Scilloideae \n2. Q: In what year was the subfamily Scilloideae first described by John Lindley? \nA: 1830 \n3. Q: Who became King of the French in 1830 and also held the title of The Duke of Orleans? \nA: Louis Philippe |
| ○ | Q: Which country, commonly visited by ships traveling from Europe and where most species of roses are native to, is located south of the remote island where Napoleon Bonaparte was exiled and eventually died? \nA: South Africa | 1. Q: Where are most species of roses native to? \nA: Asia \n2. Q: To which remote island was Napoleon Bonaparte exiled and where did he eventually die? \nA: Saint Helena \n3. Q: What country is located south of Saint Helena and was a common destination for ships traveling from Asia and Europe? \nA: South Africa |
| ○→○→○ | Q: What work did the person who shared credit with the father of the biblical figure referenced in the title of the song "Stairway to Heaven" defend his optimism in, while also famously contributing to the development of infinitesimal calculus with Newton? \nA: Theodicy | 1. Q: Who is the biblical figure that the title of the song "Stairway to Heaven" references? \nA: Jacob \n2. Q: Who was the father of Jacob? \nA: Isaac \n3. Q: Who shares credit with Isaac Newton for developing the infinitesimal calculus? \nA: Gottfried Wilhelm Leibniz \n4. Q: What work did Gottfried Wilhelm Leibniz famously defend his optimism in? \nA: Theodicy |
| ○→○ | Q: What body of water separates the country that is a key ally of the country that Iraq launched Scud missiles at during the Gulf War, from Canada? \nA: the Atlantic Ocean | 1. Q: What country did Iraq launch Scud missiles at during the Gulf War, and why is this relevant to the conflict? \nA: Israel \n2. Q: What is the native region of Carrot? \nA: the Middle East \n3. Q: What is the name of the country that is a key ally of Israel in the Middle East? \nA: the United States \n4. Q: What body of water separates the United States from Canada? \nA: the Atlantic Ocean |
| ○→○ | Q: What is the name of the town in the country that is the largest in South America after the country where Lionel Messi represents in international football competitions where Tom Brady's wife was born? \nA: Très de Maio | 1. Q: Which country does Lionel Messi represent in international football competitions? \nA: Argentina \n2. Q: What is the largest country in South America after Argentina? \nA: Brazil \n3. Q: What is the name of the model that Tom Brady is married to? \nA: Gisele Bündchen \n4. Q: What is the name of the town in Brazil where Gisele Bündchen was born? \nA: Très de Maio |
| Methods | MSQ | HotpotQA | 2Wiki | CWebQ | Average | |||||
| EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | |
| Fine-tuning methods with extra corpus | ||||||||||
| DPR† (Karpukhin et al., 2020) | 7.5 | 16.6 | 14.7 | 23.0 | 6.5 | 14.6 | 21.3 | 30.2 | 12.5 | 21.1 |
| RAG† (Lewis et al., 2020) | 3.9 | 8.2 | 9.6 | 15.5 | 15.7 | 18.6 | 13.3 | 15.8 | 10.6 | 14.5 |
| REALM† (Guu et al., 2020) | 3.7 | 8.5 | 15.3 | 22.0 | 5.1 | 9.6 | 21.1 | 27.3 | 11.3 | 16.9 |
| T5-11B-SSM (Roberts et al., 2020) | 10.6 | 16.6 | 15.4 | 22.4 | 15.6 | 20.5 | 28.5 | 35.5 | 17.5 | 23.8 |
| Retrieval-based methods with LLMs | ||||||||||
| DPR† + ChatGPT | 1.7 | 4.1 | 15.8 | 21.4 | 10.9 | 18.1 | 13.7 | 18.7 | 10.5 | 15.6 |
| DPR† + Alpaca-13B (Taori et al., 2023) | 2.2 | 8.4 | 12.0 | 21.5 | 12.6 | 21.2 | 15.9 | 27.1 | 10.7 | 19.5 |
| DPR† + Vicuna-13B (Chiang et al., 2023) | 2.2 | 7.4 | 18.6 | 25.8 | 23.9 | 27.6 | 20.3 | 27.9 | 16.2 | 22.1 |
| DPR† + WizardLM-13B (Xu et al., 2023) | 3.5 | 10.0 | 19.9 | 28.4 | 22.8 | 27.5 | 24.2 | 32.2 | 17.6 | 24.5 |
| DPR† + InstructGPT (Ouyang et al., 2022) | 4.8 | 11.6 | 26.3 | 34.8 | 23.3 | 27.1 | 34.4 | 41.6 | 22.2 | 28.8 |
| LLM-only methods on ChatGPT | ||||||||||
| Zero-Shot | 3.1 | 7.3 | 22.4 | 30.0 | 18.7 | 21.7 | 31.6 | 37.5 | 19.0 | 24.1 |
| Self-Prompting (Li et al., 2022) | 2.9 | 6.2 | 23.8 | 31.2 | 18.9 | 23.5 | 26.8 | 32.6 | 18.1 | 23.4 |
| GENREAD (Yu et al., 2023) | 8.6 | 14.6 | 33.2 | 42.6 | 30.4 | 35.3 | 33.7 | 40.1 | 26.5 | 33.2 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0 | 8.8 | 22.6 | 29.6 | 24.3 | 27.1 | 30.3 | 36.2 | 20.6 | 25.4 |
| Auto-CoT (Zhang et al., 2022) | 8.1 | 13.6 | 26.1 | 36.3 | 26.2 | 30.2 | 29.9 | 38.4 | 22.6 | 29.6 |
| Manual-CoT (random) (Wei et al., 2023) | 12.3 | 19.2 | 32.4 | 43.7 | 27.7 | 34.6 | 36.6 | 43.0 | 27.3 | 35.1 |
| SP-CoT (Ours) | 14.5 | 22.6 | 33.2 | 42.9 | 30.1 | 34.7 | 37.5 | 43.6 | 28.8 | 36.0 |
| Model | Size | Method | MSQ | HotpotQA | 2Wiki | CWebQ | Mean | Boost |
| Alpaca (Taori et al., 2023) | 13B | Zero-shot | 0.9 | 9.5 | 12.5 | 12.6 | 8.9 | - |
| Alpaca (Taori et al., 2023) | 13B | SP-CoT | 5.3 | 19.8 | 17.8 | 26.5 | 17.4 | 8.5↑ |
| Vicuna (Chiang et al., 2023) | 13B | Zero-shot | 2.3 | 11.5 | 18.1 | 18.3 | 12.6 | - |
| Vicuna (Chiang et al., 2023) | 13B | SP-CoT | 8.5 | 23.3 | 22.0 | 32.2 | 21.5 | 8.9↑ |
| WizardLM (Xu et al., 2023) | 13B | Zero-shot | 2.0 | 10.3 | 13.6 | 17.5 | 10.9 | - |
| WizardLM (Xu et al., 2023) | 13B | SP-CoT | 7.3 | 24.0 | 27.2 | 31.3 | 22.5 | 11.6↑ |
| InstructGPT (Ouyang et al., 2022) | 175B | Zero-shot | 4.4 | 25.6 | 22.8 | 38.6 | 22.9 | - |
| InstructGPT (Ouyang et al., 2022) | 175B | SP-CoT | 14.8 | 37.4 | 33.7 | 46.6 | 33.1 | 10.2↑ |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Random | 12.0±0.8 | 29.5±0.6 | 25.6±1.2 | 34.3±0.9 | 25.4±0.3 |
| Retrieve | 10.5 | 27.9 | 24.3 | 33.5 | 24.1 |
| ClusterCenter | 10.4 | 26.2 | 22.8 | 33.0 | 23.1 |
| RetrieveInTypeCluster | 11.6 | 28.7 | 23.5 | 35.3 | 24.8 |
| RetrieveInCluster | 11.5 | 30.9 | 27.8 | 34.0 | 26.1 |
| MSQ | HotpotQA | CWebQ | 2Wiki | |
| Q len. | 18.11 | 15.83 | 13.37 | 11.98 |
| A len. | 2.8 | 2.46 | 2.42 | 2.41 |
| Steps | 2.65 | 2.68 | - | 2.47 |
| Size | 2417 | 7405 | 3519 | 12576 |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Zero-shot | 3.1/7.3 | 22.4/30.0 | 18.7/21.7 | 31.6/37.5 | 19.0/24.1 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0/8.8 | 22.6/29.6 | 24.3/27.1 | 30.3/36.2 | 20.6/25.4 |
| Manual-CoT (Wei et al., 2023) (Random) | 12.3/19.2 | 32.4/43.7 | 27.7/34.6 | 36.6/43.0 | 27.3/35.1 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 13.7/21.9 | 33.9/44.7 | 31.5/38.6 | 37.2/44.0 | 29.1/37.3 |
| Auto-CoT (Zhang et al., 2022) | 8.1/13.6 | 26.1/36.3 | 26.2/30.2 | 29.9/38.3 | 22.6/29.6 |
| SP-CoT (Ours) | 14.5/22.6 | 33.2/42.9 | 30.1/34.7 | 37.5/43.6 | 28.8/36.0 |
| SP-CoT + Self-Consistency (Wang et al., 2023) | 18.3/28.3 | -/- | -/- | 47.1/54.0 | -/- |
| Method | ChatGPT | InstructGPT | Alpaca-13B | Vicuna-13B | Wizard-13B |
| Zero-shot | 20.8 | 22.9 | 8.9 | 12.6 | 10.9 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 28.7 | 33.1 | 18.0 | 23.7 | 24.7 |
| Manual-CoT (Wei et al., 2023) (Random) | 26.3 | 32.4 | 17.7 | 22.0 | 24.0 |
| Auto-CoT (Zhang et al., 2022) | 21.2 | 31.4 | 16.9 | 21.3 | 23.3 |
| SP-CoT (Ours) | 26.1 | 33.1 | 17.4 | 21.5 | 22.5 |
| ChatGPT | Alpaca-13B | Vicuna-13B | WizardLM-13B | |
| Number of Samples | 3550 | 1354 | 1562 | 1604 |
| Avg. Number of Hops | 3.04 | 2.82 | 2.92 | 3.16 |
| Avg. Question tokens | 34.61 | 25.35 | 31.05 | 34.30 |
| Avg. Answer tokens | 1.86 | 1.99 | 2.02 | 1.91 |
| Method | EM | F1 |
| Zero-shot | 12.9 | 24.3 |
| Zero-shot-CoT (Kojima et al., 2023) | 13.5 | 25.2 |
| Manual-CoT (Wei et al., 2023) (Random) | 17.7 | 30.7 |
| SP-CoT (Ours) | 16.0 | 28.2 |
| Graph | QA | Decomposition |
| ○→○ | Q: What is another common name for some members of the genus to which daffodils belong, besides daffodil? +A: Jonquil | 1. Q: What is the genus to which daffodils belong? +A: Narcissus +2. Q: What is another common name for some members of the Narcissus genus besides daffodil? +A: Jonquil |
| ○→○→○ | Q: Who held the titles of King of the French and The Duke of Orleans during the year when John Lindley first described the subfamily of the family Asparagaceae that hyacinths belong to? +A: Louis Philippe | 1. Q: What is the subfamily of the family Asparagaceae that hyacinths belong to? +A: Scilloideae +2. Q: In what year was the subfamily Scilloideae first described by John Lindley? +A: 1830 +3. Q: Who became King of the French in 1830 and also held the title of The Duke of Orleans? +A: Louis Philippe |
| ○ | Q: Which country, commonly visited by ships traveling from Europe and where most species of roses are native to, is located south of the remote island where Napoleon Bonaparte was exiled and eventually died? +A: South Africa | 1. Q: Where are most species of roses native to? +A: Asia +2. Q: To which remote island was Napoleon Bonaparte exiled and where did he eventually die? +A: Saint Helena +3. Q: What country is located south of Saint Helena and was a common destination for ships traveling from Asia and Europe? +A: South Africa |
| ○→○→○ | Q: What work did the person who shared credit with the father of the biblical figure referenced in the title of the song "Stairway to Heaven" defend his optimism in, while also famously contributing to the development of infinitesimal calculus with Newton? +A: Theodicy | 1. Q: Who is the biblical figure that the title of the song "Stairway to Heaven" references? +A: Jacob +2. Q: Who was the father of Jacob? +A: Isaac +3. Q: Who shares credit with Isaac Newton for developing the infinitesimal calculus? +A: Gottfried Wilhelm Leibniz +4. Q: What work did Gottfried Wilhelm Leibniz famously defend his optimism in? +A: Theodicy |
| ○→○ | Q: What body of water separates the country that is a key ally of the country that Iraq launched Scud missiles at during the Gulf War, from Canada? +A: the Atlantic Ocean | 1. Q: What country did Iraq launch Scud missiles at during the Gulf War, and why is this relevant to the conflict? +A: Israel +2. Q: What is the native region of Carrot? +A: the Middle East +3. Q: What is the name of the country that is a key ally of Israel in the Middle East? +A: the United States +4. Q: What body of water separates the United States from Canada? +A: the Atlantic Ocean |
| ○→○ | Q: What is the name of the town in the country that is the largest in South America after the country where Lionel Messi represents in international football competitions where Tom Brady's wife was born? +A: Très de Maio | 1. Q: Which country does Lionel Messi represent in international football competitions? +A: Argentina +2. Q: What is the largest country in South America after Argentina? +A: Brazil +3. Q: What is the name of the model that Tom Brady is married to? +A: Gisele Bündchen +4. Q: What is the name of the town in Brazil where Gisele Bündchen was born? +A: Très de Maio |
| Methods | MSQ | HotpotQA | 2Wiki | CWebQ | Average | |||||
| EM | F1 | EM | F1 | EM | F1 | EM | F1 | EM | F1 | |
| Fine-tuning methods with extra corpus | ||||||||||
| DPR† (Karpukhin et al., 2020) | 7.5 | 16.6 | 14.7 | 23.0 | 6.5 | 14.6 | 21.3 | 30.2 | 12.5 | 21.1 |
| RAG† (Lewis et al., 2020) | 3.9 | 8.2 | 9.6 | 15.5 | 15.7 | 18.6 | 13.3 | 15.8 | 10.6 | 14.5 |
| REALM† (Guu et al., 2020) | 3.7 | 8.5 | 15.3 | 22.0 | 5.1 | 9.6 | 21.1 | 27.3 | 11.3 | 16.9 |
| T5-11B-SSM (Roberts et al., 2020) | 10.6 | 16.6 | 15.4 | 22.4 | 15.6 | 20.5 | 28.5 | 35.5 | 17.5 | 23.8 |
| Retrieval-based methods with LLMs | ||||||||||
| DPR† + ChatGPT | 1.7 | 4.1 | 15.8 | 21.4 | 10.9 | 18.1 | 13.7 | 18.7 | 10.5 | 15.6 |
| DPR† + Alpaca-13B (Taori et al., 2023) | 2.2 | 8.4 | 12.0 | 21.5 | 12.6 | 21.2 | 15.9 | 27.1 | 10.7 | 19.5 |
| DPR† + Vicuna-13B (Chiang et al., 2023) | 2.2 | 7.4 | 18.6 | 25.8 | 23.9 | 27.6 | 20.3 | 27.9 | 16.2 | 22.1 |
| DPR† + WizardLM-13B (Xu et al., 2023) | 3.5 | 10.0 | 19.9 | 28.4 | 22.8 | 27.5 | 24.2 | 32.2 | 17.6 | 24.5 |
| DPR† + InstructGPT (Ouyang et al., 2022) | 4.8 | 11.6 | 26.3 | 34.8 | 23.3 | 27.1 | 34.4 | 41.6 | 22.2 | 28.8 |
| LLM-only methods on ChatGPT | ||||||||||
| Zero-Shot | 3.1 | 7.3 | 22.4 | 30.0 | 18.7 | 21.7 | 31.6 | 37.5 | 19.0 | 24.1 |
| Self-Prompting (Li et al., 2022) | 2.9 | 6.2 | 23.8 | 31.2 | 18.9 | 23.5 | 26.8 | 32.6 | 18.1 | 23.4 |
| GENREAD (Yu et al., 2023) | 8.6 | 14.6 | 33.2 | 42.6 | 30.4 | 35.3 | 33.7 | 40.1 | 26.5 | 33.2 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0 | 8.8 | 22.6 | 29.6 | 24.3 | 27.1 | 30.3 | 36.2 | 20.6 | 25.4 |
| Auto-CoT (Zhang et al., 2022) | 8.1 | 13.6 | 26.1 | 36.3 | 26.2 | 30.2 | 29.9 | 38.4 | 22.6 | 29.6 |
| Manual-CoT (random) (Wei et al., 2023) | 12.3 | 19.2 | 32.4 | 43.7 | 27.7 | 34.6 | 36.6 | 43.0 | 27.3 | 35.1 |
| SP-CoT (Ours) | 14.5 | 22.6 | 33.2 | 42.9 | 30.1 | 34.7 | 37.5 | 43.6 | 28.8 | 36.0 |
| Model | Size | Method | MSQ | HotpotQA | 2Wiki | CWebQ | Mean | Boost |
| Alpaca (Taori et al., 2023) | 13B | Zero-shot | 0.9 | 9.5 | 12.5 | 12.6 | 8.9 | - |
| Alpaca (Taori et al., 2023) | 13B | SP-CoT | 5.3 | 19.8 | 17.8 | 26.5 | 17.4 | 8.5↑ |
| Vicuna (Chiang et al., 2023) | 13B | Zero-shot | 2.3 | 11.5 | 18.1 | 18.3 | 12.6 | - |
| Vicuna (Chiang et al., 2023) | 13B | SP-CoT | 8.5 | 23.3 | 22.0 | 32.2 | 21.5 | 8.9↑ |
| WizardLM (Xu et al., 2023) | 13B | Zero-shot | 2.0 | 10.3 | 13.6 | 17.5 | 10.9 | - |
| WizardLM (Xu et al., 2023) | 13B | SP-CoT | 7.3 | 24.0 | 27.2 | 31.3 | 22.5 | 11.6↑ |
| InstructGPT (Ouyang et al., 2022) | 175B | Zero-shot | 4.4 | 25.6 | 22.8 | 38.6 | 22.9 | - |
| InstructGPT (Ouyang et al., 2022) | 175B | SP-CoT | 14.8 | 37.4 | 33.7 | 46.6 | 33.1 | 10.2↑ |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Random | 12.0±0.8 | 29.5±0.6 | 25.6±1.2 | 34.3±0.9 | 25.4±0.3 |
| Retrieve | 10.5 | 27.9 | 24.3 | 33.5 | 24.1 |
| ClusterCenter | 10.4 | 26.2 | 22.8 | 33.0 | 23.1 |
| RetrieveInTypeCluster | 11.6 | 28.7 | 23.5 | 35.3 | 24.8 |
| RetrieveInCluster | 11.5 | 30.9 | 27.8 | 34.0 | 26.1 |
| MSQ | HotpotQA | CWebQ | 2Wiki | |
| Q len. | 18.11 | 15.83 | 13.37 | 11.98 |
| A len. | 2.8 | 2.46 | 2.42 | 2.41 |
| Steps | 2.65 | 2.68 | - | 2.47 |
| Size | 2417 | 7405 | 3519 | 12576 |
| Method | MSQ | HotpotQA | 2Wiki | CWebQ | Average |
| Zero-shot | 3.1/7.3 | 22.4/30.0 | 18.7/21.7 | 31.6/37.5 | 19.0/24.1 |
| Zero-shot-CoT (Kojima et al., 2023) | 5.0/8.8 | 22.6/29.6 | 24.3/27.1 | 30.3/36.2 | 20.6/25.4 |
| Manual-CoT (Wei et al., 2023) (Random) | 12.3/19.2 | 32.4/43.7 | 27.7/34.6 | 36.6/43.0 | 27.3/35.1 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 13.7/21.9 | 33.9/44.7 | 31.5/38.6 | 37.2/44.0 | 29.1/37.3 |
| Auto-CoT (Zhang et al., 2022) | 8.1/13.6 | 26.1/36.3 | 26.2/30.2 | 29.9/38.3 | 22.6/29.6 |
| SP-CoT (Ours) | 14.5/22.6 | 33.2/42.9 | 30.1/34.7 | 37.5/43.6 | 28.8/36.0 |
| SP-CoT + Self-Consistency (Wang et al., 2023) | 18.3/28.3 | -/- | -/- | 47.1/54.0 | -/- |
| Method | ChatGPT | InstructGPT | Alpaca-13B | Vicuna-13B | Wizard-13B |
| Zero-shot | 20.8 | 22.9 | 8.9 | 12.6 | 10.9 |
| Manual-CoT (Wei et al., 2023) (Cherry-Pick) | 28.7 | 33.1 | 18.0 | 23.7 | 24.7 |
| Manual-CoT (Wei et al., 2023) (Random) | 26.3 | 32.4 | 17.7 | 22.0 | 24.0 |
| Auto-CoT (Zhang et al., 2022) | 21.2 | 31.4 | 16.9 | 21.3 | 23.3 |
| SP-CoT (Ours) | 26.1 | 33.1 | 17.4 | 21.5 | 22.5 |
| ChatGPT | Alpaca-13B | Vicuna-13B | WizardLM-13B | |
| Number of Samples | 3550 | 1354 | 1562 | 1604 |
| Avg. Number of Hops | 3.04 | 2.82 | 2.92 | 3.16 |
| Avg. Question tokens | 34.61 | 25.35 | 31.05 | 34.30 |
| Avg. Answer tokens | 1.86 | 1.99 | 2.02 | 1.91 |
| Method | EM | F1 |
| Zero-shot | 12.9 | 24.3 |
| Zero-shot-CoT (Kojima et al., 2023) | 13.5 | 25.2 |
| Manual-CoT (Wei et al., 2023) (Random) | 17.7 | 30.7 |
| SP-CoT (Ours) | 16.0 | 28.2 |
| Graph | QA | Decomposition |
| ○→○ | Q: What is another common name for some members of the genus to which daffodils belong, besides daffodil? \nA: Jonquil | 1. Q: What is the genus to which daffodils belong? \nA: Narcissus \n2. Q: What is another common name for some members of the Narcissus genus besides daffodil? \nA: Jonquil |
| ○→○→○ | Q: Who held the titles of King of the French and The Duke of Orleans during the year when John Lindley first described the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Louis Philippe | 1. Q: What is the subfamily of the family Asparagaceae that hyacinths belong to? \nA: Scilloideae \n2. Q: In what year was the subfamily Scilloideae first described by John Lindley? \nA: 1830 \n3. Q: Who became King of the French in 1830 and also held the title of The Duke of Orleans? \nA: Louis Philippe |
| ○ | Q: Which country, commonly visited by ships traveling from Europe and where most species of roses are native to, is located south of the remote island where Napoleon Bonaparte was exiled and eventually died? \nA: South Africa | 1. Q: Where are most species of roses native to? \nA: Asia \n2. Q: To which remote island was Napoleon Bonaparte exiled and where did he eventually die? \nA: Saint Helena \n3. Q: What country is located south of Saint Helena and was a common destination for ships traveling from Asia and Europe? \nA: South Africa |
| ○→○→○ | Q: What work did the person who shared credit with the father of the biblical figure referenced in the title of the song "Stairway to Heaven" defend his optimism in, while also famously contributing to the development of infinitesimal calculus with Newton? \nA: Theodicy | 1. Q: Who is the biblical figure that the title of the song "Stairway to Heaven" references? \nA: Jacob \n2. Q: Who was the father of Jacob? \nA: Isaac \n3. Q: Who shares credit with Isaac Newton for developing the infinitesimal calculus? \nA: Gottfried Wilhelm Leibniz \n4. Q: What work did Gottfried Wilhelm Leibniz famously defend his optimism in? \nA: Theodicy |
| ○→○ | Q: What body of water separates the country that is a key ally of the country that Iraq launched Scud missiles at during the Gulf War, from Canada? \nA: the Atlantic Ocean | 1. Q: What country did Iraq launch Scud missiles at during the Gulf War, and why is this relevant to the conflict? \nA: Israel \n2. Q: What is the native region of Carrot? \nA: the Middle East \n3. Q: What is the name of the country that is a key ally of Israel in the Middle East? \nA: the United States \n4. Q: What body of water separates the United States from Canada? \nA: the Atlantic Ocean |
| ○→○ | Q: What is the name of the town in the country that is the largest in South America after the country where Lionel Messi represents in international football competitions where Tom Brady's wife was born? \nA: Très de Maio | 1. Q: Which country does Lionel Messi represent in international football competitions? \nA: Argentina \n2. Q: What is the largest country in South America after Argentina? \nA: Brazil \n3. Q: What is the name of the model that Tom Brady is married to? \nA: Gisele Bündchen \n4. Q: What is the name of the town in Brazil where Gisele Bündchen was born? \nA: Très de Maio |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Prefix Tuning | 78.21.7 | 38.02.5 | 74.03.2 | 84.53.3 | 66.23.9 | 70.12.9 | 70.42.6 | 54.42.0 | 55.74.3 | 53.41.3 | 68.11.2 | 64.91.3 | 64.82.5 |
| P-Tuning-v2 | 83.10.8 | 41.71.5 | 82.30.8 | 88.70.5 | 74.82.4 | 78.12.1 | 67.55.4 | 53.91.3 | 61.41.6 | 54.52.2 | 70.71.1 | 68.01.1 | 68.71.7 |
| FT | 83.61.7 | 41.22.6 | 81.70.9 | 88.30.9 | 80.01.4 | 79.81.7 | 71.91.5 | 56.92.1 | 62.30.6 | 54.61.6 | 70.20.7 | 69.51.0 | 70.01.4 |
| PT | 71.93.4 | 37.32.8 | 73.24.5 | 84.43.5 | 61.53.3 | 65.34.2 | 58.96.1 | 53.22.8 | 55.24.8 | 53.12.4 | 66.65.3 | 63.02.6 | 62.04.0 |
| PPT | 81.22.0 | 40.25.4 | 81.20.7 | 83.67.3 | 66.83.7 | 73.42.4 | 60.77.7 | 55.41.2 | 60.43.9 | 53.61.3 | 68.00.8 | 63.10.7 | 65.63.1 |
| Unified-PPT | 76.87.7 | 44.71.7 | 79.03.3 | 87.70.6 | 64.25.8 | 68.42.9 | 63.92.9 | 54.41.1 | 56.72.4 | 54.51.6 | 67.81.1 | 67.62.9 | 65.52.8 |
| MetaPT | 85.71.3 | 45.31.1 | 82.54.5 | 88.50.3 | 73.23.7 | 78.72.2 | 65.42.4 | 56.11.7 | 58.23.1 | 54.11.3 | 69.60.6 | 68.80.9 | 68.81.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 | 71.30.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Zero-shot Inference | 89.1 | 52.3 | 83.3 | 80.6 | 57.4 | 87.2 | 76.8 | 75.8 | 85.0 | 50.5 | 77.2 | 77.5 | 74.4 |
| Few-shot Inference | 93.2 | 53.3 | 88.5 | 87.8 | 58.6 | 91.6 | 83.9 | 79.1 | 86.9 | 64.3 | 79.9 | 81.0 | 79.0 |
| Prefix Tuning | 88.62.1 | 45.42.7 | 88.72.2 | 89.11.6 | 84.63.9 | 90.62.4 | 65.46.4 | 73.42.6 | 72.34.1 | 55.51.8 | 73.92.6 | 75.02.3 | 75.22.9 |
| P-Tuning-v2 | 91.21.0 | 53.51.1 | 90.00.8 | 89.81.1 | 87.73.2 | 92.30.9 | 69.12.9 | 62.34.6 | 76.83.6 | 63.12.0 | 77.21.7 | 77.31.6 | 77.52.0 |
| FT | 92.91.4 | 53.61.9 | 89.61.3 | 91.30.5 | 88.70.9 | 93.62.0 | 77.81.6 | 76.01.7 | 84.41.5 | 63.01.8 | 77.71.3 | 82.92.9 | 81.01.6 |
| PT | 88.29.2 | 45.73.4 | 85.65.6 | 88.65.1 | 81.65.1 | 85.15.6 | 65.47.2 | 61.64.0 | 69.35.9 | 54.81.1 | 71.12.7 | 70.28.6 | 72.35.3 |
| PPT | 91.02.6 | 49.12.3 | 88.81.3 | 88.03.0 | 87.61.9 | 90.81.7 | 67.14.8 | 67.12.2 | 81.51.1 | 56.11.3 | 75.82.2 | 71.63.1 | 76.22.3 |
| Unified-PPT | 89.98.3 | 47.42.8 | 89.70.9 | 89.01.5 | 86.44.3 | 88.43.2 | 69.33.5 | 63.21.1 | 74.93.2 | 59.91.7 | 74.02.4 | 76.32.0 | 75.72.9 |
| MetaPT | 93.70.8 | 52.51.3 | 90.70.7 | 89.70.6 | 88.12.4 | 91.81.8 | 72.12.2 | 74.11.3 | 77.52.5 | 58.51.6 | 76.61.8 | 81.42.4 | 78.91.6 |
| SUPMER | 95.50.4 | 55.30.7 | 91.40.5 | 90.70.7 | 90.30.8 | 93.01.5 | 87.61.5 | 81.41.0 | 88.30.6 | 65.01.7 | 78.10.8 | 85.10.4 | 83.50.9 |
| Model: T5-base (220M) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| Prefix Tuning | 78.12.2 | 82.71.0 | 81.51.2 | 84.40.6 | 82.72.6 | 84.92.2 | 82.41.6 |
| P-Tuning-v2 | 84.00.8 | 83.60.9 | 82.41.7 | 85.90.9 | 84.21.5 | 83.82.7 | 84.01.4 |
| FT | 84.40.2 | 83.90.6 | 81.00.6 | 84.10.6 | 85.00.7 | 85.30.6 | 84.00.6 |
| PT | 79.82.5 | 75.32.2 | 76.03.2 | 79.62.0 | 79.81.9 | 83.01.8 | 78.92.3 |
| PPT | 81.91.4 | 77.93.5 | 83.61.3 | 88.41.0 | 89.11.6 | 87.81.5 | 84.81.7 |
| Unified-PPT | 80.93.0 | 82.12.0 | 76.84.0 | 81.04.5 | 82.23.9 | 84.93.3 | 81.33.5 |
| MetaPT | 86.10.3 | 84.30.8 | 84.20.7 | 89.40.9 | 90.30.6 | - | (86.90.7) |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.6 | 91.10.5 | 90.40.4 | 88.00.5 |
| Model: Flan-T5-XL (3B) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| zero-shot inference | 77.8 | 84.6 | 86.2 | 86.8 | 88.6 | 87.8 | 85.3 |
| few-shot inference | 82.5 | 90.3 | 89.7 | 92.3 | 92.2 | 89.2 | 89.4 |
| Prefix Tuning | 83.01.1 | 85.31.9 | 83.41.1 | 87.31.5 | 86.42.4 | 89.81.0 | 85.91.5 |
| P-Tuning-v2 | 85.60.4 | 86.72.3 | 86.81.8 | 88.62.1 | 90.22.3 | 88.92.6 | 87.81.9 |
| FT | 86.20.7 | 88.82.2 | 84.61.3 | 87.21.3 | 89.81.1 | 90.70.9 | 87.91.3 |
| PT | 82.61.2 | 79.03.9 | 82.51.7 | 84.22.3 | 84.52.5 | 84.82.1 | 82.92.3 |
| PPT | 85.20.9 | 83.23.3 | 89.71.8 | 92.41.7 | 93.61.5 | 89.61.1 | 89.01.7 |
| Unified-PPT | 83.00.8 | 82.53.0 | 82.04.8 | 86.03.0 | 86.62.2 | 85.62.0 | 84.32.6 |
| MetaPT | 87.10.6 | 87.42.7 | 89.51.7 | 93.80.6 | 94.31.1 | - | (90.41.3) |
| SUPMER | 87.00.2 | 89.81.4 | 91.11.2 | 95.10.8 | 95.80.9 | 91.80.8 | 91.80.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| 1 SUPMER (only labeled) | 87.5 | 47.0 | 83.8 | 89.9 | 75.4 | 79.6 | 67.9 | 56.6 | 59.0 | 54.6 | 69.2 | 69.5 | 70.0 |
| 2 SUPMER (only unlabeled) | 87.3 | 46.7 | 84.0 | 89.3 | 79.6 | 80.2 | 72.4 | 57.3 | 61.7 | 54.8 | 71.3 | 70.5 | 71.3 |
| 3 PPT (labeled + unlabeled) | 84.7 | 45.0 | 82.4 | 87.8 | 67.2 | 77.4 | 64.3 | 55.3 | 61.6 | 53.9 | 68.9 | 67.7 | 68.0 |
| 4 MetaPT (labeled + unlabeled) | 86.1 | 46.3 | 83.7 | 89.4 | 73.8 | 80.1 | 67.2 | 57.4 | 60.0 | 54.3 | 70.1 | 69.9 | 69.9 |
| 5 SUPMER (labeled + unlabeled) | 89.1 | 48.2 | 85.7 | 90.8 | 79.3 | 83.4 | 73.2 | 58.8 | 63.7 | 55.3 | 70.5 | 71.5 | 72.5 |
| Methods | Few-shot Learning | DG | |
| 1 | only sp | 66.7 | 85.7 |
| 2 | only mc | 66.9 | 85.7 |
| 3 | only ss | 67.4 | 86.7 |
| 4 | w/o ta | 68.6 | 85.3 |
| 5 | w/o curriculum | 69.9 | 87.0 |
| 6 | w/o mgr | 69.4 | 86.1 |
| 7 | SUPMER | 71.3 | 88.0 |
| Hyper-parameter | Value |
| Number of clusters for each task format | 250 |
| Tasks per batch | 4 |
| Size of support set per task | 32 |
| Size of query set per task | 32 |
| Optimizer | Adam |
| Inner loop learning rate | 0.1 |
| Outer loop learning rate | 0.1 |
| Learning rate for φ | 1e-4 |
| Scheduler | Linear scheduler |
| Warm-up steps | 0 |
| Max training steps | 100,000 |
| Validation steps | 2,000 |
| Max sequence length | 512 |
| λ | 1.0 |
| m | 2.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP |
| only sp | 83.61.5 | 42.62.2 | 81.71.8 | 86.00.6 | 65.82.8 | 71.26.6 | 64.62.1 | 57.02.5 | 58.43.3 | 53.61.5 | 69.91.3 | 66.01.0 |
| only mc | 83.41.4 | 44.51.9 | 79.35.1 | 88.30.5 | 70.54.7 | 66.41.4 | 65.93.1 | 54.91.3 | 58.71.7 | 54.21.8 | 68.80.8 | 67.61.3 |
| only ss | 84.51.5 | 45.02.0 | 81.50.7 | 88.40.5 | 73.33.1 | 79.15.8 | 62.12.6 | 53.91.0 | 56.51.4 | 53.31.3 | 67.71.3 | 63.71.7 |
| w/o ta | 84.71.0 | 40.13.3 | 81.91.8 | 87.20.8 | 73.62.8 | 78.83.7 | 66.41.9 | 56.60.9 | 59.41.8 | 54.32.4 | 69.51.1 | 70.21.1 |
| w/o curriculum | 86.80.8 | 40.82.2 | 82.31.3 | 88.40.9 | 74.83.1 | 79.71.6 | 71.02.1 | 56.50.8 | 62.61.4 | 55.41.1 | 69.70.8 | 71.31.2 |
| w/o mgr | 85.01.3 | 44.51.1 | 82.80.7 | 88.00.5 | 76.01.7 | 79.55.0 | 67.11.6 | 56.80.8 | 58.92.4 | 54.42.0 | 70.01.0 | 70.30.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 |
| Method | Source | Target | ||||
| A | B | D | E | K | R | |
| only sp | 83.41.1 | 82.11.4 | 83.00.7 | 88.51.2 | 88.90.8 | 88.10.7 |
| only mc | 84.00.6 | 82.31.2 | 81.50.9 | 88.51.0 | 89.30.7 | 88.80.7 |
| only ss | 83.60.7 | 84.70.8 | 84.20.6 | 88.90.9 | 89.70.9 | 89.00.3 |
| w/o ta | 83.40.8 | 82.01.4 | 81.71.5 | 87.80.8 | 88.20.6 | 88.60.6 |
| w/o curriculum | 84.00.5 | 84.70.8 | 83.90.6 | 89.60.5 | 90.30.8 | 89.71.1 |
| w/o mgr | 83.80.4 | 83.40.5 | 83.30.5 | 88.10.6 | 89.20.8 | 88.90.4 |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.7 | 91.10.5 | 90.40.4 |
| Method | Source | Target | AVG | |||
| A | B | D | E | K | ||
| 1 SUPMER (only labeled) | 86.4 | 84.7 | 84.8 | 90.0 | 90.7 | 87.3 |
| 2 SUPMER (only unlabeled) | 85.7 | 85.3 | 85.1 | 90.3 | 91.1 | 87.5 |
| 3 PPT (labeled + unlabeled) | 83.1 | 79.0 | 84.4 | 89.3 | 90.6 | 85.3 |
| 4 MetaPT (labeled + unlabeled) | 86.3 | 85.3 | 86.7 | 90.1 | 91.4 | 88.0 |
| 5 SUPMER (labeled + unlabeled) | 86.6 | 88.6 | 88.5 | 92.7 | 93.7 | 90.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Prefix Tuning | 78.21.7 | 38.02.5 | 74.03.2 | 84.53.3 | 66.23.9 | 70.12.9 | 70.42.6 | 54.42.0 | 55.74.3 | 53.41.3 | 68.11.2 | 64.91.3 | 64.82.5 |
| P-Tuning-v2 | 83.10.8 | 41.71.5 | 82.30.8 | 88.70.5 | 74.82.4 | 78.12.1 | 67.55.4 | 53.91.3 | 61.41.6 | 54.52.2 | 70.71.1 | 68.01.1 | 68.71.7 |
| FT | 83.61.7 | 41.22.6 | 81.70.9 | 88.30.9 | 80.01.4 | 79.81.7 | 71.91.5 | 56.92.1 | 62.30.6 | 54.61.6 | 70.20.7 | 69.51.0 | 70.01.4 |
| PT | 71.93.4 | 37.32.8 | 73.24.5 | 84.43.5 | 61.53.3 | 65.34.2 | 58.96.1 | 53.22.8 | 55.24.8 | 53.12.4 | 66.65.3 | 63.02.6 | 62.04.0 |
| PPT | 81.22.0 | 40.25.4 | 81.20.7 | 83.67.3 | 66.83.7 | 73.42.4 | 60.77.7 | 55.41.2 | 60.43.9 | 53.61.3 | 68.00.8 | 63.10.7 | 65.63.1 |
| Unified-PPT | 76.87.7 | 44.71.7 | 79.03.3 | 87.70.6 | 64.25.8 | 68.42.9 | 63.92.9 | 54.41.1 | 56.72.4 | 54.51.6 | 67.81.1 | 67.62.9 | 65.52.8 |
| MetaPT | 85.71.3 | 45.31.1 | 82.54.5 | 88.50.3 | 73.23.7 | 78.72.2 | 65.42.4 | 56.11.7 | 58.23.1 | 54.11.3 | 69.60.6 | 68.80.9 | 68.81.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 | 71.30.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Zero-shot Inference | 89.1 | 52.3 | 83.3 | 80.6 | 57.4 | 87.2 | 76.8 | 75.8 | 85.0 | 50.5 | 77.2 | 77.5 | 74.4 |
| Few-shot Inference | 93.2 | 53.3 | 88.5 | 87.8 | 58.6 | 91.6 | 83.9 | 79.1 | 86.9 | 64.3 | 79.9 | 81.0 | 79.0 |
| Prefix Tuning | 88.62.1 | 45.42.7 | 88.72.2 | 89.11.6 | 84.63.9 | 90.62.4 | 65.46.4 | 73.42.6 | 72.34.1 | 55.51.8 | 73.92.6 | 75.02.3 | 75.22.9 |
| P-Tuning-v2 | 91.21.0 | 53.51.1 | 90.00.8 | 89.81.1 | 87.73.2 | 92.30.9 | 69.12.9 | 62.34.6 | 76.83.6 | 63.12.0 | 77.21.7 | 77.31.6 | 77.52.0 |
| FT | 92.91.4 | 53.61.9 | 89.61.3 | 91.30.5 | 88.70.9 | 93.62.0 | 77.81.6 | 76.01.7 | 84.41.5 | 63.01.8 | 77.71.3 | 82.92.9 | 81.01.6 |
| PT | 88.29.2 | 45.73.4 | 85.65.6 | 88.65.1 | 81.65.1 | 85.15.6 | 65.47.2 | 61.64.0 | 69.35.9 | 54.81.1 | 71.12.7 | 70.28.6 | 72.35.3 |
| PPT | 91.02.6 | 49.12.3 | 88.81.3 | 88.03.0 | 87.61.9 | 90.81.7 | 67.14.8 | 67.12.2 | 81.51.1 | 56.11.3 | 75.82.2 | 71.63.1 | 76.22.3 |
| Unified-PPT | 89.98.3 | 47.42.8 | 89.70.9 | 89.01.5 | 86.44.3 | 88.43.2 | 69.33.5 | 63.21.1 | 74.93.2 | 59.91.7 | 74.02.4 | 76.32.0 | 75.72.9 |
| MetaPT | 93.70.8 | 52.51.3 | 90.70.7 | 89.70.6 | 88.12.4 | 91.81.8 | 72.12.2 | 74.11.3 | 77.52.5 | 58.51.6 | 76.61.8 | 81.42.4 | 78.91.6 |
| SUPMER | 95.50.4 | 55.30.7 | 91.40.5 | 90.70.7 | 90.30.8 | 93.01.5 | 87.61.5 | 81.41.0 | 88.30.6 | 65.01.7 | 78.10.8 | 85.10.4 | 83.50.9 |
| Model: T5-base (220M) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| Prefix Tuning | 78.12.2 | 82.71.0 | 81.51.2 | 84.40.6 | 82.72.6 | 84.92.2 | 82.41.6 |
| P-Tuning-v2 | 84.00.8 | 83.60.9 | 82.41.7 | 85.90.9 | 84.21.5 | 83.82.7 | 84.01.4 |
| FT | 84.40.2 | 83.90.6 | 81.00.6 | 84.10.6 | 85.00.7 | 85.30.6 | 84.00.6 |
| PT | 79.82.5 | 75.32.2 | 76.03.2 | 79.62.0 | 79.81.9 | 83.01.8 | 78.92.3 |
| PPT | 81.91.4 | 77.93.5 | 83.61.3 | 88.41.0 | 89.11.6 | 87.81.5 | 84.81.7 |
| Unified-PPT | 80.93.0 | 82.12.0 | 76.84.0 | 81.04.5 | 82.23.9 | 84.93.3 | 81.33.5 |
| MetaPT | 86.10.3 | 84.30.8 | 84.20.7 | 89.40.9 | 90.30.6 | - | (86.90.7) |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.6 | 91.10.5 | 90.40.4 | 88.00.5 |
| Model: Flan-T5-XL (3B) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| zero-shot inference | 77.8 | 84.6 | 86.2 | 86.8 | 88.6 | 87.8 | 85.3 |
| few-shot inference | 82.5 | 90.3 | 89.7 | 92.3 | 92.2 | 89.2 | 89.4 |
| Prefix Tuning | 83.01.1 | 85.31.9 | 83.41.1 | 87.31.5 | 86.42.4 | 89.81.0 | 85.91.5 |
| P-Tuning-v2 | 85.60.4 | 86.72.3 | 86.81.8 | 88.62.1 | 90.22.3 | 88.92.6 | 87.81.9 |
| FT | 86.20.7 | 88.82.2 | 84.61.3 | 87.21.3 | 89.81.1 | 90.70.9 | 87.91.3 |
| PT | 82.61.2 | 79.03.9 | 82.51.7 | 84.22.3 | 84.52.5 | 84.82.1 | 82.92.3 |
| PPT | 85.20.9 | 83.23.3 | 89.71.8 | 92.41.7 | 93.61.5 | 89.61.1 | 89.01.7 |
| Unified-PPT | 83.00.8 | 82.53.0 | 82.04.8 | 86.03.0 | 86.62.2 | 85.62.0 | 84.32.6 |
| MetaPT | 87.10.6 | 87.42.7 | 89.51.7 | 93.80.6 | 94.31.1 | - | (90.41.3) |
| SUPMER | 87.00.2 | 89.81.4 | 91.11.2 | 95.10.8 | 95.80.9 | 91.80.8 | 91.80.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| 1 SUPMER (only labeled) | 87.5 | 47.0 | 83.8 | 89.9 | 75.4 | 79.6 | 67.9 | 56.6 | 59.0 | 54.6 | 69.2 | 69.5 | 70.0 |
| 2 SUPMER (only unlabeled) | 87.3 | 46.7 | 84.0 | 89.3 | 79.6 | 80.2 | 72.4 | 57.3 | 61.7 | 54.8 | 71.3 | 70.5 | 71.3 |
| 3 PPT (labeled + unlabeled) | 84.7 | 45.0 | 82.4 | 87.8 | 67.2 | 77.4 | 64.3 | 55.3 | 61.6 | 53.9 | 68.9 | 67.7 | 68.0 |
| 4 MetaPT (labeled + unlabeled) | 86.1 | 46.3 | 83.7 | 89.4 | 73.8 | 80.1 | 67.2 | 57.4 | 60.0 | 54.3 | 70.1 | 69.9 | 69.9 |
| 5 SUPMER (labeled + unlabeled) | 89.1 | 48.2 | 85.7 | 90.8 | 79.3 | 83.4 | 73.2 | 58.8 | 63.7 | 55.3 | 70.5 | 71.5 | 72.5 |
| Methods | Few-shot Learning | DG | |
| 1 | only sp | 66.7 | 85.7 |
| 2 | only mc | 66.9 | 85.7 |
| 3 | only ss | 67.4 | 86.7 |
| 4 | w/o ta | 68.6 | 85.3 |
| 5 | w/o curriculum | 69.9 | 87.0 |
| 6 | w/o mgr | 69.4 | 86.1 |
| 7 | SUPMER | 71.3 | 88.0 |
| Hyper-parameter | Value |
| Number of clusters for each task format | 250 |
| Tasks per batch | 4 |
| Size of support set per task | 32 |
| Size of query set per task | 32 |
| Optimizer | Adam |
| Inner loop learning rate | 0.1 |
| Outer loop learning rate | 0.1 |
| Learning rate for φ | 1e-4 |
| Scheduler | Linear scheduler |
| Warm-up steps | 0 |
| Max training steps | 100,000 |
| Validation steps | 2,000 |
| Max sequence length | 512 |
| λ | 1.0 |
| m | 2.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP |
| only sp | 83.61.5 | 42.62.2 | 81.71.8 | 86.00.6 | 65.82.8 | 71.26.6 | 64.62.1 | 57.02.5 | 58.43.3 | 53.61.5 | 69.91.3 | 66.01.0 |
| only mc | 83.41.4 | 44.51.9 | 79.35.1 | 88.30.5 | 70.54.7 | 66.41.4 | 65.93.1 | 54.91.3 | 58.71.7 | 54.21.8 | 68.80.8 | 67.61.3 |
| only ss | 84.51.5 | 45.02.0 | 81.50.7 | 88.40.5 | 73.33.1 | 79.15.8 | 62.12.6 | 53.91.0 | 56.51.4 | 53.31.3 | 67.71.3 | 63.71.7 |
| w/o ta | 84.71.0 | 40.13.3 | 81.91.8 | 87.20.8 | 73.62.8 | 78.83.7 | 66.41.9 | 56.60.9 | 59.41.8 | 54.32.4 | 69.51.1 | 70.21.1 |
| w/o curriculum | 86.80.8 | 40.82.2 | 82.31.3 | 88.40.9 | 74.83.1 | 79.71.6 | 71.02.1 | 56.50.8 | 62.61.4 | 55.41.1 | 69.70.8 | 71.31.2 |
| w/o mgr | 85.01.3 | 44.51.1 | 82.80.7 | 88.00.5 | 76.01.7 | 79.55.0 | 67.11.6 | 56.80.8 | 58.92.4 | 54.42.0 | 70.01.0 | 70.30.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 |
| Method | Source | Target | ||||
| A | B | D | E | K | R | |
| only sp | 83.41.1 | 82.11.4 | 83.00.7 | 88.51.2 | 88.90.8 | 88.10.7 |
| only mc | 84.00.6 | 82.31.2 | 81.50.9 | 88.51.0 | 89.30.7 | 88.80.7 |
| only ss | 83.60.7 | 84.70.8 | 84.20.6 | 88.90.9 | 89.70.9 | 89.00.3 |
| w/o ta | 83.40.8 | 82.01.4 | 81.71.5 | 87.80.8 | 88.20.6 | 88.60.6 |
| w/o curriculum | 84.00.5 | 84.70.8 | 83.90.6 | 89.60.5 | 90.30.8 | 89.71.1 |
| w/o mgr | 83.80.4 | 83.40.5 | 83.30.5 | 88.10.6 | 89.20.8 | 88.90.4 |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.7 | 91.10.5 | 90.40.4 |
| Method | Source | Target | AVG | |||
| A | B | D | E | K | ||
| 1 SUPMER (only labeled) | 86.4 | 84.7 | 84.8 | 90.0 | 90.7 | 87.3 |
| 2 SUPMER (only unlabeled) | 85.7 | 85.3 | 85.1 | 90.3 | 91.1 | 87.5 |
| 3 PPT (labeled + unlabeled) | 83.1 | 79.0 | 84.4 | 89.3 | 90.6 | 85.3 |
| 4 MetaPT (labeled + unlabeled) | 86.3 | 85.3 | 86.7 | 90.1 | 91.4 | 88.0 |
| 5 SUPMER (labeled + unlabeled) | 86.6 | 88.6 | 88.5 | 92.7 | 93.7 | 90.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Prefix Tuning | 78.21.7 | 38.02.5 | 74.03.2 | 84.53.3 | 66.23.9 | 70.12.9 | 70.42.6 | 54.42.0 | 55.74.3 | 53.41.3 | 68.11.2 | 64.91.3 | 64.82.5 |
| P-Tuning-v2 | 83.10.8 | 41.71.5 | 82.30.8 | 88.70.5 | 74.82.4 | 78.12.1 | 67.55.4 | 53.91.3 | 61.41.6 | 54.52.2 | 70.71.1 | 68.01.1 | 68.71.7 |
| FT | 83.61.7 | 41.22.6 | 81.70.9 | 88.30.9 | 80.01.4 | 79.81.7 | 71.91.5 | 56.92.1 | 62.30.6 | 54.61.6 | 70.20.7 | 69.51.0 | 70.01.4 |
| PT | 71.93.4 | 37.32.8 | 73.24.5 | 84.43.5 | 61.53.3 | 65.34.2 | 58.96.1 | 53.22.8 | 55.24.8 | 53.12.4 | 66.65.3 | 63.02.6 | 62.04.0 |
| PPT | 81.22.0 | 40.25.4 | 81.20.7 | 83.67.3 | 66.83.7 | 73.42.4 | 60.77.7 | 55.41.2 | 60.43.9 | 53.61.3 | 68.00.8 | 63.10.7 | 65.63.1 |
| Unified-PPT | 76.87.7 | 44.71.7 | 79.03.3 | 87.70.6 | 64.25.8 | 68.42.9 | 63.92.9 | 54.41.1 | 56.72.4 | 54.51.6 | 67.81.1 | 67.62.9 | 65.52.8 |
| MetaPT | 85.71.3 | 45.31.1 | 82.54.5 | 88.50.3 | 73.23.7 | 78.72.2 | 65.42.4 | 56.11.7 | 58.23.1 | 54.11.3 | 69.60.6 | 68.80.9 | 68.81.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 | 71.30.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Zero-shot Inference | 89.1 | 52.3 | 83.3 | 80.6 | 57.4 | 87.2 | 76.8 | 75.8 | 85.0 | 50.5 | 77.2 | 77.5 | 74.4 |
| Few-shot Inference | 93.2 | 53.3 | 88.5 | 87.8 | 58.6 | 91.6 | 83.9 | 79.1 | 86.9 | 64.3 | 79.9 | 81.0 | 79.0 |
| Prefix Tuning | 88.62.1 | 45.42.7 | 88.72.2 | 89.11.6 | 84.63.9 | 90.62.4 | 65.46.4 | 73.42.6 | 72.34.1 | 55.51.8 | 73.92.6 | 75.02.3 | 75.22.9 |
| P-Tuning-v2 | 91.21.0 | 53.51.1 | 90.00.8 | 89.81.1 | 87.73.2 | 92.30.9 | 69.12.9 | 62.34.6 | 76.83.6 | 63.12.0 | 77.21.7 | 77.31.6 | 77.52.0 |
| FT | 92.91.4 | 53.61.9 | 89.61.3 | 91.30.5 | 88.70.9 | 93.62.0 | 77.81.6 | 76.01.7 | 84.41.5 | 63.01.8 | 77.71.3 | 82.92.9 | 81.01.6 |
| PT | 88.29.2 | 45.73.4 | 85.65.6 | 88.65.1 | 81.65.1 | 85.15.6 | 65.47.2 | 61.64.0 | 69.35.9 | 54.81.1 | 71.12.7 | 70.28.6 | 72.35.3 |
| PPT | 91.02.6 | 49.12.3 | 88.81.3 | 88.03.0 | 87.61.9 | 90.81.7 | 67.14.8 | 67.12.2 | 81.51.1 | 56.11.3 | 75.82.2 | 71.63.1 | 76.22.3 |
| Unified-PPT | 89.98.3 | 47.42.8 | 89.70.9 | 89.01.5 | 86.44.3 | 88.43.2 | 69.33.5 | 63.21.1 | 74.93.2 | 59.91.7 | 74.02.4 | 76.32.0 | 75.72.9 |
| MetaPT | 93.70.8 | 52.51.3 | 90.70.7 | 89.70.6 | 88.12.4 | 91.81.8 | 72.12.2 | 74.11.3 | 77.52.5 | 58.51.6 | 76.61.8 | 81.42.4 | 78.91.6 |
| SUPMER | 95.50.4 | 55.30.7 | 91.40.5 | 90.70.7 | 90.30.8 | 93.01.5 | 87.61.5 | 81.41.0 | 88.30.6 | 65.01.7 | 78.10.8 | 85.10.4 | 83.50.9 |
| Model: T5-base (220M) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| Prefix Tuning | 78.12.2 | 82.71.0 | 81.51.2 | 84.40.6 | 82.72.6 | 84.92.2 | 82.41.6 |
| P-Tuning-v2 | 84.00.8 | 83.60.9 | 82.41.7 | 85.90.9 | 84.21.5 | 83.82.7 | 84.01.4 |
| FT | 84.40.2 | 83.90.6 | 81.00.6 | 84.10.6 | 85.00.7 | 85.30.6 | 84.00.6 |
| PT | 79.82.5 | 75.32.2 | 76.03.2 | 79.62.0 | 79.81.9 | 83.01.8 | 78.92.3 |
| PPT | 81.91.4 | 77.93.5 | 83.61.3 | 88.41.0 | 89.11.6 | 87.81.5 | 84.81.7 |
| Unified-PPT | 80.93.0 | 82.12.0 | 76.84.0 | 81.04.5 | 82.23.9 | 84.93.3 | 81.33.5 |
| MetaPT | 86.10.3 | 84.30.8 | 84.20.7 | 89.40.9 | 90.30.6 | - | (86.90.7) |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.6 | 91.10.5 | 90.40.4 | 88.00.5 |
| Model: Flan-T5-XL (3B) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| zero-shot inference | 77.8 | 84.6 | 86.2 | 86.8 | 88.6 | 87.8 | 85.3 |
| few-shot inference | 82.5 | 90.3 | 89.7 | 92.3 | 92.2 | 89.2 | 89.4 |
| Prefix Tuning | 83.01.1 | 85.31.9 | 83.41.1 | 87.31.5 | 86.42.4 | 89.81.0 | 85.91.5 |
| P-Tuning-v2 | 85.60.4 | 86.72.3 | 86.81.8 | 88.62.1 | 90.22.3 | 88.92.6 | 87.81.9 |
| FT | 86.20.7 | 88.82.2 | 84.61.3 | 87.21.3 | 89.81.1 | 90.70.9 | 87.91.3 |
| PT | 82.61.2 | 79.03.9 | 82.51.7 | 84.22.3 | 84.52.5 | 84.82.1 | 82.92.3 |
| PPT | 85.20.9 | 83.23.3 | 89.71.8 | 92.41.7 | 93.61.5 | 89.61.1 | 89.01.7 |
| Unified-PPT | 83.00.8 | 82.53.0 | 82.04.8 | 86.03.0 | 86.62.2 | 85.62.0 | 84.32.6 |
| MetaPT | 87.10.6 | 87.42.7 | 89.51.7 | 93.80.6 | 94.31.1 | - | (90.41.3) |
| SUPMER | 87.00.2 | 89.81.4 | 91.11.2 | 95.10.8 | 95.80.9 | 91.80.8 | 91.80.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| 1 SUPMER (only labeled) | 87.5 | 47.0 | 83.8 | 89.9 | 75.4 | 79.6 | 67.9 | 56.6 | 59.0 | 54.6 | 69.2 | 69.5 | 70.0 |
| 2 SUPMER (only unlabeled) | 87.3 | 46.7 | 84.0 | 89.3 | 79.6 | 80.2 | 72.4 | 57.3 | 61.7 | 54.8 | 71.3 | 70.5 | 71.3 |
| 3 PPT (labeled + unlabeled) | 84.7 | 45.0 | 82.4 | 87.8 | 67.2 | 77.4 | 64.3 | 55.3 | 61.6 | 53.9 | 68.9 | 67.7 | 68.0 |
| 4 MetaPT (labeled + unlabeled) | 86.1 | 46.3 | 83.7 | 89.4 | 73.8 | 80.1 | 67.2 | 57.4 | 60.0 | 54.3 | 70.1 | 69.9 | 69.9 |
| 5 SUPMER (labeled + unlabeled) | 89.1 | 48.2 | 85.7 | 90.8 | 79.3 | 83.4 | 73.2 | 58.8 | 63.7 | 55.3 | 70.5 | 71.5 | 72.5 |
| Methods | Few-shot Learning | DG | |
| 1 | only sp | 66.7 | 85.7 |
| 2 | only mc | 66.9 | 85.7 |
| 3 | only ss | 67.4 | 86.7 |
| 4 | w/o ta | 68.6 | 85.3 |
| 5 | w/o curriculum | 69.9 | 87.0 |
| 6 | w/o mgr | 69.4 | 86.1 |
| 7 | SUPMER | 71.3 | 88.0 |
| Dataset | Task | #Train | #Test | K |
| SST-2 | Sentiment analysis | 6920 | 872 | 2 |
| SST-5 | Sentiment analysis | 8544 | 2210 | 5 |
| MR | Sentiment analysis | 8662 | 2000 | 2 |
| CR | Sentiment analysis | 1774 | 2000 | 2 |
| SUBJ | Subjectivity classification | 8000 | 2000 | 2 |
| TREC | Question classification | 5452 | 500 | 6 |
| CB | Natural language inference | 250 | 56 | 3 |
| RTE | Natural language inference | 2490 | 277 | 2 |
| QNLI | Question answering | 104743 | 5463 | 2 |
| WiC | Word sense disambiguation | 5428 | 638 | 2 |
| MRPC | Paraphrase detection | 3668 | 408 | 2 |
| QQP | Paraphrase detection | 363846 | 40430 | 2 |
| Hyper-parameter | Value |
| Number of clusters for each task format | 250 |
| Tasks per batch | 4 |
| Size of support set per task | 32 |
| Size of query set per task | 32 |
| Optimizer | Adam |
| Inner loop learning rate | 0.1 |
| Outer loop learning rate | 0.1 |
| Learning rate for φ | 1e-4 |
| Scheduler | Linear scheduler |
| Warm-up steps | 0 |
| Max training steps | 100,000 |
| Validation steps | 2,000 |
| Max sequence length | 512 |
| λ | 1.0 |
| m | 2.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP |
| only sp | 83.61.5 | 42.62.2 | 81.71.8 | 86.00.6 | 65.82.8 | 71.26.6 | 64.62.1 | 57.02.5 | 58.43.3 | 53.61.5 | 69.91.3 | 66.01.0 |
| only mc | 83.41.4 | 44.51.9 | 79.35.1 | 88.30.5 | 70.54.7 | 66.41.4 | 65.93.1 | 54.91.3 | 58.71.7 | 54.21.8 | 68.80.8 | 67.61.3 |
| only ss | 84.51.5 | 45.02.0 | 81.50.7 | 88.40.5 | 73.33.1 | 79.15.8 | 62.12.6 | 53.91.0 | 56.51.4 | 53.31.3 | 67.71.3 | 63.71.7 |
| w/o ta | 84.71.0 | 40.13.3 | 81.91.8 | 87.20.8 | 73.62.8 | 78.83.7 | 66.41.9 | 56.60.9 | 59.41.8 | 54.32.4 | 69.51.1 | 70.21.1 |
| w/o curriculum | 86.80.8 | 40.82.2 | 82.31.3 | 88.40.9 | 74.83.1 | 79.71.6 | 71.02.1 | 56.50.8 | 62.61.4 | 55.41.1 | 69.70.8 | 71.31.2 |
| w/o mgr | 85.01.3 | 44.51.1 | 82.80.7 | 88.00.5 | 76.01.7 | 79.55.0 | 67.11.6 | 56.80.8 | 58.92.4 | 54.42.0 | 70.01.0 | 70.30.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 |
| Method | Source | Target | ||||
| A | B | D | E | K | R | |
| only sp | 83.41.1 | 82.11.4 | 83.00.7 | 88.51.2 | 88.90.8 | 88.10.7 |
| only mc | 84.00.6 | 82.31.2 | 81.50.9 | 88.51.0 | 89.30.7 | 88.80.7 |
| only ss | 83.60.7 | 84.70.8 | 84.20.6 | 88.90.9 | 89.70.9 | 89.00.3 |
| w/o ta | 83.40.8 | 82.01.4 | 81.71.5 | 87.80.8 | 88.20.6 | 88.60.6 |
| w/o curriculum | 84.00.5 | 84.70.8 | 83.90.6 | 89.60.5 | 90.30.8 | 89.71.1 |
| w/o mgr | 83.80.4 | 83.40.5 | 83.30.5 | 88.10.6 | 89.20.8 | 88.90.4 |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.7 | 91.10.5 | 90.40.4 |
| Method | Source | Target | AVG | |||
| A | B | D | E | K | ||
| 1 SUPMER (only labeled) | 86.4 | 84.7 | 84.8 | 90.0 | 90.7 | 87.3 |
| 2 SUPMER (only unlabeled) | 85.7 | 85.3 | 85.1 | 90.3 | 91.1 | 87.5 |
| 3 PPT (labeled + unlabeled) | 83.1 | 79.0 | 84.4 | 89.3 | 90.6 | 85.3 |
| 4 MetaPT (labeled + unlabeled) | 86.3 | 85.3 | 86.7 | 90.1 | 91.4 | 88.0 |
| 5 SUPMER (labeled + unlabeled) | 86.6 | 88.6 | 88.5 | 92.7 | 93.7 | 90.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Prefix Tuning | 78.21.7 | 38.02.5 | 74.03.2 | 84.53.3 | 66.23.9 | 70.12.9 | 70.42.6 | 54.42.0 | 55.74.3 | 53.41.3 | 68.11.2 | 64.91.3 | 64.82.5 |
| P-Tuning-v2 | 83.10.8 | 41.71.5 | 82.30.8 | 88.70.5 | 74.82.4 | 78.12.1 | 67.55.4 | 53.91.3 | 61.41.6 | 54.52.2 | 70.71.1 | 68.01.1 | 68.71.7 |
| FT | 83.61.7 | 41.22.6 | 81.70.9 | 88.30.9 | 80.01.4 | 79.81.7 | 71.91.5 | 56.92.1 | 62.30.6 | 54.61.6 | 70.20.7 | 69.51.0 | 70.01.4 |
| PT | 71.93.4 | 37.32.8 | 73.24.5 | 84.43.5 | 61.53.3 | 65.34.2 | 58.96.1 | 53.22.8 | 55.24.8 | 53.12.4 | 66.65.3 | 63.02.6 | 62.04.0 |
| PPT | 81.22.0 | 40.25.4 | 81.20.7 | 83.67.3 | 66.83.7 | 73.42.4 | 60.77.7 | 55.41.2 | 60.43.9 | 53.61.3 | 68.00.8 | 63.10.7 | 65.63.1 |
| Unified-PPT | 76.87.7 | 44.71.7 | 79.03.3 | 87.70.6 | 64.25.8 | 68.42.9 | 63.92.9 | 54.41.1 | 56.72.4 | 54.51.6 | 67.81.1 | 67.62.9 | 65.52.8 |
| MetaPT | 85.71.3 | 45.31.1 | 82.54.5 | 88.50.3 | 73.23.7 | 78.72.2 | 65.42.4 | 56.11.7 | 58.23.1 | 54.11.3 | 69.60.6 | 68.80.9 | 68.81.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 | 71.30.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| Zero-shot Inference | 89.1 | 52.3 | 83.3 | 80.6 | 57.4 | 87.2 | 76.8 | 75.8 | 85.0 | 50.5 | 77.2 | 77.5 | 74.4 |
| Few-shot Inference | 93.2 | 53.3 | 88.5 | 87.8 | 58.6 | 91.6 | 83.9 | 79.1 | 86.9 | 64.3 | 79.9 | 81.0 | 79.0 |
| Prefix Tuning | 88.62.1 | 45.42.7 | 88.72.2 | 89.11.6 | 84.63.9 | 90.62.4 | 65.46.4 | 73.42.6 | 72.34.1 | 55.51.8 | 73.92.6 | 75.02.3 | 75.22.9 |
| P-Tuning-v2 | 91.21.0 | 53.51.1 | 90.00.8 | 89.81.1 | 87.73.2 | 92.30.9 | 69.12.9 | 62.34.6 | 76.83.6 | 63.12.0 | 77.21.7 | 77.31.6 | 77.52.0 |
| FT | 92.91.4 | 53.61.9 | 89.61.3 | 91.30.5 | 88.70.9 | 93.62.0 | 77.81.6 | 76.01.7 | 84.41.5 | 63.01.8 | 77.71.3 | 82.92.9 | 81.01.6 |
| PT | 88.29.2 | 45.73.4 | 85.65.6 | 88.65.1 | 81.65.1 | 85.15.6 | 65.47.2 | 61.64.0 | 69.35.9 | 54.81.1 | 71.12.7 | 70.28.6 | 72.35.3 |
| PPT | 91.02.6 | 49.12.3 | 88.81.3 | 88.03.0 | 87.61.9 | 90.81.7 | 67.14.8 | 67.12.2 | 81.51.1 | 56.11.3 | 75.82.2 | 71.63.1 | 76.22.3 |
| Unified-PPT | 89.98.3 | 47.42.8 | 89.70.9 | 89.01.5 | 86.44.3 | 88.43.2 | 69.33.5 | 63.21.1 | 74.93.2 | 59.91.7 | 74.02.4 | 76.32.0 | 75.72.9 |
| MetaPT | 93.70.8 | 52.51.3 | 90.70.7 | 89.70.6 | 88.12.4 | 91.81.8 | 72.12.2 | 74.11.3 | 77.52.5 | 58.51.6 | 76.61.8 | 81.42.4 | 78.91.6 |
| SUPMER | 95.50.4 | 55.30.7 | 91.40.5 | 90.70.7 | 90.30.8 | 93.01.5 | 87.61.5 | 81.41.0 | 88.30.6 | 65.01.7 | 78.10.8 | 85.10.4 | 83.50.9 |
| Model: T5-base (220M) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| Prefix Tuning | 78.12.2 | 82.71.0 | 81.51.2 | 84.40.6 | 82.72.6 | 84.92.2 | 82.41.6 |
| P-Tuning-v2 | 84.00.8 | 83.60.9 | 82.41.7 | 85.90.9 | 84.21.5 | 83.82.7 | 84.01.4 |
| FT | 84.40.2 | 83.90.6 | 81.00.6 | 84.10.6 | 85.00.7 | 85.30.6 | 84.00.6 |
| PT | 79.82.5 | 75.32.2 | 76.03.2 | 79.62.0 | 79.81.9 | 83.01.8 | 78.92.3 |
| PPT | 81.91.4 | 77.93.5 | 83.61.3 | 88.41.0 | 89.11.6 | 87.81.5 | 84.81.7 |
| Unified-PPT | 80.93.0 | 82.12.0 | 76.84.0 | 81.04.5 | 82.23.9 | 84.93.3 | 81.33.5 |
| MetaPT | 86.10.3 | 84.30.8 | 84.20.7 | 89.40.9 | 90.30.6 | - | (86.90.7) |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.6 | 91.10.5 | 90.40.4 | 88.00.5 |
| Model: Flan-T5-XL (3B) | |||||||
| Method | Source | Target | AVG | ||||
| A | B | D | E | K | R | ||
| zero-shot inference | 77.8 | 84.6 | 86.2 | 86.8 | 88.6 | 87.8 | 85.3 |
| few-shot inference | 82.5 | 90.3 | 89.7 | 92.3 | 92.2 | 89.2 | 89.4 |
| Prefix Tuning | 83.01.1 | 85.31.9 | 83.41.1 | 87.31.5 | 86.42.4 | 89.81.0 | 85.91.5 |
| P-Tuning-v2 | 85.60.4 | 86.72.3 | 86.81.8 | 88.62.1 | 90.22.3 | 88.92.6 | 87.81.9 |
| FT | 86.20.7 | 88.82.2 | 84.61.3 | 87.21.3 | 89.81.1 | 90.70.9 | 87.91.3 |
| PT | 82.61.2 | 79.03.9 | 82.51.7 | 84.22.3 | 84.52.5 | 84.82.1 | 82.92.3 |
| PPT | 85.20.9 | 83.23.3 | 89.71.8 | 92.41.7 | 93.61.5 | 89.61.1 | 89.01.7 |
| Unified-PPT | 83.00.8 | 82.53.0 | 82.04.8 | 86.03.0 | 86.62.2 | 85.62.0 | 84.32.6 |
| MetaPT | 87.10.6 | 87.42.7 | 89.51.7 | 93.80.6 | 94.31.1 | - | (90.41.3) |
| SUPMER | 87.00.2 | 89.81.4 | 91.11.2 | 95.10.8 | 95.80.9 | 91.80.8 | 91.80.9 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP | AVG |
| 1 SUPMER (only labeled) | 87.5 | 47.0 | 83.8 | 89.9 | 75.4 | 79.6 | 67.9 | 56.6 | 59.0 | 54.6 | 69.2 | 69.5 | 70.0 |
| 2 SUPMER (only unlabeled) | 87.3 | 46.7 | 84.0 | 89.3 | 79.6 | 80.2 | 72.4 | 57.3 | 61.7 | 54.8 | 71.3 | 70.5 | 71.3 |
| 3 PPT (labeled + unlabeled) | 84.7 | 45.0 | 82.4 | 87.8 | 67.2 | 77.4 | 64.3 | 55.3 | 61.6 | 53.9 | 68.9 | 67.7 | 68.0 |
| 4 MetaPT (labeled + unlabeled) | 86.1 | 46.3 | 83.7 | 89.4 | 73.8 | 80.1 | 67.2 | 57.4 | 60.0 | 54.3 | 70.1 | 69.9 | 69.9 |
| 5 SUPMER (labeled + unlabeled) | 89.1 | 48.2 | 85.7 | 90.8 | 79.3 | 83.4 | 73.2 | 58.8 | 63.7 | 55.3 | 70.5 | 71.5 | 72.5 |
| Methods | Few-shot Learning | DG | |
| 1 | only sp | 66.7 | 85.7 |
| 2 | only mc | 66.9 | 85.7 |
| 3 | only ss | 67.4 | 86.7 |
| 4 | w/o ta | 68.6 | 85.3 |
| 5 | w/o curriculum | 69.9 | 87.0 |
| 6 | w/o mgr | 69.4 | 86.1 |
| 7 | SUPMER | 71.3 | 88.0 |
| Hyper-parameter | Value |
| Number of clusters for each task format | 250 |
| Tasks per batch | 4 |
| Size of support set per task | 32 |
| Size of query set per task | 32 |
| Optimizer | Adam |
| Inner loop learning rate | 0.1 |
| Outer loop learning rate | 0.1 |
| Learning rate for φ | 1e-4 |
| Scheduler | Linear scheduler |
| Warm-up steps | 0 |
| Max training steps | 100,000 |
| Validation steps | 2,000 |
| Max sequence length | 512 |
| λ | 1.0 |
| m | 2.0 |
| Methods | SST-2 | SST-5 | MR | CR | SUBJ | TREC | CB | RTE | QNLI | WiC | MRPC | QQP |
| only sp | 83.61.5 | 42.62.2 | 81.71.8 | 86.00.6 | 65.82.8 | 71.26.6 | 64.62.1 | 57.02.5 | 58.43.3 | 53.61.5 | 69.91.3 | 66.01.0 |
| only mc | 83.41.4 | 44.51.9 | 79.35.1 | 88.30.5 | 70.54.7 | 66.41.4 | 65.93.1 | 54.91.3 | 58.71.7 | 54.21.8 | 68.80.8 | 67.61.3 |
| only ss | 84.51.5 | 45.02.0 | 81.50.7 | 88.40.5 | 73.33.1 | 79.15.8 | 62.12.6 | 53.91.0 | 56.51.4 | 53.31.3 | 67.71.3 | 63.71.7 |
| w/o ta | 84.71.0 | 40.13.3 | 81.91.8 | 87.20.8 | 73.62.8 | 78.83.7 | 66.41.9 | 56.60.9 | 59.41.8 | 54.32.4 | 69.51.1 | 70.21.1 |
| w/o curriculum | 86.80.8 | 40.82.2 | 82.31.3 | 88.40.9 | 74.83.1 | 79.71.6 | 71.02.1 | 56.50.8 | 62.61.4 | 55.41.1 | 69.70.8 | 71.31.2 |
| w/o mgr | 85.01.3 | 44.51.1 | 82.80.7 | 88.00.5 | 76.01.7 | 79.55.0 | 67.11.6 | 56.80.8 | 58.92.4 | 54.42.0 | 70.01.0 | 70.30.9 |
| SUPMER | 87.30.5 | 46.70.6 | 84.00.6 | 89.30.3 | 79.62.2 | 80.20.9 | 72.41.4 | 57.31.0 | 61.71.0 | 54.81.2 | 71.30.5 | 70.51.0 |
| Method | Source | Target | ||||
| A | B | D | E | K | R | |
| only sp | 83.41.1 | 82.11.4 | 83.00.7 | 88.51.2 | 88.90.8 | 88.10.7 |
| only mc | 84.00.6 | 82.31.2 | 81.50.9 | 88.51.0 | 89.30.7 | 88.80.7 |
| only ss | 83.60.7 | 84.70.8 | 84.20.6 | 88.90.9 | 89.70.9 | 89.00.3 |
| w/o ta | 83.40.8 | 82.01.4 | 81.71.5 | 87.80.8 | 88.20.6 | 88.60.6 |
| w/o curriculum | 84.00.5 | 84.70.8 | 83.90.6 | 89.60.5 | 90.30.8 | 89.71.1 |
| w/o mgr | 83.80.4 | 83.40.5 | 83.30.5 | 88.10.6 | 89.20.8 | 88.90.4 |
| SUPMER | 85.70.5 | 85.30.6 | 85.10.4 | 90.30.7 | 91.10.5 | 90.40.4 |
| Method | Source | Target | AVG | |||
| A | B | D | E | K | ||
| 1 SUPMER (only labeled) | 86.4 | 84.7 | 84.8 | 90.0 | 90.7 | 87.3 |
| 2 SUPMER (only unlabeled) | 85.7 | 85.3 | 85.1 | 90.3 | 91.1 | 87.5 |
| 3 PPT (labeled + unlabeled) | 83.1 | 79.0 | 84.4 | 89.3 | 90.6 | 85.3 |
| 4 MetaPT (labeled + unlabeled) | 86.3 | 85.3 | 86.7 | 90.1 | 91.4 | 88.0 |
| 5 SUPMER (labeled + unlabeled) | 86.6 | 88.6 | 88.5 | 92.7 | 93.7 | 90.0 |
| GloVe (glove.42B) | fastText (wiki-news) | |||||||||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | W(s) | W(r) | RW | ME | SE | SL | SV | |
| Raw | .70 | .57 | .39 | .74 | .57 | .37 | .22 | .82 | .62 | .51 | .80 | .64 | .44 | .35 |
| SelfExtro, k=100 | .80 | .73 | .50 | .84 | .66 | .51 | .39 | .81 | .70 | .56 | .82 | .66 | .50 | .41 |
| SelfExtro, k=200 | .79 | .72 | .49 | .84 | .66 | .51 | .39 | .81 | .70 | .57 | .82 | .65 | .51 | .42 |
| SelfExtro, k=300 | .78 | .69 | .48 | .84 | .64 | .50 | .37 | .81 | .71 | .57 | .83 | .64 | .51 | .42 |
| +WordNet(=Extro) | .80 | .74 | .49 | .83 | .66 | .49 | .36 | .78 | .67 | .51 | .80 | .63 | .50 | .40 |
| vecmap | .67 | .53 | .31 | .72 | .52 | .35 | .20 | .81 | .62 | .49 | .80 | .59 | .44 | .34 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| love | Vanilla | loved(.7745), i(.7338), loves(.7311), know(.7286), loving(.7263), really(.7196), always(.7193), want(.7192), hope(.7127), think(.7110) |
| +SelfExtro | loved(.7152), loving(.6734), loves(.6489), adore(.6348), passion(.6333), luv(.6326), hope(.6256), i(.6250), want(.6209), hate(.6181) | |
| hate | Vanilla | dont(.7318), stupid(.7193), hates(.7190), think(.7063), why(.6943), love(.6928), hating(.6927), hated(.6861), shit(.6847), know(.6825) |
| +SelfExtro | hating(.6707), hatred(.6650), dont(.6580), hates(.6529), dislike(.6447), despise(.6309), hated(.6306), stupid(.6266), think(.6249), love(.6181) | |
| forever | Vanilla | alive(.6478), gone(.6450), love(.6381), never(.6267), again(.6249), yours(.6238), life(.6171), alone(.6153), anymore(.6129), always(.6093) |
| +SelfExtro | eternally(.6297), alive(.5814), yours(.5717), life(.5532), anymore(.5524), eternity(.5488), eternal(.5474), gone(.5447), permanently(.5446), again(.5414) | |
| life | Vanilla | lives(.8053), living(.7134), things(.6869), way(.6852), mind(.6844), what(.6779), much(.6723), love(.6719), because(.6716), work(.6706) |
| +SelfExtro | lives(.7801), living(.6538), journey(.6048), everyday(.6038), mind(.5922), things(.5920), lifetime(.5892), happiness(.5749), love(.5743), way(.5728) |
| CZ | IT(s/r) | RU(s/r) | GE(s/r) | ||||
| fastText | .240 | .377 | .188 | .428 | .275 | .335 | .169 |
| k=100 | .450 | .453 | .370 | .525 | .411 | .425 | .261 |
| k=200 | .443 | .444 | .359 | .545 | .426 | .431 | .261 |
| k=300 | .451 | .456 | .359 | .539 | .426 | .449 | .271 |
| CY | ET | FI | FR | HE | PL | RU | ES | |
| fT | .187 | .213 | .455 | .314 | .371 | .283 | .305 | .385 |
| k=100 | .201 | .223 | .475 | .340 | .400 | .306 | .334 | .425 |
| k=200 | .209 | .215 | .463 | .339 | .399 | .299 | .306 | .423 |
| k=300 | .214 | .219 | .464 | .338 | .402 | .305 | .308 | .421 |
| Model(Resource) | JGA with WordNet | JGA with A-R Lexicon |
| GloVe | .798(.03) | |
| Retrofitting(Syn) | .792(.03) | .793(.01) |
| Paragraph(Syn) | .670(.05) | .657(.03) |
| Extro(Syn) | .821(.03) | .820(.01) |
| Counter-fit(Syn+Ant) | .625(.01) | .630(.03) |
| Att-Repel(Syn+Ant) | .671(.04) | .675(.03) |
| vecmap(-) | .772(.01) | |
| SelfExtro(-) | .825(.02) |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31±0.12 | 71.97±0.76 | 49.35±0.51 | 61.55±0.35 | 90.66±0.79 | 87.47±1.93 |
| Retrofit(Syn) | 81.92±5.11 | 44.72±2.77 | 22.31±0.63 | 48.87±0.40 | 82.04±1.49 | 63.40±3.53 |
| Paragram(Syn) | 86.88±0.36 | 62.37±0.36 | 40.25±0.19 | 50.04±0.20 | 80.16±0.14 | 74.77±0.43 |
| Extro(Syn) | 98.52±0.05 | 72.94±0.10 | 50.01±0.39 | 62.70±0.06 | 91.36±0.29 | 89.40±0.20 |
| Counter-fit(Syn+Ant) | 73.26±1.29 | 59.70±0.19 | 37.80±0.08 | 47.97±0.45 | 76.47±1.45 | 66.54±0.48 |
| Att-Repel(Syn+Ant) | 87.52±0.12 | 62.51±0.41 | 39.75±1.15 | 50.08±0.29 | 79.80±0.93 | 74.71±0.37 |
| PostSpec(Syn+Ant) | 76.78±0.79 | 59.45±0.15 | 37.16±0.35 | 49.16±0.64 | 76.06±0.62 | 66.11±0.79 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.44±0.05 | 72.41±0.26 | 49.79±0.37 | 62.97±0.18 | 90.93±0.11 | 89.36±0.41 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61±0.05 | 73.45±0.70 | 49.30±0.61 | 63.01±0.42 | 91.54±0.45 | 88.82±0.55 |
| Retrofit(Syn) | 98.16±0.05 | 67.15±1.73 | 43.14±1.31 | 61.74±0.33 | 89.21±1.61 | 82.77±1.27 |
| Paragram(Syn) | 98.14±0.04 | 66.32±0.27 | 42.85±0.43 | 61.96±0.22 | 89.91±0.19 | 83.14±1.15 |
| Extro(Syn) | 98.61±0.06 | 73.50±0.19 | 50.43±0.30 | 63.22±0.19 | 91.78±0.29 | 89.74±0.28 |
| Counter-fit(Syn+Ant) | 98.07±0.10 | 63.74±0.53 | 40.83±0.20 | 61.89±0.17 | 90.09±1.23 | 83.78±1.84 |
| Att-Repel(Syn+Ant) | 98.11±0.07 | 66.05±0.42 | 42.68±0.43 | 61.80±0.21 | 90.05±0.18 | 83.74±1.30 |
| PostSpec(Syn+Ant) | 98.10±0.05 | 63.88±0.36 | 40.49±0.51 | 61.84±0.13 | 90.48±0.38 | 84.54±0.38 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.58±0.02 | 73.02±0.15 | 50.05±0.23 | 63.25±0.25 | 91.61±0.08 | 89.98±0.26 |
| Threshold | Grouped Words |
| ≥ .90 | points225 points203 points166 points253··· |
| 17:20:18 09:22:30 10:22:49 07:33:04··· | |
| review10/15/2012 review3/31/2013 | |
| review11/30/2012 review12/18/2012··· | |
| ≥ .99 | !!! !! |
| january february | |
| range emailicon linkicon |
| DBpedia | Yahoo(Up) | Yahoo(Low) | AGNews | Yelp | IMDB | |
| #Train | 560,000 | 133,703 | 133,703 | 120,000 | 650,000 | 25,000 |
| #Test | 70,000 | 23,595 | 23,595 | 7,600 | 50,000 | 25,000 |
| #Class | 14 | 17 | 280 | 4 | 5 | 2 |
| #Vocab | 626,717 | 154,142 | 154,142 | 66,049 | 198,625 | 47,113 |
| Word Similarity Task | |||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | |
| GloVe | .70 | .57 | .39 | .74 | .57 | .37 | .28 |
| threshold= .95 .72 | .60 | .39 | .76 | .58 | .42 | .28 | |
| threshold= .90 .78 | .69 | .48 | .84 | .64 | .50 | .37 | |
| threshold= .50 .71 | .59 | .38 | .75 | .56 | .39 | .23 |
| Model(threshold) | DST |
| GloVe | .798 |
| SelfExtro(threshold=.95) | .800 |
| SelfExtro(threshold=.90) | .825 |
| SelfExtro(threshold=.50) | .811 |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31 | 71.97 | 49.35 | 61.55 | 90.66 | 87.47 |
| SelfExtro(thr=.95) | 98.35 | 71.61 | 49.07 | 61.01 | 90.78 | 87.23 |
| SelfExtro(thr=.90) | 98.44 | 72.41 | 49.79 | 62.97 | 90.93 | 89.36 |
| SelfExtro(thr=.50) | 98.49 | 72.74 | 50.80 | 62.10 | 90.94 | 88.33 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61 | 73.45 | 49.30 | 63.01 | 91.54 | 88.82 |
| SelfExtro(thr=.95) | 98.42 | 72.08 | 49.17 | 61.83 | 91.03 | 86.80 |
| SelfExtro(thr=.90) | 98.58 | 73.02 | 50.05 | 63.25 | 91.61 | 89.98 |
| SelfExtro(thr=.50) | 98.58 | 73.35 | 50.66 | 62.64 | 91.67 | 88.64 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| soo | Vanilla | sooo(.8394), soooo(.7938), soooo(.7715), soooo(.7359), soooo(.6844), haha(.6574), hahah(.6320), damn(.6247), omg(.6244), hahaha(.6219) |
| +SelfExtro | sooo(.8196), soooo(.7743), soooo(.7576), soooo(.7304), soooo(.6852), soooo(.6342), soooo(.6314), soooo(.6003) tooo(.5967), soooo(.00000000(.5869) | |
| elaborate | Vanilla | intricate(.7244), elaborately(.6238), extravagant(.6223), complicated(.6089), lavish(.5714), formal(.5643), sophisticated(.5639), detailed(.5623), ornate(.5619), simple(.5566) |
| +SelfExtro | intricate(.7002), elaborately(.6329), extravagant(.6109), complicated(.5651), ornate(.5586), lavish(.5533), grandiose(.5452), formal(.5326), detailed(.5235), fanciful(.5225) | |
| gratitude | Vanilla | thankfulness(.7113), generosity(.6967), kindness(.6917), appreciation(.6860) compassion(.6695), admiration(.6672), grateful(.6420), thankful(.6385), heartfelt(.6368), sympathy(.6211) |
| +SelfExtro | thankfulness(.7648), generosity(.7001), kindness(.6975), appreciation(.6934), admiration(.6830), compassion(.6738), gratefulness(.6444), reverence(.6361), thankful(.6313), sympathy(.6304) | |
| jubilate | Vanilla | exultate(.5262), deum(.4763), exultate(.4545), motet(.4054), excelsis(.3777), deo(.3721), cantata(.3651), alleluia(.3625), stabat(.3558), laudamus(.3485) |
| +SelfExtro | exultate(.5111), exultate(.4449), deum(.4269), motet(.4093), deo(.3703), choir(.3670), excelsis(.3626), mozart(.3594), stabat(.3591), cantata(.3532) | |
| elated | Vanilla | overjoyed(.7712), ecstatic(.7045), thrilled(.6576), exhilarated(.6533), gratified(.6431), enthused(.6214), saddened(.6075), flabbergasted(.6060), disheartened(.5989), giddy(.5984) |
| +SelfExtro | overjoyed(.8018), ecstatic(.7424), thrilled(.7320), excited(.6894), exhilarated(.6758), delighted(.6733), gratified(.6690), relieved(.6569), enthused(.6564), dismayed(.6342) | |
| mono-saccharide | Vanilla | disaccharide(.6470), monosaccharides(.5152), oligosaccharide(.4911), galactose(.4686), 5-carbon(.4609), n-acetylgucosamine(.4282), saccharide(.4223), sucrose(.4215), mannose(.4196), disaccharides(.4107) |
| +SelfExtro | disaccharide(.6329), monosaccharides(.4888), oligosaccharide(.4844), 5-carbon(.4617), galactose(.4414), n-acetylgucosamine(.4227), sucrose(.4206), disaccharides(.4050), carbohydrate(.4045), moieties(.4033) | |
| outlook | Vanilla | excel(.5686), recovery(.5457), export(.5368), forecast(.5229), forecasts(.5060), contacts(.5025), powerpoint(.5025), exchange(.5011), microsoft(.4944), import(.4816) |
| +SelfExtro | excel(.5145), mapssevere(.5054), icalandargoogle(.4955), export(.4883), contacts(.4877), recovery(.4753), advisorieshourly(.4673), mailbox(.4661), vcard(.4612), thunderbird(.4599) | |
| vain | Vanilla | foolish(.6229), futile(.6104), selfish(.5340), vainly(.5242), fruitless(.5133), thy(.4981), arrogant(.4976), useless(.4844), righteous(.4833), thou(.4830) |
| +SelfExtro | futil(.6273), foolish(.6079), vainly(.5454), fruitless(.5338), selfish(.5143), useless(.5044), pointless(.4910), hopeless(.4846), attempt(.4797), presumptuous(.4780) | |
| prioritize | Vanilla | prioritise(.8064), prioritizing(.7426), priorities(.6323), prioritising(.5995), evaluate(.5950), strategize(.5933), proactively(.5862), analyze(.5800), assess(.5764), identify(.5633) |
| +SelfExtro | prioritise(.8322), prioritizing(.7415), priorities(.6390), prioritising(.6225), proactively(.6075), strategize(.5990), evaluate(.5946), analyze(.5919), prioritized(.5880), manage(.5783) | |
| nomad | Vanilla | nomads(.4911), drifter(.4735), nomadic(.4574), vagabond(.4233), zen(.4161), muvo2(.4054), gypsy(.4054), jukebox(.3997), vulcan(.3906), adventurer(.3869) |
| +SelfExtro | nomads(.5254), drifter(.4916), nomadic(.4841), muvo2(.4631), vagabond(.4422), traveler(.4246), traveller(.4216), zen(.4150), jukebox(.4110), wanderer(.4073) | |
| junction | Vanilla | junctions(.5729), near(.4985), road(.4936), intersection(.4923), creek(.4911), highway(.4878), exit(.4853), bridge(.4811), rd(.4779), jct(.4772) |
| +SelfExtro | junctions(.5822), jct(.5099), highway(.4787), motorway(.4781), road(.4744), exit(.4738).hwy(.4710), rd(.4667), intersection(.4628), heafford(.4596) | |
| moss | Vanilla | peat(.5360), brady(.5090), reed(.4998), brown(.4963), fern(.4945), green(.4920), woods(.4872), pine(.4856), jones(.4840), campbell(.4824), |
| +SelfExtro | lichen(.5628), sphagnum(.5463), spagnum(.5457), peat(.5323), lichens(.5264), mosses(.5224), fern(.4934), welker(.4852), pine(.4846), ferns(.4805), |
| GloVe (glove.42B) | fastText (wiki-news) | |||||||||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | W(s) | W(r) | RW | ME | SE | SL | SV | |
| Raw | .70 | .57 | .39 | .74 | .57 | .37 | .22 | .82 | .62 | .51 | .80 | .64 | .44 | .35 |
| SelfExtro, k=100 | .80 | .73 | .50 | .84 | .66 | .51 | .39 | .81 | .70 | .56 | .82 | .66 | .50 | .41 |
| SelfExtro, k=200 | .79 | .72 | .49 | .84 | .66 | .51 | .39 | .81 | .70 | .57 | .82 | .65 | .51 | .42 |
| SelfExtro, k=300 | .78 | .69 | .48 | .84 | .64 | .50 | .37 | .81 | .71 | .57 | .83 | .64 | .51 | .42 |
| +WordNet(=Extro) | .80 | .74 | .49 | .83 | .66 | .49 | .36 | .78 | .67 | .51 | .80 | .63 | .50 | .40 |
| vecmap | .67 | .53 | .31 | .72 | .52 | .35 | .20 | .81 | .62 | .49 | .80 | .59 | .44 | .34 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| love | Vanilla | loved(.7745), i(.7338), loves(.7311), know(.7286), loving(.7263), really(.7196), always(.7193), want(.7192), hope(.7127), think(.7110) |
| +SelfExtro | loved(.7152), loving(.6734), loves(.6489), adore(.6348), passion(.6333), luv(.6326), hope(.6256), i(.6250), want(.6209), hate(.6181) | |
| hate | Vanilla | dont(.7318), stupid(.7193), hates(.7190), think(.7063), why(.6943), love(.6928), hating(.6927), hated(.6861), shit(.6847), know(.6825) |
| +SelfExtro | hating(.6707), hatred(.6650), dont(.6580), hates(.6529), dislike(.6447), despise(.6309), hated(.6306), stupid(.6266), think(.6249), love(.6181) | |
| forever | Vanilla | alive(.6478), gone(.6450), love(.6381), never(.6267), again(.6249), yours(.6238), life(.6171), alone(.6153), anymore(.6129), always(.6093) |
| +SelfExtro | eternally(.6297), alive(.5814), yours(.5717), life(.5532), anymore(.5524), eternity(.5488), eternal(.5474), gone(.5447), permanently(.5446), again(.5414) | |
| life | Vanilla | lives(.8053), living(.7134), things(.6869), way(.6852), mind(.6844), what(.6779), much(.6723), love(.6719), because(.6716), work(.6706) |
| +SelfExtro | lives(.7801), living(.6538), journey(.6048), everyday(.6038), mind(.5922), things(.5920), lifetime(.5892), happiness(.5749), love(.5743), way(.5728) |
| CZ | IT(s/r) | RU(s/r) | GE(s/r) | ||||
| fastText | .240 | .377 | .188 | .428 | .275 | .335 | .169 |
| k=100 | .450 | .453 | .370 | .525 | .411 | .425 | .261 |
| k=200 | .443 | .444 | .359 | .545 | .426 | .431 | .261 |
| k=300 | .451 | .456 | .359 | .539 | .426 | .449 | .271 |
| CY | ET | FI | FR | HE | PL | RU | ES | |
| fT | .187 | .213 | .455 | .314 | .371 | .283 | .305 | .385 |
| k=100 | .201 | .223 | .475 | .340 | .400 | .306 | .334 | .425 |
| k=200 | .209 | .215 | .463 | .339 | .399 | .299 | .306 | .423 |
| k=300 | .214 | .219 | .464 | .338 | .402 | .305 | .308 | .421 |
| Model(Resource) | JGA with WordNet | JGA with A-R Lexicon |
| GloVe | .798(.03) | |
| Retrofitting(Syn) | .792(.03) | .793(.01) |
| Paragraph(Syn) | .670(.05) | .657(.03) |
| Extro(Syn) | .821(.03) | .820(.01) |
| Counter-fit(Syn+Ant) | .625(.01) | .630(.03) |
| Att-Repel(Syn+Ant) | .671(.04) | .675(.03) |
| vecmap(-) | .772(.01) | |
| SelfExtro(-) | .825(.02) |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31±0.12 | 71.97±0.76 | 49.35±0.51 | 61.55±0.35 | 90.66±0.79 | 87.47±1.93 |
| Retrofit(Syn) | 81.92±5.11 | 44.72±2.77 | 22.31±0.63 | 48.87±0.40 | 82.04±1.49 | 63.40±3.53 |
| Paragram(Syn) | 86.88±0.36 | 62.37±0.36 | 40.25±0.19 | 50.04±0.20 | 80.16±0.14 | 74.77±0.43 |
| Extro(Syn) | 98.52±0.05 | 72.94±0.10 | 50.01±0.39 | 62.70±0.06 | 91.36±0.29 | 89.40±0.20 |
| Counter-fit(Syn+Ant) | 73.26±1.29 | 59.70±0.19 | 37.80±0.08 | 47.97±0.45 | 76.47±1.45 | 66.54±0.48 |
| Att-Repel(Syn+Ant) | 87.52±0.12 | 62.51±0.41 | 39.75±1.15 | 50.08±0.29 | 79.80±0.93 | 74.71±0.37 |
| PostSpec(Syn+Ant) | 76.78±0.79 | 59.45±0.15 | 37.16±0.35 | 49.16±0.64 | 76.06±0.62 | 66.11±0.79 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.44±0.05 | 72.41±0.26 | 49.79±0.37 | 62.97±0.18 | 90.93±0.11 | 89.36±0.41 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61±0.05 | 73.45±0.70 | 49.30±0.61 | 63.01±0.42 | 91.54±0.45 | 88.82±0.55 |
| Retrofit(Syn) | 98.16±0.05 | 67.15±1.73 | 43.14±1.31 | 61.74±0.33 | 89.21±1.61 | 82.77±1.27 |
| Paragram(Syn) | 98.14±0.04 | 66.32±0.27 | 42.85±0.43 | 61.96±0.22 | 89.91±0.19 | 83.14±1.15 |
| Extro(Syn) | 98.61±0.06 | 73.50±0.19 | 50.43±0.30 | 63.22±0.19 | 91.78±0.29 | 89.74±0.28 |
| Counter-fit(Syn+Ant) | 98.07±0.10 | 63.74±0.53 | 40.83±0.20 | 61.89±0.17 | 90.09±1.23 | 83.78±1.84 |
| Att-Repel(Syn+Ant) | 98.11±0.07 | 66.05±0.42 | 42.68±0.43 | 61.80±0.21 | 90.05±0.18 | 83.74±1.30 |
| PostSpec(Syn+Ant) | 98.10±0.05 | 63.88±0.36 | 40.49±0.51 | 61.84±0.13 | 90.48±0.38 | 84.54±0.38 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.58±0.02 | 73.02±0.15 | 50.05±0.23 | 63.25±0.25 | 91.61±0.08 | 89.98±0.26 |
| Threshold | Grouped Words |
| ≥ .90 | points225 points203 points166 points253··· |
| 17:20:18 09:22:30 10:22:49 07:33:04··· | |
| review10/15/2012 review3/31/2013 | |
| review11/30/2012 review12/18/2012··· | |
| ≥ .99 | !!! !! |
| january february | |
| range emailicon linkicon |
| DBpedia | Yahoo(Up) | Yahoo(Low) | AGNews | Yelp | IMDB | |
| #Train | 560,000 | 133,703 | 133,703 | 120,000 | 650,000 | 25,000 |
| #Test | 70,000 | 23,595 | 23,595 | 7,600 | 50,000 | 25,000 |
| #Class | 14 | 17 | 280 | 4 | 5 | 2 |
| #Vocab | 626,717 | 154,142 | 154,142 | 66,049 | 198,625 | 47,113 |
| Word Similarity Task | |||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | |
| GloVe | .70 | .57 | .39 | .74 | .57 | .37 | .28 |
| threshold= .95 .72 | .60 | .39 | .76 | .58 | .42 | .28 | |
| threshold= .90 .78 | .69 | .48 | .84 | .64 | .50 | .37 | |
| threshold= .50 .71 | .59 | .38 | .75 | .56 | .39 | .23 |
| Model(threshold) | DST |
| GloVe | .798 |
| SelfExtro(threshold=.95) | .800 |
| SelfExtro(threshold=.90) | .825 |
| SelfExtro(threshold=.50) | .811 |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31 | 71.97 | 49.35 | 61.55 | 90.66 | 87.47 |
| SelfExtro(thr=.95) | 98.35 | 71.61 | 49.07 | 61.01 | 90.78 | 87.23 |
| SelfExtro(thr=.90) | 98.44 | 72.41 | 49.79 | 62.97 | 90.93 | 89.36 |
| SelfExtro(thr=.50) | 98.49 | 72.74 | 50.80 | 62.10 | 90.94 | 88.33 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61 | 73.45 | 49.30 | 63.01 | 91.54 | 88.82 |
| SelfExtro(thr=.95) | 98.42 | 72.08 | 49.17 | 61.83 | 91.03 | 86.80 |
| SelfExtro(thr=.90) | 98.58 | 73.02 | 50.05 | 63.25 | 91.61 | 89.98 |
| SelfExtro(thr=.50) | 98.58 | 73.35 | 50.66 | 62.64 | 91.67 | 88.64 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| soo | Vanilla | sooo(.8394), soooo(.7938), soooo(.7715), soooo(.7359), soooo(.6844), haha(.6574), hahah(.6320), damn(.6247), omg(.6244), hahaha(.6219) |
| +SelfExtro | sooo(.8196), soooo(.7743), soooo(.7576), soooo(.7304), soooo(.6852), soooo(.6342), soooo(.6314), soooo(.6003) tooo(.5967), soooo(.00000000(.5869) | |
| elaborate | Vanilla | intricate(.7244), elaborately(.6238), extravagant(.6223), complicated(.6089), lavish(.5714), formal(.5643), sophisticated(.5639), detailed(.5623), ornate(.5619), simple(.5566) |
| +SelfExtro | intricate(.7002), elaborately(.6329), extravagant(.6109), complicated(.5651), ornate(.5586), lavish(.5533), grandiose(.5452), formal(.5326), detailed(.5235), fanciful(.5225) | |
| gratitude | Vanilla | thankfulness(.7113), generosity(.6967), kindness(.6917), appreciation(.6860) compassion(.6695), admiration(.6672), grateful(.6420), thankful(.6385), heartfelt(.6368), sympathy(.6211) |
| +SelfExtro | thankfulness(.7648), generosity(.7001), kindness(.6975), appreciation(.6934), admiration(.6830), compassion(.6738), gratefulness(.6444), reverence(.6361), thankful(.6313), sympathy(.6304) | |
| jubilate | Vanilla | exultate(.5262), deum(.4763), exultate(.4545), motet(.4054), excelsis(.3777), deo(.3721), cantata(.3651), alleluia(.3625), stabat(.3558), laudamus(.3485) |
| +SelfExtro | exultate(.5111), exultate(.4449), deum(.4269), motet(.4093), deo(.3703), choir(.3670), excelsis(.3626), mozart(.3594), stabat(.3591), cantata(.3532) | |
| elated | Vanilla | overjoyed(.7712), ecstatic(.7045), thrilled(.6576), exhilarated(.6533), gratified(.6431), enthused(.6214), saddened(.6075), flabbergasted(.6060), disheartened(.5989), giddy(.5984) |
| +SelfExtro | overjoyed(.8018), ecstatic(.7424), thrilled(.7320), excited(.6894), exhilarated(.6758), delighted(.6733), gratified(.6690), relieved(.6569), enthused(.6564), dismayed(.6342) | |
| mono-saccharide | Vanilla | disaccharide(.6470), monosaccharides(.5152), oligosaccharide(.4911), galactose(.4686), 5-carbon(.4609), n-acetylgucosamine(.4282), saccharide(.4223), sucrose(.4215), mannose(.4196), disaccharides(.4107) |
| +SelfExtro | disaccharide(.6329), monosaccharides(.4888), oligosaccharide(.4844), 5-carbon(.4617), galactose(.4414), n-acetylgucosamine(.4227), sucrose(.4206), disaccharides(.4050), carbohydrate(.4045), moieties(.4033) | |
| outlook | Vanilla | excel(.5686), recovery(.5457), export(.5368), forecast(.5229), forecasts(.5060), contacts(.5025), powerpoint(.5025), exchange(.5011), microsoft(.4944), import(.4816) |
| +SelfExtro | excel(.5145), mapssevere(.5054), icalandargoogle(.4955), export(.4883), contacts(.4877), recovery(.4753), advisorieshourly(.4673), mailbox(.4661), vcard(.4612), thunderbird(.4599) | |
| vain | Vanilla | foolish(.6229), futile(.6104), selfish(.5340), vainly(.5242), fruitless(.5133), thy(.4981), arrogant(.4976), useless(.4844), righteous(.4833), thou(.4830) |
| +SelfExtro | futil(.6273), foolish(.6079), vainly(.5454), fruitless(.5338), selfish(.5143), useless(.5044), pointless(.4910), hopeless(.4846), attempt(.4797), presumptuous(.4780) | |
| prioritize | Vanilla | prioritise(.8064), prioritizing(.7426), priorities(.6323), prioritising(.5995), evaluate(.5950), strategize(.5933), proactively(.5862), analyze(.5800), assess(.5764), identify(.5633) |
| +SelfExtro | prioritise(.8322), prioritizing(.7415), priorities(.6390), prioritising(.6225), proactively(.6075), strategize(.5990), evaluate(.5946), analyze(.5919), prioritized(.5880), manage(.5783) | |
| nomad | Vanilla | nomads(.4911), drifter(.4735), nomadic(.4574), vagabond(.4233), zen(.4161), muvo2(.4054), gypsy(.4054), jukebox(.3997), vulcan(.3906), adventurer(.3869) |
| +SelfExtro | nomads(.5254), drifter(.4916), nomadic(.4841), muvo2(.4631), vagabond(.4422), traveler(.4246), traveller(.4216), zen(.4150), jukebox(.4110), wanderer(.4073) | |
| junction | Vanilla | junctions(.5729), near(.4985), road(.4936), intersection(.4923), creek(.4911), highway(.4878), exit(.4853), bridge(.4811), rd(.4779), jct(.4772) |
| +SelfExtro | junctions(.5822), jct(.5099), highway(.4787), motorway(.4781), road(.4744), exit(.4738).hwy(.4710), rd(.4667), intersection(.4628), heafford(.4596) | |
| moss | Vanilla | peat(.5360), brady(.5090), reed(.4998), brown(.4963), fern(.4945), green(.4920), woods(.4872), pine(.4856), jones(.4840), campbell(.4824), |
| +SelfExtro | lichen(.5628), sphagnum(.5463), spagnum(.5457), peat(.5323), lichens(.5264), mosses(.5224), fern(.4934), welker(.4852), pine(.4846), ferns(.4805), |
| GloVe (glove.42B) | fastText (wiki-news) | |||||||||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | W(s) | W(r) | RW | ME | SE | SL | SV | |
| Raw | .70 | .57 | .39 | .74 | .57 | .37 | .22 | .82 | .62 | .51 | .80 | .64 | .44 | .35 |
| SelfExtro, k=100 | .80 | .73 | .50 | .84 | .66 | .51 | .39 | .81 | .70 | .56 | .82 | .66 | .50 | .41 |
| SelfExtro, k=200 | .79 | .72 | .49 | .84 | .66 | .51 | .39 | .81 | .70 | .57 | .82 | .65 | .51 | .42 |
| SelfExtro, k=300 | .78 | .69 | .48 | .84 | .64 | .50 | .37 | .81 | .71 | .57 | .83 | .64 | .51 | .42 |
| +WordNet(=Extro) | .80 | .74 | .49 | .83 | .66 | .49 | .36 | .78 | .67 | .51 | .80 | .63 | .50 | .40 |
| vecmap | .67 | .53 | .31 | .72 | .52 | .35 | .20 | .81 | .62 | .49 | .80 | .59 | .44 | .34 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| love | Vanilla | loved(.7745), i(.7338), loves(.7311), know(.7286), loving(.7263), really(.7196), always(.7193), want(.7192), hope(.7127), think(.7110) |
| +SelfExtro | loved(.7152), loving(.6734), loves(.6489), adore(.6348), passion(.6333), luv(.6326), hope(.6256), i(.6250), want(.6209), hate(.6181) | |
| hate | Vanilla | dont(.7318), stupid(.7193), hates(.7190), think(.7063), why(.6943), love(.6928), hating(.6927), hated(.6861), shit(.6847), know(.6825) |
| +SelfExtro | hating(.6707), hatred(.6650), dont(.6580), hates(.6529), dislike(.6447), despise(.6309), hated(.6306), stupid(.6266), think(.6249), love(.6181) | |
| forever | Vanilla | alive(.6478), gone(.6450), love(.6381), never(.6267), again(.6249), yours(.6238), life(.6171), alone(.6153), anymore(.6129), always(.6093) |
| +SelfExtro | eternally(.6297), alive(.5814), yours(.5717), life(.5532), anymore(.5524), eternity(.5488), eternal(.5474), gone(.5447), permanently(.5446), again(.5414) | |
| life | Vanilla | lives(.8053), living(.7134), things(.6869), way(.6852), mind(.6844), what(.6779), much(.6723), love(.6719), because(.6716), work(.6706) |
| +SelfExtro | lives(.7801), living(.6538), journey(.6048), everyday(.6038), mind(.5922), things(.5920), lifetime(.5892), happiness(.5749), love(.5743), way(.5728) |
| CZ | IT(s/r) | RU(s/r) | GE(s/r) | ||||
| fastText | .240 | .377 | .188 | .428 | .275 | .335 | .169 |
| k=100 | .450 | .453 | .370 | .525 | .411 | .425 | .261 |
| k=200 | .443 | .444 | .359 | .545 | .426 | .431 | .261 |
| k=300 | .451 | .456 | .359 | .539 | .426 | .449 | .271 |
| CY | ET | FI | FR | HE | PL | RU | ES | |
| fT | .187 | .213 | .455 | .314 | .371 | .283 | .305 | .385 |
| k=100 | .201 | .223 | .475 | .340 | .400 | .306 | .334 | .425 |
| k=200 | .209 | .215 | .463 | .339 | .399 | .299 | .306 | .423 |
| k=300 | .214 | .219 | .464 | .338 | .402 | .305 | .308 | .421 |
| Model(Resource) | JGA with WordNet | JGA with A-R Lexicon |
| GloVe | .798(.03) | |
| Retrofitting(Syn) | .792(.03) | .793(.01) |
| Paragraph(Syn) | .670(.05) | .657(.03) |
| Extro(Syn) | .821(.03) | .820(.01) |
| Counter-fit(Syn+Ant) | .625(.01) | .630(.03) |
| Att-Repel(Syn+Ant) | .671(.04) | .675(.03) |
| vecmap(-) | .772(.01) | |
| SelfExtro(-) | .825(.02) | |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31±0.12 | 71.97±0.76 | 49.35±0.51 | 61.55±0.35 | 90.66±0.79 | 87.47±1.93 |
| Retrofit(Syn) | 81.92±5.11 | 44.72±2.77 | 22.31±0.63 | 48.87±0.40 | 82.04±1.49 | 63.40±3.53 |
| Paragram(Syn) | 86.88±0.36 | 62.37±0.36 | 40.25±0.19 | 50.04±0.20 | 80.16±0.14 | 74.77±0.43 |
| Extro(Syn) | 98.52±0.05 | 72.94±0.10 | 50.01±0.39 | 62.70±0.06 | 91.36±0.29 | 89.40±0.20 |
| Counter-fit(Syn+Ant) | 73.26±1.29 | 59.70±0.19 | 37.80±0.08 | 47.97±0.45 | 76.47±1.45 | 66.54±0.48 |
| Att-Repel(Syn+Ant) | 87.52±0.12 | 62.51±0.41 | 39.75±1.15 | 50.08±0.29 | 79.80±0.93 | 74.71±0.37 |
| PostSpec(Syn+Ant) | 76.78±0.79 | 59.45±0.15 | 37.16±0.35 | 49.16±0.64 | 76.06±0.62 | 66.11±0.79 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.44±0.05 | 72.41±0.26 | 49.79±0.37 | 62.97±0.18 | 90.93±0.11 | 89.36±0.41 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61±0.05 | 73.45±0.70 | 49.30±0.61 | 63.01±0.42 | 91.54±0.45 | 88.82±0.55 |
| Retrofit(Syn) | 98.16±0.05 | 67.15±1.73 | 43.14±1.31 | 61.74±0.33 | 89.21±1.61 | 82.77±1.27 |
| Paragram(Syn) | 98.14±0.04 | 66.32±0.27 | 42.85±0.43 | 61.96±0.22 | 89.91±0.19 | 83.14±1.15 |
| Extro(Syn) | 98.61±0.06 | 73.50±0.19 | 50.43±0.30 | 63.22±0.19 | 91.78±0.29 | 89.74±0.28 |
| Counter-fit(Syn+Ant) | 98.07±0.10 | 63.74±0.53 | 40.83±0.20 | 61.89±0.17 | 90.09±1.23 | 83.78±1.84 |
| Att-Repel(Syn+Ant) | 98.11±0.07 | 66.05±0.42 | 42.68±0.43 | 61.80±0.21 | 90.05±0.18 | 83.74±1.30 |
| PostSpec(Syn+Ant) | 98.10±0.05 | 63.88±0.36 | 40.49±0.51 | 61.84±0.13 | 90.48±0.38 | 84.54±0.38 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.58±0.02 | 73.02±0.15 | 50.05±0.23 | 63.25±0.25 | 91.61±0.08 | 89.98±0.26 |
| Threshold | Grouped Words |
| ≥ .90 | points225 points203 points166 points253··· |
| 17:20:18 09:22:30 10:22:49 07:33:04··· | |
| review10/15/2012 review3/31/2013 | |
| review11/30/2012 review12/18/2012··· | |
| ≥ .99 | !!! !! |
| january february | |
| range emailicon linkicon |
| DBpedia | Yahoo(Up) | Yahoo(Low) | AGNews | Yelp | IMDB | |
| #Train | 560,000 | 133,703 | 133,703 | 120,000 | 650,000 | 25,000 |
| #Test | 70,000 | 23,595 | 23,595 | 7,600 | 50,000 | 25,000 |
| #Class | 14 | 17 | 280 | 4 | 5 | 2 |
| #Vocab | 626,717 | 154,142 | 154,142 | 66,049 | 198,625 | 47,113 |
| Word Similarity Task | |||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | |
| GloVe | .70 | .57 | .39 | .74 | .57 | .37 | .28 |
| threshold= .95 .72 | .60 | .39 | .76 | .58 | .42 | .28 | |
| threshold= .90 .78 | .69 | .48 | .84 | .64 | .50 | .37 | |
| threshold= .50 .71 | .59 | .38 | .75 | .56 | .39 | .23 | |
| Model(threshold) | DST |
| GloVe | .798 |
| SelfExtro(threshold=.95) | .800 |
| SelfExtro(threshold=.90) | .825 |
| SelfExtro(threshold=.50) | .811 |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31 | 71.97 | 49.35 | 61.55 | 90.66 | 87.47 |
| SelfExtro(thr=.95) | 98.35 | 71.61 | 49.07 | 61.01 | 90.78 | 87.23 |
| SelfExtro(thr=.90) | 98.44 | 72.41 | 49.79 | 62.97 | 90.93 | 89.36 |
| SelfExtro(thr=.50) | 98.49 | 72.74 | 50.80 | 62.10 | 90.94 | 88.33 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61 | 73.45 | 49.30 | 63.01 | 91.54 | 88.82 |
| SelfExtro(thr=.95) | 98.42 | 72.08 | 49.17 | 61.83 | 91.03 | 86.80 |
| SelfExtro(thr=.90) | 98.58 | 73.02 | 50.05 | 63.25 | 91.61 | 89.98 |
| SelfExtro(thr=.50) | 98.58 | 73.35 | 50.66 | 62.64 | 91.67 | 88.64 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| soo | Vanilla | sooo(.8394), soooo(.7938), soooo(.7715), soooo(.7359), soooo(.6844), haha(.6574), hahah(.6320), damn(.6247), omg(.6244), hahaha(.6219) |
| +SelfExtro | sooo(.8196), soooo(.7743), soooo(.7576), soooo(.7304), soooo(.6852), soooo(.6342), soooo(.6314), soooo(.6003) tooo(.5967), soooo(.00000000(.5869) | |
| elaborate | Vanilla | intricate(.7244), elaborately(.6238), extravagant(.6223), complicated(.6089), lavish(.5714), formal(.5643), sophisticated(.5639), detailed(.5623), ornate(.5619), simple(.5566) |
| +SelfExtro | intricate(.7002), elaborately(.6329), extravagant(.6109), complicated(.5651), ornate(.5586), lavish(.5533), grandiose(.5452), formal(.5326), detailed(.5235), fanciful(.5225) | |
| gratitude | Vanilla | thankfulness(.7113), generosity(.6967), kindness(.6917), appreciation(.6860) compassion(.6695), admiration(.6672), grateful(.6420), thankful(.6385), heartfelt(.6368), sympathy(.6211) |
| +SelfExtro | thankfulness(.7648), generosity(.7001), kindness(.6975), appreciation(.6934), admiration(.6830), compassion(.6738), gratefulness(.6444), reverence(.6361), thankful(.6313), sympathy(.6304) | |
| jubilate | Vanilla | exultate(.5262), deum(.4763), exultate(.4545), motet(.4054), excelsis(.3777), deo(.3721), cantata(.3651), alleluia(.3625), stabat(.3558), laudamus(.3485) |
| +SelfExtro | exultate(.5111), exultate(.4449), deum(.4269), motet(.4093), deo(.3703), choir(.3670), excelsis(.3626), mozart(.3594), stabat(.3591), cantata(.3532) | |
| elated | Vanilla | overjoyed(.7712), ecstatic(.7045), thrilled(.6576), exhilarated(.6533), gratified(.6431), enthused(.6214), saddened(.6075), flabbergasted(.6060), disheartened(.5989), giddy(.5984) |
| +SelfExtro | overjoyed(.8018), ecstatic(.7424), thrilled(.7320), excited(.6894), exhilarated(.6758), delighted(.6733), gratified(.6690), relieved(.6569), enthused(.6564), dismayed(.6342) | |
| mono-saccharide | Vanilla | disaccharide(.6470), monosaccharides(.5152), oligosaccharide(.4911), galactose(.4686), 5-carbon(.4609), n-acetylgucosamine(.4282), saccharide(.4223), sucrose(.4215), mannose(.4196), disaccharides(.4107) |
| +SelfExtro | disaccharide(.6329), monosaccharides(.4888), oligosaccharide(.4844), 5-carbon(.4617), galactose(.4414), n-acetylgucosamine(.4227), sucrose(.4206), disaccharides(.4050), carbohydrate(.4045), moieties(.4033) | |
| outlook | Vanilla | excel(.5686), recovery(.5457), export(.5368), forecast(.5229), forecasts(.5060), contacts(.5025), powerpoint(.5025), exchange(.5011), microsoft(.4944), import(.4816) |
| +SelfExtro | excel(.5145), mapssevere(.5054), icalandargoogle(.4955), export(.4883), contacts(.4877), recovery(.4753), advisorieshourly(.4673), mailbox(.4661), vcard(.4612), thunderbird(.4599) | |
| vain | Vanilla | foolish(.6229), futile(.6104), selfish(.5340), vainly(.5242), fruitless(.5133), thy(.4981), arrogant(.4976), useless(.4844), righteous(.4833), thou(.4830) |
| +SelfExtro | futil(.6273), foolish(.6079), vainly(.5454), fruitless(.5338), selfish(.5143), useless(.5044), pointless(.4910), hopeless(.4846), attempt(.4797), presumptuous(.4780) | |
| prioritize | Vanilla | prioritise(.8064), prioritizing(.7426), priorities(.6323), prioritising(.5995), evaluate(.5950), strategize(.5933), proactively(.5862), analyze(.5800), assess(.5764), identify(.5633) |
| +SelfExtro | prioritise(.8322), prioritizing(.7415), priorities(.6390), prioritising(.6225), proactively(.6075), strategize(.5990), evaluate(.5946), analyze(.5919), prioritized(.5880), manage(.5783) | |
| nomad | Vanilla | nomads(.4911), drifter(.4735), nomadic(.4574), vagabond(.4233), zen(.4161), muvo2(.4054), gypsy(.4054), jukebox(.3997), vulcan(.3906), adventurer(.3869) |
| +SelfExtro | nomads(.5254), drifter(.4916), nomadic(.4841), muvo2(.4631), vagabond(.4422), traveler(.4246), traveller(.4216), zen(.4150), jukebox(.4110), wanderer(.4073) | |
| junction | Vanilla | junctions(.5729), near(.4985), road(.4936), intersection(.4923), creek(.4911), highway(.4878), exit(.4853), bridge(.4811), rd(.4779), jct(.4772) |
| +SelfExtro | junctions(.5822), jct(.5099), highway(.4787), motorway(.4781), road(.4744), exit(.4738).hwy(.4710), rd(.4667), intersection(.4628), heafford(.4596) | |
| moss | Vanilla | peat(.5360), brady(.5090), reed(.4998), brown(.4963), fern(.4945), green(.4920), woods(.4872), pine(.4856), jones(.4840), campbell(.4824), |
| +SelfExtro | lichen(.5628), sphagnum(.5463), spagnum(.5457), peat(.5323), lichens(.5264), mosses(.5224), fern(.4934), welker(.4852), pine(.4846), ferns(.4805), |
| GloVe (glove.42B) | fastText (wiki-news) | |||||||||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | W(s) | W(r) | RW | ME | SE | SL | SV | |
| Raw | .70 | .57 | .39 | .74 | .57 | .37 | .22 | .82 | .62 | .51 | .80 | .64 | .44 | .35 |
| SelfExtro, k=100 | .80 | .73 | .50 | .84 | .66 | .51 | .39 | .81 | .70 | .56 | .82 | .66 | .50 | .41 |
| SelfExtro, k=200 | .79 | .72 | .49 | .84 | .66 | .51 | .39 | .81 | .70 | .57 | .82 | .65 | .51 | .42 |
| SelfExtro, k=300 | .78 | .69 | .48 | .84 | .64 | .50 | .37 | .81 | .71 | .57 | .83 | .64 | .51 | .42 |
| +WordNet(=Extro) | .80 | .74 | .49 | .83 | .66 | .49 | .36 | .78 | .67 | .51 | .80 | .63 | .50 | .40 |
| vecmap | .67 | .53 | .31 | .72 | .52 | .35 | .20 | .81 | .62 | .49 | .80 | .59 | .44 | .34 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| love | Vanilla | loved(.7745), i(.7338), loves(.7311), know(.7286), loving(.7263), really(.7196), always(.7193), want(.7192), hope(.7127), think(.7110) |
| +SelfExtro | loved(.7152), loving(.6734), loves(.6489), adore(.6348), passion(.6333), luv(.6326), hope(.6256), i(.6250), want(.6209), hate(.6181) | |
| hate | Vanilla | dont(.7318), stupid(.7193), hates(.7190), think(.7063), why(.6943), love(.6928), hating(.6927), hated(.6861), shit(.6847), know(.6825) |
| +SelfExtro | hating(.6707), hatred(.6650), dont(.6580), hates(.6529), dislike(.6447), despise(.6309), hated(.6306), stupid(.6266), think(.6249), love(.6181) | |
| forever | Vanilla | alive(.6478), gone(.6450), love(.6381), never(.6267), again(.6249), yours(.6238), life(.6171), alone(.6153), anymore(.6129), always(.6093) |
| +SelfExtro | eternally(.6297), alive(.5814), yours(.5717), life(.5532), anymore(.5524), eternity(.5488), eternal(.5474), gone(.5447), permanently(.5446), again(.5414) | |
| life | Vanilla | lives(.8053), living(.7134), things(.6869), way(.6852), mind(.6844), what(.6779), much(.6723), love(.6719), because(.6716), work(.6706) |
| +SelfExtro | lives(.7801), living(.6538), journey(.6048), everyday(.6038), mind(.5922), things(.5920), lifetime(.5892), happiness(.5749), love(.5743), way(.5728) |
| CZ | IT(s/r) | RU(s/r) | GE(s/r) | ||||
| fastText | .240 | .377 | .188 | .428 | .275 | .335 | .169 |
| k=100 | .450 | .453 | .370 | .525 | .411 | .425 | .261 |
| k=200 | .443 | .444 | .359 | .545 | .426 | .431 | .261 |
| k=300 | .451 | .456 | .359 | .539 | .426 | .449 | .271 |
| CY | ET | FI | FR | HE | PL | RU | ES | |
| fT | .187 | .213 | .455 | .314 | .371 | .283 | .305 | .385 |
| k=100 | .201 | .223 | .475 | .340 | .400 | .306 | .334 | .425 |
| k=200 | .209 | .215 | .463 | .339 | .399 | .299 | .306 | .423 |
| k=300 | .214 | .219 | .464 | .338 | .402 | .305 | .308 | .421 |
| Model(Resource) | JGA with WordNet | JGA with A-R Lexicon |
| GloVe | .798(.03) | |
| Retrofitting(Syn) | .792(.03) | .793(.01) |
| Paragraph(Syn) | .670(.05) | .657(.03) |
| Extro(Syn) | .821(.03) | .820(.01) |
| Counter-fit(Syn+Ant) | .625(.01) | .630(.03) |
| Att-Repel(Syn+Ant) | .671(.04) | .675(.03) |
| vecmap(-) | .772(.01) | |
| SelfExtro(-) | .825(.02) |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31±0.12 | 71.97±0.76 | 49.35±0.51 | 61.55±0.35 | 90.66±0.79 | 87.47±1.93 |
| Retrofit(Syn) | 81.92±5.11 | 44.72±2.77 | 22.31±0.63 | 48.87±0.40 | 82.04±1.49 | 63.40±3.53 |
| Paragram(Syn) | 86.88±0.36 | 62.37±0.36 | 40.25±0.19 | 50.04±0.20 | 80.16±0.14 | 74.77±0.43 |
| Extro(Syn) | 98.52±0.05 | 72.94±0.10 | 50.01±0.39 | 62.70±0.06 | 91.36±0.29 | 89.40±0.20 |
| Counter-fit(Syn+Ant) | 73.26±1.29 | 59.70±0.19 | 37.80±0.08 | 47.97±0.45 | 76.47±1.45 | 66.54±0.48 |
| Att-Repel(Syn+Ant) | 87.52±0.12 | 62.51±0.41 | 39.75±1.15 | 50.08±0.29 | 79.80±0.93 | 74.71±0.37 |
| PostSpec(Syn+Ant) | 76.78±0.79 | 59.45±0.15 | 37.16±0.35 | 49.16±0.64 | 76.06±0.62 | 66.11±0.79 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.44±0.05 | 72.41±0.26 | 49.79±0.37 | 62.97±0.18 | 90.93±0.11 | 89.36±0.41 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61±0.05 | 73.45±0.70 | 49.30±0.61 | 63.01±0.42 | 91.54±0.45 | 88.82±0.55 |
| Retrofit(Syn) | 98.16±0.05 | 67.15±1.73 | 43.14±1.31 | 61.74±0.33 | 89.21±1.61 | 82.77±1.27 |
| Paragram(Syn) | 98.14±0.04 | 66.32±0.27 | 42.85±0.43 | 61.96±0.22 | 89.91±0.19 | 83.14±1.15 |
| Extro(Syn) | 98.61±0.06 | 73.50±0.19 | 50.43±0.30 | 63.22±0.19 | 91.78±0.29 | 89.74±0.28 |
| Counter-fit(Syn+Ant) | 98.07±0.10 | 63.74±0.53 | 40.83±0.20 | 61.89±0.17 | 90.09±1.23 | 83.78±1.84 |
| Att-Repel(Syn+Ant) | 98.11±0.07 | 66.05±0.42 | 42.68±0.43 | 61.80±0.21 | 90.05±0.18 | 83.74±1.30 |
| PostSpec(Syn+Ant) | 98.10±0.05 | 63.88±0.36 | 40.49±0.51 | 61.84±0.13 | 90.48±0.38 | 84.54±0.38 |
| vecmap(-) | 98.01±0.27 | 68.99±0.68 | 44.57±1.14 | 62.22±0.25 | 89.87±0.88 | 83.80±1.33 |
| SelfExtro(-) | 98.58±0.02 | 73.02±0.15 | 50.05±0.23 | 63.25±0.25 | 91.61±0.08 | 89.98±0.26 |
| Threshold | Grouped Words |
| ≥ .90 | points225 points203 points166 points253··· |
| 17:20:18 09:22:30 10:22:49 07:33:04··· | |
| review10/15/2012 review3/31/2013 | |
| review11/30/2012 review12/18/2012··· | |
| ≥ .99 | !!! !! |
| january february | |
| range emailicon linkicon |
| DBpedia | Yahoo(Up) | Yahoo(Low) | AGNews | Yelp | IMDB | |
| #Train | 560,000 | 133,703 | 133,703 | 120,000 | 650,000 | 25,000 |
| #Test | 70,000 | 23,595 | 23,595 | 7,600 | 50,000 | 25,000 |
| #Class | 14 | 17 | 280 | 4 | 5 | 2 |
| #Vocab | 626,717 | 154,142 | 154,142 | 66,049 | 198,625 | 47,113 |
| Word Similarity Task | |||||||
| W(s) | W(r) | RW | ME | SE | SL | SV | |
| GloVe | .70 | .57 | .39 | .74 | .57 | .37 | .28 |
| threshold= .95 .72 | .60 | .39 | .76 | .58 | .42 | .28 | |
| threshold= .90 .78 | .69 | .48 | .84 | .64 | .50 | .37 | |
| threshold= .50 .71 | .59 | .38 | .75 | .56 | .39 | .23 |
| Model(threshold) | DST |
| GloVe | .798 |
| SelfExtro(threshold=.95) | .800 |
| SelfExtro(threshold=.90) | .825 |
| SelfExtro(threshold=.50) | .811 |
| Freezed Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.31 | 71.97 | 49.35 | 61.55 | 90.66 | 87.47 |
| SelfExtro(thr=.95) | 98.35 | 71.61 | 49.07 | 61.01 | 90.78 | 87.23 |
| SelfExtro(thr=.90) | 98.44 | 72.41 | 49.79 | 62.97 | 90.93 | 89.36 |
| SelfExtro(thr=.50) | 98.49 | 72.74 | 50.80 | 62.10 | 90.94 | 88.33 |
| Trainable Vectors | DBpedia | Yahoo(Up) | Yahoo(Low) | Yelp | AGNews | IMDB |
| GloVe | 98.61 | 73.45 | 49.30 | 63.01 | 91.54 | 88.82 |
| SelfExtro(thr=.95) | 98.42 | 72.08 | 49.17 | 61.83 | 91.03 | 86.80 |
| SelfExtro(thr=.90) | 98.58 | 73.02 | 50.05 | 63.25 | 91.61 | 89.98 |
| SelfExtro(thr=.50) | 98.58 | 73.35 | 50.66 | 62.64 | 91.67 | 88.64 |
| Cue Word | Method | Top-10 Nearest Words(Cosine Similarity Score) |
| soo | Vanilla | sooo(.8394), soooo(.7938), soooo(.7715), soooo(.7359), soooo(.6844), haha(.6574), hahah(.6320), damn(.6247), omg(.6244), hahaha(.6219) |
| +SelfExtro | sooo(.8196), soooo(.7743), soooo(.7576), soooo(.7304), soooo(.6852), soooo(.6342), soooo(.6314), soooo(.6003) tooo(.5967), soooo(.00000000(.5869) | |
| elaborate | Vanilla | intricate(.7244), elaborately(.6238), extravagant(.6223), complicated(.6089), lavish(.5714), formal(.5643), sophisticated(.5639), detailed(.5623), ornate(.5619), simple(.5566) |
| +SelfExtro | intricate(.7002), elaborately(.6329), extravagant(.6109), complicated(.5651), ornate(.5586), lavish(.5533), grandiose(.5452), formal(.5326), detailed(.5235), fanciful(.5225) | |
| gratitude | Vanilla | thankfulness(.7113), generosity(.6967), kindness(.6917), appreciation(.6860) compassion(.6695), admiration(.6672), grateful(.6420), thankful(.6385), heartfelt(.6368), sympathy(.6211) |
| +SelfExtro | thankfulness(.7648), generosity(.7001), kindness(.6975), appreciation(.6934), admiration(.6830), compassion(.6738), gratefulness(.6444), reverence(.6361), thankful(.6313), sympathy(.6304) | |
| jubilate | Vanilla | exultate(.5262), deum(.4763), exultate(.4545), motet(.4054), excelsis(.3777), deo(.3721), cantata(.3651), alleluia(.3625), stabat(.3558), laudamus(.3485) |
| +SelfExtro | exultate(.5111), exultate(.4449), deum(.4269), motet(.4093), deo(.3703), choir(.3670), excelsis(.3626), mozart(.3594), stabat(.3591), cantata(.3532) | |
| elated | Vanilla | overjoyed(.7712), ecstatic(.7045), thrilled(.6576), exhilarated(.6533), gratified(.6431), enthused(.6214), saddened(.6075), flabbergasted(.6060), disheartened(.5989), giddy(.5984) |
| +SelfExtro | overjoyed(.8018), ecstatic(.7424), thrilled(.7320), excited(.6894), exhilarated(.6758), delighted(.6733), gratified(.6690), relieved(.6569), enthused(.6564), dismayed(.6342) | |
| mono-saccharide | Vanilla | disaccharide(.6470), monosaccharides(.5152), oligosaccharide(.4911), galactose(.4686), 5-carbon(.4609), n-acetylgucosamine(.4282), saccharide(.4223), sucrose(.4215), mannose(.4196), disaccharides(.4107) |
| +SelfExtro | disaccharide(.6329), monosaccharides(.4888), oligosaccharide(.4844), 5-carbon(.4617), galactose(.4414), n-acetylgucosamine(.4227), sucrose(.4206), disaccharides(.4050), carbohydrate(.4045), moieties(.4033) | |
| outlook | Vanilla | excel(.5686), recovery(.5457), export(.5368), forecast(.5229), forecasts(.5060), contacts(.5025), powerpoint(.5025), exchange(.5011), microsoft(.4944), import(.4816) |
| +SelfExtro | excel(.5145), mapssevere(.5054), icalandargoogle(.4955), export(.4883), contacts(.4877), recovery(.4753), advisorieshourly(.4673), mailbox(.4661), vcard(.4612), thunderbird(.4599) | |
| vain | Vanilla | foolish(.6229), futile(.6104), selfish(.5340), vainly(.5242), fruitless(.5133), thy(.4981), arrogant(.4976), useless(.4844), righteous(.4833), thou(.4830) |
| +SelfExtro | futil(.6273), foolish(.6079), vainly(.5454), fruitless(.5338), selfish(.5143), useless(.5044), pointless(.4910), hopeless(.4846), attempt(.4797), presumptuous(.4780) | |
| prioritize | Vanilla | prioritise(.8064), prioritizing(.7426), priorities(.6323), prioritising(.5995), evaluate(.5950), strategize(.5933), proactively(.5862), analyze(.5800), assess(.5764), identify(.5633) |
| +SelfExtro | prioritise(.8322), prioritizing(.7415), priorities(.6390), prioritising(.6225), proactively(.6075), strategize(.5990), evaluate(.5946), analyze(.5919), prioritized(.5880), manage(.5783) | |
| nomad | Vanilla | nomads(.4911), drifter(.4735), nomadic(.4574), vagabond(.4233), zen(.4161), muvo2(.4054), gypsy(.4054), jukebox(.3997), vulcan(.3906), adventurer(.3869) |
| +SelfExtro | nomads(.5254), drifter(.4916), nomadic(.4841), muvo2(.4631), vagabond(.4422), traveler(.4246), traveller(.4216), zen(.4150), jukebox(.4110), wanderer(.4073) | |
| junction | Vanilla | junctions(.5729), near(.4985), road(.4936), intersection(.4923), creek(.4911), highway(.4878), exit(.4853), bridge(.4811), rd(.4779), jct(.4772) |
| +SelfExtro | junctions(.5822), jct(.5099), highway(.4787), motorway(.4781), road(.4744), exit(.4738).hwy(.4710), rd(.4667), intersection(.4628), heafford(.4596) | |
| moss | Vanilla | peat(.5360), brady(.5090), reed(.4998), brown(.4963), fern(.4945), green(.4920), woods(.4872), pine(.4856), jones(.4840), campbell(.4824), |
| +SelfExtro | lichen(.5628), sphagnum(.5463), spagnum(.5457), peat(.5323), lichens(.5264), mosses(.5224), fern(.4934), welker(.4852), pine(.4846), ferns(.4805), |
| Difficulty | NatSQL + RAT-SQL | Din-SQL GPT-4 | GPT3.5-turbo |
| Easy | 91.6% | 91.1% | 87.7% |
| Medium | 75.2% | 79.8% | 75.1% |
| Hard | 65.5% | 64.9% | 72.5% |
| Extra Hard | 51.8% | 43.4% | 53.9% |
| Overall | 73.7% | 74.2% | 74.3% |
| Operator | Description |
| Scan | Scan all rows in a table with optional filtering predicate |
| Aggregate | Aggregate a stream of tuples using a grouping criterion into a stream of groups |
| Filter | Remove tuples from a stream that do not match a predicate |
| Sort | Sort a stream according to a sorting expression |
| TopSort | Select the top-K tuples from a stream according to a sorting expression |
| Join | Perform a logical join operation between two streams based on a join condition |
| Except | Compute the set difference between two streams of tuples |
| Intersect | Compute the set intersection between two streams of tuples |
| Union | Compute the set union between two streams of tuples |
| Spider Difficulty | Q → QPL | Q+QD → QPL | Q → SQL | Support |
| Simple Schema Encoding | ||||
| Easy | 87.5% | 84.7% | 91.9% | 248 |
| Medium | 84.3% | 72.2% | 76.9% | 446 |
| Hard | 66.7% | 62.0% | 64.9% | 174 |
| Extra Hard | 54.8% | 45.1% | 44.6% | 166 |
| Overall | 77.4% | 69.1% | 73.3% | 1034 |
| Rich Schema Encoding | ||||
| Easy | 93.5% | 91.5% | 248 | |
| Medium | 89.0% | 88.3% | 446 | |
| Hard | 74.7% | 69.5% | 174 | |
| Extra Hard | 65.1% | 63.9% | 166 | |
| Overall | 83.8% | 82.0% | 1034 |
| Type | Gold label | Time | Correct |
| QPL | Incorrect query | 89% | 53% |
| Correct query | 100% | 79% | |
| 67% | |||
| SQL | Incorrect query | 102% | 50% |
| Correct query | 101% | 25% | |
| 34% | |||
| Avg time | QPL | 123s | |
| SQL | 132s |
| Error Type | Error | Err.% | Explanation |
| Wrong aggregate | 35 | 21% | Error in Aggregate (sum, avg, count, max, min) |
| Join | 31 | 19% | Wrong join (e.g., not on Primary/Foreign key) |
| Wrong column | 17 | 12% | Output does not include the right columns |
| Missing filter | 17 | 10% | Filter stage is missing |
| Wrong constant | 15 | 9% | Compare with wrong constant (e.g., isOfficial = 'Y' vs. 'T') |
| Wrong predicate | 12 | 7% | Error in selection predicate (e.g., > instead of <) |
| Lost | 12 | 7% | Predicted QPL is completely wrong |
| Typing | 7 | 4% | Compare with constant of wrong type (e.g., age = 'old' vs. 20) |
| Extremum | 4 | 2% | Error in selecting top value (min vs. max for example) |
| Intersect | 4 | 2% | Error in Intersect operation |
| Wrong structure | 3 | 2% | QPL plan is not a connected tree |
| Syntax issue | 3 | 2% | Predicted QPL is not syntactically valid |
| Except | 2 | 1% | Error in Except operation |
| Distinct | 1 | 1% | Missing distinct flag |
| Wrong table | 1 | 1% | Refers to the wrong table in the schema |
| Grand Total | 167 |
| Hyperparameter | Value |
| # Epoch | 10 |
| Dropout Prob. | 0.05 |
| Batch Size | 1 |
| Learning Rate | 0.0002 |
| Adaptation (r) | 16 |
| LoRA α | 32 |
| QPL Length | Q → QPL | Q+QD → QPL | Support |
| 1 | 87.3% | 78.3% | 189 |
| 2 | 86.6% | 83.4% | 277 |
| 3 | 85.3% | 78.0% | 191 |
| 4 | 75.0% | 62.9% | 124 |
| 5 | 67.1% | 54.4% | 164 |
| 6 | 48.2% | 25.9% | 27 |
| 7 | 31.8% | 22.7% | 44 |
| ≥8 | 11.9% | 24.4% | 18 |
| Overall | 77.4% | 69.1% | 1034 |
| QD-QPL Alignment | Support | Correct | Exec Acc | Avg QPL Gold Len. |
| [0.0, 0.4] | 1 | 0 | 0 | 5.0 |
| (0.4, 0.5] | 19 | 4 | 21.1 | 4.6 |
| (0.5, 0.6] | 9 | 2 | 22.2 | 6.3 |
| (0.6, 0.7] | 43 | 9 | 20.9 | 5.0 |
| (0.7, 0.8] | 63 | 21 | 33.3 | 4.2 |
| (0.8, 0.9] | 93 | 33 | 35.5 | 4.6 |
| (0.9, 1] | 779 | 624 | 80.1 | 2.8 |
| Schema ID | Errors | Samples | Error Rate |
| battle_death | 1 | 16 | 6% |
| car_1 | 27 | 92 | 29% |
| concert_singer | 5 | 45 | 11% |
| course_teach | 0 | 30 | 0% |
| cre_Doc_Package_Mgt | 8 | 84 | 10% |
| dog_kennels | 12 | 82 | 15% |
| employee_hire_evaluation | 1 | 38 | 3% |
| flight_2 | 15 | 80 | 19% |
| museum.visit | 2 | 18 | 11% |
| network_1 | 9 | 56 | 16% |
| orchestra | 0 | 40 | 0% |
| pets_1 | 2 | 42 | 5% |
| poker_player | 0 | 40 | 0% |
| real_estate_property | 1 | 4 | 25% |
| singer | 1 | 30 | 3% |
| student_transcripts_tracking | 16 | 78 | 21% |
| tvshow | 5 | 62 | 8% |
| voter_1 | 3 | 15 | 20% |
| world_1 | 44 | 120 | 37% |
| wta_1 | 15 | 62 | 24% |
| Grand Total | 167 | 1034 | 16% |
| Difficulty | NatSQL + RAT-SQL | Din-SQL GPT-4 | GPT3.5-turbo |
| Easy | 91.6% | 91.1% | 87.7% |
| Medium | 75.2% | 79.8% | 75.1% |
| Hard | 65.5% | 64.9% | 72.5% |
| Extra Hard | 51.8% | 43.4% | 53.9% |
| Overall | 73.7% | 74.2% | 74.3% |
| Operator | Description |
| Scan | Scan all rows in a table with optional filtering predicate |
| Aggregate | Aggregate a stream of tuples using a grouping criterion into a stream of groups |
| Filter | Remove tuples from a stream that do not match a predicate |
| Sort | Sort a stream according to a sorting expression |
| TopSort | Select the top-K tuples from a stream according to a sorting expression |
| Join | Perform a logical join operation between two streams based on a join condition |
| Except | Compute the set difference between two streams of tuples |
| Intersect | Compute the set intersection between two streams of tuples |
| Union | Compute the set union between two streams of tuples |
| Spider Difficulty | Q → QPL | Q+QD → QPL | Q → SQL | Support |
| Simple Schema Encoding | ||||
| Easy | 87.5% | 84.7% | 91.9% | 248 |
| Medium | 84.3% | 72.2% | 76.9% | 446 |
| Hard | 66.7% | 62.0% | 64.9% | 174 |
| Extra Hard | 54.8% | 45.1% | 44.6% | 166 |
| Overall | 77.4% | 69.1% | 73.3% | 1034 |
| Rich Schema Encoding | ||||
| Easy | 93.5% | 91.5% | 248 | |
| Medium | 89.0% | 88.3% | 446 | |
| Hard | 74.7% | 69.5% | 174 | |
| Extra Hard | 65.1% | 63.9% | 166 | |
| Overall | 83.8% | 82.0% | 1034 |
| Type | Gold label | Time | Correct |
| QPL | Incorrect query | 89% | 53% |
| Correct query | 100% | 79% | |
| 67% | |||
| SQL | Incorrect query | 102% | 50% |
| Correct query | 101% | 25% | |
| 34% | |||
| Avg time | QPL | 123s | |
| SQL | 132s |
| Error Type | Error | Err.% | Explanation |
| Wrong aggregate | 35 | 21% | Error in Aggregate (sum, avg, count, max, min) |
| Join | 31 | 19% | Wrong join (e.g., not on Primary/Foreign key) |
| Wrong column | 17 | 12% | Output does not include the right columns |
| Missing filter | 17 | 10% | Filter stage is missing |
| Wrong constant | 15 | 9% | Compare with wrong constant (e.g., isOfficial = 'Y' vs. 'T') |
| Wrong predicate | 12 | 7% | Error in selection predicate (e.g., > instead of <) |
| Lost | 12 | 7% | Predicted QPL is completely wrong |
| Typing | 7 | 4% | Compare with constant of wrong type (e.g., age = 'old' vs. 20) |
| Extremum | 4 | 2% | Error in selecting top value (min vs. max for example) |
| Intersect | 4 | 2% | Error in Intersect operation |
| Wrong structure | 3 | 2% | QPL plan is not a connected tree |
| Syntax issue | 3 | 2% | Predicted QPL is not syntactically valid |
| Except | 2 | 1% | Error in Except operation |
| Distinct | 1 | 1% | Missing distinct flag |
| Wrong table | 1 | 1% | Refers to the wrong table in the schema |
| Grand Total | 167 |
| Hyperparameter | Value |
| # Epoch | 10 |
| Dropout Prob. | 0.05 |
| Batch Size | 1 |
| Learning Rate | 0.0002 |
| Adaptation (r) | 16 |
| LoRA α | 32 |
| QPL Length | Q → QPL | Q+QD → QPL | Support |
| 1 | 87.3% | 78.3% | 189 |
| 2 | 86.6% | 83.4% | 277 |
| 3 | 85.3% | 78.0% | 191 |
| 4 | 75.0% | 62.9% | 124 |
| 5 | 67.1% | 54.4% | 164 |
| 6 | 48.2% | 25.9% | 27 |
| 7 | 31.8% | 22.7% | 44 |
| ≥8 | 11.9% | 24.4% | 18 |
| Overall | 77.4% | 69.1% | 1034 |
| QD-QPL Alignment | Support | Correct | Exec Acc | Avg QPL Gold Len. |
| [0.0, 0.4] | 1 | 0 | 0 | 5.0 |
| (0.4, 0.5] | 19 | 4 | 21.1 | 4.6 |
| (0.5, 0.6] | 9 | 2 | 22.2 | 6.3 |
| (0.6, 0.7] | 43 | 9 | 20.9 | 5.0 |
| (0.7, 0.8] | 63 | 21 | 33.3 | 4.2 |
| (0.8, 0.9] | 93 | 33 | 35.5 | 4.6 |
| (0.9, 1] | 779 | 624 | 80.1 | 2.8 |
| Schema ID | Errors | Samples | Error Rate |
| battle_death | 1 | 16 | 6% |
| car_1 | 27 | 92 | 29% |
| concert_singer | 5 | 45 | 11% |
| course_teach | 0 | 30 | 0% |
| cre_Doc_Package_Mgt | 8 | 84 | 10% |
| dog_kennels | 12 | 82 | 15% |
| employee_hire_evaluation | 1 | 38 | 3% |
| flight_2 | 15 | 80 | 19% |
| museum.visit | 2 | 18 | 11% |
| network_1 | 9 | 56 | 16% |
| orchestra | 0 | 40 | 0% |
| pets_1 | 2 | 42 | 5% |
| poker_player | 0 | 40 | 0% |
| real_estate_property | 1 | 4 | 25% |
| singer | 1 | 30 | 3% |
| student_transcripts_tracking | 16 | 78 | 21% |
| tvshow | 5 | 62 | 8% |
| voter_1 | 3 | 15 | 20% |
| world_1 | 44 | 120 | 37% |
| wta_1 | 15 | 62 | 24% |
| Grand Total | 167 | 1034 | 16% |
| Difficulty | NatSQL + RAT-SQL | Din-SQL GPT-4 | GPT3.5-turbo |
| Easy | 91.6% | 91.1% | 87.7% |
| Medium | 75.2% | 79.8% | 75.1% |
| Hard | 65.5% | 64.9% | 72.5% |
| Extra Hard | 51.8% | 43.4% | 53.9% |
| Overall | 73.7% | 74.2% | 74.3% |
| Operator | Description |
| Scan | Scan all rows in a table with optional filtering predicate |
| Aggregate | Aggregate a stream of tuples using a grouping criterion into a stream of groups |
| Filter | Remove tuples from a stream that do not match a predicate |
| Sort | Sort a stream according to a sorting expression |
| TopSort | Select the top-K tuples from a stream according to a sorting expression |
| Join | Perform a logical join operation between two streams based on a join condition |
| Except | Compute the set difference between two streams of tuples |
| Intersect | Compute the set intersection between two streams of tuples |
| Union | Compute the set union between two streams of tuples |
| Spider Difficulty | Q → QPL | Q+QD → QPL | Q → SQL | Support |
| Simple Schema Encoding | ||||
| Easy | 87.5% | 84.7% | 91.9% | 248 |
| Medium | 84.3% | 72.2% | 76.9% | 446 |
| Hard | 66.7% | 62.0% | 64.9% | 174 |
| Extra Hard | 54.8% | 45.1% | 44.6% | 166 |
| Overall | 77.4% | 69.1% | 73.3% | 1034 |
| Rich Schema Encoding | ||||
| Easy | 93.5% | 91.5% | 248 | |
| Medium | 89.0% | 88.3% | 446 | |
| Hard | 74.7% | 69.5% | 174 | |
| Extra Hard | 65.1% | 63.9% | 166 | |
| Overall | 83.8% | 82.0% | 1034 | |
| Type | Gold label | Time | Correct |
| QPL | Incorrect query | 89% | 53% |
| Correct query | 100% | 79% | |
| 67% | |||
| SQL | Incorrect query | 102% | 50% |
| Correct query | 101% | 25% | |
| 34% | |||
| Avg time | QPL | 123s | |
| SQL | 132s |
| Error Type | Error | Err.% | Explanation |
| Wrong aggregate | 35 | 21% | Error in Aggregate (sum, avg, count, max, min) |
| Join | 31 | 19% | Wrong join (e.g., not on Primary/Foreign key) |
| Wrong column | 17 | 12% | Output does not include the right columns |
| Missing filter | 17 | 10% | Filter stage is missing |
| Wrong constant | 15 | 9% | Compare with wrong constant (e.g., isOfficial = 'Y' vs. 'T') |
| Wrong predicate | 12 | 7% | Error in selection predicate (e.g., > instead of <) |
| Lost | 12 | 7% | Predicted QPL is completely wrong |
| Typing | 7 | 4% | Compare with constant of wrong type (e.g., age = 'old' vs. 20) |
| Extremum | 4 | 2% | Error in selecting top value (min vs. max for example) |
| Intersect | 4 | 2% | Error in Intersect operation |
| Wrong structure | 3 | 2% | QPL plan is not a connected tree |
| Syntax issue | 3 | 2% | Predicted QPL is not syntactically valid |
| Except | 2 | 1% | Error in Except operation |
| Distinct | 1 | 1% | Missing distinct flag |
| Wrong table | 1 | 1% | Refers to the wrong table in the schema |
| Grand Total | 167 |
| Hyperparameter | Value |
| # Epoch | 10 |
| Dropout Prob. | 0.05 |
| Batch Size | 1 |
| Learning Rate | 0.0002 |
| Adaptation (r) | 16 |
| LoRA α | 32 |
| QPL Length | Q → QPL | Q+QD → QPL | Support |
| 1 | 87.3% | 78.3% | 189 |
| 2 | 86.6% | 83.4% | 277 |
| 3 | 85.3% | 78.0% | 191 |
| 4 | 75.0% | 62.9% | 124 |
| 5 | 67.1% | 54.4% | 164 |
| 6 | 48.2% | 25.9% | 27 |
| 7 | 31.8% | 22.7% | 44 |
| ≥8 | 11.9% | 24.4% | 18 |
| Overall | 77.4% | 69.1% | 1034 |