# Label efficiency (E7, design split; mean over 3 seeds x draws) ## dbpedia_accuracy Labelled facts used (mean): 1% = 75 (11 errors), 5% = 374 (55 errors), 10% = 748 (112 errors), 25% = 1869 (275 errors), 100% = 7475 (1116 errors) ### AP by label budget (mean ± sd over seeds x draws) | method | 0.01 | 0.05 | 0.1 | 0.25 | 1.0 | |:-------------------------------------------------------------|:--------------|:--------------|:--------------|:--------------|:--------------| | KGGuard+W (cross-fit, + web evidence) | 0.149 ± 0.000 | 0.332 ± 0.022 | 0.357 ± 0.019 | 0.398 ± 0.022 | 0.470 ± 0.000 | | KGGuard+ (cross-fit) | 0.149 ± 0.000 | 0.306 ± 0.021 | 0.337 ± 0.018 | 0.380 ± 0.020 | 0.454 ± 0.001 | | KGGuard+W (cross-fit, + web evidence) [combiner=fixed] | 0.319 ± 0.029 | 0.360 ± 0.024 | 0.371 ± 0.016 | 0.391 ± 0.012 | 0.439 ± 0.001 | | KGGuard+W (cross-fit, + web evidence) [combiner=logistic] | 0.279 ± 0.041 | 0.349 ± 0.023 | 0.364 ± 0.023 | 0.394 ± 0.013 | 0.456 ± 0.000 | | PredicatePrior (cross-fit) | 0.179 ± 0.009 | 0.230 ± 0.021 | 0.280 ± 0.021 | 0.323 ± 0.022 | 0.404 ± 0.000 | | GBDT (cross-fit, all features) | 0.220 ± 0.027 | 0.297 ± 0.031 | 0.344 ± 0.021 | 0.382 ± 0.024 | 0.493 ± 0.003 | | GBDT (cross-fit, all features + web) | 0.236 ± 0.034 | 0.312 ± 0.034 | 0.362 ± 0.022 | 0.401 ± 0.028 | 0.494 ± 0.008 | | KGGuard+W (cross-fit, + web evidence) [CV-selected combiner] | 0.305 ± 0.046 | 0.356 ± 0.024 | 0.371 ± 0.016 | 0.391 ± 0.019 | 0.470 ± 0.000 | ### Runs beating the best label-free baseline (LLM zero-shot (Qwen2.5-7B), AP 0.257) | method | 0.01 | 0.05 | 0.1 | 0.25 | 1.0 | |:-------------------------------------------------------------|:-------|:-------|:------|:-------|:------| | KGGuard+W (cross-fit, + web evidence) | 0/15 | 15/15 | 15/15 | 15/15 | 3/3 | | KGGuard+ (cross-fit) | 0/15 | 15/15 | 15/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [combiner=fixed] | 14/15 | 15/15 | 15/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [combiner=logistic] | 11/15 | 15/15 | 15/15 | 15/15 | 3/3 | | PredicatePrior (cross-fit) | 0/15 | 2/15 | 13/15 | 15/15 | 3/3 | | GBDT (cross-fit, all features) | 1/15 | 14/15 | 15/15 | 15/15 | 3/3 | | GBDT (cross-fit, all features + web) | 4/15 | 14/15 | 15/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [CV-selected combiner] | 12/15 | 15/15 | 15/15 | 15/15 | 3/3 | ### Label-free baselines (E1, no labels) | method | AP | |:------------------------------------|------:| | LLM zero-shot (Qwen2.5-7B) | 0.257 | | FactCheck-Consensus (4 LLMs, GIV-F) | 0.255 | | FactCheck-RAG (Gemma2-9B, web) | 0.233 | | FactCheck-GIV-F (Gemma2-9B) | 0.219 | | SDValidate | 0.182 | | PaTyBRED | 0.177 | | FactCheck-DKA (Gemma2-9B) | 0.157 | | TransE | 0.151 | | CAGED | 0.142 | | ComplEx | 0.139 | ### Combiner chosen by the CV rule | budget | fixed | gam | logistic | |---------:|--------:|------:|-----------:| | 0.01 | 11 | 0 | 4 | | 0.05 | 13 | 1 | 1 | | 0.1 | 15 | 0 | 0 | | 0.25 | 8 | 5 | 2 | | 1 | 0 | 3 | 0 | ## factbench Labelled facts used (mean): 1% = 15 (7 errors), 5% = 76 (40 errors), 10% = 153 (78 errors), 25% = 382 (195 errors), 100% = 1530 (780 errors) ### AP by label budget (mean ± sd over seeds x draws) | method | 0.01 | 0.05 | 0.1 | 0.25 | 1.0 | |:-------------------------------------------------------------|:--------------|:--------------|:--------------|:--------------|:--------------| | KGGuard+W (cross-fit, + web evidence) | 0.510 ± 0.000 | 0.510 ± 0.000 | 0.896 ± 0.013 | 0.919 ± 0.006 | 0.928 ± 0.000 | | KGGuard+ (cross-fit) | 0.510 ± 0.000 | 0.510 ± 0.000 | 0.834 ± 0.011 | 0.861 ± 0.009 | 0.876 ± 0.000 | | KGGuard+W (cross-fit, + web evidence) [combiner=fixed] | 0.636 ± 0.074 | 0.880 ± 0.020 | 0.899 ± 0.007 | 0.909 ± 0.005 | 0.918 ± 0.000 | | KGGuard+W (cross-fit, + web evidence) [combiner=logistic] | 0.609 ± 0.066 | 0.895 ± 0.012 | 0.914 ± 0.005 | 0.918 ± 0.004 | 0.925 ± 0.000 | | PredicatePrior (cross-fit) | 0.510 ± 0.004 | 0.508 ± 0.006 | 0.508 ± 0.005 | 0.511 ± 0.006 | 0.509 ± 0.000 | | GBDT (cross-fit, all features) | 0.510 ± 0.000 | 0.791 ± 0.026 | 0.856 ± 0.019 | 0.893 ± 0.009 | 0.934 ± 0.001 | | GBDT (cross-fit, all features + web) | 0.510 ± 0.000 | 0.897 ± 0.022 | 0.938 ± 0.008 | 0.960 ± 0.007 | 0.981 ± 0.000 | | KGGuard+W (cross-fit, + web evidence) [CV-selected combiner] | 0.617 ± 0.074 | 0.883 ± 0.023 | 0.908 ± 0.008 | 0.912 ± 0.006 | 0.928 ± 0.000 | ### Runs beating the best label-free baseline (FactCheck-RAG (Gemma2-9B, web), AP 0.834) | method | 0.01 | 0.05 | 0.1 | 0.25 | 1.0 | |:-------------------------------------------------------------|:-------|:-------|:------|:-------|:------| | KGGuard+W (cross-fit, + web evidence) | 0/15 | 0/15 | 15/15 | 15/15 | 3/3 | | KGGuard+ (cross-fit) | 0/15 | 0/15 | 6/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [combiner=fixed] | 0/15 | 14/15 | 15/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [combiner=logistic] | 0/15 | 15/15 | 15/15 | 15/15 | 3/3 | | PredicatePrior (cross-fit) | 0/15 | 0/15 | 0/15 | 0/15 | 0/3 | | GBDT (cross-fit, all features) | 0/15 | 0/15 | 13/15 | 15/15 | 3/3 | | GBDT (cross-fit, all features + web) | 0/15 | 15/15 | 15/15 | 15/15 | 3/3 | | KGGuard+W (cross-fit, + web evidence) [CV-selected combiner] | 0/15 | 14/15 | 15/15 | 15/15 | 3/3 | ### Label-free baselines (E1, no labels) | method | AP | |:------------------------------------|------:| | FactCheck-RAG (Gemma2-9B, web) | 0.834 | | FactCheck-Consensus (4 LLMs, GIV-F) | 0.774 | | LLM zero-shot (Qwen2.5-7B) | 0.741 | | FactCheck-GIV-F (Gemma2-9B) | 0.669 | | SDValidate | 0.657 | | FactCheck-DKA (Gemma2-9B) | 0.656 | | PaTyBRED | 0.617 | | TransE | 0.568 | | CAGED | 0.533 | | ComplEx | 0.513 | ### Combiner chosen by the CV rule | budget | fixed | gam | logistic | |---------:|--------:|------:|-----------:| | 0.01 | 4 | 1 | 10 | | 0.05 | 9 | 0 | 6 | | 0.1 | 7 | 0 | 8 | | 0.25 | 6 | 4 | 5 | | 1 | 0 | 3 | 0 |