Document concatenated and non-concatenated versions
Browse files
README.md
CHANGED
|
@@ -9,7 +9,7 @@
|
|
| 9 |
</div>
|
| 10 |
|
| 11 |
<div align="center">
|
| 12 |
-
<img src="humdial-bench.png" alt="HumDial-EIBench pipeline and task overview" width="90%">
|
| 13 |
<p><em>Figure: Three-stage data pipeline and the four evaluation tasks in HumDial-EIBench.</em></p>
|
| 14 |
</div>
|
| 15 |
|
|
@@ -88,6 +88,34 @@ Test robustness when text sentiment contradicts vocal affect (for example, sarca
|
|
| 88 |
|
| 89 |
- The dataset is publicly available on HuggingFace: [HumDial-EIBench](https://huggingface.co/datasets/ASLP-lab/HumDial-EIBench)
|
| 90 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
---
|
| 92 |
|
| 93 |
## Evaluation Usage
|
|
|
|
| 9 |
</div>
|
| 10 |
|
| 11 |
<div align="center">
|
| 12 |
+
<img src="assets/humdial-bench.png" alt="HumDial-EIBench pipeline and task overview" width="90%">
|
| 13 |
<p><em>Figure: Three-stage data pipeline and the four evaluation tasks in HumDial-EIBench.</em></p>
|
| 14 |
</div>
|
| 15 |
|
|
|
|
| 88 |
|
| 89 |
- The dataset is publicly available on HuggingFace: [HumDial-EIBench](https://huggingface.co/datasets/ASLP-lab/HumDial-EIBench)
|
| 90 |
|
| 91 |
+
### Audio Input Versions
|
| 92 |
+
|
| 93 |
+
The release provides two audio-input versions with the same dialogue IDs,
|
| 94 |
+
annotations, and evaluation tasks:
|
| 95 |
+
|
| 96 |
+
- **Concatenated version (`en/`, `zh/`)**: the final-turn audio contains the
|
| 97 |
+
original question followed by the multiple-choice prompt. The final-turn
|
| 98 |
+
`text` field also contains the complete concatenated text. This version is
|
| 99 |
+
suitable for audio-only inference.
|
| 100 |
+
- **Non-concatenated version (`no_concat/en/`, `no_concat/zh/`)**: the final-turn
|
| 101 |
+
audio contains only the original question. The multiple-choice prompt is
|
| 102 |
+
provided separately in the final turn's `prompt_text` field, so a model can
|
| 103 |
+
receive it as text. The original question remains in `text`.
|
| 104 |
+
|
| 105 |
+
For example:
|
| 106 |
+
|
| 107 |
+
```json
|
| 108 |
+
{
|
| 109 |
+
"text": "Please summarize how my emotions changed in this conversation.",
|
| 110 |
+
"prompt_text": "Choose one option and provide a reason:\nA. ...\nB. ...\nC. ...\nD. ...",
|
| 111 |
+
"answer": "C. ..."
|
| 112 |
+
}
|
| 113 |
+
```
|
| 114 |
+
|
| 115 |
+
In both versions, previous dialogue turns and all annotations remain aligned.
|
| 116 |
+
The only intended difference is whether the final multiple-choice prompt is
|
| 117 |
+
included in the audio or supplied as a separate text field.
|
| 118 |
+
|
| 119 |
---
|
| 120 |
|
| 121 |
## Evaluation Usage
|