|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- PKU-ONELab/NLG-Eval |
|
language: |
|
- en |
|
base_model: |
|
- meta-llama/Meta-Llama-3-8B |
|
--- |
|
# Themis |
|
|
|
Themis: A Reference-free NLG Evaluation Language Model with Flexibility and Interpretability |
|
|
|
Paper: https://aclanthology.org/2024.emnlp-main.891 |
|
|
|
Github: https://github.com/PKU-ONELab/Themis |
|
|
|
## Introduction |
|
|
|
We propose **Themis**, an 8B-parameter large language model (LLM) specifically designed and trained for NLG evaluation with more comprehensive capabilities. |
|
|
|
Our Themis can evaluate various NLG tasks, including uncommon ones like question-answering evaluation (**Versatility**), in a reference-free manner (**Independence**). Moreover, it allows for specific and customized evaluation aspects and criteria, including overall quality and more fine-grained aspects (**Flexibility**), and its evaluation contains corresponding analysis and explanation together with the rating (**Interpretability**). |
|
|
|
We believe that an ideal evaluator should be convenient to use and possess these characteristics. The comparison between related methods and Themis is shown in the table below. |
|
|
|
| Method | Versatility | Independence | Flexibility | Interpretability | Open-source | |
|
| :---------------: | :---------: | :----------: | :---------: | :--------------: | :---------: | |
|
| UniEval | β | β | βοΈ | β | βοΈ | |
|
| G-Eval | βοΈ | βοΈ | βοΈ | βοΈ | β | |
|
| X-Eval | βοΈ | β | βοΈ | β | β | |
|
| Prometheus | βοΈ | β | βοΈ | βοΈ | βοΈ | |
|
| Auto-J | βοΈ | βοΈ | β | βοΈ | βοΈ | |
|
| InstructScore | βοΈ | β | β | βοΈ | βοΈ | |
|
| TIGERScore | βοΈ | βοΈ | β | βοΈ | βοΈ | |
|
| **Themis (Ours)** | βοΈ | βοΈ | βοΈ | βοΈ | βοΈ | |
|
|
|
## Performance |
|
|
|
We implement experiments on several common NLG evaluation tasks and datasets to compare our Themis with other methods, including SummEval for summarization, Topical-Chat for dialogue response generation, SFRES&SFHOT for data-to-text, QAGS for factuality, MANS for story generation, and WMT23 zh-en for machine translation. Experimental results show that our Themis achieves better overall evaluation performance over other evaluation models, including GPT-4. |
|
|
|
| Method | SummEval | Topical-Chat | SFHOT& SFRES | QAGS | MANS | WMT23 | Average Spearman | |
|
| -------------------- | :-------: | :----------: | :---------: | :-------: | :-------: | :-------: | :------------: | |
|
| BLEU | 0.075 | 0.388 | 0.024 | - | 0.032 | 0.021 | - | |
|
| ROUGE | 0.152 | 0.412 | 0.101 | - | -0.002 | 0.151 | - | |
|
| BARTScore | 0.329 | 0.086 | 0.208 | 0.425 | 0.350 | 0.118 | 0.253 | |
|
| BERTScore | 0.231 | 0.394 | 0.139 | - | 0.285 | 0.219 | - | |
|
| BLEURT | 0.152 | 0.388 | 0.244 | - | 0.138 | 0.263 | - | |
|
| CometKiwi | 0.228 | 0.340 | 0.251 | 0.094 | 0.251 | 0.343 | 0.251 | |
|
| UniEval | 0.474 | 0.577 | 0.282 | - | - | - | - | |
|
| G-Eval (GPT-3.5) | 0.409 | 0.585 | - | 0.461 | - | - | - | |
|
| G-Eval (GPT-4) | 0.523 | 0.588 | - | 0.611 | - | - | - | |
|
| GPT-3.5 Turbo | 0.416 | 0.578 | 0.306 | 0.431 | 0.328 | 0.347 | 0.401 | |
|
| GPT-4 Turbo | 0.511 | **0.746** | 0.320 | 0.637 | 0.473 | **0.437** | 0.521 | |
|
| X-Eval | 0.480 | 0.605 | 0.303 | 0.578 | - | - | - | |
|
| Prometheus-13B | 0.163 | 0.434 | 0.173 | - | 0.007 | 0.129 | - | |
|
| Auto-J-13B | 0.198 | 0.425 | 0.141 | 0.226 | 0.380 | 0.104 | 0.246 | |
|
| TIGERScore-13B | 0.384 | 0.346 | 0.200 | 0.504 | 0.231 | 0.248 | 0.319 | |
|
| InstructScore-7B | 0.258 | 0.241 | 0.247 | - | 0.298 | 0.219 | - | |
|
| **Themis-8B (ours)** | **0.553** | 0.725 | **0.333** | **0.684** | **0.551** | 0.405 | **0.542** | |
|
|
|
We further conduct more in-depth analyses, including generalization tests on unseen tasks like the instruction-following evaluation as well as aspect-targeted perturbation tests, and our Themis also exhibits superior evaluation performance. For more experimental results and details, please refer to our paper. |
|
|
|
## Requirements and Usage |
|
|
|
Please refer to our [github repo](https://github.com/PKU-ONELab/Themis) for more details. |
|
|
|
## Citation |
|
|
|
``` |
|
@inproceedings{hu2024themis, |
|
title={Themis: A Reference-free NLG Evaluation Language Model with Flexibility and Interpretability}, |
|
author={Hu, Xinyu and Lin, Li and Gao, Mingqi and Yin, Xunjian and Wan, Xiaojun}, |
|
booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing}, |
|
pages={15924--15951}, |
|
year={2024} |
|
} |
|
``` |