Automatically evaluating the quality of language generation is critical. Although recent learned metrics show high correlation with human judgement, these metrics can not explain their verdict or associate the scores with defects in generated text. To address this limitation, we present InstructScore, an explainable evaluation metric for text generation. By harnessing both explicit human instruction and the implicit knowledge of GPT-4, we fine-tune a text evaluation metric based on LLaMA, producing both a score for generated text and a human readable diagnostic report. We evaluate InstructScore on a variety of generation tasks, including translation, captioning, data-to-text and commonsense generation. Experiments show that our 7B model surpasses all other unsupervised metrics, including those based on 175B GPT-3 and GPT-4. Surprisingly, our InstructScore, even without direct supervision from human-rated data, achieves performance levels on par with state-of-the-art metrics like COMET22, which were fine-tuned on human ratings.
Fields | Explanation Failure Mode | Description |
---|---|---|
Local Failure Mode | ||
Error Type | Inconsistency to explanation | M1: Error type descriptions are not consistent with explanation |
Error Location | Inconsistency to explanation | M2: Error locations are not consistent with the explanation |
Error location hallucination | M3: Error locations are not referred in the output text | |
Major/Minor | Major/Minor disagreement | M5: Major and minor labels do not correspond to the correct severity levels |
Explanation | Error location hallucination | M4: Error locations can not refer to the output text |
Explanation failure | M6: The explanation is wrong. However, error at a specified location does exist | |
Global Failure Mode | ||
All 4 Fields | False negative error | G1: Error described in the explanation is not an error |
Repetition | G2: One error is mentioned more than once among explanations | |
Phrase misalignment | G3: Incorrect phrase and correct phrase are not correctly aligned | |
Mention multiple errors | G4: One error span mentions multiple errors | |
Common failure modes of the explanation output of first step Exp-Generator (Fine-tuned LLaMA on synthetic data without refinement). Local errors are field-specific, which only correspond to the error of the local field. Global errors can affect all four fields, such as error type, error location, major/minor, and explanation. The observation of the failures modes at first step Exp-Generator is the main motivation for us to perform refinement with automatic feedback.
@article{xu2023instructscore,
title={Instructscore: Towards explainable text generation evaluation with automatic feedback},
author={Xu, Wenda and Wang, Danqing and Pan, Liangming and Song, Zhenqiao and Freitag, Markus and Wang, William Yang and Li, Lei},
journal={The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) 2023},
year={2023}
}