[Paper Review] INSTRUCTSCORE: Explainable Text Generation Evaluation with Finegrained Feedback

Wenda Xu, Danqing Wang+5 more|arXiv (Cornell University)|May 23, 2023

Natural Language Processing Techniques4 citations

TL;DR

InstructScore is a fine-grained, explainable text generation evaluation metric that leverages GPT-4's implicit knowledge to generate natural language diagnostic reports—identifying error types, locations, severities, and justifications—without requiring human-annotated ratings. Its 7B LLaMA-based model outperforms all unsupervised metrics and rivals supervised state-of-the-art models like COMET22, achieving 13.7% higher human alignment after iterative refinement.

ABSTRACT

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.

Motivation & Objective

• To develop an explainable text generation evaluation metric that provides detailed, human-readable diagnostic reports for generated text.
• To enable fine-grained error detection by identifying error types, locations, severities, and justifications without relying on human-annotated scores.
• To improve model alignment with human judgment through self-supervised feedback loops using GPT-4 as a meta-evaluator.
• To create a generalizable, unsupervised training pipeline that is adaptable across diverse text generation tasks.
• To surpass existing unsupervised metrics in correlation with human judgments while maintaining interpretability.

Proposed method

• Construct a synthetic dataset by prompting GPT-4 to generate structured diagnostic reports for text generation errors, using explicit instruction templates.
• Design a feedback pipeline that queries GPT-4 on error explanations to assess correctness, consistency, and alignment with human judgment.
• Fine-tune a 7B LLaMA-based model on self-generated diagnostic outputs, optimizing for feedback scores to improve explanation quality.
• Iteratively refine the model by sampling failure modes from real machine-generated outputs and re-evaluating explanations via GPT-4 feedback.
• Use a multi-turn prompt framework to validate error annotations, including error location extraction, phrase alignment, and severity classification.
• Apply automated evaluation with GPT-4 to score diagnostic reports, assigning 1 for correct/no-error claims and 0 for incorrect or inconsistent claims.

Figure 1: InstructScore generates a comprehensive error diagnostic report for text generation tasks, including error type, location, severity label, and explanation. Based on this report, InstructScore counts the number of major errors (each worth $-5$ ) and minor errors (each worth $-1$ ), ultimate

Experimental results

Research questions

• RQ1 Can an unsupervised text evaluation metric generate detailed, human-readable diagnostic reports that explain numerical scores?
• RQ2 How well can a 7B LLaMA-based model perform compared to 175B closed-source models in text generation evaluation?
• RQ3 To what extent can self-supervised feedback from GPT-4 improve alignment with human judgment without human-annotated data?
• RQ4 Can the metric generalize across diverse text generation tasks such as translation, captioning, and commonsense generation?
• RQ5 What failure modes emerge in diagnostic reports, and how can they be systematically detected and corrected?

Key findings

• The 7B InstructScore model surpasses all existing unsupervised metrics, including those based on GPT-3 (175B) and GPT-4, across multiple text generation tasks.
• InstructScore achieves performance on par with supervised state-of-the-art metrics like COMET22 in machine translation, despite being trained without human-rated data.
• InstructScore outperforms supervised BLEURT in 6 out of 9 translation directions, demonstrating strong zero-shot generalization.
• After iterative refinement using GPT-4 feedback, InstructScore improves human judgment correlation by 13.7%.
• The method successfully identifies and corrects common failure modes in diagnostic reports, such as incorrect error locations and misclassified severities.
• The diagnostic reports generated by InstructScore are highly consistent with human annotations, with GPT-4 confirming 98% of no-error claims as correct.

Figure 2: Our InstructScore pipeline consists of three components: First, we construct synthetic data from GPT-4 and use it to fine-tune a 7B LLAMA model. Second, we sample from real-world machine-generated distribution to trigger InstructScore ’s failure modes. We query GPT-4 on each failure mode a