Improving Model Factuality with Fine-grained Critique-based Evaluator

2024-10-24

arXiv.org

AI Summary

This work addresses the factual inaccuracies prevalent in large language model (LLM) outputs by proposing FenCE, a statement-level, interpretable framework for factual evaluation and optimization. Methodologically: (1) it introduces the first fine-grained, statement-level factual evaluator supporting multi-source document grounding and textual critique; (2) it designs an evaluator-driven hallucination-free response revision and preference training paradigm that jointly models factual judgment, critique generation, and response refinement. Experiments show FenCE improves accuracy by 2.9% on LLMAggreFact; after optimization, Llama2-7B-chat and Llama3-8B-chat achieve 16.86% and 14.45% absolute gains in factual consistency on FActScore, respectively—outperforming existing fine-tuning approaches. The core contribution lies in establishing the first interpretable, document-grounded, statement-level factual evaluation mechanism and closing the evaluation–revision–preference optimization loop.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Factuality evaluation aims to detect factual errors produced by language models (LMs) and hence guide the development of more factual models. Towards this goal, we train a factuality evaluator, FenCE, that provides LM generators with claim-level factuality feedback. We conduct data augmentation on a combination of public judgment datasets to train FenCE to (1) generate textual critiques along with scores and (2) make claim-level judgment based on diverse source documents obtained by various tools. We then present a framework that leverages FenCE to improve the factuality of LM generators by constructing training data. Specifically, we generate a set of candidate responses, leverage FenCE to revise and score each response without introducing lesser-known facts, and train the generator by preferring highly scored revised responses. Experiments show that our data augmentation methods improve the evaluator's accuracy by 2.9% on LLM-AggreFact. With FenCE, we improve Llama2-7B-chat and Llama3-8B-chat's factuality rate by 16.86% and 14.45% on FActScore, outperforming state-of-the-art factuality finetuning methods by 8.83% and 6.96%.

Problem

• Detect factual errors in language model outputs
• Train evaluator for claim-level factuality feedback
• Improve model factuality via critique-based training

Innovation

• Fine-grained critique-based evaluator for factuality
• Data augmentation on public judgment datasets
• Framework leveraging evaluator to improve LM factuality