Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Verify Summary Facts with Fine-Grained LLM Feedback

Jihwan Oh, Jeonghwan Choi, Nicole Hee-Yeon Kim, Taewon Yun, Hwanjun Song

TL;DR

The paper addresses the high cost and limited scalability of human-labeled fact verification data for summaries by introducing FineSumFact, a large-scale dataset of fine-grained LLM-generated feedback. It presents a four-stage pipeline that generates diverse summaries from 10 LLMs, collects sentence-level feedback via an off-the-shelf verifier (FineSurE) powered by Llama-3-70B-Instruct, and trains a lightweight verifier (Llama-3-8B-Instruct) through sequence-level distillation (QLoRA). Empirical results show that training with extensive LLM feedback yields higher agreement with human judgments than human-only baselines and outperforms QA/NLI-based evaluators, with added benefits from explainable feedback such as reasoning and error localization. The approach also indicates substantial gains in efficiency, achieving near-teacher performance with significantly faster inference and lower API costs, suggesting a scalable path for domain-general fact verification of summaries.

Abstract

Training automatic summary fact verifiers often faces the challenge of a lack of human-labeled data. In this paper, we explore alternative way of leveraging Large Language Model (LLM) generated feedback to address the inherent limitation of using human-labeled data. We introduce FineSumFact, a large-scale dataset containing fine-grained factual feedback on summaries. We employ 10 distinct LLMs for diverse summary generation and Llama-3-70B-Instruct for feedback. We utilize this dataset to fine-tune the lightweight open-source model Llama-3-8B-Instruct, optimizing resource efficiency while maintaining high performance. Our experimental results reveal that the model trained on extensive LLM-generated datasets surpasses that trained on smaller human-annotated datasets when evaluated using human-generated test sets. Fine-tuning fact verification models with LLM feedback can be more effective and cost-efficient than using human feedback. The dataset is available at https://github.com/DISL-Lab/FineSumFact.

Learning to Verify Summary Facts with Fine-Grained LLM Feedback

TL;DR

The paper addresses the high cost and limited scalability of human-labeled fact verification data for summaries by introducing FineSumFact, a large-scale dataset of fine-grained LLM-generated feedback. It presents a four-stage pipeline that generates diverse summaries from 10 LLMs, collects sentence-level feedback via an off-the-shelf verifier (FineSurE) powered by Llama-3-70B-Instruct, and trains a lightweight verifier (Llama-3-8B-Instruct) through sequence-level distillation (QLoRA). Empirical results show that training with extensive LLM feedback yields higher agreement with human judgments than human-only baselines and outperforms QA/NLI-based evaluators, with added benefits from explainable feedback such as reasoning and error localization. The approach also indicates substantial gains in efficiency, achieving near-teacher performance with significantly faster inference and lower API costs, suggesting a scalable path for domain-general fact verification of summaries.

Abstract

Training automatic summary fact verifiers often faces the challenge of a lack of human-labeled data. In this paper, we explore alternative way of leveraging Large Language Model (LLM) generated feedback to address the inherent limitation of using human-labeled data. We introduce FineSumFact, a large-scale dataset containing fine-grained factual feedback on summaries. We employ 10 distinct LLMs for diverse summary generation and Llama-3-70B-Instruct for feedback. We utilize this dataset to fine-tune the lightweight open-source model Llama-3-8B-Instruct, optimizing resource efficiency while maintaining high performance. Our experimental results reveal that the model trained on extensive LLM-generated datasets surpasses that trained on smaller human-annotated datasets when evaluated using human-generated test sets. Fine-tuning fact verification models with LLM feedback can be more effective and cost-efficient than using human feedback. The dataset is available at https://github.com/DISL-Lab/FineSumFact.

Paper Structure

This paper contains 34 sections, 4 equations, 5 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fact Verification Datasets.
  4. Fact Verification Methods.
  5. Preliminary
  6. Dataset with Human Feedback.
  7. Learning with LLM Feedback
  8. Evaluation
  9. Agreement with Humans
  10. Factuality Error Localization
  11. Ablation on Feedback Granularity
  12. Ablation on Feedback Size
  13. Inference Latency
  14. Understanding Why It Works
  15. Conclusion
  16. ...and 19 more sections

Figures (5)

  • Figure 1: Pipeline: our evaluator is trained with LLM feedback generated on diverse input texts and summaries and then tested on an unseen test set.
  • Figure 2: Prompt for fact verification ("Binary Label" in Table \ref{['table:granularity-feedback']}).
  • Figure 3: Prompt for fact verification ("Binary Label + Reasoning" in Table \ref{['table:granularity-feedback']}).
  • Figure 4: Prompt for fact verification ("Binary Label + Reasoning + Error Localization" in Table \ref{['table:granularity-feedback']}, which is exactly the same with FineSurE song2024finesure).
  • Figure 5: Error category distribution of summaries with LLM feedback for each summarizer, where the error category is estimated using the automated fact verification.