Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Factual Consistency of News Summarization by Contrastive Preference Optimization

Huawen Feng, Yan Fan, Xiong Liu, Ting-En Lin, Zekun Yao, Yuchuan Wu, Fei Huang, Yongbin Li, Qianli Ma

TL;DR

Faced with factual hallucinations in news summarization by large language models, the paper introduces Contrastive Preference Optimization (CPO) to disentangle faithful versus imaginative content and a Probing-based Specific Training (PST) to identify and train weak layers. It also introduces LESSON, a sentence-level annotated dataset for training and evaluation. Empirical results across diverse backbones show CPO+PST substantially improves factual consistency and outperforms RL-based and other fine-tuning strategies, with PST offering stability and efficiency. These contributions enable more reliable, domain-adaptable news summarization in practice.

Abstract

Despite the recent progress in news summarization made by large language models (LLMs), they often generate summaries that are factually inconsistent with original articles, known as "hallucinations" in text generation. Unlike previous small models (e.g., BART, T5), current LLMs make fewer silly mistakes but more sophisticated ones, such as imposing cause and effect, adding false details, overgeneralizing, etc. These hallucinations are challenging to detect through traditional methods, which poses great challenges for improving the factual consistency of text summarization. In this paper, we propose Contrastive Preference Optimization (CPO) to disentangle the LLMs' propensities to generate faithful and fake content. Furthermore, we adopt a probing-based specific training method to improve their capacity of distinguishing two types of propensities. In this way, LLMs can execute the instructions more accurately and have enhanced perception of hallucinations. Experimental results show that CPO significantly improves the reliability of summarization based on LLMs.

Improving Factual Consistency of News Summarization by Contrastive Preference Optimization

TL;DR

Faced with factual hallucinations in news summarization by large language models, the paper introduces Contrastive Preference Optimization (CPO) to disentangle faithful versus imaginative content and a Probing-based Specific Training (PST) to identify and train weak layers. It also introduces LESSON, a sentence-level annotated dataset for training and evaluation. Empirical results across diverse backbones show CPO+PST substantially improves factual consistency and outperforms RL-based and other fine-tuning strategies, with PST offering stability and efficiency. These contributions enable more reliable, domain-adaptable news summarization in practice.

Abstract

Despite the recent progress in news summarization made by large language models (LLMs), they often generate summaries that are factually inconsistent with original articles, known as "hallucinations" in text generation. Unlike previous small models (e.g., BART, T5), current LLMs make fewer silly mistakes but more sophisticated ones, such as imposing cause and effect, adding false details, overgeneralizing, etc. These hallucinations are challenging to detect through traditional methods, which poses great challenges for improving the factual consistency of text summarization. In this paper, we propose Contrastive Preference Optimization (CPO) to disentangle the LLMs' propensities to generate faithful and fake content. Furthermore, we adopt a probing-based specific training method to improve their capacity of distinguishing two types of propensities. In this way, LLMs can execute the instructions more accurately and have enhanced perception of hallucinations. Experimental results show that CPO significantly improves the reliability of summarization based on LLMs.
Paper Structure (25 sections, 7 equations, 8 figures, 11 tables, 1 algorithm)

This paper contains 25 sections, 7 equations, 8 figures, 11 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Evaluating Factual Consistency
  4. Probing for Truthfulness
  5. Methodology
  6. Sentence-Level Data Collection
  7. Contrastive Preference Optimization
  8. Probing-based Specific Training
  9. Experiments
  10. Experimental Details
  11. Datasets
  12. Backbones
  13. Experimental Settings
  14. RQ1: Does CPO with PST improve LLM's factual consistency?
  15. RQ2: Does CPO decouple LLMs' different propensities successfully?
  16. ...and 10 more sections

Figures (8)

  • Figure 1: The diagram of the LLMs' propensities to generate faithful and fake content. In abstractive summarization, the model is supposed to generate a factually consistent summary with the preference to be faithful to the context. However, it often hallucinates with the preference to over-imagine with internal knowledge.
  • Figure 2: The diagram of our approach compared with methods based on reinforcement learning.
  • Figure 3: The diagram of our method. Based on LESSON annotated by ChatGPT and GPT-4, we adopt Incentive Loss and Penalty Loss to optimize LLMs' contrastive preferences. Meanwhile, we dynamically calculate the probing scores of each layer and employ probing specific training to select weak layers to remedy their insensitivity.
  • Figure 4: The win rate of CPO+PST on factual consistency under human evaluation.
  • Figure 5: The head-level probing results. Darker green means higher accuracy.
  • ...and 3 more figures