Reducing Hallucinations in Summarization via Reinforcement Learning with Entity Hallucination Index

2025-07-30

AI Summary

To address named entity hallucination—i.e., the generation of entities absent from or inconsistent with the source text—in abstractive summarization, this paper proposes a reinforcement learning framework that requires no human-annotated factual labels. The core contribution is the Entity Hallucination Index (EHI), a computationally tractable and differentiable reward signal quantifying factual deviation by automatically extracting and aligning named entities between summaries and source documents. Built upon a pre-trained language model, our method jointly optimizes EHI alongside standard language modeling objectives during summary generation. Experiments demonstrate significant reductions in EHI, substantial improvements in entity accuracy and factual consistency, and preservation of summary fluency and informativeness. To ensure reproducibility, we release both source code and a fully executable Colab notebook.

Technology Category

Natural Language Processing: SummarizationMachine Learning: Large Multimodal Models (LMMs)Data Mining & Knowledge Management: Data Visualization & Summarization

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Fairness-aware retrieval and rankingGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Reducing hallucinations in abstractive summarization remains a critical challenge for deploying language models (LMs) in real-world settings. In this work, we introduce a rewarddriven fine-tuning framework that explicitly optimizes for Entity Hallucination Index (EHI), a metric designed to quantify the presence, correctness, and grounding of named entities in generated summaries. Given a corpus of meeting transcripts, we first generate baseline summaries using a pre-trained LM and compute EHI scores via automatic entity extraction and matching. We then apply reinforcement learning to fine-tune the model parameters, using EHI as a reward signal to bias generation toward entity-faithful outputs. Our approach does not rely on human-written factuality annotations, enabling scalable fine-tuning. Experiments demonstrate consistent improvements in EHI across datasets, with qualitative analysis revealing a significant reduction in entity-level hallucinations without degradation in fluency or informativeness. We release a reproducible Colab pipeline, facilitating further research on hallucination-aware model fine-tuning using lightweight, hallucintion metrics like EHI.

Problem

• Reducing entity hallucinations in abstractive summarization models
• Optimizing summaries using Entity Hallucination Index (EHI) metric
• Fine-tuning models via reinforcement learning without human annotations

Innovation

• Reinforcement learning optimizes Entity Hallucination Index
• Automatic entity extraction enables scalable fine-tuning
• Lightweight EHI metric reduces hallucinations without quality loss