# Fine-grained Hallucination Detection and Editing for Language Models

**Authors:** Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Graham Neubig, Yulia Tsvetkov, Hannaneh Hajishirzi  
**arXiv:** [2401.06855](https://arxiv.org/abs/2401.06855)  
**Date:** 2024-01-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.06855v1/

## Key points

- Authors address the issue of hallucinations in large language models (LMs) that generate factually incorrect statements
- Introduce a novel task of automatic fine-grained hallucination detection to address the issue
- Present a comprehensive taxonomy categorizing hallucinations into six hierarchically defined types for more nuanced error detection and correction
- Propose FAVA (Fine-grained Automatic VAlidation model) to detect and correct fine-grained hallucinations, outperforming ChatGPT in both automatic and human evaluations
- Research contributes to improving factuality in LM-generated text by enhancing FActScores by 5-10%
- Offer a detailed and comprehensive method for detecting and correcting hallucinations in language models at finer levels of granularity
- Use carefully designed synthetic data for accurate evaluation and improvement of fine-grained hallucination detection methods

## Summary

In their paper "Fine-grained Hallucination Detection and Editing for Language Models," Abhika Mishra, Akari Asai, Vidhisha Balachandran, Yizhong Wang, Graham Neubig, Yulia Tsvetkov, and Hannaneh Hajishirzi address the issue of hallucinations in large language models (LMs). These LMs often generate factually incorrect statements, known as hallucinations, which can hinder their practical use. <br><br>

 The authors introduce a novel task of automatic fine-grained hallucination detection to address the issue of hallucinations in large language models.<br><br>

While existing approaches focus on coarse-grained automatic hallucination detection or editing, this paper presents a comprehensive taxonomy that categorizes hallucinations into six hierarchically defined types. This allows for a more nuanced approach to detecting and correcting errors at finer levels of granularity. <br><br>

To evaluate the effectiveness of their approach, the researchers introduce a new benchmark that includes fine-grained human judgments on outputs from two LMs across various domains. Their analysis reveals that popular models like ChatGPT and Llama 2-Chat exhibit hallucinations in a significant percentage of their outputs. Moreover, many of these hallucinations fall into categories that have not been extensively explored. <br><br>

In response to these findings, the authors propose FAVA (<b>Fine-grained Automatic VAlidation model</b>), a retrieval-augmented LM trained to detect and correct fine-grained hallucinations. Through carefully designed synthetic data generations, FAVA significantly outperforms ChatGPT in fine-grained hallucination detection according to both automatic and human evaluations. <br><br>

 This research contributes to addressing the challenge of hallucinations in language generation models by introducing a more nuanced approach to detecting and correcting errors at finer levels of granularity.<br><br>

While there is room for further improvement, FAVA's suggested edits also enhance the factuality of LM-generated text by improving FActScores by 5-10%. <br><br>

 The authors' approach offers a more detailed and comprehensive method for detecting and correcting hallucinations in language models, leading to improved factuality in generated text.<br><br>

 Through their proposed model, FAVA, the authors not only address the issue of hallucinations but also improve the overall factuality of language model-generated text.<br><br>

 The use of carefully designed synthetic data allows for more accurate evaluation and improvement of fine-grained hallucination detection methods. Overall, this research contributes to addressing the challenge of hallucinations in language generation models by introducing a more nuanced approach to detecting and correcting errors at finer levels of granularity.