The Effects of Hallucinations in Synthetic Training Data for Relation Extraction

Research output: Contribution to book/Conference proceedings/Anthology/ReportConference contributionContributedpeer-review

Contributors

Abstract

Relation extraction is crucial for constructing knowledge graphs, with large high-quality datasets serving as the foundation for training, fine-tuning, and evaluating models. Generative data augmentation (GDA) is a common approach to expand such datasets. However, this approach often introduces hallucinations, such as spurious facts, whose impact on relation extraction remains underexplored. In this paper, we examine the effects of hallucinations on the performance of relation extraction on the document and sentence levels. Our empirical study reveals that hallucinations considerably compromise the ability of models to extract relations from text, with recall reductions between 19.1% and 39.2%. We identify that relevant hallucinations impair the model's performance, while irrelevant hallucinations have a minimal impact. Additionally, we develop methods for the detection of hallucinations to improve data quality and model performance. Our approaches successfully classify texts as either'hallucinated' or'clean,' achieving high F1-scores of 83.8% and 92.2%. These methods not only assist in removing hallucinations but also help in estimating their prevalence within datasets, which is crucial for selecting high-quality data. Overall, our work confirms the profound impact of relevant hallucinations on the effectiveness of relation extraction models.

Details

Original languageEnglish
Title of host publicationKBC-LM-LM-KBC 2024 - Joint proceedings of the KBC-LM workshop and the LM-KBC challenge 2024
Number of pages17
Publication statusPublished - 2024
Peer-reviewedYes

Publication series

SeriesCEUR Workshop Proceedings
Volume3853
ISSN1613-0073

Workshop

Title2nd Workshop on Knowledge Base Construction from Pre-Trained Language Models
Abbreviated titleKBC-LM 2024
Duration12 November 2024
Website
LocationLive! Casino & Hotel Maryland
CityBaltimore
CountryUnited States of America

External IDs

ORCID /0000-0001-5458-8645/work/193180545

Keywords

ASJC Scopus subject areas