Towards Factual Large Language Models in Low-Resource Domains

ACL ARR 2025 May Submission7301 Authors

Abstract: Direct Preference Optimization (DPO) over automatically generated factuality preference rankings has been shown to significantly improve the factuality of large language models (LLMs). However, existing approaches often rely on assumptions, such as access to comprehensive reference or a strong correlation between model confidence and factuality, that do not hold in low-resource domains. To address these limitations, we propose a method for automatically constructing factuality preference datasets from domain-specific resources such as terminologies and knowledge graphs. We introduce two novel factuality estimators: one that links entities from arbitrary domain resources to Wikipedia entries, using their articles as proxy evidence, and another that uses a judge model to estimate factuality in the absence of reliable evidence. We also conduct a systematic study of key factors affecting factuality gains in representative domains, including estimator type, verification set, preference set size, and model scale. Experiments demonstrate significant improvements in in-domain factuality without degrading downstream task performance, while showing evidence of acquired domain knowledge.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: NLP in resource-constrained settings, factuality, domain adaptation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Publicly available software and/or pre-trained models, Data resources

Languages Studied: English

Submission Number: 7301