Contrastive Error Attribution for Finetuned Language Models

Faisal Ladhak, Esin Durmus, Tatsunori Hashimoto


Abstract
Recent work has identified noisy and misannotated data as a core cause of hallucinations and unfaithful outputs in Natural Language Generation (NLG) tasks. Consequently, identifying and removing these examples is a key open challenge in creating reliable NLG systems. In this work, we introduce a framework to identify and remove low-quality training instances that lead to undesirable outputs, such as faithfulness errors in text summarization. We show that existing approaches for error tracing, such as gradient-based influence measures, do not perform reliably for detecting faithfulness errors in NLG datasets. We overcome the drawbacks of existing error tracing methods through a new, contrast-based estimate that compares undesired generations to human-corrected outputs. Our proposed method can achieve a mean average precision of 0.93 at detecting known data errors across synthetic tasks with known ground truth, substantially outperforming existing approaches. Using this approach and re-training models on cleaned data leads to a 70% reduction in entity hallucinations on the NYT dataset and a 55% reduction in semantic errors on the E2E dataset.
Anthology ID:
2023.acl-long.643
Volume:
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Month:
July
Year:
2023
Address:
Toronto, Canada
Editors:
Anna Rogers, Jordan Boyd-Graber, Naoaki Okazaki
Venue:
ACL
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
11482–11498
Language:
URL:
https://aclanthology.org/2023.acl-long.643
DOI:
10.18653/v1/2023.acl-long.643
Bibkey:
Cite (ACL):
Faisal Ladhak, Esin Durmus, and Tatsunori Hashimoto. 2023. Contrastive Error Attribution for Finetuned Language Models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11482–11498, Toronto, Canada. Association for Computational Linguistics.
Cite (Informal):
Contrastive Error Attribution for Finetuned Language Models (Ladhak et al., ACL 2023)
Copy Citation:
PDF:
https://preview.aclanthology.org/improve-issue-templates/2023.acl-long.643.pdf
Video:
 https://preview.aclanthology.org/improve-issue-templates/2023.acl-long.643.mp4