Improving Faithfulness by Augmenting Negative Summaries from Fake Documents

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11913 - 11921 December 7-11, 2022 ©2022 Association for Computational Linguistics Improving Faithfulness by Augmenting Negative Summaries from Fake Documents Tianshu Wang1, Faisal Ladhak2, Esin Durmus3, He He1 1New York University, 2 Columbia University, 3 Stanford University {tw2112, hhe}@nyu.edu, faisal@cs.columbia.edu, esdurmus@stanford.edu Abstract Current abstractive summarization systems tend to hallucinate content that is unfaithful to the source document, posing a risk of mis- information. To mitigate hallucination, we must teach the model to distinguish halluci- nated summaries from faithful ones. However, the commonly used maximum likelihood train- ing does not disentangle factual errors from other model errors. To address this issue, we propose a back-translation-style approach to augment negative samples that mimic fac- tual errors made by the model. Specifically, we train an elaboration model that generates hallucinated documents given the reference summaries, and then generates negative sum- maries from the fake documents. We incorpo- rate the negative samples into training through a controlled generator, which produces faith- ful/unfaithful summaries conditioned on the control codes. Additionally, we find that adding textual entailment data through multi- tasking further boosts the performance. Ex- periments on three datasets (XSum, GigaWord, and WikiHow) show that our method consis- tently improves faithfulness without sacrific- ing informativeness according to both human and automatic evaluation.1 1 Introduction Despite the fast progress in fluency and coherence of text summarization systems, a common chal- lenge is that the generated summaries are often unfaithful to the source document, containing hal- lucinated, non-factual content (Cao et al., 2018; Falke et al., 2019, inter alia). Current summa- rization models are usually trained by maximum likelihood estimation (MLE), where unfaithful and faithful summaries are penalized equally if they both deviate from the reference. As a result, when 1Our code is available at https://github.com/ COFE2022/CoFE. Controlled Generator zimbabwe crisis deal faces international doubters Summarizer zimbabwe crisis deal faces international doubters Summary Elaborator zimbabwe’s president robert mugabe and opposition leader morgan tsvangirai will sign a deal... Document [ENT] [CON] + zimbabwe’s president, robert mugabe, and the opposition leader morgan tsvangirai reached a deal on thursday ... Fake Document mugabe and opposition leader reached deal to end crisis Figure 1: Overview of CoFE. The original and fabri- cated document-summary pairs are shown in blue and red respectively. The trained elaborator first generates fake documents from the summary. Then, the summa- rizer generates summaries from the fake documents, which are likely to contain hallucinated information (underlined). A controlled generator is then trained to produce the original (faithful) and the fabricated (un- faithful) summaries depending on the control codes. the model fails to imitate the reference, it is likely to “over-generalize” and produce hallucinated con- tent. In this work, we address the issue by explicitly teaching the model to discriminate between posi- tive (groundtruth) and negative (unfaithful) sum- maries. The key challenge is to generate realis- tic negative samples. Existing work on negative data augmentation mainly focuses on corrupting the reference (e.g., replacing entities) or sampling …

Content truncated to 500 words