Learning to Revise References for Faithful Summarization

Abstract

In many real-world scenarios with naturally occurring datasets, reference summaries are noisy and contain information that cannot be inferred from the source text. On large news corpora, removing low quality samples has been shown to reduce model hallucinations. Yet, this method is largely untested for smaller, noisier corpora. To improve reference quality while retaining all data, we propose a new approach: to revise-not remove-unsupported reference content. Without ground-truth supervision, we construct synthetic unsupported alternatives to supported sentences and use contrastive learning to discourage/encourage (un)faithful revisions. At inference, we vary style codes to over-generate revisions of unsupported reference sentences and select a final revision which balances faithfulness and abstraction. We extract a small corpus from a noisy source-the Electronic Health Record (EHR)-for the task of summarizing a hospital admission from multiple notes. Training models on original, filtered, and revised references, we find (1) learning from revised references reduces the hallucination rate substantially more than filtering (18.4% vs 3.8%), (2) learning from abstractive (vs extractive) revisions improves coherence, relevance, and faithfulness, (3) beyond redress of noisy data, the revision task has standalone value for the task: as a pretraining objective and as a post-hoc editor 1 . * This project was completed during an NLP research internship with Amazon Comprehend Medical.1 Code to recreate all MIMIC-III (Johnson et al., 2016) data and models (corruption, revision, and summarization) is available at https://github.com/amazon-research/ summary-reference-revision.

MisCaltral: Reducing Numeric Hallucinations of Mistral with Precision Numeric Calculation

“…This issue is particularly pronounced in tasks that require factual accuracy or domain-specific knowledge, where LLMs have been shown to fabricate details or make assertions without factual basis [14], [15], [16]. Moreover, the research has looked into the impact of training methodologies on the propensity of LLMs to hallucinate, and revealed that the choice of training datasets, the diversity of content, and the methods of data pre-processing play crucial roles in shaping the models' tendency to generate spurious outputs [17], [18], [19]. In addition to training data, the architectural aspects of LLMs have also been scrutinized, as certain designs may be more susceptible to hallucinations, suggesting a need for architectural innovations to mitigate this issue [10], [20].…”