Analyzing and Evaluating Faithfulness in Dialogue Summarization

TL;DR

This paper investigates the faithfulness of dialogue summaries, revealing that over 35% contain factual inconsistencies, and introduces a new model-level evaluation method to improve assessment of factual correctness.

Contribution

It provides a systematic human analysis of faithfulness in dialogue summarization and proposes a novel evaluation approach using multi-choice questions for better factual assessment.

Findings

35% of summaries are factually inconsistent

Proposed evaluation method correlates well with human judgment

Released dataset and toolkit for future research

Abstract

Dialogue summarization is abstractive in nature, making it suffer from factual errors. The factual correctness of summaries has the highest priority before practical applications. Many efforts have been made to improve faithfulness in text summarization. However, there is a lack of systematic study on dialogue summarization systems. In this work, we first perform the fine-grained human analysis on the faithfulness of dialogue summaries and observe that over 35% of generated summaries are faithfully inconsistent respective the source dialogues. Furthermore, we present a new model-level faithfulness evaluation method. It examines generation models with multi-choice questions created by rule-based transformations. Experimental results show that our evaluation schema is a strong proxy for the factual correctness of summarization models. The human-annotated faithfulness samples and the evaluation toolkit are released to facilitate future research toward faithful dialogue summarization.