Analyzing and Evaluating Faithfulness in Dialogue Summarization
This work addresses the critical issue of faithfulness in dialogue summarization for practical applications, providing tools and data to support future research, though it is incremental in nature.
The paper tackles the problem of factual errors in dialogue summarization by conducting a fine-grained human analysis that reveals over 35% of generated summaries are inconsistent with source dialogues, and it introduces a new model-level evaluation method using rule-based transformations to assess faithfulness.
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.