Salience-aware Dialogue Summarization via Parallel Original-Extracted Streams
Abstract: In dialogue summarization, traditional approaches often concatenate utterances in a linear fashion, overlooking the dispersion of actions and intentions inherent in interactive conversations. This tendency frequently results in inaccurate summary generation. In response to this challenge, we formulate dialogue summarization as an extract-then-generate task. To tackle the extraction phase, we introduce an algorithm designed to identify Utterances Most related to speakers' key Intents (UMIs). These UMIs serve as labels to train an extraction model. Moving to the generation phase, we view a dialogue as parallel original-extracted streams. Correspondingly, we present a model named Row-Column Fusion Dual-Encoders and Utterance Prefix for Dialogue Summarization, abbreviated as RCUPS, with the goal of enhancing the model's ability to discern utterances and align with our sentence-level extraction. RCUPS integrates the row-column wise fusion module, which amalgamates vector representations from a dual-branch encoder. In the decoding stage, an utterance-level prefix is strategically employed to emphasize crucial details, while weight decay is applied to non-UMIs to mitigate their influence. To assess the effectiveness of RCUPS, comprehensive experiments on SAMSum, DialogSum, and TODSum datasets show significant improvements over robust baselines.
Paper Type: Long
Research Area: Summarization
Research Area Keywords: Dialogue Summarization
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 806
Loading