Multi-Granularity Information Interaction Framework for Incomplete Utterance Rewriting

TL;DR

The paper addresses incomplete utterance rewriting by identifying the need to locate source words and avoid irrelevant utterances. It introduces a multi-task information interaction framework with context selection, edit matrix construction, relevance merging, and intention check to capture multi-granularity semantic information. The approach leverages BERT-based context encoding and a token-level edit matrix guided by sentence-level relevance, achieving new state-of-the-art on Restoration-200K and competitive results on CANARD. This work demonstrates that combining sentence-level relevance with token-level edits can improve rewrite quality, albeit with higher computation, and points toward future integration with LLMs for efficiency and scalability.

Abstract

Recent approaches in Incomplete Utterance Rewriting (IUR) fail to capture the source of important words, which is crucial to edit the incomplete utterance, and introduce words from irrelevant utterances. We propose a novel and effective multi-task information interaction framework including context selection, edit matrix construction, and relevance merging to capture the multi-granularity of semantic information. Benefiting from fetching the relevant utterance and figuring out the important words, our approach outperforms existing state-of-the-art models on two benchmark datasets Restoration-200K and CANAND in this field. Code will be provided on \url{https://github.com/yanmenxue/QR}.

Figures (1)

• Figure 1: Model pipeline. Our model contains 4 parts: context selection, edit matrix construction, relevance merging and intention check.