Enhanced Dialogue Flow: Non-Sentential Utterance Rewrite for Co-reference and Ellipsis Resolution

Dialogue systems have emerged as a critical area of research given their vast potential in various applications. However, current models often fall short of expectations in practical multit-urn dialogues. A significant challenge lies in the frequent occurrence of utterances lacking a complete sentence structure, characterized by co-references and ellipsis. This limitation hampers the models' ability to comprehend genuine intentions, thereby impacting the performance of dialogue systems. In this paper, we advocate for the inclusion of a pre-processing step in dialogue systems to rewrite non-sentential utterances. We propose a context-aware utterance rewriter that enhances multi-turn dialogue modeling. Leveraging a substantial number of parameters and extensive training data, Large Language Models (LLMs) have demonstrated remarkable proficiency in understanding and generating intricate linguistic patterns. Consequently, we present a dialogue rewriting scheme that fine-tunes the LLM with specific instructions, enabling co-reference resolution and ellipsis completion based on contextual cues. Experimental results validate the effectiveness of our approach in improving rewriting tasks, and the proposed LLM-rewriter significantly enhances the response quality of multiturn dialogue systems.