Scholarlink

<p>Recently, multi-party conversation (MPC) analysis has received increased attention due to its widespread use in understanding retrieval-based or generation-based complex dialog systems. Even though multiple work has been conducted under MPC-based utterance structure modelling, much work remains to be done for MPC-based utterance semantics modelling to address the research question, <i>“How can the discourse level semantics be used to improve utterance-interlocutor MPC modelling?”</i>. In this study, we propose <i>WSW 2.0</i>, a novel pre-trained language model for MPC-based utterance semantics modelling. The objective is to design new self-supervised tasks of utterance semantics modelling using the complex discourse structures to capture retrieval-based semantics, facilitating several downstream tasks such as reply utterance selection (RUS) and speaker identification (SI) for MPC understanding. To our knowledge, this is the first attempt to use the semantic relevance of sub-level utterances and the exact root of multiple shared sub-level utterances of an MPC for graphical-based MPC understanding. Experimental results show that our model achieves new state-of-the-art performance for RUS and SI, which translates to improvements of up to 21.64% in recall (R<sub><i>n</i></sub><i>@k</i>) and 16.23% in Precision@1 (P@1).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Who Says What (WSW) 2.0: Utterance Semantic Modelling for Speaker Identification in Text-Based Multi-Party Conversations

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Y. H. P. P. Priyadarshana,
  • Zilu Liang,
  • Ian Piumarta

摘要

Recently, multi-party conversation (MPC) analysis has received increased attention due to its widespread use in understanding retrieval-based or generation-based complex dialog systems. Even though multiple work has been conducted under MPC-based utterance structure modelling, much work remains to be done for MPC-based utterance semantics modelling to address the research question, “How can the discourse level semantics be used to improve utterance-interlocutor MPC modelling?”. In this study, we propose WSW 2.0, a novel pre-trained language model for MPC-based utterance semantics modelling. The objective is to design new self-supervised tasks of utterance semantics modelling using the complex discourse structures to capture retrieval-based semantics, facilitating several downstream tasks such as reply utterance selection (RUS) and speaker identification (SI) for MPC understanding. To our knowledge, this is the first attempt to use the semantic relevance of sub-level utterances and the exact root of multiple shared sub-level utterances of an MPC for graphical-based MPC understanding. Experimental results show that our model achieves new state-of-the-art performance for RUS and SI, which translates to improvements of up to 21.64% in recall (Rn@k) and 16.23% in Precision@1 (P@1).