Towards Neural Speaker Modeling in Multi-Party Conversation: The Task, Dataset, and Models

Towards Neural Speaker Modeling in Multi-Party Conversation: The Task, Dataset, and Models Zhao Meng,1,2 Lili Mou,1,3 Zhi Jin1,∗ 1Key Laboratory of High Confidence Software Technologies, MoE; Software Institute, Peking University 2Department of Computer Science, ETH Zurich 3David R. Cheriton School of Computer Science, University of Waterloo zhmeng@student.ethz.ch, doublepower.mou@gmail.com, zhijin@sei.pku.edu.cn Abstract Neural network-based dialog systems are attracting increasing attention in both academia and industry. Recently, researchers have begun to realize the importance of speaker modeling in neural dialog systems, but there lacks established tasks and datasets. In this paper, we propose speaker classification as a surrogate task for general speaker modeling, and collect massive data to facilitate research in this direction. We further investigate temporal-based and content-based models of speakers, and propose sev- eral hybrids of them. Experiments show that speaker classification is feasible, and that hybrid models outperform each single component. Keywords: Speaker Classification, Speaker Modeling, Multi-Party Conversation 1. Introduction Human-computer conversation has long attracted attention in both academia and industry. Researchers have developed a variety of approaches, ranging from rule-based systems for task-oriented dialog (Ferguson et al., 1996; Graesser et al., 2005) to data-driven models for open-domain conversa- tion (Ritter et al., 2011). A simple setting in the research of dialog systems is context-free, i.e., only a single utterance is considered dur- ing reply generation (Shang et al., 2015). Other stud- ies leverage context information by concatenating several utterances (Sordoni et al., 2015) or building hierarchi- cal models (Yao et al., 2015; Serban et al., 2016). The above approaches do not distinguish different speakers, and thus speaker information would be lost during conversation modeling. Speaker modeling is in fact important to dialog systems, and has been studied in traditional dialog research. How- ever, existing methods are usually based on hand-crafted statistics and ad hoc to a certain application (Lin and Walker, 2011). Another research direction is speaker modeling in a multi-modal setting, e.g., acoustic and vi- sual (Uthus and Aha, 2013), which is beyond the focus of this paper. Recently, neural networks have become a prevailing tech- nique in both task-oriented and open-domain dialog sys- tems. After single-turn and multi-turn dialog studies, a few researchers have realized the role of speakers in neural con- versational models. Li et al. (2016) show that, with speaker identity information, a sequence-to-sequence neural dialog system tends to generate more coherent replies. In their approach, a speaker is modeled by a learned vector (also known as an embedding). Such method, unfortunately, re- quires massive conversational data for a particular speaker to train his/her embedding, and thus does not generalize to rare or unseen speakers. Ouchi and Tsuboi (2016) formalize a new task of addressee selection on online forums: by leveraging either the tem- ∗Corresponding author. poral or utterance information, they predict whom a post is talking to. While tempting for benchmarking speaker mod- eling, the task requires explicit speaker ID mentions, which occurs occasionally, and thus is restricted. In this paper, we propose a speaker classification task that predicts the speaker of an utterance. …

Content truncated to 500 words