Conversation Initiation by Diverse News Contents Introduction

Proceedings of NAACL-HLT 2019, pages 3988–3998 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3988 Conversation Initiation by Diverse News Contents Introduction Satoshi Akasaki∗ The University of Tokyo akasaki@tkl.iis.u-tokyo.ac.jp Nobuhiro Kaji Yahoo Japan Corporation nkaji@yahoo-corp.jp Abstract In our everyday chit-chat, there is a conver- sation initiator, who proactively casts an ini- tial utterance to start chatting. However, most existing conversation systems cannot play this role. Previous studies on conversation systems assume that the user always initiates conversa- tion, and have placed emphasis on how to re- spond to the given user’s utterance. As a result, existing conversation systems become passive. Namely they continue waiting until being spo- ken to by the users. In this paper, we consider the system as a conversation initiator and pro- pose a novel task of generating the initial utter- ance in open-domain non-task-oriented con- versation. Here, in order not to make users bored, it is necessary to generate diverse utter- ances to initiate conversation without relying on boilerplate utterances like greetings. To this end, we propose to generate initial utterance by summarizing and chatting about news arti- cles, which provide fresh and various contents everyday. To address the lack of the training data for this task, we constructed a novel large- scale dataset through crowd-sourcing. We also analyzed the dataset in detail to examine how humans initiate conversations (the dataset will be released to facilitate future research activ- ities). We present several approaches to con- versation initiation including information re- trieval based and generation based models. Experimental results showed that the proposed models trained on our dataset performed rea- sonably well and outperformed baselines that utilize automatically collected training data in both automatic and manual evaluation. 1 Introduction Conversation1 systems are becoming increasingly important as a means to facilitate human-computer ∗This work was done during research internship at Yahoo Japan Corporation. 1“Conversation” in this paper refers to open-domain non- task-oriented conversations and chit-chat. Figure 1: Conversation initiation task. The system in this example is given a news post about “iPhone” and generates an initial utterance for chatting about it. communication. However, most of the studies on conversation systems have been based on the as- sumption that a human always initiates conver- sation. As a result, the systems are designed to be passive (Yan, 2018), meaning that they keep waiting until they are spoken to by the hu- man and will never speak to the human proac- tively. For example, popular encoder-decoder models (Sutskever et al., 2014; Vinyals and Le, 2015) are designed to respond to input utterances provided by humans, and it is difficult for them to proactively initiate the conversation. Although some systems are able to initiate conversations, they basically adopt template-based generation methods and thus lack diversity. This paper investigates generating the very first utterance in a conversation. We feel strongly that conversation systems should not always be pas- sive; sometimes, they have to proactively initiate the conversation to enable more natural conversa- tion. In addition, it is …

Content truncated to 500 words