Automatic Utterance Generation in Consideration of Nominatives and Emoticon Annotation

Automatic Utterance Generation in Consideration of Nominatives and Emoticon Annotation Yusuke Nishio Mirai Miura Dongli Han Graduate School of Integrated Basic Sciences Graduate School of Information Science Department of Information Science Nihon University, Tokyo, JAPAN Nara Institute of Science and Technology Nara, JAPAN College of Humanities and Sciences Nihon University, Tokyo, JAPAN han@chs.nihon- u.ac.jp Abstract The demand is increasing recently for non- task-oriented conversation system in various scenes. Previous studies provide various strategies to enrich the methods for generating utterances, thus making the conversation systems or agents appear more interesting. However, most previous works tend to rely on templates and therefore are not able to perform flexible conversation- utterance generation. We propose here in this paper a thorough modification to a previous work to address this problem. Specifically, we introduce an automatic utterance generation in consideration of the embedded structure of sentences based on the principle of nominative maintenance. Moreover, emotion presumption has been implemented to add entertaining elements into the conversation with a user. The experimental results show that our approach proposed in this study has helped improve the performance of a conversation system. 1 Introduction Non-task-oriented conversation systems have been developed during the past decade. They pay more attention to continuing the conversation by any means rather than the rigorousness of the utterance’s content in comparison with task- oriented ones. However, the insufficiency in methods for generating utterances still remains as a critical issue unsolved. For example, Higuchi et al. concentrate on modalities appearing in human’s utterances, and try to incorporate them into the process of utterance generation (Higuchi et al., 2008). Song et al. (Song et al., 2009) and Han et al. (Han et al., 2010) present a strategy to provide new topics for users in a free conversation system at the point the system “considers” that the user has lost interest in the current topic. As just described, most previous studies provide various strategies to enrich the methods for generating utterances, so that the user might feel interested in the system and intend to continue the conversation. However, none of them could escape from the fact that they all generate utterances depending mainly on some particular kinds of templates or augmented templates. As a case study to cope with this problem, Han et al. develop a free conversation system employing Markov sequences as shown in Figure 1 (Han et al., 2011; Nishio and Han, 2012). They use the topic-word pair extracted beforehand to search the Twitter for snippets that contain a noun in the beginning and a verb or adjective in the end, and then generate an utterance employing a Two-starting-word style Markov connection. Although this approach has been proven quite effective in promoting the human-like qualities of utterances, a significant problem has been observed simultaneously: Utterance Focus PACLIC-27 66 Copyright 2013 by Yusuke Nishio, Mirai Miura, and Dongli Han 27th Pacific Asia Conference on Language, Information, and Computation pages 66－74 usually spreads around too widely. This seems to come from the nature of the two-starting-word …

Content truncated to 500 words