Context-Dependent Automatic Response Generation Using Statistical Machine Translation Techniques

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1345–1350, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Context-Dependent Automatic Response Generation Using Statistical Machine Translation Techniques Andrew Shin Ryohei Sasano Hiroya Takamura Manabu Okumura Tokyo Institute of Technology shin@lr.pi.titech.ac.jp {sasano,takamura,oku}@pi.titech.ac.jp Abstract Developing a system that can automatically respond to a user’s utterance has recently be- come a topic of research in natural language processing. However, most works on the topic take into account only a single preceding ut- terance to generate a response. Recent works demonstrate that the application of statisti- cal machine translation (SMT) techniques to- wards monolingual dialogue setting, in which a response is treated as a translation of a stim- ulus, has a great potential, and we exploit the approach to tackle the context-dependent re- sponse generation task. We attempt to extract relevant and significant information from the wider contextual scope of the conversation, and incorporate it into the SMT techniques. We also discuss the advantages and limitations of this approach through our experimental re- sults. 1 Introduction Various approaches have been applied to the re- sponse generation task, each with its own merits and drawbacks. While one of the main concerns on the topic has been the semantic relevance of the re- sponse, it has mostly been discussed in terms of a limited conversational scope, mostly a single utter- ance. This provides us with a room for research on a wider scope of conversation, which reflects not only a single preceding utterance, but the overall context of the current conversation. SMT-based data-driven approach to the response generation task was recently introduced by Ritter et al. (2011). They demonstrated that it was better-suited for response generation than some of the previous approaches, including information re- trieval approach. We exploit this model to address the above-mentioned problem of reflecting a wider scope of conversation. We present a context-dependent model where we attempt to generate more semantically relevant and diverse responses by adding the semantically impor- tant words from previous utterances to the most re- cent one. By doing so, we hope not only to diversify the responses, but also to be able to take semantics from broader scope of the conversation into account. 2 Response Generation using SMT 2.1 Overview Ritter et al. (2011) remarked that stimulus-response pairs in the same language often have a strong struc- tural resemblance, as shown in the example conver- sation below, that may be exploited in SMT plat- forms. In the usual SMT setting, a string f in a source language is translated into a string e in a target language according to probability distribution p(e|f) (Brown et al., 1993). Ritter et al. applied the SMT techniques to monolingual conversation set- ting, and treated the response as the translation of the stimulus. Stimulus: What is your hobby? Response: My hobby is hiking. 2.2 Challenges Although the application of SMT to the response generation task demonstrates potentials, it has a …

Content truncated to 500 words