Unsupervised Approach for Dialogue Act Classification

Unsupervised Approach for Dialogue Act Classification Kiyonori Ohtake National Institute of Information and Communications Technology (NICT)/ Advanced Telecommunications Research Institute International (ATR) 2-2-2 Hikaridai, Keihanna Science City 619-0288, JAPAN kiyonori.ohtake (at) nict go.jp Abstract. This paper presents an unsupervised approach for dialogue act (DA) classification. We used a latent variable model to compress the dimensions of the feature vector. We introduced a paraphraser to reduce the variety of expressions and to solve the pragmatic problem for DA classification. The paraphraser seemed to work well on some DA classifications in the unsupervised approach. The results obtained by the unsupervised approach were compared with the manually annotated labels. A preliminary experiment for semi-supervised tagging was also carried out, and we discuss these results. Keywords: Unsupervised, Dialogue Act, Paraphrasing, Latent Variable Model. 1. Introduction Recognizing the intentions of a user in a dialogue system is very important. So far, many methods have been developed to infer a user’s intention in a dialog situation. To infer the user’s intention in an utterance, the utterance can be categorized into given classes. Therefore, many studies have designed the classes called dialogue act (DA) labels that approximate a speaker’s intention. They annotated the labels on a corpus to analyze the phenomena for DA interaction or to develop a DA tagger in order to infer the DA label from a speech segment (e.g. some utterances, an utterance, or a part of an utterance). Most studies on DA taggers were based on a supervised method (e.g., (Stolcke et al., 2000; Tanaka and Yokoo, 1999)). The labels used in a DA tagger have to be predefined, and supervised methods require a corpus that is manually annotated by the labels. On the other hand, it is difficult to design a tag set (labels) that can be used to annotate a corpus because the design of a tag set depends on the domain and the task. Therefore, we have to redesign the tag set and construct a corpus annotated with a new tag set if we apply our system to different domains or tasks. In addition, designing a tag set that can be used in any domain or task is very difficult. However, we have to annotate DA tags on a corpus, because many applications require predefined DA tags. This paper discusses an unsupervised approach to infer the user’s intention in a situation by using a dialog system. Unsupervised approach may not achieve highly accurate results when compared to the supervised approach. However, in any domain or task, the unsupervised approach can yield human DA annotators with machine judgments of the DA classification that may be useful to keep the consistency of DA annotation results for a corpus. In addition, annotating a corpus with given labels is very time-consuming. An unsupervised method is independent of annotation and designing the tag set. In order to achieve an unsupervised method, we need an unsupervised clustering method. So far, many clustering methods have been proposed and discussed for applications in natural language processing (NLP), such …

Content truncated to 500 words