Improving Korean Speech Acts Analysis by Using Shrinkage and Discourse Stack

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 731 –741, 2005. © Springer-Verlag Berlin Heidelberg 2005 Improving Korean Speech Acts Analysis by Using Shrinkage and Discourse Stack Kyungsun Kim1, Youngjoong Ko2, and Jungyun Seo3 1 Information Retrieval Division, Diquest.Inc, Seocho-dong, Seocho-gu, Seoul, 137-070, Korea kksun@diquest.com 2 Dept. of Computer Engineering, Dong-A University, 840, Hadan 2-dong, Saha-gu, Busan, 604-714, Korea yjko@dau.ac.kr 3 Dept. of Computer Science and Interdisciplinary Program of Integrated Biotechnology, Sogang University, Seoul, 121-742, Korea seojy@sogang.ac.kr Abstract. A speech act is a linguistic action intended by a speaker. It is impor- tant to analyze the speech act for the dialogue understanding system because the speech act of an utterance is closely tied with the user’s intention in the utter- ance. This paper proposes to use a speech acts hierarchy and a discourse stack for improving the accuracy of classifiers in speech acts analysis. We first adopt a hierarchical statistical technique called shrinkage to solve the data sparseness problem. In addition, we use a discourse stack in order to easily apply discourse structure information to the speech acts analysis. From the results of experi- ments, we observed that the proposed model made a significant improvement for Korean speech acts analysis. Moreover, we found that it can be more useful when training data is insufficient. 1 Introduction To understand a natural language dialogue, a dialogue system must be able to make out the speaker’s intentions indicated by utterances. Since the speech act of an utter- ance is very important in understanding a speaker’s intentions, it is an essential part of a dialogue system. However, it is difficult to infer the speech act from a surface utter- ance because the utterance may represent more than one speech act according to the context [5][7]. Various machine learning models have been used to efficiently classify speech acts such as MEM (Maximum Entropy Model) [1], HMM (Hidden Markov Model) with Decision Tree [8][11], Neural Network Model [5]. And there are also studies on methods of automatically selecting efficient features with useful information for speech acts analysis [5][10]. Since the machine learning models can efficiently ana- lyze a large quantity of data and consider many different feature interactions, they can provide a means of associating features of utterances with particular speech acts. Generally, it is hard to create enough the number of examples for each speech act in the training examples. Thus this situation has been one of the main causes for errors occurred in speech acts analysis. That is, the sparse data problem from low 732 K. Kim, Y. Ko, and J. Seo frequency of some speech acts has commonly occurred in the previous research [8]. Due to the problem, the accuracy of each speech act in previous research tends to be proportional to the frequency of each speech act in the training data. Therefore, we first focus on how to scale up statistical learning methods to solve the sparseness problem of training data in speech acts analysis. Then we propose to construct …

Content truncated to 500 words