您的位置: 首页  科研动态

Two-phase reanalysis model for understanding user intention.[Pattern Recognition Letters]

发布日期: 2014-02-27   浏览次数 17

Sangwoo Kang a, Jungyun Seo b
a Department of Computer Science, Sogang University, South Korea
b Department of Computer Science and Interdisciplinary Program of Integrated Biotechnology, Sogang University, South Korea

    Literature on dialogue systems for end-user applications has seen a rapid growth in recent years. The aim of a dialogue system is to perform tasks based on natural language processing, and thus has the potential to significantly improve interaction between users and machines. A conventional dialogue system consists of the following components: natural language understanding, dialogue management, and response generation. Natural language understanding is intended to interpret user intention as accurately as possible, by analyzing input sentences, which is required for dialogue management to generate adequate responses to users. User intention is translated into three attributes, namely of Speech Act (SA), Concept Sequence (CS), and a set of Arguments (ARGs). SA describes an utterance that attempts to affect the addressee, while CS consists of a set of concepts representing domain-dependent actions, and ARGs denote essential details of CS in pairs comprising data type and corresponding value.
    This paper proposes a two-phase reanalysis model for understanding user intention in utterances, by considering the correlative characteristics between the three attributes relating to user intention. The proposed model comprises two phases. In the first phase, each attribute is analyzed in the optimized sequence. The results of the analysis are then used as features that undergo reanalysis in the second phase, with the assumption that the relationship between the attributes is correlative. The experiments conducted showed that the proposed model improves user intention analysis over the baseline model, with an error reduction rate in Speech Act, Concept Sequence, and Arguments of 0.64%, 14.78%, and 5.84%, respectively.