Proceedings of the 13th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pages 70–73, Seoul, South Korea, 5-6 July 2012. c 2012 Association for Computational Linguistics Rapid Development Process of Spoken Dialogue Systems using Collaboratively Constructed Semantic Resources Masahiro Araki Department of Information Science Kyoto Institute of Technology Matsugasaki, Sakyo-ku, Kyoto 606-8585, Japan araki@kit.ac.jp Abstract We herein propose a method for the rapid development of a spoken dialogue system based on collaboratively constructed semantic resources and compare the proposed method with a conventional method that is based on a relational database. Previous development frameworks of spoken dialogue systems, which presuppose a relational database management system as a background application, require complex data definition, such as making entries in a task-dependent language dictionary, templates of semantic frames, and conversion rules from user utterances to the query language of the database. We demonstrate that a semantic web oriented approach based on collaboratively constructed semantic resources significantly reduces troublesome rule descriptions and complex configurations in the rapid development process of spoken dialogue systems. 1 Introduction There has been continuing interest in the development methodology of spoken dialogue systems (SDS). In recent years, statistical methods, such as Williams et al. (2007) and Hori et al. (2009), have attracted a great deal of attention as a data-driven (i.e., corpus-driven) approach, which can reduce the troublesome manual coding of dialogue management rules. Statistical methods can also be applied to other components of SDS, such as semi-automatic construction of semantic interpreters and response generators. However the overall SDS development process still requires some hand coding, for example to establish the connection to the underlying application. Another data-driven approach was designed to provide all of the SDS components with the goal of rapidly constructing the entire system (Kogure et al., 2001 Heinroth et al., 2009). This approach starts from a data model definition (and so can be regarded as a data-modeling driven approach) and adds rules and templates, which are used as task- dependent knowledge in an SDS. As a data model definition, Kogure et al. (2001) used a relational database (RDB) schema and Heinroth et al. (2009) used OWL, which is an ontology definition language in semantic web applications. Although these data-modeling schemata are familiar to developers of web applications, additional definition of rules and templates needed for an SDS is troublesome for ordinary web developers because such SDS-related rules require specialized knowledge of linguistics and speech application development. We herein propose a new data-modeling driven approach for rapid development of SDS that is based on collaboratively constructed semantic resources (CSRs). We present an automatic generation mechanism of code and data for a simple SDS. In addition, we compare the proposed approach with an ordinary data-modeling driven approach that is based on a RDB. By using CSRs and the Rails framework of web application development, the troublesome definitions of rules and templates for SDS can be reduced significantly. 70

The remainder of the present paper is organized as follows. Section 2 describes the proposed approach to a data-modeling driven development process for SDS based on CSRs. Section 3 compares the proposed approach with the previous RDB-based approach. In Section 4, the paper concludes with a discussion of future research. 2 Data-modeling driven approach based on CSRs In this section, we explain our previous data- modeling driven approach and describe additional new functionality based on CSRs. 2.1 Object-oriented SDS development framework We previously proposed a data-modeling driven framework for rapid prototyping of SDS (Araki et al., 2011). This includes a class library that is based on the class hierarchy and the attribute definitions of an existing semantic web ontology, i.e., Schema.org1. This class library is used as a base class of an application-specific class definition. An example class definition is shown in Figure 1. Figure 1: Example of class definition extending existing class library. In this example, the MyBook class inherits all of the attributes of the Book class of Schema.org in the same manner as object-oriented programming languages. The developer can limit the attributes that are used in the target application by listing them in the constraints section. On the other hand, the developer can add additional attributes (in this class, ranking attributes as the type of integer) in the definition of the class. 1 http://schema.org/ The task type and dialogue initiative type are indicated as annotations at the beginning of the class definition. In this example, the task type is DB search and the initiative type is user initiative. This information is used in generating the controller code and view code of the target SDS. Using Grails2, which is a Rails web application framework, the proposed framework generates the dialogue controller code of the indicated task type and the view code, which have speech interaction capability on the HTML5 code from this class definition. The overall concept of the object- oriented framework is shown in Figure 2. Data model definition Mix‐in of traits embed application logic State definition generate convert Grails Data model definition Groovy generate HTML5 code Model Controller View Figure 2: Overview of the object-oriented SDS development framework. 2.2 Usage of CSRs The disadvantage of our previous framework, described in the previous subsection, is the high dependence on the dictation performance of the speech recognition component. The automatically generated HTML5 code invokes dictation API, irrespective of the state of the dialogue and initiative type. In order to improve speech recognition accuracy, grammar rules (in system initiative dialogue) and/or the use of a task/domain-dependent language model (LM) (in mixed/user initiative dialogue) are necessary. In our previous framework, the developer had to prepare these ASR-related components using language resources, which are beyond the proposed data-driven framework. In order to overcome this defect, we add the Freebase3 class library, which is based on large- scale CSRs, because Freebase already includes the 2 http://grails.org/ 3 http://www.freebase.com/ @DBSearch @SystemInitiative class MyBook extends Book { int ranking static constraints = { name(onsearch:"like") author(onsearch:"like") publisher() ranking(number:true) } } 71