Analysis of Implicit Conditions in Database Search Dialogues

Analysis of Implicit Conditions in Database Search Dialogues Shun-ya Fukunaga Hitoshi Nishikawa Hikaru Yokono Tetsuro Takahashi Takenobu Tokunaga Tokyo Institute of Technology Fujitsu Laboratories Ltd. fukunaga.s.ab@m.titech.ac.jp, {hitoshi, take}@c.titech.ac.jp {yokono.hikaru, takahashi.tet}@jp.fujitsu.com Abstract This paper reports an annotation to a corpus of database search dialogues on real estate, and the analysis on implicit information in the utterances for constructing database queries. Two annotators annotated 50 dialogues with a set of database field tags, resulting in a high inter-annotator agreement (Cohen’s κ=0.79), and the analysis revealed that 10% of the utterances included non-database-field information. We further investigated these utterances to find that more than 93% of them included useful information for figuring out search conditions, which we call the implicit condition. The contribution of this paper is to present the existence and importance of the implicit conditions in the database search dialogues and both qualitative and quantitative analysis of them. Our corpus can provide a fundamental language resource for constructing a dialogue system which can utilise the implicit conditions. The paper concluded with possible approaches to achieve our long-term goal, extracting the implicit conditions in the database search dialogues and utilising them to construct queries. Keywords: database search dialogue, implicit condition, corpus annotation 1. Introduction The information that a dialogue system should extract from user utterances highly depends its backend applica- tion. When being used as a natural language interface of a database system, the dialogue system should be able to extract pieces of information corresponding to the record fields of the database for constructing a query. There have been several attempts to extract this type of information from user utterances in database search dialogues. For instance, several studies (Raymond and Riccardi, 2007; Mesnil et al., 2015; Liu and Lane, 2016) tried to extract values for the database field defined in the ATIS (The Air Travel Information System) corpus (Hemphill et al., 1990; Dahl et al., 1994) from the user utterances. The ATIS cor- pus includes a set of dialogues between users and an air travel system that were collected through the Wizard-of-Oz method. The tags corresponding to the backend database fields (e.g. departure city, arrival date, etc.) were annotated to the expressions in the user utterances. However, the ut- terances in real dialogues include information that does not always directly correspond to a database field but provides useful information for constructing database queries. It will be more efficient and natural if the dialogue system can utilise this type of information for retrieving the database. For instance, in real estate search dialogues, which is our target domain, the number of family members provides use- ful information for deciding the size of a house, but it is rarely a field of the real estate database. The number of family members is an attribute of the customer rather than that of houses. We call this type of information the implicit condition. Our long-term goal is to establish a method that is capable of extracting implicit conditions from the user utterances in the database search dialogue …

Content truncated to 500 words