Implementing Nature Language Interfaces To Chinese GIS Based On Semantic Parsing

Geographical Information System (GIS) and its technologies have already been widely used in many fields, such as commerce, science, planning and decision-making, and they have been a very useful tool in people’s production and life. Compared to most of the existing GIS systems, the GIS system with natural language interface is much more easy and convenient to use. However, most of previous work employed rule-based approach. Obviously, the rule-based approaches are difficult to solve the flexibility of natural language expression.On the other hand, semantic parsing has received much attention in natural language processing domain in recent years. Semantic parsing is the task of mapping a natural language (NL) sentence into a complete, formal meaning representation (MR) in a meaning representation language (MRL). Therefore, in this paper, we perform an exploratory study of applying the semantic parsing methods to implementing natural language interface (NLI) to GIS system. Specifically, we take Nanjing city map query as a specific application domain, and the corresponding research involves three major tasks:(1) We devise a formal meaning representation language (MRL) related to a specific GIS application and then create a corresponding corpus. First, we devise a formal MRL called GISQL, which is oriented to Nanjing city map query application; Second, based on the formal language GISQL, we further develop a corpus annotated with Chinese semantic parsing, which lay the foundation for further research.(2) For the preprocessing of Chinese natural language questions, we develop an effective joint model to process word segmentation and Geographical Entity Recognition (GER). The experimental results show that the joint model achieves significant performance improvement compared with the baseline system based on CRF.(3) In order to implement the mapping from natural language questions into formal meaning representation, we propose a discriminative latent structural prediction approach for semantic parsing with the use of latent variable perceptron. The experimental results show that the proposed method significantly outperforms the baseline method.