Chinese Semantic Parsing Based On Machine Translation

Semantic parsing is the task of transforming a natural language utterance into a machine-interpretable logical expression. In recent years, semantic parsing in English has made great progress. Specially, the machine translation-based semantic parsing method has achieved a lot of good results in English semantic parsing task. Chinese semantic analysis researches mainly focused on shallow semantic analysis tasks, such as information extraction, semantic role labeling, etc., however, little work has been done in deep Chinese semantic parsing task. There are inherent differences between Chinese and English, therefore, one cannot simply apply methods that are feasible for English to Chinese. In this paper, we propose a machine translation-based approach for Chinese semantic parsing, which considers the characteristics of Chinese. The main contribution of this paper includes:1) We translate the popular English dataset GEOQUERY into Chinese. English GEOQUERY includes 880 natural language sentences and their corresponding logical expressions, in order to build the training dataset for Chinese semantic parsing, we translate it manually into Chinese. To reduce the training errors brought by improper Chinese word segmentation, we segmented the Chinese natural language sentences manually.2) We propose a machine translation-based method for Chinese semantic parsing, and build the Chinese Semantic Parsing Model (CSPM). Experiments on the Chinese GEOQUERY dataset show that CSPM can achieve good results (F1 is 63.35%), which means that the machine translation-based semantic parsing method is suitable for Chinese semantic parsing task.3) We also propose the Improved Chinese Semantic Parsing Model (ICSPM) to fix the problem that many logical expressions generated by CSPM break the standard structure of logical expressions. ICSPM uses the language model to filter the nonstandard logical expressions. Experiments on the Chinese GEOQUERY dataset show that our ICSPM can achieve better results (F1 is 75.60%), which means that using language model to filter the logical expressions can effectively improve the performance of Chinese Semantic Parsing Model.