Semantic Dependency Graph Parsing of Financial Domain Questions Based on Deep Learning

Abstract In order to effectively solve the problems of limited accuracy of semantic parsing and ambiguous semantic recognition in the current Question-Answer System based on Chinese Knowledge Map in intention recognition. This paper presents a method about Dependency Reduction in language dependency analysis(DR-BLSTM-CRF) based on Semantic Dependency Graph Parsing(SDGP), Bidirectional Long Short-term Memory(BLSTM) model and Conditional Random Field(CRF) of Xunfei Open Platform. The semantic dependency graph dependency reduction method is as follows:1) Semantic Dependency Graph Parsing of interrogative sentences based on Web API of Xunfei Open Platform to obtain a sentence representation containing semantic dependency information. 2) Named Entity Recognition (NER) algorithm combined with BLSTM and CRF is used to identify the named entity of the interrogative sentence to gain a sequence containing character label information, and then combine the analysis results of semantic dependency graph to obtain a more accurate semantic dependency graph through dependency reduction. The experimental results show that the precision, recall and F1 value of the model proposed in this paper are 33.4%, 33.9% and 34.2% higher than those of LTP in terms of the semantic dependency analysis on the 140,000 self-built data sets of financial domain questions. The model can effectively analyze the semantic dependency of financial domain questions.