Solving the Chinese Physical Problem Based on Deep Learning and Knowledge Graph

摘要

In recent years, problem solving, automatic proof and human-like test-tasking have become ahot spot of research. This paper focus on the study of solving physical problem in Chinese.Based on the analysis of physical corpus, it is found that the physical problem are made up of ntupleswhich contain concepts and relations between concepts, and the n-tuples can beexpressed in the form of UP-graph (The graph of understanding problem), which is the semanticexpression of physical problem. UP-graph is the base of problem solving which is generated byusing physical knowledge graph (PKG). However, current knowledge graph is hard to be usedin problem solving, because it cannot store methods for solving problem. So this paper presentsa model of PKG which contains concepts and relations, in the model, concepts and relations aresplit into terms and unique IDs, and methods can be easily stored in the PKG as concepts.Based on the PKG, DKP-solving is proposed which is a novel approach for solving physicalproblem. The approach combines rules, statistical methods and knowledge reasoning effectivelyby integrating the deep learning and knowledge graph. The experimental results over the dataset of real physical text indicate that DKP-solving is effective in physical problem solving.