Research on Geometry Problem Text Understanding Based on Bidirectional LSTM-CRF
Wei Duo Zhou, Ruixi Xu, Hao Guan, Jietong Zhao, Yongsheng Rao · 2022
The challenge of enabling a computer to understand a geometry problem in natural language has been attracting increasingly attention. As society becomes more data oriented, knowledge graphs (KG) that represents structured relations between entities are also gaining ground. However, it is still brand new that to combine knowledge graph with geometry problem understanding. Focusing on understanding geometry problem text in Chinese natural language, we propose two methods: (1) three-level representation normalization; (2) acquiring structured geometric relation by integrating entity extraction using Bilstm-CRF and constructing relation triad using template mapping. Based on these, we construct a system that implements obtaining geometric relation triads and finally put them in knowledge graph. With our three-level representation normalization, the accuracy in its subsequent processing, entity recognition and extraction, has significantly and constantly improved about 3 percent compared with directly processing on the raw problem text with simple data cleaning. The results of the final test confirm the possibility that the system successfully recognizes and obtains all geometry relation triads of each problem sentence in the entire testing set, achieving complete transformation from an unstructured geometry problem text to a structured relation triad set.