| Plane geometry problem understanding is an important research problem in the field of natural language processing and intelligent education.It has become a research hotspot in recent years and has important application prospects in the field of education.Plane geometry problem understanding is the foundation of human-like problem-solving and automatic generation of geometric figures.Different from general natural language understanding,geometry problem understanding has higher requirements for the correctness,accuracy and completeness of information extraction.The variety of different geometry representations is a difficult problem in the automatic geometry problem understanding.The existing methods of geometry problem understanding mainly include template-based methods and machine-learning-based methods.The former has high accuracy of geometry information extraction,but low flexibility and low recall rate.The latter has good adaptability,but its accuracy is limited and easy to cause error propagation.Focusing on the understanding plane geometry problem text,This paper proposes an automatic plane geometry problem understanding method based on the three-level Representation Normalization.The main research contents include the following parts:(1)Analysis of the diversity of geometry problems and geometry information characteristics.Firstly,the diversity of geometry representation is analyzed in detail.Secondly,multidimensional analysis is carried out on the descriptive characteristics of geometry entities and geometry relations in the geometry problem text.Then,categorize and analyze the representations of geometry relations in geometry problems;Finally,according to the geometry features,the standard geometry problem representation rules are formulated.(2)This paper propose the three-level Representation Normalization.This method divides the text representation of geometry questions into three levels: character-and-words,single geometry relation and compound geometry relation.Firstly,the character-and-words level text is normalized by rules.Secondly,the template library of the single geometric relation is constructed,and the normal representation of the single geometric relation is matched by the template to realize the text normalization of the single geometric relation level.Finally,through information splitting,complementing and rectifying of the compound geometry relation text,it is decomposed into multiple single geometry relation text,so as to achieve the representation normalization of the compound geometry relation text.This method realizes the standard representation of geometry problems and solves the problem of representation diversity of geometry problems.(3)This paper propose a geometry entity and relation extraction method based on the three-level Representation Normalization and Bi-LSTM CRF.Firstly,the data set was constructed through the nomalized representation of geometry problems.Secondly,a geometry entity extraction model based on Bi-LSTM CRF was constructed.Then,the geometry entities extracted from the model are fused and duplicated by using the results of the three-level Representation Normalization,and the dictionary of geometry entities and their representations is constructed to realize geometry entity extraction.Then,the entity relation type is extracted by matching the standard representation of the problem clause,and the corresponding geometry entities are determined by querying the entity representation dictionary to realize the geometric entity relation extraction.Tested by 928 plane geometry problems,there are 837 problems are understanded completely and correctly,and the success rate of 90.194%.The experiment shows that this method can realize the geometry problem understanding task well. |
