Entity Recognition in Arithmetic Word Problem Based on BERT and Boundary Detection
Rao Peng, Chensi Li, Zebin Wu, Xinguo Yu · 2023
The arithmetic word problem solving is a hot issue in the field of educational technology. The quantitative relation is the key step to solve the problem, and the entity is the base unit in relation. This paper proposes a method of recognizing entities in arithmetic word problem based on BERT model and boundary detection, and combines syntax-semantic (S2) model to extract relations. The first step is to input the problem text into the BERT model to obtain the vector sequence. Then the vector sequence is fed into two different neural networks to identify the boundary of all entities. Thirdly, for each boundary token, the nearest one is combined to obtain the target entity. The identified entities will serve as anchor points to assist the S2 model in extracting quantitative relations. In the experiment, 1500 mixed examples collected previously are used for test. The results show our entity recognition method reaches 77.5% f1 value on simple problem and 74.5% f1 value on complex problem.