Research on the CRF-based Sequence Labeling Algorithm Used in Reference Resolution of Mathematical Word Problem Understanding

Abstract

Resolving the reference phenomenon in application problems is a key step to realize the understanding of mathematical problems. Compared with the Chinese corpus in the general field, this study analyzed the characteristics of reference resolution in mathematical stratified sampling word problems, and explored key factors by combining the C4.5 decision tree algorithm. On this basis, the study proposed a CRF-based sequence labeling algorithm, which was exploited to identify the to-be-solved items of the stratified sampling problem and resolved them. The experimental data is the stratified sampling word problems collected from the math problems in the Chinese college entrance examination and the test problems in the textbook from 2012 to 2020. The results show that the F-values for the antecedents and anaphors of the mathematical stratified sampling word problems based on CRF sequence labeling can reach 81.29% and 90.69%, respectively, and the F-values for reference resolution can reach 84.84%, which is higher than the traditional Bayesian method.