An automatic problem-solving method for word problems based on graph neural network

Abstract

The invention discloses a method for automatically solving problems based on a graph neural network. The invention firstly uses a cyclic neural network to encode the input text of the problem, and simultaneously constructs a numerical unit graph and a numerical comparison graph, and the output of the cyclic neural network (word level representation) are used as node features. The node features are fed into the graph neural network-based encoder together with the two constructed graphs to learn the graph representation features of the topic, enabling the final graph features to contain numerical textual relations and size information. And a pooling term is used to aggregate different sets of graph features into one, resulting in the output of the graph transformer. Finally, the output graph features are used as the input of a tree-structure-based decoder to generate the final solved expression tree. The invention improves task performance by enriching numerical representations in problems, and can achieve better problem-solving effects.

Description

technical field [0001] The invention relates to the field of computational linguistics, in particular to an automatic problem-solving method for applied problems based on a graph neural network. Background technique [0002] Solving mathematical problems, that is, automatically answering mathematical problems based on text descriptions, has attracted the attention of researchers since the 1960s and is an important natural language understanding task. A typical math word problem is to give a description of the problem and give a short statement about the problem with the unknown quantity. Earlier studies attempted to design automatic solvers through statistical machine learning methods and semantic parsing methods. However, these methods have poor generalization because it takes a lot of effort to design appropriate functions and expression templates. [0003] In recent years, automatic solvers based on deep learning have begun to appear. These deep learning methods can au...

AI Technical Summary

Problems solved by technology

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an automatic problem-solving method based on a graph neural network to solve the problem of low correctness of the problem-solving expressions generated by the existing deep learning model

Images