Abstract
The long-standing problem of geometric problem solving in artificial intelligence education has attracted widespread attention. It is necessary to combine geometry diagrams and text descriptions to form a logical representation. This involves combining the knowledge of mathematical theorems, generating a solution sequence, and executing to obtain the answer. However, deficiencies in the feature extraction of geometry diagrams and the fusion of diagram text information can lead to poor performance in solving geometry problems. To effectively extract geometry diagram features, this study proposes an improved diagram parser DenseNet, and enhances the semantic representation of cross-modal information by adding auxiliary tasks. A structural and semantic pre-training strategy was used to parse the text description to avoid different problem solving schemes due to subtle differences in the interpretation of text content. Information fusion was realized by connecting the two modal labels, and then the information was sent to the encoder for fusion. The geometric knowledge was generated under the guidance of multi-modal information, and these programs were executed to obtain the results. Additionally, the performance of the proposed geometric neural solution method on the PGPS9K dataset is improved by 1.3% on average. Compared with the Geometry3K dataset, the effectiveness was proven.