Semantic interaction-enhanced encoding network for math word problem solving

Abstract

Solving math word problems (MWPs) requires machines to understand not only the literal meaning of text but also the abstract logic and mathematical reasoning embedded within it. However, existing models often lack explicit reasoning capabilities for semantic information, particularly when dealing with complex math word problem texts. Additionally, these models tend to embed all kinds of information without fine-grained selection, which may introduce unexpected noise for mathematical expression generation. To address these challenges, we propose a Semantic Interaction-Enhanced Encoding Network (SIEN) for math expression generation is proposed in this paper. Firstly, SIEN constructs a semantic role interaction graph for each problem and employs a graph attention neural network to learn interaction and semantic information, offering a more structured and enriched view of the math word problem text. Secondly, SIEN introduces a multi-channel adapter module that simultaneously learns comprehensive contextual information from numeric information channel, hierarchical semantic information channel, and interaction information channel. Furthermore, SIEN introduces a dynamic weighting mechanism that adjusts the information weight from each channel to prioritize relevant information and reduce noise. Experimental results on three public benchmark datasets demonstrate that SIEN achieves significant performance improvement over other state-of-the-art baseline models.