Enhancing Math Word Problem Solving Through Salient Clue Prioritization: A Joint Token-Phrase-Level Feature Integration Approach

Recently, deep learning-based methods, such as sequence-to-sequence and sequence-to-tree models, have shown remarkable success in Math Word Problem (MWP) solving. However, a predominant drawback of them is their insufficient attention to the inherent characteristics of math problems, including the varying significance of words within the problem. To address this issue, our paper introduces a novel approach termed “Salient Clue Prioritization (SCP)” for MWP. This approach revolves around understanding MWP from a fresh perspective by emphasizing salient clues. Our proposed approach involves two key training procedures: integrating a coarse-fine representation for multi-level comprehension during the first training phase, and dynamically adjusting attention weights based on identified keywords in the second training phase. Through extensive experimentation, our approach overcomes the limitations of existing methods by effectively identifying and adaptively utilizing salient clues to solve MWP. Our results indicate superior performance compared to other existing models, demonstrating the enhanced understanding of both the MWP itself and the relationship between words and final mathematical solutions.