A Scene-Attention Relation-Centric Algorithm for Solving Arithmetic Word Problems

Rao Peng, Xinguo Yu, Chuanzhi Yang, Xiaopan Lyu

Abstract: This paper presents the Scene-Attention Relation-Centric Algorithm (SARC), a novel approach to solving Arithmetic Word Problems (AWPs). Building on a decade of advancements in relation-centric methodologies, this study integrates scene knowledge to enhance both relation acquisition and inference, resulting in a high-performance algorithm that excels in accuracy and generating didactic solutions. For relation acquisition, the paper introduces the Scene-Syntax-Semantic (S3<math><msup is="true"><mrow is="true"><mi is="true">S</mi></mrow><mrow is="true"><mn is="true">3</mn></mrow></msup></math>) method, which extracts explicit and implicit relations from scene-labeled text. This is achieved using two dedicated S3<math><msup is="true"><mrow is="true"><mi is="true">S</mi></mrow><mrow is="true"><mn is="true">3</mn></mrow></msup></math> model pools: one for acquiring intra-scene relations and another for bridging relations between scenes. In relation inference, a Scene-Guided Symbolic Solver is developed to categorize the acquired relations by their associated scenes or scene pairs, enabling the inference process to operate in two modes: intra-scene and inter-scene. This scene-guided strategy enhances both the accuracy and efficiency of relation inference by increasing the certainty of inference operations. Experimental evaluations on five authoritative datasets demonstrate that the proposed algorithm surpasses all baseline models. It achieves an overall accuracy improvement of 1.7% and a 4.1% increase on problems involving multiple scenes. Additionally, qualitative analyses of illustrative cases reveal that the algorithm generates more didactic solutions compared to baseline methods. These findings underscore the effectiveness of incorporating scene knowledge into relation-centric approaches, significantly advancing both problem-solving accuracy and explanatory quality.