.png){kind=icon}
.png){kind=icon}
Return
Number-enhanced relational graph encoding with hierarchical tree decoding for numerical reasoning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114755.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the advent of the information age, massive amounts of data are stored in structured table formats. In real-world applications, the hybrid form that integrates tables and text has gained increasing attention, with numerical reasoning problems emerging as one of the most challenging research focuses due to their complexity. Numerical reasoning tasks require models to accurately infer numerical answers or explanations by understanding numerical relationships and performing computational reasoning based on given context information and questions. To address the limitations of existing methods-namely, the encoder’s difficulty in effectively capturing intricate relationships among numbers, table structures, and textual content, and the decoder’s inflexibility in handling diverse reasoning patterns-we propose a Number-Enhanced Relational Graph Encoding with Hierarchical Tree Decoding (NRGHT) for numerical reasoning questions. Our method constructs three types of relational graphs from tabular view, numerical view and relation view and adopts a novel relation-enhanced graph encoder module, enhancing the encoder’s capacity to model cross-modal and numerical dependencies. In the decoding phase, a hierarchical recursive tree-structured decoder is employed to generate an expression tree, which is then executed to produce the final answers. Experimental results on the TAT-QA benchmark demonstrate that the proposed NRGHT framework achieves competitive performance and establishes a strong baseline for hybrid numerical reasoning tasks.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W