Scholarlink

As a critical component of intelligent table applications, table semantic parsing requires simultaneous comprehension of structural and semantic dimensions – a challenging but essential research focus in table retrieval. Current approaches face dual limitations: intricate cell layouts hinder effective semantic extraction, while existing models suffer from excessive parameters and computational overhead. To address these challenges, we propose TaQ-Former, a structure-aware Querying Transformer that achieves cross-modal alignment through two key designs: (1) structural features extraction via cross-attention with frozen table encoder while ensuring computational stability, and (2) iterative semantic refinement through alternating processing of table topology and textual markups. Combined with our graph-enhanced pre-training framework for text-attributed graph representations. Our method enables robust semantic transfer under fixed computational constraints. Experimental results demonstrate that our method outperforms existing approaches on two benchmark datasets. Additional experiments validate the superior performance of our framework in cross-dataset generalization, complex table processing, and table retrieval under diverse query intents.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Table Semantic Retrieval Method Using Self-supervised Pre-training Querying Transformer and Text-Attributed Graphs

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yinxuan Shen,
  • Linchao Ye,
  • Zhenping Xie

摘要

As a critical component of intelligent table applications, table semantic parsing requires simultaneous comprehension of structural and semantic dimensions – a challenging but essential research focus in table retrieval. Current approaches face dual limitations: intricate cell layouts hinder effective semantic extraction, while existing models suffer from excessive parameters and computational overhead. To address these challenges, we propose TaQ-Former, a structure-aware Querying Transformer that achieves cross-modal alignment through two key designs: (1) structural features extraction via cross-attention with frozen table encoder while ensuring computational stability, and (2) iterative semantic refinement through alternating processing of table topology and textual markups. Combined with our graph-enhanced pre-training framework for text-attributed graph representations. Our method enables robust semantic transfer under fixed computational constraints. Experimental results demonstrate that our method outperforms existing approaches on two benchmark datasets. Additional experiments validate the superior performance of our framework in cross-dataset generalization, complex table processing, and table retrieval under diverse query intents.