🤖 AI Summary
This work addresses the limitation of autoregressive table recognition methods, whose sequential generation process induces cell representations that are dependent on generation order, thereby compromising global structural consistency. To overcome this, the authors propose a structure refinement module incorporating non-causal attention within a unified multi-task framework, enabling the learning of order-invariant, cell-level feature representations while simultaneously performing table structure prediction, cell localization, and content recognition. By eliminating reliance on autoregressive ordering, the approach substantially enhances both global coherence and parallel inference efficiency. Experimental results on two large-scale datasets demonstrate significant improvements in cell localization and end-to-end recognition accuracy, along with an approximately threefold reduction in overall inference time.
📝 Abstract
Multi-task table recognition jointly addresses table structure prediction, cell localization, and cell content recognition within a unified framework. Existing approaches often rely on autoregressive decoders to generate table structures and reuse their hidden states for cell localization and content recognition. This autoregressive generation process can make cell representations order-dependent, degrading global consistency across cells. This paper proposes a structural refinement module that produces order-independent cell features through non-causal attention. This design enables parallel inference of cell contents while conditioning each cell on global context encoded in the refined features. Experiments on two large datasets demonstrate consistent gains in cell localization and end-to-end recognition, while reducing overall inference time by around threefold.