A Hybrid Probabilistic Approach for Table Understanding

Tables of data are used to record vast amounts socioeconomic, scientific, and governmental information. Although humans create tables using underlying organizational principles, unfortunately AI systems struggle understand the contents these tables. This paper introduces an end-to-end system for table understanding, process capturing relational structure in We introduce models that identify cell types, group cells into blocks serve a similar functional role, predict relationships between blocks. hybrid, neuro-symbolic approach, combining embedded representations learned from thousands with probabilistic constraints capture regularities how organize Our model is better able positional invariants headers enforce homogeneity types. One limitation this research area lack rich datasets evaluating so we new benchmark dataset comprised 431 diverse data.gov. The evaluation results show our achieves state-of-the-art performance on type classification, block identification, relationship prediction, improving over prior efforts by up 7% macro F1 score.