Multi-Stage Attentional U-Net (MSAU) architecture [9] for its proven effectiveness in entity segmentation from the char-grid embedding [12]. We go beyond its limitation of having no Administrative documents are one of the most widely used direct supervision from the linking between entities by adding medium for data storage and communication, which raises Part-Intensity Fields and Part-Association Fields (PIF-PAF) the need for an automated solution for form understanding. [13] which results in an additional two-head branch for entity However, this task remains largely difficult due to the layout linking and prediction from document images. In addition variations between forms for different purposes, organizations, to original segmentation output, the Part-Intensity Field head etc. Owning to its complexity, classical works like heuristicbased produces of confidence score of each entity's key points, [1], [2], or recent deep learning-based methods [3], while the Part-Association Fields head allows the associations [4] often break the problem into several sub-components: between entities. Apart from the major addition of PIF-PAF text-line detection, entity recognition, relation extraction, etc. module, our work also improved upon MSAU by adding The rationale of the division is that we may have better Corner Pooling [14] to solve the issue of long-range distance control of the input-output correspondence while cascading between entities, enabling the wider-range propagation of information for later steps.