Abstract: A method comprises identifying entities in a form document; for each of the entities, generating a vector for the entity that comprises a plurality of values describing content of the entity and a position of the entity within the form document; identifying, by the computing system, pairs of the entities that satisfy a spatial relationship requirement; for each pair of entities satisfying the spatial relationship requirement, applying a classification machine-learned (ML) model that takes as input the vectors for the pair of entities and outputs a category for a first entity of the pair of entities and a category for a second entity of the pair of entities, wherein a combination of the category for the first entity and the category for the second entity represents a semantic relationship between the first entity and the second entity; and processing the form document based on the semantic relationships between the entities.