Semi-Structured Document Annotation Using Entity and Relation Types

Abstract

Semi-structured documents such as web-pages and reports contain text units with complex structure connecting these. We motivate and address the problem of annotating such semi-structured documents using a knowledge graph schema with entity and relation types. This poses significant challenges not addressed by the existing literature. The latent document structure needs to be recovered, and paths in the latent structure need to be jointly annotated with entities and relationships. We present a two stage solution. First, the most likely document structure is recovered by structure search using a probabilistic graphical model. Next, nodes and edges in the recovered document structure are jointly annotated using a probabilistic logic program, considering logical constraints as well as uncertainty. We additionally discover new entity and relation types beyond those in the specified schema. We perform experiments on real webpage and complex table data to show that our model outperforms existing table and webpage annotation models for entity and relation annotation.