Document Intelligence (DI) refers to the ability to read, understand and interpret documents. It has grown in the last years as a core component in fintech, legaltech, insurtech, etc. for managing automatically the information of document workflows, integrating both reading an understanding. It is challenging due to big inter-class variety and intra-class similarity, huge number of classes/layouts, heterogeneity (printed, handwritten, graphics), and ambiguity (domain-dependent meaning, language).
As communication objects, documents can be parsed according to compositional rules. Language model (LM) architectures that NLP applies to text sentences, become a valid paradigm to be adopted by DI. LMs in DI can be backed by Graphs, that robustly represent objects and relations. Relational reasoning for document parsing involves manipulating structured representations of semantically meaningful document objects (titles, tables, figures...) and relations, using compositional rules. Modern AI approaches use a relational inductive bias, imposing constraints among entities in learning time. It is a fundamental task in semantic segmentation, document classification, document object recognition, document visual Q&A...
In this talk we review some proposed Graph-based models to learn representations for the language of documents. The document language model refers to the compositional description in terms of building objects, relations and their properties. We illustrate the proposed contributions in solving tasks like semantic segmentation, classification, document object recognition and link discovery.
