Joint Learning of Text-based Categories

The fundamental problem addressed by this SBIR proposal is the capability to uncover new entity types, relations and topics that are implicit in text documents.Joint inference offers an opportunity to make discoveries that are more accurate and more nuanced than traditional pipelined approaches, but the hypothesis space is more complex and extraordinarily large.In this project we propose to develop a system, called JTL (Joint Text Learner), that leverages and extends previous work on probabilistic soft logic (PSL).PSL enables a modeler to specify statistical relational models using a simple logical language, which encodes both structure and latent variables that can represent topics and/or context.Our approach combines two recent successful developments in PSL: a general and extensible framework for topic modeling referred to as latent topic networks and 2) new techniques for knowledge graph identification which are able to jointly reason over extractions about entities and relations, and incorporate both statistical and ontological domain knowledge to infer a knowledge graph from text.We combine these ideas by developing a factored model for topics, entities and relations expressed using latent topic networks, which can be used for joint inference.