Learning To Understand Entities In Text

Abstract

Real world entities such as people, organizations and countries play a critical role in text. Reading text offers rich information about these entities, both explicit, such as historical facts and scientific findings, and implicit, such as social relationships and personal opinions. Automatically extracting rich entity information promises exciting opportunities, such as question answering systems which can reason across facts mentioned in different documents, and analytic models which can help us understand constantly evolving social relations. This dissertation studies how machines can read natural language text, gather rich entity information, and map this information to a structured format. We present three independent studies, each focusing on different aspects of entity centric text understanding. First, we introduce a semantic parser populating entity attributes embedded in noun phrases to a large scale knowledgebase. Our method addresses the challenges arise from incompleteness in schema (i.e., existing ontologies cannot represent the meaning of many English phrases) and in KB (i.e., most KB misses many facts). Second, we study rich entity categories that can be inferred from the sentence that the entity occurs in. Our new formulation with virtually unrestricted types allows us to expand the standard KB-based training methodology with typing information from Wikipedia definitions and naturally-occurring head-word supervision. Lastly, we introduce a document-level model to infer dynamic entity-entity relationships. Unlike prior work which mostly focused on factual relationships, our work considers sentiment relationship between a pair of entities and presents a model which considers the document and social context jointly. For each of these studies, we address the limited annotated data challenge via crowdsourcing and/or harvesting large-scale naturally occurring weak supervision. Each study presents a new model and learning framework exploiting new sources of supervision to organize entity information. Together, this thesis expands the scope of information that can be learned about entities from text and points towards future work for entity centric document understanding.