Fine-grained entity typing system - design and analysis

Named entity recognition (NER) is a natural language processing (NLP) task that involves identifying mentions (spans of text) denoting entities in a given text document and assigning them a semantic category/type from a given taxonomy. It is considered to be one of the fundamental tasks in NLP and forms the basis for higher level understanding. In this thesis, we deal with fine-grained entity type recognition, which is a variant of the classic NER task where the usual types are sub-divided into fine-grained types. We show that the current approaches, which address this problem using only local context, are insufficient to completely address the problem. We systematically identify the fundamental challenges and misconceptions that underlie the assumptions, approaches and evaluation methodologies of this task and propose improvements and alternatives. We do this by first analyzing the role of context and background knowledge in the task of fine-grained entity typing. Second, we introduce a modular architecture for fine-grained typing of entities and show that a rather simple instantiation of these modules reaches the state-of-the-art performance