Noun-based attention mechanism for Fine-grained Named Entity Recognition

Fine-grained Named Entity Recognition is a challenging Natural Language Processing problem as it requires on classifying entity mentions into hundreds of types that can span across several domains and be organized in several hierarchy levels. This task can be divided into two subtasks: Fine-grained Named Entity Detection and Fine-grained Named Entity Typing. In this work, we propose solutions for both of these subtasks. For the former, we propose a system that uses a stack of Byte-Pair Encoded vectors in combination with Flair embeddings, followed by a BILSTM-CRF network, which allowed us to improve the current state of the art for the 1k-WFB-g dataset. In the second subtask, attention mechanisms have become a common component in most of the current architectures, where the patterns captured by these mechanisms are generic, so in theory, they could attend to any word in the text indistinctly, regardless of its syntactic type, often causing inexplicable errors. To overcome this limitation we propose an attention mechanism based specifically on the use of elements of the noun syntactic type. We have compared our results to those obtained with a generic attention mechanism, where our method presented better results.