Reconhecimento de entidades nomeadas para o português usando redes neurais

Abstract

Modern approaches to Named Entity Recognition (NER) use Neural Networks to automatically extract text features and incorporate them into the classification process. Word Embeddings, a type of Language Model (LM), are a key ingredient for improving the perfor- mance of NER systems. More recently, Contextualized LM, which adapt according to the context in which the word appears, have also proved indispensable. This master’s thesis shows how different combinations of Word Embeddings and Contextualized LM impact the NER task in Portuguese. The impact of textual diversity and size of the training corpus used in the construction of LMs were explored by the results of this task. Also, a compar- ative study of 16 combinations of different LMs, contextualized and Word Embeddings, is presented. Evaluations were performed in the Mini-HAREM corpus, widely adopted in the Portuguese NER task. The best result achieved in this research surpasses the state-of- the-art approach by 5.99% in a five-category scenario and 4.31% when considering the ten HAREM categories. In addition to the HAREM assessments, specific domains of this task were also studied. The results in these cases were evaluated in Clinical, Police and Geolog- ical context corpora. Superior or competitive results were obtained for all corpora in relation to other approaches.