Avaliação da influência do uso de embeddings no reconhecimento de entidades nomeadas para o português

Abstract

Named Entity Recognition (NER) is a crucial task in today's context, as there is an increasing demand for automated tools capable of extracting information from texts. Although NER is a widely studied area, it remains complex and requires a high cost in terms of time and computational resources, especially in languages with few resources, such as Portuguese, which lacks annotated data and high- performance trained models. In order to contribute to the development of this task in Portuguese language, this work applied the Deep Learning (DL) architecture, BiLSTM-CRF, using the FlairNLP Framework, testing different types of Word Embeddings, Static Embeddings, Contextualize Embeddings, and their concatenations, to analyze their influence on the results obtained. The trained models showed a significant gain in the use of Contextualize Embeddings and Embeddings concatenation, achieving a new state of the art with an F1 score of 98.88% for the general domain in the MultiWikiNER multilingual corpus, in the legal domain 89.31% for the LeNER-Br Corpus and 86.86% for the Pl-Corpus, which is part of the UlyssesNER-Br.