• JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
  • JoomlaWorks Simple Image Rotator
 
  Bookmark and Share
 
 
Master's Dissertation
DOI
https://doi.org/10.11606/D.8.2020.tde-06042021-192617
Document
Author
Full name
Andressa Vieira e Silva
E-mail
Institute/School/College
Knowledge Area
Date of Defense
Published
São Paulo, 2020
Supervisor
Committee
Lopes, Marcos Fernando (President)
Barretto, Marcos Ribeiro Pereira
Felippo, Ariani Di
Ferreira, Marcelo Barra
Title in Portuguese
Um modelo de classificação para o Reconhecimento de Entidades Nomeadas
Keywords in Portuguese
Reconhecimento de Entidades Nomeadas
Redes neurais
Representações word embeddings
Traços de representação linguística
Abstract in Portuguese
O Reconhecimento de Entidades Nomeadas (REN) é uma tarefa de Proces- samento de Linguagem Natural (PLN) que busca identificar as Entidades Nomeadas de um texto, tais como nomes de pessoas, cidades e organiza- ções, classificando-as em um conjunto pré-definido de categorias. Essa é considerada uma tarefa difícil, pois as Entidades Nomeadas constituem uma classe gramatical com muita variação lexical e de baixa frequência quando comparadas à massa total de dados textuais. Recentemente, as pesquisas com redes neurais profundas têm mostrado excelentes resultados em diversas aplicações de PLN, incluindo o REN. Nesta pesquisa, foram investigadas duas arquiteturas de redes neurais para o REN no Harem, um corpus de língua portuguesa: BERT (devlin et al., 2018) e uma rede neural bidirecional LSTM (BiLSTM). O objetivo principal foi explorar traços baseados na distribuição contextual das entidades, através de representações vetoriais word embeddings associadas a traços linguísticos. Foram usados traços de etiquetagem morfossintática, forma ortográfica da palavra e recursos lexicais. Esses traços foram concatenados às representações word embeddings para alimentar a BiLSTM. Os resultados mostraram uma melhora estatisticamente significativa no desempenho desse modelo em comparação à BiLSTM apenas com os word embeddings. O modelo BERT, por sua vez, obteve medidas próximas ao estado da arte no Harem.
Title in English
A classification model for Named Entity Recognition
Keywords in English
Linguistic feature representation
Named Entity Recognition
Neural networks
Word embeddings
Abstract in English
Named Entity Recognition (NER) is a Natural Language Processing (NLP) task that aims at identifying Named Entities in a text, such as person, city, and organization names, classifying them into a pre-defined set of categories. NER is considered a hard task as Named Entities are a grammatical class with lots of lexical variation and relatevely low frequency if compared to the total mass of textual data. Nevertheless, deep neural network researches have recently shown excelent results in several NLP applications, including NER. In this work, two neural network archictetures were investigated for Harem, a corpus of Portuguese: BERT (devlin et al., 2018) and a bidirectional neural network LSTM (BiLSTM). The main goal was to explore features based on the entities contextual distribution by means of word embeddings vectors associated with linguistic features. We used as features part-of-speech tagging, spelling formats, and lexical resources. Those features were concatened with word embeddings vectors and fed into the BiLSTM. Our results showed a significant performance improvement with this model if compared to a BiLSTM using only word embeddings. On the other hand, BERT model obtained scores close to the Harem state-of-the-art.
 
WARNING - Viewing this document is conditioned on your acceptance of the following terms of use:
This document is only for private use for research and teaching activities. Reproduction for commercial use is forbidden. This rights cover the whole data about this document as well as its contents. Any uses or copies of this document in whole or in part must include the author's name.
Publishing Date
2021-04-06
 
WARNING: Learn what derived works are clicking here.
All rights of the thesis/dissertation are from the authors
CeTI-SC/STI
Digital Library of Theses and Dissertations of USP. Copyright © 2001-2024. All rights reserved.