• 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.3.2023.tde-10012024-092904
Document
Author
Full name
Flávio Nakasato Cação
E-mail
Institute/School/College
Knowledge Area
Date of Defense
Published
São Paulo, 2023
Supervisor
Committee
Costa, Anna Helena Reali (President)
Carvalho, Aline Marins Paes
Roman, Norton Trevisan
Title in Portuguese
Uma abordagem de aprendizado por reforço profundo para respostas a perguntas complexas de domínio aberto.
Keywords in Portuguese
Aprendizado computacional
Aprendizagem profunda
Processamento de linguagem natural
Abstract in Portuguese
Recentemente, modelos compostos por apenas módulos neurais de Recuperação de Informação e Compreensão de Leitura de Máquina/Gerador de Texto baseados em modelos de linguagem pré-treinados alcançaram o estado da arte em vários conjuntos de dados desafiadores de processamento de linguagem natural. No entanto, ainda há espaço significativo para melhorias na capacidade de raciocínio desses sistemas, especialmente no domínio de perguntas e respostas complexas de domínio aberto (CODQA - Complex Open-Domain Question Answering). Neste projeto, propomos uma arquitetura que combina as principais características desses modelos dentro de uma configuração de Aprendizado por Reforço, com a capacidade extra de realizar múltiplos saltos entre documentos para responder às perguntas dos usuários. Um sistema com esta capacidade é fundamental para construir agentes conversacionais capazes de responder a perguntas complexas que requerem múltiplas consultas em uma base de conhecimento não-estruturada. Nossos sistemas alcançaram um F1-score máximo de 0.13 ± 0.3 no conjunto de teste, usando em média apenas 47% das passagens de texto totais disponíveis.
Title in English
A deep reinforcement learning approach to complex open-domain question answering.
Keywords in English
Complex open-domain question answering
Conversational agents
Deep reinforcement learning
Abstract in English
Recently, models composed of only a neural Information Retrieval and a Machine Reading Comprehension/Text Generator modules based on pretrained language models have reached the state of the art in several challenging natural language processing datasets. However, there is still significant room for improvement in the reasoning capacity of these systems, especially in the realm of complex open-domain question answering (CODQA) datasets. In this project, we propose an architecture that combines the main features of these models within a Reinforcement Learning setting, with the extra ability to perform multiple hops among documents to answer to users questions. A system with this capability is critical for building conversational agents able to answer difficult questions that require multiple queries on a non-structured database. Our systems achieved a maximum F1-score of 0.13±0.3 on the test set, using on average only 47% of the total available text passages.
 
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
2024-01-11
 
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.