Loading...
Publication
Modelling semantic relations with distributitional semantics and deep learning : question answering, entailment recognition and paraphrase detection
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 426.17 KB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Nesta dissertação apresenta-se uma abordagem à tarefa de modelar relações semânticas
entre dois textos com base em modelos de semântica distribucional e em aprendizagem
profunda. O presente trabalho tira partido de várias disciplinas da ciência
cognitiva, com especial relevo para a computação, a linguística e a inteligência artificial,
e com fortes influência da neurociência e da psicologia cognitiva.
Os modelos de semântica distribucional (também conhecidos como ”word embeddings”)
são usados para representar o significado das palavras. As representações
semânticas das palavras podem ainda ser combinadas para obter o significado de
um excerto de um texto recorrendo ao uso da aprendizagem profunda, isto é, com o
apoio das redes neurais de convolução.
Esta abordagen é utilizada para replicar a experiência realizada por Bogdanova
et al. (2015) na tarefa de deteção de perguntas que podem ser respondidas as mesmas
respostas tal como estas foram respondidas em fóruns on-line. Os resultados do
desempenho obtidos pelas experiências apresentadas nesta dissertação são equivalentes
ou melhores que os resultados obtidos no trabalho de referência mencionado
acima.
Apresentao também um estudo sobre o impacto do pré-processamento apropriado
do texto, tendo em conta os resultados que podem ser obtidos pelas abordagens
adotadas no trabalho de referência supramencionado. Este estudo é levado a cabo
removendo-se certas pistas que podem levar o sistema, indevidamente, a detetar
perguntas equivalentes. Essa remoção das pistas leva a uma diminuição significativa
no desempenho do sistema desenvolvido no trabalho de referência.
Nesta dissertação é ainda apresentado um estudo sobre o impacto que os word
embeddings treinados previamente têm na tarefa de detetar perguntas semanticamente
equivalentes. Substituindo-se, aleatoriamente, word embeddings previamente
treinados por outros melhora-se o desempenho do sistema.
Além disso, o modelo foi utilizado na tarefa de reconhecimento de implicações
para Português, onde mostrou uma taxa de acerto similar à da baseline. Este trabalho também reporta os resultados da aplicação da abordagem adotada
numa competição para a deteção de paráfrases em Russo. A configuração final apresenta
duas melhorias: usa character embeddings em vez de word embeddings e usa
vários filtros de convolução. Esta configuração foi testado na execução padrão da
Tarefa 2 da competição relevante, e mostrou resultados competitivos.
This dissertation presents an approach to the task of modelling semantic relations between two texts, which is based on distributional semantic models and deep learning. The present work takes advantage of various disciplines of cognitive science, mainly computation, linguistics and artificial intelligence, with strong influences from neuroscience and cognitive psychology. Distributional semantic models (also known as word embeddings) are used to represent the meaning of words. Word semantic representations can be further combined towards obtaining the meaning of a larger chunk of a text using a deep learning approach, namely with the support of convolutional neural networks. These approaches are used to replicate the experiment carried out, by Bogdanova et al. (2015), for the task of detecting questions that can be answered by exactly the same answer in online user forums. Performance results obtained by my experiments are comparable or better than the ones reported in that referenced work. I present also a study on the impact of appropriate text preprocessing with respect to the results that can be obtained by the approaches adopted in that referenced work. Removing certain clues that can unduly help the system to detect equivalent questions leads to a significant decrease in system’s performance supported by that referenced work. I also present a study of the impact that pre-trained word embeddings have in the task of detecting the semantically equivalent questions. Replacing pre-trained word embeddings by randomly initialised ones improves the performance of the system. Additionally, the model was applied to the task of entailment recognition for Portuguese and showed an accuracy on a level with the baseline. This dissertation also reports on the results of an experimental study on the application of the adopted approach to the shared task of sentence paraphrase detection in Russian. The final set up contained two improvements: it uses several convolutional filters and it uses character embeddings instead of word embeddings. It was tested in Task 2 standard run of the relevant shared task and it showed competitive results.
This dissertation presents an approach to the task of modelling semantic relations between two texts, which is based on distributional semantic models and deep learning. The present work takes advantage of various disciplines of cognitive science, mainly computation, linguistics and artificial intelligence, with strong influences from neuroscience and cognitive psychology. Distributional semantic models (also known as word embeddings) are used to represent the meaning of words. Word semantic representations can be further combined towards obtaining the meaning of a larger chunk of a text using a deep learning approach, namely with the support of convolutional neural networks. These approaches are used to replicate the experiment carried out, by Bogdanova et al. (2015), for the task of detecting questions that can be answered by exactly the same answer in online user forums. Performance results obtained by my experiments are comparable or better than the ones reported in that referenced work. I present also a study on the impact of appropriate text preprocessing with respect to the results that can be obtained by the approaches adopted in that referenced work. Removing certain clues that can unduly help the system to detect equivalent questions leads to a significant decrease in system’s performance supported by that referenced work. I also present a study of the impact that pre-trained word embeddings have in the task of detecting the semantically equivalent questions. Replacing pre-trained word embeddings by randomly initialised ones improves the performance of the system. Additionally, the model was applied to the task of entailment recognition for Portuguese and showed an accuracy on a level with the baseline. This dissertation also reports on the results of an experimental study on the application of the adopted approach to the shared task of sentence paraphrase detection in Russian. The final set up contained two improvements: it uses several convolutional filters and it uses character embeddings instead of word embeddings. It was tested in Task 2 standard run of the relevant shared task and it showed competitive results.
Description
Keywords
Semântica Processamento da linguagem natural Linguística cognitiva Teses de mestrado - 2017