Experiments in neural question answering

In this thesis, we apply deep learning methods to tackle the tasks of finding duplicate questions, learning to rank the answers of a Multiple Choice Question (MCQ) and classifying the answers to a question coming from the context of a paragraph. We draw our attention toward the problems related to sentence-sentence similarity. We used siamese architecture for better representation of the question and answers. The basic aim of all the methods proposed in this thesis is to build the word embeddings of question and answers and feed them to a deep neural architecture. We used several such architectures like Long-Short term memory (LSTM) and Convolutional Neural Network (CNN). We have also implemented an attention mechanism to put more focus on the sentence-word relationship. Our goal was to extract a refined representation of the question and answers through different combination of these deep learning techniques. We generated a representation of a sentence according to the context of another sentence for solving our tasks. We provide some simple but efficient deep learning models to solve our tasks. As neural models are data-driven, we train our model extensively by making pairs such as question-question and question-answer over a large-scale real-life dataset. We used three different datasets to solve our three different tasks. The Quora dataset of several question-answer pairs was used for the task of finding duplicate question. The OpenTriviaQA question answering dataset for the ranking of multiple answers. Lastly, we use the SQuAD dataset for the answer classification of reading comprehension task. We evaluate our models based on metrics like accuracy, precision, recall, F1 scores. Our methods and experiments demonstrate some significant improvements over the state-of-the-art methods.

Author

• Islam, Rafat-Al-

• convolutions
• convolutional neural network
• long-short term memory
• multiple choice questions
• questions and answers

Identifier

hdl:10133/4998

OAI identifier oai:identifier

oai:opus.uleth.ca:10133/4998