Triple-attention computation model for question answering

• Main Author: Yu, Sicheng
• Other Authors: Mao Kezhi
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/75953
• Institution: Nanyang Technological University

In order to assess the degree of intelligence the machine, the machine's understanding of the language is an indispensable and important aspect. The question answering system is an important task for the machine to understand human language. Thesis proposes a question and answer system model based on three kinds of attention calculations. Comparing with other existing models, the calculation of the three attention fully extracted the information between the context and the question in different aspects, making the neural network better learn the context-based representation of the question. The model consists of three layers, embedding layer, attention layer, and predict layer. The role of embedding layer is to vectorize the words in the context and question. The attention layer first calculates the mutual attention between the context and the question, and then calculates the Self-Attention. Finally, the predictive layer is used to predict the start and end of the answer. Through experiments on the SQuAD dataset, the performance of the model using the different RNN architectures is better than that of the main reference model in both EM and F1 values. In addition, the performance of this model has performed well in many question answering models proposed in recent years, surpassing many classical models, and has strong competitiveness. Keywords: Natural language processing, Recurrent neural network, Question answering, Attention mechanism