QA System By Deep Learning

QA System is now the most heated topic in the field of natural language processing.It not only enables users to directly ask questions by natural language instead of groups of keywords,but also returns precise and concise answers rather than bunch of related web pages.In recent years there have been many breakthrough in applying deep learning in NLP.Many outstanding papers and frameworks make deep learning-based QA system the most exciting field of research in NLP,such as SyntaxNet,which is very popular in massive QA system development.In this paper,we apply mainstream deep learning algorithms to QA system,and our contribution is four-fold:1.Proposing a high-performance task-oriented QA model based on Word2Vec and CNN.We improved and extended CNN-based sentence classification methods into QA Question classification and retrieval.2.Presenting a bi-directional LSTM sequence to sequence open-domain QA model with attention mechanism.This framework out-performs related uni-directional RNN Seq2Seq model in terms of context understanding and grammar accuracy.3.Performing a series of experiments on Facebook bAbI datasets,Ubuntu Dialog datasets and other dialog datasets to compare the performance measures of proposed task-oriented QA model and open-domain QA model with respect to previous work.From the results we can see that proposed models in this thesis out-performs related models in all performance measures.4.Presenting the first maintenance-free,DevOps-friendly solution for TensorFlow NLP service by Docker without using TensorFlow Serving.Major contributions in this paper:1.Finding and proving the relation between CNN-based QA algorithm and dimension of Word2Vec input.2.Providing a copy-and-interpolation solution for solving degraded performance of CNN-based QA algorithm when increasing number of classes in question set.3.Using BLSTM and attention mechanism to optimize RNN-based Seq2Seq QA system.4.The first paper to create QA micro-service with TensorFlow and Docker instead of TensorFlow Serving.