基于深度学习的问答系统技术研究

发布时间：2018-05-31 16:55

本文选题：问答系统 + 词向量　；参考：《浙江大学》2017年硕士论文

【摘要】：问答系统是目前自然语言处理领域中的研究热点,它既能让用户通过自然语言直接发问,又能直接向用户返回精确、简洁的答案,而不是一系列相关网页。近年来,深度学习技术为问答系统领域带来诸多突破,基于深度学习技术的问答算法研究成为了自然语言处理最热门的研究方向,诞生了大量优秀的文章与开发框架,如Google在2016年推出的SyntaxNet,大大降低了高性能问答系统的开发成本。本文应用了深度学习技术进行问答系统的构造,开展的工作如下:1.利用词向量与卷积神经网络搭建了一套高准确率的面向具体任务问答系统,改进了已有的卷积神经网络问句分类算法,探索了模型初始化参数与模型性能的关系。2.基于双向长短时记忆模型与注意力机制搭建了一套端到端开放领域问答系统,改进了前人基于单向长短时记忆模型的端到端问答算法在问句语义表征上的缺点。3.在Facebook bAbI、Ubuntu Dialogue Corpus等常用数据集上进行了实验对比,通过实验结果对比突出了本文设计的问答算法的有效性与合理性,并对实验结果做了较为详细的分析。4.利用TensorFlow、Docker构建了一套维护成本低、部署方便的问答系统微服务,解决了 TensorFlow框架线上服务部署困难的问题。本文的主要贡献如下:1.创造性地发现了基于词向量和卷积神经网络的问句语义相似度算法性能与词向量维数之间的关系,并通过实验加以验证。2.尝试了通过复制插值的方式扩展基于词向量和卷积神经网络的问句语义相似度算法中词向量输入部分的维数,解决了问句类别数上升时模型性能下降的问题。3.使用双向长短时记忆模型与注意力机制改进了现有基于循环神经网络的端到端问答算法模型,提高了平均问答长度等性能指标。4.基于TensorFlow与Docker实现了一整套问答系统微服务,创新性地使用Spring Boot包装算法脚本,解决了TensorFlow Serving的兼容性问题,实现了弹性部署与扩容,维护成本低。
[Abstract]:Question and answer system is a hot research topic in the field of natural language processing. It can not only let users directly ask questions through natural language, but also return accurate and concise answers to users directly, rather than a series of related web pages. In recent years, deep learning technology has brought many breakthroughs to the field of question and answer system. The research of question and answer algorithm based on deep learning technology has become the most popular research direction of natural language processing, and a large number of excellent articles and development frameworks have been born. SyntaxNet, for example, launched by Google in 2016, has greatly reduced the cost of developing a high-performance question-and-answer system. In this paper, the deep learning technology is used to construct the Q & A system, and the work is as follows: 1. Using word vector and convolutional neural network, a set of quizzes oriented question answering system with high accuracy is set up. The existing convolutional neural network question classification algorithm is improved, and the relationship between model initialization parameters and model performance is explored. An end-to-end open domain question-and-answer system based on bidirectional long and short term memory model and attention mechanism is constructed, which improves the shortcomings of the previous end-to-end question answering algorithm based on one-way long and short term memory model on the semantic representation of question sentences. The experimental results are compared on Facebook bAbIbuntu Dialogue Corpus and other common data sets. The validity and rationality of the question and answer algorithm designed in this paper are highlighted by the comparison of experimental results, and the experimental results are analyzed in detail. By using Tensor flow Docker, a question-and-answer system micro-service with low maintenance cost and convenient deployment is constructed, which solves the problem of difficult service deployment on the TensorFlow framework. The main contributions of this paper are as follows: 1. The relationship between the performance of semantic similarity algorithm based on word vector and convolution neural network and the dimension of word vector is found out creatively. This paper attempts to extend the dimension of word vector input in the semantic similarity algorithm of question sentence based on word vector and convolutional neural network by replicating and interpolating, and solves the problem of deterioration of model performance when the number of question categories increases. The existing end-to-end question-and-answer algorithm model based on cyclic neural network is improved by using bidirectional long and short time memory model and attention mechanism, and the performance index of average question length is improved. Based on TensorFlow and Docker, a set of question answering system micro-service is implemented, and Spring Boot packaging algorithm script is innovatively used, which solves the compatibility problem of TensorFlow Serving, realizes flexible deployment and expansion, and has low maintenance cost.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】