面向自动问答的中学历史知识库构建

发布时间：2018-04-05 20:06

本文选题：知识库　切入点：word2vec　出处：《哈尔滨工业大学》2017年硕士论文

【摘要】：随着人工智能在围棋方面的突破,“阿尔法狗”进入了大众的视野并且热度居高不下,与此同时人工智能以及深度学习越来越被大众熟知,家喻户晓。那么人工智能的一大目标就是让机器理解人的自然语言,也是目前的难点重点。所以对于问答系统的研究是必要的,现在有很多种问答系统,比如百度知道或搜狗问问的社区问答、一站到底等知识问答及小度聊天机器人等,这些系统尤其是知识问答系统往往都有知识库。因而对知识库的构建进行研究是必要的。对于高中历史学科问答系统来说研究的意义在于探究人工智能的极限,之所以这么说的原因是,让机器理解人的自然语言是目前的一大难点。到目前为止不论是传统的机器学习方法还是目前大热的深度学习,都无法真正的了解人类的自然语言,而高考历史的问答恰恰是在检测人工智能在理解人类的自然语言方面的进展。本文研究的主要研究内容如下:一.对中学历史题进行分析,针对中学历史题的特点进行知识库结构的设计。二.构建历史分词词表,因为对于历史材料难以精准分词,经常容易把相关事件、专业名词等在分词时分隔开,因而需要构建一个分词词表。在构建词表时运用了新词发现技术。三.高中历史知识库的扩充。为了能够全面准确地回答高考历史题,仅仅采用历史课本内容构建知识库是不够的,因此必须对其进行扩充。待扩充的资料为爬取的百度百科历史相关条目。使用了word2vec、doc2vec模型训练语义向量,与LR、SVM、Random Forest、GBDT等四种统计分类模型结合。最终利用最优的分类器选出在诸多条目中适合扩充进中学历史知识库的条目。四.高中历史知识库内容分类标注。通过对高考历史题的分析,每个历史实体的属性大致可以分为七类如影响类、背景类等。所以在历史实体扩充进知识库的时候,需要将其内容进行分类。使用训练好的词向量作为分类模型的输入,分类模型选择朴素贝叶斯、LSTM、BLSTM、C-LSTM等四种模型进行对比实验,最终选出最优的分类器进行高中历史知识库内容分类标注。
[Abstract]:With the breakthrough of artificial intelligence in go, "AlphaGo" has entered the public's field of vision and high fever, at the same time, artificial intelligence and in-depth learning are becoming more and more familiar to the public, household name.So one of the major goals of artificial intelligence is to make machines understand human natural language, which is also a difficult point at present.Therefore, it is necessary to study the question and answer system. Now there are many kinds of question answering systems, such as the community question and answer system that Baidu knows or Sogou asks, the one stop waiting for knowledge question answering and small chat robot, and so on.These systems, especially the Q & A system, often have a knowledge base.Therefore, it is necessary to study the construction of knowledge base.The significance of the research is to explore the limits of artificial intelligence for senior high school history subject question and answer system. The reason for this is that it is difficult for machines to understand human natural language.So far, neither the traditional machine learning methods nor the current hot deep learning have been able to truly understand the natural language of human beings.And the history of the gaokao quiz is precisely to test artificial intelligence in understanding the natural language of human progress.The main contents of this paper are as follows: 1.This paper analyzes the history problems in middle schools and designs the knowledge base structure according to the characteristics of the history questions in middle schools.II.Because it is difficult to be precise participle for historical materials, it is often easy to separate the relevant events and specialized nouns in the segmentation time, so it is necessary to construct a word segmentation list.The new word discovery technology is used in the construction of the lexicon.III.The expansion of high school history knowledge base.In order to answer the history questions of college entrance examination comprehensively and accurately, it is not enough to construct the knowledge base by using the contents of history textbooks, so it is necessary to expand them.To expand the data for crawling Baidu encyclopedia history related items.The word2vectdoc2vec model is used to train the semantic vector, which is combined with four statistical classification models, such as LRX SVM random Forester GBDT and so on.Finally, the optimal classifier is used to select the items that are suitable to be extended to the knowledge base of history in middle school.IV.High school history knowledge base content classification annotation.Through the analysis of the history questions of the college entrance examination, the attributes of each historical entity can be divided into seven categories, such as influence class, background class and so on.Therefore, when the historical entity is expanded into the knowledge base, it is necessary to classify its contents.The trained word vector is used as the input of the classification model, and the naive Bayesian LSTM / BLSTM C-LSTM model is selected to carry out the comparative experiments. Finally, the optimal classifier is selected for the content classification and tagging of the high school history knowledge base.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP18

【相似文献】