【作者】 张俊；

【导师】 李青； 骆祥峰；

【作者基本信息】 上海大学 ， 计算机应用技术， 2016， 博士

【摘要】 随着网络的迅速发展,网络中的文本信息呈现海量、快速、多样、稀疏的特征,从而使得网络用户经常出现迷航现象。在此情况下,我们需要一种面向机器的文本语义理解方法来辅助人们高效地组织、管理这些大规模网络文本信息,进而为人们提供优质的网络服务。但是,目前的机器文本语义理解方法存在以下问题,包括:(1)人类文本语义表达能力与机器文本语义理解能力之间存在较大鸿沟;(2)机器文本语义理解复杂性与网络文本数据海量性之间的矛盾;(3)机器的文本语义理解过程缺乏背景知识指导。针对上述三个问题,本文提出了基于人类认知过程的文本语义理解HTSC模型,本模型的基本思想为“认知框架+轻量级算法实现”,通过模拟人类的认知过程,以使机器文本语义理解结果更符合人类的认知结果,从而解决问题(1);通过轻量级的算法实现,以使本模型能够适应、处理大规模的网络文本环境,从而解决问题(2);在认知框架下,通过当前文本语义与背景知识间循环往复的动态理解过程,以从全局的角度来理解文本语义,从而解决问题(3)。本文具体研究以下内容:1.基于人类的记忆认知过程和文本信息加工过程理论,提出由感觉记忆、知觉关联记忆、工作记忆、短期情景记忆和长时记忆等几部分构成的文本语义理解HTSC模型,为面向机器的文本语义理解提供框架指导。2.在HTSC模型总体框架基础上,提出了HTSC模型中各个记忆系统下的文本语义表征方法,从而为机器的文本语义理解提供理论载体,包括:基于课文多水平表征理论,提出基于工作记忆的文本语义表征方法;基于关键词、篇章、话题三层文本语义网络,提出HTSC模型中基于短期情景记忆和长时记忆的文本语义表征方法;基于人类概念学习过程,提出文本语义幂级数表征方法。3.在语言学配价理论的基础上,提出了词之间语义关系的挖掘方法,避免了传统关联规则挖掘方法语义信息低下、挖掘效果参数依赖性高的缺陷,为机器文本语义理解中语义关系的发现提供了可能,包括:基于配价理论以及句子压膜理论,提出文本中动词依赖集挖掘方法;在动词依赖集基础上,结合关联规则挖掘方法,提出词之间原子语义关系挖掘方法;基于词之间原子语义关系,并结合文本中动词依赖集,提出词之间具体语义关系挖掘方法。4.在人类认知过程的基础上,提出了面向机器的文本语义动态理解过程,以使得机器的文本语义理解更符合人类的认知结果,包括:基于连通图理论,对文本语义连贯范式进行形式化定义;基于文本语义连贯范式,提出当前文本语义中激活线索的产生方法;基于语义链网络,提出背景知识激活方法以及融合方法。5.基于人类概念学习过程中的复杂度度量理论,提出了文本理解的复杂度度量方法,从而为不同层次用户提供个性化服务提供了可能,包括:基于概念代数复杂度度量,提出文本代数复杂度ACT;通过分析文本语义理解复杂度的影响因子,提出关键词及其之上关联关系的复杂度度量方法;基于关键词及其之上关联关系复杂度,提出文本代数复杂度改进方法,包括扩展型文本代数复杂度EACT和泛化型文本代数复杂度GEACT。6.基于HTSC模型及其之上的复杂度度量方法,研究与开发了两个应用演示系统:基于认知过程的智能交互系统和基于复杂度度量的搜索演示系统,从而为本文提出的HTSC模型提供了真实、合理的应用场景。本文以基于人类认知过的文本语义理解模型研究为突破,探索建立面向机器文本语义理解的理论与方法体系,进一步从理论上研究了文本语义的表征方法、词之间语义关系发现方法、当前文本语义与背景知识的动态交互方法,以及复杂度度量方法等。本研究成果能够直接应用于语义搜索、个性化推荐、用户关系发现、网络舆情监测等方面。更多还原

【Abstract】 With the rapid development of the Web,online text information tends to be volume,velocity,variety and value,which make Web users always be lost on the Web.Under this situation,it requires a machine-oriented textual semantic comprehension method to help users efficiently organize and manage those massive textual information,and then to offer proper Web services.However,there are some problems that the current textual semantic comprehension methods are facing,including:(1)there is a great gap between the cognitive ability of humans and the comprehensive ability of machines;(2)the contradiction between the high complexity of comprehension and the volume feature of the texts on the Web;(3)the comprehension process lacks of the guidance of the background knowledge.To deal with the above problems,this paper proposes a textual semantic comprehension model based on human cognitive process(HTSC),whose basic idea is “cognitive framework + implementation with lightweight algorithms”.Through mimicking human cognitive process,the comprehension results of HTSC are much close to the results of human cognition,which is properly to solve problem(1);Implementation with lightweight algorithms makes HTSC model capable of dealing with large-scale online texts,which is properly to solve problem(2);HTSC model supports the dynamic interactions between the current text and the background knowledge,which is properly to solve problem(3).The main research work of this paper is as below:1.On the basis of human memory process and text information processing,proposing the HTSC model that is comprised of sensory memory,perceptual associative memory,working memory,transient episodic memory and long-term memory,so as to provide theoretical support for machine-oriented textual semantic comprehension.2.Under the framework of HTSC model,proposing textual semantic representations in different memory systems,so as to provide semantical carriers for machine-oriented textual semantic comprehension,including: working memory based multi-level textual semantic representation;transient episodic memory and long-term memory based textual semantic representation,which is composed of three levels: word,discourse and topic;power series representation model is proposed based on human concept learning.3.Based on the valency theory in linguistics,proposing a semantical relation mining algorithm between word,which avoids the defects of weak semantics and strong dependency on parameters carried by traditional relation mining algorithms.This part of work including: on the basis of valency theory and the restrict of memory capacity,proposing the verb dependency set mining algorithm;based on verb dependency set and association relation mining algorithm,proposing the elemental relation mining algorithm between words;combined with verb dependency set and elemental relations,proposing the specific relation mining algorithm.4.Inspired by human cognitive process,proposing a method for dynamic comprehension of textual semantics,so as to make the comprehension results much close to human cognition.Works including: based on the theory of connected graph,giving formalized definition of coherence interruptions of textual semantics;based on coherence interruptions of textual semantics,proposing the generation method of activation cue;under semantic link network,proposing the activation method of background knowledge and its fusion algorithm.5.On the basis of human concept learning,proposing the complexity measurement of textual semantics,which is capable of providing personalized services for users having different background knowledge.Works including: in view of the algebraic complexity of concept,proposing the algebraic complexity measurement of text,ACT;through analyses of those facts that have influences on the complexity measurement,giving the complexity measurement for words and relations;on the basis of the complexity measurement of words and relations,proposing two kinds of expanding methods for ACT,which are extension of algebraic complexity of text,EACT,and general extension of algebraic complexity of text,GEACT.6.Under the theory of HTSC model and its complexity measurements,developing two application systems: cognitive process based intelligent interaction system and complexity measurement based search engine demo,which provide real,suitable application scenarios for our proposed HTSC model.This paper is willing to discover the theory and methodology of machine-oriented textual semantic comprehension,by the means of human cognitive process.It gives theoretical analyses and experiments on textual semantic representation,semantic relation discovery between words,dynamic interaction algorithm between the current text and the background knowledge,and complexity measurement.The research achievement can be directly applied to semantic search,personalized recommendation,relations discovery between users and online public sentiment monitoring,etc.更多还原