CASIA OpenIR  > 毕业生  > 博士学位论文
基于认知记忆激活的语义表示方法及其应用研究
白洁1,2
学位类型工学博士
导师曾大军
2018-05-27
学位授予单位中国科学院研究生院
学位授予地点北京
关键词语义表示 认知心理学 文本表示 网络表示学习 信息传播
摘要
语义表示与分析是人工智能领域的基础研究问题。随着互联网数据的爆炸式增长,如何实现高效准确的语义表示,进而从大规模数据中挖掘有价值的语义信息并加以利用,具有十分重要的研究价值。然而,面对日益复杂的海量动态数据计算需求,新的研究挑战也在不断涌现。例如在文本表示领域,由于社交媒体平台的用户生成文本具有语义模糊、新词、非规范表述等特性,如何更有效地挖掘和表示文本的隐含语义内容?在网络表示学习领域,如何将大规模网络中包含的抽象语义信息高效地表示出来? 
由于人脑在知识表示与语义分析方面的天然优势,借鉴认知心理学中人脑的记忆激活机制完成语义表示与分析成为一个重要的研究思路。然而目前已有的方法大多需要借助专家经验或已有知识,无法适应大规模、动态化的互联网计算环境。基于以上现状,本论文有针对性地探索了如何更有效地将记忆激活理论应用于语义表示工作,并服务于现实应用。具体来说,本论文将认知心理学中的 ACT (Adaptive Control of Thought)记忆激活理论体系引入文本和网络的自动语义表示工作中,分别探索了 ACT 理论体系中的关联激活理论和扩散激活理论在文本和网络语义表示中的应用价值,提出了相应的文本增强表示方法和网络表示学习方法。并进一步将所提出的方法联合建模,应用于模因爆发预测的实际任务当中。 
本论文的主要研究内容与贡献包括: 
1. 将 ACT 关联激活理论应用于文本表示,提出了基于关联激活的文本增强表示方法AADE。 
从信息论“熵”的角度,论证了利用现有信息激活隐含语义概念的关键因素,并通过这些因素与 ACT 关联激活理论中激活长期记忆所需要素的一致性,论证 ACT 关联激活理论在文本隐含语义表示和分析中的应用原理与价值。在此基础上,提出基于关联激活的文本增强表示方法 AADE,包括一个总体框架以及三个具体的 AADE 模型。通过在长、短文本,中、英文数据集上进行的文本极性分析、主题建模、文本检索、分类、聚类等一系列基础性文本分析实验,证明了 AADE 方法能够在线性时间内显著提升多个文本分析任务的效果。对比当前的主流文本表示方法,AADE 能够有效挖掘文本中的隐含语义信息并进行显式化表示,兼具有效性、可解释性、高效性与兼容性。 
2. 将 ACT 扩散激活理论应用于网络表示学习,提出了基于扩散激活的网络表示学习方法 Spread-gram。 
首先论证了 ACT 扩散激活公式用于网络表示学习的可行性,同时提出 Spread-gram 模型的目标函数。然后根据 ACT 扩散激活理论中的节点传播模式实现网络的扩散激活式节点搜索与向量更新策略。在此基础上,根据网络的不同类型,分别提出针对同质网络和异质网络的 Spread-gram 模型。在同质网络和异质网络上完成的节点分布可视化、节点分类、链接预测等一系列实验表明,Spread-gram 训练获得的网络节点向量能够有效地表示网络中的语义信息,在多个任务中取得显著效果提升,且仅需很少的迭代次数便可以使模型达到收敛。对比已有方法,Spread-gram 实现了网络的全局输入,避免了随机游走等方法可能带来的局部输入和输入偏差问题,同时具有模型快速收敛的特性。 
3. 探索了 AADE 和 Spread-gram 方法在信息传播领域的联合应用,提出了模因爆发实时预测方法 SNT。 
该方法以信息传播单元 —“模因”作为研究对象,通过观察模因产生初期的传播情况,预测其未来是否会呈现爆发式传播。SNT 整合了模因传播过程中的文本语义特征、网络空间特征和时序特征,并对这三个领域特征进行联合建模,实现模因爆发的实时预测。在模因传播的公开数据集以及真实采集数据集上,通过真实场景的实验模拟以及典型案例分析,分析了 SNT 的有效性。相比已有方法,SNT 充分利用了模因传播过程中的语义与时空信息,在对不同类型的模因进行实时爆发预测任务时都能取得较好的效果,且预测有效性不依赖于监测时间,这对于信息传播分析和早期预警具有十分重要的意义。
 
其他摘要
Semantic representation and analysis are fundamental research questions of artificial intelligence. Specifically, with the Internet data explosion, it is of great importance to represent semantics efficiently and accurately, so as to acquire and utilize valuable information from large-scale data. However, new challenges are emerging along with the increasingly complex demand of data analysis. For example, in the field of text representation, the user generated contents usually have characteristics such as semantic ambiguity, neologism, and non-canonical expressions. Based on these facts, how to mine the implicit semantics from text effectively? In the field of network representation learning, how to effectively represent the abstract semantic information from the large-scale networks? 
As the nature of human brain in knowledge representation and semantic analysis, one of the potential ideas of semantic representation and analysis is to take advantages of the human memory activation mechanism. However, most existing related researches rely on the expert experience or prior knowledge, which can not fit for the large-scale, dynamic computing environment adaptively.This thesis explores the application of memory activation theories to semantic representation, as well as their further real-world practices. Specifically, this thesis study the feasibility of applying the ACT theory in the field of text and network semantic representation, and put forward a text representation method and a network representation learning method accordingly. Furthermore, the proposed method is adopted jointly on the meme burst prediction task. 
The main researches and contributions of this thesis include: 
1. Apply the ACT association activation theory to the text representation question, and propose a text enrichment method named AADE. 
From the perspective of ``entropy'', we study the key factors to enable activation of implicit information, which demonstrates the possibility of applying ACT association activation theory in textual implicit semantic representation analysis. Based on the demonstration above, we propose an overall paradigm AADE and a set of follow-up AADE models. The experimental results show that the AADE models significantly improve multiple text mining tasks on various text with linear time consumption. Compared with the existing methods, AADE discover and represent the implicit semantics in text, which is more effective, more explainable, more efficient and more compatible for text representation. 
2. Apply the ACT spreading activation theory to the network representation learning question, and propose Spread-gram. 
First, we demonstrate the feasibility of the ACT spreading activation formula for network representation learning, and propose the objective function of the Spread-gram model. Then, based on the spreading activation mechanism in the ACT theory, the network's spreading-activating search strategy is implemented. According to the different types of networks, the specific Spread-gram models for homogeneous networks and heterogeneous networks are proposed separately. A series of experiments on both homogeneous and heterogeneous networks shows that significant improvements are achieved in multiple tasks within only a small number of iterations. Compared with the existing methods, Spread-gram learns network representations through global input, which avoids the possible deviation caused by local input methods, and achieves convergence soon in the model training. 
3. Explore the joint application of AADE and Spread-gram in the field of information dissemination, and propose a real-time meme burst prediction method named SNT. 
Taking the information dissemination unit ``meme'' as the research object, SNT predicts whether a meme will burst in the future by observing its initial propagation. SNT investigats and incorporates semantics-based features, network-based features and temporal features during meme propagations to predict meme burst in real-time. The effectiveness of SNT is evaluated through multiple tasks conducted on a public dataset and a real-world dataset. Compared with the existing methods, SNT takes full advantages of the semantic and spatio-temporal information in meme propagation, and is effective for various kind of memes. The effectiveness of SNT prediction is independent of monitoring time, which is of great importance for information dissemination analysis and early warning.
 
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/21023
专题毕业生_博士学位论文
作者单位1.中国科学院自动化研究所
2.中国科学院大学
推荐引用方式
GB/T 7714
白洁. 基于认知记忆激活的语义表示方法及其应用研究[D]. 北京. 中国科学院研究生院,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于认知记忆激活的语义表示方法及其应用研(6621KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[白洁]的文章
百度学术
百度学术中相似的文章
[白洁]的文章
必应学术
必应学术中相似的文章
[白洁]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。