【作者】 黄焕；

【导师】 刘清堂；

【作者基本信息】 华中师范大学 ， 教育技术学， 2014， 博士

【摘要】 近年来,随着多媒体技术和网络技术的发展,e-Learning已成为人们学习的另一种主要方式。e-Learning打破了传统教学在时间和空间上的限制,给学习者提供了一个宽松、自由、开放的学习环境。然而,在e-Learning中,学习者以自主学习为主,长时间处于孤立的学习状态而缺乏情感支持,这已经成为制约e-Learning学习效果的一个关键因素。构建一个具有情感能力的智能学习系统,在学习过程中为学习者提供一定的情感帮助和支持服务,是解决e-Learning中情感缺失的一种有效方法。学习者模型是学习者的个体特征在计算机系统中的一种抽象表示,代表了智能学习系统所理解和认识的学习者。因此,学习者情感建模是实现智能学习系统具有情感能力的关键,具有重要的研究意义。微博是当前最流行的一种社交网络平台,具有简单性、便利性和实时性,受到了青少年学生的广泛青睐。它不仅为学习者提供了一种简单、即时的沟通方式,而且为他们提供了一个自由抒发个人情感的空间,为实现学习者情感建模提供了一条新途径。但是,微博信息噪音较多、内容简短、语法松散,这也给学习者微博的情感分析带来了很多挑战。本文围绕着学习者情感模型的构建,重点研究了基于微博分析的学习者情感建模技术,具体研究内容包括五个方面：(1)融合情感特征的学习者模型；(2)学习者情感型微博的识别；(3)学习者微博的情感语义描述；(4)学习者微博的情感类别判定；(5)学习者情感建模的实现及其在适应性学习系统的应用。在以上研究的基础上,本文取得的研究成果主要体现在以下几个方面：(1)针对现有的学习者模型在特征选择上没有考虑学习者的情感,论文提出了一个融合情感特征的学习者模型,并给出了一个基于微博分析的学习者情感建模框架。该学习者模型包括了学习者五个方面的个体特征,并详细给出了每个特征的形式化描述方法；对于情感状态的形式化描述,基于心理学中的分类表示模型和空间表示模型,提出了一种以基本情感为维度的空间表示方法；通过分析微博在学习者情感建模中的可行性和优越性,提出了一个以微博情感分类为基本思路的学习者情感建模框架。(2)针对学习者微博中包含较多没有表达情感的噪音数据,且难以形成具有代表性的样本数据集,论文提出了一种基于单类分类的情感型微博识别方法,以解决一般的二元分类算法无法有效识别出情感型微博的问题。首先,通过对大量的实例进行分析与总结,提炼出了五条非情感型微博过滤规则,以过滤掉那些易于识别的非情感型微博,缩小情感型微博识别的范围；然后,在现有研究的基础上,定义了26个区分情感型微博和非情感型微博的分类特征,并使用单类学习算法来进一步识别出情感型微博,提高了情感型微博识别的准确性。(3)针对学习者微博的内容简短,使用以关键词作为特征的向量空间模型来描述微博的情感语义会导致严重的数据稀疏问题,论文提出了一种基于情感特征词聚类的微博情感语义描述方法,从特征约简和权重计算两个方面来缓解数据稀疏问题。首先,通过停用词过滤、低频词过滤和基于信息增益的过滤三种策略,从微博情感分类语料库中选择出情感特征词；然后,基于词语的上下文计算词语间的语义相似度,采用层次聚类算法将语义相似的情感特征词聚为一簇,使用词簇作为微博情感语义描述的特征；最后,提出了一种词簇特征的权重计算方法,以组成词语的权重之和为基础,并考虑词语的词簇代表能力和词簇的类别区分能力。(4)针对学习者微博中情感表达具有模糊性,一种分类算法往往不能有效区分所有的情感类别,论文提出了一种基于元学习的微博情感类别判定算法,以充分利用分类算法之间的互补性,从整体上提高学习者微博情感分类的准确性。首先,分别使用朴素贝叶斯、最大熵和支持向量机算法学习得到三个不同的基情感分类器；然后,通过叠加策略和N折交叉验证的方式得到元训练集,并使用元学习算法从中学习得到一个元情感分类器；最后,依次使用基情感分类器和元情感分类器判定学习者微博的情感类别,并将元情感分类器的结果作为最终结果。更多还原

【Abstract】 In recent years, with the rapid development of multimedia technology and the Internet, e-Learning has become another important way for learning. E-Learning breaks the limitation of time and space in traditional teaching mode, and provides learners with a loose, free and open learning environment. However, in e-Learning environment, learners mostly learn by themselves, which result in the feeling of isolation. That leads learner’s lack of emotional support, and then limits the effect of e-Learning.To give emotional support for learners, we should develop an intelligent learning system which can understand learners’ emotion and give them some appropriate responses. Learner Model, as one of the most important components of intelligent learning system, stores the individual characteristics of learners, which are the basis of implementing personalized and adaptive learning service. Therefore, modeling learner’s emotion is of great value in powering the intelligent learning system with emotional capability.Micro-blog is currently the most popular online social network, with simplicity, convenience and instantaneity, widely used by the young students. It provides learners with not only a simple, instant way for communication, but also a free space for emotion expression. So it can become a new way of modeling learners’ emotion. However, learners’ micro-blogs usually include many non-emotional data, and the content is very short, and the syntax is also very loose, which bring many challenges for emotion mining from learner’s micro-blogs.Around modeling learner’s emotion, this paper focus on the emotion modeling technologies based on analysis of his/her micro-blogs. The main research content includes five aspects:(1) The learner model that integrated emotion characteristics;(2) The recognition of learner’s emotional micro-blogs;(3) The description of emotional semantic in learner’s micro-blogs;(4) The emotion classification algorithm for learner’s micro-blog;(5) The implementation and application of learner emotion modeling. Based on the above studies, the contributions of this thesis are mainly reflected in the following aspects:(1)Aiming at that the characteristics of existing learner models are not systematic, especially do not consider the learner’s emotion, we put forward a learner model with emotion characteristic and an emotion modeling framework based on micro-blog analysis in this dissertation. The learner model includes five characteristics of learners, and offers their formalization description respectively. As for the formalization description of learner’s emotion, based on the category model and space model in psychology, a space model using the basic emotions as the dimensions was proposed. Furthermore, we analyzed the feasibility and advantages of micro-blog for learner’s emotion modeling, and proposed a learner emotion modeling framework.(2)As there are much noise data in learners’ micro-blogs, we study on the automatic recognition of emotional micro-blogs. Generally, we can use a binary classification algorithm to classify the micro-blogs into emotional class or non-emotional class, but it is hard to form a statistically-representative sample set of the non-emotional class. So we proposed a new approach based on one-class classification to recognize the emotional micro-blogs. First, based on the analysis of a large number of non-emotional micro-blogs, we summarized five rules to filter out the most easily identifiable non-emotional micro-blogs. Then26features were selected to represent the micro-blogs and the one-class classification algorithm was used to further recognize the emotional micro-blogs.(3)The learners’ micro-blogs are usually very short. So if we use the general vector space model based on keywords to describe the emotional semantic of micro-blogs, it will lead to a very large, sparse feature space. In order to resolve this problem, we proposed a new approach based on words clustering to describe the emotional semantic of micro-blogs. First, we selected the words which are useful for emotion expression from the micro-blogs corpus by three strategies:stop words filtering, low-frequency words filtering and filtering based on the information gain. Then we computed the semantic similarity of words based on their contexts, and proposed a hierarchical clustering algorithm to group words into several clusters according their semantic similarity. Finally, we used the word clusters as features, and proposed a feature weighting method which includes three factors:the sum of weighting values of all component words, the representation degree of component words, and the discrimination degree of word cluster.(4)There are many classification algorithms, but most existing studies only use one algorithm to predict the emotion category of micro-blogs, and do not leverage the complementarities between them. To improve the emotion classification accuracy of learners’ micro-blogs, an ensemble approach based on meta-learning was proposed. First, we respectively used the Naive Bayes, Maximum Entry and Support Vector Machine to learn three component classifiers from the micro-blog corpus. Then we employed the stacking method and N-fold cross validation to produce the training samples for meta-learning, and used the Maximum Entry or Support Vector Machine to learn a meta-classifier. Finally, we used the component classifiers and the meta-classifier to predict the emotion category of micro-blogs successively, and the result of the meta-classifier is the final result.更多还原