Font Size: a A A

Research On Technologies Of Emotion Recognition And Topic Mining For Course Comments

Posted on:2015-06-28Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z LiuFull Text:PDF
GTID:1227330467460393Subject:Education Technology
Abstract/Summary:PDF Full Text Request
In recent years, many open interactive learning platforms emerge in the field of education. These platforms provide abundant interactive contents, in which includes learners’ comments towards courses. These comments record problems learners encounter in learning, opinions towards course resources and teachers, etc. These opinions are quite valuable for assisting learners selecting courses, improving teaching quality and platform support. However, currently, the feedback information formed in large amounts of course comments is still not well exploited. How to exploit learners’ feedbacks to provide decision support for teaching has become a realistic problem that educational researchers are considering. Therefore, it is an urgent need for an effective means of text mining to process these comments, and to provide the intuitive, accurate and effective information for analysts.This paper mainly focuses on the key techniques of emotion recognition and topic mining of course comments. The purpose is to apply these techniques for learning behavior analysis, evaluation of online learning resources, etc., and to realize the deep combination of information technique and education. This study takes the emotion information hidden in course comments as the starting point. To deal with the problems of high-dimensionality of text data, high annotation cost of training corpus in emotion recognition and uncertainty of topic mining, we propose the technical route of "data acquisition-text feature selection-semi-supervised emotion recognition-topic emotion mining". According to the technical route, we first reduce feature dimensionality of the comment set and obtain some discriminative features, thus each comment sample is represented as a set of frequencies of selected features in the sample. Then we adopt a semi-supervised learning method to tackle the training set with large amounts of unlabeled samples, and verify the effectiveness of constructed emotion recognition model on testing samples. Finally, we use the emotion recognition model to obtain emotion labels of testing samples in the step of topic mining, and construct topic-emotion model with the known emotion labels. Besides, the topic-emotion model is respectively used to mine the key topics-emotion information for each course unit and learner. The work in this dissertation has been supported by the project "Research on key technologies of college youth online behavior oriented emotion recognition"(14BGL131) in National Social Science Foundation and the project "Research on key technologies of network behavioral security and adolescent development"(No.200603110400) in National Key Technology R&D Program in the12th Five year Plan of China.The main research works and innovations of this dissertation are as follows:(1) Aiming at the high dimensionality and redundancy of feature space of comment data, we propose a multi-space particle swarm optimization based feature selection method to select the discriminative N-gram features in a large feature space. Through the equal sized cross-division of the training sample space, we counstruct a group of particle swarms on each sample subspace to seek the best solution for feature selection. After several iterations, multi diverse feature selection strategies are formed, and then these selection strategies are fused to form a final feature evaluation result. Experimental results indicate that, compared with conventional feature selection methods, the proposed method can select more discriminative features, and the acceptable recognition accuracy can be achieved with the lower dimentional features.(2) Aiming at the high annotation cost of training samples in emotion recognition, we propose the adaptive multi-view selection (AMVS) based semi-supervised emotion recognition algorithm. The algorithm uses the constructed emotion vocabulary to calculate the emotion strength of each N-gram feature. Then, the distribution of feature emotion strength is used select feature views, and the importance distribution of feature dimensionality is adaptively constructed to determine the dimensionality of each view in an iterative way. Thus, a small number of views are formed to pick the credible unlabeled samples in an ensemble way. In the picking process, the unlabeled samples with highest credibility are selected and annotated to update the training set in each iteration. The semi-supervised training will be terminated after several iterations. Experiments results show that, compared with conventional multi-view semi-supervised learning methods, AMVS achieve the higher diversity among different views, and picked unlabeled samples are more credible, finally obtain the higher recognition accuracy. In addition, we also apply the recognition results to predict the support rate of online courses. Compared with user rating, the emotion recognition method can achieve a more accurate prediction closer to the actual emotion distribution of comments.(3) Since the different aspects about a course exist in comment set, to deal with the detection of local topics, we propose the deterministic emotional information based topic model (DEI-TM) to extract the local topic-emotion information hidden in comments. The method utilizes the emotion recognition model trained in the second step to obtain the emotion labels of testing samples, and establishes the dependencies of "emotion-text" and "text-topic-sentence" to train the topic-emotion model. To detect the local topics in comments, the DEI-TM models the key emotion sentences in comments instead of comments. In this model, it is assumed that a sentence only involves a topic and an emotion class, which fully considers the association between different sentences. Experimental results show that, compared with author-topic model, the proposed model can achieve the better generalization ability. Specifically, the inter-topic correlation measurement result shows that the similarity among different topics is lower, and the intra-topic entropy measurement result shows that words representing each topic have the higher polymerization degree. Finally, the proposed model is applied to the topic-emotion mining of course unit and individual learner, the probability distribution of emotion-topic-word is utilized to display key topic-emotion information for each course unit or individual learner.
Keywords/Search Tags:Course Reviews, Emotion Recognition, Feature Selection, Multi-ViewSemi-Supervised Learning, Topic Mining
PDF Full Text Request
Related items