多场景网络学习中基于行为-情感-主题联合建模的学习者兴趣挖掘关键技术研究

With the in-depth integration of information technology and education, the multi-context and large-scale development of online learning have urged the generation of massive learning behavioral data, and also offer a great number of clues for exploring learners' interest topics in interactive learning. Taking the dependencies among discourse behavior, textual sentiment and topic as the starting point, the project studies the forming mechanism and mining approach of learners’ interest topics to offer the decision support for personalized teaching in multi-context online education. With focuses of describability and computability of learners’ interests and comprehensibility of behavior-sentiment-topic joint model, this research will involve the following three aspects: (1) Studying multi-course context oriented representation approach of learners’ interests to comprehensively describe the interest topics of learners and dependencies among various factors driving interest topics; (2) Studying the generation and solution approach of behavior-sentiment-topic model to accurately compute the joint probability distribution among behaviors, sentiments and topics of individuals in different courses; (3) Relying on actual multi-context online education environment, conducting the typical experimental design and analytics of learners’ interest mining to fully verify the validity and reliability of proposed theories, approaches and technologies, and to offer valuable evaluation and feedback for further improvement. Research outcomes will be used to support personalized intelligent tutoring in massive online education as key technologies.

随着信息技术与教育的深度融合，网络学习的多场景化与规模化发展正促使着海量学习行为数据的动态生成，也为探索学习者在互动学习中的兴趣主题信息提供了大量线索。课题拟以话语行为、文本情感与主题间的依赖关系为切入点，研究学习者兴趣主题的形成机理与挖掘方法，为个性化网络教学提供决策支持，围绕学习者兴趣的可描述性、可计算性以及行为-情感-主题联合模型的可理解性问题，展开三方面的研究内容：1）研究面向多课程场景的学习者兴趣表示方法，全面描述学习者的兴趣主题及驱动其产生的各要素依赖关系；2）研究行为-情感-主题模型的生成与求解方法，准确计算个体在不同课程中的行为、情感与主题的联合概率分布；3）依托实际多场景网络学习环境，开展学习者兴趣挖掘的典型实例设计与分析，全面验证本项目所提理论、方法与技术的有效性与可靠性，并进行评估与反馈改进。研究成果有望形成面向大规模网络学习的个性化智能导学关键技术支撑。

随着网络学习场景和规模的快速发展，学习者在参与学习活动时则会产生大量文本信息，本项目以话语行为、文本情感和主题间的依赖关系为切入点，对学习者的兴趣主题形成机理和相关影响进行了研究，旨在加强网络学习环境中教师对学生的个性化指导。整个项目期间，相关研究成果主要包括以下几个部分：1.针对当前文本数据无法反映学习过程情绪变化趋势的问题，提出一种无监督的论坛文本挖掘模型，即时序-情感主题模型。该模型可将情感与时间联合建模，以捕获时间维度上的情感变化。在此模型基础上引入兴趣主题，可发现互动话语中学习者关键兴趣主题及其演化趋势。2.提出基于学习者行为和情感多特征的话题模型，该话题模型可实现对学习者个人关注的话题进行检测，并可识别话题关联的行为和情感内容。提出的模型在泛化能力和输出结果上，表现出更好的性能优势。3.基于论坛文本的实证研究与分析。本项目基于提出的模型，以项目负责人所在学校的云课堂学习者话语行为与主要数据来源，并开展了学习者话语差异、认知行为模式、情绪特征和兴趣对学习成效影响的相关性实验。