【作者】 李旭；

【导师】 陈文智； 欧阳璠；

【作者基本信息】 浙江大学 ， 电子信息（专业学位）， 2024， 博士

【摘要】 结构化知识(命名实体)是构成知识图谱的主要元素,是描述、理解和分析文本的关键数据资源。构建教育知识图谱的重要前提是实现大规模学科知识的结构化组织。这些学科结构化知识可以用来呈现学科关键知识,描述不同知识之间的关联关系,为学习诊断、个性化学习、协作学习等提供智能支持。然而现有开源的知识图谱数据资源中包含与教育领域相关的细粒度结构化知识还非常少,导致这些开源数据难以支持面向教育情境的算法和工具研发。大量知识存在于无结构化的教育文本中,现有通过正则化技术和爬虫技术识别领域(半)结构化信息的技术方案难以迁移到不同的学科情境、无法有效支持面向学习过程的赋能需求。本研究认为要实现知识驱动的教育创新,必须结合深度学习技术来构建知识识别技术框架,并在不同教学情境中验证知识智能驱动的教育实践效果。总之,实现知识智能与教育的深度融合并非易事,本研究需要解决以下几个方面的关键挑战:(1)理论框架构建。构建技术赋能教育的理论框架是指导具体技术研发和教育实践的必要步骤。在这方面,该研究的主要工作包括:a)梳理支撑知识智能赋能教育的理论基础;b)构建并阐述理论框架的内容及特点;c)阐述理论框架与后续子研究之间的关联关系。(2)学科结构化知识数据资源构建。高质量数据集是构建基于深度学习技术的人工智能算法的基础性资源。教育领域缺乏结构化知识识别的开放数据集资源,阻碍了下游算法和相关智能应用的研发。在这方面,该研究的主要工作包括:a)邀请学科专家共同界定了教育学科的结构化知识范式(知识实体的类型);b)聘请学科专业人员标注、校正,构建了一个面向教育领域,具有较大规模且开源的结构化知识识别数据集;c)本研究基于该数据集进行大量的基准实验,评估现有人工智能模型的性能。(3)学科结构化知识识别算法研发。每个领域都有自己独有的术语体系、符号系统。在开放领域数据集上获得良好性能的算法模型迁移到教育领域会面临严重的性能损失,降低了算法模型的可用性。针对这个问题,该研究的主要工作包括:a)分析教育领域数据集的主要特征,尤其是数据集中领域词汇和歧义性词汇对知识类型识别和知识实体边界定位带来的影响;b)本研究提出了基于晶格结构的多粒度信息融合策略来增强对输入文本的语义表示,帮助模型更准确地识别领域词汇的类型和提升学科知识实体的边界定位能力;c)本研究通过大量实验验证了该模型的有效性。d)该模型提升了学科结构化知识识别的性能,为开发面向教育情境的知识智能工具提供了核心组件。(4)知识智能驱动的协作学习实证研究。基于大规模结构化知识研发的智能教学工具可以有效识别和表征学科知识及其相互之间的关联关系。研究人员需要研究如何利用知识智能教学工具的知识识别和分析技术,更好地支持知识建构、知识创造等知识密集型的教学实践,并探索该类工具在协作学习情境中的有效实践模式。在这方面,该研究主要的工作包括:a)针对教学写作情境中自动化评估面临的可解释性差、评估效度低等挑战,本研究基于结构化知识识别和表示技术开发了知识智能驱动的写作评估工具(Edu NERScore),并通过实证研究1探索了该工具对写作评估效度、评估效率和评估结果可解释性的影响。b)此外,针对学生在协作写作过程中对高质量反馈内容的诉求,本研究整合了结构化知识对比的教学策略,升级了Edu NERScore工具的相关功能。并通过实证研究2探究了学科结构化知识作为反馈内容对在线讨论、同伴互评和协作写作质量等方面的影响。综上,该研究的主要贡献包括以下三个方面:1)探究了“数据-模型-应用”的研究和实践路径,为人工智能+教育的深度融合提供了参照范式。2)在关键技术层面,本研究为教育领域贡献了一个结构化知识识别的数据集和人工智能模型,推动了教育知识图谱的发展。3)在应用层面,本研究探究了知识智能教学工具在写作评估和协作写作两个场景中的应用效果,揭示了结构化知识在增强写作评估可解释性、支持面向学习过程的高阶协作等方面的积极影响,为知识智能驱动的教育创新提供了实践案例。更多还原

【Abstract】 Structured knowledge(Named Entity)is the main element that forms the Knowledge Graph and is the key data resource for describing,understanding and analyzing texts.An important prerequisite for constructing an educational knowledge graph is to achieve a structured organization of large-scale disciplinary knowledge.These disciplinary structured knowledge can be used to present the key knowledge of the corresponding discipline,describe the association relationship between different knowledge,and provide intelligent support for learning diagnosis,personalized learning,and collaborative learning.However,existing open-source knowledge graph resources contain very little fine-grained structured knowledge related to the education,making it difficult for these open-source data to support the development of algorithms and tools for real educational contexts.A large amount of knowledge exists in unstructured educational texts,and existing technical solutions for identifying(semi-)structured information in the domain through regularization and crawling techniques are difficult to transfer to different disciplinary contexts,and do not effectively support the need for learning processoriented empowerment.This research believe that in order to realize knowledge-driven educational innovation,it is necessary to combine deep learning technologies to build the knowledge recognition technology framework,and to verify the effectiveness of knowledge intelligencedriven educational practice in different educational contexts.In conclusion,realizing the deep integration of knowledge intelligence and education is not an easy task,and this research need to address the following key challenges:(1)Theoretical framework construction.Building a theoretical framework for technologyenabled education is a required step to guide specific technology development and educational practice.In this regard,the main work of the research includes: a)sorting out the fundamental theories supporting knowledge-based intelligence empowering intelligent education; b)constructing and elaborating the content and features of the theoretical framework; and c)illustrating the correlation between the theoretical framework and the subsequent studies.(2)Discipline structured knowledge data resource construction.High-quality datasets are foundational resources for constructing AI algorithms based on deep learning techniques.The lack of open dataset resources for structured knowledge recognition in the education domain hinders the development of downstream algorithms and related intelligent applications.In this regard,the main work of this research includes: a)this research invited subject experts to define the structured knowledge paradigm(Named Entity schema)in the education discipline; b)this research hired subject professionals to label,correct,and construct the first large-scale,opensource Named Entity Recognition dataset for the education; and c)this research conducted a large number of benchmark experiments based on the dataset to evaluate the performance of existing models.(3)Discipline structured knowledge recognition algorithm development.Each domain has its own unique terminology system,symbol system.Algorithmic models that get good performance on open domain datasets migrated to the education domain will face serious performance loss and reduce the usability of the algorithmic models.To address this problem,the main work of this research includes: a)analyzing the main features of the education domain dataset,especially the impact brought by domain terminology and ambiguous vocabulary in the dataset on the identification of entity types and the boundary localization of knowledge entities; b)this research propose a multi-granularity information fusion strategy based on the lattice structure to enhance semantic representations of the input text,which helps the model to better accurately identify the types of domain vocabulary and to improve the boundary localization of disciplinary knowledge entities; c)this research verify the effectiveness of the model through extensive experiments; d)the model improves the performance of disciplinary structured knowledge recognition and provides a core component for the development of knowledge intelligence tools oriented to the educational context.(4)Empirical study of collaborative learning driven by knowledge intelligence.Intelligent teaching tools developed based on large-scale structured knowledge can effectively identify and characterize disciplinary knowledge and their interrelated relationships.Researchers need to study how to utilize the knowledge recognition and analysis techniques of knowledge intelligent teaching tools to better support knowledge-intensive teaching practices,and to explore effective practice pattern for knowledge intelligence-driven collaborative learning.In this regard,the main work of this study includes: a)in response to the challenges of poor interpretability and low assessment validity of automated assessments in academic writing contexts,this research developed a knowledge intelligence-driven writing assessment tool(Edu NERScore)based on structured knowledge recognition and representation techniques,and explored the tool’s impact on the validity of writing assessments,the efficiency of the assessment,and the interpretability of the assessment results through the empirical study 1.(b)in response to students’ demand for high-quality feedback content during collaborative writing,this research integrated the instructional strategy of structured knowledge comparison and upgraded the relevant functions of the Edu NERScore.And through the empirical study 2,this research explored the impact of disciplinary structured knowledge as feedback content on the quality of online discussion,peer review,and collaborative writing.In summary,the main contributions of this research include the following three aspects:1)this research have explored the research and practice path of "data-model-application" in a complete pattern,which provides a reference paradigm for the in-depth integration of AI+Education.2)from the perspective of key technologies,our work contributes the first data resource and intelligent model for structured knowledge recognition in the education domain,and pushes forward the development of the educational knowledge graph.3)from the perspective of education practice,this research explore the effectiveness of knowledge intelligence teaching tools in two scenarios: writing assessment and collaborative writing,revealing the positive impact of structured knowledge in enhancing the interpretability of writing assessment and supporting higher-order collaboration oriented to the learning process,which provides practical cases for future knowledge intelligence-driven educational innovation.更多还原