联系我们

论文查询结果	返回搜索

论文编号：	16150
作者编号：	2120243878
上传时间：	2026/6/9 22:11:36
中文题目：	“搜索即学习”视角下GenAI可解释性对大学生学习效果的
英文题目：	Impact of GenAI Explainability on Student Learning: SAL Perspective
指导老师：	李颖
中文关键字：	信息搜寻情境；搜索即学习；GenAI可解释性；学习效果；认知负荷
英文关键字：	information seeking context; Search as Learning; GenAI explainability; learning outcomes; cognitive load
中文摘要：	随着生成式人工智能（GenAI）技术的快速发展，以ChatGPT、Kimi等为代表的对话式搜索工具正在深刻改变着大学生的学术信息搜寻行为。然而，GenAI在用户端的生成过程与推理逻辑往往呈现不够透明，容易引发用户的信任不足与认知偏差，这也使其可解释性及其在教育领域的影响成为学界关注的热点。本研究从“搜索即学习”（Search as Learning, SAL）视角出发，聚焦大学生学术信息搜寻情境，旨在探究生成式AI的可解释性如何影响用户的信息行为和学习效果。 研究以Kumarage和Saarela提出的生成式AI可解释性三角框架（生成机制透明性、用户可理解性、评估保真度）和认知负荷理论为理论基础，构建了“可解释性-认知负荷-学习效果”的理论框架。采用双组前后测准实验设计，以“算法偏见”为搜索主题，招募了78名有GenAI使用经验的大学生，并将其随机分为高可解释性组与低可解释性组。实验通过定制化的Kimi界面操纵可解释性水平，采用概念图前后测评估被试的实际学习效果（知识增益），并结合问卷测量其主观学习体验（感知满意度、有用性、相关性、准确性）及认知负荷。 研究得出以下发现：（1）生成式AI的可解释性水平会对被试的实际学习效果产生正向影响，实验结果表明，高可解释性组在概念图总分及丰富性、有效性、层次性、探索性维度上的增益都要明显高于低可解释性组，但在对专业性的增益方面二者无明显差别；（2）可解释性对于被试的主观学习体验的影响在具体维度方面存在一定差异，其中，高可解释性组在感知满意度和感知准确性上的得分均明显高于低可解释性组，但在感知有用性和相关性方面二者并无明显差异；（3）研究证实了认知负荷在可解释性与部分学习效果指标之间的中介作用，高可解释性设计通过降低用户的外在认知负荷，释放认知资源，从而间接提升了概念图增益、满意度和准确性；（4）可解释性三角框架的三个维度对学习效果存在差异化影响：评估保真度对感知准确性影响最强，用户可理解性对满意度影响最突出，生成机制透明性则对概念图增益影响较为明显。 本研究拓展了“搜索即学习”理论在智能信息环境下的适用边界，揭示了GenAI可解释性影响学习效果的认知机制。研究建议，高校应将AI可解释性认知纳入信息素养教育，GenAI工具开发者应构建多维度协同的可解释性设计，并推动相关学术规范与行业标准的建立，以引导学生在智能时代形成善用AI而不依赖AI的批判性信息素养。
英文摘要：	With the rapid development of Generative Artificial Intelligence (GenAI) technology, conversational search tools represented by ChatGPT and Kimi are profoundly transforming university students' academic information-seeking behaviors. However, the generation process and reasoning logic of GenAI are often insufficiently presented to users, which may lead to trust deficits and cognitive biases. This has made explainability and its impact on education a focal point of academic attention. From the perspective of "Search as Learning" (SAL), this study focuses on the context of university students' academic information seeking, aiming to explore how the explainability of Generative AI influences users' information behaviors and learning outcomes. Grounded in Kumarage and Saarela's triangular framework of GenAI explainability (transparency of generation mechanisms, user comprehensibility, and evaluation fidelity) and Cognitive Load Theory, this research constructs a theoretical framework of "Explainability-Cognitive Load-Learning Outcomes." Employing a two-group pretest-posttest quasi-experimental design with "algorithmic bias" as the search topic, the study recruited 78 university students with GenAI experience and randomly assigned them to either a high explainability group or a low-explainability group. The experiment manipulated explainability levels through a customized Kimi interface, assessed participants' actual learning outcomes (knowledge gain) using pre- and post-test concept mapping, and measured their subjective learning experiences (perceived satisfaction, usefulness, relevance, and accuracy) along with cognitive load through questionnaires. The study yielded the following findings: (1) The level of GenAI explainability positively impacts participants' actual learning outcomes. Results indicate that the high explainability group demonstrated significantly greater gains than the low-explainability group in total concept map scores and in the dimensions of richness, effectiveness, hierarchical structure, and exploratory depth, though no significant difference was found in the professionalism dimension. (2) The impact of explainability on participants' subjective learning experiences varies across specific dimensions. The high explainability group scored significantly higher in perceived satisfaction and perceived accuracy compared to the low-explainability group, while no significant differences were observed in perceived usefulness and relevance. (3) The study confirmed the mediating role of cognitive load between explainability and certain learning outcome indicators. High explainability designs indirectly enhanced concept map gains, satisfaction, and accuracy by reducing users' extraneous cognitive load and freeing up cognitive resources. (4) The three dimensions of the explainability triangular framework exert differentiated influences on learning outcomes: evaluation fidelity had the strongest impact on perceived accuracy, user comprehensibility most prominently influenced satisfaction, and transparency of generation mechanisms showed a more pronounced effect on concept map gains. This study extends the boundaries of "Search as Learning" theory in intelligent information environments and reveals the cognitive mechanisms through which GenAI explainability affects learning outcomes. The research recommends that universities incorporate AI explainability literacy into information literacy education, that GenAI tool developers construct multi-dimensional collaborative explainability designs, and that academic standards and industry guidelines be established to guide students in developing critical information literacy for the intelligent era—enabling them to use AI wisely without becoming dependent on it.
查看全文：	预览  下载(下载需要进行登录)