Invited Talks

Abstract

Publication-oriented academic writing remains challenging for many Chinese postgraduates, partly because writing instruction often emphasizes how to construct individual manuscript sections without making explicit how a paper is ultimately judged for scholarly merit and publishability. This lecture presents the development and pilot validation of a Dual-Layer Assessment Framework for Academic Writing (DFAFAW) designed to bridge these two levels of evaluation.

The framework integrates two complementary rubrics: the Manuscript Structure & Quality Rubric (MSQR), which evaluates eight core sections of a research article, and the Scholarly Merit Rubric (SMR), which addresses seven higher-order dimensions, including originality, significance, claim–evidence alignment, methodological rigor, clarity, contribution, and positioning in related work. Evidence was gathered through expert review, cognitive pretesting, a postgraduate survey (403 valid responses), and classroom implementation with 27 students.

Findings provide preliminary support for the framework's conceptual coherence, interpretability, and instructional usability, while also revealing difficulties novice writers face in evaluating higher-order scholarly qualities. The lecture will discuss how this dual-layer approach can make publication expectations more visible and support self-assessment, feedback, and strategic manuscript revision.

Abstract

Knowledge tracing (KT) is a core technique for personalized learning, which models the dynamic changes of learners’ knowledge states to predict future learning performance. Nevertheless, practical KT applications still suffer from inherent limitations, including the conflict between stable long-term knowledge structures and volatile short-term learning behaviors, sparse interaction records for new learners in cold-start scenarios, and poor cross-domain transferability of learning patterns, which severely restrict the comprehensive performance of intelligent educational systems. This talk will introduce a progressive integrated research framework to enhance the completeness, robustness, and generalizability of KT models systematically.

To address the multi-scale characteristics of learning behaviors, the framework first establishes a graph-based representation learning paradigm to jointly model long-term stable knowledge structures and short-term dynamic learning activities, equipped with an adaptive fusion strategy to dynamically balance the weights of dual temporal states for accurate prediction. Furthermore, to alleviate the data sparsity problem in cold-start settings, the framework leverages large language models to enrich multidimensional semantic attributes of knowledge concepts and adopts a dual-view contrastive learning mechanism to align behavioral and semantic features, eliminating redundant and unreliable generated information and strengthening the robustness of sparse sequence modeling.

On this basis, the framework optimizes cross-domain KT by retaining complete student learning trajectories rather than relying on single final knowledge states, and constructs a discrete cognitive codebook from continuous knowledge state sequences to capture universal cross-context cognitive prototypes. The learned codebook serves as transferable prior knowledge to realize implicit domain alignment and improve generalization in data-scarce target domains.