A Comparative Study of Knowledge Component Granularity in Intelligent Language Tutoring Systems

Intelligent Language Tutoring Systems (ILTS) aim to provide adaptive language-learning experiences tailored to individual students. A major challenge is determining the appropriate granularity of Knowledge Components (KCs) for predicting student performance. Accurate prediction is essential for identifying activities within the Zone of Proximal Development, which students can complete with scaffolded feedback. Although much prior research focuses on STEM and higher-education contexts, this study investigates KC granularity in school-based language learning. We analyze 153,819 first-attempt responses from 1,059 seventh-grade students in Germany, collected while they were using the web-based ILTS FeedBook. Using the Additive Factors Model (AFM) and Performance Factor Analysis (PFA), we evaluate 19 KC models spanning 11 levels of the domain hierarchy, from fine-grained item-level representations to semantically grounded abstractions. Cross-validation shows that the preferred granularity depends on the evaluation setting. Fine-grained item-level models perform best for new students on known content, while semantic mid-level models perform consistently strongly when generalizing to new content. Regarding model fit and complexity, the Akaike Information Criterion favors fine-grained models in both implementations, while the Bayesian Information Criterion rankings diverge because the implementations represent and penalize model complexity differently. Learning rates and learning curves provide related views of opportunity-based practice patterns, with near-zero rates for some fine-grained models partly reflecting limited repeated opportunities. Overall, the findings show that KC model selection depends on the intended modeling objective and highlight the value of semantically grounded mid-level representations for generalizable and interpretable learner modeling in school-based ILTS.