This paper presents a novel method for assisting learners from a Roman alphabet background in learning challenging Japanese vocabulary, particularly Kanji. To overcome the black-box limitations of existing large-scale language model (LLM)-based keyword association techniques, we propose a generative framework that explicitly models the process of associative memory formation using Kanji components. This framework uses a novel expectation-maximization algorithm to learn latent structures and compositional rules from associative memory data generated by learners on an online platform. This enables the generation of interpretable and systematic associative memories, and demonstrates particularly strong performance in cold-start environments for new learners.
