Personalizing Kanji Memorization: Designing Adaptive Mnemonics Based on Learner Preferences Using Large Language Models

Kanji learning poses persistent challenges for learners of Japanese, requiring the memorization of visually complex characters alongside their meanings. Keyword mnemonics are widely used to reduce this burden by associating kanji components with memorable cues, yet their effectiveness varies substantially across learners due to differences in prior knowledge, visual interpretation, and narrative preference. We present personalized kanji mnemonic generation using a large language model (LLM), emphasizing learner-controllable operations that allow users to iteratively shape the mnemonic to their preferences. Informed by a formative interview study, we design and implement an iOS prototype that generates initial mnemonics and enables learners to iteratively adjust component keywords, narrative style, and story constraints.We evaluate this approach in a small within-subjects user study comparing personalized LLM-generated mnemonics with standardized human-authored mnemonics, measuring user preference ratings and meaning recall both immediately and after a delay. Participants rated personalized mnemonics higher on average, while immediate meaning recall was comparable across conditions. Delayed recall outcomes showed greater variance. Finally, our qualitative analysis highlights opportunities for improving personalized kanji mnemonics.