In exploring anime, users often find it challenging to formulate their preferred narrative features as explicit search conditions. We present a character-centered exploration vision in which familiar characters serve as cross-title queries connecting users’ preferences with unfamiliar narrative content. As an initial text-based proof of concept, we introduce Chara2Vec, which constructs fixed-length character representations from textual descriptions. Both Chara2Vec and SBERT showed positive mean rank correlations with human judgments, although the design did not support a controlled comparison between the methods. These results suggest that description-based representations can capture aspects of perceived character similarity and motivate future research on multimodal and spoiler-controllable character representations.
