Learning Hierarchical Representations of Stories by Using Multi-layered Structures in Narrative Multimedia

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Digital Humanities

Background:

• Existing narrative embedding methods often focus solely on the story level.
• Analyzing narrative structures requires understanding relationships across multiple granularities (e.g., characters, scenes, movies).
• A unified framework is needed to represent and compare narrative elements at various scales.

Purpose of the Study:

• To develop a hierarchical representation learning framework for narrative utterances.
• To enable comparison of narrative elements at different granularity levels (e.g., scenes, characters, movies).
• To propose methods for learning embeddings that capture multi-layered narrative structures.

Main Methods:

• Proposed a four-layered structure for movies: character roles, characters, scenes, and movies.
• Introduced Char2Vec to represent characters based on dynamic role changes within character networks.
• Developed Scene2Vec to learn scene representations from character interactions.
• Utilized Hierarchical Story2Vec to capture story meaning from the sequential order of scenes.

Main Results:

• The proposed methods successfully generated embeddings for narrative utterances at different levels.
• Evaluation demonstrated the model's ability to estimate similarities between narrative elements in real movies.
• The hierarchical approach effectively captures the relationships between characters, scenes, and stories.

Conclusions:

• The hierarchical representation learning framework provides a robust method for analyzing narrative structures.
• This approach allows for a more comprehensive understanding and comparison of narrative works.
• The developed methods (Char2Vec, Scene2Vec, Hierarchical Story2Vec) offer valuable tools for computational narrative analysis.