Story Embedding: Learning Distributed Representations of Stories based on Character Networks (Extended Abstract)

O-Joun Lee, Jason J. JungIJCAI 2020 · DOI

TL;DR

This paper introduces a novel method for representing entire stories as fixed-length vectors by analyzing character interactions within narrative works. By treating character networks as dynamic social networks and applying graph embedding techniques, the authors propose models that capture the essence of a story, enabling better comparison and analysis of narratives.

What Problem Does This Paper Address?

Representing complex narratives like stories in a way that computers can understand and analyze is a significant challenge. Previous approaches often focused on specific events or manually crafted features, which are difficult to generalize across different types of media or tasks. The authors highlight that while character networks (who interacts with whom) are easier to extract than detailed event-based models, they lack semantic richness. The core problem this paper addresses is the lack of a robust, generalizable method for creating story representations that can capture narrative content and allow for meaningful comparisons between different stories, moving beyond media-specific or task-specific heuristics.

Methods

The paper proposes two main story embedding models, Story2Vec-F (flow-oriented) and Story2Vec-D (denouement-oriented), which are then combined into a unified model, Story2Vec-U.

  • Character Networks: Stories are first transformed into dynamic character networks. Nodes represent characters, and edges represent interactions (measured by dialogue frequency). These networks evolve over scenes.
  • Social Roles (Subgraphs): To capture local structures within character networks, the authors adapt the Weisfeiler-Lehman (WL) relabeling process. This process iteratively assigns labels to nodes based on their neighbors' labels, effectively identifying subgraphs that represent "social roles" of characters at different scales. A "proximity-aware WL relabeling" is introduced to incorporate interaction weights (proximity) alongside adjacency.
  • Graph Embedding (Doc2Vec): The identified social roles (subgraphs) are then embedded into vector representations using Doc2Vec.
  • Story2Vec-F: This model learns representations by predicting the co-occurrence of social roles within temporal-radial neighborhoods (considering both scene progression and local graph structure). It aims to capture the dynamic flow of a story.
  • Story2Vec-D: This model focuses on the final scene of the character network, aiming to capture the climax or denouement of the story. It uses a similar prediction mechanism but only considers the last scene's network.
  • Story2Vec-U: This unified model alternates between training Story2Vec-F and Story2Vec-D, attempting to leverage both dynamic and static aspects of the narrative.

Datasets

The evaluation was performed on 142 real movies. Character networks were extracted using a tool called CharNet-Extractor from movie scripts obtained from IMSDb. Ground truth for story similarity was collected from 50 human evaluators via a web application. Genre similarity was calculated using Jaccard index on genre sets from IMDB.

Key Insights and Ideas

  • Character interactions as narrative proxies: The fundamental assumption that the frequency and patterns of character interactions within a story can serve as a proxy for the story's content is a key insight. This allows for a more computationally tractable approach than detailed event extraction.
  • "Social Roles" as meaningful subgraphs: Adapting graph embedding techniques to identify and represent local subgraph structures ("social roles") around characters provides a way to capture nuanced relationships and character dynamics within the network.
  • Temporal and structural embedding: The combination of temporal (scene-based) and structural (graph-based) information through models like Story2Vec-F and Story2Vec-D offers a more comprehensive representation of narrative.
  • Denouement's importance: The finding that Story2Vec-D (focusing on the story's end) achieved the highest accuracy suggests that the climax and resolution of a story may disproportionately influence human perception of its similarity to other stories.

Key Results

The proposed models, particularly Story2Vec-D, outperformed genre similarity in estimating story similarity. The absolute errors for Story2Vec-D were lower than for Story2Vec-F and Story2Vec-U. While Story2Vec-U improved the variance of standalone models, it did not consistently outperform Story2Vec-D in accuracy. The paper presents experimental results in Figure 3, showing box-and-whisker plots of absolute errors for Story2Vec-F, Story2Vec-D, Story2Vec-U, and Genre, indicating the superiority of the proposed methods.

Limitations and Future Work

  • Integration of dynamic and static approaches: The authors acknowledge that the alternating training method for Story2Vec-U was not optimal and suggest the need for a better integration strategy for dynamic and static story features.
  • Granularity of stories: The current approach embeds entire narrative works. Future work could explore learning representations at different levels of granularity (e.g., scenes, chapters).
  • Multi-modal representation: The current models focus solely on textual narrative. Future research should incorporate other physical features of narrative multimedia (e.g., visual, auditory) for a more comprehensive representation.

Summary updated

BibTeX
@inproceedings{lee2020story,
  author = {O-Joun Lee and Jason J. Jung},
  title = {{Story Embedding: Learning Distributed Representations of Stories based on Character Networks (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/709},
}
Figure