GraphLit: Learning Text-Enriched Dynamic Character Network Representations for Literary Study

Technical Summary: GraphLit: Learning Text-Enriched Dynamic Character Network Representations for Literary Study

Problem Statement
Existing computational methods for modeling fictional characters generally fall into two categories, both of which suffer from significant limitations. First, Character Network (CN) approaches model characters as nodes and their interactions as edges (e.g., co-occurrence or dialogue). While effective for capturing social structures, these methods often treat characters as static functions, overlooking the rich textual context (signifiers like names, descriptions, and speech patterns) in which characters appear. Second, text-based methods and Large Language Models (LLMs) capture textual nuances but fail to integrate the social dynamics inherent in character networks. Furthermore, LLM-based representations often struggle with generalization to unseen novels and rely on copyrighted data that hinders reproducibility. There is a need for a framework that explicitly represents both textual and relational information within a reproducible, public-domain corpus.

Methodology
The authors propose GraphLit, a self-supervised learning framework built upon Dynamic Heterogeneous Character Networks (DHCNs).

1. Data Construction (DHCNs):

   • The framework processes long novels from Project Gutenberg (approx. 20,000 books) by dividing them into temporally localized text blocks.
   • Within each block, a Heterogeneous Character Network (HCN) is constructed containing two node types: Character nodes and Segment nodes (textual chunks).
   • Edges are defined as:
     • Character-Character: Based on co-occurrence within a restrictive window (20 words).
     • Character-Segment: Connecting a character to the specific text segments where they are mentioned.
   • The full DHCN for a book is a sequence of these disconnected HCNs, preserving narrative time.

2. Representation Learning (GraphLit):

   • Initialization: Character nodes are initialized with prompts containing names and aliases; Segment nodes are initialized using Qwen3-Embeddings.
   • Architecture: The model adapts the Heterogeneous Graph Masked Autoencoder (HGMAE) framework.
     • A Transformer Decoder processes the ordered sequence of segment attributes within a block to preserve text coherence.
     • A Heterogeneous Graph Transformer (HGT) backbone refines node representations by aggregating messages from neighbors, distinguishing between node and edge types.
   • Training Objectives: The model is trained via self-supervision to reconstruct corrupted graphs:
     • Edge Reconstruction: Predicting masked character-character and character-segment links.
     • Attribute Reconstruction: Reconstructing masked initial node attributes (names/text) using a scaled cosine loss.
     • Ordering Loss: A ListMLE loss is applied to ensure the model learns the correct narrative order of blocks.
     • Global Link Prediction: A task to dissociate characters across different books to encode book-level information.
   • Pooling: Global representations for characters, blocks, and entire books are generated using heterogeneous multi-head attention layers.

Key Contributions

1. DHCN Formulation: The introduction of Dynamic Heterogeneous Character Networks, which extend Dynamic Character Networks by grounding characters in their localized textual contexts via segment nodes.
2. GraphLit Framework: A self-supervised learning framework that learns expressive representations from DHCNs, outperforming text-only and graph-only baselines.
3. Digital Humanities Application: A demonstration of how DHCN-derived topological measures correlate with narrative non-linearity (circuitousness), suggesting that social discrepancies are a key component of non-linear storytelling.

Results
The framework was evaluated on 12 character-related tasks from the Character Embedding Benchmark (CEB) and a Quotation Attribution task.

• Performance Gains: GraphLit consistently outperforms text-only baselines (e.g., Qwen3 embeddings) and graph-only baselines (DCNs). The improvements are most pronounced in contextual tasks (e.g., character identification from masked descriptions, quotation attribution), where understanding the interplay between character and text is crucial.
• Ablation Insights:
  • Context over Social Structure: Surprisingly, removing explicit character-character edges (No-CN variant) resulted in competitive or slightly superior performance on book-level and contextual tasks. This suggests that the primary source of information is the grounding of characters in their dynamic textual contexts, rather than the social network structure itself.
  • Role Prediction: Explicit character networks provided the most significant gains for socially grounded tasks like role prediction and protagonist identification.
  • Dynamics: Static CN structures performed on par with dynamic ones for book-level tasks, indicating that contextual dynamics within blocks are more critical than the evolution of the network across blocks for certain tasks.
• Narrative Non-Linearity: Regression analysis revealed that topological features derived from DHCNs (transitivity, new character ratios) significantly correlate with narrative non-linearity. GraphLit embeddings captured social discrepancies more effectively than semantic-only embeddings, showing that non-linear narratives often involve shifts in social focalization.

Significance
The paper claims that GraphLit offers a reproducible framework for learning rich literary representations that explicitly bridge the gap between formalist approaches (network structures) and textualist approaches (signifiers). By demonstrating that character representations grounded in textual context are more informative for downstream tasks than those relying solely on social interaction graphs, the work supports the literary critical view that characters emerge through evolving textual signifiers. Furthermore, the ability to link social network dynamics to narrative non-linearity provides a new quantitative tool for Digital Humanities, suggesting that the evolution of character relationships is a measurable indicator of narrative structure. The authors note that while their results are robust, the framework relies on imperfect character name clustering and that future work could explore coreference resolution and broader downstream applications.