Telling Stories through Multi-User Dialogue by Modeling Character Relations

• Character-driven continuation must maintain persona consistency, track relationships, and logically advance the story while generating character interactions.Source: §1 · p. 1
• The task predicts the next character response using dialogue-only continuation over collaborative Dungeons and Dragons narratives.Source: §1 · pp. 1–2
• The extended CRD3 dataset adds automatically extracted character personas and pairwise relationship types from Critical Role Wiki descriptions.Source: §1 · p. 2
• Training jointly on character-related sub-tasks significantly improves story continuation accuracy.Source: §1 · p. 2
• The primary contributions are the extended CRD3 methodology and a multi-task system leveraging character relations and persona information.Source: §1 · p. 2

• Symbolic storytelling systems enforce plot coherence through explicit rules, whereas neural systems generally learn plot-driven generation from story corpora.Source: §2 · p. 2
• Modern plot-generation pipelines often generate a high-level plot before filling in the story under plot constraints.Source: §2 · p. 2
• Existing storytelling tasks generally omit dialogue and character interactions, including Story Cloze’s prediction of a story ending.Source: §2 · p. 2
• Persona-focused dialogue systems model agent consistency and contextual information but typically lack story or plot and use less long-term context.Source: §2 · p. 2

3.1 CRD3 Automated Dataset Extension

• CRD3 contains 159 episodes, 398,682 turns, and 34,243 human-written summary dialogue chunks spanning dialogue and first- and second-person narration.Source: §3.1 · pp. 2–3
• The dataset extension clusters Critical Role Wiki relation descriptions with TF-IDF and K-means, then assigns positive, negative, or neutral labels to dialogue chunks.Source: §3.1 · p. 3

3.2 Multi-task Learning

• The multi-task model predicts both the next character and the next story response from summaries, pairwise relationship labels, and the last n interaction turns.Source: §3.2 · p. 3
• The architecture uses separate task-specific output layers over shared pre-trained transformer blocks.Source: Figure 1 · p. 3, §3.2 · p. 3
• Joint training combines ranking and classification losses, sharing network parameters between response retrieval and next-character prediction.Source: §3.2 · p. 3

4 Evaluation

• The base evaluation model is a BERT-like bi-encoder ranker, and the experiments use the architecture described for jointly modeling the two subtasks.The reported ablations build on this transformer ranker.Source: §4 · p. 3
• The evaluation varies prior summary chunks and compares base versus relation-enriched inputs under single-task and joint multi-task training.Story continuation uses Hits@N, while character prediction uses frequency-weighted F1.Source: §4.1 · p. 4, §4.2 · p. 4, §4 · p. 4

4.1 Historical Context Ablations

• Performance continues increasing with added historical context through the maximum tested length of 10, indicating that CRD3 requires unusually long contexts.The largest average gain occurs when increasing context from one chunk to two, with smaller gains from additional chunks.Source: §4.1 · p. 4, Table 4 · p. 4

4.2 Multi-task Ablations

• The relation labels remain fixed across the entire story, effectively averaging relationship changes over time.More fine-grained time-varying labels might perform better but would require extensive additional human annotation.Source: §4.2 · p. 4
• For story continuation, the Base+Summary+Relations multi-task model outperforms all other tested models, while multi-task models consistently beat their single-task counterparts.The authors attribute the improvement to combining relationship-cluster inputs with next-character prediction, supporting more accurate persona-consistent continuation.Source: §4.2 · pp. 4–5
• Character prediction is strongest for Base+Summary multi-task and Base+Relations single-task models, closely matching Base+Summary+Relations multi-task performance.Thus, the relation-aware multi-task configuration is most clearly advantageous for story continuation rather than character prediction.Source: §4.2 · p. 4

• Ablation studies support the hypothesis that relationship information improves character-driven story completion.Source: §5 · p. 5
• Multi-task training on character dialogue and automatically extracted relationship information is identified as a particularly efficient way to provide this information.Source: §5 · p. 5

• Character-driven storytelling may inform learning agents that communicate naturally while maintaining a consistent character or persona.Source: §6 · p. 5
• Because the system learns from fantasy role-playing interactions, it may reproduce data biases, including non-normative language; retrieval mitigates but does not eliminate some risks.Source: §6 · p. 5

• Table 5 contains randomly selected CRD3 extended examples.Source: Table 5 · p. 7
• Both tables provide example-based views of the extended CRD3 data.Source: Table 5 · p. 7, Table 6 · p. 7
• Table 6 contains randomly selected CRD3 extended examples.Source: Table 6 · p. 7

• The multi-task model’s hyperparameters were adapted from Urbanek et al. (2019) and not tuned further.Source: Table 7 · p. 8
• All experiments were conducted on 4 Nvidia TitanX GPUs.Source: Table 7 · p. 8
• The experiments take less than 24 hours.Source: Table 7 · p. 8

• Table 8 presents pair-wise character relationship information.Source: Table 8 · p. 8
• The relationship information uses clustering.Source: Table 8 · p. 8
• The relationship information uses VADER.Source: Table 8 · p. 8