Introduction
Children with Autism Spectrum Disorder (ASD) frequently encounter difficulties with narrative coherence—the ability to construct logically linked, causally dependent, and socially contextualized stories. In natural settings, storytelling is often a collaborative effort where parents "scaffold" the child's narrative through targeted questioning. This research applies Behavioral Signal Processing (BSP) to quantify how different parental elicitation strategies—specifically "context-oriented" versus "action-oriented" prompts—influence the structural organization and coherence of stories told by children with ASD.
Methodology: An Entity-Based Approach
The researchers developed a computational pipeline focused on "entities" (characters and objects) to model narrative structure. The process begins with Coreference Resolution to group mentions of the same entity. From these groups, several key features are extracted:
- Frequency Descriptors: The distribution of entity mentions is modeled using exponential and Generalized Pareto (GP) distributions. A steep distribution curve indicates a narrative focused on a few central characters, whereas a flatter curve suggests a more fragmented story.
- Markov Chain Modeling: By treating the dominant entity of each sentence as a state, the system calculates transition probabilities, such as the likelihood of the child maintaining focus on the primary character.
- Entity Evolution: This tracks the "path" of entities throughout the transcript. Mentions increase an entity's score, while their absence decreases it. The "directional distance" between these paths reveals how characters interact or compete for prominence.
Key Insights: Scaffolding and Narrative Focus
The analysis revealed distinct structural "fingerprints" associated with different prompting styles.
Action-Oriented Scaffolding: Questions like "What is the boy doing?" encourage children to follow a specific character's trajectory. This results in narratives centered around a few dominant entities, creating a "common axis" that enhances coherence. As seen in the entity evolution sequences, action narratives typically show a primary entity that receives sustained attention.
Context-Oriented Scaffolding: Conversely, prompts focusing on setting or appearance (e.g., "What color is the balloon?") tend to produce "associative" narratives. These stories often mention many entities briefly without establishing a clear hierarchy, leading to a flatter and less cohesive structure.
Results and Statistical Significance
The entity-based features were evaluated using a K-Nearest Neighbor (K-NN) classifier to distinguish between context, action, and mixed narratives.
- Classification Performance: The system achieved a 62.88% unweighted accuracy. Notably, Markov-based "EntityTransition" features alone reached 59.85% accuracy, significantly outperforming simple entity counts (52.27%).
- Structural Indicators: Action narratives demonstrated a significantly higher probability ($p < 0.05$) of returning to the primary and secondary entities in consecutive sentences compared to context narratives. This validates the hypothesis that action-prompting fosters a tighter, more focused narrative loop.
Conclusion and Future Directions
While limited by a small sample size (11 subjects), this study demonstrates the power of entity-based modeling in capturing the nuances of communicative interactions in ASD. Future work aims to integrate semantic and syntactic data—such as verb usage and emotional tone—into the feature set. Additionally, researchers plan to compare parent-moderated stories with those elicited by Embodied Conversational Agents (ECAs) to explore how virtual peers might be programmed to provide optimal scaffolding for narrative development.
Summary updated