Inception Inference: Nested Probabilistic Reasoning over Story Graphs from Text

Gokul Srinath Seetha Ram

Keywords: Story Graphs, Probabilistic Inference, Narrative Understanding, Uncertainty Quantification, Counterfactual Analysis, Structured Text Analysis

TL;DR: We introduce Inception Inference, a 3-layer probabilistic framework that extracts story graphs from narratives, quantifies uncertainty with confidence scoring, and tests robustness via counterfactual analysis.

Abstract: We introduce Inception Inference, a 3-layer hierarchical framework for extracting and reasoning over story graphs from narrative text. The framework combines base-level event extraction, meta-level confidence scoring, and counterfactual-level robustness analysis to reveal narrative arcs and causal relationships. We provide a complete implementation with evaluation metrics, baseline comparisons, and visualization tools. Our approach achieves strong performance across Graph-F1, BLEU, calibration error, and robustness metrics, with statistically significant improvements over baselines. Inception Inference produces interpretable outputs with quantified uncertainty and structural stability, advancing probabilistic approaches to story understanding in alignment with the SPIGM workshop’s focus on structured inference and generative modeling.

Submission Number: 97