Work

Dynamic Epistemic–Temporal Graph Reasoning With Uncertainty Calibration for Evidence-Consistent Forensic Interrogation Narratives

Year: 2026
Type: article
Abstract: Forensic interrogation narratives are dynamic and evolve over time, across claims, events, locations, and evidential sources, so statement-level analysis alone is inadequate to support the reliability assessment of claimed consistency.The proposed dynamic epistemic-temporal graph reasoning framework with uncertainty calibration for evidence-consistent forensic interrogation narratives aims to provide a robust framework for generating forensic interrogation narratives that are evidence-consistent and incorporate uncertainty.It combines ModernBERT for contextual language representation, DeBERTa-v3 as a natural language inference model to model support and contradiction, a Dynamic Epistemic-Temporal Heterogeneous Graph Transformer for multi-hop reasoning, and conformal prediction to provide uncertainty-aware decision support.Interrogation discourse becomes structured representations of claims, entities, events, temporal expressions, locations, and relations of evidence.These elements are embedded in a dynamic heterogeneous graph that represents semantic, temporal, and evidential relationships between elements of the evolving story.The proposed Dynamic Epistemic-Temporal Heterogeneous Graph Transformer aims to detect inconsistencies that might not be captured by single-or two-part analysis, especially when evidence is spread across various claims, events, time references, places, and sources of evidence.Natural language inference facilitates semantic relation assessment, and temporal reasoning deals with event order, duration, overlap, and timeline compatibility.Conformal prediction is used to separate reliable consistency judgments from cases that are more ambiguous and require human review.It goes beyond the classification of deception and provides a framework of dynamic, evidence-based narrative reasoning designed to complement, not supplant, expert forensic judgment.The evaluation emphasizes evidence-consistency classification, semantic, temporal, evidential, and multi-hop reasoning, as well as uncertainty calibration.The future goals of crossdataset generalization and real-world robustness are left for future work.
Institution: Magadh University
Language: English
Field: Psychology
Open Access status: hybrid
Estimated APC paid:
OpenAlex ID: W7225611327
FWCI:
Cites: 34
Cited by:
Related to: 10