IR-KG-LLM-Alignment

Historical Narrative Conflict Detection

IR-KG-LLM-Alignment tackles the challenging problem of detecting conflicting claims across historical narrative sources, where the same events are described with different perspectives, omissions, and outright contradictions. The system combines information retrieval, knowledge graph construction, and LLM-based belief graph reasoning to identify and categorize narrative conflicts.

The pipeline begins with BM25 retrieval to gather relevant passages across a corpus of historical texts, then constructs a knowledge graph where entities are historical actors, events, and locations, and edges encode claimed relationships (participated in, caused, contradicts). An LLM-powered belief graph reasoner operates over this KG to identify inconsistencies, temporal impossibilities, and perspective-dependent claims.

The belief graph formalism distinguishes between factual claims (verifiable), interpretive claims (perspective-dependent), and mythological claims (culturally constructed), enabling nuanced conflict detection that respects the nature of historical discourse. The system can identify when two sources agree on facts but disagree on interpretation, versus when they make contradictory factual claims.

This work has been submitted to IJCAI-ECAI, CoNLL, and ACL venues, contributing to the intersection of NLP, knowledge representation, and digital humanities. The evaluation dataset covers Ancient, Medieval, and Modern history across three languages.

• BM25 + Knowledge Graph + LLM belief graph reasoning pipeline for narrative conflict detection
• Submitted to IJCAI-ECAI, CoNLL, and ACL
• Three-tier claim taxonomy: factual, interpretive, and mythological
• Evaluation across Ancient, Medieval, and Modern history in three languages
• Distinguishes factual contradictions from perspective-dependent interpretive disagreements