From Narrative to Network

SagaGraph audits how reliably modern NLP reconstructs the social world of the Icelandic family sagas — and how much the errors in that reconstruction distort the historical and literary conclusions drawn from it. Every edge in every network here traces back to the sentence that licenses it.

Research questions

• RQ1How accurately can modern NLP reconstruct saga social networks?
• RQ2Which relation types are hardest to extract?
• RQ3How sensitive are network conclusions to extraction errors?
• RQ4Which structural properties recur across sagas?
• RQ5Does network centrality correspond to narrative prominence?

Early answer to RQ1: on the hand-annotated Njáls saga sample, the rule-based extractor with coreference reaches 45% F1 on relation extraction (lenient, chapter-level), against a large-language-model baseline at 40% — see the full breakdown.

The pipeline

Raw text → segmentation → named-entity detection → entity normalization → coreference → relation and event extraction → confidence scoring → human correction → network construction → analysis. Four independent extraction methods are compared against a manually annotated gold subset: sentence co-occurrence (a naive baseline), a dependency-rule classifier, the same classifier with pronoun coreference, and structured extraction by a large language model.

Corpus