resolucao-correferencia-e-aliases-literarios

Use a sliding window of the previous 7 sentences to find candidate PERSON entities, then score them by syntactic agency: subjects of speech verbs get +2.0, subjects of physical actions +1.5, and objects +0.5. The highest combined score wins the pronoun link.

Compare each detected name against canonical entities using Jaro-Winkler similarity with a default threshold of 0.85. Matches like "Sra. Silva" to "Helena Silva" are merged under the same canonical_id automatically.

You need structured JSON output from a named entity recognition step containing entity text, category, and character offsets, plus a Portuguese dependency parser such as UDPipe or an adapted SpaCy model for subject and object role extraction.

Scenes with multiple same-gender speakers create ambiguous candidates within the sentence window. When the top two scores differ by less than 0.15, the system consults the scene's cooccurrence graph and picks the character with highest centrality.

The validation target is a minimum F1-score of 75% on alias clustering using an adapted LitBank Coreference dataset. Simple nominal ambiguity cases like "Maria Silva" and "Sra. Silva" must merge into one canonical_id in 100% of regression tests.