Entity Salience

Salience is not merely frequency; it measures how structurally central an entity is to the document's narrative.

• Title & Headers: Entities appearing in titles or section headers receive higher salience scores.
• Subject Position: Entities occupying the grammatical subject position in key sentences are weighted more heavily.
• Coreference Chains: The length and density of a coreference chain (all mentions of the same entity) directly boost salience.

A person mentioned once in passing has low salience; the same person as the subject of 15 sentences has high salience.

Salience decouples importance from raw term frequency to solve the common-word problem.

• TF-IDF Limitation: Standard TF-IDF overweights generic but frequent entities (e.g., "United States" in a news corpus).
• Salience Correction: A salience model learns that "United States" is often background context, not the topic.
• Mechanism: Models use features like syntactic depth, discourse role, and document position to assign a normalized score between 0.0 and 1.0.

This allows a system to correctly identify that a rare entity mentioned in the lede paragraph is more salient than a frequent entity in boilerplate text.

Salience models exploit the rhetorical structure of a document to assign importance.

• Lede Proximity: News articles follow an inverted pyramid; entities in the first 50 words are exponentially more likely to be salient.
• Section Density: In academic papers, an entity concentrated in the Abstract and Conclusion is more salient than one scattered evenly.
• Syntactic Prominence: Entities in main clauses are weighted higher than those in subordinate clauses or parentheticals.

These features allow a model to distinguish between a primary topic and a tangential reference without any external knowledge base.

A critical distinction in information retrieval: salience is intrinsic to the document, while relevance is query-dependent.

• Salience: "Elon Musk" is the most salient entity in his biography, regardless of the user's search.
• Relevance: "Elon Musk" becomes relevant to a query about "electric vehicles," even in documents where he is not salient.
• Application: Search systems use salience for document summarization and indexing, and relevance for ranking against a query.

Confusing these two concepts leads to poor keyphrase extraction that misses the document's actual focus.

Modern systems train neural classifiers to predict salience directly from contextual embeddings.

• Input: BERT or similar transformer embeddings for each entity mention, concatenated with positional and syntactic features.
• Architecture: A feedforward layer on top of the mention embedding predicts a binary or continuous salience label.
• Training Data: Models are trained on datasets where human annotators have marked entities as primary, secondary, or tertiary in importance.
• Output: A ranked list of entities with calibrated confidence scores, enabling downstream filtering (e.g., keep top-K salient entities).

This approach generalizes across domains without hand-crafted rules.

Salience acts as a critical pruning mechanism in keyphrase extraction pipelines.

• Candidate Overgeneration: NER and noun phrase chunking often produce hundreds of candidate entities per document.
• Salience Thresholding: Entities with a salience score below 0.3 are discarded, dramatically reducing noise.
• Weighted Scoring: Final keyphrase scores are computed as a weighted product of salience and relevance: FinalScore = Salience * 0.6 + Relevance * 0.4.
• Result: A document about Apple's earnings call will surface "iPhone revenue" and "Tim Cook" while suppressing "Cupertino" and "NASDAQ."

Without salience filtering, keyphrase extraction degenerates into a noisy list of all mentioned entities.