Salience Scoring

Assigns higher weights to entities appearing in structurally prominent document zones. An entity in a title tag (H1) or opening paragraph receives a higher score than one buried in a footer.

• HTML Semantics: Parses <title>, <h1>-<h6>, and <strong> tags.
• Positional Bias: Applies a decay function to entities appearing later in the document.
• Document Object Model (DOM): Analyzes visual hierarchy for prominence cues.

Balances local frequency against global rarity. A term that appears frequently in a specific document but rarely across the entire corpus is deemed highly salient.

• Term Frequency (TF): Raw count or logarithmic scaling of local occurrences.
• Inverse Document Frequency (IDF): Penalizes common stop words and generic terms.
• Sublinear Scaling: Prevents high-frequency entities from dominating the score disproportionately.

Models the document as a lexical co-occurrence network where nodes are entities and edges represent semantic connections. Algorithms like TextRank compute centrality to identify the most interconnected, and therefore salient, entities.

• Eigenvector Centrality: Measures influence based on connections to other high-scoring nodes.
• Random Walks: Simulates information flow through the entity graph.
• Window Size: Defines the co-occurrence span (typically 2-10 words).

Uses transformer models like BERT to generate dynamic vector representations. Salience is calculated as the cosine similarity between an entity's contextualized embedding and the document's overall topic embedding.

• Dynamic Representations: Resolves polysemy by analyzing surrounding context.
• Attention Weights: Aggregates token-level importance from the final transformer layer.
• Centroid Proximity: Measures distance from the entity vector to the document's semantic centroid.

Evaluates grammatical roles to determine importance. An entity acting as the nominal subject (nsubj) of a main clause is weighted more heavily than one in a prepositional modifier.

• Dependency Parsing: Constructs a tree of binary grammatical relations.
• Role Hierarchy: Prioritizes agents, subjects, and direct objects over adjuncts.
• Root Distance: Calculates the syntactic path length from the entity to the sentence's root verb.

Aggregates mentions across pronouns and noun phrases to calculate a unified salience score. The entity 'Apple' and its coreferent 'it' or 'the tech giant' are linked into a single chain.

• Mention Clustering: Groups all textual references to the same real-world entity.
• Chain Length: Longer, uninterrupted chains indicate higher discourse centrality.
• Anaphora Resolution: Resolves backward references to calculate cumulative prominence.

The foundational information extraction subtask that locates and classifies named entities in unstructured text into pre-defined categories such as person, organization, location, and medical codes. Salience scoring depends on accurate NER as its first step—you cannot weight an entity's importance if you cannot identify it. Modern systems use contextualized embeddings from transformer models to resolve ambiguous mentions like 'Apple' (company vs. fruit) based on surrounding context.

The NLP task of connecting textual entity mentions to their corresponding unique, unambiguous entries in a knowledge base like Wikidata or DBpedia. This disambiguation step is critical for salience scoring because it consolidates multiple surface forms ('NYC', 'New York', 'Big Apple') into a single canonical entity ID. Without entity linking, a salience algorithm fragments importance scores across lexical variants rather than aggregating them correctly.

A numerical statistic reflecting how important a word is to a document in a collection. Calculated by multiplying term frequency (how often a term appears) by inverse document frequency (how rare the term is across the corpus). While TF-IDF operates at the lexical level rather than the entity level, it serves as the statistical ancestor of modern salience scoring. Its core insight—that rarity amplifies importance—directly informs entity weighting algorithms.

A graph-based ranking algorithm for NLP that identifies the most important sentences or keywords in a document based on their co-occurrence links. TextRank builds a graph where nodes are lexical units and edges represent co-occurrence within a window. It then applies a PageRank-inspired iterative voting mechanism to score each node's centrality. This graph-theoretic approach directly parallels how entity salience is computed in knowledge graphs.

A statistical measure of how frequently two distinct named entities appear together within a defined context window. High co-occurrence signals semantic relatedness and is used to build and enrich knowledge graphs. In salience scoring, co-occurrence patterns help distinguish central entities (those that co-occur with many others) from peripheral ones. This metric underpins graph centrality calculations like degree centrality and betweenness centrality.

The NLP task of finding all expressions in a text that refer to the same real-world entity, linking pronouns ('she', 'it') and definite noun phrases ('the company', 'the CEO') back to their named antecedents. Salience scoring requires coreference resolution to aggregate all mentions of an entity—explicit and anaphoric—into a single importance weight. Without it, an entity referenced primarily through pronouns would be incorrectly scored as low salience.