Entity Salience Scoring

The foundational statistical method that calculates salience by multiplying Term Frequency (how often an entity appears) by Inverse Document Frequency (how rare the entity is across a corpus).

• Penalizes common entities like 'company' or 'year'
• Rewards distinctive, domain-specific entities
• Serves as the bag-of-words baseline against which neural models are benchmarked

Example: 'Apple' in a technology article scores higher than in a fruit farming guide due to IDF weighting.

Applies network theory metrics to the document's entity co-occurrence graph to identify the most structurally important nodes.

• Degree Centrality: Entities with the most connections to other entities
• Betweenness Centrality: Entities that bridge distinct topical clusters
• Eigenvector Centrality: Entities connected to other high-importance entities (like PageRank for documents)

These algorithms reveal which entity acts as the conceptual hub of the text.

Uses transformer-based models like BERT to generate contextualized vector representations of each entity mention and compares them to the document's overall semantic embedding.

• Cosine similarity between entity embedding and document centroid
• Captures semantic relevance beyond surface-level frequency
• Disambiguates homonyms through surrounding context

This method identifies entities that are semantically central even when mentioned infrequently.

Assigns higher salience to entities appearing in syntactically prominent positions within the document structure.

• Title and headings: Entities in H1/H2 tags receive maximum weight
• Subject position: Entities acting as grammatical subjects of main clauses
• Lead paragraph: Entities in the opening 100 words
• Topic sentences: Entities at the start of paragraphs

This heuristic mirrors how human readers identify the primary subject of a text.

Measures the discourse persistence of an entity by tracking how many times it is referenced through pronouns, nominal phrases, and name variations across the document.

• Long coreference chains indicate sustained topical focus
• Entities with short chains are likely tangential
• Requires robust coreference resolution as a preprocessing step

An entity mentioned once by name but referenced 20 times via 'it' or 'the company' is highly salient.

Augments local document salience with external authority signals from knowledge bases like Wikidata and Google's Knowledge Graph.

• PageRank of the entity's Wikipedia page
• In-degree in Wikidata (number of incoming relationships)
• Notability flags and external identifier counts
• SameAs assertions linking the entity across multiple authoritative sources

This prevents locally frequent but globally obscure entities from being over-weighted.

The foundational NLP process that identifies textual mentions of entities and maps them to their unique, unambiguous entries in a target knowledge base like Wikidata. This disambiguation is a critical prerequisite for salience scoring, as a system must first know which 'Paris' is being discussed before it can assess the entity's importance to the document. Without accurate entity linking, salience scores are assigned to ambiguous surface forms rather than distinct real-world objects, corrupting the entire analysis.

The algorithmic assignment of a numerical importance score to an entity (node) within a graph. In the context of a document's entity graph, node weighting is the direct implementation of salience scoring. Algorithms like PageRank or TextRank are adapted to this purpose, where nodes represent entities and edges represent co-occurrence or semantic relationships. A higher weight signifies a more central, and therefore more salient, entity. This score is what downstream systems use to prioritize information.

The NLP task of identifying all expressions in a text that refer to the same entity. For accurate salience scoring, a system must link not just proper nouns but also pronouns ('she', 'it', 'they') and nominal phrases ('the company', 'the new policy') back to their correct antecedent. Failing to resolve coreferences leads to a fragmented entity graph where a single, highly salient entity is incorrectly split into multiple low-importance nodes, severely underestimating its true contextual significance.

The assignment of a numerical value to a relationship (edge) in a graph to represent its strength or semantic distance. In entity salience, edge weights between co-occurring entities can be based on factors like proximity in the text, the specificity of the relationship verb, or the frequency of co-occurrence. A strong edge weight between a central entity and a peripheral one can elevate the peripheral entity's salience score, reflecting the influence of contextual association.

The specific sub-task of entity linking that resolves which distinct real-world entity a textual mention refers to when the name is ambiguous. This is the critical decision point that directly impacts salience. For example, in an article about the tech industry, the system must disambiguate 'Apple' as the corporation (Q312) rather than the fruit (Q89). A mis-disambiguation assigns all salience signals to the wrong entity node, producing a completely erroneous measure of the document's core topics.

The technique of encoding a knowledge graph's structural information into dense, low-dimensional vectors and integrating them into machine learning models. Modern salience scoring often uses graph neural networks (GNNs) to generate entity embeddings that capture not just an entity's attributes but its multi-hop neighborhood. Injecting these context-aware embeddings into a scoring model allows for a far more nuanced assessment of salience than simple frequency-based methods, capturing the latent importance of an entity based on its role in the document's overall semantic structure.