Informativeness

Informativeness evaluates a phrase's ability to represent the unique topical domain of a document rather than common language. It penalizes generic terms that appear frequently across many documents.

• Uses TF-ICF (Term Frequency-Inverse Corpus Frequency) to measure domain specificity
• A phrase like "machine learning" scores high in an ML paper but low in a general news corpus
• Contrasts with phraseness, which measures linguistic well-formedness independently of topical relevance
• Often combined with TF-IDF to balance local importance against corpus-wide frequency

Informativeness scores are typically derived from statistical distributions of terms within and across documents, using frequency-based heuristics to identify salient phrases.

• TF-IDF: Weights terms by their frequency in a document divided by their prevalence across the corpus
• TF-ICF: Adapts TF-IDF by substituting inverse corpus frequency for inverse document frequency
• YAKE uses five statistical features including term frequency, term relatedness to context, and sentence position
• RAKE computes informativeness through word degree and frequency in a co-occurrence graph

Modern informativeness scoring leverages dense vector embeddings to measure how semantically similar a candidate phrase is to the overall document representation.

• KeyBERT computes cosine similarity between candidate phrase embeddings and the document embedding
• EmbedRank uses sentence embeddings to rank phrases by their semantic alignment with the document
• This approach captures conceptual relevance beyond exact keyword matching
• Effective for identifying keyphrases that capture the document's core theme even with varied vocabulary

Informativeness scoring often incorporates the positional distribution of candidate phrases within a document, based on the assumption that key concepts appear in structurally significant locations.

• Phrases appearing in titles, abstracts, and section headings receive higher weights
• First-occurrence position: Earlier mentions indicate greater topical importance
• Sentence-level position: Terms in topic sentences or concluding sentences are weighted more heavily
• Combined with frequency metrics to prevent over-weighting of incidental early mentions

Informativeness and phraseness are complementary scoring dimensions in keyphrase extraction. Informativeness measures topical relevance, while phraseness evaluates linguistic well-formedness.

• Phraseness: Is the candidate a valid multi-word expression? (e.g., "artificial intelligence" vs. "intelligence artificial")
• Informativeness: Does the candidate capture the document's subject matter? (e.g., "neural network" vs. "recent years")
• Both scores are combined in systems like YAKE and RAKE for final candidate ranking
• A high-phraseness, low-informativeness phrase is linguistically valid but topically irrelevant

Informativeness is inherently corpus-relative—a phrase's score depends on the background document collection used for comparison. Proper calibration requires a representative corpus.

• Background corpus selection critically impacts IDF and ICF calculations
• Domain-specific corpora (e.g., PubMed for biomedical text) yield more accurate informativeness scores
• KP20k and other benchmark datasets provide standardized corpora for evaluation
• Mismatched background corpora can inflate scores for common domain terms or suppress rare but relevant phrases

A scoring component that measures how linguistically well-formed a candidate sequence is as a phrase, independent of its topical relevance. It penalizes incomplete or ungrammatical n-grams.

• Evaluates syntactic validity using Part-of-Speech (POS) tagging
• Filters candidates matching patterns like (JJ)*(NN)+ for noun phrases
• Ensures extracted units are valid standalone concepts, not fragments

A statistical measure evaluating the importance of a word to a document within a corpus. It balances Term Frequency (TF)—how often a word appears—against Inverse Document Frequency (IDF)—how rare it is across all documents.

• High weight assigned to terms frequent in a specific document but rare in the general collection
• Foundational baseline for sparse retrieval and candidate scoring
• Often used as a feature within composite informativeness scores

A weighting scheme adapting TF-IDF by replacing Inverse Document Frequency with Inverse Corpus Frequency (ICF). It measures domain specificity by comparing term frequency in a target corpus against a general reference corpus.

• Identifies terms that are characteristic of a specific domain
• Useful when distinguishing domain jargon from general language
• Directly quantifies the 'domain specificity' aspect of informativeness

The process of assigning a numerical weight to each candidate phrase based on features like frequency, position, and semantic similarity. Informativeness is a primary feature in this calculation.

• Combines multiple signals: phraseness, informativeness, and positional weight
• Final ranking determines which keyphrases represent the document
• Scoring functions range from linear combinations to learned models

A measure of the prominence or importance of a named entity within a document. It is often used to filter or weight entity-based keyphrases based on their centrality to the core narrative.

• Calculated using graph centrality in co-occurrence networks
• High salience entities are typically more informative about the document's subject
• Complements statistical informativeness with structural importance

A re-ranking algorithm that balances a phrase's relevance to the document against its redundancy with already selected keyphrases. It ensures the final set is both informative and diverse.

• Penalizes candidates too similar to higher-ranked selections
• Uses cosine similarity to measure redundancy
• Prevents keyphrase lists from being dominated by a single topic cluster