Co-occurrence Graph

Edges represent a symmetric relationship between two terms, with weights quantifying the strength of their association. The weight is typically the raw count or a normalized metric of how often the two words appear together within a defined sliding window. A higher weight implies a stronger semantic or topical bond, making these edges critical for graph-based ranking algorithms like TextRank.

The graph is built by scanning text with a fixed-size window (e.g., 2-10 words). An edge is created or its weight incremented every time two words fall within this window. This method captures local syntactic and semantic relationships without requiring deep parsing. The window size is a crucial hyperparameter: smaller windows capture fixed phrases, while larger windows capture broader thematic connections.

In algorithms like RAKE, the graph is built specifically from content words, using stopwords as natural boundary markers. Candidate keyphrases are sequences of contiguous words that are not separated by a stopword. The co-occurrence graph then links these content words, and the final score of a candidate phrase is derived from the sum of the degrees or weighted degrees of its constituent words.

Once constructed, the graph is used to score word importance via centrality algorithms like PageRank or Degree Centrality. A node's score is determined by the number and weight of its connections. In TextRank, a word is important if it is linked to many other important words. The final keyphrase score is often the sum of its constituent word scores, identifying the most salient phrases in the document.

A co-occurrence graph is a sparse representation of text, capturing explicit, first-order statistical relationships. This contrasts with dense vector embeddings (like BERT), which capture latent semantic similarity. While embeddings excel at synonymy, co-occurrence graphs provide transparent, easily interpretable evidence for why a phrase was selected, based purely on explicit frequency and proximity data.

Since the graph relies solely on token co-occurrence statistics and stopword lists, it is inherently language-agnostic. It requires no pre-trained models, linguistic parsers, or external corpora. This makes it highly effective for domain-specific jargon (e.g., medical or legal text) where pre-trained NLP models may fail to recognize novel compound terms, as it builds a custom statistical model for each input document.

A family of algorithms that represent text units as vertices and their semantic or co-occurrence relationships as edges. Salience is determined by structural centrality measures rather than purely statistical frequency.

• PageRank adaptation: Runs on lexical networks to find authoritative nodes
• Vertex centrality: Measures like degree, closeness, and eigenvector centrality identify key terms
• Iterative convergence: Scores propagate through the graph until a stable ranking is achieved

A graph-based ranking algorithm that builds a word or phrase co-occurrence network from a document and applies PageRank to identify the most salient keyphrases. It is fully unsupervised and domain-independent.

• Window size: Typically uses a co-occurrence window of 2-10 words to establish edges
• Weighted edges: Edge weights reflect co-occurrence frequency within the sliding window
• Post-processing: Top-ranked words are collapsed into multi-word keyphrases if they appear adjacently in text

An unsupervised, domain-independent algorithm that extracts keyphrases by analyzing word co-occurrence within stopword-delimited sequences. It does not require a pre-built graph but constructs an implicit co-occurrence matrix.

• Candidate generation: Splits text on stopwords and punctuation to form contiguous phrase candidates
• Scoring: Each word is scored by its degree (co-occurrence count) divided by its frequency
• Phrase score: Candidate keyphrase score is the sum of its constituent word scores

An evaluation metric computing the harmonic mean of precision and recall for the top-K predicted keyphrases against a gold-standard set. It is the primary benchmark for comparing keyphrase extraction systems.

• Precision@K: Proportion of top-K predictions that are correct
• Recall@K: Proportion of all gold-standard keyphrases captured in the top-K
• Stemming matching: Typically uses Porter stemmer to match morphological variants between predictions and ground truth

The process of assigning a numerical weight to each candidate phrase extracted from a co-occurrence graph or statistical model. Scores determine the final ranked output of keyphrases.

• Frequency-based: Term frequency and TF-IDF variants provide baseline salience signals
• Position-based: Phrases appearing early in a document or in section headers receive higher weights
• Graph centrality: Node degree, betweenness, and PageRank scores derived from the co-occurrence structure

Two orthogonal scoring dimensions used to evaluate candidate keyphrases. Phraseness measures linguistic well-formedness, while informativeness measures topical relevance.

• Phraseness: Assessed via POS patterns (e.g., adjective-noun sequences) and n-gram probability
• Informativeness: Quantified by TF-IDF, domain specificity (TF-ICF), or semantic similarity to the document centroid
• Combined scoring: Final score often multiplies or linearly combines both dimensions for balanced extraction