EmbedRank

The core mechanism of EmbedRank calculates the cosine similarity between a document-level embedding and the embedding of each candidate phrase. This allows the algorithm to select phrases that are semantically central to the document's meaning, not just statistically frequent. Unlike graph-based methods like TextRank, it does not require building a co-occurrence network, making it robust to vocabulary variation and synonymy.

A direct extension of the base algorithm, EmbedRank++ applies Maximal Margin Relevance (MMR) to the ranked candidate list. This re-ranking step penalizes new candidates that are too similar to already selected phrases. The result is a diverse set of keyphrases that cover multiple distinct aspects of the document, avoiding the redundancy common in pure similarity-based extraction.

EmbedRank is agnostic to the specific embedding model but typically relies on powerful Sentence-BERT (SBERT) variants. By using Siamese and triplet network structures fine-tuned for semantic textual similarity, the algorithm captures nuanced contextual meaning. This allows it to identify high-quality keyphrases even when they use different wording than the main document body.

The initial phase extracts candidate phrases using Part-of-Speech (POS) tagging patterns, typically selecting sequences of zero or more adjectives followed by one or more nouns. This linguistic filtering ensures that only grammatically well-formed noun phrases are considered, significantly reducing the candidate pool to meaningful concepts before the computationally intensive embedding and similarity scoring step.

A foundational statistical measure for information retrieval. It quantifies a word's importance by multiplying its frequency in a document (Term Frequency) by the logarithm of its rarity across the corpus (Inverse Document Frequency).

• Mechanism: Penalizes common stop words and rewards domain-specific terms.
• Limitation: Purely frequency-based; cannot capture semantic meaning or multi-word phrase boundaries.
• Use Case: Often serves as a strong sparse baseline against which neural models like EmbedRank are compared.

A graph-based ranking algorithm inspired by Google's PageRank. It constructs a network where nodes represent words or phrases, and edges represent co-occurrence within a sliding window.

• Mechanism: Runs iterative voting until convergence; the highest-scoring nodes are selected as keyphrases.
• Key Difference from EmbedRank: Relies on statistical co-occurrence graphs rather than semantic vector similarity.
• Advantage: Fully unsupervised and language-agnostic, requiring no external corpora or pre-trained models.

A method leveraging BERT embeddings to extract keywords. It generates document and candidate phrase embeddings, then ranks candidates by their cosine similarity to the document embedding.

• Similarity to EmbedRank: Both use embedding similarity for ranking.
• Key Difference: KeyBERT typically uses simpler candidate generation (n-grams) and often embeds candidates independently rather than using sentence-level context.
• Variants: Supports Max Sum Similarity and Maximal Marginal Relevance (MMR) to increase diversity among the top-N results.

A re-ranking algorithm crucial for producing a diverse set of keyphrases. It balances a phrase's relevance to the source document against its redundancy with already selected phrases.

• Formula: Iteratively selects candidates that maximize λ * Similarity(Doc, Candidate) - (1-λ) * max(Similarity(Selected, Candidate)).
• Integration: Often applied as a post-processing step to EmbedRank or KeyBERT outputs to prevent near-duplicate phrases.
• Parameter: The λ parameter controls the trade-off between accuracy and diversity.

A lightweight, unsupervised extractor that relies on statistical features from a single document. It requires no external corpora or pre-trained embeddings.

• Features: Computes term weighting based on casing, position, frequency, and context relatedness.
• Contrast with EmbedRank: YAKE is purely statistical and computationally cheaper, while EmbedRank leverages external semantic knowledge from pre-trained models.
• Advantage: Excellent for offline or resource-constrained environments where loading large transformer models is infeasible.

The standard metric for evaluating keyphrase extraction systems. It computes the harmonic mean of precision and recall for the top-K predicted phrases against a gold-standard author list.

• Calculation: F1@K = 2 * (Precision@K * Recall@K) / (Precision@K + Recall@K).
• Benchmarking: EmbedRank and other neural methods are typically evaluated on datasets like KP20k or Inspec using F1@5, F1@10, and F1@15.
• Nuance: Requires exact surface-form matching unless stemming is applied, which can penalize semantically correct but lexically variant predictions.