RAKE

RAKE segments text into candidate phrases using stopwords (e.g., 'the', 'and', 'of') as boundaries. The algorithm splits the document at each stopword or punctuation delimiter, treating the resulting contiguous word sequences as candidate keyphrases. This approach is computationally efficient and language-agnostic, requiring only a stopword list rather than a trained part-of-speech tagger. For example, in the sentence 'The extraction of keyphrases from documents,' the candidates become 'extraction,' 'keyphrases,' and 'documents.'

RAKE constructs a word co-occurrence graph where vertices represent distinct words and edges connect words that appear together within candidate phrases. Each word receives a score based on its degree (number of co-occurring words) and frequency (total occurrences). The word score is calculated as deg(w) / freq(w), rewarding words that appear in diverse contexts while penalizing common words. Candidate phrase scores are the sum of their constituent word scores, effectively identifying phrases with high semantic density.

Unlike supervised methods that require labeled training data, RAKE operates without any prior knowledge of the document's domain. It relies solely on statistical properties of the input text, making it immediately applicable to new domains without retraining. This domain independence is achieved by using generic stopword lists and frequency-based scoring rather than domain-specific features or external knowledge bases. The algorithm performs consistently across scientific articles, news reports, and technical documentation without parameter tuning.

RAKE achieves O(n) time complexity relative to the number of words in the document, making it suitable for real-time and large-scale applications. The algorithm requires only a single pass through the text for phrase segmentation and a second pass for scoring. Memory overhead is minimal, as the co-occurrence graph stores only unique words rather than full n-gram combinations. This efficiency enables RAKE to process thousands of documents per second on commodity hardware, contrasting with embedding-based methods that require GPU acceleration.

RAKE includes a mechanism for handling interior stopwords within multi-word phrases. After initial scoring, the algorithm identifies candidate phrases that appear adjacent to each other separated only by a stopword and evaluates whether combining them produces a higher-scoring phrase. For example, 'bag of words' may initially split into 'bag' and 'words,' but RAKE can recognize the combined form as a valid keyphrase if it appears frequently. This step recovers common technical phrases that contain internal stopwords.

RAKE has been evaluated against human-assigned keyphrases on standard datasets, demonstrating strong precision for present keyphrase extraction. In the original paper, RAKE achieved precision scores comparable to supervised systems when extracting author-assigned keywords from technical abstracts. The algorithm performs particularly well on documents with clear topical focus and distinct vocabulary, though it may underperform on highly ambiguous or conversational text where stopword patterns are less reliable indicators of phrase boundaries.