Supervised Keyphrase Extraction

The core mechanism treats each candidate phrase as an independent data point to be classified as a keyphrase or non-keyphrase. A model is trained on features extracted from labeled documents to learn a decision boundary.

• Input: A candidate phrase and its surrounding context.
• Output: A probability score indicating keyphraseness.
• Training Data: Requires documents where human annotators have tagged the ground-truth keyphrases.
• Common Algorithms: Naive Bayes, Support Vector Machines (SVM), and Random Forests were standard before deep learning dominance.

Instead of classifying isolated candidates, this method classifies every token in the source document using a labeling scheme, typically BIO (Beginning, Inside, Outside) or BILOU tags. This transforms extraction into a token-level prediction problem.

• B-Tag: Marks the token that begins a keyphrase.
• I-Tag: Marks tokens inside a multi-word keyphrase.
• O-Tag: Marks tokens that are not part of any keyphrase.
• Advantage: Naturally handles keyphrase boundary detection without a separate candidate generation step.

Before deep learning, the success of supervised models depended heavily on hand-crafted features that capture phraseness and informativeness. These features numerically encode linguistic and statistical observations.

• Term Frequency-Inverse Document Frequency (TF-IDF): Measures corpus-level specificity.
• Positional Features: The first occurrence position of a phrase, as keyphrases often appear early.
• Part-of-Speech (POS) Patterns: Filters for noun phrases matching specific syntactic sequences.
• Length & Capitalization: Metrics that correlate with proper nouns and technical terminology.

Modern supervised systems use neural networks to bypass manual feature engineering. BiLSTM-CRF architectures were the state-of-the-art for sequence labeling, using a Conditional Random Field (CRF) layer to enforce valid tag transitions.

• BiLSTM: Captures bidirectional context for each token.
• CRF Layer: Ensures an I-tag cannot follow an O-tag without a B-tag.
• BERT/RoBERTa: Fine-tuned transformer encoders now dominate, using the contextualized embedding of each token to predict its BIO tag with a linear classification head.

The primary constraint of supervised extraction is the absolute requirement for high-quality, domain-specific training data. Model performance is directly correlated with the size and consistency of the annotated corpus.

• KP20k: A standard benchmark containing 20,000 scientific abstracts with author-assigned keyphrases.
• Domain Shift: A model trained on computer science abstracts will perform poorly on medical case reports without fine-tuning.
• Annotation Cost: Creating ground-truth labels requires domain experts and strict inter-annotator agreement guidelines.

Strictly defined, supervised extraction is limited to present keyphrases—phrases that appear verbatim in the text. Sequence labeling models cannot generate tokens that do not exist in the input.

• Present Extraction: The model highlights contiguous spans of text.
• Absent Limitation: To generate abstractive keyphrases not found in the text, one must switch to a Keyphrase Generation model (a sequence-to-sequence task).
• Hybrid Systems: Some architectures combine extraction for present phrases and generation for absent ones.

The foundational supervised approach where each candidate phrase is treated as an independent data point. A model is trained to output a probability score indicating whether the candidate is a valid keyphrase.

• Features: Models use lexical features (TF-IDF, position), semantic features (embeddings), and statistical features.
• Training Data: Requires a corpus of documents annotated with positive (keyphrase) and negative (non-keyphrase) examples.
• Algorithms: Common classifiers include Support Vector Machines (SVM), Random Forests, and Gradient Boosting Machines.

Frames extraction as a token-level classification task, typically using BIO (Begin, Inside, Outside) or BILOU tagging schemes. This method excels at identifying multi-word keyphrases with precise boundaries.

• Architecture: BiLSTM-CRF and Transformer-based models (like BERT) are standard.
• Advantage: Naturally handles keyphrase boundary detection without needing a separate candidate generation step.
• Output: Each token receives a label, and contiguous spans of 'B' and 'I' tokens form the final keyphrases.

Supervised models are benchmarked against a gold-standard set of author or expert-assigned keyphrases using strict or lenient matching.

• F1@K: The harmonic mean of precision and recall calculated only on the top-K predicted keyphrases. Common values for K are 5, 10, and 15.
• Mean Reciprocal Rank (MRR): Evaluates the ranking quality by averaging the reciprocal of the rank of the first correctly predicted keyphrase.
• Exact vs. Partial Match: Exact match requires a character-level match, while partial match allows for stemming or sub-phrase overlap.

After candidate generation, supervised models assign a relevance score to each phrase. The final step is ranking candidates by this score and selecting the top-N.

• Scoring Features: A combination of phraseness (linguistic quality) and informativeness (topical relevance) features.
• Re-ranking: Techniques like Maximal Marginal Relevance (MMR) can be applied post-hoc to the ranked list to increase diversity and reduce redundancy among the final keyphrase set.

Standard datasets are crucial for training and comparing supervised models. They provide the necessary annotated keyphrases.

• KP20k: The largest standard benchmark, containing over 20,000 scientific article abstracts from computer science with author-assigned keyphrases.
• Inspec: A smaller dataset of 2,000 abstracts from scientific journals, often used for domain-specific evaluation.
• SemEval: A series of shared tasks that have provided datasets for keyphrase extraction, including the SemEval-2010 Task 5 dataset.

Supervised models must be explicitly designed to handle two distinct types of keyphrases.

• Present Keyphrases: Phrases that appear verbatim in the source text. Sequence labeling models are naturally suited for this task.
• Absent Keyphrases: Relevant phrases that do not appear in the text. This is a generative task, often handled by Keyphrase Generation models using sequence-to-sequence architectures, which are a separate but related supervised paradigm.