Absent Keyphrase Extraction

Unlike present extraction, this task uses encoder-decoder architectures (e.g., T5, BART) to generate keyphrases token by token. The encoder reads the source document into a latent representation, while the decoder autoregressively produces phrases that capture the document's core themes—even if those exact words never appear in the text.

• One2Set and One2One are common training paradigms
• Models learn to abstract beyond surface-level n-grams
• Requires paired training data: (document, keyphrase set)

The defining feature is the ability to synthesize concepts not lexically present. For example, a paper describing 'gradient descent optimization with adaptive step sizes' might generate the absent keyphrase 'Adam Optimizer'—a concept the model infers from the description without seeing the term.

• Requires world knowledge encoded during pre-training
• Bridges the gap between descriptive text and canonical terminology
• Critical for tagging documents with standardized subject headings

Advanced models employ hybrid pointer-generator networks that learn when to copy a token from the source (present keyphrases) and when to generate a novel token from the vocabulary (absent keyphrases). This is governed by a dynamic switching probability at each decoding step.

• Copy mechanism: Handles verbatim multi-word expressions
• Generate mechanism: Produces novel, abstracted terminology
• Balances extraction fidelity with generative flexibility

Evaluating absent keyphrases is inherently harder than present extraction. A generated phrase like 'Neural Machine Translation' may be perfectly correct but absent from the gold-standard list, leading to false negatives. This motivates the use of soft evaluation metrics and semantic matching.

• F1@K penalizes valid absent predictions if not in reference set
• BERTScore and embedding-based metrics offer softer comparisons
• Human evaluation remains the gold standard for generative quality

Models require large corpora where documents are paired with both present and absent keyphrases. KP20k and OpenKP are standard benchmarks, typically sourced from scientific abstracts where author-assigned keywords frequently include terms not in the abstract body.

• Author keywords serve as weak supervision for absent phrases
• Data augmentation can artificially mask present phrases to force generation
• Domain adaptation is critical for specialized vocabularies

To prevent hallucinated or irrelevant keyphrases, modern systems incorporate constrained decoding techniques. These restrict the output vocabulary to valid keyphrase candidates from a knowledge base or enforce structural constraints like phrase length and part-of-speech patterns during beam search.

• Prefix-constrained generation limits outputs to known concepts
• Coverage mechanisms prevent repetitive or redundant predictions
• Ensures generated phrases are well-formed and domain-appropriate