Generative Keyword Extraction
A keyword extraction paradigm based on generative models such as large language models, which formulates keyword generation as a conditional text generation task. The output is not constrained to the vocabulary of the source text and may produce semantically equivalent but more standardized phrases.
Timeline (last 90 days)
The generative KE paradigm does not require keywords to appear verbatim in the source text; instead, the model produces semantically equivalent but more standardized phrases.
Evaluation of generative KE requires more detailed human judgment or semantic similarity measures, and traditional matching-based metrics struggle to measure the semantic correctness of generated keywords.
Generative methods formulate the KE task as conditional text generation, with the model producing keywords autoregressively word by word; the output is not constrained by the source vocabulary, at the cost of losing the fidelity guarantee to the source text.
All Facts (3)
The generative KE paradigm does not require keywords to appear verbatim in the source text; instead, the model produces semantically equivalent but more standardized phrases.
50%UnverifiedEvaluation of generative KE requires more detailed human judgment or semantic similarity measures, and traditional matching-based metrics struggle to measure the semantic correctness of generated keywords.
50%UnverifiedGenerative methods formulate the KE task as conditional text generation, with the model producing keywords autoregressively word by word; the output is not constrained by the source vocabulary, at the cost of losing the fidelity guarantee to the source text.
50%