metadata

license: cc
task_categories:
  - text-generation
  - text2text-generation
language:
  - en
tags:
  - keyphrase-generation
  - domain-adaptation
  - paleontology
  - astrophysics
  - natural-language-processing
size_categories:
  - 1K<n<10K
configs:
  - config_name: paleo
    data_files:
      - split: train
        path:
          - paleo/train.jsonl
      - split: test
        path: paleo/test.jsonl
  - config_name: nlp
    data_files:
      - split: train
        path:
          - nlp/train.jsonl
      - split: test
        path: nlp/test.jsonl
  - config_name: astro
    data_files:
      - split: train
        path:
          - astro/train.jsonl
      - split: test
        path: astro/test.jsonl

`silk` synthetic training samples and human-labeled test sets for domain adaptation in keyphrase generation

This dataset contains the synthetic samples generated by 🧵 silk, a method that leverages citation contexts to create synthetic samples of documents paired with silver-standard keyphrases for adapting keyphrase generation models to new domains. We applied silk on three domains: Natural Language Processing (nlp), Astrophysics (astro) and Paleontology (paleo).

This dataset also includes three human-labeled test sets to assess the performance of keyphrase generation across these domains.

Citation

If you use this dataset, please cite the following paper:

Florian Boudin and Akiko Aizawa.
Unsupervised Domain Adaptation for Keyphrase Generation using Citation Context,
Proceedings of EMNLP 2024 (Findings).

silk synthetic training samples and human-labeled test sets for domain adaptation in keyphrase generation

Citation

`silk` synthetic training samples and human-labeled test sets for domain adaptation in keyphrase generation