taln-ls2n/taln-archives

TALN-Archives是一个用于关键词提取和生成模型基准测试的数据集。该数据集包含1207篇法文科学论文的摘要，关键词由作者在非受控环境下标注。部分文档（456篇完全翻译和719篇部分翻译）提供了英文翻译，支持跨语言/多语言关键词生成实验。数据集还采用了PRMU方案对参考关键词进行分类，并提供了文本预处理和词干提取的详细信息。

原始信息汇总

TALN-Archives Benchmark Dataset for Keyphrase Generation

Overview

• Languages: French (fr), English (en)
• License: CC-BY-4.0
• Multilinguality: Multilingual
• Task Categories: Text-mining, Text-generation
• Task IDs: Keyphrase-generation, Keyphrase-extraction
• Size: 1K<n<10K
• Pretty Name: TALN-Archives

Dataset Details

Content

• Composition: 1207 abstracts of scientific papers in French.
• Annotations: Keyphrases annotated by authors in an uncontrolled setting.
• Translations: English translations available for 456 fully- and 719 partially-translated documents.

Structure

• Data Fields:
  • id: Unique identifier of the document.
  • title: Title of the document.
  • abstract: Abstract of the document.
  • keyphrases: List of reference keyphrases.
  • prmu: List of Present-Reordered-Mixed-Unseen categories for reference keyphrases.
  • translation: Translations of title, abstract, and keyphrases in English if available.

Statistics

• Test Split:

  Split	# documents	#words	# keyphrases	% Present	% Reordered	% Mixed	% Unseen
  Test	1207	138.3	4.12	53.83	12.32	21.69	12.16

Processing

• Text Pre-processing: Tokenization using spacy (fr_core_news_sm model) with a rule to keep words with hyphens as one token.
• Stemming: Applied using Snowball stemmer for French in nltk before matching reference keyphrases against the source text.

References

• Boudin, 2013: [TALN Archives : a digital archive of French research articles in Natural Language Processing].
• Boudin and Gallina, 2021: [Redefining Absent Keyphrases and their Effect on Retrieval Effectiveness].