LugolBis/KeywordsExtraction

French Keyword Extraction

Dataset Description

This dataset is a collection of French phrase–keyword pairs designed for training and evaluating keyword extraction models, prompt engineering, or query expansion systems. It is composed of three distinct subsets (configurations), each reflecting a specific linguistic register and use case.

All inputs are in French and contain between 1 and 5 sentences, offering varied contextual lengths. The dataset was generate with Claude Sonnet 5 (High / Medium).

Supported Configurations

Config Name	Description	Input Length	Keyword Separator
prompt	Prompts addressed to AI assistants (interrogative/imperative forms).	1 to 5 sentences	", " (comma + space)
sentences	Affirmative or declarative statements, not necessarily questions.	1 to 5 sentences	", " (comma + space)
searching	Internet search queries, often containing significant background context.	1 to 4 sentences (dense context)	" " (simple space)

---

Dataset Structure

Data Fields

All configurations share the same two-column CSV structure:

• Input (string): The original French text (prompt, sentence, or search query).
• Output (string): The list of extracted keywords.

Important: Pay attention to the keyword separator for each config:

• prompt & sentences → keywords are separated by a comma and a space (e.g., "intelligence artificielle, éthique, régulation").
• searching → keywords are separated by a simple space (e.g., "meilleur restaurant paris 2024").

Data Splits

Currently, this dataset is provided as a single split (train) per configuration. If you wish to create train/validation splits, we recommend doing so locally using datasets or sklearn.model_selection.

---

Usage Example (Python)

Load a specific configuration using the 🤗 datasets library:

from datasets import load_dataset

# Load the "prompt" subset
dataset_prompt = load_dataset("LugolBis/KeywordsExtraction", "prompt", split="train")

# Load the "sentences" subset
dataset_sentences = load_dataset("LugolBis/KeywordsExtraction", "sentences", split="train")

# Load the "searching" subset
dataset_searching = load_dataset("LugolBis/KeywordsExtraction", "searching", split="train")

# Explore an example
print(dataset_prompt[0])
# Output example:
# {'Input': 'Quels sont les impacts du réchauffement climatique sur la biodiversité marine ?', 
#  'Output': 'réchauffement climatique, biodiversité marine, impacts'}

---

Dataset Creation

Curation Rationale

The dataset was built to cover three distinct modalities of natural language queries:

1. Prompts: Capturing the instructional/inquisitive tone used when interacting with Large Language Models.
2. Sentences: Covering descriptive and factual statements to extract core concepts.
3. Searching: Mimicking real-world search engine behaviors where context is key, and keywords often act as boolean/concatenated phrases (hence the space separator).

Source Data

All the Input sentences (prompts, qearch queries, etc.) were generated with an LLM : Claude Sonnet 5 (High / Medium).

Annotations

All the Output keywords extracted from the sentences were generated with an LLM : Claude Sonnet 5 (High / Medium).

---

Considerations for Using the Data

Social Impact & Limitations

• The dataset is exclusively in French, making it suitable for Francophone NLP applications but limited for multilingual models.
• The keyword extraction logic might carry inherent biases depending on the annotator or the LLM used for generation.
• The searching subset uses space-separated keywords, which might represent query expansion tokens rather than strict semantic concepts.

Recommended Use Cases

• Fine-tuning small/medium LLMs for French keyword generation.
• Training embedding models for document retrieval.
• Evaluating prompt engineering techniques for summarization.
• Building query suggestion systems for search engines.

Citation

If you use this dataset in your research, please cite it as follows :

@misc{your_name_2024_french_keywords,
  author = {LugolBis},
  title = {French Keyword Extraction},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/LugolBis/KeywordsExtraction}}
}