CSKE 0.2.0

Efficient and powerful class-specific keyword extraction.

CSKE: Class-Specific Keyword Extraction

CSKE is a high-performance Python library designed for iterative, class-specific keyword extraction. Unlike generic extractors, CSKE maintains coherence to a user-defined class, and leverages clustering techniques to ensure that expanded keyword sets remain semantically anchored to the target class.

This code was introduced in the KONVENS 2024 paper titled: An Improved Method for Class-specific Keyword Extraction: A Case Study in the German Business Registry

Key Features

• Iterative Expansion: Automatically discovers new keywords by walking through your dataset starting from a small seed list (defined by you!).
• Drift Prevention: Uses clustering and filtering to weed out out "semantic drift" keywords that do not fit will to your defined domain.
• Weighted Extraction: Balance the influence between the local document context and the global seed keywords.
• Hardware Accelerated: Built on top of PyTorch and Sentence Transformers with automatic support for CUDA.

Installation

pip install cske

Usage

import pandas as pd
from cske import CSKE

df = pd.DataFrame({
    "text_content": [
        "Neural networks are a subset of machine learning.",
        "The transformer architecture revolutionized NLP.",
        "Deep learning models require significant GPU resources.".
        "..."
    ]
})

# 2. Initialize the extractor, using any sentence transformer model, i.e., from Hugging Face
extractor = CSKE(embedding_model="all-MiniLM-L6-v2")

# 3. Run the pipeline
keywords = extractor.keyword_pipeline(
    starting_seed=["machine learning", "neural networks"],
    df=df,
    df_col_to_extract="text_content",
    n_iterations=3, # number of iterations (partitions of your data)
    number_newseed=2, # maximum number of "new" seeds per iteration
    do_filter=True  # whether to perform filtering to prevent drift
)

print(f"Expanded Keyword Set: {keywords}")

Key Parameters, and what they mean

Parameter	Default	Description
n_iterations	5	How many rounds of expansion to perform.
seed_weight	1.0	Importance given to the original seed keywords.
doc_weight	0.0	Importance given to document context.
do_filter	True	Whether or not to apply filtering.
topk	None	If set, limits the final output to the top-k keywords.

---

Citation

If you find CSKE useful or utilize it for your work, please considering citing:

@inproceedings{meisenbacher-etal-2024-improved,
    title = "An Improved Method for Class-specific Keyword Extraction: A Case Study in the {G}erman Business Registry",
    author = "Meisenbacher, Stephen  and
      Schopf, Tim  and
      Yan, Weixin  and
      Holl, Patrick  and
      Matthes, Florian",
    editor = "Luz de Araujo, Pedro Henrique  and
      Baumann, Andreas  and
      Gromann, Dagmar  and
      Krenn, Brigitte  and
      Roth, Benjamin  and
      Wiegand, Michael",
    booktitle = "Proceedings of the 20th Conference on Natural Language Processing (KONVENS 2024)",
    month = sep,
    year = "2024",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.konvens-main.18/",
    pages = "159--165"
}

This work makes use of the KeyBERT library, so please also consider citing it: KeyBERT