kex

Light/easy keyword extraction from documents.


Keywords
keyword-extraction, nlp, information-retrieval, nlp-machine-learning
License
MIT
Install
pip install kex==2.0.4

Documentation

license PyPI version PyPI pyversions PyPI status

KEX

Kex is a python library for unsurpervised keyword extractions, supporting the following features:

Our paper got accepted by EMNLP 2021 main conference 🎉 (camera-ready is here):
This paper has proposed three new algorithms (LexSpec, LexRank, TFIDFRank) and conducted an extensive comparison/analysis over existing keyword extraction algorithms with the proposed methods. Our algorithms are very simple and fast to compute yet established very strong baseline across the dataset (the best MRR/Precision@5 in the average over all the datasets). The TFIDFRank is based on the SingleRank algorithm but with the TFIDF as the population term and the LexSpec and LexRank are based on the lexical specificity where we write a short introduction to lexical specificity here as it is less popular than TFIDF. To reproduce all the results in the paper, please follow these instructions.

Get Started

Install via pip

pip install kex

Extract Keywords with Kex

Built-in algorithms in kex is below:

Basic usage:

>>> import kex
>>> model = kex.SingleRank()  # any algorithm listed above
>>> sample = '''
We propose a novel unsupervised keyphrase extraction approach that filters candidate keywords using outlier detection.
It starts by training word embeddings on the target document to capture semantic regularities among the words. It then
uses the minimum covariance determinant estimator to model the distribution of non-keyphrase word vectors, under the
assumption that these vectors come from the same distribution, indicative of their irrelevance to the semantics
expressed by the dimensions of the learned vector representation. Candidate keyphrases only consist of words that are
detected as outliers of this dominant distribution. Empirical results show that our approach outperforms state
of-the-art and recent unsupervised keyphrase extraction methods.
'''
>>> model.get_keywords(sample, n_keywords=2)
[{'stemmed': 'non-keyphras word vector',
  'pos': 'ADJ NOUN NOUN',
  'raw': ['non-keyphrase word vectors'],
  'offset': [[47, 49]],
  'count': 1,
  'score': 0.06874471825637762,
  'n_source_tokens': 112},
 {'stemmed': 'semant regular word',
  'pos': 'ADJ NOUN NOUN',
  'raw': ['semantic regularities words'],
  'offset': [[28, 32]],
  'count': 1,
  'score': 0.06001468574146248,
  'n_source_tokens': 112}]

Compute a statistical prior

Algorithms such as TF, TFIDF, TFIDFRank, LexSpec, LexRank, TopicalPageRank, and SingleTPR need to compute a prior distribution beforehand by

>>> import kex
>>> model = kex.SingleTPR()
>>> test_sentences = ['documentA', 'documentB', 'documentC']
>>> model.train(test_sentences, export_directory='./tmp')

Priors are cached and can be loaded on the fly as

>>> import kex
>>> model = kex.SingleTPR()
>>> model.load('./tmp')

Supported language

Currently algorithms are available only in English, but soon we will relax the constrain to allow other language to be supported.

Benchmark on 15 Public Datasets

Users can fetch 15 public keyword extraction datasets via kex.get_benchmark_dataset.

>>> import kex
>>> json_line, language = kex.get_benchmark_dataset('Inspec')
>>> json_line[0]
{
    'keywords': ['kind infer', 'type check', 'overload', 'nonstrict pure function program languag', ...],
    'source': 'A static semantics for Haskell\nThis paper gives a static semantics for Haskell 98, a non-strict ...',
    'id': '1053.txt'
}

Please take a look an example script to run a benchmark on those datasets.

Implement Custom Extractor with Kex

We provide an API to run a basic pipeline for preprocessing, by which one can implement a custom keyword extractor.

import kex

class CustomExtractor:
    """ Custom keyword extractor example: First N keywords extractor """

    def __init__(self, maximum_word_number: int = 3):
        """ First N keywords extractor """
        self.phrase_constructor = kex.PhraseConstructor(maximum_word_number=maximum_word_number)

    def get_keywords(self, document: str, n_keywords: int = 10):
        """ Get keywords

         Parameter
        ------------------
        document: str
        n_keywords: int

         Return
        ------------------
        a list of dictionary consisting of 'stemmed', 'pos', 'raw', 'offset', 'count'.
        eg) {'stemmed': 'grid comput', 'pos': 'ADJ NOUN', 'raw': ['grid computing'], 'offset': [[11, 12]], 'count': 1}
        """
        phrase_instance, stemmed_tokens = self.phrase_constructor.tokenize_and_stem_and_phrase(document)
        sorted_phrases = sorted(phrase_instance.values(), key=lambda x: x['offset'][0][0])
        return sorted_phrases[:min(len(sorted_phrases), n_keywords)]

Reference paper

If you use any of these resources, please cite the following paper:

@inproceedings{ushio-etal-2021-kex,
    title={{B}ack to the {B}asics: {A} {Q}uantitative {A}nalysis of {S}tatistical and {G}raph-{B}ased {T}erm {W}eighting {S}chemes for {K}eyword {E}xtraction},
    author={Ushio, Asahi and Liberatore, Federico and Camacho-Collados, Jose},
        booktitle={Proceedings of the {EMNLP} 2021 Main Conference},
    year = {2021},
    publisher={Association for Computational Linguistics}
}