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):
In our paper, we conducted an extensive comparison and analysis over existing keyword extraction algorithms and proposed new algorithms LexRank
and LexSpec
that
achieve very competitive baseline with very low complexity. Our proposed algorithms are based on the lexical specificity and we write a short introduction to the
lexical specificity here.
To reproduce all the result in the paper, please follow these instructions.
Get Started
Install via pip
pip install kex
Built-in algorithms in kex is below:
Basic usage:
>>> import kex
>>> model = kex.SingleRank()
>>> 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.
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}
}