Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Multilingual Latent Dirichlet Allocation (LDA) Pipeline

This project is for text clustering using the Latent Dirichlet Allocation (LDA) algorithm. It can be adapted to many languages provided that the Snowball stemmer, a dependency of this project, supports it.

Usage

from artifici_lda.lda_service import train_lda_pipeline_default


FR_STOPWORDS = [
    "le", "les", "la", "un", "de", "en",
    "a", "b", "c", "s",
    "est", "sur", "tres", "donc", "sont",
    # even slang/texto stop words:
    "ya", "pis", "yer"]
# Note: this list of stop words is poor and is just as an example.

fr_comments = [
    "Un super-chat marche sur le trottoir",
    "Les super-chats aiment ronronner",
    "Les chats sont ronrons",
    "Un super-chien aboie",
    "Deux super-chiens",
    "Combien de chiens sont en train d'aboyer?"
]

transformed_comments, top_comments, _1_grams, _2_grams = train_lda_pipeline_default(
    fr_comments,
    n_topics=2,
    stopwords=FR_STOPWORDS,
    language='french')

print(transformed_comments)
print(top_comments)
print(_1_grams)
print(_2_grams)

Output:

array([[0.14218195, 0.85781805],
       [0.11032992, 0.88967008],
       [0.16960695, 0.83039305],
       [0.88967041, 0.11032959],
       [0.8578187 , 0.1421813 ],
       [0.83039303, 0.16960697]])

['Un super-chien aboie', 'Les super-chats aiment ronronner']

[[('chiens', 3.4911404011996545), ('super', 2.5000203653313933)],
 [('chats',  3.4911393765493255), ('super', 2.499979634668601 )]]

[[('super chiens', 2.4921035508342464)],
 [('super chats',  2.492102155345991 )]]

How it works

See Multilingual-LDA-Pipeline-Tutorial for an exhaustive example (intended to be read from top to bottom, not skimmed through). For more explanations on the Inverse Lemmatization, see Stemming-words-from-multiple-languages.

Supported Languages

Those languages are supported:

  • Danish
  • Dutch
  • English
  • Finnish
  • French
  • German
  • Hungarian
  • Italian
  • Norwegian
  • Porter
  • Portuguese
  • Romanian
  • Russian
  • Spanish
  • Swedish
  • Turkish

You need to bring your own list of stop words. That could be achieved by computing the Term Frequencies on your corpus (or on a bigger corpus of the same language) and to use some of the most common words as stop words.

Dependencies and their license

numpy==1.14.3           # BSD-3-Clause and BSD-2-Clause BSD-like and Zlib
scikit-learn==0.19.1    # BSD-3-Clause
PyStemmer==1.3.0        # BSD-3-Clause and MIT
snowballstemmer==1.2.1  # BSD-3-Clause and BSD-2-Clause
translitcodec==0.4.0    # MIT License
scipy==1.1.0            # BSD-3-Clause and MIT-like

Unit tests

Run pytest with ./run_tests.sh. Coverage:

----------- coverage: platform linux, python 3.6.7-final-0 -----------
Name                                       Stmts   Miss  Cover
--------------------------------------------------------------
artifici_lda/__init__.py                       0      0   100%
artifici_lda/data_utils.py                    39      0   100%
artifici_lda/lda_service.py                   31      0   100%
artifici_lda/logic/__init__.py                 0      0   100%
artifici_lda/logic/count_vectorizer.py         9      0   100%
artifici_lda/logic/lda.py                     23      7    70%
artifici_lda/logic/letter_splitter.py         36      4    89%
artifici_lda/logic/stemmer.py                 60      3    95%
artifici_lda/logic/stop_words_remover.py      61      5    92%
--------------------------------------------------------------
TOTAL                                        259     19    93%

License

This project is published under the MIT License (MIT).

Copyright (c) 2018 Artifici online services inc.

Coded by Guillaume Chevalier at Neuraxio Inc.

About

A Multilingual Latent Dirichlet Allocation (LDA) Pipeline with Stop Words Removal, n-gram features, and Inverse Stemming, in Python.

Topics

Resources

License

Packages

No packages published