polyglot3

polyglot3 is a maintained, Python 3 only continuation of polyglot, a natural language pipeline for massively multilingual applications. It keeps polyglot's API and models under the package name polyglot3, and removes the native dependencies that made polyglot hard to install. It is maintained by Dcipher Analytics AB.

• Free software: GPLv3 license
• Source: https://github.com/dcipheranalytics/polyglot3
• Install: pip install polyglot3 (Python 3.9+, no system libraries)

Features

• Tokenization: word and sentence segmentation for all languages
• Language detection: 176 languages
• Named entity recognition: 25 languages
• Part of speech tagging: 7 languages
• Sentiment analysis: 136 languages
• Word embeddings: 25 languages

See Supported languages for the exact lists.

What changed from polyglot

• New name. Import polyglot3 instead of polyglot (from polyglot3.text import Text) and run the polyglot3 command. The API is otherwise the same, models and ~/polyglot_data are shared, and polyglot3 can be installed next to the original package.
• No PyICU. Word and sentence segmentation use a pure-Python port of ICU's break iterators, generated from ICU 78.3's own rules, including the Thai dictionary engine. Outside Chinese, Japanese, Korean, Lao, Khmer and Burmese (which ICU segments with dictionaries that are not bundled), the boundaries match ICU's exactly, so the existing NER and POS models see the tokens they were trained on.
• No pycld2. Language detection uses fastText's lid.176 model via fasttext-langdetect.
• Python 3.9 to 3.13 only; Python 2 support and six are gone.
• New model host. Models are downloaded from a Hugging Face dataset and verified against checksums. The original Stony Brook server is offline.
• Removed: morphological analysis and transliteration (their models are no longer available), and the TextBlob/NLTK leftovers that never worked.
• Fixed: tokens after emoji and other non-BMP characters were garbled; the sentiment subcommand did nothing; several downloader bugs.

Quick tutorial

Download the models for the languages you need (once):

polyglot3 download LANG:tr LANG:cs

from polyglot3.text import Text, Word

Language detection

text = Text("Bugün hava çok güzel, parka gidelim.")
print(text.language.code, text.language.name)

tr Turkish

Tokenization

zen = Text(
    "Beautiful is better than ugly. "
    "Explicit is better than implicit. "
    "Simple is better than complex."
)
print(zen.words)
print(zen.sentences)

['Beautiful', 'is', 'better', 'than', 'ugly', '.', 'Explicit', 'is', 'better', 'than', 'implicit', '.', 'Simple', 'is', 'better', 'than', 'complex', '.']
[Sentence("Beautiful is better than ugly."), Sentence("Explicit is better than implicit."), Sentence("Simple is better than complex.")]

Part of speech tagging

text = Text("Petr Novák pracuje v Praze pro společnost Škoda.")
for word, tag in text.pos_tags:
    print("{:<16}{}".format(word, tag))

Petr            PROPN
Novák           PROPN
pracuje         ADJ
v               ADP
Praze           NOUN
pro             ADP
společnost      NOUN
Škoda           PROPN
.               PUNCT

Named entity recognition

text = Text(
    "Mustafa Kemal Atatürk, 1923 yılında Ankara'da Türkiye Cumhuriyeti'ni ilan etti."
)
print(text.entities)

[I-PER(['Mustafa', 'Kemal', 'Atatürk']), I-LOC(['Türkiye'])]

Polarity

text = Text("Bu film harika ama son kısmı berbat.")
for w in text.words:
    print("{:<16}{:>2}".format(w, w.polarity))

Bu               0
film             0
harika           1
ama             -1
son              0
kısmı            0
berbat          -1
.                0

Embeddings

from polyglot3.load import load_embeddings

embeddings = load_embeddings("tr")
print(embeddings.nearest_neighbors("Ankara"))

['İstanbul', 'İzmir', 'Bursa', 'Antalya', 'Adana', 'Trabzon', 'Balıkesir', 'Konya', ...]

Supported languages

Credits and citation

polyglot was created by Rami Al-Rfou with contributions from Yingtao Tian and others at the Data Science Lab, Stony Brook University. The models come from their research. If you use them, please cite the papers:

• Word embeddings and part of speech tagging: Polyglot: Distributed Word Representations for Multilingual NLP (CoNLL 2013)
• Named entity recognition: POLYGLOT-NER: Massive Multilingual Named Entity Recognition (SDM 2015)
• Sentiment: Building sentiment lexicons for all major languages (ACL 2014)

@InProceedings{polyglot:2013:ACL-CoNLL,
  author    = {Al-Rfou, Rami and Perozzi, Bryan and Skiena, Steven},
  title     = {Polyglot: Distributed Word Representations for Multilingual NLP},
  booktitle = {Proceedings of the Seventeenth Conference on Computational Natural Language Learning},
  month     = {August},
  year      = {2013},
  address   = {Sofia, Bulgaria},
  publisher = {Association for Computational Linguistics},
  pages     = {183--192},
  url       = {https://aclanthology.org/W13-3520}
}

@inproceedings{polyglotner,
  author    = {Al-Rfou, Rami and Kulkarni, Vivek and Perozzi, Bryan and Skiena, Steven},
  title     = {{Polyglot-NER}: Massive Multilingual Named Entity Recognition},
  booktitle = {Proceedings of the 2015 {SIAM} International Conference on Data Mining},
  month     = {April},
  year      = {2015},
  publisher = {SIAM}
}

@inproceedings{chen2014building,
  title     = {Building sentiment lexicons for all major languages},
  author    = {Chen, Yanqing and Skiena, Steven},
  booktitle = {Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Short Papers)},
  pages     = {383--389},
  year      = {2014}
}

Segmentation is derived from ICU (Unicode License V3) and language detection uses fastText's lid.176 model.

polyglot3 is maintained by Dcipher Analytics AB (info@dcipheranalytics.com). This repository has been updated by the Dcipher Analytics dev team with help from Claude Opus. Many thanks to the original polyglot authors for their work.