Available Models & Languages

Table of contents

Stanza provides pretrained NLP models for a total 66 human languages. On this page we provide detailed information on these models.

Pretrained models in Stanza can be divided into two categories, based on the datasets they were trained on:

1. Universal Dependencies (UD) models, which are trained on the UD treebanks, and cover functionalities including tokenization, multi-word token (MWT) expansion, lemmatization, part-of-speech (POS) and morphological features tagging and dependency parsing;
2. NER models, which support named entity tagging for 8 languages, and are trained on various NER datasets.

Available UD Models

The following table lists all UD models supported by Stanza and pretrained on the Universal Dependencies v2.8 datasets. You can find more information about the POS tags, morphological features, and syntactic relations used on the Universal Dependencies website. We recommend you always use the lastest released models. However, you can still use these earlier models by downloading them and putting them in the correct directory. You can find performance of all available models on the System Performance page.

Table Notes

1. marks models which have very low unlabeled attachment score (UAS) when evaluated end-to-end (from tokenization all the way to dependency parsing). Specifically, their UAS is lower than 50% on the Universal Dependencies 2.8 test set. The output of these models is unlikely to be accurate enough for serious syntactic analysis. In some cases, these models are resource constrained. If you are interested in improving the accuracy of these languages, please get in contact with the treebank authors regarding expanding the training data.
2. marks the default package for a language, which is the package trained on the largest treebank available for that language.
3. The copyright and licensing status of machine learning models is not very clear (to us). We list in the table below the Treebank License of the underlying data from which each language pack (set of machine learning models for a treebank) was trained. To the extent that The Trustees of Leland Stanford Junior University have ownership and rights over these language packs, all these Stanza language packs are made available under the Open Data Commons Attribution License v1.0.

Language	Language code	Package	Version	Treebank License	Treebank Doc	Notes
Afrikaans	af	afribooms	1.0.0
Ancient Greek	grc	proiel	1.0.0
	grc	perseus	1.0.0
Arabic	ar	padt	1.0.0
Armenian	hy	armtdp	1.0.0
Basque	eu	bdt	1.0.0
Belarusian	be	hse	1.0.0
Bulgarian	bg	btb	1.0.0
Buryat	bxr	bdt	1.0.0
Catalan	ca	ancora	1.0.0
Chinese (simplified)	zh / zh-hans	gsdsimp	1.0.0
Chinese (traditional)	zh-hant	gsd	1.0.0
Classical Chinese	lzh	kyoto	1.0.0
Coptic	cop	scriptorium	1.0.0
Croatian	hr	set	1.0.0
Czech	cs	cac	1.0.0
	cs	cltt	1.0.0
	cs	fictree	1.0.0
	cs	pdt	1.0.0
Danish	da	ddt	1.0.0
Dutch	nl	alpino	1.0.0
	nl	lassysmall	1.0.0
English	en	ewt	1.0.0
	en	gum	1.0.0
	en	lines	1.0.0
	en	partut	1.0.0
Estonian	et	edt	1.0.0
	et	ewt	1.0.0
Finnish	fi	ftb	1.0.0
	fi	tdt	1.0.0
French	fr	gsd	1.0.0
	fr	partut	1.0.0
	fr	sequoia	1.0.0	LGPLLR
	fr	spoken	1.0.0
Galician	gl	ctg	1.0.0
	gl	treegal	1.0.0	LGPLLR
German	de	gsd	1.0.0
	de	hdt	1.0.0
Gothic	got	proiel	1.0.0
Greek	el	gdt	1.0.0
Hebrew	he	htb	1.0.0
Hindi	hi	hdtb	1.0.0
Hungarian	hu	szeged	1.0.0
Indonesian	id	gsd	1.0.0
Irish	ga	idt	1.0.0
Italian	it	isdt	1.0.0
	it	partut	1.0.0
	it	postwita	1.0.0
	it	twittiro	1.0.0
	it	vit	1.0.0
Japanese	ja	gsd	1.0.0
Kazakh	kk	ktb	1.0.0
Korean	ko	gsd	1.0.0
	ko	kaist	1.0.0
Kurmanji	kmr	mg	1.0.0
Latin	la	ittb	1.0.0
	la	proiel	1.0.0
	la	perseus	1.0.0
Latvian	lv	lvtb	1.0.0
Lithuanian	lt	alksnis	1.0.0
	lt	hse	1.0.0
Livvi	olo	kkpp	1.0.0
Maltese	mt	mudt	1.0.0
Marathi	mr	ufal	1.0.0
North Sami	sme	giella	1.0.0
Norwegian (Bokmaal)	no / nb	bokmaal	1.0.0
Norwegian (Nynorsk)	nn	nynorsk	1.0.0
	nn	nynorsklia	1.0.0
Old Church Slavonic	cu	proiel	1.0.0
Old French	fro	srcmf	1.0.0
Old Russian	orv	torot	1.0.0
Persian	fa	seraji	1.0.0
Polish	pl	lfg	1.0.0
	pl	pdb	1.0.0
Portuguese	pt	bosque	1.0.0
	pt	gsd	1.0.0
Romanian	ro	nonstandard	1.0.0
	ro	rrt	1.0.0
Russian	ru	gsd	1.0.0
	ru	syntagrus	1.0.0
	ru	taiga	1.0.0
Scottish Gaelic	gd	arcosg	1.0.0
Serbian	sr	set	1.0.0
Slovak	sk	snk	1.0.0
Slovenian	sl	ssj	1.0.0
	sl	sst	1.0.0
Spanish	es	ancora	1.0.0
	es	gsd	1.0.0
Swedish	sv	lines	1.0.0
	sv	talbanken	1.0.0
Swedish Sign Language	swl	sslc	1.0.0
Tamil	ta	ttb	1.0.0
Telugu	te	mtg	1.0.0
Turkish	tr	imst	1.0.0
Ukrainian	uk	iu	1.0.0
Upper Sorbian	hsb	ufal	1.0.0
Urdu	ur	udtb	1.0.0
Uyghur	ug	udt	1.0.0
Vietnamese	vi	vtb	1.0.0
Wolof	wo	wtb	1.0.0

Other Available Models for Tokenization

We have trained a couple Thai tokenizer models based on publicly available datasets. The Inter-BEST dataset had some strange sentence tokenization according to the authors of pythainlp, so we used their software to resegment the sentences before training. As this is a questionable standard to use, we made the Orchid tokenizer the default.

Available NER Models

The following table lists all NER models supported by Stanza, pretrained on various NER datasets. Again, you can find performance of all available models on the System Performance page.

Table Notes

1. marks the default package for a language.
2. For packages with 4 named entity types, supported types include PER (Person), LOC (Location), ORG (Organization) and MISC (Miscellaneous) 2b. The Vietnamese VLSP model spells out the entire tag, though: PERSON, LOCATION, ORGANIZATION, MISCELLANEOUS.
3. For packages with 18 named entity types, supported types include PERSON, NORP (Nationalities/religious/political group), FAC (Facility), ORG (Organization), GPE (Countries/cities/states), LOC (Location), PRODUCT,EVENT, WORK_OF_ART, LAW, LANGUAGE, DATE, TIME, PERCENT, MONEY, QUANTITY, ORDINAL and CARDINAL (details can be found on page 21 of this OntoNotes documentation).
4. The BSNLP dataset(s) use EVENT, LOCATION, ORGANIZATION, PERSON, PRODUCT.
5. The Italian FBK dataset uses LOC, ORG, PER
6. The Myanmar UCSY dataset uses LOC (Location), NE (Misc), ORG (Organization), PNAME (Person), RACE, TIME, NUM

Available Sentiment Models

A description of the sentiment tool and the models available for that tool can be found here.

Training New Models

To train new models, please see the documents on training and adding a new language.