UralicNLP is a natural language processing library targeted mainly for Uralic languages.
UralicNLP can produce morphological analyses, generate morphological forms, lemmatize words and give lexical information about words in Uralic and other languages. The languages we support include the following languages: Finnish, Russian, German, English, Norwegian, Swedish, Arabic, Ingrian, Meadow & Eastern Mari, Votic, Olonets-Karelian, Erzya, Moksha, Hill Mari, Udmurt, Tundra Nenets, Komi-Permyak, North Sami, South Sami and Skolt Sami. The information originates mainly in FST tools and dictionaries developed in the GiellaLT infrastructure. Currently, UralicNLP uses nightly builds for most of the supported languages.
See the catalog of supported languages
The library can be installed from PyPi.
pip install uralicNLP
If you want to use the Constraint Grammar features (from uralicNLP.cg3 import Cg3), you will also need to install VISL CG-3.
If you are using Linux and you run into problems with installing HFST, you can find some help on a blog post on installing hfst
On Windows, HFST depends on 32 bit Microsoft Visual C++ Redistributable 2017. Although, I would recommend using Windows subsystem for Linux.
Arabic and English FSTs require Foma.
The API is under constant development and new languages will be added to the nightly builds system. That's why UralicNLP provides a functionality for looking up the list of currently supported languages. The method returns 3 letter ISO codes for the languages.
from uralicNLP import uralicApi
uralicApi.supported_languages()
>>{'cg': ['vot', 'lav', 'izh', 'rus', 'lut', 'fao', 'est', 'nob', 'ron', 'olo', 'bxr', 'hun', 'crk', 'chr', 'vep', 'deu', 'mrj', 'gle', 'sjd', 'nio', 'myv', 'som', 'sma', 'sms', 'smn', 'kal', 'bak', 'kca', 'otw', 'ciw', 'fkv', 'nds', 'kpv', 'sme', 'sje', 'evn', 'oji', 'ipk', 'fit', 'fin', 'mns', 'rmf', 'liv', 'cor', 'mdf', 'yrk', 'tat', 'smj'], 'dictionary': ['vot', 'lav', 'rus', 'est', 'nob', 'ron', 'olo', 'hun', 'koi', 'chr', 'deu', 'mrj', 'sjd', 'myv', 'som', 'sma', 'sms', 'smn', 'kal', 'fkv', 'mhr', 'kpv', 'sme', 'sje', 'hdn', 'fin', 'mns', 'mdf', 'vro', 'udm', 'smj'], 'morph': ['vot', 'lav', 'izh', 'rus', 'lut', 'fao', 'est', 'nob', 'swe', 'ron', 'eng', 'olo', 'bxr', 'hun', 'koi', 'crk', 'chr', 'vep', 'deu', 'mrj', 'ara', 'gle', 'sjd', 'nio', 'myv', 'som', 'sma', 'sms', 'smn', 'kal', 'bak', 'kca', 'otw', 'ciw', 'fkv', 'nds', 'mhr', 'kpv', 'sme', 'sje', 'evn', 'oji', 'ipk', 'fit', 'fin', 'mns', 'rmf', 'liv', 'cor', 'mdf', 'yrk', 'vro', 'udm', 'tat', 'smj']}
The dictionary key lists the languages that are supported by the lexical lookup, whereas morph lists the languages that have morphological FSTs and cg lists the languages that have a CG.
If you have a lot of data to process, it might be a good idea to download the morphological models for use on your computer locally. This can be done easily. Although, it is possible to use the transducers over Akusanat API by passing force_local=False.
On the command line:
python -m uralicNLP.download --languages fin eng
From python code:
from uralicNLP import uralicApi
uralicApi.download("fin")
When models are installed, generate(), analyze() and lemmatize() methods will automatically use them instead of the server side API. More information about the models.
Use uralicApi.model_info(language) to see information about the FSTs and CGs such as license and authors. If you know how to make this information more accurate, please don't hesitate to open an issue on GitHub.
from uralicNLP import uralicApi
uralicApi.model_info("fin")
To remove the models of a language, run
from uralicNLP import uralicApi
uralicApi.uninstall("fin")
A word form can be lemmatized with UralicNLP. This does not do any disambiguation but rather returns a list of all the possible lemmas.
from uralicNLP import uralicApi
uralicApi.lemmatize("вирев", "myv")
>>['вирев', 'вирь']
uralicApi.lemmatize("luutapiiri", "fin", word_boundaries=True)
>>['luuta|piiri', 'luu|tapiiri']
An example of lemmatizing the word вирев in Erzya (myv). By default, a descriptive analyzer is used. Use uralicApi.lemmatize("вирев", "myv", descrpitive=False) for a non-descriptive analyzer. If word_boundaries is set to True, the lemmatizer will mark word boundaries with a |. You can also use your own transducer
Apart from just getting the lemmas, it's also possible to perform a complete morphological analysis.
from uralicNLP import uralicApi
uralicApi.analyze("voita", "fin")
>>[['voi+N+Sg+Par', 0.0], ['voi+N+Pl+Par', 0.0], ['voitaa+V+Act+Imprt+Prs+ConNeg+Sg2', 0.0], ['voitaa+V+Act+Imprt+Sg2', 0.0], ['voitaa+V+Act+Ind+Prs+ConNeg', 0.0], ['voittaa+V+Act+Imprt+Prs+ConNeg+Sg2', 0.0], ['voittaa+V+Act+Imprt+Sg2', 0.0], ['voittaa+V+Act+Ind+Prs+ConNeg', 0.0], ['vuo+N+Pl+Par', 0.0]]
An example of analyzing the word voita in Finnish (fin). The default analyzer is descriptive. To use a normative analyzer instead, use uralicApi.analyze("voita", "fin", descrpitive=False). You can also use your own transducer
From a lemma and a morphological analysis, it's possible to generate the desired word form.
from uralicNLP import uralicApi
uralicApi.generate("käsi+N+Sg+Par", "fin")
>>[['kättä', 0.0]]
An example of generating the singular partitive form for the Finnish noun käsi. The result is kättä. The default generator is a regular normative generator. uralicApi.generate("käsi+N+Sg+Par", "fin", dictionary_forms=True) uses a normative dictionary generator and uralicApi.generate("käsi+N+Sg+Par", "fin", descrpitive=True) a descriptive generator. You can also use your own transducer
If you need to get a lower level access to the HFST transducer object, you can use the following code
from uralicNLP import uralicApi
sms_generator = uralicApi.get_transducer("sms", analyzer=False) #generator
sms_analyzer = uralicApi.get_transducer("sms", analyzer=True) #analyzer
The same parameters can be used here as for generate() and analyze() to specify whether you want to use the normative or descriptive analyzers and so on. The defaults are get_transducer(language, cache=True, analyzer=True, descrpitive=True, dictionary_forms=True).
Note this requires the models to be installed (see above) and VISL CG-3. The disambiguation process is simple.
from uralicNLP.cg3 import Cg3
sentence = "Kissa voi nauraa"
tokens = sentence.split(" ") #Do a simple tokenization for the sentence
cg = Cg3("fin")
print cg.disambiguate(tokens)
>>[(u'Kissa', [<Kissa - N, Prop, Sg, Nom, <W:0.000000>>, <kissa - N, Sg, Nom, <W:0.000000>>]), (u'voi', [<voida - V, Act, Ind, Prs, Sg3, <W:0.000000>>]), (u'nauraa', [<nauraa - V, Act, InfA, Sg, Lat, <W:0.000000>>])]
The return object is a list of tuples. The first item in each tuple is the word form used in the sentence, the second item is a list of Cg3Word objects. In the case of a full disambiguation, these lists have only one Cg3Word object, but some times the result of the disambiguation still has some ambiguity. Each Cg3Word object has three variables lemma, form and morphology.
disambiguations = cg.disambiguate(tokens)
for disambiguation in disambiguations:
possible_words = disambiguation[1]
for possible_word in possible_words:
print possible_word.lemma, possible_word.morphology
>>Kissa [u'N', u'Prop', u'Sg', u'Nom', u'<W:0.000000>']
>>kissa [u'N', u'Sg', u'Nom', u'<W:0.000000>']
>>voida [u'V', u'Act', u'Ind', u'Prs', u'Sg3', u'<W:0.000000>']
>>nauraa [u'V', u'Act', u'InfA', u'Sg', u'Lat', u'<W:0.000000>']
The cg.disambiguate takes in remove_symbols as an optional argument. Its default value is True which means that it removes the symbols (segments surrounded by @) from the FST output before feeding it to the CG disambiguator. If the value is set to False, the FST morphology is fed in to the CG unmodified.
The default FST analyzer is a descriptive one, to use a normative analyzer, set the descriptive parameter to False cg.disambiguate(tokens,descrpitive=False).
It is possible to run one CG with tags produced by transducers of multiple languages.
from uralicNLP.cg3 import Cg3
cg = Cg3("fin", morphology_languages=["fin", "olo"])
print(cg.disambiguate(["Kissa","on","kotona", "."], language_flags=True))
The code above will use the Finnish (fin) CG rules to disambiguate the tags produced by Finnish (fin) and Olonets-Karelian (olo) transducers. The language_flags parameter can be used to append the language code at the end of each morphological reading to identify the transducer that produced the reading.
It is also possible to pipe multiple CG analyzers. This will run the initial morphological analysis in the first CG, disambiguate and pass the disambiguated results to the next CG analyzer.
from uralicNLP.cg3 import Cg3, Cg3Pipe
cg1 = Cg3("fin")
cg2 = Cg3("olo")
cg_pipe = Cg3Pipe(cg1, cg2)
print(cg_pipe.disambiguate(["Kissa","on","kotona", "."]))
The example above will create a CG analyzer for Finnish and Olonets-Karelian and pipe them into a Cg3Pipe object. The analyzer will first use Finnish CG with a Finnish FST to disambiguate the sentence, and then Olonets-Karelian CG to do a further disambiguation. Note that FST is only run in the first CG object of the pipe.
UralicNLP makes it possible to obtain the lexicographic information from the Giella dictionaries. The information can contain data such as translations, example sentences, semantic tags, morphological information and so on. You have to define the language code of the dictionary.
For example, "sms" selects the Skolt Sami dictionary. The word used to query, however, can appear in any language. If the word is a lemma in Skolt Sami, the result will appear in "exact_match", if it's a word form for a Skolt Sami word, the results will appear in "lemmatized", and if it's a word in some other language, the results will appear in "other_languages", i.e if you search for cat in the Skolt Sami dictionary, you will get a result of a form {"other_languages": [Skolt Sami lexical items that translate to cat]}
An example of querying the Skolt Sami dictionary with car.
from uralicNLP import uralicApi
uralicApi.dictionary_search("car", "sms")
>>{'lemmatized': [], 'exact_match': [], 'other_languages': [{'lemma': 'autt', ...}, ...]
It is possible to list all lemmas in the dictionary:
from uralicNLP import uralicApi
uralicApi.dictionary_lemmas("sms")
>> ['autt', 'sokk' ...]
By default, UralicNLP uses a TinyDB backend. This is easy as it does not require an external database server, but it can be extremely slow. For this reason, UralicNLP provides a MongoDB backend.
Make sure you have both MongoDB and pymongo installed.
First, you will need to download the dictionary and import it to MongoDB. The following example shows how to do it for Komi-Zyrian.
from uralicNLP import uralicApi
uralicApi.download("kpv") #Download the latest dictionary data
uralicApi.import_dictionary_to_db("kpv") #Update the MongoDB with the new data
After the initial setup, you can use the dictionary queries, but you will need to specify the backend.
from uralicNLP import uralicApi
from uralicNLP.dictionary_backends import MongoDictionary
uralicApi.dictionary_lemmas("sms",backend=MongoDictionary)
uralicApi.dictionary_search("car", "sms",backend=MongoDictionary)
Now you can query the dictionaries fast.
UralicNLP comes with tools for parsing and searching CoNLL-U formatted data. Please refer to the Wiki for the UD parser documentation.
UralicNLP provides semantic models for Finnish (SemFi) and other Uralic languages (SemUr) for Komi-Zyrian, Erzya, Moksha and Skolt Sami. Find out how to use semantic models
When your NLP needs grow out of what UralicNLP can provide, we have your back! Rootroo offers consulting related to a variety of NLP tasks. We have a strong academic background in the state-of-the-art AI solutions for every NLP need. Just contact us, we won't bite.
If you use UralicNLP in an academic publication, please cite it as follows:
Hämäläinen, Mika. (2019). UralicNLP: An NLP Library for Uralic Languages. Journal of open source software, 4(37), [1345]. https://doi.org/10.21105/joss.01345
@article{uralicnlp_2019,
title={{UralicNLP}: An {NLP} Library for {U}ralic Languages},
DOI={10.21105/joss.01345},
journal={Journal of Open Source Software},
author={Mika Hämäläinen},
year={2019},
volume={4},
number={37},
pages={1345}
}
For citing the FSTs and CGs, see uralicApi.model_info(language).