Skip to main content

SpaCy support to split affixes for Freeling-like affixes rules and dictionaries

Project description

SpaCy Affixes Documentation Status

SpaCy support for affixes splitting for Freeling-like affixes rules and dictionaries.


This library was born to split clitics from verbs so POS tagging works out-of-the-box with spaCy models.

from spacy_affixes import AffixesMatcher
nlp = spacy.load("es")
affixes_matcher = AffixesMatcher(nlp, split_on=["VERB"])
nlp.add_pipe(affixes_matcher, name="affixes", before="tagger")
for token in nlp("Yo mismamente podría hacérselo bien."):

The output will be

Hay Hay AUX AUX__Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin False None None None 0
que que SCONJ SCONJ___ False None None None 0
hacér hacer VERB  True suffix_selo suffix hacer 2
se se PRON PRON__Person=3 False None None None 0
lo el PRON PRON__Case=Acc|Gender=Masc|Number=Sing|Person=3|PronType=Prs False None None None 0
todo todo PRON PRON__Gender=Masc|Number=Sing|PronType=Ind False prefix_todo None None 0
, , PUNCT PUNCT__PunctType=Comm False None None None 0
y y CONJ CCONJ___ False None None None 0
rápidamente rápidamente ADV ADV___ False suffix_mente None None 0
además además ADV ADV___ False prefix_a None None 0
. . PUNCT PUNCT__PunctType=Peri False None None None 0

However, words with suffixes could also be split if needed, or virtually any word for which a rule matches, just by passing a list of Universal Dependency POS’s to the argument split_on. Passing in split_on="*" would make AffixesMatcher() try to split on everything it finds.

Rules and Lexicon

Due to licensing issues, spacy-affixes comes with no rules nor lexicons by default. There are two ways of getting data into spacy-affixes:

  1. Create the rules and lexicon yourself with the entities you are interested on, and pass them in using AffixesMatcher(nlp, rules=<rules>, dictionary=<dictionary>). The format for these is as follows.

    • rules: Dictionary of rules for affixes handling. Each dict uses a key that contains the pattern to match and the value is a list of dicts with the corresponding rule parameters:

      • pattern: Regular expression to match, (ex. r"ito$") If a match is found, it gets removed from the token


      • pos_re: EAGLE regular expression to match, (ex. r"V")

      • strip_accent: Boolean indicating whether accents should be stripped in order to find the rest of the token in the lexicon

      • affix_add: List of strings to add to the rest of the token to find it in the lexicon. Each element in the list is tried separately, as in an OR condition. The character * means add nothing (ex. ["*", "io"])

      • affix_text: List of Strings with the text to the rest of the token as individual tokens. For example, a rule for dígamelo might have ["me", "lo"] as its affix_text

    • lexicon: Dictionary keyed by word with values for lemma, EAGLE code, UD POS, and UD Tags.

  2. Convert the Freeling data. Take into account that if you use Freeling data you are effectively agreeing to their license, which might have implications in the release if your own code. If installed, spacy-affixes will look for the environment variables FREELINGDIR or FREELINGSHARE to find the affixes rules and dictionary files and will process them. If you don’t have Freeling installed you can always run the download command:

python -m spacy_affixes download <lang> <version>

Where lang is the 2-character ISO 639-1 code for a supported language, and version an tagged version in their GitHub repository.


  • Some decisions might feel idiosyncratic since the purpose of this library at the beginning was to just split clitics in Spanish texts.


0.1.0 (2019-04-02)

  • First release on PyPI.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

spacy_affixes-0.1.4.tar.gz (28.5 kB view hashes)

Uploaded source

Built Distribution

spacy_affixes-0.1.4-py2.py3-none-any.whl (15.6 kB view hashes)

Uploaded py2 py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page