Skip to main content

An Ancient Greek Morphology Tagger

Project description

AnGEL Morphology

An Ancient Greek Morphology Tagger

Angel takes in Ancient Greek plain text and returns morphology tags for each token. More specifically, it tags part-of-speech, person, number, tense, mood, voice, gender, case, and degree. The morphology tagging mostly follows the AGDT 2.0 style. The exception would be the inclusion of the "g" tag for the part-of-speech "particle" which is not included in the AGDT 2.0 documentation, but is used by some annotators within the AGDT 2.1 treebank collection. For a list of all tags available within each element, refer to the AGDT 2.1 tagset


pip install angel-tag


The input should be a string. The output is a tuple of tuples.

from angel import tag

greek_string = 'ὧν ἐς πολὺ μὲν οὐκ ἐπῄσθοντο Ῥωμαῖοι διὰ τὰς ἐν ἄστει κρίσεις τε καὶ στάσεις·'
results = tag(greek_string)


(('ὧν', 'p-p---ng-'), ('ἐς', 'r--------'), ('πολὺ', 'a-s---na-'), ('μὲν', 'd--------'), ('οὐκ', 'd--------'), ('ἐπῄσθοντο', 'v3paim---'), ('Ῥωμαῖοι', 'n-p---mn-'), ('διὰ', 'r--------'), ('τὰς', 'l-p---fa-'), ('ἐν', 'r--------'), ('ἄστει', 'n-s---nd-'), ('κρίσεις', 'n-p---fa-'), ('τε', 'd--------'), ('καὶ', 'c--------'), ('στάσεις', 'n-p---fa-'), ('·', 'u--------'))

If you just need to tag one sentence, then the above example is fine. But if you want to tag an entire document, don't feed it to the tagger one sentence at a time; or even worse, one token at a time. It'll take forever that way and accuracy will suffer. Just give it the entire document all at once.

from angel import tag

with open('xenophon-hellenica.txt', 'r', encoding='utf-8') as infile:
    entire_book =

results = tag(entire_book)

It is possible that an excessively large string may cause memory issues. If you run into that problem, then perhaps split the text in half and try that. This is an issue that will be addressed in later releases.


This novel architecture utilizes no rules or morphology lookup tables. Rather, it examines individual token morphology and each token's context within the sentence using a series of neural networks. Furthermore, because of the varying tendencies of the many human annotators which are found among the AGDT treebanks, Angel considered annotators as a feature during training. Consequently, while running inference, an annotator must be chosen for the tagger to emulate. "Vanessa Gorman" is the default choice as her annotation style is up to date and she is currently the single most prolific annotator.


Partially imitating the assessment criteria used by Barbara McGillivray and Alessandro Vatri in the development of their state of the art (91% POS accuracy) tagger they used in their Diorisis corpus, Angel was trained on 26 works in the AGDT 2.1 treebank while 7 works were reserved for validation during training. Though Diorisis trained on roughly 50% more data (from the PROIEL treebanks), Angel outperformed it scoring 95.5% accuracy predicting parts-of-speech in the validation set. That score was further confirmed by testing upon the first five works within the Gorman treebanks wherein it scored 95.7% part-of-speech accuracy, earning it state of the art status by a significant margin.

Further Development

  • Exceptionally large works may cause memory issues. This will be addressed in subsequent updates.
  • Annotation inconsistencies will be addressed. Some obsolete tags still exist in the current version of AGDT.
  • Lemmatization and dependency tagging integration is planned. Other paradigms for morphology tagging may be implemented as well.
  • A lot more training data is available. It will be used.
  • Much hyperparameter tinkering to be done.
  • OS testing.


Copyright (c) 2021 Chris Drymon under the MIT license.

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

angel-tag-0.0.9.tar.gz (9.1 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page