Lamon, The Latin POS Tagger & Lemmatizer

.. image:: https://badge.fury.io/py/lamonpy.svg :target: https://pypi.python.org/pypi/lamonpy .. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.4091536.svg :target: https://doi.org/10.5281/zenodo.4091536

Lamon (LAtin MOrphological tools, pronounced /leɪmən/) is a simple POS tagger & lemmatizer for Latin written in C++ and Lamonpy is a Python package of Lamon. You can easily obtain lemma and tag of each word in given text using Lamonpy.

Getting Started

You can install Lamonpy easily using pip. (https://pypi.org/project/lamonpy/) ::

$ pip install --upgrade pip
$ pip install lamonpy

The supported OS and Python versions are:

• Linux (x86-64) with Python >= 3.5
• macOS >= 10.13 with Python >= 3.5
• Windows 7 or later (x86, x86-64) with Python >= 3.5
• Other OS with Python >= 3.5: Compilation from source code required (with c++11 compatible compiler)

Here is a simple example using Lamonpy to analyze Latin texts. ::

from lamonpy import Lamon
lamon = Lamon()
score, tagged = lamon.tag('In principio creavit Deus caelum et terram.')[0]
print(tagged)
# `tagged` is a list of tuples `(start_pos, end_pos, lemma, tag)`
# [(0, 2, 'in', 'r--------'), 
#  (3, 12, 'principium', 'n-s---nb-'), 
#  (13, 20, 'creo', 'v3sria---'), 
#  (21, 25, 'deus', 'n-s---mn-'), 
#  (26, 32, 'caelus', 'n-s---ma-'), 
#  (33, 35, 'et', 'c--------'), 
#  (36, 42, 'terra', 'n-s---fa-'), 
#  (42, 43, '.', '---------')]

Tagging Model and Its Accuracy

Lamon's tagging model is based on BiLSTM network trained with Perseus Latin Dependency Treebanks <https://perseusdl.github.io/treebank_data/>_ (4,000 sentences) and self-trained with raw Latin corpora (440,000 sentences) collected by Latina Vivense <https://latina.bab2min.pe.kr/xe/text>_.

Since there is no available standard for evaluating Latin taggers, we built own test set named vivens of 900 sentences. The result of evaluation is shown below:

+-------------------+---------------------+---------------------+ | | vivens (900 sents) | Perseus (4000 sents)|

•               +-------+------+------+-------+------+------+
  

| | lemma | tag | both | lemma | tag | both | +===================+=======+======+======+=======+======+======+ |Lamon | 94.6 | 83.0 | 81.1 | 89.4 | 80.2 | 76.6 | +-------------------+-------+------+------+-------+------+------+ |Lamon (large) | 94.2 | 83.3 | 81.3 | 89.7 | 81.9 | 78.3 | +-------------------+-------+------+------+-------+------+------+ |Lamon (uv.) | 94.4 | 82.6 | 80.7 | 87.7 | 77.9 | 73.8 | +-------------------+-------+------+------+-------+------+------+ |Backoff | 88.1 | | | 92.4 | | | +-------------------+-------+------+------+-------+------+------+ |123 POS | | 58.1 | 54.8 | | 83.8 | 79.6 | +-------------------+-------+------+------+-------+------+------+ |CRF POS | | 69.1 | 63.4 | | 77.3 | 72.9 | +-------------------+-------+------+------+-------+------+------+

• Lamon : base size (embedding_size:80, hidden_size:160)
• Lamon (large) : large size (embedding_size:160, hidden_size:320)
• Lamon (uv.) : large size without Perseus' dataset
• Backoff : cltk.lemmatize.latin.backoff.BackoffLatinLemmatizer <https://docs.cltk.org/en/latest/latin.html#lemmatization-backoff-method>_
• 123 POS : cltk.tag.pos.POSTag.tag_ngram_123_backoff <https://docs.cltk.org/en/latest/latin.html#gram-backoff-tagger>_
• CRF POS : cltk.tag.pos.POSTag.tag_crf <https://docs.cltk.org/en/latest/latin.html#crf-tagger>_
• For calculating both score of 123 POS and CRF POS, Backoff's results are used.

Since Lamon and all cltk's tagger are trained with Perseus' dataset, the scores for Perseus are not significant for confirming the actual accuracy of each model. Rather, it shows that 123 POS and CRF POS are overfitting to Perseus's dataset.

Because the size of the vivens dataset is small, the results of this evaluation can be inaccurate. We plan to acquire larger dataset for evaluation and publish the dataset to make more accurate evaluation.

Tagset

Lamon supports three types of tagset.

1. perseus ::

   1: part of speech

   n noun v verb a adjective d adverb c conjunction r adposition p pronoun m numeral i interjection e exclamation u punctuation

   2: person

   1 first person 2 second person 3 third person

   3: number

   s singular p plural

   4: tense

   p present i imperfect r perfect l pluperfect t future perfect f future

   5: mood

   i indicative s subjunctive n infinitive m imperative p participle d gerund g gerundive

   6: voice

   a active p passive d deponent

   7: gender

   m masculine f feminine n neuter

   8: case

   n nominative g genitive d dative a accusative v vocative b ablative l locative

   9: degree

   p positive c comparative s superlative

2. vivens ::

   Moods

   D: indicative S: subjunctive I: imperative T: infinitive L: participle

   Tenses

   0M: present 0E: perfect RM: imperfect RE: pluperfect FM: future FE: future perfect

   Voices

   A: active P: passive

   Participle (combination of mood, tense & voice)

   L0A: present participle LRP: past participle LFA: future active participle LFP: gerundive

   Persons

   1: first 2: second 3: third

   Genders

   m: masculine f: feminine n: neuter

   Numbers

   s: singular p: plural

   Cases

   o: nominative g: genitive d: dative a: accusative b: ablative v: vocative x: adverbial

   Degrees

   (positive isn't marked explicitly.) c: comparative u: superlative

   etc

   r: preposition j: conjunction

3. raw ::

   ...

Demo

https://latina.bab2min.pe.kr/xe/lTagger (Korean)

Larger Models

Due to the package size limit of pypi, the distributed wheel package contains base model only. We provide larger models by Google-drive links.

• Large model : https://drive.google.com/file/d/1u8LdvD-zKtrj7kDRs6CjQw74ZG6aT8jS/view?usp=sharing
• Large model (unsupervised) : https://drive.google.com/file/d/1nw8LO_1o0O894gXzgQ7Hx5Fyikvy1w2u/view?usp=sharing

You can use these models by passing the model path to Lamon.__init__ as arguments. ::

from lamonpy import Lamon
lamon = Lamon(dict_path='dict.large.bin', tagger_path='tagger.large.bin')

License

Lamonpy is licensed under the terms of MIT License, meaning you can use it for any reasonable purpose and remain in complete ownership of all the documentation you produce.

History

• 0.2.0 (2020-10-16)
  • [NUM] token for Roman numeral was added.
  • The accuracy was slightly increased by introducing joint lemma-tag layer.
• 0.1.0 (2020-09-26)
  • the first version of lamonpy

Citation

::

@software{bab2min_2020_4091536,
  author       = {bab2min},
  title        = {bab2min/lamonpy: 0.2.0},
  month        = oct,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v0.2.0},
  doi          = {10.5281/zenodo.4091536},
  url          = {https://doi.org/10.5281/zenodo.4091536}
}