As with most python packages, simply use pip to install from PyPi.
pip install pyconll
pyconll is also available as a conda package on the pyconll
channel. Only package 2.2.0 and newer are available for conda at the
conda install -c pyconll pyconll
This package is designed for, and only tested with python 3.4 and up and
will not be backported to python 2.x.
This tool is intended to be a minimal, low level, and
functional library in a widely used programming language. pyconll
creates a thin API on top of raw CoNLL annotations that is simple and
intuitive in a popular programming language.
In my work with the Universal Dependencies project, I saw a
dissapointing lack of low level APIs for working with the CoNLL-U
format. Most tooling focuses on graph transformations and DSLs for
terse, automated changes. Tools such as Grew and
Treex are very powerful and
productive, but have a learning curve and are limited the scope of their
DSLs. CL-CoNLLU is simple and
low level, but Common Lisp is not widely used in NLP, and difficult to
pickup for beginners. UDAPI is in python
but it is very large and has little guidance. pyconll attempts to fill
the gaps between what other projects have accomplished.
Other useful tools can be found on the Universal Dependencies
Hopefully, individual researchers find pyconll useful, and will use it
as a building block for their tools and projects. pyconll affords a
standardized and complete base for building larger projects without
worrying about CoNLL annotation and output.
# This snippet finds what lemmas are marked as AUX which is a closed class POS in UD
UD_ENGLISH_TRAIN = './ud/train.conll'
train = pyconll.load_from_file(UD_ENGLISH_TRAIN)
aux_lemmas = set()
for sentence in train:
for token in sentence:
if token.upos == 'AUX':
Uses and Limitations
This package edits CoNLL-U annotations. This does not include the
annotated text itself. Word forms on Tokens are not editable and
Sentence Tokens cannot be reassigned or reordered. pyconll focuses
on editing CoNLL-U annotation rather than creating it or changing the
underlying text that is annotated. If there is interest in this
functionality area, please create a github issue for more visibility.
This package also is only validated against the CoNLL-U format. The
CoNLL and CoNLL-X format are not supported, but are very similar. I
originally intended to support these formats as well, but their format
is not as well defined as CoNLL-U so they are not included. Please
create an issue for visibility if this feature interests you.
Lastly, linguistic data can often be very large and this package
attempts to keep that in mind. pyconll provides methods for creating in
memory conll objects along with an iterate only version in case a corpus
is too large to store in memory (the size of the memory structure is
several times larger than the actual corpus file). The iterate only
version can parse upwards of 100,000 words per second on a 16gb ram
machine, so for most datasets to be used on a dev machine, this package
will perform well. The 2.2.0 release also improves parse time and memory
footprint by about 25%!
Contributions to this project are welcome and encouraged! If you are
unsure how to contribute, here is a
from Github explaining the basic workflow. After cloning this repo,
please run make hooks and pip install -r requirements.txt to
properly setup locally. make hooks setups up a pre-push hook to
validate that code matches the default YAPF style. While this is
technically optional, it is highly encouraged.
pip install -r requirements.txt sets up environment dependencies
like yapf, twine, sphinx, etc.
For packaging new versions, please use setuptools version 24.2.0 or
greater for creating the appropriate packaging that recognizes the
README and CHANGELOG
When changing either of these files, please change the Markdown version
and run make docs so that the other versions stay in sync.
Code formatting is done automatically on push if githooks are setup
properly. The code formatter is
YAPF, and using this ensures that
coding style stays consistent over time and between authors.