Averaged perceptron part-of-speech tagger (à la Collins '02)
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


Perceptronix Point Never

Perceptronix Point Never (PPN) is an implementation of a part of speech tagger using a hidden Markov model, the averaged perceptron classifier, and a greedy decoding scheme. The classifier features are based loosely on those used by Ratnaparkhi 1996 and Collins 2002. Following Collins, the same features, including orthographic features, are used regardless of word frequency.

PPN has been tested on CPython 3.4 and PyPy3 (2.3.1, corresponding to Python 3.2); the latter is much, much faster. It requires three third-party packages: nltk and jsonpickle from PyPI and my own nlup library, available from GitHub; see requirements.txt for the versions used for testing.


usage: python -m PPN [-h] [-v] [-V] (-t TRAIN | -r READ)
                     (-u TAG | -w WRITE | -e EVALUATE) [-E EPOCHS] 
                     [-O ORDER]

Perceptronix Point Never, by Kyle Gorman

optional arguments:
  -h, --help            show this help message and exit
  -v, --verbose         enable verbose output
  -V, --really-verbose  even more verbose output
  -t TRAIN, --train TRAIN
                        training data
  -r READ, --read READ  read in serialized model
  -u TAG, --tag TAG     tag unlabeled data
  -w WRITE, --write WRITE
                        write out serialized model
  -e EVALUATE, --evaluate EVALUATE
                        evaluate on labeled data
  -E EPOCHS, --epochs EPOCHS
                        # of epochs (default: 10)
  -O ORDER, --order ORDER
                        Markov order (default: 2)

The included PPN-wsj.json.gz is a tagging model trained on the Wall St. Journal portion of the Penn Treebank.

For anything else, UTSL.


MIT License (BSD-like); see source.


If you use Perceptronix Point Never in your research, please consider linking readers to this repository and/or directing them to the following paper, which describes the tagger and reports an evaluation on the "standard split":

K. Gorman, S. Bedrick, G. Kiss, E. Morley,. R. Ingham, M. Mohammed, K. Papadakis, and J. van Santen. 2015. Automated morphological analysis of clincial languaeg samples. In CLPsych, 108-116.

What's with the name?

It is an homage to experimental musician Daniel Lopatin, who performs under the name Oneohtrix Point Never.

Bugs, comments?

Contact Kyle Gorman.


M. Collins. 2002. Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms. In EMNLP, 1-8.

A. Ratnaparkhi. 1996. A maximum entropy model for part-of-speech tagging. In EMNLP, 133-142.