Apertag 1.0.2

Averaged Perceptron Sequence Tagger

Apertag: Averaged Perceptron Tagger

Apertag is a sequence tagger based on an averaged perceptron model. In order to avoid making assumptions about what kind of sequence data you are labeling, or the format of your features, the input to the tagger is simply sequences of feature value sets. Each set of values represent an observation to receive a tag. A feature value can be any python type, as long as it can be hashed, but it’s important to note that the the values are used only in a binary fashion, i.e. either they exist in the context of the item being tagged or not; the nature of the value has no impact on the decision.

A simple example illustrating an NP-chunker:

>>> t = Tagger()
>>> t.train([[['POS=DT','WRD=the'],['POS=NN','WRD=dog']]],[['NP-B','NP-I']])
>>> t.tag([['POS=DT','WRD=the'],['POS=NN','WRD=dog']])
['NP-B', 'NP-I']

There is one crucial exception to all this featuratory freedom: Any features wishing to make use of the actual output tags need to signal this by formatting their value as a string with special tags that will be replaced by the corresponding tags from the current context during tagging. The tag format is “<Tn>”, where n is the negative index of the tag relative to the current position. For example, if you are training a POS-tagger and you have a feature that looks at the current word and the previous output tag, and the current word is “dog”, the feature could be encoded as “<T1>:dog”. The tagger will expand this using its predicted label context into something like “DT:dog” (depending on your tag set and feature format, of course).

An example illustrating a POS-tagger with output label features:

>>> t = Tagger()
>>> t.train([[['POS -1:<T-1>','W:the'],['POS -1:<T-1>','W:dog']]],[['DT','NN']])
>>> t.tag([['POS -1:<T-1>','W:the'],['POS -1:<T-1>','W:dog']])
['DT', 'NN']

It is most likely a good idea to use this format for training as well, even though you (hopefully) have the output tags yourself at that point, to ensure the features are identical across training and tagging.

If you don’t require output tags for any of your features, you can slightly increase performance (especially for non-string features) by setting expand_features=False.

Where do I put my columns?

The tagger module can be run as a standalone script, which takes its input as good old files of tab delimited columns, where each row is an observation consisting of feature values, followed by the tag in the last column. For more info run:

$ python apertag.py {train,tag} -h

References and acknowledgments

• http://people.csail.mit.edu/mcollins/papers/tagperc.pdf

• https://github.com/lmjohns3/py-tagger/

• https://honnibal.wordpress.com/2013/09/11/a-good-part-of-speechpos-tagger-in-about-200-lines-of-python/