frankier/apertiumhmm2dot

Draws diagrams of apertium-tagger HMM model files using Graphviz

What?

This program draws Apertium's HMM models in a way inspired by https://en.wikipedia.org/wiki/Hidden_Markov_model#Inference

Requirements

• Apertium's library files and header files.
• An Apertium pair with trained HMM tagger data.

Usage

$ make
$ ./hmm2dot.sh /path/to/an/apertium/tagger/model.prob

Then resulting files are:

1. msm.svg: the internal Markov model. Nodes are "coarse" part-of-speech tags. Edge weights reflect the probability of one part of speech following another.
2. obs.svg: a bipartite graph reflecting the probability a certain ambiguity set will be observed given a particular actual part of speech tag.

I have tested with apertium-en-ca/en-ca.prob.

If you're using Debian or Ubuntu's Graphviz, layout will be poor due to being compiled without a triangulation library. It is (slightly) better with one. Install libgts-dev and then compile GraphViz yourself for a better layout. If anyone can coax Graphviz into producing more readable layouts, a pull request would be very welcome!

Also included

The following programs might be useful as small examples for those wanting to start working with Apertium's taggers.

• trace-tagger-model - dump some info about a hmm tagger model
• trace-tagger-spec - unfinished
• trace-streamed-types - parse Apertium's stream format using m5w's parser and print the result (used by the unigram taggers and the perceptron tagger)
• trace-tagger-words - parse Apertium's stream format using the older HMM parser and print the result. Needs a tsx.

Caveats/TODO

• Should parse command line options better.
• Shouldn't put implementation in header files.
• Different Graphviz options could possibly produce better graph layout.