Dagger - Dataset Aggregation for Learning to Search

A learning implementation of "A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning" (https://arxiv.org/abs/1011.0686).

Requirements

numpy scikit-learn

TODO

A lot of this could be refactored to make it easier to plug in both new Oracles and different learning algorithms. It shouldn't be particularly hard to extend this to other learning tasks either.

Data

While Dagger can be used for any type of structured learning task, this implementation was mostly intended for POS. However, the amount of time to modify it for other tasks should be fairly trivial.

The format is pretty simple. batches of <feature> <class>, separated by a carriage return:

the article
quick adjective
brown adjective
fox noun
jumped verb
over preposition
the the
lazy adjective
dog noun

i pronoun
love verb
this determiner
time noun
of preposition
year noun

Running

You can train a POS tagger fairly easily:

python dagger.py <train_file> <model>

HYou can compare it to a baseline easily as well:

python baseline.py <train_file> <model>