Unification-Based Learning

This is a copy of code from Tom Kwiatkowski's page at Edinburgh, which no longer exists. It is forked from Jim White's Github repository. The code corresponds to Tom Kwiatkowski's 2010 EMNLP paper:

Inducing probabilistic CCG grammars from logical form with higher-order unification. Tom Kwiatkowski, Luke Zettlemoyer, Sharon Goldwater and Mark Steedman. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cambridge, MA, 2010 http://www.aclweb.org/anthology/D10-1119.pdf

This fork includes a few changes to make it easier to run the experiments.

Building

You will need to have a Java Development Kit in order to build the application and a Java Runtime Environment to run the experiments.

The following command can be used to build the application from the root of the repository:

javac -classpath src src/tom_kwiatkowski/ubl/**/*.java src/RunTrain.java

Running the Experiments

The default experiment can be run using the following command:

java -classpath src RunTrain

In addition, the following arguments can be used to configure the experiments. All of the command-line arguments are of the form --argument=value.

--lf

Sets the logical form to be used, which automatically sets the data root to the appropriate directory for training and testing. Possible values are funql and lambda.

--corpus

Sets the corpus to be used, which automatically sets the data root to the appropriate directory for training and testing. Currently only geo880 is supported with the main script.

--root

The root directory containing the data. This should be one of the directories immediately under "experiments".

--task

Set to either test or dev to determine the data split to be used.

--split

If running the dev experiments, the cross-validation split to use.

--probs

The file (relative to the root directory) containing the GIZA probs.

--fixedlex

The file containing the fixed lexicon.

--output

The name of the output file containing the resulting lexicon.

--epochs

The number of training epochs to use.

--alpha

Alpha0, used to define the temperature of parameter updates.

--c

C, used to define the temperature of parameter updates.

--maxlen

The maximum sentence length in processing.

--multiplier

The initial weight multiplier.

--lexiconweight

The initial lexicon weight.

--prune

The maximum cell size for pruning.