Graph-based Dependency Parser

The data is in Conll06 format.

The parser components are:

• Structured Perceptron
• CLE Decoder
• Feature Model

Training

To train, ensure that all arguments in the file run_training.sh are set as desidered and run script:

bash run_training.sh

Predicting + Evaluating

To make predictions using a trained model, ensure all arguments in the file run_experiments.sh are set as desired and run script:

bash run_experiments.sh

This will create a .pred file in the preds directory. It will also load the newly created file and evaluate again a gold standard (if not wanted, comment out).

Results

English

N_EPOCHS	LR	INIT	FEATS	UAS	UCM
3	0.3	zeros	basic	0.841	0.143
10	0.3	zeros	basic	0.857	0.173
10	0.5	zeros	basic	0.857	0.167
10	0.5 + decay	zeros	basic	0.867	0.195
10	0.5 + decay	random	basic	0.867	0.182
20 + ES 17	0.5 + decay	zeros	basic	0.867	0.197
20 + ES 14	0.5 + decay	zeros	additional	0.873	0.19
20 + ES 16	0.5 + decay	zeros	all	0.878	0.216

German

N_EPOCHS	LR	INIT	FEATS	UAS	UCM
3	0.3	zeros	basic	0.867	0.336
10	0.3	zeros	basic	0.883	0.357
10	0.5	zeros	basic	0.882	0.355
10	0.5 + decay	zeros	basic	0.894	0.385
10	0.5 + decay	random	basic	0.893	0.387
20 + ES 15	0.5 + decay	zeros	basic	0.894	0.381
20 + ES 16	0.5 + decay	zeros	additional	0.897	0.392
20 + ES 13	0.5 + decay	zeros	all	0.904	0.416