Hi @stevenbird and all,
Here is some results from our arc-eager and arc-standard implementation compared with Malt Parser on part of English PTB with

training data : 181 sentences (165: projective, 16: non-projective)
testing data : 214 sentences

Our implementation :
A. Arc-standard (UAS,LAS)

Training data : (0.961, 0.956)
Test data : (0.737, 0.708)

B. Arc-Eager (UAS,LAS)

Training data : (0.954, 0.948)
Test data : (0.726, 0.696)

Malt parser :
A. Arc-Standard (UAS,LAS)

Test data : (0.756, 0.725)

B. Arc-Eager

Test data (0.759, 0.726)

The discrepancies between our implementation and Malt parser are

Different feature sets: For simplicity we only use basic features.
Learning algorithm: Our implementation is slower mainly because scikit-learn libsvm wrapper haven't implemented SVC.decision_function for sparse matrix. I have to used SVC.predict_proba which use cross-validation.

Some points of implementation I would like to discuss :
(1) Currently, the code only accept ConLL shared task data format for both training and testing. What else should I implement ?
(2) Should I add more complicated features for better performance
(3) Should we focus on improving the learning speed by

Use different wrapper for libsvm
Separate the training data to different bin based on some features. Train different classifiers for each bin

Transition-based dependency parsing algorithms #694

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions