This repository contains the code to reproduce the experiment result of the paper CRF autoencoder for unsupervised dependency parsing on WSJ data set and PASCAL dataset.
- Java 8
- sbt 0.13.13
sbt assembly
- Experiments on WSJ dataset.
java -cp target/scala-2.12/CRFAE-Dep-Parser-assembly-1.0.jar edu.shanghaitech.nlp.crfae.parser.DepParserLauncher --train-file wsj10-train.txt --test-file wsj-test.txt --model-type projective --training-type hard --reg-type L1 --rules-type wsj --km-type decoder --gd-num-passes 10 --em-num-passes 10 --batch-size 200 --init-rate 0.1 --lambda 2.5 --prior-weight 9.0
- Experiments on Pascal dataset.
java -cp target/scala-2.12/CRFAE-Dep-Parser-assembly-1.0.jar edu.shanghaitech.nlp.crfae.parser.DepParserLauncher --train-file pascal-{}-train.txt --test-file pascal-{}-test.txt --model-type projective --training-type hard --reg-type L1 --rules-type ud --km-type joint --gd-num-passes 2 --em-num-passes 2 --batch-size 200 --init-rate {} --lambda {} --prior-weight {}
| language | learning rate (init-rate) | lambda | prior weight |
|---|---|---|---|
| basque | 0.01 | 2.5 | 0.1 |
| czech | 0.01 | 1.0 | 1.0 |
| danish | 0.05 | 2.5 | 25.0 |
| dutch | 0.1 | 2.5 | 9.0 |
| protuguese | 0.05 | 2.5 | 5.0 |
| slovene | 0.01 | 0.5 | 1.0 |
| swedish | 0.05 | 0.5 | 25.0 |
No wonder
DT NN
DT NN
2 0
So he adjusts
RB PRP VBZ
RB PRP VBZ
3 3 0
- Sequence of words.
- Sequence of POS.
- Sequence of Universal POS. (If it is not available, duplicate the sequence of POS.)
- Sequence of Head index.
If you found this repository helpful, you could cite
@inproceedings{cai2017crf,
title={CRF Autoencoder for Unsupervised Dependency Parsing},
author={Cai, Jiong and Jiang, Yong and Tu, Kewei},
booktitle={Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
pages={1639--1644},
year={2017}
}
This code is distributed under GPLv3.0 LICENSE.