Cross-lingual Dependency Parsing Based on Distributed Representations

Jiang Guo, jiangfeng1124@gmail.com

This parser can be used both for monolingual and cross-lingual dependency parsing. The monolingual component has been optimized and integrated into LTP.

Compile

Require: cmake (available in most Linux systems)

• ./configure
• make

Resources

Alignment dictionaries:

• resources/align/PROJ (target->source)
• resources/align/CCA (source->target).

Unzip the embeddings:

• cd resources && unzip acl15-cl-wemb.zip && mv acl15-cl-wemb/cca cca && mv acl15-cl-wemb/projected projected && rm -r acl15-cl-wemb && cd ..

Note that the EN word embeddings in PROJ are used only for initialization, they get updated/finetuned while training.

Running the executable

For robust projection: ./scripts/train-PROJ.sh

• set fix_word_embeddings=false (conf/nndep.cfg)

For CCA: ./scripts/train-CCA.sh

Learning parameters are defined in conf/nndep.cfg.

Reference

@InProceedings{Guo:2015:clnndep,
  author    = {Guo, Jiang and Che, Wanxiang and Yarowsky, David and Wang, Haifeng and Liu, Ting},
  title     = {Cross-lingual Dependency Parsing Based on Distributed Representations},
  booktitle = {Proceedings of ACL},
  year      = {2015},
}