This fork of UUParser contains the code developed for the following paper:
Miryam de Lhoneux, Miguel Ballesteros and Joakim Nivre. 2019. Recursive Subtree Composition in LSTM-based Dependency Parsing. To appear at NAACL 2019.
Instructions on how to train the parser can be found on the original repository. Options specific to the paper:
--use-recursive-composition: setTreeLSTMto use LSTM composition andRecNNto use a recurrent cell.--unidir-lstm: set toforwardorbackwardto replace the BiLSTM with a forward or a backward LSTM.
All experiments were run with --disable-rlmost and --k 2. We set the POS embedding size to 25 by setting --pos-emb-size to 25 for the +pos models (POS embeddings are disabled by default) and disable character embeddings in the -char models by setting --char-emb-size to 0.
The statistics for Figure 3. were collected with:
python src/treebanks_stats.py --include [treebank iso codes] --datadir [UD data dir]
For the ensemble, run
python src/ensemble.py outfile file1 file2 file3 ...
where file1, file2, etc. are model predictions and the output of the ensemble is written to outfile.
The ensemble uses the dependency decoder from the LxMLS toolkit copied in src/dependency_decoder.py.
This software is released under the terms of the Apache License, Version 2.0.
For questions and usage issues, please contact miryam dot de underscore lhoneux at lingfil dot uu dot se