Baseline UDPipe models for CoNLL 2017 Shared Task in UD Parsing, and supplementary material.
The models require UDPipe version at least 1.1 and are evaluated using the official evaluation script.
The models are trained on a slightly different split of the official UD 2.0 CoNLL 2017 training data, so called baselinemodel split, in order to allow comparison of models even during the shared task. This baselinemodel split of UD 2.0 CoNLL 2017 training data is available for download.
Furthermore, we also provide UD 2.0 CoNLL 2017 training data with automatically predicted morphology. We utilize the baseline models on development data and perform 10-fold jack-knifing (each fold is predicted with a model trained on the rest of the folds) on the training data.
Finally, we supply all required data and hyperparameter values needed to replicate the baseline models.
Trained models for UDPipe used to produce our final submission to the Vardial 2017 CLP shared task (https://bitbucket.org/hy-crossNLP/vardial2017). The SK model was trained on CS data, the HR model on SL data, and the SV model on a concatenation of DA and NO data. The scripts and commands used to create the models are part of separate submission (http://hdl.handle.net/11234/1-1970).
The models were trained with UDPipe version 3e65d69 from 3rd Jan 2017, obtained from
https://github.com/ufal/udpipe -- their functionality with newer or older versions of UDPipe is not guaranteed.
We list here the Bash command sequences that can be used to reproduce our results submitted to VarDial 2017. The input files must be in CoNLLU format. The models only use the form, UPOS, and Universal Features fields (SK only uses the form). You must have UDPipe installed. The feats2FEAT.py script, which prunes the universal features, is bundled with this submission.
SK -- tag and parse with the model:
udpipe --tag --parse sk-translex.v2.norm.feats07.w2v.trainonpred.udpipe sk-ud-predPoS-test.conllu
A slightly better after-deadline model (sk-translex.v2.norm.Case-feats07.w2v.trainonpred.udpipe), which we mention in the accompanying paper, is also included. It is applied in the same way (udpipe --tag --parse sk-translex.v2.norm.Case-feats07.w2v.trainonpred.udpipe sk-ud-predPoS-test.conllu).
HR -- prune the Features to keep only Case and parse with the model:
python3 feats2FEAT.py Case < hr-ud-predPoS-test.conllu | udpipe --parse hr-translex.v2.norm.Case.w2v.trainonpred.udpipe
NO -- put the UPOS annotation aside, tag Features with the model, merge with the left-aside UPOS annotation, and parse with the model (this hassle is because UDPipe cannot be told to keep UPOS and only change Features):
cut -f1-4 no-ud-predPoS-test.conllu > tmp
udpipe --tag no-translex.v2.norm.tgttagupos.srctagfeats.Case.w2v.udpipe no-ud-predPoS-test.conllu | cut -f5- | paste tmp - | sed 's/^\t$//' | udpipe --parse no-translex.v2.norm.tgttagupos.srctagfeats.Case.w2v.udpipe
Tools and scripts used to create the cross-lingual parsing models submitted to VarDial 2017 shared task (https://bitbucket.org/hy-crossNLP/vardial2017), as described in the linked paper. The trained UDPipe models themselves are published in a separate submission (https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-1971).
For each source (SS, e.g. sl) and target (TT, e.g. hr) language,
you need to add the following into this directory:
- treebanks (Universal Dependencies v1.4):
SS-ud-train.conllu
TT-ud-predPoS-dev.conllu
- parallel data (OpenSubtitles from Opus):
OpenSubtitles2016.SS-TT.SS
OpenSubtitles2016.SS-TT.TT
!!! If they are originally called ...TT-SS... instead of ...SS-TT...,
you need to symlink them (or move, or copy) !!!
- target tagging model
TT.tagger.udpipe
All of these can be obtained from https://bitbucket.org/hy-crossNLP/vardial2017
You also need to have:
- Bash
- Perl 5
- Python 3
- word2vec (https://code.google.com/archive/p/word2vec/); we used rev 41 from 15th Sep 2014
- udpipe (https://github.com/ufal/udpipe); we used commit 3e65d69 from 3rd Jan 2017
- Treex (https://github.com/ufal/treex); we used commit d27ee8a from 21st Dec 2016
The most basic setup is the sl-hr one (train_sl-hr.sh):
- normalization of deprels
- 1:1 word-alignment of parallel data with Monolingual Greedy Aligner
- simple word-by-word translation of source treebank
- pre-training of target word embeddings
- simplification of morpho feats (use only Case)
- and finally, training and evaluating the parser
Both da+sv-no (train_ds-no.sh) and cs-sk (train_cs-sk.sh) add some cross-tagging, which seems to be useful only in
specific cases (see paper for details).
Moreover, cs-sk also adds more morpho features, selecting those that
seem to be very often shared in parallel data.
The whole pipeline takes tens of hours to run, and uses several GB of RAM, so make sure to use a powerful computer.