When replicating the results of the CoNLL 2017 shared task from the models here https://nlp.stanford.edu/data/CoNLL17_Parser_Models/ , I ran into an issue as I could not get the same performance as is reported in the paper. My steps in reproducing, I showcase English LinES treebank:
1) download ud-treebanks-conll2017 that was distributed among participants of CoNLL 2017 shared task
2) download respective model from https://nlp.stanford.edu/data/CoNLL17_Parser_Models/
3) download test sets ud-test-v2.0-conll2017
4) tag test set file "en_lines.conllu" in ud-test-v2.0-conll2017 by the downloaded model "English-LinES-Tagger"
5) parse tagged in step 4 test set file "en_lines.conllu" located in "English-LinES-Tagger" by downloaded model "English-LinES-Parser"
6) evaluate by the conll17_ud_eval.py script with the gold data and parsed result.
The result is different than the one reported in the paper. So what could have gone wrong?
When replicating the results of the CoNLL 2017 shared task from the models here https://nlp.stanford.edu/data/CoNLL17_Parser_Models/ , I ran into an issue as I could not get the same performance as is reported in the paper. My steps in reproducing, I showcase English LinES treebank:
1) download ud-treebanks-conll2017 that was distributed among participants of CoNLL 2017 shared task 2) download respective model from https://nlp.stanford.edu/data/CoNLL17_Parser_Models/ 3) download test sets ud-test-v2.0-conll2017 4) tag test set file "en_lines.conllu" in ud-test-v2.0-conll2017 by the downloaded model "English-LinES-Tagger" 5) parse tagged in step 4 test set file "en_lines.conllu" located in "English-LinES-Tagger" by downloaded model "English-LinES-Parser" 6) evaluate by the conll17_ud_eval.py script with the gold data and parsed result.
The result is different than the one reported in the paper. So what could have gone wrong?