pmulcaire / rosita Goto Github PK

Hi, thanks for the great repo!

I was working on training SRL models on the Arabic part of ontonotes using AllenNLP. I follow the instruction here to do the pre-processing. I also remove the diacritics as describe in your paper. The training part can successfully be finished.

However, the program crashes at the validation step. The validation code will finally call the conll 2005 evaluation script srl-eval.pl to compute metrics such as precision / recall / f1, and srl-eval.pl report an error: phrase_set->load_SE_tagging: some phrases are unclosed! . There is a similar issue report here.

I was thinking this bug is due to there are some complex or wrong labels in the gold label that srl-eval.pl can't handle. I was wondering have you experience similar problems?

Any help will be appreciated!