search

oudalab / Arabic-NER

32 stars 11 forks source link

use prodigy with mixedin data and pretrained spacy ner model get error. #13

Open YanLiang1102 opened 6 years ago

YanLiang1102 commented 6 years ago

I created an issue in Spacy for this.

YanLiang1102 commented 6 years ago

https://github.com/explosion/spaCy/issues/2496

YanLiang1102 commented 6 years ago

well I hard coded and put the model path inside of spacy source code in here File "/usr/local/lib/python3.5/dist-packages/spacy-2.0.12.dev0-py3.5-linux-x86_64.egg/spacy/util.py", line 119, in load_model but after training for like 30 mins got and segmentation fault error.....:(

YanLiang1102 commented 6 years ago

well that is becasue it uses up all the memory since the big pretrained vectors, so use the our own vector and finally get it trained, I made virtual environment and rebuild all the dependecy and finally get it to run. here is the perfomance with only the ldcdaat+prodigy data not perfect, since it is using ontoNotes orginal format, it dose not take advantage of the tag clustering stuff, we have and the anercorp data is not merged in too. image @ahalterman

YanLiang1102 commented 6 years ago

@andy what is that BEFORE accuracy mean Andy? the 0.029 the mixed in ratio I am using is 5 so prodigy has 801 records, we will mix in the first 4005 records of onto_notes to the augmented data set.

YanLiang1102 commented 6 years ago

also I need to manually copy the ar model into the dist-package forder for the virtual environment in order to fix the problem

ahalterman commented 6 years ago

The BEFORE is built in to Prodigy, I think. It just applies the existing model to the entities and reports the accuracy. So that must be the accuracy on model9.

YanLiang1102 commented 6 years ago

I mean that accuracy is definitely not right on model9 it should be around 55% I guess that is with "the new format data" on model9?