French Demo

[mcavdar]

• +spacy fr model
• check vocabulary stuff
• see embedding performance

[mcavdar]

f1_score_vs_epoch_for_all_classes.pdf

[mcavdar]

First output with general purpose fr word embedding

[mcavdar]

First output with medical(fr) word embedding

[mcavdar]

With wikifr word embedding:
f1_conll_vs_epoch_for_all_classes.pdf
f1_score_vs_epoch_for_all_classes.pdf
If we see different classes' performance, we realize that there is a huge difference between some classes'(in the test set) score. But why?

Output of system:
Evaluate model on the test set
processed 12240 tokens with 1203 phrases; found: 1257 phrases; correct: 767.
accuracy: 92.88%; precision: 61.02%; recall: 63.76%; FB1: 62.36
ANAT: precision: 58.37%; recall: 89.71%; FB1: 70.72 209
CHEM: precision: 54.14%; recall: 35.21%; FB1: 42.67 266
DEVI: precision: 22.95%; recall: 77.78%; FB1: 35.44 61
DISO: precision: 49.50%; recall: 65.36%; FB1: 56.34 202
GEOG: precision: 77.78%; recall: 58.33%; FB1: 66.67 9
LIVB: precision: 86.49%; recall: 77.73%; FB1: 81.88 222
OBJC: precision: 35.71%; recall: 55.56%; FB1: 43.48 14
PHEN: precision: 14.29%; recall: 30.00%; FB1: 19.35 21
PHYS: precision: 48.94%; recall: 63.89%; FB1: 55.42 47
PROC: precision: 76.21%; recall: 90.75%; FB1: 82.85 206

Confusion matrix:
confusion_matrix_for_epoch_0017_in_valid_token_evaluation.pdf

[mcavdar]

After POS Tags added:

processed 12240 tokens with 1203 phrases; found: 1351 phrases; correct: 823.
accuracy: 92.89%; precision: 60.92%; recall: 68.41%; FB1: 64.45
ANAT: precision: 58.62%; recall: 87.50%; FB1: 70.21 203
CHEM: precision: 52.35%; recall: 46.21%; FB1: 49.09 361
DEVI: precision: 43.48%; recall: 55.56%; FB1: 48.78 23
DISO: precision: 44.54%; recall: 69.28%; FB1: 54.22 238
GEOG: precision: 81.82%; recall: 75.00%; FB1: 78.26 11
LIVB: precision: 86.32%; recall: 81.78%; FB1: 83.99 234
OBJC: precision: 16.67%; recall: 44.44%; FB1: 24.24 24
PHEN: precision: 27.27%; recall: 30.00%; FB1: 28.57 11
PHYS: precision: 57.78%; recall: 72.22%; FB1: 64.20 45
PROC: precision: 77.11%; recall: 89.60%; FB1: 82.89 201

confusion_matrix_for_epoch_0022_in_valid_token_evaluation.pdf
f1_conll_vs_epoch_for_all_classes.pdf
f1_score_vs_epoch_for_all_classes.pdf