Arabic-NER

Model uses Named Entity Recognition concept to tag words in arabic text.

Architecture

It consists of Bi-directional GRU units(One forward and the other is backward) and a CRF layer ^{referenced from https://arxiv.org/pdf/1508.01991v1.pdf}

Specifications

Model is trained on ANERCorp dataset.^more. And uses FastText's Arabic vectors for word embedding.

No. epochs: 20

Accuracy: 94.2%

Classification report:

              precision    recall  f1-score   support

         LOC       0.99      0.99      0.99     11055
        PERS       0.74      0.65      0.69       824
         ORG       0.64      0.46      0.54       503
        MISC       0.63      0.38      0.47       237
 avg / total       0.95      0.94      0.94     12619

F1_score: 95.0%

Sample

Input

ماذا يفعل طلال عبد الهادي في دبي بعد ما رجع من برلين؟ كان يعمل هناك في شركة فولكسفاجن، صحيح؟

Output