Improved Deep Persian Named Entity Recognition
Mohammad Hadi Bokaei, Maryam Tayefeh Mahmoudi · 2018
Named Entity Recognition (NER) is a challenging task specifically for low resource languages like Perisan. In this work we evaluate the result of deep learning structures in NER. Specifically we use the only publicly available corpus (ArmanPersoNER) to train a model in which features are extracted using recurrent and convolutional neural networks. Extracted features vectors are then passed to a simple fully connected network and finally a conditional random field layer is used to find the best sequence tag for the input word sequence. Results show statically significant improvement over the available results on the corpus. Specifically the best structure has 81.50% and 76.79% word level and phrase level f1 score respectively. We also conduct an error analysis on the output of the best model and show that this corpus also have much errors and must be reviewed to have a more accurate one to further improve the performance.