Boosting Arabic Named Entity Recognition with K-Fold Cross Validation on LSTM and Bi-LSTM Models

Abstract

Named-Entity-Recognition (NER) is one of the most important Information-Extraction (IE) use cases, which is used to improve the performance of Natural Languages Processing (NLP) tasks, such as Relation-Extraction (RE), Question-Answering (QA). Recently, Arabic NER is tackled in different ways by researchers. In this study, we assess the performance of two widely used models, namely, LSTM and Bi-LSTM on the NER task in the Arabic language and perform a comparative study between these models. In contrast to the traditional data partition technique widely used during the training, we employ the technique of k-fold cross-validation to improve the performance of each model. The experimental results reveal that the performance of all models is improved when k-fold cross-validation is applied. Additionally, according to our experiment results, the Bi-LSTM model outperforms the LSTM model in terms of our evaluation metric. We achieve the best F1 score of 94.17% with CNN-Bi-LSTM-CRF. An ablation study on k-fold cross-validation demonstrates that the F1 score increased from 87.28 to 94.17%.