Using LSTM and GRU With a New Dataset for Named Entity Recognition in the Arabic Language
This addresses the lack of structured data for Arabic NER, an incremental contribution to domain-specific NLP.
The authors tackled Named Entity Recognition (NER) for Arabic by creating a new structured dataset of over 36,000 records and using LSTM and GRU models, achieving approximately 80% accuracy.
Named entity recognition (NER) is a natural language processing task (NLP), which aims to identify named entities and classify them like person, location, organization, etc. In the Arabic language, we can find a considerable size of unstructured data, and it needs to different preprocessing tool than languages like (English, Russian, German...). From this point, we can note the importance of building a new structured dataset to solve the lack of structured data. In this work, we use the BIOES format to tag the word, which allows us to handle the nested name entity that consists of more than one sentence and define the start and the end of the name. The dataset consists of more than thirty-six thousand records. In addition, this work proposes long short term memory (LSTM) units and Gated Recurrent Units (GRU) for building the named entity recognition model in the Arabic language. The models give an approximately good result (80%) because LSTM and GRU models can find the relationships between the words of the sentence. Also, use a new library from Google, which is Trax and platform Colab