An Integrated Deep Learning Framework Combining LSTM-CRF, GRU-CRF, and CNN-CRF with Word Embedding Techniques for Arabic Named Entity Recognition

Abstract

Named entity recognition (NER) is the main function of natural language processing (NLP) and has many applications. Arabic NER systems aim to identify and classify Arabic NEs in Arabic text, which provide unique problems due to the language's complex morphology and syntactic structures. This paper provides an integrated deep learning system that incorporates three deep learning architectures—LSTM-CRF, GRU-CRF, and CNN-CRF—as well as three word embedding techniques: GloVe, Word2Vec, and FastText, all trained on Arabic corpus. To develop NER state-of-the-art in Arabic language, the present paper proposed a 3-stage process of pre-processing, feature extraction, and a combination of various deep network schemes. In the preprocessing section, operations such as removing irrelevant words, correcting words, etc. will be used to improve the system's efficiency. In the feature extraction section, three-word embedding methods, Glove, word2vec, and fasttext, which are trained with Arabic texts, are used, and finally, three LSTM-CRF, GRU-CRF, and CNN-CRF models are trained with each word embedding, and the results they are combined. Experimental results on benchmark dataset, ANERcorp show that our methodology is effective, with an accuracy of 94.39%, which outperforms other cutting-edge methods. However, combining multiple deep learning models with word embeddings increases computational complexity and resource requirements, potentially complicating implementation in resource-constrained contexts. Future efforts will concentrate on optimizing the framework to lower computational costs while keeping good performance.