A DEEP LEARNING MODEL FOR A MULTI LEVEL ARABIC TEXT POSTAGGING

Abstract

A portion of Assigning grammatical categories to every word in a sentence is the goal of speech (POS) tagging, a basic task in natural language processing (NLP). The multi-level POS tagging method for Arabic, a morphologically rich and syntactically complex language, is presented in this project. Our approach focuses on two levels of annotation: XPOS+FEATS (language-specific POS tags and morphological features) and UPOS (universal POS tags). We apply and contrast three models: a refined XLM-ROBERTA transformer model, a Conditional Random Field (CRF) model based on handcrafted features, and a BiLSTM model that learns from sequential patterns. The Arabic-PADT corpus, which follows the Universal Dependencies framework, was used to train and assess all models. According to experimental findings, deep learning models perform noticeably better than conventional methods. Because of its deep contextualized embeddings and subword tokenization capabilities, the XLM-ROBERTA model performs best at both tagging levels and is able to capture the morphological richness and syntactic structure of Arabic. With possible uses in information extraction, syntactic parsing, and machine translation, this research advances the creation of more precise Arabic language processing systems.