Hybrid deep learning based POS tagger for Roman Urdu

Parts-of-Speech (POS) tagging is a highly encouraged research topic in the field of Natural Language Processing. POS entails numerous practical applications such as text indexing, information retrieval, corpus tagging for research, and linguistic work. This paper outlines multiple methods for part-of-speech tagging in Roman Urdu. Sufficient work and relevant required corpora are not available for Roman Urdu. We have identified that there are several parts-of-speech classes in the Urdu Language, with limited access to a well-annotated corpus. A manually verified corpus has been used to evaluate and report multiple methods for the said task. Our experiments deal with twenty-three unique parts-of-speech classes based on the contextual requirements of the Urdu Language. Our experiments include several methods built upon artificial neural networks, based on approaches such as multi-layered neural networks, feedback recurrent networks, and self-attention models. The corpus we used is not domain specific and covers several topics of Pakistani interest. Our experiments varied to a certain degree in the success demonstrated and outperformed numerous baseline models of machine learning and deep learning.