Exploring Computational Models for Syntactic Analysis in Sindhi Part-of-Speech Tagging

Abstract

Identifying part-of-speech tags is basically a crucial aspect of language tagging. It facilitates the introduction of very important applications, for example, machine translation and sentiment analysis. Low-resource languages typically get deprived of these resources mainly because of their complex morphology, scarcity of annotated datasets, and difficulties they present when writing from right to left. This paper tries to tackle these issues by thoroughly going over the different POS tagging methods to set a reliable benchmark for performance assessment of the Sindhi language. We conducted experiments using a well-balanced and standardized dataset with five different models, i.e., Hidden Markov Models (HMM), Bidirectional Long Short-term Memory (BiLSTM), Naive Bayes, Support Vector Machines (SVM), and Conditional Random Fields (CRF). Results revealed that Naive Bayes was the best among others, as it used morphological suffix patterns effectively to reach a level of accuracy of 97%. CRF and HMM both followed closely behind Naive Bayes and secured accuracy results of 92.29% and 93.37%, respectively. SVM encountered difficulties with repetitive tags, resulting in a lower accuracy of 84.64%, which gave it a lower accuracy of 84.64%, the The BiLSTM model was capable of using contextual information, thus reaching the accuracy of 91.32%. These results indicate that, in fact, in the case of languages with regular morphological patterns, simple statistical methods might be highly effective even if neural networks were more advanced. This paper lays a strong groundwork for the future progress of natural language processing for Sindhi and other minor languages.