Kurdish Adaptive Hybrid Sentiment Analysis (KA-HSA)

Abstract

Sentiment analysis is important to capture user sentiment but is underdeveloped in low-resource languages like Kur-dish, which suffers from dialectal diversity, double scripts, and limited annotated data. Sentiment analysis for Kurdish is limited due to linguistic challenges and a lack of sufficient annotated data. In this paper, we introduce Kurdish Adaptive Hybrid Sentiment Analysis (KA-HSA). The XLM-Roberta model, with CNN-BiLSTM and a meta-classifier incorporated, has been applied to many applications within natural language processing; however, this research targets its application and adaptation to the Kurdish language. A main focus of this work is to overcome obstacles such as dialectal variability, the use of two different scripts, and a lack of adequate data issues not properly addressed with current approaches to sentiment analysis. A novel deep learning framework that integrates a multilingual transformer (XLM-Roberta) with a CNN-BiLSTM hybrid model. These are integrated using weighted voting and a meta-learning classifier to enhance sentiment prediction. KA-HSA was trained and evaluated on the newly developed KurdiSent corpus, which contains manually labeled Kurdish texts in Sorani and Kurmanji dialects. The model achieved 89.5% accuracy and 89.1% F1-score, outperforming the state-of-the-art models: the Sentiment Analysis in Low-Resource Contexts study (by 4.4% accuracy and 4.7% F1) and the KMD Emotion Dataset Study (by 14.1% accuracy and14.2% F1). The framework also has a strong preprocessing pipeline that is specifically designed to address the linguistic features of Kurdish. The outputs indicate that KA-HSA can accurately encode both syntactic and semantic features and provide a reproducible way of conducting sentiment analysis on various low-resource languages.