Enhancing Low-Resource Sentiment Analysis: A Transfer Learning Approach

Abstract

The identification and extraction of subjective information from text, known as sentiment analysis, has seen advancements in employing cross-lingual approaches. However, the effective implementation and evaluation of sentiment analysis systems necessitate language-specific data to account for diverse sociocultural and linguistic variations. This paper outlines the process of collecting and annotating a dataset for sentiment analysis in Central Kurdish. We investigate classical machine learning and neural network-based techniques for this purpose. Furthermore, we adopt a transfer learning approach to enhance performance by leveraging pre-trained models for data augmentation. Our results demonstrate that despite the challenging nature of the task, data augmentation contributes to achieving high F1 scores and accuracy.