Multi-Scale Domain Adaptive Transfer Based Sentiment Analysis for Morphological Rich Tamil Text

Sentiment analysis is the procedure of extracting information from given text about several functions, human beings, systems and facts. Sentimental analysis utilizes data mining and natural language processing (NLP) techniques to unearth, discover, retrieve and refine the information from the World Wide Web’s limitless textual information. The sentiment analyzers for several high-resource languages and certain Indic languages are completely developed. However, Tamil that is considered as a morphological rich language has not experienced these advancements. This paper proposes a novel method called, Multi-Scale Attention and Domain Adaptive Transfer Learning (MSA-DATL) for performing intelligent sentiment analysis as positive, negative or neutral. The proposed method includes three complementary components. They pre-processing, keyword extraction and classification for performing morphological rich sentiment analyses using Tamil text. First, Multilingual Text-To-Text-Transfer-Transformer-based Pre-processing is performed. Following which Multi-Scale Attention Mechanism-based Keyword Extraction using Sliding Window for short range contextual dependencies and Sparse Attention function for long range contextual dependencies are done. Finally, Domain Adaptive Transfer Learning-based Classifier is applied for efficient sentiment analysis. Experimental results on the dataset used in our work show that the proposed MSA-DATL method achieves 29% accuracy, with weighted-average precision, recall of 29%, 28% respectively, significantly outperforming baseline methods. These findings illustrate the robustness and efficiency of our method making a notable contribution to intelligent analysis of morphological rich Tamil text.