Ontology based Semantic Analysis framework in Sindhi Language

Abstract

 Sentiment analysis, identifying polarity information (Positive, Negative, or Neutral sentiment) from textual data, is a crucial aspect of natural language understanding. However, its implementation in low resource languages like Sindhi presents significant challenges due to linguistic diversity and a limited amount of labeled data. This work addresses these challenges by proposing an ontology-driven sentiment analysis framework that integrates domain-specific ontological knowledge with the power of the Distil-BERT model for efficient sentiment classification. We constructed a custom Sindhi sentiment dataset, comprising 123 sentences annotated with three sentiment classes: Positive, Negative, and Neutral. The Distil-BERT model was employed for tokenization and sequence classification, leveraging its efficiency and adaptability for resource-constrained settings. Using Pytorch and the Hugging Face Transformers library, we trained the model with supervised pre-training arguments using the Trainer API. Additionally, a domain-specific ontology was developed to capture complex linguistic relationships and enrich the model’s semantic understanding, enabling it to handle diverse sentiment- bearing expressions effectively. Experimental results highlight the efficacy of our approach. The ontology-driven model achieved an impressive accuracy of 93%, significantly outperforming the baseline model, which achieved 82%. This improvement underscores the importance of integrating ontological knowledge, particularly in addressing the nuances of low-resource languages like Sindhi. Performance evaluation metrics, including precision, recall, and F1 Score, further validate the superior performance of the ontology-driven framework. This study presents a robust solution for sentiment analysis in Sindhi, laying the groundwork for future research in Natural Language Processing (NLP) for low-resource languages. Expanding the ontology to include more sentiment contexts and exploring hybrid deep learning approaches for sentiment classification offer promising directions for future work.