XAI-SENTIFORMER: A DEEP TRANSFORMER FRAMEWORK FOR SEMANTIC SENTIMENT UNDERSTANDING IN LOW-RESOURCE MULTILINGUAL NLP

Babar Bakht Khan; Ghulam Muhy Ud Deen Raee; Anum Shafeeq; Mubasher Hussain Malik; Hamid Ghous

Authors

Babar Bakht Khan
Ghulam Muhy Ud Deen Raee
Anum Shafeeq
Mubasher Hussain Malik
Hamid Ghous

Abstract

Social media platforms generate massive volumes of multilingual content containing mixed languages, slang, abbreviations, emojis, and informal expressions, creating substantial challenges for conventional sentiment analysis systems. This study presents XAI-SentiFormer, a deep transformer-based framework designed for semantic sentiment understanding in low-resource multilingual NLP environments. The proposed framework evaluates the effectiveness of transformer-based language modeling, particularly multilingual Bidirectional Encoder Representations from Transformers (mBERT), for sentiment classification across English, Hindi, and Spanish social media data. A synthetic dataset comprising 10,000 multilingual social media posts was constructed to emulate realistic online communication patterns, including code-switching, noisy text, informal vocabulary, and emoji-rich expressions. The proposed transformer framework was compared with conventional machine learning and deep learning baselines, including logistic regression with TF-IDF features and Long Short-Term Memory (LSTM) networks. Performance evaluation was conducted using accuracy, precision, recall, F1-score, and receiver operating characteristic (ROC) analysis. Experimental results demonstrated that the proposed mBERT-based framework achieved a peak accuracy of 91%, outperforming baseline approaches by approximately 14–18%. The findings highlight the effectiveness of transformer architectures in capturing multilingual semantic relationships and contextual sentiment representations, particularly in low-resource and code-switched language settings. Furthermore, the study emphasizes the importance of explainable AI (XAI) mechanisms for improving interpretability and transparency in transformer-driven sentiment analysis systems. Overall, the proposed framework demonstrates strong potential for robust multilingual social media analytics and intelligent sentiment understanding in real-world NLP applications.