Review Radar: Context-Aware Sentiment Classification using RoBERTa on Multi-Domain Review Analysis

Abstract

The growing reliance on sentiment analysis for decision making in e-commerce, social media, and customer service is challenged by the complexity of sarcasm, which often misleads conventional models. This research presents ReviewRadar, a sarcasm aware sentiment analysis framework fine-tuned using the RoBERTa transformer architecture. Our dataset integrates a stable multi-domain corpus with targeted sarcasm augmentation, encompassing diverse review categories such as technology, fashion, travel, food delivery, and customer feedback. Unlike general sentiment models, ReviewRadar emphasizes the nuanced detection of sarcastic remarks, which are specially challenging due to contextual and lingual ambiguity. Preprocessing involved noise removal, tokenization, and stable class distribution, followed by fine tuning RoBERTa with optimized hyperparameters. trial evaluation demonstrated an overall accuracy exceeding 90%, with important improvement in sarcasm classification compared to baseline models such as DistilBERT. The proposed approach offers enhanced reliability for real-world sentiment monitoring systems, enabling businesses to better interpret user opinions and respond effectively.

Introduction

Traditional sentiment analysis systems struggle to detect sarcasm, especially in customer reviews, where ironic expressions can invert literal sentiment. This leads to misclassification and reduces the reliability of automated feedback interpretation.

Research Goal

To address this, the authors introduce ReviewRadar, a sentiment analysis framework fine-tuned specifically to detect sarcastic expressions across multiple review domains, using the RoBERTa transformer model.

Key Contributions

1. Multi-domain sarcasm-enriched dataset with balanced sentiment distribution.

2. Customized fine-tuning of RoBERTa for improved sarcasm recognition.

3. Performance comparison showing superiority over baseline models like DistilBERT.

4. A deployable framework for practical applications in real-time sentiment monitoring.

Methodology Overview

1. Dataset Construction

• ~15,000 customer reviews collected across domains (tech, fashion, travel, food, services).

• 2,000 additional sarcastic reviews added (mainly in negative/neutral categories).

• Preprocessing: Cleaning, tokenization (BPE), deduplication, and balanced partitioning (80/10/10 split).

2. Model Architecture

• RoBERTa-base used due to superior contextual understanding.

• Training used AdamW optimizer, cross-entropy loss, class weighting, dropout (0.1), and early stopping.

3. Evaluation Metrics

• Accuracy, precision, recall, F1-score, confusion matrix, and ablation testing.

Experimental Results

Additional Findings:

• Sarcasm augmentation improved F1-score by 3.1%.

• Best improvement in negative sentiment detection, where sarcasm is most common.

• Confusion matrix showed fewer misclassifications between neutral/negative classes compared to the baseline.

Implications

ReviewRadar’s enhanced sarcasm detection:

• Enables more accurate sentiment analysis in real-world applications.

• Is valuable for:

  • E-commerce feedback analysis

  • Social media monitoring

  • Customer service systems

  • Market research

Conclusion

This research successfully developed ReviewRadar, a sarcasm-aware sentiment classification framework that addresses critical limitations in existing sentiment analysis systems. By combining targeted dataset curation with optimized RoBERTa fine-tuning, the approach achieves 91.8% classification accuracy while demonstrating particular effectiveness in sarcastic content recognition. The methodology\'s strength lies in its multi-domain applicability and systematic approach to sarcastic pattern learning, making it suitable for diverse real-world deployment scenarios where accurate sentiment interpretation drives business decisions.

References

[1] Cambria, E., Poria, S., Bajpai, R., & Schuller, B. (2022). Affective Computing and Sentiment Analysis: A Review. Artificial Intelligence Review. https://doi.org/10.1007/s10462-022-10144-1 [2] Poria, S., Cambria, E., Bajpai, R., & Hussain, A. (2020). Multimodal Sentiment Analysis: A Survey. https://arxiv.org/abs/1708.07150 [3] Sajid, M., Majid, M., et al. (2023). Deep Learning Approaches for Sentiment Analysis. Applied Sciences, 13(10), 4550. https://doi.org/10.3390/app13104550 [4] González, J. A., & Cervantes, F. (2012). Feature-specific sentiment analysis in product reviews. CICLing 2012. https://cicling12.org (paper retrieved from conference proceedings) [5] Srivastava, N., et al. (2022). Sentiment Analysis of Customer Reviews Using a Hybrid Evolutionary SVM-Based Approach. https://doi.org/10.1007/978-981-13-0617-4_61 [6] Kaushik, M., et al. (2022). Machine Learning and NLP Techniques for Sentiment Analysis of E-commerce Data. Sustainability, 14(18), 10844. https://doi.org/10.3390/su141810844 [7] Kumar, A., & Sharma, A. (2021). Aspect-based Sentiment Analysis Using Attention based Deep Learning Models. Artificial Intelligence Review. https://doi.org/10.1007/s10462-021-09973-3 [8] Abbas, M., et al. (2023). A Survey on Sentiment Analysis Using Hybrid Deep Learning Models. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2023.119259 [9] Ajitha, R., et al. (2022). Sentiment Analysis for E-Commerce Platforms Using Python Libraries. Springer. https://link.springer.com/chapter/10.1007/978-981-16-7921-2_11 [10] Ali, W., et al. (2022). Machine Learning-based Sentiment Classification of Customer Reviews. Elsevier – Expert https://doi.org/10.1016/j.eswa.2022.118893 [11] Joshi, A., Sharma, V., & Bhattacharyya, P. (2017). Harnessing context incongruity for sarcasm detection. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, 757-762. [12] Liu, Y., et al. (2019). RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692. [13] Ghosh, H., Rakshit, G., & Naskar, S. K. (2021). Transformer-based approaches for social media sarcasm detection. IEEE Access, 9, 95509-95520. [14] Chauhan, M., Kumar, R., & Singh, P. (2022). Joint learning frameworks for sentiment and sarcasm detection tasks. Expert Systems with Applications, 192, 116294. [15] Misra, R., & Arora, P. (2019). Sarcasm detection using contextual word embeddings. Proceedings of the Workshop on Noisy User-generated Text, 135-143. [16] Bedi, K., Sharma, A., & Gupta, V. (2022). Data augmentation strategies for improving sarcasm detection performance. Journal of Computational Linguistics, 48(3), 587-612.

Copyright

Copyright © 2025 Rudresh HH, Kavyashree GJ. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.