Copyright © 2026 Authors retain the copyright of this article. This article 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.

@article{208465,
        author = {Ms. Anjali Jawale and Mr. Shreyash Dodekar and Mrs. Shital Pashankar},
        title = {Robustness Analysis Of Lightweight Machine Learning Models For Sentiment Classification Under Noisy Text},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {no},
        pages = {203-211},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=208465},
        abstract = {Nowadays, sentimental analysis is an important application of natural language processing and widely used to understand whether a review expresses a positive or negative opinion. Many sentiment classification models are trained using clean data. Real-world reviews often contain spelling mistakes, missing or repeated characters, abbreviations and informal language. A model which executes well on clean data may lose accuracy on noisy text, making robustness essential for evaluation. Previous research has mainly focused on the robustness of neural and transformer-based NLP models, while lightweight machine-learning models have received less attention. These traditional models are simple, fast, resource-efficient and easier to interpret compared to complex deep-learning models. Therefore, this study focuses on evaluating the robustness of lightweight machine-learning classifiers when handling noisy text. 
The study assesses Multinomial Naive Bayes, Logistic Regression, Linear SVM using TF-IDF representation for fair comparison. The Amazon Cell-Phone Reviews dataset is used to classify reviews into positive and negative sentiments. To simulate real-world text, five types of noise—deletion, insertion, substitution, swapping and abbreviation are introduced at 10%, 20% and 30% severity levels. The noise is programmatically added while keeping the original sentiment labels unchanged. The models are evaluated on noisy and clean data to measure their sturdiness. Performance is examined using precision, recall, accuracy, performance degradation and F1 score. The study addresses which model remains most robust and which type, level of noise causes greatest performance loss. The study helps to determine the level and type of noise which leads to highest performance loss. This finding helps developers and researchers in choosing lightweight model for sentiment analysis.},
        keywords = {Machine Learning, Noisy Text, Robustness, Sentiment Analysis, NLP},
        month = {September},
        }