Implementation of Lightweight Machine Learning Models for Real-time Text Classification on Resource-Constrained Devices

Abstract

This paper addresses the growing need for implementing intelligent Natural Language Processing (NLP) systems on low-power, memory-limited devices such as Raspberry Pi, mobile phones, and IoT edge hardware. As edge computing and smart devices proliferate, there is an urgent need for more advanced NLP technology that does not require constant cloud access and is efficient in computing and provides results in real time. While deep learning and cloud-based models typically offer high text-classification accuracy and have demonstrated exceptional performance across a range of NLP tasks, they are often too resource-intensive for real-time deployment in constrained environments. To overcome these limitations, we explore a set of lightweight machine learning (ML) models—Multinomial Naive Bayes, Logistic Regression, and Decision Tree—to perform sentiment classification on a subset of the Amazon Reviews Polarity dataset. Following thorough data preprocessing and Term Frequency-Inverse Document Frequency (TF-IDF) vectorization, two optimization techniques are employed: feature selection via Chi-Squared tests and simulated post-training quantization. Our experimental results show that resource consumption can be substantially reduced, with minimal accuracy loss, thereby demonstrating feasibility for edge-based text analytics and offline functionality. We provide a detailed comparative analysis that highlights how classical ML models remain viable in scenarios where modern deep learning architectures cannot be efficiently deployed.