Combination of Synonym Replacement Data Augmentation and Transformer Architecture with Swish Activation for Sentiment Classification in E-Commerce Product Review Data
Arjun Elvas Janggiara

Sriwijaya University

Abstract

In the digital era, e-commerce platforms are a primary source for consumer product reviews, which contain valuable sentiment reflecting customer satisfaction. Sentiment analysis is crucial for extracting these opinions to provide strategic insights for businesses. However, this task faces challenges, including limited labeled data and linguistic diversity. Data augmentation, particularly techniques like synonym replacement, has emerged as an effective solution to enrich training data without manual relabeling. Concurrently, Transformer architectures like BERT have revolutionized Natural Language Processing (NLP) due to their superior ability to capture deep, bidirectional contextual meaning, though they often demand high computational resources.
This research aims to develop a sentiment analysis model for e-commerce product reviews by combining synonym replacement data augmentation with a Transformer architecture that utilizes the Swish activation function. The study seeks to investigate the impact of synonym replacement on training data variety and model performance, evaluate the effectiveness of the Swish activation function in improving classification accuracy compared to standard functions, and determine the extent to which this combined approach can produce a more accurate and efficient sentiment analysis model. The research is confined to publicly available e-commerce reviews in Indonesian or English, focusing solely on synonym replacement for augmentation and a customized Transformer with Swish activation. Performance will be evaluated using standard classification metrics. The findings are expected to contribute academically to the development of more efficient NLP models and offer a practical, resource-conscious solution for businesses to automatically and accurately understand customer sentiment.

Keywords:

Topic: Mathematics and Applications