Adaptive Thresholding for Sentiment Analysis Across Online Reviews Based on BERT Model BERT-based Adaptive Thresholding for Sentiment Analysis

Online product reviews contain a wealth of information about user feedback and evaluations of products. This information can serve as a reference for merchants to improve their products and also as a guide for consumers looking to purchase goods. However, the existing evaluation systems mainly implement user comments combined with a rating score, where the score is an absolute value, and the comment text is the subjective opinion of the user. At the same time, for different categories of products, the language used in user reviews varies, making a generalized classification standard imprecise. Therefore, a method that identifies the optimal threshold for categorizing review texts of various types of products, taking into account their differences, is necessary. This task is called Adaptive Thresholding for Sentiment Analysis (ATSA) by this paper. A model based on BERT is built in this paper, using the ratings as labels and the text as input. It aims to enhance the accuracy of sentiment analysis for this specific type of text-product reviews-by finding an adaptive classification threshold suitable for the overall characteristics of the comment dataset. Tests carried out using a dataset sourced from JD.com show that this technique is highly efficient.