Identifying positive and negative comments underneath technology videos based on Bert's model

Abstract

With the gradual popularization of the Internet and the increasing richness of content, more and more popular science videos appear in peoples vision. However, people have different attitudes towards them and cannot make explicit judgments. This paper used comments from some popular science videos on Bilibili as a dataset, trained the Bert model using parallel training methods such as (Distributed Data parallel) DDP, and selected different proportions of datasets (20%, 30%, and 40%) for training and testing data to objectively determine the positive and negative attitudes of people in comments. Research has found that different scale datasets have similar performance, with an 86% -87% accuracy rate. Besides, multi-card parallel processing can bring higher returns. The results help people comprehensively understand audience feedback and emotional preferences towards technical content. At the same time, it reveals the complex relationship between the emotions reflected in comments and various factors such as audience engagement, content strategies, and demographics.