An Al-BERT-Bi-GRU-LDA algorithm for negative sentiment analysis on Bilibili comments

Abstract

The number of online self-learning users has been increasing due to the promotion of various lifelong learning programs. Unstructured commentary text related to their real learning experience regarding the learning process is generally published by users to show their opinions and complaints. The article aims to utilize the dataset of real text comments of 10 high school mathematics courses participated by high school students in the Bilibili platform and construct a hybrid algorithm called the Artificial Intelligence-Bidirectional Encoder Representations from Transformers (BERT) + Bidirectional Gated Recurrent Unit (BiGRU) and linear discriminant analysis (LDA) to crunch data and extract their sentiments. A series of tests regarding algorithm comparison were conducted on the educational review datasets. Comparative analysis found that the proposed algorithm achieves higher precision and lower loss rates than other alternative algorithms. Meanwhile, the loss ratio of the proposed algorithm was kept at a low level. At the topic mining level, the topic clustering of negative comments found that the barrage content was very messy and the complexity of the course content was generally reported by the students. Some problems related to videos were also mentioned. The outcomes are promising that the fundamental issues underlined by the students can be effectively resolved to improve curriculum and teaching quality.