Year: 2025
Type: conference-paper
Abstract: To improve the accuracy and generalizability of current models for text sentiment detection, this paper presents a self-learning framework for adaptive sentiment prediction. The framework uses sentiment text as the research object and combines it with the Transformer-based BERT model. It improves the quality of pseudo-labeling by changing the confidence threshold dynamically and makes better use of unlabeled data. This method provides a good solution for small-sample learning when labeling is expensive but there is plenty of unlabeled data. The framework is based on the BERT base model and incorporates Twitter social media big data. It achieves 91.8% test accuracy on the SemEval-2018 sentiment dataset and high confidence distribution on the Twitter unlabeled dataset. The experimental results demonstrate that, in comparison with the conventional fixed-threshold pseudo-labeling method, the adaptive method proposed in this paper exhibits significant advantages and signifies considerable application value in real-world scenarios where labeling resources are limited.
Authors: Beiyue Chen, Ziyun Huang, Shangjiang Lin, Wenhui Liao
Institution: Zhuhai College of Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414931864