Application of PCA-Whitening Enhanced BERT for Chinese Semantic Understanding in Short Text Product Name Classification
Yifan Zhu, Hanping Hou, Wei Zhou, Ruiyang Qi, Qiuxia Zhang, Shiyu Jiang · 2024
This study examines the utility of PCA-Whitening enhanced BERT for Chinese short text product name classification within the context of extreme multi-class mapping. Although traditional machine learning models offer strong interpretability and rapid execution, they are ill-equipped to handle the complexities of semantic relationships inherent in extensive product catalogues. In contrast, BERT exhibits superior semantic comprehension, particularly in the context of polysemy and intricate phrasing. By employing keyword extraction techniques, the input data is effectively reduced in noise, thereby enhancing mapping accuracy. The PCA-Whitening technique optimizes the original embedding space, resulting in a more uniform vector distribution and improved computational efficiency. The experimental results demonstrate that the PCA-Whitening optimized BERT model attains a top-5 accuracy of 95% on brick level (the smallest classifications). Moreover, when the model was applied to a high-quality dataset, the top-1 accuracy reached 99%, which serves to illustrate the considerable impact that data quality has on model performance. The modular design of the model permits prospective enhancements, including multi-standard mapping and multilingual adaptability. In conclusion, this research contributes to the efficiency and effectiveness of product classification systems, with substantial implications for e-commerce, supply chain management, and digital standardization.