Research on BERT-based deep multilayer fusion Chinese short text classification method

The categorization of brief Chinese texts, a critical area for extracting insights from data with limited information content, presents unique challenges such as limited word count, ambiguity, and non-standardized information. These factors complicate the extraction and representation of textual features. This study introduces the BERT-based BRLC (BERT Recurrent Layer Composition) model, customized for brief text categorization assignments. The model employs BERT as its foundation, utilizing the Transformer architecture to encode input text and produce a semantically rich vector representation. The BRLC component then captures both global and local sequential features of the text, culminating in classification outcomes through a non-linear transformation process. Evaluated against contemporary deep learning models, the BRLC model demonstrates superior classification performance and efficiency. Utilizing the THUNews dataset, our model attains a precision level of 94.92%, outperforming the original BERT by 2.19%. Comparative experiments confirm the BRLC model's effectiveness in enhancing the accuracy of short text classification tasks.