DAEPK:Domain-Adaptive Text Feature Enhancement Technology Integrating Prior Knowledge Domain In Text Classification

Abstract The scarcity of resources, lack of labeled texts, and insufficient corpora pose significant challenges to many specialized classification problems. It increases the difficulty of small language and domain-specific classification problems. In this paper, we propose a new approach to address this problem: by assigning multi-dimensional additional weights to words using external knowledge, thereby enhancing text features in low-resource domains. This is achieved by introducing the concepts of 'prior domains' and 'Adjusted Term Frequency Vectors (Adjust-TFs). Then, we propose a Domain Adaptive Text Feature Enhancement Method that combines prior domains and utilizes knowledge transfer to solve the classification problem. We conducted experiments across various data sizes on multilingual text categorization datasets. Experimental results demonstrate that our method notably enhances classification accuracy across diverse sample conditions, particularly in low-resource scenarios. This method still performs well in the few-shot setting, which, to some extent, alleviates the training problem of DNNs under this condition. Furthermore, we identify conditions that lead to improved performance in low-resource settings, providing valuable insights for future research.