Multi-Label Text Classification Based on Multi-Task Dynamic Decoding

Recent advances in deep learning-based multi-label text classification methods have shown substantial progress in both theoretical research and practical applications. However, current studies rely exclusively on threshold-based decoding methods to interpret the model-generated label probabilities. This approach overlooks valuable information about the number of labels within a given text, leading to the prediction of numerous irrelevant labels and ultimately compromising the model's classification performance. To address this issue, we propose a novel approach, the DDA-RoBERTa model, which leverages the probability vectors for dynamically decoding labels in a multi-task architecture. Our model employs the RoBERTa model to generate document representation vectors. Simultaneously, we model two tasks: multi-label text classification and label quantity classification. By asynchronously and dynamically decoding the label probabilities generated by the multi-label text classification task using the predicted results of the label quantity classification task, we take into account latent label quantity information within the text, resulting in improved classification performance. Furthermore, we introduce a learnable loss function based on the homoscedasticity-uncertainty assumption to balance the training scale across multiple tasks. Our experimental results on English public datasets, such as Reuters-21578 and AAPD, demonstrate significant enhancements in the overall F1 value for the model's performance, with an increase of 0.30% and 1.94%, respectively, compared to the RoBERTa model. Additionally, recall improved by 1.76% and 4.38% for these datasets.