A simple interpolation-based data augmentation method for implicit sentiment identification

Implicit sentiment identification tasks occupy an important place in the field of natural language processing (NLP). In recent years, deep learning models have made significant progress in most tasks in NLP, but implicit sentiment identification still faces two major challenges. First, implicit sentiment texts are more complex, and it is difficult to understand the semantics due to the lack of sentiment words compared to explicit sentiment texts. Second, implicit sentiment text is prone to overfitting the model due to the lack of training data, which requires a lot of specialized domain knowledge for the label annotation work, resulting in a lack of training data. To address these issues, we propose a simple but effective interpolation-based data augmentation method for implicit sentiment identification (ISIMIX). The method performs interpolation operations in hidden space without label mixing, which effectively alleviates the data scarcity problem. In addition, ISIMIX interpolates in hidden space to solve the validity problem of previous data augmentation and better perform implicit sentiment identification. Our approach consists of first generating multiple augmentation samples from the raw data through a series of underlying augmentation techniques and blending these augmentation samples. Subsequently, the original data is combined with the generated augmented data, while ensuring that the blending weight of the original data is higher than that of the augmented data, to find a balance between moderate perturbation and avoiding the introduction of too much noise. In order to reduce the discrepancy between the predicted distributions of the raw and augmented data, we introduced a Jensen-Shannon divergence regularization term. Experimental results on three publicly available implicit sentiment datasets show that our method exhibits superior performance compared to both other data augmentation techniques and mainstream text classification methods. These results fully validate the effectiveness and wide application potential of ISIMIX in implicit sentiment identification tasks.