Text Classification Based on Multilingual Back-Translation and Model Ensemble

Pretraining-and-finetuning have demonstrated excellent performance in the field of text classification. However, for downstream tasks, having an ample amount of training data remains crucial. When manually annotated training data is insufficient, a common approach is to perform data augmentation, and back translation using machine translation techniques is a commonly used text augmentation method. However, machine translation may introduce low-quality samples, thereby affecting the effectiveness of text augmentation. In this study, we use pre-trained models to generate sentence embeddings and employ sentence vector similarity comparison to automatically filter back-translated samples, aiming to enhance the quality of back-translation. Combining the ensemble learning approach, we enhance the classification accuracy by integrating multiple sets of diverse models. Our method achieved an accuracy of 90.7 in the 2023 CHIP evaluation task, securing the third-place position.