Claim Stance Classification Optimized by Data Augment

Abstract

As online fora increasingly become the main media for argument and debate, the automatic processing of such data is rapidly becoming more and more important. Stance classification, which aims to classify the stance of the claims towards the given topic, can be applied in many application areas such as users' feelings about services and products. We propose a ensemble model for stance classification with data augment for small sample scenarios, multi-sample dropout for low training speed scenarios, focal loss for imbalance sample scenarios, pseudo labels for self-supervised training scenarios, adversarial training for low robustness scenarios, and all the above can be used in normal scenarios. Besides, the ensemble model is composed of task-specific RoBERTa and MacBERT, which can make more reasonable predictions. We used dataset from NLPCC to validate the model and it worked well.