Tsingriver at SemEval-2023 Task 10: Labeled Data Augmentation in Consistency Training

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 782–786 July 13-14, 2023 ©2023 Association for Computational Linguistics Tsingriver at SemEval-2023 Task 10: Labeled Data Augmentation In Consistency Training Yehui Xu Yunnan University yehuixu24@163.com Haiyan Ding* Yunnan University teidhy@163.com Abstract Semi-supervised learning has promising perfor- mance in deep learning, one of the approaches is consistency training on a large amount of unlabeled data to constrain model predictions to be invariant to input noise. However, The degree of correlation between unlabeled data and task objective directly affects model pre- diction performance. This paper describes our system designed for SemEval-2023 Task 10: Explainable Detection of Online Sexism. We utilize a consistency training framework and data augmentation as the main strategy to train a model. The score obtained by our method is 0.8180 in subtask A, ranking 57 in all the teams. 1 Introduction Sexism is an increasingly serious network problem. It may harm targeted women, make cyberspace inaccessible and unpopular, and perpetuate social asymmetry and injustice. Explainable Detection of Online Sexism task targets to predict whether a sentence is sexism (Jane, 2014). A small amount of labeled data limits the improve- ment of supervised learning model performance. In contrast, semi-supervised and self-supervised learning methods can learn from massive unlabeled data (Chen et al., 2020), even better than supervised learning. Bidirectional Encoder Representation from Transformers(BERT) and Robustly optimized BERT approach(RoBERTa) (Zhuang et al., 2021) are classic models that are trained by unlabeled data in unsupervised learning tasks. In this work, we aim to apply a semi-supervised learning method to accomplish the task, using labeled data as the train data of supervised training and augmented labeled data as the train data of unsupervised training(consistency training), respectively. Due to the difficulty of selecting domain-relevance unlabeled data in consistency learning and the fact that the labeled data is domain-relevance data that matches with task objective, we decide to augment labeled data instead of unlabeled data that mismatch the distributions of labeled data. Improving model performance is as important as ensuring model stability (Gunel et al., 2020), which is why we choose to use consistency learn- ing. Our method guarantees the same prediction result for the same sentence expressed with differ- ent words in a certain probability. In other words, it improves the model’s performance in detecting synonymous sentences. To not lose BERT’s origi- nal performance as much as possible, we combine two strategies: consistency training and synonym replacement method based on TF-IDF to fine-tune BERT, instead of directly retraining BERT (Gha- landari et al., 2022). Compared with the super- vised learning method training our baseline model in this paper, our main training strategies have significantly improved model performance from 0.8004 to 0.8180. We have released the code at https://github.com/vincent-hyx/task_10. 2 Background There are two main directions of research work, one is consistency training (Tarvainen and Valpola, 2017), and the other is unsupervised data augmen- tation(UDA) (Xie et al., 2020). The consistency training method relies on unmarked data training to overcome the weakness of …

Content truncated to 500 words