Zhegu at SemEval-2023 Task 9: Exponential Penalty Mean Squared Loss for Multilingual Tweet Intimacy Analysis

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 318–323 July 13-14, 2023 ©2023 Association for Computational Linguistics Zhegu at SemEval-2023 Task 9: Exponential Penalty Mean Squared Loss for Multilingual Tweet Intimacy Analysis Pan He1, Yanru Zhang1,2∗ 1University of Electronic Science and Technology of China, Chengdu 2Shenzhen Institute for Advanced Study, UESTC newtonysls@gmail.com yanruzhang@uestc.edu.cn Abstract We present the system description of our team Zhegu in SemEval-2023 Task 9 Multilingual Tweet Intimacy Analysis. We propose EPM (Exponential Penalty Mean Squared Loss) for the purpose of enhancing the ability of learning difficult samples during the training process. Meanwhile, we also apply several methods (frozen Tuning & contrastive learning based on Language) on the XLM-R multilingual lan- guage model for fine-tuning and model ensem- ble. The results in our experiments provide strong faithful evidence of the effectiveness of our methods. Eventually, we achieved a Pear- son score of 0.567 on the test set. 1 Introduction Text intimacy is considered as the essential compo- nent in emotional communication and social rela- tionships (Pei and Jurgens, 2020; Hovy and Yang, 2021; Sullivan, 2013). At the same time, intimacy exists in any language in the world. However, both data sets and methods of this area are still rare. Valuable research on the multilingual regression task is necessary to facilitate the sentiment analysis of language. To further investigate the intimacy in text, Pei et al. propose the MINT, a new Multi- lingual intimacy analysis dataset. In this case, we describe the model system of our team Zhegu in SemEval-2023 Task 9. Since the mean of intimacy is different from the "positive" or "negative" in sentiment analysis, it is difficult to have a clear definition. There exist many hard samples, which can not be learned by the model easily. Lin et al. propose Focal loss for learning the hard samples for model only in classi- fication. Therefore, we present EPM (Exponential Penalty Mean Squared Loss) for the regression task, which set a penalty term for hard samples dur- ing calculating the training loss to accelerate the ∗*Corresponding author model focus on those hard samples. We also con- struct a language-based contrastive loss function to draw closer the distance of the text in the same language and distinguish in the different languages. The test set in this task includes four new languages text data additionally, which do not exist in the training data. In order to achieve better Zero-shot inference on new languages data, we also adopt the frozen tuning to fine tune the model, which means freezing some parameters of the model for training. All in all, the main contributions of our paper are as follows. • We first propose EPM, which is an exponential penalty mean squared loss function for better learning hard sample in regression. • We introduce a language-based contrastive loss for the multilingual problem. • We apply frozen tuning for increasing the ca- pability of model generalization. All methods we mentioned above will be presented in detail in the Section 3 …

Content truncated to 500 words