ZBL2W at SemEval-2023 Task 9: A Multilingual Fine-tuning Model with Data Augmentation for Tweet Intimacy Analysis

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 770–775 July 13-14, 2023 ©2023 Association for Computational Linguistics ZBL2W at SemEval-2023 Task 9: A Multilingual Fine-tuning Model with Data Augmentation for Tweet Intimacy Analysis Hao Zhang∗, Youlin Wu∗, Junyu Lu, Zewen Bai, Jiangming Wu, Hongfei Lin, Shaowu Zhang† School of Computer Science and Technology, Dalian University of Technology, China {zh373911345,wuyoulin,dutljy,986427968, wjm374733751}@mail.dlut.edu.cn {hflin,zhangsw}@dlut.edu.cn Abstract This paper describes our system used in the SemEval-2023 Task 9 Multilingual Tweet Inti- macy Analysis. There are two key challenges in this task: the complexity of multilingual and zero-shot cross-lingual learning, and the diffi- culty of semantic mining of tweet intimacy. To solve the above problems, our system extracts contextual representations from the pretrained language models, XLM-T, and employs vari- ous optimization methods, including adversar- ial training, data augmentation, ordinal regres- sion loss and special training strategy. Our sys- tem ranked 14th out of 54 participating teams on the leaderboard and ranked 10th on predict- ing languages not in the training data. Our code is available on Github 1. 1 Introduction A theoretical approach to understanding the level of intimacy, human emotion, and social connec- tions in communication is called intimacy in lan- guage. Quantifying the intimacy expressed in lan- guage plays a key role in revealing important social norms in various context (Pei and Jurgens, 2020). SemEval-2023 Task 9 aims at mining intimacy in tweets. Pei et al. (2023) provides a new multi- lingual intimacy analysis dataset covering 13,372 tweets in 10 languages. Text intimacy is labeled on a continuum scale, in the range of [1, 5]. Among them, the participants are only given the training set in six languages (English, Spanish, Italian, Por- tuguese, French, and Chinese), and the model per- formance will be evaluated on the test set of ten languages, including 4 languages not in the training set (Hindi, Arabic, Dutch and Korean). Given that the SemEval-2023 Task 9 is a com- plex challenge that integrates language transfer and *These authors contributed equally to this work. †Corresponding author. 1https://github.com/HelloHank-Hub/Multilingual-Tweet- Intimacy-Analysis semantic mining, we propose the intimacy predic- tion system based on pre-trained representations and the fusion of multiple optimization methods. First, we chose XLM-T (Barbieri et al., 2021) as the encoder of the model, which is pre-trained on millions of tweets in multiple languages. Then, ac- cording to the form of annotated dataset, we change the calculation method of model loss and adjust the training strategy of the model to make it more suitable for the intimacy prediction of zero-shot language. Finally, we introduce adversarial train- ing and data augmentation to enhance the model generalization. Our system obtained Pearson cor- relation scores of 0.7029 on seen language section and 0.4359 on unseen language section. 2 Related Work Intimacy computing is an emerging field of natural language processing. Pei et al. (2023) propose to encode social messages of intimacy through topics and other more subtle cues such as swear words. A new computational framework for studying the expression of intimacy …

Content truncated to 500 words