Lazybob at SemEval-2023 Task 9: Quantifying Intimacy of Multilingual Tweets with Multi-Task Learning

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 927–933 July 13-14, 2023 ©2023 Association for Computational Linguistics lazybob at SemEval-2023 Task 9: Quantifying Intimacy of Multilingual Tweets with Multi-Task Learning Mengfei Yuan The Ohio State University Sunwoda Electronic Co., Ltd. ymf924@gmail.com Cheng Chen Ping An Insurance oscarhscc@gmail.com Abstract This study presents a systematic method for analyzing the level of intimacy in tweets across ten different languages, using multi-task learn- ing for SemEval 2023 Task 9: Multilingual Tweet Intimacy Analysis. The system begins with the utilization of the official training data, and then we experiment with different fine- tuning tricks and effective strategies, such as data augmentation, multi-task learning, etc. Through additional experiments, the approach is shown to be effective for the task. To enhance the model’s robustness, different transformer- based language models and some widely-used plug-and-play priors are incorporated into our system. Our final submission achieved a Pear- son R of 0.6160 for the intimacy score on the official test set, placing us at the top of the leader board among 45 teams. 1 Introduction Intimacy is a fundamental social aspect of language. Understanding the dynamics of intimacy in con- versation is important for a variety of practical applications. Research on intimacy in conversa- tions between people, particularly in the context of chance encounters, such as the phenomenon of the "stranger on the train," can provide valuable in- sights into human communication and relationships (Griffin, 2006; West et al., 2010; Farber, 2003). Studying the content on different social platform can help researchers to better understand how peo- ple establish and develop relationships through con- versation. Social workers can also benefit from an understanding of how intimacy is established and maintained in conversation to help clients build stronger relationships with others (Labov, 1972; Brown and Levinson, 1978). The multilingual tweet intimacy task was hosted by the University of Michigan and Snap Inc. This task focuses on predicting the intimacy of tweets in 10 languages (Pei et al., 2022), where 6 lan- guages with training data (English, Spanish, Por- tuguese, Italian, French, Chinese), and 4 zero-shot languages (Korean, Arabic, Hindi, Dutch). Our submission achieved the first place overall score in these 10 languages. In this paper, we demonstrate the following con- tributions: (1) We designed a novel multi-language and multi-task training framework that can simul- taneously train data from all languages. (2) Based on 10 million crawled Twitter, book, movie, and Reddit data in various languages, we conducted domain pre-training of multiple pre-training mod- els to enhance the model’s prediction ability in that domain. (3) During the training process, we also adopted contrastive learning strategies such as group-layerwise learning rate decay (GLLRD), ad- versarial weight perturbation (AWP) to enhance the semantic understanding and robustness of language models. (4) For zero-shot language prediction, we tried bidirectional translation data augmentation to enhance the model’s prediction ability for small languages in this task. 2 Background Sociologists and linguists have conducted many studies on how people use language to measure the relationships …

Content truncated to 500 words