Arizonans at SemEval-2023 Task 9: Multilingual Tweet Intimacy Analysis with XLM-T

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1656–1659 July 13-14, 2023 ©2023 Association for Computational Linguistics Arizonans at SemEval-2023 Task 9: Multilingual Tweet Intimacy Analysis with XLM-T Nimet Beyza Bozdag, Tugay Bilgis, Steven Bethard University of Arizona {nbbozdag, tbilgis, bethard}@arizona.edu Abstract This paper presents the systems and approaches of the Arizonans team for SemEval 2023 Task 9: Multilingual Tweet Intimacy Analysis. We finetune the Multilingual RoBERTa model trained with about 200M tweets, XLM-T. Our final model ranked 9th out of 45 overall, 13th in seen languages, and 8th in unseen languages. 1 Introduction Intimacy is an integral part of human relationships and an essential component of language (Pei and Jurgens, 2020). Pei and Jurgens also propose that intimacy can be modeled with computational meth- ods and that analyzing textual intimacy can po- tentially reveal information about social norms in different contexts. With social media taking an increasingly large part in people’s lives, intimacy has found itself a new platform for display. In these digital social con- texts’, Twitter is an important source for analyzing textual intimacy. To further promote computational modeling of textual intimacy, Pei et al. (2023) make use of publicly available Twitter data to create and annotate a textual intimacy dataset called MINT from the following 6 languages: English, Spanish, French, Portuguese, Italian, and Chinese. They also annotate a smaller set of tweets in Dutch, Hindi, Ko- rean, and Arabic to allow for zero-shot prediction experiments. SemEval 2023 Task 9: Multilingual Tweet Intimacy Analysis (Pei et al., 2023), asks participants to predict the intimacy of tweets in the aforementioned 10 languages. In this paper, we explain our approach to inti- macy analysis: finetuning XLM-T, the Multilingual RoBERTa model trained with about 200M tweets. To improve performance we experimented with different hyperparameters and changed the train- ing data label distribution by duplicating tweets with certain labels. We discuss the issues we ran into while working on our model, such as the im- balanced range of intimacy scores in the provided training dataset. Our final submission ranked 9th out of 45 overall, 13th in seen languages, and 8th in unseen languages. We release our code at https: //github.com/beyzabozdag/tweet-intimacy. 2 Background Intimacy is an essential part of language, however resources on textual intimacy analysis remain rare (Pei et al., 2023). The first textual intimacy dataset was annotated by Pei and Jurgens (2020) with 2,397 English questions collected mostly from social me- dia posts and fictional dialogues. However, Pei et al. (2023) claim that models trained over phrases in the question structure might not generalize well to text in other forms of languages. To work on a more generalizable dataset and pro- mote computational modeling of textual intimacy, Pei et al. (2023) proposed their new dataset MINT, which is a Multilingual intimacy analysis dataset covering 13,384 tweets in 10 languages including English, French, Spanish, Italian, Portuguese, Ko- rean, Dutch, Chinese, Hindi, and Arabic. Each tweet in the dataset has a label between 1 and 5 as an …

Content truncated to 500 words