iREL at SemEval-2023 Task 9: Improving understanding of multilingual Tweets using Translation-Based Augmentation and Domain Adapted Pre-Trained Models

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2052–2057 July 13-14, 2023 ©2023 Association for Computational Linguistics iREL at SemEval-2023 Task 9: Improving understanding of multilingual Tweets using Translation-Based Augmentation and Domain Adapted Pre-Trained Models Bhavyajeet Singh, Ankita Maity, Pavan Kandru, Aditya Hari and Vasudeva Varma IIIT Hyderabad {bhavyajeet.singh, ankita.maity, siri.venkata, aditya.hari}@research.iiit.ac.in vv@iiit.ac.in Abstract This paper describes our system (iREL) for Tweet intimacy analysis shared task of the Se- mEval 2023 workshop at ACL 2023. Our sys- tem achieved an overall Pearson’s r score of 0.5924 and ranked 10th on the overall leader- board. For the unseen languages, we ranked third on the leaderboard and achieved a Pear- son’s r score of 0.485. We used a single multilingual model for all languages, as dis- cussed in this paper. We provide a detailed description of our pipeline along with multiple ablation experiments to further analyse each component of the pipeline. We demonstrate how translation-based augmentation, domain- specific features, and domain-adapted pre- trained models improve the understanding of intimacy in tweets. The code can be found at https://github.com/bhavyajeet/Multilingual- tweet-intimacy 1 Introduction Social media platforms have become a ubiqui- tous source of communication, with millions of users generating and sharing content daily. Among these, Twitter has emerged as a popular platform for sharing short messages or tweets, which can express a wide range of emotions, opinions, and sentiments.(Liew and Turtle, 2016) (Cislaru, 2015) Such tweets can also provide valuable insights into various social and political issues, making them a valuable resource for researchers and policymakers alike. Intimacy in language refers to the degree of emo- tional closeness or familiarity between individuals, which is often reflected in the choice of words, tone, and context of communication. (Wynne and Wynne, 1986). The degree of intimacy can vary based on the relationship between individuals, and it can significantly impact on the results of social interactions. In recent years, the analysis of intimacy in social media data has gained some attention (Pei and Ju- rgens, 2020a), as it can provide valuable insights into various aspects of human behavior, such as the formation of social networks, the spread of infor- mation, and the dynamics of online communities. However, the analysis of intimacy in social me- dia data is challenging, as it requires sophisticated methods for processing and interpreting large vol- umes of unstructured text data. Multilingual tweet intimacy analysis is even more challenging, as it involves the complexity of multiple languages and cross-cultural communication. The task (Pei et al., 2022a) aims at quantifying the intimacy of tweets from a total of 10 different languages in the form of a score between 1 to 5. This paper describes our system used for analysing the intimacy of tweets belonging to 10 different languages. We utilise domain-specific features and domain-adapted pre-trained models in order to im- prove the understanding of intimacy in tweets. We further utilise a translation-based data augmenta- tion pipeline which proves effective in significantly improving the scores for unseen languages. This paper also discusses …

Content truncated to 500 words