Billy-Batson at SemEval-2023 Task 5: An Information Condensation based System for Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 177–183 July 13-14, 2023 ©2023 Association for Computational Linguistics HULAT at SemEval-2023 Task 9: Data Augmentation for Pre-trained Transformers Applied to Multilingual Tweet Intimacy Analysis Isabel Segura-Bedmar Computer Science Department, Universidad Carlos III de Madrid, Leganés, Spain isegura@inf.uc3m.es Abstract This paper describes our participation in SemEval-2023 Task 9, Intimacy Analysis of Multilingual Tweets. We fine-tune some of the most popular transformer models with the training dataset and synthetic data generated by different data augmentation techniques. Dur- ing the development phase, our best results were obtained by using XLM-T. Data aug- mentation techniques provide a very slight im- provement in the results. Our system ranked in the 27th position out of the 45 participat- ing systems. Despite its modest results, our system shows promising results in languages such as Portuguese, English, and Dutch. All our code is available in the repository https: //github.com/isegura/hulat_intimacy. 1 Introduction The Intimacy Analysis of Multilingual Tweets (Pei et al., 2023) is a very novel task whose goal is to estimate the level of intimacy of a text. If we can detect the degree of intimacy of a text, this could help us better understand the social norms that exist within a culture or society. Furthermore, from a computational point of view, this task will allow us to assess whether current computational models are capable of identifying the intimacy level of a text, distinguishing between texts with very intimate content and texts without any intimate information. Although the automatic analysis of textual in- timacy in language could be very beneficial for politicians, governments, sociologists, anthropol- ogists, companies, and so on, very few research efforts have been devoted to this challenging task to now (Pei and Jurgens, 2020). Another important aspect of this task is its multilingual setting. The dataset used in the task is the MINT dataset (Pei and Jurgens, 2020) with a total of 13,384 tweets in 10 languages including English, French, Spanish, Italian, Portuguese, Korean, Dutch, Chinese, Hindi, and Arabic. However, the training subset does not include tweets written in Arabic, Dutch, Korean, or Hindi. The task can be modeled as a regression task. We explore some of the most successful transform- ers such as BERT (Devlin et al., 2019), DistilBERT (Sanh et al., 2019), MiniLM (Wang et al., 2020), XLM-RoBERTa (Conneau et al., 2020), and XLM- T(Barbieri et al., 2022). In addition, we also study some data augmentation techniques to increase the training dataset. Based on our experiments during the development phase of the competition, we de- cided to combine the XLM-T transformer model and the data augmentation techniques to estimate our predictions for the test set during the test phase. Our team, HULAT, obtained an overall correla- tion (Pearson’s r) of 0.55. The highest correlation was 0.616, while the lowest correlation was 0.03. Our system ranked in the 27th position out of the 45 participating systems. Although our results are modest, our system performs very well for some languages such as …

Content truncated to 500 words