I2C at SemEval-2022 Task 4: Patronizing and Condescending Language Detection using Deep Learning Techniques

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 459 - 463 July 14-15, 2022 ©2022 Association for Computational Linguistics Abstract Patronizing and Condescending Language is an ever-present problem in our day-to- day lives. There has been a rise in patronizing language on social media platforms manifesting itself in various forms. This paper presents two performing deep learning algorithms and results for the “Task 4: Patronizing and Condescending Language Detection.” of SemEval 2022. The task incorporates an English dataset containing sentences from social media from around the world. The paper focuses on data augmentation to boost results on various deep learning methods as BERT and LSTM Neural Network. 1 Introduction Nowadays, Patronizing and Condescending Language (PCL) (Pérez-Almendros, Espinosa- Anke, and Schockaert 2022) is used to refer to a forced kindness that derives from a perceived superiority towards another person. A subtle form of bullying, being patronized can leave a person feeling infuriated and impotent. It is a type of behaviour that cuts across generations. An older person can talk down to a younger colleague, but it can just as easily happen the other way around. Men can patronize women at work and vice versa. But what they have in common is power play, with one individual exerting their authority or seniority over another. People could feel discriminated by condescending comments. Considering the relevance of equality and respect, and it is important to note that nobody should be treated in such a way as to feel intimidated or different. In SemEval-2022: Task 4 Subtask 1, participants must determine whether a phrase presents PCL or not. The idea behind the proposed solution was to compare three models based on deep learning. The first model is a Long short-term memory (LSTM) neural network (Zhou et al. 2015). The second, a LSTM neural network with embedding pretrained layer. The third one uses BERT (Devlin et al. 2018). The latter yielded the best results. For all the models we have used data augmentation to balance the training dataset. The F1-score in our final submission (BERT) was 0.4134 on the test dataset. Compared to the results of the winning team, the difference is not extremely large. Nevertheless, our model obtained a good result of accuracy (0.61) compared to other participants. 2 Background This paper is focused on Subtask 1: Binary classification. The corpus provided to perform subtask 1 (Pérez-Almendros, Espinosa-Anke, and Schockaert 2020) was composed of 10473 documents with 6 features: “id”, “doc id”, “keyword”, “country code”, “text” and “label”. The dataset was imbalanced with respect to the class “label”. Out of 10473 rows, 9470 were from the negative class and 993 from the positive class. English data augmentation was applied to the original dataset to balance them (Shorten, Khoshgoftaar, and Furht 2021). I2C at SemEval-2022 Task 4: Patronizing and Condescending Language Detection using Deep Learning Techniques Laura Vázquez Ramos Adrián Moreno Monterde Escuela Técnica Superior de Ingeniería Escuela Técnica Superior de Ingeniería Universidad de Huelva (Spain) Universidad de Huelva (Spain) laura.vazquez005@alu.uhu.es adrian.moreno521@alu.uhu.es Victoria …

Content truncated to 500 words