AliEdalat at SemEval-2022 Task 4: Patronizing and Condescending Language Detection using Fine-tuned Language Models, BERT+BiGRU, and Ensemble Models

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 387 - 393 July 14-15, 2022 ©2022 Association for Computational Linguistics Abstract This paper presents the AliEdalat team’s methodology and results in SemEval-2022 Task 4: Patronizing and Condescending Language (PCL) Detection. This task aims to detect the presence of PCL and PCL categories in text in order to prevent further discrimination against vulnerable communities. We use an ensemble of three basic models to detect the presence of PCL: fine-tuned bigbird, fine-tuned mpnet, and BERT+BiGRU. The ensemble model performs worse than the baseline due to overfitting and achieves an F1-score of 0.3031. We offer another solution to resolve the submitted model’s problem. We consider the different categories of PCL separately. To detect each category of PCL, we act like a PCL detector. Instead of BERT+BiGRU, we use fine-tuned roberta in the models. In PCL category detection, our model outperforms the baseline model and achieves an F1-score of 0.2531. We also present new models for detecting two categories of PCL that outperform the submitted models. 1 Introduction Increasing internet access rates and the development of a diverse range of online forums have allowed people around the world to engage in a tremendous range of topics. This has been accompanied by an increase in unhealthy online texts whose negative effects on people have been significant. One type of such unhealthy texts is a text with patronizing and condescending language (PCL). When a person's language expresses a superior attitude towards others or describes their situation in a benevolent way that creates a sense of pity, the person has used this type of language. In the media, vulnerable communities seem to be a good target for this type of language. However, this type of language can normalize discrimination. We believe that unfair treatment of vulnerable groups leads to greater deprivation and inequality for these groups. Therefore, recognizing the existence of this type of language and its variations is important and can prevent these problems. So far, significant work has been done on modeling the language that deliberately and openly undermines others, such as offensive language or hate speech, but little has been done on the language of humiliation and pity. This language of humiliation and pity is used in the media subtly and indirectly and different from other types of unhealthy languages. The special focus on the language of humiliation and compassion for vulnerable communities has been noted only in the work of Pérez-Almendros et al., (2020). In this work, a dataset to identify this type of language is presented, but no significant work has been done in designing a model to classify this type of text. Unhealthy text papers usually focus on obvious and aggressive phenomena such as detecting fake news, fact-checking, modeling offensive language, and spreading rumors. There has been a few work on PCL recently. Wang and Potts (2019) introduced compassion modeling in direct communication from the perspective of natural language analysis. They created and tagged a dataset with social media …

Content truncated to 500 words