Tesla at SemEval-2022 Task 4: Patronizing and Condescending Language Detection using Transformer-based Models with Data Augmentation

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 394 - 399 July 14-15, 2022 ©2022 Association for Computational Linguistics Tesla at SemEval-2022 Task 4: Patronizing and Condescending Language Detection using Transformer-based Models with Data Augmentation Sahil Manoj Bhatt Manish Shrivastava Language Technologies Research Center (LTRC) International Institute of Information Technology, Hyderabad sahil.bhatt@research.iiit.ac.in m.shrivastava@iiit.ac.in Abstract This paper describes our system for Task 4 of SemEval 2022: Patronizing and Condescend- ing Language (PCL) Detection. For sub-task 1, where the objective is to classify a text as PCL or non-PCL, we use a T5 Model fine- tuned on the dataset. For sub-task 2, which is a multi-label classification problem, we use a RoBERTa model fine-tuned on the dataset. Given that the key challenge in this task is clas- sification on an imbalanced dataset, our models rely on an augmented dataset that we generate using paraphrasing. We found that these two models yield the best results out of all the other approaches we tried. 1 Introduction Detecting the presence of patronizing and conde- scending elements in text is an important task for NLP because of the social impact it has. PCL is characterised by a superior attitude towards others, or a manner of speech that seems to portray others in a pitying way. It is different from other problems in the field of text classification such as detection of hate speech or abusive comments because it is not necessarily done on purpose. Having an automated system that is capable of understanding and classi- fying language that contains PCL elements would be the first step towards making people and entities, such as media publications, aware of the kind of language they use when talking about vulnerable communities, and as a result, prevent discrimina- tion, stereotypes and harm that could potentially arise from the use of such language. The PCL Detection task at SemEval-2022 (Pérez- Almendros et al., 2022) aims to solve this prob- lem by exploring different systems that are capable of detecting features that indicate and categorize PCL in the Don’t Patronize Me! dataset (Pérez- Almendros et al., 2020). Sub-task 1 has been for- mulated as a binary classification problem, where the goal is to identify whether a given text falls under the category of PCL or not. Sub-task 2 is a multi-label classification problem, where a given text is either free of PCL or belongs to one or more of the seven PCL categories described in the dataset provided. This paper describes the system developed by team Tesla for SemEval-2022 Task 4. One of the key challenges of this task is the unequal class distribu- tion across the dataset for both sub-tasks. In this paper, we introduce a system to detect PCL using i) a T5 (Raffel et al., 2019) model for subtask- 1 and ii) a RoBERTa (Liu et al., 2019) model for sub-task 2. For our final submission, we use these two models finetuned on an augmented dataset, which we create by generating paraphrases of the sentences belonging …

Content truncated to 500 words