LastResort at SemEval-2022 Task 4: Towards Patronizing and Condescending Language Detection using Pre-trained Transformer Based Models Ensembles

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 352 - 356 July 14-15, 2022 ©2022 Association for Computational Linguistics LastResort at SemEval-2022 Task 4: Towards Patronizing and Condescending Language Detection using Pre-trained Transformer Based Model Ensembles Samyak Agrawal Radhika Mamidi International Institute of Information Technology Hyderabad samyak.agrawal@research.iiit.ac.in, radhika.mamidi@iiit.ac.in Abstract This paper presents our solutions for Task4 at SemEval2022: Patronizing and Condescend- ing Language Detection. This shared task con- tains two sub-tasks. The first sub-task is a bi- nary classification task whose goal is to predict whether a given paragraph contains any form of patronising or condescending language(PCL). For the second sub-task, given a paragraph, we have to find which PCL categories express the condescension. Here we have a total of 7 overlapping sub-categories for PCL. Our pro- posed solution uses BERT based ensembled models with hard voting and techniques applied to take care of class imbalances. Our paper de- scribes the system architecture of the submitted solution and other experiments that we con- ducted. Our best performing models achieve an F1 score of 59.4 and 15.7 on sub-tasks 1 and 2 respectively. 1 Introduction Patronizing and condescending attitude in language generally denotes the writer’s sense of superiority over others. If someone is patronizing or conde- scending, it means what they write/say is accompa- nied by a sense of pity or compassion. Often, usage of PCL is relatively unconscious, and the intent of the writer is not to hurt a particular group or person they are referring to. So while being harmless in its intention. Usage of PCL still poses a risk of harming vulnerable people or groups by stereotyp- ing them or normalizing specific behaviour towards them. Task4 at SemEval-2022 (Pérez-Almendros et al. (2022)), Patronizing and Condescending Language Detection provides two sub-tasks. The goal of sub- task1 is to identify if the given paragraph contains PCL. The goal of sub-task2 is to determine which subcategory of PCL expresses the condescension. The seven subcategories are Unbalanced power re- lations, Shallow solution, Presupposition, Author- ity voice, Metaphor, Compassion and The poorer, the merrier. A given paragraph can show instances of multiple subcategories. We experimented with multiple transformer- based models. We used focal loss and Weighted Random Sampling to address the class imbalance; we also tried out ensembling models with hard vot- ing, which improved the accuracy over the baseline models for both the sub-tasks. The paper is structured as follows: Section 2: describes the dataset and related work. Section 3: describes our system and model architecture. Section 4 has information regarding the dataset size and splits with libraries used. Section 5 discusses the findings from our experiments, and section 6 concludes our paper. 2 Background There has been work done to detect potentially harmful forms of language. Liu et al. (2019a) used BERT and LSTM based models to detect offensive language in the dataset, Offensive Language Identi- fication Dataset (OLID) provided by Zampieri et al. (2019) at SemEval 2019. Indurthi et al. (2019) used InferSent (Conneau et al. (2018)) …

Content truncated to 500 words