University of Bucharest Team at Semeval-2022 Task4: Detection and Classification of Patronizing and Condescending Language

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 509 - 514 July 14-15, 2022 ©2022 Association for Computational Linguistics University of Bucharest Team at Semeval-2022 Task4: Detection and Classification of Patronizing and Condescending Language Raluca Andreea Gînga, Bogdan Dobre, Tudor-Andrei Dumitra¸scu and Bogdan Radu Silviu Sielecki University of Bucharest {gingaraluca, dobrebogdan98, tudorandrei.dumitrascu, sieleckiradu}@gmail.com Abstract This paper details our implementations for find- ing Patronizing and Condescending Language in texts, as part of the SemEval Workshop Task 4. We have used a variety of methods from simple machine learning algorithms applied on bag of words, all the way to BERT models, in order to solve the binary classification and the multi-label multi-class classification. 1 Introduction The Patronizing and Condescending Language De- tection Task (Pérez-Almendros et al., 2022) is based on the paper Don’t Patronize Me! (P’erez- Almendros et al., 2020), which is an annotated Dataset with Patronizing and Condescending Lan- guage Towards Vulnerable Communities. The aim of this task is to identify PCL, and to categorize the language used to express it, specifi- cally when referring to communities identified as being vulnerable to unfair treatment in the media. Participants were provided with sentences in con- text (paragraphs), extracted from news articles, in which one or several predefined vulnerable com- munities are mentioned. The challenge is divided into two subtasks. 1. Subtask 1: Binary classification. Given a para- graph, a system must predict whether or not it contains any form of PCL. 2. Subtask 2: Given a paragraph, a system must identify which PCL categories express the condescension. The PCL taxonomy was de- fined based on previous works on PCL (i.e. Unbalanced power relations, Shallow solution, Presupposition, Authority voice, Metaphor, Compassion, The poorer, the merrier. ) 2 Background The dataset used for this SemEval 2022 task was Don’t Patronize Me! (P’erez-Almendros et al., 2020), which contains a suite of sentences that men- tion some vulnerable communities and published in media in a lot of English speaking countries. The paragraphs were manually annotated to show 1) whether the text contains any kind of PCL, and 2) if it contains PCL, what linguistic techniques (categories) are used to express the condescension. The paragraphs, according to (P’erez-Almendros et al., 2020), were extracted from News on Web (NoW) corpus (Davies, 2013), being annotated by three expert annotators, with backgrounds in com- munication, media and data science. The dataset for subtask 1 (binary classification) contained a number of 10.636 paragraphs and 2.792 instances were used for the categories classification subtask. In Figure 1, it can be seen that for the first sub- task, there are almost 1000 texts that contain PCL. This means that the dataset is highly imbalanced and this problem needs to be addressed. Figure 1: Classes Distribution for Binary Classification problem (Subtask 1) For task 2, the paragraphs from task 1 are split according to the type of PCL speech category into sentences, resulting in 950 samples. 3 System Overview 1. Subtask 1 (Binary Classification) Because the dataset was very imbalanced, we …

Content truncated to 500 words