Text Augmentation for Fairer PCL Classification: Tackling Class Imbalance Effectively

Abstract

Condescending and patronizing language can occasionally be interpreted positively or badly. Patronizing someone can imply "supporting." One definition of condescension is "superior attitude towards others."The suggested method analyses news
stories from different nations and determines whether or not they use patronizing or condescending language and categorizes the detected PCL into various groups like 1) Unbalanced power relations(UPR), 2) Shallow solution(SS), 3) Presupposition(PS), 4)
Authority voice(AV), 5) Metaphor(MP), 6) Compassion(CP) and 7) The poorer, the merrier(TPTM). This workuse deep learning (DL) techniques to address this issue, approaching it like a typical multi label text classification problem. It is suggested to use a
pre-trained model Distil-Bert with Data Augmentation methods to achieve best classification. It is considered macro f1 as the metric, The model that has been suggested DistilBERT attained a score of 47.01. The utilization of back translation in data augmentation resulted in a score of 47.34. The utilization of contextual word embedding for data augmentation
resulted in a score of 47.80, while the implementation of synonym replacement for data augmentation yielded a score of 49.26.