This paper addresses patronizing and condescending language (PCL), a harmful form of speech that belittles vulnerable groups and reinforces negative stereotypes. Unlike overt hate speech, PCL is subtle and often escapes human annotation. The authors propose two detection systems. The first combines static word embeddings with a convolutional neural network (CNN) for binary classification. The second fine‑tunes a pre‑trained BERT model on the PCL task. Results show that the BERT‑based approach outperforms the CNN baseline, achieving an F1 score of 0.528 on the development set. The study offers an initial step toward multilingual harmful‑language detection and suggests that larger datasets and more advanced architectures could further improve performance.
Review