Context-Aware Stereotype Detection: Conversational Thread Analysis on BERT-based Models

Conversational context plays a pivotal role in disambiguating messages in human communication. In this study, we investigate the impact of contextual information on detecting stereotypes related to immigrants using various BERT-based models. We use two Spanish corpora containing news comments and tweets, together with their conversational threads, annotated with stereotypes related to immigrants in Spain. The results show that the influence of context on stereotype detection varies across different models, corpora and context levels. Although context can enhance performance in specific scenarios, it does not consistently improve stereotype detection across all the levels of contexts. Our comprehensive evaluation underscores the complex relationship between context and stereotype identification when we use BERT-based Language Models. In particular, we found that the number of texts benefiting from contextual analysis may be too limited for the models to effectively learn from.