Machines Do See Color: A Guideline to Classify Different Forms of Racist Discourse in Large Corpora (2401.09333v2)

Abstract: Current methods to identify and classify racist language in text rely on small-n qualitative approaches or large-n approaches focusing exclusively on overt forms of racist discourse. This article provides a step-by-step generalizable guideline to identify and classify different forms of racist discourse in large corpora. In our approach, we start by conceptualizing racism and its different manifestations. We then contextualize these racist manifestations to the time and place of interest, which allows researchers to identify their discursive form. Finally, we apply XLM-RoBERTa (XLM-R), a cross-lingual model for supervised text classification with a cutting-edge contextual understanding of text. We show that XLM-R and XLM-R-Racismo, our pretrained model, outperform other state-of-the-art approaches in classifying racism in large corpora. We illustrate our approach using a corpus of tweets relating to the Ecuadorian ind\'igena community between 2018 and 2021.

• The paper introduces a comprehensive methodology using Transformer-based models like XLM-RoBERTa to detect both overt and covert racist discourse within expansive text corpora.
• It employs a four-step process that fuses robust theoretical frameworks with advanced BERT-based architectures to capture nuanced racial language in context.
• Empirical analysis in the Ecuadorian context reveals that overt racist content is more likely from peripheral network actors, highlighting social risks and mobilization dynamics.

Introduction

The analysis of racist discourse within large textual corpora has been beset by methodological challenges, partially due to the nuanced and evolving nature of language underpinning such discourse. Traditional methodologies often focus on overt forms of racism and discount subtler manifestations, which has created a significant research gap in the domain of computational political science. Filling this gap, a comprehensive methodology combines a fundamental theoretical framework with state-of-the-art machine learning techniques, notably the bidirectional Transformers models like XLM-RoBERTa (XLM-R), to detect and categorize instances of covert and overt racism within expansive text corpora adeptly.

Conceptual Foundation

A solid theoretical footing is integral to discerning racism in text. The presented research draws on established definitions, viewing racism not just as a dyadic interaction between individuals but as a pervasive structural phenomenon informed by historical and societal context. The conceptualization of racism adopted amalgamates ideologies of racial domination and prevailing racial stereotypes, which in turn dictate how racist language manifests itself. Consequent to these conceptual grounds, the creation of a coding scheme to define overt and covert racism within the analyzed corpora is pivotal, allowing context-sensitive machinery to navigate the intricacies of racist discourse.

Methodological Advancements

The paper proposes a four-step process that guides researchers to contextualize racism aptly, identify explicit and subsurface forms of racial discourse, and apply cutting-edge NLP technique. The use of BERT-based models, particularly adept at comprehending context, supplements traditional methods. These models leverage Transformer architectures, advanced by their self-attention mechanisms that consider the interdependence of words within textual segments, thereby generating dynamic word representations essential for classifying nuanced discourse.

Empirical Application and Insights

Empirical analysis in the Ecuadorian context elucidates the profundity of the approach. Specialized versions of the XLM-R model are finely tuned, exhibiting superior proficiency in identifying covert and overt racist discourse over more traditional NLP methods. The methodology's robustness is further illuminated through its application to a corpus of tweets pertaining to the Ecuadorian indígena community. Interestingly, the research reveals that overt—and to some degree, covert—racist content is less likely to emanate from central figures within social networks, in alignment with the social risks that overt racial language entails. Furthermore, the paper demonstrates that challenges to the socio-racial status quo incite a marked increase in racial discourse, pointing to the use of such language as a mobilization tool against perceived threats to in-group hegemony.

Conclusion

The innovative approach advocated in the paper transcends beyond the detection of racist language, illustrating the utility of Transformer-based models in interpreting complex socio-political phenomena through large text corpora. The paper underscores the importance of fusing robust theoretical foundations with advanced computational models to yield insights that resonate with contemporary racism's covert and explicit modes. Notably, the template established here serves as a springboard for the exploration of other complex discursive patterns within political science and beyond.