Detecting biased language in court decisions

Abstract

The linguistic expression of subjectivity is a complex phenomenon that has been the object of reflection by several sub-areas of Linguistics and, more recently, of Computational Linguistics. Linguistic subjectivity, in terms of the linguistic expression of the speaker's opinions and attitudes, affects all levels of discourse organization and is present, to different degrees, in diverse textual genres. Subjectivity and bias are connected, in the sense that the presence of bias in discourse has been related, both in Linguistics and Computational Linguistics, to the occurrence of signs of subjectivity.

Court decisions are an argumentative text genre that may convey traces of subjectivity but should not be biased. As a discourse that represents the State’s position on social matters, it should reflect the principle of Equality. Nonetheless, a preliminary analysis of cases of gender violence reveals that this is not always the case. The research proposed in this paper aims to study the linguistic formulations that convey subjectivity and bias in court decisions on gender violence against women. The goal is to develop a linguistic model to detect these instances of bias, with a future possibility of application in a tool for automatic detection of gender bias in discourse, fueled by Artificial Intelligence (AI) and Natural Language Processing (NLP) techniques. A corpus of court decisions on gender violence has been extracted from the public access database of Instituto de Gestão Financeira e Equipamentos da Justiça (IGFEJ), and has been subject to analysis. A set of examples has been compiled in the analytical section of this study, demonstrating the possibility of connecting certain linguistic features, such as mitigation and intensification mechanisms, evidential expressions and counter-argumentative movements, to the presence of subjectivity and bias in discourse.