Language Models (LMs) have been shown to inherit undesired biases that might hurt minorities and underrepresented groups if such systems were integrated into real-world applications without careful fairness auditing.This paper proposes FairBelief, an analytical approach to capture and assess beliefs, i.e., propositions that an LM may embed with different degrees of confidence and that covertly influence its predictions. With FairBelief, we leverage prompting to study the behavior of several state-of-the-art LMs across different previously neglected axes, such as model scale and likelihood, assessing predictions on a fairness dataset specifically designed to quantify LMs’ outputs’ hurtfulness.Finally, we conclude with an in-depth qualitative assessment of the beliefs emitted by the models.We apply FairBelief to English LMs, revealing that, although these architectures enable high performances on diverse natural language processing tasks, they show hurtful beliefs about specific genders. Interestingly, training procedure and dataset, model scale, and architecture induce beliefs of different degrees of hurtfulness.

The varied backgrounds and experiences of human annotators inject different opinions and potential biases into the data, inevitably leading to disagreements. Yet, traditional aggregation methods fail to capture individual judgments since they rely on the notion of a single ground truth. Our aim is to review prior contributions to pinpoint the shortcomings that might cause stereotypical content generation. As a preliminary study, our purpose is to investigate state-of-the-art approaches, primarily focusing on the following two research directions. First, we investigate how adding subjectivity aspects to LLMs might guarantee diversity. We then look into the alignment between humans and LLMs and discuss how to measure it. Considering existing gaps, our review explores possible methods to mitigate the perpetuation of biases targeting specific communities. However, we recognize the potential risk of disseminating sensitive information due to the utilization of socio-demographic data in the training process. These considerations underscore the inclusion of diverse perspectives while taking into account the critical importance of implementing robust safeguards to protect individuals’ privacy and prevent the inadvertent propagation of sensitive information.

Understanding and quantifying the bias introduced by human annotation of data is a crucial problem for trustworthy supervised learning. Recently, a perspectivist trend has emerged in the NLP community, focusing on the inadequacy of previous aggregation schemes, which suppose the existence of single ground truth. This assumption is particularly problematic for sensitive tasks involving subjective human judgments, such as toxicity detection. To address these issues, we propose a preliminary approach for bias discovery within human raters by exploring individual ratings for specific sensitive topics annotated in the texts. Our analysis’s object consists of the Jigsaw dataset, a collection of comments aiming at challenging online toxicity identification.

The 2022 edition of LT-EDI proposed two tasks in various languages. Task Hope Speech Detection required models for the automatic identification of hopeful comments for equality, diversity, and inclusion. Task Homophobia/Transphobia Detection focused on the identification of homophobic and transphobic comments. We targeted both tasks in English by using reinforced BERT-based approaches. Our core strategy aimed at exploiting the data available for each given task to augment the amount of supervised instances in the other. On the basis of an active learning process, we trained a model on the dataset for Task i and applied it to the dataset for Task j to iteratively integrate new silver data for Task i. Our official submissions to the shared task obtained a macro-averaged F₁ score of 0.53 for Hope Speech and 0.46 for Homo/Transphobia, placing our team in the third and fourth positions out of 11 and 12 participating teams respectively.

Current abusive language detection systems have demonstrated unintended bias towards sensitive features such as nationality or gender. This is a crucial issue, which may harm minorities and underrepresented groups if such systems were integrated in real-world applications. In this paper, we create ad hoc tests through the CheckList tool (Ribeiro et al., 2020) to detect biases within abusive language classifiers for English. We compare the behaviour of two BERT-based models, one trained on a generic hate speech dataset and the other on a dataset for misogyny detection. Our evaluation shows that, although BERT-based classifiers achieve high accuracy levels on a variety of natural language processing tasks, they perform very poorly as regards fairness and bias, in particular on samples involving implicit stereotypes, expressions of hate towards minorities and protected attributes such as race or sexual orientation. We release both the notebooks implemented to extend the Fairness tests and the synthetic datasets usable to evaluate systems bias independently of CheckList.