Mirko Lai

Marco A. Stranisci, Simona Frenda, Mirko Lai et al.

Inside the NLP community there is a considerable amount of language resources created, annotated and released every day with the aim of studying specific linguistic phenomena. Despite a variety of attempts in order to organize such resources has been carried on, a lack of systematic methods and of possible interoperability between resources are still present. Furthermore, when storing linguistic information, still nowadays, the most common practice is the concept of "gold standard", which is in contrast with recent trends in NLP that aim at stressing the importance of different subjectivities and points of view when training machine learning and deep learning methods. In this paper we present O-Dang!: The Ontology of Dangerous Speech Messages, a systematic and interoperable Knowledge Graph (KG) for the collection of linguistic annotated data. O-Dang! is designed to gather and organize Italian datasets into a structured KG, according to the principles shared within the Linguistic Linked Open Data community. The ontology has also been designed to account for a perspectivist approach, since it provides a model for encoding both gold standard and single-annotator labels in the KG. The paper is structured as follows. In Section 1 the motivations of our work are outlined. Section 2 describes the O-Dang! Ontology, that provides a common semantic model for the integration of datasets in the KG. The Ontology Population stage with information about corpora, users, and annotations is presented in Section 3. Finally, in Section 4 an analysis of offensiveness across corpora is provided as a first case study for the resource.

Mirko Lai, Alessandra Urbinati, Simona Frenda et al.

Current research primarily focuses on model performance, while comparatively less attention has been devoted to uncertainty estimation, particularly in settings where LLMs are increasingly used to generate annotated data. We introduce a framework combining conformal prediction with Collaborative Filtering-style annotators' representation to model LLM behavior in relation to human annotators and to analyze patterns of agreement and disagreement. Using Non-Conformity Scores, we introduce the Ghost Prediction metric and the Ghost Annotator representation to quantify cases in which model predictions diverge from all available human annotations. We compute cosine similarity measures to explore differences in model behavior across sociodemographic axes. We evaluated four LLMs of different size and families across four content moderation datasets. Our finding shows that while we find that all models uncertainty increases with annotator disagreement, larger models tend to be more confident in the classification of texts that are not aligned with any human annotation. Finally, the Ghost Annotator framework reveals a consistent and robust pattern of demographic misalignment, suggesting a structural bias likely rooted in pretraining corpora.

Alessandra Urbinati, Mirko Lai, Simona Frenda et al.

Automatic content moderation is crucial to ensuring safety in social media. Language Model-based classifiers are being increasingly adopted for this task, but it has been shown that they perpetuate racial and social biases. Even if several resources and benchmark corpora have been developed to challenge this issue, measuring the fairness of models in content moderation remains an open issue. In this work, we present an unsupervised approach that benchmarks models on the basis of their uncertainty in classifying messages annotated by people belonging to vulnerable groups. We use uncertainty, computed by means of the conformal prediction technique, as a proxy to analyze the bias of 11 models against women and non-white annotators and observe to what extent it diverges from metrics based on performance, such as the $F_1$ score. The results show that some pre-trained models predict with high accuracy the labels coming from minority groups, even if the confidence in their prediction is low. Therefore, by measuring the confidence of models, we are able to see which groups of annotators are better represented in pre-trained models and lead the debiasing process of these models before their effective use.