This article presents a methodological approach to automatically characterize language registers in French. The method of emerging sequential patterns is described and the obtained results demonstrated first on a corpus of tweets, then more broadly on registers in French. As both a premise and a result of the present approach, the definition presented in this paper for the notion of a language register highlights the notion of a linguistic norm.
Textual corpora are central to Digital Humanities, but their large-scale use remains limited by the lack of suitable tools, particularly for Named Entity Recognition. To address this, we present a fully annotated corpus of seven 19th-century French novels. This resource enables the training and evaluation of tools robust to literary specificities. We propose a CamemBERT-based NER model, effective on such texts, and demonstrate its relevance through character network extraction, opening new perspectives on narrative dynamics. All resources are freely available online.
Abusive language and the propagation of harmful stereotypes have unfortunately become commonplace occurrences on various social media platforms, partly due to users' freedom, anonymity and the lack of regulation provided by these platforms. The sheer volume and often implicit nature of such unwanted content make manual moderation of these user spaces a formidable task. Various scientific communities (Computational Social Science, Natural Language Processing and Computational Linguistics) interested in its at least partial automation have taken up the problem over the past ten years. This special issue aims to encourage interdisciplinary submissions in the field of abusive language detection discussing the limitations of the current approaches and directions for future work.
One under-researched avenue for hate speech and offensive language detection is the integration of knowledge related to speech acts. In previous work, we investigated whether the distribution of speech acts differs across offensive and non-offensive language. Our findings revealed supporting evidence. In the present article, we fine-tune several BERT models and LLMs on the German Speech Acts Dataset. Our goals are two-fold: we want to contribute relevant research results to speech act theory by developing and providing models that detect and classify speech acts in documents or other types of discourse such as tweets. We hope that detected speech acts can be used in a beneficial way as additional features in the detection of hate speech. Our best-performing model achieves a macro-averaged F-1-score of 68.68%.
This paper presents an extended version of CyberAgressionAdo, a French open-access dataset for online hate detection in multiparty conversations. The annotation process was improved with refined guidelines and a two-phase inter-annotator agreement study. A new adaptation of the Weirdness Index is introduced to analyze annotator disagreements. Now structured as a perspectivist corpus, with annotations provided by multiple annotators, CyberAgressionAdo-Large constitutes an enriched resource for the computational analysis of online hate situations in French.
