TREMoLo-Tweets: A Multi-Label Corpus of French Tweets for Language Register Characterization

Proceedings of Recent Advances in Natural Language Processing, pages 950–958 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_108 950 TREMoLo-Tweets: a Multi-Label Corpus of French Tweets for Language Register Characterization Jade Mekki1,2 Gw´enol´e Lecorv´e1,3 Delphine Battistelli2 Nicolas B´echet4 1Univ Rennes, CNRS, IRISA / Lannion-Vannes, France 2Universtit´e Paris Nanterre, CNRS, MoDyCo / Nanterre, France 3Orange Labs / Lannion, France 4Universit´e de Bretagne Sud, CNRS, IRISA / Vannes, France firstName.lastName @ {irisa.fr, orange.com, parisnanterre.fr} Abstract The casual, neutral, and formal language reg- isters are highly perceptible in discourse pro- ductions. However, they are still poorly stud- ied in Natural Language Processing (NLP), especially outside English, and for new tex- tual types like tweets. To stimulate re- search, this paper introduces a large corpus of 228,505 French tweets (6M words) annotated in language registers. Labels are provided by a multi-label CamemBERT classifier trained and checked on a manually annotated subset of the corpus, while the tweets are selected to avoid undesired biases. Based on the cor- pus, an initial analysis of linguistic traits from either human annotators or automatic extrac- tions is provided to describe the corpus and pave the way for various NLP tasks. The cor- pus, annotation guide and classifier are avail- able on http://tremolo.irisa.fr. 1 Introduction Language registers are of particular interest in (socio-)linguistics because (1) they are a highly perceptible characteristics of discourse produc- tions ; (2) they represent a significant source of information about the writer/speaker, the relation- ship between interlocutors, or other elements of the communication context ; (3) they are a con- cept known to all (advantage when running per- ceptual tests). Among the possible perceptions of this phenomenon, the partitioning into casual, neutral, and formal registers is probably the most used as it is found in many situations of everyday life. While corpora like GYAFC—where these variations are referred to as “formality level”— have recently popularized the domain (Rao and Tetreault, 2018), it is still poorly studied overall in Natural Language Processing (NLP), especially outside English. Moreover, current work largely focuses on textual types for which registers are al- ready known from the linguistic literature1 whereas many new types, with their peculiarities, arise from the Computer-Mediated Communications (CMCs) (e.g., SMS, tweets...). Therefore, the analysis of CMC corpora in terms of language registers is a challenge both in terms of descriptive linguistics and applications in NLP. As part of the TREMoLo project focusing on language registers2, this paper tries to go beyond these limits and presents the corpus TREMoLo- Tweets, gathering 228,505 tweets (6M words), in French, with multi-label annotations among the ca- sual, neutral and formal registers. The annotations come from a CamemBERT (Martin et al., 2020) model fine-tuned on a manually annotated subset (a.k.a. seed) of the whole corpus. After a state of the art in Section 2, the corpus creation is presented in Section 3. Then, Section 4 provides first linguistic conclusions derived from statistics on manually and automatically-derived linguistic traits. Finally, possible tasks opened by the proposed corpus are listed in the conclusion. 2 Background and …

Content truncated to 500 words