Abusive language in Spanish children and young teenager’s conversations: data preparation and short text classification with contextual word embeddings

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1533–1537 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1533 Abusive language in Spanish children and young teenager’s conversations: data preparation and short text classification with contextual word embeddings Marta R. Costa-juss`a, Esther Gonzalez, Asuncion Moreno, Eudald Cumalat TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona marta.ruiz@upc.edu,esthergonzalezbenitez@gmail.com asuncion.moreno@upc.edu, eudaldcumalat@gmail.com Abstract Abusive texts are reaching the interests of the scientific and social community. How to automatically detect them is one question that is gaining interest in the natural language processing community. The main contribution of this paper is to evaluate the quality of the recently developed ”Spanish Database for cyberbullying prevention” for the purpose of training classifiers on detecting abusive short texts. We compare classical machine learning techniques to the use of a more ad- vanced model: the contextual word embeddings in the particular case of classification of abusive short-texts for the Spanish language. As contextual word embeddings, we use Bidirectional Encoder Representation from Transformers (BERT), we show that it mostly outperforms classical techniques. Far beyond the experimental impact of our research, this project aims at planting the seeds for an innovative technological tool with a high potential social impact and aiming at being part of the initiatives in artificial intelligence for social good. 1. Introduction Many studies show that the unappropriated behaviour of children and young teenagers is increasing, due to their easy access to Internet information, among other reasons (Saba, 2018). Facebook, Twitter or Instagram are some of the many applications that they are using nowadays, without taking into account how other peo- ple may feel about their posts, making cyberbullying one of the problems for young people. There are many studies focusing on automatic abusive or cyberbully- ing text detection (Salawu et al., 2017). Approaches vary from using classical machine learning algorithms (Van Hee et al., 2018) to using neural networks mod- els and specifically using different word embedding features (Chen et al., 2017), among others. Addition- ally, there is recent work (Zhu et al., 2019) using con- textual word embeddings (Devlin et al., 2018) which is reaching great success in many downstream Natural Language Processing (NLP) applications. While most of the research is done for English, there is a recent work for Spanish abusive text classification (Mercado et al., 2018) focusing on large texts and using Support Vector Machines. Differently from previous work, the purpose of this study is to automatically analyse the teenager’s mood according to their short texts for Spanish and using latest NLP techniques (BERT on its multilingual ver- sion). In turn, we want to know if we can achieve this objective by training on the database created in the framework of the project “Safeguarding children online” where project partners were: d-LAB, Safe- ToNet (London, UK), Orange and the Innovation and Technology Center of the Universitat Polit`ecnica de Catalunya (CIT-UPC). The database has the following annotated sentiments: aggression, anxiety, depression, distress, substances, sexuality and violence. While …

Content truncated to 500 words