Cyberbullying Detection on Twitter for the Colombian Population Using Artificial Intelligence Techniques Detección de Ciberacoso en Twitter para la Población Colombiana Usando Técnicas de Inteligencia Artificial

Resumen

Cyberbullying is an increasingly relevant issue in contemporary society that can have serious consequences on the emotional and psychological well-being of victims. Currently, due to the vast exchange of digital interactions, it is challenging and labor-intensive for online platform moderators to manually detect and remove all cyberbullying comments. Therefore, there is a need for automatic models that employ artificial intelligence techniques to detect cyberbullying. This article proposes machine learning-based models and language-based models for cyberbullying detection on the social network Twitter. The machine learning techniques used are XGBoost, logistic regression, and random forests. On the language model side, a fine-tuning process was applied to a masked language model based on a transformer named roberta-base-bne. Although there are currently various models for this purpose, most of them are developed using English. In the case of Spanish, there are very few studies, and in the particular case of Colombian Spanish, there is no precedent for contributions in this area.
Additionally, this article introduces a corpus comprising tweets written in Colombian Spanish, meticulously annotated by a qualified occupational therapist. Two distinct datasets stem from this corpus. Dataset 1 is characterized by the annotation of a tweet as cyberbullying and another as non-cyberbullying, both containing the same word, carefully considering the context in which each word is employed. In contrast, Dataset 2 features different words in cyberbullying and non-cyberbullying tweets. The rationale behind utilizing these two datasets lies in capturing diverse language expressions and their contextual nuances, assessing the effectiveness of the applied techniques in discerning such context. Results from dataset 1 reveal that models achieve an area under the ROC curve of 0.797, 0.796, 0.785, and 0.910 with logistic regression, random forests, XGBoost, and roberta-base-bne, respectively. Meanwhile, employing Dataset 2 yields area under the ROC curve values of 0.983, 0.978, 0.971, and 0.996, respectively. Finally, we introduce a web application named AI Cyberbullying Detector tailored for therapists, empowering them to leverage artificial intelligence in cyberbullying-related studies.