Adrián Pastor López-Monroy

Cyberbullying has experienced a surge in recent years, mainly due to the widespread adoption of social media platforms. This trend manifests in multiple ways, with hostile language being one of the most common. The latter underscores the urgent need for robust detection methods to address this issue effectively. To address this problem, we propose a novel pipeline to enhance hostile language detection in social media. Our approach consists of a combination of two ideas: First, we propose conducting a Domain Adaptation procedure to specialize the knowledge of a pre-trained BERT, making it more specialized in the domain of social media. For this adaptation, we modify the traditional random Masked Language Modeling technique and propose three novel strategies for selecting the subset of tokens to mask out cleverly. Second, we tailor an Adversarial Regularizer when fine-tuning the adapted BERT for specific hostile-language datasets. We evaluate the performance of our method for detecting hate speech, aggressiveness, offensiveness, and sexism. Our results show that the Domain Adaptation procedure significantly outperforms vanilla BERT, and the Adversarial Regularizer can lead to more robust fine-tuning, thereby enhancing performance. Moreover, we demonstrate that these methods can be used together to achieve an even more significant performance boost.

Self-harm is an increasing public health problem with high prevalence rates in adolescents. Furthermore, it can be an indicator of different mental health disorders (e.g., depression). Recently, diverse computational methods have leveraged user-generated data from social media to study and identify this issue following a text classification perspective. In this context, previous work has shown that personal statements -phrases containing first-person pronouns- contain valuable information for modeling the profiles of authors, including some mental health conditions. Motivated by these discoveries, this paper examines the relevance of personal statements to tackle the self-harm detection task on social media. Furthermore, we adapted approaches that pay special attention to words in this type of sentence to reveal characteristics of users, specifically, self-harming behavior. Currently, these approaches assign the same level of importance to words contained in personal phrases without distinguishing which are more associated with the personal contexts of authors. Hence, we introduced a novel weighting factor that exploits the proximity between personal pronouns and words to quantify their relevance in the task. This novel weighting factor, inspired by findings from author profiling and depression detection studies, is being evaluated for the first time in the context of self-harm detection. Our experimental results demonstrated significant improvements over state-of-the-art methods in self-harm detection, including transformer-based methods and pretrained language models for mental healthcare. This refined approach not only surpasses the previous weighting factor in its application to depression detection, but also exceeds it by a difference of more than 3.9%.