Symbolic knowledge representations for time-sensitive offensive language detection

Increasing online communication made it necessary to create automatic tools for the detection of abusive language to protect the users. The development of these tools over the years has followed a well-known trend in NLP: from feature-based linear classifiers to neural networks and fine-tuned pre-trained language models. Although the latter have the best performance, these systems face several challenges. The main aim of this project is to improve offensive language models by making them more context-aware and semantically dynamic. Current NLP classifiers for abusive language tend to rely on biased datasets, that in turn propagate the bias in the classification, and specific keywords, thus missing the underlying context (e.g., some words will likely trigger the algorithm into an abusive classification even though the content is not offensive). To overcome these limits, recent studies adopted Explainable AI (XAI) techniques to understand bias drivers and proposed to leverage cultural world knowledge, with the assumption that its incorporation into the systems increases their cultural awareness, which is crucial for culture-dependent tasks such as hate speech detection. This leads to the third challenge: identifying the best approach to inject knowledge into the systems. Many methods exist - from joint pretraining to post-hoc finetuning and adapter-based approaches - but we lack tools to evaluate them. Fourthly, offensive language detection systems rely on datasets from fixed time spans, thus they do not consider the semantic change of words over time.

Given these open challenges, the project will investigate what kind of improvements the injection of external world knowledge into the classifier will lead to. Preliminary results in literature have been promising, leading to the adoption of Knowledge Graphs (KGs) in a wide range of sectors. However, the field of abusive language is still new to their application. Hence, this project aims to address the research...