Transfer Learning for Hate Speech Detection

• Applicant: Professor Dr. Alexander Fraser
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 505660769

Performing classification for a socially relevant task is an important area of research. Classification critically depends on training data. But there are significant challenges when the training data is in the wrong language, focuses on the wrong domain, is from the wrong time or is taken from the wrong societal context. In this project, we will study transfer learning focusing on these four important types of training data mismatch for the interesting and socially relevant task of Hate Speech Detection. Lack of training data is a significant issue for many NLP tasks limiting theset of application scenarios which can be supported by machine learning techniques. This problem is important for hate speech detection as it is crucial to protect all communities. In the project we will develop transfer learning approaches in order to support the above mentioned four scenarios. Cross-lingual transfer learning techniques enable building a system on source language data and applying it to the target language without the need for target language training data. On the other hand, the source data available often differs from the requirements of the target language application area, e.g., in the defined label set, thus making them difficult to use. Training a system on out-of-domain data leads to significant performance drops compared to setups where in-domain data is available. Previous work applied domain adaptation techniques, improving accuracy in some cases, but the proposed techniques are sensitive to the choice of used source and target domains which limits their applicability. Language and the discussed topics change over time, especially on social media, where hate speech is a major problem. It has been shown that even recent state-of-the-art systems degrade over time but little work has been done to mitigate the issue without requiring expensive data annotation. Previous work on hate speech detection focused on the societal context aspect by proposing datasets specific to given groups, such as people of a single religion or ethnicity. However, this increases the data scarcity issue because supporting a wide range of groups then requires further data annotation. We will develop transfer learning techniques to overcome these issues. Our goal is to leverage all of the available data from mismatched language, domain, time and societal contexts in order to efficiently transfer maximal knowledge to the target application scenario. By building robust training methods we will eliminate the need for careful training data selection and the need for additional expensive data annotation, by using all of the data sources that are already available. In the evaluation of the outlined methods we will focus on hate speech detection, since it is crucial to support a large number of communities. But in the project the methods we develop will be generally applicable to transfer learning for a wide range of NLP applications.

• DFG Programme: Research Grants
• Co-Investigator: Dr. Viktor Hangya