The Open University
Addressing Bias in Hate Speech Detection: Enhancing Target Group Identification with Semantics
Abstract
dc:description.abstractThe high volume of online content to be reviewed by moderators requires the creation of automated models using AI. However, one of the main challenges of using AI for hate speech detection is that these models can be biased and discriminatory. <br></br><br></br> In this thesis, we consider bias as <i>the unfair, inaccurate, or prejudice outcome when automatically identifying and classifying hate speech</i>. This often leads to disproportionately flagging content from specific demographic groups, misclassifying non-hateful content as hate speech, or failing to recognise hate speech due to cultural or linguistic nuances. <br></br><br></br> At the core of this problem is the training data, and its imbalanced representation of certain groups or types of hate speech. To address this issue, previous work relied on the automatic identification of groups targeted by hate speech (for instance, to ensure balanced datasets, or adapt models to the specific target groups). <br></br><br></br> Target group identification has turned an essential task in hate speech detection. However, the use of slang, non-standard, vague, and domain-specific language to target specific groups with hate speech makes this task particularly challenging. <br></br><br></br> To address this problem, this thesis investigates the use of semantics, such as knowledge graphs and linguistic resources like dictionaries, to <i>enhance hate speech training data by improving the understanding of specialised, slang and domain-specific language for target group recognition by both humans and machines</i>. <br></br><br></br> We start conducting a systematic literature review on state-of-the-art techniques using semantics to address bias in AI, which inform our research. After reviewing the main challenges identified through a comprehensive literature review of bias in hate speech detection, we focus on enhancing the task of target group identification. <br></br><br></br> We make two key contributions to the field of bias in hate speech detection. First, we demonstrate that enriching training data with explicit semantic information enhances automated models for target group identification, providing interpretability and generalisation to unseen data. Furthermore, we conduct the first study to investigate the role of semantics in hate speech annotation tasks, which demonstrates that providing background knowledge improves the performance of human annotators in identifying target groups. <br></br><br></br> Overall, this thesis offers technical solutions to address bias in hate speech detection, by exploiting semantic resources for an improved understanding of the complex language frequently used in hate speech. <br></br><br></br> <b>Content warning.</b> <i>This thesis aims to tackle hate speech and contains examples of triggering and harmful language.</i>
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Reyero Lobo, Paula
Rights
- Language dc:language
- en