In recent years, the internet, especially social media platforms, has experienced a rapid expansion. This growth has empowered individuals to create and share content of their choosing, regardless of its appropriateness. Unfortunately, this freedom has also led to the proliferation of hate speech – offensive language or threats directed at individuals based on various factors such as ethnicity, religion, or sexual orientation.
Hate speech detection models have emerged as crucial tools in moderating online content and combatting the spread of harmful speech, particularly on social media platforms. These computational systems are designed to identify and categorize online comments as either hate speech or non-hate speech. Assistant Professor Roy Lee from the Singapore University of Technology and Design (SUTD) emphasizes the significance of these models in maintaining a safe and inclusive online environment.
While evaluating the performance of hate speech detection models is essential, traditional methods, such as using held-out test sets, often fall short in accurately assessing the models’ efficacy due to inherent biases within the datasets. To address this limitation, researchers introduced HateCheck and Multilingual HateCheck (MHC) as functional tests that simulate real-world scenarios, capturing the complexity and diversity of hate speech effectively.
Building on the foundations of HateCheck and MHC, Asst. Prof. Lee and his team developed SGHateCheck, an AI-powered tool tailored to the linguistic and cultural context of Singapore and Southeast Asia. By incorporating large language models (LLMs) and translating test cases into Singapore’s main languages, SGHateCheck offers over 11,000 meticulously annotated test cases, enabling a nuanced evaluation of hate speech detection models. Unlike MHC, SGHateCheck provides a more regionally specific focus, addressing the unique linguistic features of Southeast Asia, such as Singlish.
The research team discovered that LLMs trained on monolingual datasets tend to exhibit biases toward non-hateful classifications. In contrast, LLMs trained on multilingual data sets demonstrate a more balanced performance and improved accuracy in detecting hate speech across different languages and cultural contexts. This highlights the importance of incorporating culturally diverse and multilingual training data in developing hate speech detection applications for multilingual regions like Southeast Asia.
SGHateCheck has the potential to play a significant role in enhancing the detection and moderation of hate speech in online environments in Southeast Asia. By addressing the specific linguistic and cultural nuances of the region, SGHateCheck is poised to contribute to creating a more respectful and inclusive online space. Asst. Prof. Lee envisions the implementation of SGHateCheck in various online platforms, including social media, online forums, and news websites, to combat hate speech effectively.
As part of his future plans, Asst. Prof. Lee intends to develop a content moderation application utilizing SGHateCheck and expand its capabilities to include other Southeast Asian languages like Thai and Vietnamese. SGHateCheck exemplifies SUTD’s commitment to integrating cutting-edge technology with thoughtful design principles to address real-world challenges. By prioritizing cultural sensitivity in the development of hate speech detection tools, SGHateCheck showcases the importance of a human-centered approach in technological innovation and development.