Understanding Adversarial Queries in African Languages
Recent research efforts led by Makerere AI Lab and Google Research have produced important insights into the use of adversarial queries—essential for enhancing the safety and cultural relevance of language models. This innovative work highlights the need for targeted datasets in addressing potential risks within various sensitive domains.
What is the Pilot Study?
The pilot study collected 8,091 annotated adversarial queries in English and six African languages, including Pidgin English, Luganda, Swahili, and Chichewa. These queries are specifically crafted to elicit unsafe responses from language models (LLMs), allowing researchers to better understand and mitigate risks.
Dataset Accessibility
The dataset, which is open-source, can be explored via this GitHub link. It serves as a crucial resource for evaluating different AI models with an emphasis on both safety and cultural relevance across languages.
Expert Annotations
Experts from seven sensitive domains, including culture, religion, and employment, have annotated these queries. They classified the data into ten topics, such as “corruption and transparency” related to politics, five generative AI themes like public interest and misinformation, and 13 sensitive characteristics, including age and tribe, tailored to the African context.
Dominant Domains in the Data
Among the queries, the health domain stands out with 2,076 entries, followed closely by education with 1,469 entries. Notable topics include chronic disease at 373 and education assessment at 245. A staggering 80% of the queries address issues like misinformation, disinformation, and stereotypes, playing a crucial role in public welfare discussions.
Social Group Concerns
The majority of the queries focus on various social groups, often categorized by factors such as gender (e.g., “Chibok girls”), age (e.g., “newborns”), religion (e.g., “Traditional African religions”), and education level (e.g., “uneducated”). This strategic focus ensures that the data is relevant and contextually significant for the African populace.
Conclusion
The collaborative effort between Makerere AI Lab and Google Research not only enhances our understanding of adversarial queries but also emphasizes the importance of culturally aware AI models. By utilizing this newly available dataset, researchers and developers can work towards more safe and effective AI solutions for African languages.
Related Keywords: adversarial queries, language models, open-source datasets, cultural relevance, machine learning, AI safety, sensitive domains.
