Multilingual NMT with Gender Bias Detection

Overview

This research team is developing AI-powered translation tools for low-resource Ethiopian languages building technology that doesn't just break language barriers, but challenges the gender biases embedded within them.

Machine translation tools are silently reinforce harmful stereotypes defaulting occupational terms like "doctor" or "engineer" to masculine forms across languages like Amharic, Oromo, and Tigrinya. These biases embed and amplify them in the digital tools communities increasingly depend on.

The project develops a multilingual neural machine translation prototype spanning English, Amharic, Ge'ez, and Awi, using transfer learning to detect and mitigate gender bias at the model level. Women are embedded throughout the process, leading data collection, preprocessing, and labeling, while linguistics experts ensure cultural and grammatical accuracy for each target language.

By building AI translation from the ground up, the initiative is laying the foundation for technology that reflects the full diversity of Ethiopian society. If scaled, this model could reshape how low-resource African languages are represented in AI.