ofdn/OpenSpeaks-Before-AI

A set of frameworks for creating the AI/ML building blocks for low-resource languages. # OpenSpeaks Before AI* _A set of frameworks for creating the AI/ML building blocks for low-resource languages_. This experimental initiative by O Foundation aims to understand and document the challenges many low-resourced language speakers face. We implemented the initiative in August-September 2021 as a part of the MozFest Trustworthy AI Working Groups program, a cohort by the Mozilla Foundation around building trustworthy AI. AI and ML affect everybody, whether or not they get to participate in building the massive datasets that drive AI/ML research and development. However, historically-oppressed communities, including but not limited to Black, ethnolinguistic minority communities, and poor-income communities all around the world, Dalit-Bahujan-Adivasi communities in South Asia, often have very little say about the data that is collected to build AI/ML models. Those in the research/solution-building side of the equation do not often have interest, know-how, or resource. Our initial goal during the Working Group program was to study a few existing platforms by open-auditing. By looking at our findings from the audits, we planned to create a set of frameworks so others can identify the sociolinguistic and other anthropological gaps critical to people and data. Our understanding of digital biases is influenced by a wide range of feminist literature, particularly that of collectives such as WhoseKnowledge? and Equality Labs. Two studies inspired our approach: a seminal paper titled "Datasheets for datasets" (Gebru et al. 2021), focusing on identifying gaps and biases in datasets, and our research on the monetization of web content in two Adivasi languages of India (Ho and Santali). During this process, we asked ourselves, "how could speakers of low-resource languages audit a platform based on their socioeconomic context?". ## Context on colonial and post-colonial caste oppression The caste system in South Asia prevalent in the Hindu-majority societies befor …