HIP Domain Translate: A Linguistically-Grounded HCI Approach to Domain Adaptation

Abstract

Machine translation (MT) has enabled immense communication and shared knowledge throughout the world, yet it still falls short. Specifically, when these systems lack data in a particular domain, their accuracy plummets. Considerable work in data-based and model-based approaches have improved the accuracy of domain-specific MT, but these do not directly address ambiguities and decisions of style that are context-dependent. In this thesis, I present Human-Intelligence Powered (HIP) Domain Translate, a system that leverages post-edits made by users to machine translated texts by generating "fixes" within a domain for users to apply. Though interviews, the collection and analysis of post-edits, and a preliminary six-person user study on the prototype, I demonstrate that this system shows promise as a method to quickly and effectively capture contextual information within a domain that is missing in machine translation outputs.