📄 The past and future of Fairslator

Summary

Machine translation systems often make unjustified choices about the people mentioned in a text, such as their gender or the formality of address, when the source sentence leaves these open. Fairslator is a web tool, started as a personal project in 2022, that works on top of any MT engine. It spots such ambiguities, asks the user to resolve them through an interface, and then regenerates the translation to match. It currently handles English into German, French, Czech and Irish.

This short paper sets out how the existing proof of concept, now supported by the University of Vienna (2026), will be rebuilt from scratch as an open-source, language-agnostic library. The new design processes each source–translation pair in three stages:

• Preparation: parsing, named entities, coreference and alignment.
• Detection: matching against a catalogue of bias patterns, which yields user questions and proposed rewrites.
• Rewriting: applying the user's choices.

Beyond fairness, the tool aims to reduce tedious post-editing through one-click accept/reject rewrites. It also invites contributors to add modules for other languages. The paper is a roadmap and reports no evaluation results.

Key points

• The current Fairslator covers four directed pairs: English→German, French, Czech and Irish.
• The old implementation in C# and XML carries technical debt. It will be replaced by a Python version built on spaCy, with optional offloading to a Large Language Model or Grammatical Framework.
• Scope goes beyond gender to any arbitrary choice an MT system makes about what (mostly human) referents mean or how they should be addressed in context, such as forms of address.
• Related efforts include gender and formality features from Google, Microsoft, DeepL, Amazon and Apple. There are also academic rewriters for Arabic, Hebrew and Greek, and "automatic post-editing" features in TMS products (XTM, Intento, TextUnited).
• Fairslator positions itself as complementary to these efforts. Unlike most of them, it is open and language-independent, serving as a platform where researchers can experiment.