NMT Training & Feasibility Pilot Project: StarPigeon
This pilot project explored whether a custom-trained neural machine translation (NMT) engine could support StarPigeon, a subtitling group for StarCraft II commentary videos. Working with a team, I helped train MT engines using Microsoft Custom Translator and Systran, and conducted multiple rounds of automatic and human evaluation.
Along the way, I gained hands-on experience in:
Building and cleaning bilingual datasets
Training MT engines and interpreting BLEU scores
Coordinating post-editing tests and user acceptance testing
Facilitating team meetings and project timelines
Although the final model didn’t meet our quality and timing goals, this project gave me valuable insight into MT feasibility assessment in a niche, jargon-heavy domain like esports.
