Castia vs. Google Translate.

What Google Translate actually does

Google's own published research describes Translate's engine as a neural machine translation (NMT) system - a deep encoder-decoder network trained to produce one fluent, accurate translation of the text it's given. That's a single-pass job: read the source, decode the most probable target-language sentence, and return it. There's no documented step where the system drafts several differently-angled versions of a line and scores them against each other before picking one - the architecture isn't designed to do that, because sentence-level fluency and accuracy is the problem it's solving.

For most translation needs - reading a menu, understanding an email, getting the gist of a foreign-language article - that's exactly the right tool, and a fast, free, extremely well-engineered one.

Where a single accurate pass isn't the same job

A song lyric or a character's line of dialogue has constraints a menu item doesn't: rhyme, rhythm, singability, idiom, a character's established voice. Translation researcher Peter Low's "Pentathlon Principle" names five demands a singable lyric translation has to balance at once - singability, sense, naturalness, rhythm, and rhyme - and argues that treating literal semantic accuracy as the only priority actively produces worse lyric translations, not better ones. A model optimized purely for translation accuracy has no mechanism for making that tradeoff on purpose.

This is also where automated translation-quality metrics run into their own well-documented limits: Callison-Burch et al.'s EACL 2006 paper on BLEU (the standard automated MT metric) shows that a higher BLEU score doesn't reliably track a better human-judged translation, precisely because BLEU rewards closeness to one reference wording, not whether a differently-worded line captures the same feeling better. Castia's Judge stage exists for the same reason a metric like BLEU falls short here: it scores rewrites against what the line actually says (the Translator's literal anchor), not against surface closeness to any single wording.

Side by side