NeuroBiSpell (FR/EN demonstrator)
Translation Checker (QA)
using Deep Learning
Application to the human-in-the-loop charge optimization:
- identification of risky sentences to be checked first by a human, and trustworthy sentences that may be ignored in a frugal resource approach.
- ability to produce an auto-corrected ready-to-use version of the sentence that could be applied with a single-click rather than complex edition operations.
- computer-aided Translation Memory curation.
Result (see screenshot below):
- semantic and structural comparison score between source and target sentences (similarity percentage).
- auto-corrections and highlighted doubts on the source sentence, produced by the monolingual neural auto-corrector.
- possible alerts produced by the rule-base spell and grammar checker on the source sentence.
- auto-corrections and highlighted doubts on the target sentence, according to the source sentence, produced by the bilingual neural auto-corrector.
- possible alerts produced by the rule-base spell and grammar checker on the target sentence.
- numbers (one, two, etc), ordinals (first, second, etc) and multipliers (hundred, thousand, etc), both digital or alphabetical,
even if not in the same form between source and target.
- days of week, and months.
- negations (currently, may be wrong on a non-explicit negation).
- missing/added/wrong articles.
- lowercased/uppercased letters or words, accents.
- fused/cut words
- spelling and grammar
All features should still work even if the sentence is strongly rewritten (similar meaning with different words, not the same order of parts, added/removed parts, etc)