Freelance Linguistic Editor, AI Alignment (KO/EN)
Nov 2024 – presentKorean data for the human-annotation stage of DeepL's RLHF pipeline: preference ranking, minimal and full edits, cross-lingual judgments and transcript correction.
- Honorific register (jondaemal / banmal)
- Topic vs subject particles (은/는 · 이/가)
- Sentence-ending choice
- Ellipsis and ambiguity
Details
- Ranking
- Compared model-generated Korean candidates for contextual fit, nuance and pragmatic naturalness, with quantitative and qualitative rankings.
- Editing
- Minimal edits that fix only what breaks Korean grammar, so the model learns a precise error signal; full rewrites that keep the semantic core and rebuild structure and word choice as gold-standard data.
- KO–EN
- Cross-lingual tasks: syntactic classification, correcting flawed source text, and judging which translation is better. Also English metalinguistic comments for the research team on why a sentence is ambiguous, with new decision criteria where the guidelines could not decide.
- Speech
- Corrected phonological and contextual errors in Korean speech-recognition and spoken-language transcripts.
Ranking picks the better of existing candidates; editing produces new text, grounded in linguistic reasoning, that goes beyond them.
Beyond the contract
- Proposed a pragmatic framework (Sentiment, Formality, Friendliness) internally for modeling social distance in controlled generation; subsequently developed the unadopted research proposal into an independent, sole-authored PANDORA paper.
- During the v3 API beta, tested the pre-release formality feature with a few example sentences from legal, medical and automotive contexts, and reported where the register slipped as public feedback in DeepL's developer community.
- DeepL Bridges Early Access: nuance research and the Grue Divide survey platform featured on the DeepL Blog. Voluntary feedback, at the community team's request, on the /v3 API (suggesting that a call list the available parameters, with each description fetched in a follow-up call), developer community strategy and editor UX, with a working "Nudge Editor" prototype; plus ongoing suggestions on documentation clarity and product proposals through official channels.