IdeaWell

Project ideas from Hacker News discussions.

Our Amish Language

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

3 Dominant Themes

# Theme Key Take‑away Representative Quote
1 LLM mistranslation of “hooche Leit” Users point out that an LLM rendered the Pennsylvania‑Dutch phrase as “hohe Leute” (standard German) but stripped away its original meaning of “high/fancy people”. The error is described as plain wrong rather than a subtle connotation shift. “It’s plain wrong.” – usrnm
2 Dialect status & cultural bias The discussion highlights how German speakers treat “dialect” as a low‑brow label, draws parallels to Nazi critiques of Yiddish, and notes the saying a shprakh iz a dyalekt mit armey un flot (a language is a dialect with an army and navy). “Germans are used to having a single source of correct grammar and vocabulary.” – abstractspoon
3 Language documentation & community attitudes Several commenters stress that Pennsylvania Dutch has no standardized orthography, is mainly oral, and that concepts like “love” are expressed differently. They also note the lack of writing tradition among Amish and the resulting challenges for preservation. “It’s hard to ask for a soft favor. Difficult to communicate affection, impossible to say the word love.” – pantalaimon

These three themes capture the bulk of the conversation: erroneous AI output, linguistic hierarchy and German‑centric stigma, and the practical hurdles of keeping a largely oral, unstandardized dialect alive.

🚀 Project Ideas

Pennsylvania Dutch Corpus Builder

Summary

  • A collaborative web platform where speakers upload audio recordings of Pennsylvania Dutch (Deitsch) speech, community members transcribe and align the text, and the resulting corpus is version‑tagged by dialect variant and speaker demographic.
  • Core value: creates a high‑quality, openly licensed text‑audio corpus that enables better LLM training, linguistic research, and preservation of an endangered oral language.

Details

Key Value
Target Audience Pennsylvania Dutch speakers, linguists, heritage learners, NLP researchers
Core Feature Upload → crowd‑sourced transcription → forced‑alignment → downloadable dataset (CSV/JSONL) with dialect tags
Tech Stack Frontend: React + TypeScript; Backend: Node.js/Express; Storage: AWS S3 + PostgreSQL; Alignment: Montreal Forced Aligner (MFA) wrapper; Auth: Auth0
Difficulty Medium
Monetization Hobby

Notes

  • HN users lamented “Due to the lack of a standard orthography don’t expect LLMs to do anything remotely usable…” (trollbridge) – a curated corpus directly addresses this gap.
  • Provides the data needed for the fine‑tuning ideas discussed (e.g., “I’d recommend giving it a squiz… Facebook's No Language Left Behind project”) and can be cited in future academic or preservation projects.

DialectLLM Helper

Summary

  • A fine‑tunable LLM service that lets users submit Pennsylvania Dutch phrases and receive corrected translations, with an optional feedback loop where speakers can upvote/downvote outputs and suggest better renderings.
  • Core value: closes the translation quality gap for low‑resource dialects by continuously improving model outputs through community verification.

Details

Key Value
Target Audience Heritage speakers, researchers, translators, anyone needing PA‑Dutch ↔ English translation
Core Feature Dialect‑specific LLM endpoint (base model + LoRA adapters) + UI for correction feedback and variant selection
Tech Stack Base model: HuggingFace Transformers (Llama‑2 or Mistral); Fine‑tuning: PEFT/LoRA; API: FastAPI; Frontend: Svelte; Feedback store: Redis + Postgres
Difficulty High
Monetization Revenue-ready: subscription tier for API calls ($0.001 per 1k tokens) + free community tier

Notes

  • Directly responds to the criticism that LLMs “explain away mistakes” and give incorrect translations (e.g., the “hooche Leit” discussion). Users noted “the output of your LLM… is plain wrong” (carschno); this tool lets the community correct such errors in real time.
  • Enables practical utility for genealogists, historians, and Amish‑descendant communities seeking reliable language tools.

Deitsch Learning Companion

Summary

  • A mobile‑first app offering bite‑sized lessons, audio flashcards, cultural notes (e.g., meanings of “Ordnung”, expressions of affection), and a community‑driven phrasebook for Pennsylvania Dutch, employing spaced‑repetition and user‑generated content.
  • Core value: lowers the barrier for heritage learners and outsiders to acquire functional PA‑Dutch while preserving cultural context often missing from generic language apps.

Details

Key Value
Target Audience Heritage learners, Amish‑descendant youth, linguists, curious outsiders
Core Feature Lesson library + audio recordings + cultural annotation + SRS flashcards + community phrase submissions
Tech Stack Mobile: React Native (Expo); Backend: Firebase (Auth, Firestore, Storage); Audio processing: FFmpeg via Cloud Functions; SRS algorithm: open‑source SM‑2
Difficulty Medium
Monetization Hobby

Notes

  • HN commenters highlighted the difficulty expressing love and grief in PA‑Dutch (“We have no distinct word for it…”) and the desire to preserve the language despite cultural shifts (panative, skissane). This app directly tackles those pain points by providing contextual examples and affective phrases and audio from native speakers.
  • Encourages discussion on dialect variation and can serve as a living repository that evolves with community input, aligning with the thread’s emphasis on preserving oral traditions.

Read Later