Some tentative first steps towards a Star Trek universal communicator
summary
Builds a tiny-data pipeline that extracts lexicons and grammar rules instead of training on millions of sentence pairs. Leaftop currently auto-extracts about 300 words from each of 1,400 languages, and a universal grammar extractor learns plural inflection from singulars for 11% of the world’s nouns (it rediscovers Latin declensions). Uses UTF‑8 + Hensel-style periodic distance, Python tooling (pandas, numpy, nltk, scikit‑learn, random forest, count vectorizer) and modest CPU runs (tens of thousands of CPU hours) to localize LibreOffice fragments and get useful single-word and token-by-token translations. Results are noisy but usable for low‑resource languages.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:58The Sydney Language
- 1:35Language Loss
- 1:55Cycle of Language Death
- 3:59Language and Technology
- 7:04Universal Translation
- 12:19Translations
- 14:19Single word translation
- 30:49Grammar
- 33:44Periodic Numbers
- 35:19Linear Regression
- 37:14Translation Rules
- 38:39Limitations
- 40:19Questions
related talks
23mMultilingual Natural Language Processing
37mSpeech to text with Elasticsearch
32mSetting up a machine translation service for Timor-Leste
31mBuilding Large-scale Localized Language Models
48mIntroduction to Natural Language Processing with Python
33mScalably Internationalizing Millions of Latent Semantic Labels
43mBreaking the language barrier: how do we quickly add multilanguage support in our AI application?
11mStanfordNLP