A self learning word prediction module for handling multilingual input
summary
A working keyboard prediction module for Firefox OS implements a JavaScript predictor that ships with precomputed frequency lists (built from an ~11 GB corpus with roughly 500k Bengali words) and learns on-device by bumping word weights every time a user retypes a token. It combines rule-based roman→native mapping, n-gram-style lookups and regular-expression transliteration rules, trains the model in ~32–40 minutes on a cluster, runs on low-end 128 MB devices but is noticeably slow there and needs memory/engine optimizations; it preserves privacy by avoiding server-side collection.
AI-written summary · about
chapters (16)
- 0:00Start
- 3:09What do we have today?
- 3:29How does Firefox OS fit in?
- 4:37So how does it work right now?
- 6:09Test Case 1: bn (Bengali)
- 6:46Previous Implementation
- 7:43Implemented Changes
- 11:17Any Demo?
- 15:19What about the local languages?
- 15:44Phonetic Mapping Ambiguities
- 17:13Multiple Input & Phonetic Variants
- 17:54Alphabet Mapping Definition
- 18:41Training Data Generation
- 20:44Dealing with Foreign Words
- 22:52Acknowledgements
- 23:30Questions and Suggestions!
related talks
21mThe Impact of Word, Multiple Word, and Sentence Input on Virtual Keyboard Decoding Performance
16mSpeech Recognition on the Web
37mSpeech to text with Elasticsearch
29mMachine learning for front-end developers
12mOn Suggesting Phrases vs. Predicting Words for Mobile Text Composition
22mUsing Machine Learning to Improve UI/UX
20mModelling Learning of New Keyboard Layouts
25mUsing Machine Learning To Improve UI/UX