nextWord: faster typing. A Data Science approach.
Alejandro Morales
December, 2014.
What if we could... type less and faster?
- nextWord does exactly that!
- core design pinciples: simple, fast and easy-to-use.
- uses millions of words from different sources.
- leverages powerful data mining and NLP algorithms
- nextWord allows the user to cut up to 15% of typing time.
It is estimated that the average cellphone user sends or recieves up to 40 text messages per day with younger users can send or recive up to a hundred messages (Pew Research Center: how americans text, 2011)
Usage
- Side panel:
- phrase captured through the text box.
- 3 next-word suggestions appear after a short delay.
- leftmost being most probable .
- Details: detailed predictions.
- How to: instructions.
- About: explains the algorithms and data used.
Algorithm
nextWord uses a trigram language model with interpolation and Kneser-Ney smoothing.
The last two words are used to predict the next word, e.g.:
may no [nextWord] --> 1. longer
If that combination of words is not existent, we back off to a bigram model, e.g.:
no [nextWord] --> 1.one, 2.longer, 3.matter
Data & Predictive power
Data
- Corpora: mix of twitter, news and blog sources [1].
- Cleaned for profanities.
- Improved accuracy via Kneser-Ney smoothing of lower order n-grams [2].
[1] Coursera data
[2] N-gram language model
Predictive power
- 15% with single most probable suggestion.
- up to 20% when considering all three suggestions*.
*When tested against a large text.