TheNextWord
Bunny Laden
December 2014
1: TheNextWord: A Shiny App
- User inputs a phrase
- App responds with up to three word suggestions
- Uses n-grams of 5, 4, 3, and 2 words
- Adds previously constructed user-based n-grams to the core n-grams to make the app more responsive to an individual user
- N-grams are stored in data frames as frequency and individual words
- N-grams are presorted from high to low frequency
- Saves unmatched words for later learning
2: The Matching Strategy
3: Learning from the User
- Unmatched words can be processed offline and the resulting n-grams added to the model when the app launches in the future
- Great feature for people with unusual vocabularies, such as data scientists, physicians, and lawyers
4: Predictions: TheNextWord and SwiftKey
| Phrase (from NY Times) | Actual | TheNextWord | SwiftKey |
|---|---|---|---|
| Despite reforms in the | military's | united, world, middle | world, morning, past |
| You've roasted and rested | your | on, his | to, the, I |
| The average price of an | individual | old, individual, american | old, issue, email |
| intended to defuse | pressure | on, of, the | to, the, I |
| Winter offers a slower | less | than | than, pace, and |
| Demonstrators marched intermittently along | the | with, the, a | the, with, to |
5: Future Improvements
- Make the matching algorithm more efficient
- Finish implementing the learning feature
- Add more n-grams
- Use a Hadoop cluster for processing
- Add parts-of-speech tagging (POS) to help when guessing is necessary
- Investigate a hybrid semantic-word frequency model for making predictions based on semantic context