Word Predictor

Kai Lun

This presentation introduces a prototype application which
which predicts the next word in a sentence.


A project requirement for
Coursera-John Hopkins Data Science Specialization
TitlePage

Word Prediction App

The app interactively suggests the ‘next word’ coherent to the user input text. Its predictive model is able to present hints to one or more ‘next’ words as specified.

App Features

  • Stable and optimized prototype
  • Live responsiveness to text input
  • Autocompletes words
  • Displays top 1-5 word possibilities
  • Built in English but can be adapted to other languages

Instructions & Functions

Description of Key Algorithm

Why Stupid Backoff?

  • Inexpensive. It requires few resources compared to Katz Backoff.
  • Simple but Accurate. It approaches the quality of Kneser-Ney.
  • Depends on relative freqencies rather than absolute scores.
  • No discounting applied.

sbo sbo2

Conclusion

The following are future enhancements planned;

  1. Explore the use of paid servers to improve speed of app.
  2. Include other sources of corpora (i.e. dictionaries) to cover a larger range of vocabulary.
  3. Explore other models and smoothing techniques to enhance performance of prediction model.

Word Predictor is still at its prototyping stage with a lot of growth potential!

For more insight into the app and its development,
  • Click here for milestone report of the project; and
  • Click here for more behind the scenes info.