App for next word prediction

This presentation explains an App for predicting the next word of an user input. The top requirements where:

• intuitive and minimalistic design
• high efficiency/ small processing delays
• possible use on mobile devices

Online-Demo at https://fschoen.shinyapps.io/NextWordPrediction/

• Instructions
  • The user has to enter text in the input box
  • The prediction for the next word automatically updates (no button-clicking necessary)
  • User may switch languages

• App uses Maximum-Likelihood-Estimate
  • Prediction based on frequencies of previous observations
• Basis for the model in the prototype: corpus of twitter posts
• App selects most likely next word accoding to most common bi- and trigrams in corpus
  • Bi- and trigrams = phrases with two and three words
• Trigram-matches are preffered over bigram-matches

Further information on data preprocessing: https://rpubs.com/FSCHOEN/MilestoneReportCapstoneDataScience

• Example
  • User enters good
  • Most common bigram starting with good is good morning
  • Prediction for next word is morning
• BUT
  • User enters a good
  • Most common trigram starting with a good is a good day
  • Prediction for next word is day as trigram-matches are preferred

• Current status
  • well working prototype for English, German and Finnish
  • decent accuracy (although model was created with only a relativly small text corpus on a home PC)
• Future optimisation
  • performance enhancement by hashing
    
  • increasing accuracy by using a larger text corpus
• Additional possible features
  • model improvement by learning on-use(esp. new words)
  • support for more languages
  • adding case-sensitivity
  • grammar detection