Word-a-Tron

Predictive Text Application

JHU DDS Capstone Project; Spring 2015

• Built off of three, unstructured text files
  • Blogs (1,010,243 lines)
  • News (899,289 lines)
  • Twitter (2,360,149 lines)
• Must predict a word after a word/phrase typed in
• Must have a profanity filter for predicted word

Task is modeling a language with the above corpuses

• Combine, clean, and tokenize files
• Build n-gram tables for prediction
• Prune, aggregate, and simplify tables
• Utilized Domino Data Labs to run concurrent scripts to speed up iterations

What didn't work

• Smoothing (no discernible improvement)
• Interpolation (no discernible improvement)
• Tf-idf ranking (n-grams too short)
• Cosine similarity (n-grams too short)

Keep it Simple Stupid!!!

Katz (Stupid) Backoff

1. Take (at most) last tri-gram, and find highest frequency occurrence
2. If tri-gram non-existent, back-off and use bi-gram
3. If bi-gram non-existent, back-off and use uni-gram

Predicted words are scanned for profanity, and any profane words are substituted with @#$%, but left in corpus for prediction purposes

Application is located at : Word-a-Tron

1. Enter a word or phrase in the text box on the left
2. Predicted word will appear on the right after the word/phrase
3. Use a check box to see the top 3 words predicted or for verbal input of word/phrase

Improvements

• Adaptive Learning with user input over time
• Integration of Information Retrieval techniques for larger phrases

Voice Control powered by annyang!