Natural Language Generation (NLG)

• Subset of Natural Language Processing
• Utilizes Grammars and Statistical Models
• Extracted from Human Written Texts

SwiftKey®

• Uses NLG to Predict Next Word as User Types

Task

• Extract and Build Datasets from Representative Text
• Model that Text to Simulate SwiftKey’s Process
• Recreate this Functionality in R

Provided Datasets

• 4 Languages - German, English, Finnish and Russian
  Focus is on English Text for this Project
• Each further broken down into News, Blogs and Twitter Feeds

Descriptive Data Analysis

• 100 Million Words
• Varying Word Choices between Feeds
• For instance, "Dimora"" is only seen in the News Feed

Detailed Report: [https://rpubs.com/yxes/40858]

Individual Word Extraction

• Convert Line to Lower Case
• Remove Non A - Z characters, Spaces or Apostrophe
• Exception: Twitter Retains # Symbol
• Clean up Multiple Apostrophe’s or Hash Marks(#)
• Remove Multiple Spaces
• Split the Line into a Series of Words Separated by Spaces

Grams

• Groups of Words (n-grams) Counted from Resulting Lines
• Groups Consisting of 1-4 Words Counted and Ranked
• Tables Created with the Highest Word Count of a Given Group

Processing

• Based on Modified Markov Model
• Locate Optimal Sequence of Tags for a Given Word Sequence
• Last Four Words are Extracted from Input String
• This Combination is Searched in the Quadgram Table
• If Found - Return the next Word in the List
  • Not Found - Use the latest 3 words and Perform Search of Trigram table
• Loop through Each Table until there’s a Match
• If no Match, Return the Most Common English Word “the”