NLP Model for word prediction

• Naive(stupid) back-off model
• Good-Turing smoothing https://class.coursera.org/nlp/lecture/32
• Katz's back-off model https://en.wikipedia.org/wiki/Katz%27s_back-off_model
• Kneser-Ney smoothing https://en.wikipedia.org/wiki/Kneser%E2%80%93Ney_smoothing

Analyse of algorithm implementation:

• CPU usage for all models are the same. Each approach requires N-1 cluster index searches for N-gram. It is possible to reduce cluster index searches for all models if we add a second index to our table.
  
• Memory usage for all models are almost the same.

Preprocess of input text is required. The following functions will help to build best results

• Remove punctuation
• Remove strip whitespaces
• Remove special symbols, especially emotions.
• Transform to lower

Example:

  transFuncs<- list(removePunctuation, stripWhitespace, removeNumbers, content_transformer( function(x) iconv(enc2utf8(x), sub = " ")), 
                    content_transformer(tolower))
  corpus <- tm_map(corpus, tm_reduce, tmFuns = transFuncs)

Preprocessing steps

1. Partitioning data to reduce memory consumption.
2. Building n-grams for n=1:4. Writing results to file.
3. Filter all n-grams with frequency less than 2.
4. Adding columns for each n-gram about start, last (n+1)-gram and mid (n+2) gram. This step significantly improves performance for Kneser-Ney smoothing.
5. Building tables for Good-Turing Smoothing.