The app predicts the next word in a partial sentence based on previous words. The app uses n-grams and stupid backoff algorithm to predict the next word. The following steps were used to generate the n-grams

• Datasets from three sources twitter, blogs, and news were available from swiftkey
• The blog, news, and twitter datasets were sampled (10%).
• The datasets were converted into a corpus and cleaned.
• The sentences were tokenized into unigrams, bigrams, trigrams, quadragrams, and pentagrams and their frequencies estimated
• Very low frequency n-grams (frequency<2) were removed and the n-grams were written to datasets

• The app uses the unigram, bigram, trigram, quadragram, and pentagram datasets to predict.
• The datasets are loaded when the app starts
• The user enters a partial sentence (request is to use at least 2 words)
• Based on the sentence next words are predicted with a stupid-backoff algorithm which calculates scores based on n-gram frequencies and penalty for backoff.

The details of the scoring for the algorithm can be found at http://www.aclweb.org/anthology/D07-1090.pdf

The user inputs a partial sentence and then hits the submit button.The predicted words and their back-off scores are presented in the table whereas the top 5 words are plotted in the main panel along with a wordcloud of upto the top 100 predictions.

• The app predicts the next words but can have improvements in terms of the processing time
• The other enhancements would be adding word completion component to the app as the user is typing the word
  

References

1. Gendron G.(2015). MODSIM World (13): 1:10
   
2. Jurafsky D and Martin J. (2014). Speech and Language Processing
3. Chen S and Goodman J. (1998). An empirical study of smoothing techniques for language modeling
   

The app is hosted at PredictNextWord
The github repository with the R-code is located at Github