Application: https://rimdzius.shinyapps.io/textpredictor/

The application allows a user to type into the text box. It is an interactive application, and does not require any button to begin the analysis.

The text box itself can be clicked and dragged to change its size, if needed.

Once the user stops typing, the application will run the predictive algorithm. The predicted word will be shown in the button below the input box. A user can simply click the button to add it to the end of their text.

The script files can all be viewed at https://github.com/rimdzius/PredictiveText.

1. The processing begins before the user begins typing: if there is no input in the text box, the function will return a sample() of a list of the top 100 most common words in the english language, obtained from wikipedia.

2. While processing, the application will tokenize the user text, removing any numbers, punctuation or symbols. It then loops over the 2-, 3- and 4-gram datasets that have been preprocessed. It also loops over 1, 2 and 3 of the final words input in each of these ngram datasets.

3. These preprocessed ngram datasets (viewable on github) are ordered in decreasing frequency. Any matches found will be saved with the word and probability data, based on the frequency of the match vs the frequency of the previous words. (This is based on the Katz back-off model.)

4. The highest probability set is used, returning the last word in the set. If no matches are returned at all, the top100 words dataset (mentioned in step 1) will be sampled from again.

One thing to mention is that this application itself does not utilize set.seed(). This was done purposefully to allow a bit more randomization to the sampling of the top100 words. However, all preprocessing steps are reproducible.

There are several .R scripts on the github page that show the preprocessing steps.

• preprocessing.R: This is the main function call, preprocessing()
• download_data.R: This simply downloads and unzips the data, if it’s not already present.
• create_tokens.R: This reads the data in, tokenizes and cleans up the tokens.
• create_ngram_matrix.R: This is the bulk of the work. This will create the ngram datasets. It will return a dataframe including these ngram terms, and the frequencies that they are used.
• write_files.R: Simply writes the dataframes into an “ngram.txt” file, and a “ngram_freq.txt” file.

Text prediction is a bit difficult in its own right: getting a large enough sample dataset, including large n-gram files, and keeping the overall size of the app down.

Since this app would be designed for a user’s phone, the clear answer is to utilize the user’s own inputs (with permission!) to train the algorithm.

Instead of using an enormous dataset of the English language, including news articles, blog and twitter posts, we can use a smaller dataset to set the framework, and then use the user’s own data entry to store a database of their most commonly used words, terms and phrases.