This application was developed as a final submission to the capstone program for the Johns Hopkins University “Data Science and Machine Learning” specialization certificate. It demonstrates the flexibility of the R language and Shiny application development framework by showcasing a text-prediction algorithm. The application returns a next-word suggestion along with an array of additional suggestions as long as there is one word present in the text box and a match for the input has been found in the database used for the algorithm.

The application is available for use at https://5yf0hd-such70sam-pichardo.shinyapps.io/predictr/

After the app initializes (this may take up to 10 seconds as the word databases are loaded in), the user is prompted to type at least one word. A default is placed automatically. The app will not load any data if the text box is left empty.

The algorithm will provide a suggestion for the next word along with a table of the top 5 suggestions, if available. Uncommon misspellings and inputs for which there are no matches in the databases will not return any results.

This application operates on a “stupid backoff” model. When an input is detected, the app utilizes a function devoted to searching a database of 3-grams extracted from a sample of 5% of the english-language corpus of tweets blog posts, and news articles available to download through this link. When the last two words of the input can be matched with a set of 3-grams in the corpus, the array with the top 5 most common configurations is shown along with the top result as a suggestion in the sidebar.

To parse the input, the application extracts the words into a single-column data frame of length n. The nth and (n-1)th word are extracted for use in the 3-gram function, which calls on a 2-gram function first if the (n-1)th word is not present and second if no matches are available for both words.