Description: This is a shiny app that predicts the next word of a provided text based on the previous 3 words. The words are generated by analyzing the texts of news, blogs and twitter.

How it functions: The app consists of a field that accepts user input of a text, a refresh button to generate the prediction and an output field where the single predicted word appears.

Instructions:

1. Load the app: https://meri-galindo.shinyapps.io/PredictWord/
2. Enter the text in the box indicated
3. Hit ENTER or click on the “Predict word” button
4. The predicted word appears on the right panel, within the box provided.

• 10% of the training data were subset and analyzed to generate 4-Grams, 3-Grams, 2-Grams and most frequent single words using the quanteda package for efficiency.
• The nGrams column was divided into 2 columns, where the first column is the search feature (n-1_Gram) and the second column has the last word (word to predict).
• The size of the resulting dataframes was reduced by removing non-used columns and repeated n-1_Grams (only the highest frequency last word is kept). This resulted in using only 40% of the initial memory.
• The data frames were ordered and saved for later use.
• The shiny app loaded the data frames instead of calculating them every time, reducing the memory and time to run the app (only 11 MB of data to load).

• The prediction algorithm first searches the 4-Grams, then the 3-Grams, then the 2-Grams and finally the single most frequent word to predict the next word.

• To speed up the search the ordered data frames were searched using a binary search, which reduced the search time dramatically from seconds/minutes to miliseconds.

• https://quanteda.io/

• https://web.stanford.edu/~jurafsky/NLPCourseraSlides.html

• https://web.stanford.edu/~jurafsky/

• https://docs.posit.co/shinyapps.io