Predictograph

by Deadpool
6th of December, 2018

What's a predictograph?

Predictograph is an algorithm that predicts the next word from the two words you provide. It's main features are:

  • input can be written in any case, with or without punctuation
  • numerics are not supposed to be provided as input or expected as output
  • easy to use with elaborate explanation of functions
  • allows weighted predictions for simpler predictive models

More on the underlying learning principle: https://en.wikipedia.org/wiki/Katz%27s_back-off_model

Inputs of the app

  • Line to be finished [text]: 2-word phrase that should be completed by the algorithm

  • Style [drop-down menu]: different users have different styles of communication, thus we formed different styles (formal, mixed, informal) to match our users' preferences

  • Weight [slider]: predictions by simpler model tend to be less accurate, thus users can move the slider to make the correction more or less strict

The function that does most of the work

FindTrigram7 <- function(inp, data3, data2, mutation = 50){
  trig_a <- data.frame(ngram = character(), freq = numeric())
  trig_b <- data.frame(ngram = character(), freq = numeric())
  inp <- tolower(inp)
  inp <- gsub("[[:punct:]]", "", inp)
  regex <- gsub(" ", "_", sprintf("%s%s%s", "^", inp, "_"))
  indices <- grep(regex, data3$feature)
  if(length(indices) > 2) {
    trig_a <- data3[indices, ]
    trig_a$probability_percent <- round(((trig_a$frequency/sum(trig_a$frequency))*100), 2)
    trig_a$target[1] <- strsplit(as.character(trig_a$feature), "_")[[1]][[3]]
    trig_a$target[2] <- strsplit(as.character(trig_a$feature), "_")[[2]][[3]]
    trig_a$target[3] <- strsplit(as.character(trig_a$feature), "_")[[3]][[3]]
    trig_a2 <- trig_a[1:3, c("target", "probability_percent")]
    print(trig_a2)
  } else {
    y <- paste(strsplit(inp," ")[[1]][2], collapse=" ")
    regex2 <- gsub(" ", "_", sprintf("%s%s%s", "^", y, "_"))
    indices2 <- grep(regex2, data2$feature)
    if(length(indices2) < 1) {
    x <- c("no_data", "no_data", "no_data")
    y <- c(0, 0, 0)
    trig_b2 <- data.frame(target = x, probability_percent = y)
    } else {
    trig_b <- data2[indices2, ]
    trig_b$probability_percent <- round(((trig_b$frequency/sum(trig_b$frequency))*mutation), 2)
    trig_b$target[1] <- strsplit(as.character(trig_b$feature), "_")[[1]][[2]]
    trig_b$target[2] <- strsplit(as.character(trig_b$feature), "_")[[2]][[2]]
    trig_b$target[3] <- strsplit(as.character(trig_b$feature), "_")[[3]][[2]]
    trig_b2 <- trig_b[1:3, c("target", "probability_percent")]}
    print(trig_b2) 
  }
}

Sample output

plot of chunk screen

Find the app here: https://deadpool.shinyapps.io/Predictograph/