TypeLess - the best prediction app for your typewriter

• goal of the project was to build input prediction web application using data sets obtained from the HC Corpora
• all three subsets (Twitter, Blogs, News) have been used to train and test the application

• mark sentence boundaries
• replace “bad” words, urls, hashtags, nicknames and emails using tags
• identify and replace with tags dates, currency amounts and parts of the addresses
• replace groups of emoticons with tags
• tokenize remaining text removing punctuation and remaining special characters.

• pre-process input text using the same rules as for the main corpora
• predict MLE method predict next using \( n, n - 1, n -2, ..., 1 \) trailing tokens (with default \( n = 5 \))
• combine models using exponential penalty for decreasing length (for example 5-gram based prediction has 10 times higher score than a one based on 4-grams)

Future directions:

• applying spelling correction to reduce noise and the size of the model
• test context based predictions using initial clustering by topic and sentiment using emoticons as the noisy labels
• test alternative storage modes

Interested?

• Visit TypeLess on shinyapps.io