German SwiftKey

• Use data from diffent sources: Blogs, news and twitter.
• Cleaning and counting word sequences.
• Selecting only the most common.
• Giving priority to different options: sequences of 3 words, 2 words, etc.
• By default showing the most common words.

Here we can see how the model predicts based on the text:

text <- "ich"
nextWords(text)

[1] "bin"  "habe" "mich"

text <- "ich bin ein"
nextWords(text)

[1] "großer" "sehr"   "freund"

text <- ". "
nextWords(text)

[1] "die" "der" "und"

Go to the link here

And type!

You will see always 3 words as suggestions for the next one!

And the final prediction! Easy! Isn't it?

After finishing this project, the team can conclude:

• The predictions are fast and accurate enought to help users.
• Analysis of German language was useful to reduce the number of words and by doing so, the memory required to save the data.
• The model can be applied to other language with no major effort.

Out team needs the support to continue with these ideas:

• Use word embeddings like word2vec to get better predictions for words that are not in the corpora.
• Include corpora with part of the speech tags to take probabilities of sequences (for example Verb -> Object -> Preposition) into consideration.
• Implementation of Recurrent Neural Networks (RNN) to predict the next word.