I used this Stanford university text to understand the theory.
This is my summer vacation june 2024 ML practice in which I aimed to make a words auto-complete program that is similar to how your mobile keyboard suggests word based on already typed part of message.
This is based on conditional probability where probability of next word given already typed words is calculated and the highest probability word is predicted. For example:
p(time | a person who thinks all the) = probability of 'time' given 'a person who thinks all the'
We know that in data, there is only one this absurd quote so the pribability of 'time' is highest given existing string.
nis the number of last words it will consider. If n is 3, it will only try to predict the sequence starting from last 3 words (i.e., it will predict the 4th word after the last 3 words).- I rebuilt the entire model every iteration, which results in slow compute. If you only build model once, it will be extremely fast (but I believe it would be inaccurate).
To personalize it for your friends group, make sure to extract all group-chat messages and put them in a .txt file and feed the file to this model. Currently it uses shakespeare and alice in wonderland text so it gives weird results.