Build a bigram next-word predictor with weighted sampling

You are given a training set of token sequences (sentences), for example: ``` [["a","b","c"], ["a","s","d"]] ``` 1) Train a simple **next-word prediction** model that, for each word `w`, counts which words most frequently appear **immediately after** `w` (a bigram / 1st-order Markov model). 2) At inference time, given a current word `w`, output a **random next word** sampled **proportionally to the observed counts** after `w` (i.e., weighted by frequency). 3) Discuss what you would do if the vocabulary and/or number of distinct next-words per token is very large (memory and latency constraints).