We have considered recurrent models that work a word at a time, and models that work a character at a time. It is also possible to use a subword representation, in which, say, the word “searchability” is represented as two tokens, the word “search” followed by the suffix “-ability.”

Run a notebook such as www.tensorflow.org/text/guide/subwords_ tokenizer that allows you to build a subword tokenizer. Train it on some text. What are the tokens that get represented? What words and what nonwords? How can you visualize the results?