#word-level

Question 1: Single choice

A character-level LSTM language model generates strings such as "the tiee traveller", with misspelled words. You want it to produce only real words. What is the main change?

Tokenize the text into words, so its inputs and outputs range over a word vocabulary1
Add more LSTM layers, so that the model memorizes the spelling of every word2
Always pick the most likely next character, so that no unlikely letters are generated3
Increase num_steps, so that every training window contains only whole words4

Timed test