fn word(lexicon: &Lexicon, token: &str) -> Option<String>
One word as the model learns it, or none (normalize_token).
normalize_token