 * Be better about normalizing input text.
 * Fix english assumtions about single-letter words besides I and a.
 * See about making the parsing into phrases and words more configurable.
 * Investigate using longer-order chains to improve generation quality.
 * Try to use less memory!
