search

pteichman / cobe

A Markov chain based text generation library and MegaHAL style chatbot
http://teichman.org/blog/
MIT License
242 stars 51 forks source link

Babbled replies are longer than non-babbled on average #2

Closed pteichman closed 14 years ago

pteichman commented 14 years ago

When cobe doesn't recognize any word in reply()'s input, it generates replies based on random tokens in the database. The resulting replies are likely to be much longer than replies seeded with a known word.

This might be: 1) Scoring rating those sentences higher? 2) Using random tokens rather than random words? (i.e. babbling on punctuation/space tokens)

pteichman commented 14 years ago

Generating a single random token and using that as the pivot for all replies seems to improve this behavior, but I'm not sure if shorter replies are better.

pteichman commented 14 years ago

Time to close this one.