3 ms·
There's a non-zero probability of transitioning from "had" to "had", but I don't understand how there could be a non-zero probability of "had" following "had ha
by staticautomatic 6y ago
There's a non-zero probability of transitioning from "had" to "had", but I don't understand how there could be a non-zero probability of "had" following "had had" or of "had had" following "had had" unless the model is being trained on its own output.
- ninjanomnom 6y agoTraditionally only the last output is known by a markov chain and when it comes to chat bot generation the output is usually one word at a time. The bot doesn't track that the last 2+ words it output are the same, all it knows is that the last word it said is 'had', so what next? There are countless markov chain chat bots made though and plenty bend the rules, to some extent though they are all susceptible to this. In an old irc chat of mine we had a lot of fun abusing one such bot to get it stuck in these kinds of loops, in our case it was the word 'fuck'. You'll have to forgive our teenager sense of humor.
- JHonaker 6y agoIt's perfectly fine, theoretically for the Markov chain to depend on more than the previous token. When you start to do that though, unless you can write down a function (or approximation, like an NN) the number of states in your state space that you have to enumerate explodes very quickly. This is the simple motivation behind stuff like GPT or neural reinforcement learning.