This 2009 GrokItBot:// A Python, AIML and Bayesian IM bot just popped up on my radar….  It seems to be a novel application:

>>With a Bayesian guessing filter, GrokItBot should take a guess at the topic of conversation and, if it finds no suitable response, ask for clarification.<<

- What exactly is a “Bayesian guesser”?

- And, are there other examples of this type of application?

Hi Marcus,

I’ll try and resist quoting wikipedia on this. Bayesian “prediction” or inference is based on a statistical probability distribution that uses existing “beliefs” to predict future outcomes.  I’ve never seen it used in any practical applications, but it comes up in lots of theoretical AI papers. It is possible to write natural language parsers built on compression algorithms that also use the Bayes’ theorem to pick out pertinent bits of information from a time series or input stream.

Robby.

Bayes rule has been used in identifying spam in email. It also has applications in sentiment analysis and recommendation engines.

I did some prototyping predicting concepts/responses based on input. The key is a large enough corpus to get the statistics right.

So, would that be like Siri, who, when reading the keywords “hungry” and “dinner”, assumes that food is the topic because those words frequently occur in combination therewith? (I’m not sure if Siri uses Bayesian methods, but from the descriptions it does sound similar)

I find that Bayes Theorem is more useful for decision making or limited classification.

If you have a smaller set of choices, and a training set of examples that select from those choices, you can use it. It answers the question, what is the probability that a given choice among the choices is the correct one.
This is good for sentiment analysis (positive|negative) or Email validity (spam|not spam).

https://class.coursera.org/machlearning-001/lecture/243
http://www.programminglogic.com/bayes-theorem-with-examples/

For topic selection, because of the sparseness of Natural language, and the fact that some words which may show up less frequently may influence topic selection more, I like TF-IDF better.
http://en.wikipedia.org/wiki/Tf–idf

Topic Modeling
http://videolectures.net/icml07_mimno_moht/