Robert: That definitely looks useful. One aspect of chatbots I never wanted to bother much with was basic math (especially arithmetic and algebra), since it seems like such a “solved problem”. But since it definitely comes up in conversation—and since it’s critical to processing the many “word problems” that come up in everyday situations—having a good math module is probably essential to a realistic bot.

As for the other structured facts WA can produce, I’m hoping my bot will be able to cull this itself from NL. But never look a gift horse in the mouth. And who knows? Could be useful for NLP training as well.

Carl: Thanks for the link, I’ll look into it!

I believe in the future most bots will use external resources to supplement their “on-board” knowledge. There is little need for every bot or botmaster to duplicate and curate wikipedia or a dictionary. Most of the major search engines provide apis that you can use to generate queries. I find that different search engines provide different (some better some worse) results. In Skynet, I use a cascading search engine approach where in some cases if an answer is not available via one engine, it tries a second.

Beyond the search engines, DBPedia and Freebase are good for “facts”.

C R Hunt - Jun 17, 2011:

Robert: That definitely looks useful. One aspect of chatbots I never wanted to bother much with was basic math (especially arithmetic and algebra), since it seems like such a “solved problem”. But since it definitely comes up in conversation—and since it’s critical to processing the many “word problems” that come up in everyday situations—having a good math module is probably essential to a realistic bot.

I am still working on my Skynet-AI math module. The problem with sending a math query to a search engine is that they often require specific formatting (minimal natural language) to handle the problem. Most can’t do word problems.

Robert: did you find any repositories of UMIA data in your travels? I wanted to take a look at some of the standard queries (I believe they are regular expressions).

@C R Hunt: I got the idea for a math agent from the many irc bots that have calculator functions. The difference is, my agent doesn’t require a trigger such as “.c”, so I can ask it “what is 75 F in Centigrade”, “convert 2 m/s to mph”, etc., which is more like natural language. Google Calculator makes a nice math agent that I’ve been using; now I’ll add Wolfram Alpha and see which of them I prefer. My architecture is designed so that users can customize which agents respond to which inputs. So if Google Calc proves better at some types of math while Wolfram Alpha is better at others (Google doesn’t do calculus or solve algebra equations, for example, but is probably faster for simple math expressions), I can “teach” the system through feedback to use one agent for some inputs, and another agent for other inputs.

@Merlin: The download for the UIMA framework and SDK (http://uima.apache.org/downloads.cgi) includes several annotators, which I’ve played with a bit and found not really compelling enough at this time for me to mess with Java again :)

For example they have a SimpleNameAnnotator that seems to be a regex that looks for two titlecased words in succession, so it recognized “Get Involved” as a name; also it seems to look for a capitalized word followed by another capitalized word with a period at the end followed by a third capitalized word, so it recognizes “Thomas J. Watson”, but if a name ends a sentence the first word of the next sentence will also be included in the recognized name: in other words “John saw Thomas Watson. He asked…” will give the name as “Thomas Watson. He”. Also, it doesn’t recognize single names like “John”. So I think I can do better without the overhead of their cumbersome UIMA framework :)

@Robert: Google calculator is very impressive at picking out when a search query can be “solved”, even when sloppy users (and I’m probably among the worst) try to trip it up.

Definitely keep me informed about how incorporating WA works out.