Javelin

After I replied to a comment about my earlier post regarding artificial intelligence, I promptly forgot about the topic for several hours. This morning, I picked up one of the books that I'm reading -- The Last Ship -- and, wondering if one of the tools mentioned repeatedly in there, a sighting device called Big Eyes, really did exist, I did a web search, finding that yep, it sure did. Somehow, being at the PC reminded me of my search the other day regarding the IBM Piquant system, and I thought it might be fun to look to see what else might be readily available out there on the same topic. In this manner, I came to a product known as Javelin -- Justification-based Answer Validation through Language Interpretation. Its a pretty slick concept, and puts the lie -- well, partially -- to my glib statement about AI being mostly language parsing. The article cited includes this:

(P)attern-based approaches don’t represent the meaning of the patterns they use, and it is not clear whether they can be generalized for more difficult, non-factoid questions....

Practically useful question answering in non-factoid domains e.g., intelligence analysis) requires more sophisticated question decomposition, reasoning, and answer synthesis. For these hard questions, QA (Question Answering) architectures must define relationships among entities, gather information from multiple sources, and reason over the data to produce an effective answer. As QA functionality becomes more sophisticated, the set of decisions made by a system will not be captured by pipelined architectures or multi-pass constraint relaxation, but must be modeled as a step-by-step decision flow, where the set of processing steps is determined at run time for each question.

Now, I have to admit that it took me five minutes to repeatedly read that before I understood it, and even then, I'm sure I didn't get it all. But what I did get was this: systems such as Piquant could be useful for answering straightforward quetions. (Not sure about the reference to factoids, which I've always taken to mean 'stuff that could be true, could be made up, but in either case is of a trivial nature'; they seem to mean 'its true, but perhaps not all provable'. Thats my guess. ).

But when you get to the requirement to evaluate information -- not just regurgitate it -- you need the ability to dynamically establish a structure where you can meld information from multiple sources, rate it based on the perceived (detected) validity of the source, and then use it to create an answer to the question. As the methodology evolves, it could become self-modifiying so that the steps it takes to answer one question might change as the next question comes along.

My first reaction is, sure this stuff will work on any but the most constrained scale -- aka 'then a miracle occurs' -- but actually, its kind of cool. Obviously, there are some pretty bright people out there.