2011/3/3 Petr Baudis <[email protected]>: > On Thu, Mar 03, 2011 at 04:44:46PM -0500, Colin Kern wrote: >> On Thu, Mar 3, 2011 at 4:19 PM, Alain Baeckeroot >> <[email protected]> wrote: >> > >> > Le 03/03/2011 13:47, Don Dailey a écrit : >> > >> >> I am rather curious how they designed the software to answer trivia >> >> questions and what the problems were. >> >> >> > >> > see http://lwn.net/Articles/427665/ the last 2 comments (#5 and #6) give >> > references to the tech part. >> > >> >> There's also a whole book about it by Stephen Baker of Numerati fame. >> It's a good read. >> http://www.amazon.com/Final-Jeopardy-Machine-Quest-Everything/dp/0547483163/ref=sr_1_1?ie=UTF8&qid=1299188628&sr=8-1 > > Unfortunately, when looking for actual technical details, every popular > article seems to be really shallow. Something like > >> Unfortunately, Baker's book does not provide details on how these >> Linux-managed algorithms and processors are able to work together and in >> sequence. Neither do we learn much about how AI takes eg. any subject >> and determines what types of words are most and least associated with >> it, or Watson's use of some 100+ algorithms to analyze a question in >> different ways, the results of which are then ranked by another set of >> algorithms (eg. "Was the contender alive that year?"). (New York Times, >> 6/20/2010). > > from the Amazon page is pretty much the best I have been able to find. > Did anyone find any further details on the actual "100+ algorithms"? > > -- > Petr "Pasky" Baudis > Computer science education cannot make an expert programmer any more > than studying brushes and pigment can make an expert painter. --esr > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
Hmm, I think the book was more detailed than that snippet lets on, but I might be misremembering. I've been following Watson pretty closely, so my understanding of how it works is put together from a bunch of sources. I'll give a summary of what I've gathered on how it works. First, Watson takes the text of the question and parses it to identify the type of answer it's looking for, e.g. "author" or "decade". It also parses out the keywords and phrases from the question. Next thing it does it search its database, which includes wikipedia, imdb, etc., for sentences that contain these keywords and phrases from the question, to compile a huge list of "candidate" answers. Next, those candidate answers have to ranked and confidence scores assigned. This is where the "100+" algorithms come in. Each algorithm evaluates the candidate answer using a different method or criteria, which might not even be relevant for the question being asked. The outputs from all these algorithms are then used in a weighted voting scheme, which uses a neural network to determine the weights. The network was trained by processing a corpus of past Jeopardy questions. If the top scored candidate answer is above a threshold, Watson will buzz in and answer. Otherwise, Watson doesn't buzz. Colin _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
