Ivan Herman recently offered some insight into how Watson actually works. Herman reports, “I was at Chris Welty’s keynote yesterday at the WWW2012 Conference. His talk was on Jeopardy/Watson and, although this is not the first time I heard/saw something on Watson, some things really became clear only at his keynote. Namely: what is really the central paradigm that made the question answering mechanism so successful in the case of Watson? Well… query answering in Watson is not some sort of a deterministic algorithm that turns a natural language question into a query into a huge set of data. This approach does not work.”

He continues, “Instead, a question is analyzed and, based on search in various set of data, a large set of possible answers is extracted. These ‘candidate’ answers are analyzed separately along a whole series of different dimensions (geographical or temporal dimensions, or, which I found the most interesting, putting back candidate answers into the original question and search that again against various sources of information to rank them again). The result is a vector of numerical values representing the results of the analysis along those different dimensions. That ‘vector’ is summed up into one final value using a weight values for each dimension. The weights themselves are obtained through a prior training process (in this case using a number of stored Jeopardy question/answers). Finally, the answer with the highest value (I presume over a certain threshold value) is returned.”

Read more here.

Image: Courtesy Flickr/ LeWEB11