Photosynth looks like an incredible piece of technology. Essentially, this new product from Microsoft's Live Labs, is a next generation photo stitching system which analyses digital pictures for signatures and then uses those signatures to build whole environments in 3d (or something close to it).
It struck me that these signatures are like facts describing the scene. In order for the system to work, these facts must be something like invariants - so that when the pictures of the same objects are analysed, the constellation of facts are constrained enough to provide strong enough clues for matching heterogeneous images.
Now, imagine you have a big collection of text document and you mine them for facts. These assertions about the real world can then be combined and strung together to form more comprehensive accounts of things, people, places etc. For example, I read one document and it tells me that Toronto is in Canada. Another document tells me that Canada is in North America. I can combine them and infer that Toronto is in North America.
One of the challenges with this metaphor is the domain of input. With photos, and it looks like Photosynth is best suited to outdoor scenes with buildings and city scapes, there is a more direct relationship between reality and the resulting image. However, with text, the distance between reality and the information in the document is pretty large. One has to sift objective material from subjective material, facts from opinions, etc.
The above idea - of mining information from multiple documents and combining it - is not new (of course). However, I find the photosynth application appealing as a metaphor when considering text mining.
Comments