World-wise web? Finally on the horizon are computers that can reason
Richard Waters, Financial Times, March 3, 2008
Excerpts from the article:
...As Google shows, being able to return a string of websites in response to a query can give rise to a multi-billion dollar business. With so much at stake, even small incremental improvements on the road to AI may create big business opportunities. "It isn't about being perfect," says Barney Pell, chief executive of Powerset, an ambitious new search company. "It's about being able to differentiate enough to make a commercial product. People are realising that the goals of AI may be way out, but in the field of AI the time is here for really exciting applications."
...Although the companies trying to employ natural language processing admit it is far from perfect, they maintain that technical advances in recent years have at least given it a level of practical application. By using software to "read" text, services such as Powerset and Mr Spivack's Twine aim to add tags to data automatically. The natural language approach also raises the possibility of new applications, for example being able directly to answer questions posed by a user - which has long been a dream in web search.
Powerset has become the most visible champion of this approach. The plunging cost of computing and the wealth of data available on the web have combined to breathe new life into this technology, according to Mr Pell. "One of the big problems was just a lack of computing resources," he says of earlier attempts. Also, refining a natural language search engine requires "a tremendous amount of 'tuning'; you need data to improve these systems". Thanks to the explosion of information on the web, data are not in short supply.
Powerset is using technology licensed from PARC - the famed Silicon Valley research laboratories formerly owned by Xerox - to try to solve the problems of natural language processing. The software is based on similar ideas to those in quantum physics, says Mr Pell. A number of potential meanings for all the elements in the text are allowed to co-exist as equally accurate during the "reading", until the most likely answer is singled out at the end.
Even supporters of this type of natural language analysis limit their claims for the technology, though they say it does not need to be perfect to be useful. According to Mr Spivack, an accuracy level of 70 per cent in analysing and tagging text has its uses...
|
