October 10, 2003
Bio·IT World's Mark D. Uehling spoke with William Hayes, project lead for text mining and bioinformatics at AstraZeneca.
Q: What's changed in the vendor community?
A: Bioinformatics and drug discovery and text mining have started taking off. There are about a hundred companies out there trying to get our attention. Overall, the market is headed in the right direction. There are a lot of very good point solutions. But there aren't any very good integrated solutions.
What do you recommend?
You really want to pull in as much text, have as many different sources as you can, and run your analyses across all of them, because the results reinforce each other. The statistics get better the more data you have.
Text search. I thought that would be one of the easiest projects. It was really hard. It seems like a mature technology. I consider 4 gigabytes of text not that large, especially when Google can take the entire Internet and turn it around in a few seconds. We have the requirement that we do a term search and get back the result, on average, in under a second. It was quite daunting to find something that could manage that.
Can vendors deal with the quantities of text you're expecting?
They're not able to scale up to all the text we're interested in. The few integrated solutions I've seen so far have been looking at a few hundred to a few thousand documents. We're looking at 40 gigs-plus — over 10 million documents.
What solutions do you like for text categorization?
For our particular purposes, Reel Two fitted us quite well. We've not been interested in doing really large-scale categorization. Their interface lends itself well to the domain experts who build up a small, focused, filtering process computationally. We've also been looking at using Reel Two's engine for term disambiguation.
What stands out in what you've seen so far?
The NLP [natural language processing] solution we're testing right now is pretty impressive. I consider it the crown jewel of our text-mining capabilities. We can do a protein-protein interaction query and get very accurate results.
Back to Digging Into Digital Quarries