Magazine article AI Magazine

The Value of AI Tools: Some Lessons Learned

Magazine article AI Magazine

The Value of AI Tools: Some Lessons Learned

Article excerpt

Artificial intelligence is a hot commodity, and for the first time in decades we are seeing the emergence of companies whose primary business purports to be AI. I'm curious to see whether these companies can sustain themselves in the long run, especially since AI has traditionally delivered most of its value under the hood. Today, AI plays an important role behind the scenes in industries from finance to manufacturing, but it's only recently that there has been enough interest to raise its visibility to the forefront of a company's value proposition.

One area with a history of commercial interest is web-based information extraction. The advent of the World Wide Web in the 1990s created a new set of opportunities for applied AI, and for many people, Google's highly public focus on AI and machine learning helped validate the growing importance of these technologies for business. In fact, there have been many other companies, big and small, that have also worked on technologies for understanding and extracting web data. In this column I will comment on some of the thorny business issues, as well as opportunities, that I have seen in this arena, which continues to be a strong area of interest for both AI researchers as well as practitioners.

Personally, I became involved in research on web data extraction at the University of Southern California's Information Sciences Institute in the mid-1990s, when I worked on a series of research projects for extracting data from semistructured sites, with Craig Knoblock and other colleagues. Late in 1999, at the height of the Internet boom, we launched Fetch Technologies, Inc., to commercialize this work. At the time, companies with less capable technology were selling for many hundreds of millions of dollars, and we were looking forward to becoming very wealthy. Unfortunately, just as we brought our first product to market - a machine-learning tool for extracting data - the Internet bust caught up with us and we ran out of cash.

Rather than shut down the company completely, we kept a few people, and with money from federal research grants, we were able to limp along. After the bust, there wasn't much interest in our initial product, but we were able to continue baking the technology and after several years of hard work produced a very solid product. With just a few examples, the Fetch Agent Platform could be taught how to navigate through a website and extract semistructured data (for example, data on web pages). So, for instance, one could easily show the system how to extract from a telephone directory site, including filling out the relevant forms on the site, and the system would effectively reverse-engineer the site, extracting a database of harvested data. To achieve this, we developed a series of increasingly sophisticated machine-learning methods for data extraction (Gazen and Minton 2005; Knoblock et al. 1998; Lerman, Minton, and Knoblock 2003; Minton, Ticrea, and Beach 2003; Muslean, Minton, and Knoblock 2006) as well as a data flow architecture that could harvest and extract data at scale (Barish et al. 2000).

The Fetch Agent Platform was a very impressive AI system, one that I'm still proud of, but our business never quite hit that inflection point we were looking for. On the plus side, we eventually did have some important successes. For instance, Fetch was used by many background-checking companies to harvest criminal records from county and state court sites; and the company was acquired by Connotate, which still uses the technology today. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.