Computers to Read Handwriting
Computers to Read Handwriting
Spokesman-Review, February 16, 1997, p. A4
Post Office is spending millions on program to read addresses: Associated Press Merrifield, Va.
Computers are being trained to do the impossible: Read sloppy handwriting. Already in 34 cities, computer software is helping read handwritten addresses. But the system still has a long way to go.
The Remote Computer Reader likes numbers written in third-grade, block style. But the computer finds it easier to read cursive&md;as long as it's not too showy. And no Old English, thank you.
The reader's memory is somewhere inside a blue box of computer circuit boards in an out-of-the-way, gray room at the Postal Service's Engineering and Development Center in suburban Virginia.
"Here it is," program manager Al Lawson says, showing off the read-your-writing machine. "It's not very exciting. It's a box."
About $6 million has already been spent to get the Remote Control Reader up and running in the 34 cities. By the end of the year, the software will be at all 254 of the post office's main processing sites, sorting nearly one-quarter of the estimated 8 billion pieces of handwritten mail that move through the Postal Service every year, Lawson said.
Right now, it reads 10 percent to 14 percent. Although it correctly sorted four of 20 test letters penned by employees of The Associated Press&md;a 20 percent success rate.
It was confused by a sample letter addressed to "200 West Ave. E." Another had too much space between the numbers in the 45140 ZIP code of Loveland, Ohio. The first two numbers in the 61937 ZIP to Lovington, Ill., were touching on a third, but the machine still got this one right.
In all, the machine couldn't sort 16 of the 20 test letters. They would have had to be routed by one of the 22,000 workers nationwide who manually key in destinations.
Tom Fahey, communications director for the American Postal Workers Union, said the new technology could mean the loss of jobs in upcoming years. But he said the workers might be needed for other postal work by the time the system is perfected.
That's Stanley Turk's job.
The systems analyst plopped down in front of a color computer monitor and called up a fictional test letter addressed to: John Johnson on Saddle Notch Drive in Loveland, Colo., 80537.
Sloppy writing caused instant problems.
The first two digits of the ZIP code ran together. Also, the sender wrote an oversized, European-style No. 7&md;one with a horizontal slash through its vertical line.
"It read the seven as a two," Turk said.
The machine also was stumped on a sample letter being sent to 1 Lakeshore Drive in Valentine, Neb., 69201. The machine thought a small, partial circle meant to be a zero was an extraneous mark on the envelope, maybe ink from a messy pen.
Turk says the machines are programmed to filter out other things, too, like postmarks, stamps, logos&md;even "Ed McMahon staring you in the face saying you've already won."
In a half-second, the computer goes through an extensive checklist, gaining confidence at each step about sending a letter to a certain destination.
First, it searches for the address. Then it breaks it into units of information like the post office box number or ZIP code.
"It took more than five years of research just to get the computer to find the address block," Turk says. "It's still not perfect."
Next, the computer looks for a five-digit ZIP, comparing each number with ones etched in its memory. If it can decode the ZIP, it automatically knows the city and state.
Then instead of trying to identify the street name, it tries to decipher the street number. That's easier.
If the street number is 10, for example, the computer searches its memory for all addresses within the ZIP that begin with that number&md;10 Main St., 10 Commerce Ave., 10 E. Greenway Blvd. This list includes all variations such as 10 East Greenway Boulevard.
The next step is a bit tricky.
The computer traces the handwritten street name, keeping track of how many times upstrokes become downstrokes; downstrokes become upstrokes. It approximates the number of characters. Then it tries to find a probable address match on the list.
"If it's too close to call," Turk says, "the computer says: &ls;Let an operator key it.'"