WORLD SCIENCE

"Long before it's in the papers"
December 19, 2005

Aug. 31, 2005
By Susan S. Lang/Cornell University
and World Science staff

Researchers say they have developed a system that lets a computer scan text in various languages, figure out the grammatical rules behind it, then spit out simple, but sensible sentences of its own making. 

The method also works for such data as sheet music or genetic code, the researchers said, and has implications for speech recognition and genomics.

It could also provide new insights into language learning, according to the scientists.

The program “can take a body of text, abstract from it a collection of recurring patterns or rules and then generate new material,” said Shimon Edelman, a computer scientist and professor of psychology at Cornell University in Ithaca, N.Y. 

Edelman co-wrote a paper describing the findings in the Aug. 16 issue of the research journal Proceedings of the National Academy of Sciences.

The system, called Automatic Distillation of Structure (ADIOS), identifies complex patterns in texts by repeatedly aligning sentences and looking for overlapping parts. 

For example, the sentences “I would like to book a first-class flight to Chicago,” “I want to book a first-class flight to Boston,” and “Book a first-class flight for me, please” might tip off the computer that the following is a useable string of text: “book a first-class flight.”

The researchers said in one experiment, they “trained” the system with a group of thousands of normal sentences of this type, all relating to airline travel. As a result, the computer generated sentences including the following: “I would like to book the first class”; “I plan to make a round trip”; “what kind of food would be served”; and “how many flights does Continental have.” 

Because the program can identify patterns that contain other patterns, which in turn contain still more, the computer’s knowledge grows “as a sort of forest of branching trees of possibilities,” said Edelman.

The researchers said they also successfully tested the method with translations of the Bible in several languages.

ADIOS relies on processes believed to occur in human language learning, Edelman added. “This may eventually help researchers understand how children, who learn language in a similar item-by-item fashion and with very little supervision, eventually master the full complexities of their native tongue.”

The system can also help scientists decipher sequences of chemical units in genetic code, the researchers said. For instance, if the code contains information for producing a given molecule, the system can help predict the structure and function of that molecule—a useful service to many biologists. 

The system has also been tested on musical notation, added Edelman, who developed it with researchers at Cornell and at Tel Aviv University.

* * *

Send us a comment on this story, or send it to a friend

WORLD SCIENCE

WORLD SCIENCE