Feb. 6, 2003
Vol. 22 No. 9

current issue
archive / search

    Niyogi uses computers to analyze language evolution

    By Steve Koppes
    News Office

    If a computer could master language as well as a child does, the feat would rank as one of the greatest technological achievements of our time. But so far, computers fall far short of the capability.

    “How do children learn the language of their parents with seemingly effortless ease?” asks Partha Niyogi, Associate Professor in Computer Science, Statistics and the Physical Science Collegiate Division. Linguists, psychologists and computer scientists specializing in artificial intelligence would all like to know how to answer that question.

    The computational analysis of how language evolves may well hold the answer, suggests Niyogi, who is completing a book on the topic. That is because children imperfectly learn the language of their parents.

    “If language acquisition is perfect, then language is successfully transmitted from generation to generation. You would expect that languages would remain stable over time. They wouldn’t change,” he said.

    But languages do change with time. Today’s English, for example, reads and sounds much differently than it did 1,000 years ago.

    “The changes that I’m talking about are deep,” Niyogi said. “It’s not like a word takes on a new meaning because you have computers today and you didn’t have computers 1,000 years ago.”

    In modern English, for example, a verb precedes its object. Modern English speakers say, “I eat apples.” But centuries ago, speakers of old English would have said, “I apples eat.” Old English shares this trait with many languages, including German, Japanese and Niyogi’s native language, Bengali. “This is entirely arbitrary. Both choices are equally efficient,” Niyogi said.

    Linguists and philosophers have known since the 19th century about the relationship between language learning and language evolution. But only in recent years have researchers such as Niyogi developed the computational tools to quantify this relationship.

    Language lends itself to computational analysis because it is a formal system, Niyogi said. Its words, phrases and sentences conform to a set of rules. Niyogi’s computational method for studying language evolution combines a given language theory with a given learning theory, both numerically expressed, and deduces their evolutionary consequences.

    “What we want to study is the interplay between learning by individuals and evolution in populations,” he said, just as biologists study how the transmission of genes from parent to child affects population genetics.

    Even Charles Darwin commented on the similarities between linguistic and biological evolution in The Origin of Species. And although Darwin presented a compelling case for how organisms evolve, he stated his case without the benefit of statistics, Niyogi said. It remained for–among others–the late Ronald Fisher, an English statistician and biologist, to develop a statistical model for Darwin’s theory of natural selection.

    “That really revolutionized evolutionary biology by clarifying its basic principles,” Niyogi said. And just as biology benefited from the development of mathematical models, linguistics did so from the models Noam Chomsky developed in his work. But only in recent years have scholars such as Niyogi attempted to apply computation to understanding language evolution.

    Niyogi’s ultimate goal is to build computer systems that can interact with and learn from humans. The first step is to teach computers how to translate sounds into words. Niyogi has devoted considerable effort toward developing a numerical method for detecting stop consonants in speech, such as the “d” sound in “dark.”

    “There’s a period of silence where there’s no sound coming out, and then there’s a sudden, sharp, noisy burst. That burst can be as short as 10 milliseconds,” Niyogi said.

    “It’s a feeble, fleeting event in the acoustic stream, and yet you have to be able to detect it if you want to understand speech, because these consonants occur a lot.”

    Niyogi is attempting to understand the principles necessary to build a speech recognition system. Although his interest in the system is intellectual rather than commercial, the work has technological implications for further development of hand-held computers. If users could simply speak into their hand-held computers, they would no longer need to struggle with tiny keyboards.

    “It would be a tremendous advancement in user interfaces,” Niyogi said.