Can computers comprehend?
Computers can fly airplanes, predict the paths of hurricanes, and beat the best human chess players, but they can't yet do something even small children can do: hold a natural-sounding conversation. To help the rest of us understand the challenge, Associate Professor of Computer Science Lillian Lee demonstrates her point in a 2001 essay with a sentence that can be interpreted in at least ways: At last, a computer that understands you like your mother.
“That very sentence illustrates the host of difficulties that arise in trying to analyze human utterances,” writes Lee. “A moment's reflection reveals that the sentence admits at least three different interpretations … .”
- It could refer to a computer that understands you as well as your mother does.
- It could also refer to a computer that appreciates your affection for your mother.
- And, finally, it could be referring to a computer that understands you as well as it understands your mother.
Lee continues, “…the sentence is ambiguous, and yet we humans seem to instantaneously rule out all the alternatives except the first … . We do so based on a great deal of background knowledge. How are we to get such information into a computer?”
How indeed? Programming a computer with the totality of physical, social, and cultural knowledge available to humans would be a stupendous, perhaps impossible task. Researchers in the field of Natural Language Processing (NLP) have therefore preferred to seek more efficient approaches, either by parsing language into a formal, logical system, or by studying and applying the statistical structure of actual utterances. Lee has been taking the latter approach in her NLP work, using what she calls “knowledge lean” methods. The techniques sound opaque—“Subspace Projection,” “Iterative Residual Re-Scaling,” “Boolean Matrix Multiplication” (papers)—but the upshot is to help computer systems acquire linguistic expertise without expensive, tedious intervention by human interpreters.
The effort has already borne fruit: Lee notes that sophisticated statistical models, large samples of real human speech, and ever-improving techniques for computer learning have “led to our achieving the milestone of commercial-grade speech recognition products capable of handling continuous speech ranging over a large vocabulary.”
Lee has also been focusing on such applications as automated construction of thesauri and teaching computers to find word boundaries in “segmentless” languages like Japanese, characterized by hard-to-find spaces between words. Computers must be taught to compensate for this. There's still much work to do, however, before users can actually converse with their computers, a la HAL-9000 in the classic film 2001: A Space Odyssey. Understanding human utterances, after all, entails more than observing word frequencies, including grasping matters like subtext and rhetorical devices such as sarcasm and understatement.
“That is language ‘in the wild,'” she says. “You can't get away from it.”