Teaching computers to speak plainly Analysis/Commentary My Fair Lady's Professor Higgins asked, "Why can't the English teach their children how to speak?" Today we might ask, "Why can't scientists teach their computers how to speak plainly?" While Professor Higgins was struggling with the problem of local dialect, today's computational linguistics professors struggle with the problem of localized thinking in computers. Most computers understand language very narrowly. Unlike humans, they don't grasp context when searching for meaning, they simply match facts. A computer can answer a simple natural language query, but usually can't sift out the core meaning of a complex natural language sentence. It has trouble distinguishing between "time flies" and "horse flies." For a human, one is a metaphor for time passing quickly, and the other refers to pesky insects. According to James P. Hogan's book Mind Matters (1997), computers must acquire a sensitivity to connotations that allows them to give a sentence appropriate meaning. Hogan writes that experimenters found that humans recognize similar meanings fast. Human response time increases for words like "spy," "eavesdrop," and "lurk." Grammar seems hardwired in the human understanding, leaving an individual's attention free to focus on clusters of words and their related meanings within a context. Advanced computers understand complex sentences only by first parsing them -- dividing them up grammatically into component parts -- and then looking at individual components for meaning and relation within the larger rule-governed whole. One computer strength gives hope to experimenters -- the power to branch indefinitely, and to keep meanings in suspense until needed. Given a better focus on meaning, computers may eventually acquire better understanding skills. Already many software programs use word-matching effectively to answer text queries -- for example, the Net search program askJeeves.com. Given a larger base of understanding of the relations between words and their context, computers might even eventually learn to grasp idioms and far-fetched metaphors. But there are pitfalls. A few years ago, Joseph Weizenbaum devised a software psychotherapist, a robot program called ELIZA, that gave the illusion of grasping subtle meanings. In reality, the program mimicked understanding, just as a human psychotherapist might pronounce "I see," while not seeing or solving anything. Natural language research plods on undeterred after thirty years of often unsatisfactory results. But computer scientists tend to see the glass as half full, and already there are signs that the glass is getting fuller. Microsoft and IBM vow to bring the magic of natural language processing -- plain English -- to their software, and still other companies offer "personal assistant" programs that provide reliable online help for expert queries. July 21, 1999