Automated Natural Spoken Dialog

E ngineers have long sought to design systems that understand and act upon spoken language. Extracting meaning from natural, unconstrained speech over the telephone is technically challenging, and quantifying semantic content is crucial for engineering and evaluating such systems. Traditional menu-driven speech recognition systems force users to learn the machine’s jargon, but many people are unwilling or unable to navigate such highly structured interactions. AT&T’s “How May I Help You?” (HMIHY) technology (http://www. research.att.com/~algor/hmihy) shifts the burden to the machine by requiring it to adapt to human language and understand what people actually say rather than what a system designer expects them to say. The intuition underlying our approach is that for a given task, some linguistic events are more salient—crucial to recognize and understand—than others. Researchers have already developed methods to automatically extract named entities such as phone and credit card numbers embedded in natural spoken language and to translate utterances into Spanish and Japanese. Building on these and other early laboratory experiments, we have developed algorithms that automatically learn the salient words, phrases, and grammar fragments for a given task far more reliably than other methods.