Keynoter: Natural Language Processing Has Huge Potential

NEW YORK (SpeechTEK 2010) — Using natural language processing (NLP) techniques to expand the range of speech offers a “huge amount of promise,” said University of Rochester professor James Allen during his August 3 keynote at SpeechTEK 2010. 

Among the examples cited by Allen, author of Natural Language Understanding and the John H. Dessauer Chair of Computer Science at the University of Rochester, N.Y., are interactive planning, human-robot interaction, medical advising (managing patient medications), and interactive data exploration. “This is the future. It’s what we really want systems to do,” he said.

Today’s systems, and what the technology will be able to do in the future, represent a “dramatic change” from the old days when speech application developers “built everything by hand—slowly and rather painfully,” he said.

NLP experienced “a huge revolution in the 1980s” with the introduction of statistical modeling, Allen said; since then, NLP has achieved up to 97 percent accuracy for parsing parts of speech and speech tagging.

One of the earliest applications for NLP was machine translation, which when it first came out in the 1980s used very complex models and a lot of computational power, according to Allen. Since then, NLP has been used for information extraction, message understanding, sentiment analysis, question answering, summarization, and more. “There are a wide range of applications, and the field has exploded,” he said.

NLP, he explained, “has a fairly robust ability to abstract meaning from sentiments,” with success rates in the range of 85 percent.

But even today, NLP has its challenges, the least of which is customer sentiment, according to Allen. “From a consumer perspective, people find these systems frustrating most of the time,” he said, “and from a developers’ framework, the tasks that can be accomplished are fairly limited.”

He added that one of the problems with current NLP technology is “there’s a lot of off-the-shelf NLP capabilities, but often shallow processing,” something that can be overcome.

NLP can improve modern speech systems by combining them with other technologies, creating opportunities for speech-to-speech translation, spoken dialogue systems, and “more complex systems handling more complex tasks,” Allen stated.

To show what’s possible with NLP systems, Allen demonstrated a medical and a mapping application that he built, both of which received high performance rates.

But in the final analysis, “significant progress has been made in the last few decades, significant opportunities are ready for adding speech to applications, and robust, mixed-initiative dialogue systems are more robust,” he concluded. 

Other conference speakers agreed, saying NLP has advanced far in the past 20 to 30 years since it was first introduced into airline flight reservation systems. Srinivas Bangalore, principal technician for voice-enabled services at AT&T Labs, said NLP—or, as he called it, spoken language understanding—has evolved. Among its widest uses, he said during a session August 4, is in mobile search applications. Previously, it had only been used in specific call center apps with limited semantical representations.

For the technology to truly advance, though, Bangalore said it will need to include context, such as geospatial and temporal factors, user preferences, and the interaction context itself, and to normalize semantic variations and support inferences.

During the same session, Kurt Fuqua, CEO of Cambridge Mobile, suggested that for NLP to truly work as it was intended, grammars will need to be more scalable and vocabularies will need to be more dynamic.