Dr. Hossein Eslambolchi
March 2012
The application of Natural Language Understanding (NLU) was once limited to the bridge of USS Enterprise. But although NLU is still not ready for major deployment in advanced applications – think call center applications – the technology has matured into a real-world product.
Currently the main barrier to progress on NLU is computing power and not artificial intelligence. Today’s computing platform is inherently limited by Moore’s Law; and this limit does not afford the computing power necessary to apply NLU.
Quantum computing environments, however, allow much faster computing, and will enable computers to achieve a much deeper, real-time understanding of vocabulary in the next few decades.
KEY POINTS
• NLU will someday allow machines to fully understand unconstrained human languages as well as people do; however, it will be awhile before this goal is achieved.
• Recent advances in applied NLU have come almost entirely from statistical approaches.
• NLU technology can be used with both spoken language and text.
• Progress in NLU has enabled task-specific applications to accept natural language input. Global service providers are offering advanced voice services based on NLU today; similar services are also slowly being deployed in smart phones.
• NLU can partially automate the retrieval of customer service information. Search functions can also benefit from NLU dialogues.
• Prompt-constrained, directed-dialogue systems can simulate some features of NLU by enhancing existing speech-recognition keyword-driven systems.
• NLU technology for spoken dialogue is significantly different than what is commonly used for text-based search engines.
THE TECHNOLOGY
NLU refers to technologies that enable or assist computers in understanding the spoken word or text.
Natural language understanding of human speech goes beyond the capabilities of a speech recognizer that is given spoken input. It takes the output of a speech recognizer, which is a sequence of words, and identifies relevant topics.
One of the primary goals of NLU is to allow a machine to carry on a dialogue with a human.
NLU as exemplified by the shipboard computer of Star Trek or the more ominous computer Hal in the movie 2001: A Space Odyssey remains in the realm of science fiction. However, recent progress has made it possible for computers to receive human language input in natural form and carry out useful actions, including partial dialogue automation.
In theory, NLU can depend on detailed knowledge: lexical, syntactic, semantic, pragmatic and world-knowledge. This approach has made only slow progress, because these tasks are extremely difficult to program. Early attempts to create NLU technology based on language parsing achieved little of practical use.
Modern advances in natural language understanding are based on statistical modeling of the relationships between distinctive utterances and conversational intent. The statistical approach has shown decent progress when applied to specific tasks.
Specific domain keywords and sentences can be understood in any given task, and ambiguities and conflicting information can be resolved. However, realizing a truly comprehensive NLU system may require systems that approximate human knowledge.
Currently, NLU is being applied in information retrieval, question answering and automated dialogue systems for customer service requests, such as those deployed in global service providers’ NLU applications. These systems are built using large datasets collected from user input, and thus require significant customized work. With this “user utterance data”, it is possible to engineer a natural language system to achieve a reasonably high success rate.
The word natural in the phrase “natural language” is inherently ambiguous. In this field, natural generally means machine understanding of normal, unconstrained human speech or text.
However, when many vendors use the word natural, they mean it in a looser sense. For them, natural is equivalent to “conversational”. Most currently-deployed, speech recognition systems that claim to be natural language systems are simply prompt-constrained systems that employ a conversational style as a façade for a system looking for key words or phrases. Speech recognition alone suffices to accomplish these tasks. To a casual observer these systems may appear to be using NLU technology, but there is no actual understanding on the part of the system.
Many vendors boast natural language processing engines that are designed specifically for text input rather than speech. Information retrieval based on natural language queries is one important application of these engines. For example, text-input engines can be used to find resources on the Internet using a natural language query. Unfortunately, the results of these queries are often unsatisfactory given the limits of today’s technology.
There are important differences between text-input NLU systems and speech-input NLU systems. Speech-input engines are designed to be tolerant of the recognition errors that are inherent to any ASR engine and the subjectivities of conversational speech. Text-input engines have not been tested for use with speech and are expected to output text exclusively. Their performance could be catastrophic if they are used to interpret speech.
A natural language semantic markup has been developed by the W3C Voice Browser Working Group and has been incorporated into Voice XML 2.0, which is an extensible markup language (XML) for the creation of automated speech recognition (ASR) and interactive voice response (IVR) applications.
THE PLAYERS
• Nuance provides NLU in its Say Anything offering. Nuance has deployed this technology for Sprint PCS.
• AT&T utilizes NLU in their call center applications.
POTENTIAL IMPACT
Global service providers have been a leader in speech technologies including speech/audio coding, signal processing, speech recognition, speaker recognition, text-to-speech synthesis and natural language understanding.
In 2000 IBM was the first to deploy a high-volume natural language understanding system for customer care operations. Since then, business has committed to commercializing and extending these technologies to revolutionize the electronic contact market.
Tier 1 service global service providers began introducing NLU services in 2007; several commercial customer applications and internal applications have also been deployed. Compared to current touchtone and directed dialogue applications, NLU applications provide greater customer satisfaction and are much faster to use.
Global service provider’s NLU capability is based on advanced machine learning technology, which computes statistics for associating strings of spoken or written words to different possible user intentions. It also includes a named entity model for detecting and extracting phrases that can facilitate complete dialogue automation. This approach to NLU is likely to become the standard in the industry over the next several years and remain the standard – until the next generation of closer-to-natural language understanding is developed.
The market for NLU integrated voice response is likely to emerge dramatically over the next several years. Microsoft and Google have an early lead in the necessary technologies, but competitors have advantages in marketing and in the development tools required for speedy execution.
Only with continuing investment in this area will service providers be able to leverage their “first to market” position.
