Natural language processing
Speech recognition takes sounds and turns them into words, or at least identifies particular sounds and associates them with certain words.
Understanding the meaning conveyed by a human language is another matter. The two are connected, as you might want a system to make sense of a stream of words generated by speech recognition software (and that is what happens in some of the systems outlined above), but they are separate functions.
Just as you might put a voice interface in parallel with a Web interface to a back-end system, a natural language analyser can be fed with text produced by a speech recognition subsystem or merely typed in at a keyboard.
A feed from a speech recognition module might give better results if it was able to detect and provide information about which syllables were stressed and where any pauses fell in a particular sentence.
Consider the sequence of words -Woman without her man is nothing"â€"supposedly half the population parse it as -Woman, without her man, is nothing" while the remainder take it to mean -Woman! Without her, man is nothing". An old chestnut, admittedly, but it does illustrate the problem, because none of us would have any trouble determining which meaning was intended if the sentence was spoken naturally.
In less contrived situations, interpreting our first language seems quite straightforward to most of usâ€"after all, we've been doing it for most of our lives. Getting a computer to do it is a different matter.
One of the problems is that words and phrases can have multiple meanings, and very often the sentence containing such a word does not provide sufficient context to determine which meaning is intended. A whole class of jokesâ€"including -the world's funniest joke" (see www.laughlab.co.uk/winner.html)â€"relies on this, or at least on the existence of homonyms with very different meanings.
While natural language processing (NLP) is still a work in progress for computer scientists, there are some tools existâ€"such as Simplis' Zlangâ€"that can make it easier for an application developer to support natural language queries.
It's relatively easy to do NLP when the context is restricted, and that's why speech interfaces for telephone betting services have proved successful. The wider the domain covered, the harder it gets. For hands-on experience of how NLP systems can react, try MIT's START system â€"-what's the weather in Sydney?" is answered sensibly, but -Did Australia beat England in the test match?" isn't understood at all.
Microsoft has a substantial NLP research group, and its work has provided the technology behind the Office grammar checker, Encarta's ability to answer questions, and the IntelliShrink feature in Mobile Information Server that extracts unnecessary words and characters from a message and applies abbreviations to it before relaying the condensed message to a mobile device.
The company stresses that its NLP technology uses an automated knowledge base. Just as the use of a DBMS in a conventional application separates data storage from data processing, this approach means new NLP algorithms can be grafted onto an existing knowledge base.
Systems built on hand-coded data do not have this flexibility. Microsoft's experimental MindNet system contains information about interrelationships between words that it has generated from an analysis of texts including two dictionaries and Microsoft's own Encarta encyclopaedia.
