Where Am I? I'm Right Here!
We are, or are soon to be, entering the age of natural language interfaces. One area in which they are already appearing is in the specialised application interface space. For example, there are a number of online calendaring systems which take expressions like 'go swimming every Thursday at 4pm' and do the right thing with them. One could argue that these interfaces know a few good tricks that they teach the user and then get away with limited actual NLP technology. However, they are doing one important thing: changing the way in which people expect to interact with systems.
One of the great advantages of deploying these systems in limited domains is that they require a limited set of linguistic capabilities to do a small set of tasks reasonably well. This is something that should be exploited as much as possible.
One domain in which there is a comprehensive failure to do this is in the area of driving directions.
A key aspect of dialogue is the notion of context. I can get away with saying 'it' because I have already established the point of reference and conventions tell the listener how to resolve 'it'. Contextual knowledge leads to efficiency in communication. If, for example, I want driving directions, the standard web interface works like this:
- Enter the starting address.
- Enter the target address.
- Press a button.
- Enjoy your map.
If I enter '123 Foo Street, Pittsburgh, PA, 15206', when I enter '456 Bar Street' can you guess which city that would be in? Take a wild guess? Atlanta? Ok, imagine that there are 3 cities with Bar Street: {Atlanta, Pittsburgh, San Jose}, could you possibly select the most likely from that set? Google, Microsoft and Yahoo! can none of them figure it out. (The answer is Pittsburgh BTW).
Comments