Tuesday 15 September 2015

Is a phrase the natural unit of language?

If we look at our most basic communication needs, a phrase is the natural unit. When we need to just get by in a foreign language, we don't try to understand structure or grammar, we just have a book of phrases - a phrase book.

  • "two beers please"
  • "pour aller a la gare s'il vous plait"
  • "my hovercraft is full of eels"

A phrase is the simplest verbal structure to convey a full message. A young child may use individual words in a semi formed way, but he/she is using a phrase. Just a very simple, ungrammatical one that becomes more formal as learning takes place.

  • "bikkit"
  • "bikkit plis"
  • "could I have a biscuit please?"
A connectionist model gives a really good account of this. Connectionist models are good at abstracting what's common between different stimuli. Presented with many repeats of the phrases:

  • "two beers please"
  • "two biscuits please"
  • "two pies please"
  • "three beers please"
  • "four beers please"

a connectionist network is probably going to be able to abstract the 'magnitude - noun - otherthing' phrase structure from the commonalities.

I'd love to spend some time working on that - I'm pretty sure that a fairly simple network would be able to generate novel utterances from repeated phrase presentation. Maybe a bidirectional model, with a semantic module linked to a language module. Sadly work gets in the way. Perhaps some kindly soul at Google would like to give me a job researching natural language interpretation by neural nets ;-)

No comments:

Post a Comment