[lojban] Re: some thoughts about lojban use and future

[Rob Speer <lojban-out@lojban.org>]

On Wed, Jan 05, 2005 at 12:36:33AM -0800, xah lee wrote:
>   The task of having computers understanding natural languages at 
> parsing level is basically solved.

*sputter*

There are thousands of people in my field (natural language processing) who
will be very very surprised to hear that.

Have you seen any NLP applications that work and aren't carefully rigged demos?
There aren't many. Here's the state of the art in NLP:

* We can recognize spoken text with accuracy on about 95% of words. The
programs that do this don't use any real linguistics; they work based on the
finite-state model of language, which was discredited 50 years ago, because
it's easy to program and they have bigger problems to deal with first.

* We can translate text, badly, with results that tend to be more amusing than
informative, between languages in the same family. Companies like Xerox write
their manuals in an artificial dialect of English designed to survive machine
translation.

* We can make phone-answering systems to replace touch-tone systems, like
airports use. I maintain that this isn't natural language, though, because
these systems work in the finite world of city names, yes/no responses, and
tracking numbers, and you can't use any grammar in your responses.

* We can convert Japanese kana to kanji and back. This requires a significant
amount of understanding of syntax, and is to me the most impressive practical
achievement of NLP so far.

* We can recognize handwriting, usually.

* We can make clever, overhyped demos that fool the public and get grant money.

Here are the things we can't do outside of demos:
1. We can't translate between vastly different languages (like English and
Korean) with any success at all.
2. We can't parse arbitrary sentences.
3. Even if we have parse trees, we can't turn them into accurate semantic
representations.
4. Even if we have accurate semantic representations, we can't put them together
and hold a natural discourse.

I'm working on a project to use Lojban to get a foot in the door on 2, 3, and
4. The problem is convincing a professor that artificial language processing
will help natural language processing, but it's not because the things I'm
doing in Lojban have already been accomplished with natural languages. It's
because it seems that we won't accomplish these things in natural languages for
another 15 to 50 years, and most people have given up on them.

-- 
Rob Speer