Notes on computational linguistics.pdf - UCLA Department of ...

Stabler - Lx 185/209 2003
The best models of human language processing are based on the programmatic hypothesis that
human language processes are (at least, in large part) computational. That is, the hypothesis is that
understanding or producing a coherent utterance typically involves changes of neural state that
can be regarded as a calculation, as the steps in some kind of derivation.
We could try to understand what is going on by attempting to map out the neural responses to
linguistic stimulation, as has been done for example in the visual system of the frog (Lettvin et al.,
1959, e.g.). Unfortunately, the careful in vitro single cell recording that is required for this kind of
investigation of human neural activity is impractical and unethical (except perhaps in some unusual
cases where surgery is happening anyway, as in the studies of Ojemann?).
Another way to study language use is to consider how human language processing problems could
possibly be solved by any sort of system. Designing and even building computational systems with
properties similar to the human language user not only avoids the ethical issues (the devices we
build appear to be much too simple for any kind of “robot rights” to kick in), but also, it allows
us to begin with systems that are simplified in various respects. That this is an appropriate initial
focus will be seen from the fact that many problems are quite clear and difficult well before we get
to any of the subtle nuances of human language use.
So these lecture notes briefly review some of the basic work on how human language processing
problems could possibly be solved by any sort of system, rather than trying to model in detail the
resources that humans have available for language processing. Roughly, the problems we would
like to understand include these:
perception: given an utterance, compute its meaning(s), in context. This involves recognition of
syntactic properties (subject, verb, object), semantic properties (e.g. entailment relations, in
context), and pragmatic properties (assertion, question,…).
production: given some (perhaps only vaguely) intended syntactic, semantic, and pragmatic properties,
create an utterance that has them.
acquisition: given some experience in a community of language users, compute a representation
of the language that is similar enough to others that perception/production is reliably consistent
across speakers
Note that the main focus of this text is “computational linguistics” in this rather scientific sense,
as opposed to “natural language processing” in the sense of building commercially viable tools
for language analysis or information retrieval, or “corpus linguistics” in the sense of studying the
properties of collections of texts with available tools. Computational linguistics overlaps to some
extent with these other interests, but the goals here are really quite different.
The notes are very significantly changed from earlier versions, and so the contributions of the
class participants were enormously valuable. Thanks especially to Dan Albro, Leston Buell, Heidi
Fleischhacker, Alexander Kaiser, Greg Kobele, Alex MacBride, and Jason Riggle. Ed Keenan provided
many helpful suggestions and inspiration during this work.
No doubt, many typographical errors and infelicities of other sorts remain. I hope to continue
revising and improving these notes, so comments are welcome!
stabler@ucla.edu
3