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

Stabler - Lx 185/209 2003
11 Semantics, discourse, inference
A logic has three components: a language, a semantics, and an inference relation. As discussed in §1, a
computational device may be able to recognize a language and compute the inferences, but it does not even
make sense to say that it would compute the semantics. The semantics relates expressions to things in the
world, and those things are only relevant to a computation to the extent that they are represented. For example,
when the bank computes the balance in your account, the actual dollars do not matter to the computation; all
that matters is the representations that are in the bank’s computer. The interpretation function that maps the
numbers to your dollars is not computed. So typically when “semantics” is discussed in models of language
processing, what is really discussed is the computation of representations for reasoning. The semantics is
relevant when we are thinking about what the reasoning is about, and more fundamentally, when we are
deciding whether the state changes in a machine should be regarded as reasoning at all.
Standard logics are designed to have no structural ambiguity, but as we have seen, human language allows
extensive ambiguity. (In fact, S6.6.3 shows that the number of different derivations cannot be bounded by any
polynomial function of the number of morphemes in the input.) The different derivations often correspond
to different semantic values, and so linguists have adopted the strategy of interpreting the derivations (or
sometimes, the derived structures). But it is not the interpretation that matters in the computational model;
rather it is the syntactic analysis itself that matters.
With this model of human language use, if we call the representation of percieved sounds PF (for ‘phonetic’
or ‘phonological form’) and the representation of a completed syntactic LF (for ‘logical form’), the basic picture
of the task of the grammar is to define the LF-PF relation. The simplest idea, and the hypothesis adopted here,
is that LF simply is the syntactic analysis. We find closely related views in passages like these:
PF and LF constitute the ‘interface’ between language and other cognitive systems, yielding direct
representations of sound, on the one hand, and meaning on the other as language and other systems
interact, including perceptual and production systems, conceptual and pragmatic systems.
(Chomsky, 1986, p68)
The output of the sentence comprehension system…provides a domain for such further transformations
as logical and inductive inferences, comparison with information in memory, comparison with
information available from other perceptual channels, etc...[These] extra-linguistic transformations
are defined directly over the grammatical form of the sentence, roughly, over its syntactic structural
description (which, of course, includes a specification of its lexical items). (Fodor et al., 1980)
…the picture of meaning to be developed here is inspired by Wittgenstein’s idea that the meaning of
a word is constituted from its use – from the regularities governing our deployment of the sentences
in which it appears…understanding a sentence consists, by definition, in nothing over and above
understanding its constituents and appreciating how they are combined with one another. Thus the
meaning of the sentence does not have to be worked out on the basis of what is known about how
it is constructed; for that knowledge by itself constitutes the sentence’s meaning. If this is so, then
compositionality is a trivial consequence of what we mean by “understanding” in connection with
complex sentences. (Horwich, 1998, pp3,9)
In these passages, the idea is that reasoning is defined “directly” over the syntactic analyses of the perceived
language. Understanding an expression is nothing more than having the ability to obtain a structural analysis
over basic elements whose meanings are understood.
It might seem that this makes the account of LF very simple. After all, we already have our syntactic analyses.
For example, the grammar from the previous chapter, gh5.pl provides an analysis of Titus be -s human:
231