Minds and Computers : An Introduction to the Philosophy of Artificial Intelligence

particular language. The initial sound in the word ‘pat’ and the initial
sound in the word ‘bat’ are two phonemes which di
ffer only in terms
of voicing. The phoneme /p/ is not voiced but /b/ is. This means that
when you produce the phoneme /b/ your vocal chords vibrate but
when you produce /p/ air merely passes noiselessly through. Try
putting your fingers on your voice box as you pronounce ‘pat’ and
‘bat’ and you should be able to tell the di
fference in voicing. We can
tell from the fact that ‘pat’ and ‘bat’ mean di
fferent things that voicing
is a distinctive contrast in English and that /p/ and /b/ are distinct
phonemes.
This is a very rough and ready account of phonemes but for present
purposes we merely need to appreciate that one of the mediating tasks
in interpreting a spoken utterance involves determining, from the
phonetic properties of the utterance, a representation of the distinct-
ively contrasting sound units which are meaningful in the language
uttered.
Another mediating task involves parsing the phonemic representa-
tion into a syntactic structure. In other words, we need to determine
how the string of meaningful speech sounds breaks up into words,
phrases, clauses and sentences. Finally, we need to work out what
these words, phrases and clauses mean.
So in all, there are three distinct levels of representation implicated
in understanding an utterance – phonemic representations, syntactic
representations and semantic representations.
Producing an utterance involves these same three representational
transformations. We intend a particular meaning, determine the syn-
tactic structure which encodes that meaning, convert the syntactic rep-
resentation to a phonemic representation and finally produce phonetic
output for each phoneme. This production of phonetic output is gov-
erned by the phonemic environment in which a phoneme is situated –
a feature of language production we will examine in Chapter 16.
Although the sequential manner in which I’ve enumerated these
transformational stages might lead you to think that each stage of
processing occurs separately and in sequence, there is good evidence
to suppose that all three occur in concert.
146
  

---

For instance, it seems that the determination of syntactic structure
is influenced by expectations governed by the semantic representation
we are constructing as we syntactically parse an utterance. To make
this clear, consider the following three sentences.
[1]
The horse raced past the barn.
[2]
The horse raced past the barn yesterday.
[3]
The horse raced past the barn fell.
Sentence [1] fits the prototypical syntactic pattern of an English sen-
tence – it is composed of a noun phrase followed immediately by a
verb phrase which consists of a transitive verb and its complement.
Sentence [2] also fits this pattern – it merely adds an adjunct after the
verb phrase.
Sentence [3], on the other hand, presents initially as unusual and
potentially ungrammatical. We hear a noun phrase ‘the horse’ fol-
lowed immediately by what we take to be the main verb of the sen-
tence – ‘raced’ – and expect the remainder of the sentence (or at least
the clause) to consist of the complement of the verb (past the barn /
in the three o’clock at Flemington) and possibly one or more adjuncts
(rather quickly / last Tuesday).
Consequently, when ‘past the barn’ is followed by the verb ‘fell’ we
are taken aback as we were expecting maybe an adjunct but not
another verb. We had already determined a semantic representation
for the sentence up until that point but the presence of the final verb
shows us that either we were wrong about the meaning (in thinking it
was the meaning of [1]) or that the sentence is ungrammatical.
In other words, the meaning we construct as we interpret the sen-
tence – in combination with what we implicitly know about the syntax
of typical English sentences – conditions our expectations of syntac-
tic structure. We assume the most likely syntactic structure and con-
struct our semantic representation accordingly, which then influences
our expectations of the remaining syntax.
It can, in fact, be quite di
fficult to get past the seeming ungram-
maticality of [3] until we realise that the sentence doesn’t have the
typical syntactic structure we initially assumed. The verb ‘fell’ is the
main verb of [3] and ‘raced past the barn’ merely indicates which horse
it was that fell – the one which was raced past the barn. Note that we
can substitute ‘painted purple stripes’ for ‘raced past the barn’ to
achieve a similar e
ffect, although the effect will be weaker as we are
less likely to assume the horse to be the painting agent than the racing
agent. If, however, we substitute ‘running quickly’ or ‘belonging to
Anne’ for ‘raced past the barn’, the sentence becomes unambiguous.
  
147

---

Sentences with ambiguous syntax provide us with evidence that the
construction of semantic representations and syntactic representa-
tions occur in concert and a
ffect each other. Homophonic phrases
provide further evidence that parsing phonemic representations into
syntactic elements is influenced by both syntactic and semantic expec-
tations.
Homophones are sequences of phonemes that have more than one
semantic interpretation, such as ‘pair’ and ‘pear’. A special case of
homophony is homonymy. Homonyms are homophones which also
have the same orthographic representation – i.e. they’re written the
same way, such as the verb ‘bank’ and the noun ‘bank’. There are very
many homonyms in English as many words can be interpreted as nouns
or as verbs – e.g. paint, spring, void, power, urge, whisper, sleep, etc.
Homophonic phrases are sequences of phonemes which can be
parsed as more than one distinct syntactic structure. An example
homophonic phrase is the sequence of phonemes which can be inter-
preted as ‘way up high’ or ‘weigh a pie’.
We can interpret a line sung by Judy Garland in The Wizard Of Oz
as either of the following two sentences:
[4]
Somewhere, over the rainbow, way up high.
[5]
Somewhere, over the rainbow, weigh a pie.
Both of these are perfectly grammatical syntactic constructions – [4]
is a description and [5] is an imperative. Consequently, syntactic con-
siderations alone are not su
fficient, in this case, to disambiguate
between the two possible interpretations.
Our semantic interpretation, however, inclines us strongly towards
[4]. The semantic interpretation of ‘Somewhere, over the rainbow’
makes it a much more likely hypothesis that Judy Garland goes on to
further describe this place and ipso facto a much less likely hypothesis
that she is issuing a pie weighing imperative.
In cases such as the following, however, it is not clear if it is syn-
tactic or semantic considerations driving the disambiguation of the
homophonic phrase:
[6]
I went the baker’s to weigh a pie
[7]
* I went to the baker’s to way up high
Sentence [7] is ungrammatical (as the asterisk indicates) so, in this
instance, syntax alone is su
fficient to disambiguate the homophone;
however, it is di
fficult to separate the syntactic consideration from the
influence of the semantic association. Fortunately, this is not some-
thing we have to rule on here.
148
  

---

What I want to concentrate on in the remainder of this chapter is the
fact that we readily and automatically recognise [7] as an ungrammat-
ical sentence. I want to examine how we might employ symbol systems
methods to rule on the grammaticality of strings of written language.
14.2 GENERATIVE GRAMMAR
Noam Chomsky revolutionised the discipline of linguistics in the
1950s by taking a new approach to the study of grammar.
Grammar, before Chomsky, involved little more than taxonomis-
ing parts of speech and enumerating prescriptive principles that stu-
dents of grammar should abide by. For instance, students of the
prescriptive grammarians were told to never split an infinitive.
Chomsky, in contrast, took the grammar of a language to be the
mechanism by which all and only the grammatical strings of the lan-
guage can be generated. Chomsky argued that our internalised know-
ledge of the systematic generative grammar of our language accounts
for the infinite productivity of language.
Language is infinitely productive in that we are able to produce,
and rule on the grammaticality of, an infinite number of sentences,
despite only ever having been exposed to a finite number. As a native
speaker, you can immediately tell that the sentence ‘Michelle’s grand-
mother sells drugs to bikers in Belarus’ is a grammatical sentence of
English, even though it is unlikely that you’ve ever encountered that