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

Exercise 13.6
Generate a bottom-up search in our kinship system for the
states:
(a) mother (m);
(b) grandfather (j).
Expert systems can be usefully deployed in weak artificial intelligence
projects. One such well known system is MYCIN which was developed
in the 1970s at Stanford. MYCIN can make probabilistic diagnoses of
pathologies and recommend medication based on the results of blood
tests. Its resident information encodes heuristic procedures followed
by medical doctors in making rough and ready diagnoses in the
absence of developed cultures. Its rule application procedure appeals
to probabilistic reasoning principles so it is able to suggest a number
of possible diagnoses with a certainty factor attached to each.
More interestingly for our purposes, there are some who hold
out hope that expert systems alone can give rise to strong artificial
 
143
Figure 13.1
Bottom-up search.
father ( j)
parent ( j), male ( j) 
parent_of ( j, j), male ( j)
parent_of ( j, m), male ( j)
parent_of ( j, h), male ( j)
male ( j)

---

intelligence. The Cyc project, founded by Douglas Lenat and devel-
oped under the auspices of the private research institution Cycorp,
aims at precisely this. The conviction held by Cyc researchers is that
if they can encode – in the resident information of the system – all (or
much of) the information that you and I take to be common know-
ledge, and develop a su
fficiently sophisticated inference engine for
making deductions, they will thereby develop a strong artificial intel-
ligence artefact.
Whether or not expert systems methods are su
fficient for artificial
intelligence remains an open empirical question. There are reasons,
however, to believe that there may be problems with this approach.
Some of these we will consider in Chapter 15 and some in Chapter 19.
In defence of the Cyc project, however, it appears – from what I can
gather of their operations – that they are responsive to at least some
of these concerns and are augmenting the traditional expert system
model accordingly.
Perhaps the most significant and di
fficult computational problem
which needs to be solved on the way to true artificial intelligence is the
problem of interpreting and producing natural language. It is to this
issue that we turn our attention in the next chapter.
144
  

---

C H A P T E R 1 4
MACHINES AND LANGUAGE
Devising computational procedures for handling natural language is
arguably the most significant problem facing artificial intelligence
researchers. In this chapter we’re going to begin by considering the
various computational problems which need to be solved in order to
facilitate this.
As is the case with machine reasoning, a comprehensive survey of
computational methods for facilitating linguistic interpretation and
production would require dedicated volumes. We’re going to concen-
trate here on just one of these methods, in service of one of the functions
constitutive of linguistic competence – determining grammaticality.
We will lend further consideration to the procedures governing lin-
guistic activity in Chapter 16 and will return to examine comput-
ational methods for implementing further functions subserving the
linguistic facility in Chapter 19.
14.1 INTERPRETING LANGUAGE
Let’s reflect on the various procedures involved in the comprehension
of a spoken utterance.
Spoken language is generally delivered in a continuous phonetic
stream which does not readily reveal its linguistic properties. While it
might seem that individual words are easily discernible from phonetic
properties of an utterance alone, this is generally not the case unless
one is speaking – very – slowly – and – carefully. This is clear if you
look at a visualisation of a waveform of a recorded utterance.
Furthermore, the absolute phonetic properties of an utterance – such
as pitch and volume – will di
ffer significantly from speaker to speaker.
Interpreting a spoken utterance is a non-trivial function – in fact it is
quite a complex procedure.
There are several tasks that mediate the interpretation of audi-
tory input as a sentence of natural language. One of these tasks
145

---

involves converting the phonetic input to a phonemic representa-
tion.
Phonemes are the atomic meaningful speech sounds of which a
spoken language is constituted. We will learn a lot more about
phonemes in Chapter 16 but for now a rough description will serve.
Phonemes are idealisations to which actual speech sounds –
phones – approximate and which represent distinctive contrasts in a