¿ìè¶ÌÊÓƵ

Take the sentences: ‘Time flies like an arrow.’ and ‘Fruit flies like
a banana.’ After a bit of thought, it is obvious that the first sentence
can be understood only if it is recognised as a simile, while the second
is a statement of fact. Anyone with common sense can parse these sentences
correctly.

For the past thirty years, however, bizarre sentences of this type have
proved a major obstacle to designing a computer that can understand everyday
language. The problem is that computers have no common sense, or, at least,
that was the case. A research team working in Austin, Texas, believes it
has now found the key to providing computers with common sense.

Their optimism is fuelled by the results from the Large Common Sense
Knowledge Base project (Cyc), which is nearing completion after ten years.
Cyc started life as part of a project to create a computerised encyclopedia
– dreamt up by one of computing’s legendary figures, Alan Kay (‘The man
who made computers personal’, ¿ìè¶ÌÊÓƵ, 19 June 1993). While he was
working at Atari’s research centre, Kay asked Douglas Lenat, consulting
professor of computer science at nearby Stanford University, to add something
original to this project. Lenat suggested ‘automating the white space’.
This was his shorthand for computerising the knowledge that makes everyday
life possible – knowing that someone cannot be in two places at once, for
example, or that an object will fall if it is not supported.

Do vampires exist?

But Atari hit financial difficulties in the early 1980s, Kay left,
and Lenat was persuaded to relocate his idea. Lenat moved to the newly established
Microelectronics and Computer Technology Corporation (MCC), a cooperative
research and development organisation in search of a prestige project.
The MCC is based in Austin and its shareholders and associates include Apple
Computer, American Express, Microsoft and the US Department of Defense.

The MCC researchers planned Cyc as a decade-long project because they
knew that finding a way to represent, organise and access a huge volume
of assorted and often inexact knowledge would not be easy. From the start
Cyc was planned to be several orders of magnitude larger than existing expert
systems, which tend to break down when they hold more than 5000 to 10 000
rules pertaining to a particular area of expertise. As Lenat recognised,
there would be no short cut to making a machine ‘intelligent’, no equations
of artificial thought waiting to be discovered. And since Cyc researchers
would have to accumulate knowledge over a long period they needed a robust
structure that could adapt to changes.

They decided to represent knowledge in ‘frames’, an approach developed
by Marvin Minsky, one of the pioneers of artifical intelligence. Frames
are mini-databases which collect related pieces of information together
in a structured way. The top level of a frame has fixed information which
is always true while the lower levels, or ‘slots’, hold details of specific
examples of the fixed information. Returning to our sentences, the frame
‘flies (noun)’, for example, could contain the fixed knowledge that flies
are insects and that they have wings, while a slot that holds information
on type could be filled with ‘fruit’, and the slot ‘likes-to-eat’ could
be filled with ‘banana’.

But frames soon proved too limited to express the taken-for-granted
knowledge Cyc was attempting to gather – such as what a ‘person’ is, or
what ‘today’ means. So Lenat worked on alternatives with R. V. Guha, a
postgraduate student of mechanical engineering who joined Cyc in 1987, and
eventually became co-director of the project. They combined frames with
a form of logic called ‘first-order predicate calculus’.

Predicates, when used in connection with logic, are statements about
individuals or things, and are true or false when applied to particular
individuals or things. For example, the predicate ‘is-a-flying-insect’ is
true when applied to fruit flies but false when applied to time. Predicate
calculus uses variables and quantifiers, such as ‘for all’ and ‘there exists’,
which allow general statements to be made about classes of objects. In the
case of fruit flies, for example, one of the statements in the database
may be that ‘for all fruit flies there exists a part of anatomy that is
used for flying’.

By extending first-order predicate calculus and combining it with frames
and other techniques, Lenat and Guha created a language capable of representing
the everyday world in a way that computers can make sense of. Called CycL,
the language enables common-sense knowledge to be gathered as a collection
of assertions that are true by default rather than by definition. The assertions
are usually true – for example, instead of holding that ‘all birds fly’,
CycL holds that birds are creatures which ‘usually fly’.

In CycL reasoning is done by argument rather than strict logic. If Cyc
is given a sentence, it will try to find arguments for and against the statements
in that sentence, and try to resolve the arguments using such criteria as
whether they are based on causal relationships, or whether there is evidence
for the argument. Logic-based systems, in contrast, would attempt to apply
absolute values of truth and falsehood.

To solve the problems of searching through a vast database and then
reasoning quickly, Cyc is organised a little like a library where the knowledge
exists in the books and the databases, but it may be faster to ask a librarian
for some kinds of information. In Cyc, the ‘epistemological level’ corresponds
to the books on the shelves; knowledge is represented as assertions and
structured in the simplest way possible. Most of this knowledge at the epistemological
level exists again at the heuristic level where related knowledge is grouped
together and a range of algorithms enables Cyc to quickly make inferences
about information. The heuristic level is equivalent to the librarian who
is familiar with the organisation of the books and has a set of rules of
thumb that he or she uses to find knowledge quickly. An interface called
Tell-Ask sits between the two and automatically translates sentences.

The use of a concept known as ‘microtheories’ helps to keep the mass
of knowledge in Cyc in some form of order. CycL treats a specific set of
assertions as a theory, so the sentence: ‘The personnel department of a
large company organises the hiring of staff’, could be considered part
of a microtheory on large companies. Lenat says the major advantage of microtheories
is that they allow for local consistency without requiring global consistency.
Assertions within a microtheory must be consistent, but they need not be
consistent with those of another microtheory. This enables Cyc to cope
with a dialogue such as: ‘Who was Dracula?’ ‘A vampire.’ ‘Do vampires exist?’
‘No.’ Dracula and vampires can exist in a microtheory concerning Bram Stoker,
but not in a modern scientific microtheory. Assertions can, however, be
imported from one microtheory to another. So, for example, if Dracula dropped
a wine glass it would be expected to shatter.

Tutors for computers

This method of handling consistency is what has enabled Cyc to grow
way beyond the boundaries of any previous expert system. Cyc now has the
equivalent of around two million rules. Almost all this knowledge is expressed
in terms of assertions which are true only by default, rather than by definition.
It includes thousands of microtheories, and a variety of models of such
abstractions as time and space, each under some circumstances but not others.
So, for example, Cyc knows that if a piece of wood is chopped up, each piece
is still a ‘piece of wood’. It also recognises that if a table is chopped
up each piece is not still a table.

Lenat says: ‘We’ve got adequate ways of handling time, space, causality,
beliefs, substance, human emotions, capabilities and so on – where adequate
means a set of partial solutions which cover the common cases. ‘We don’t
seek completeness in Cyc. That would be as hopeless, absurd and unnecessary
as if any human being sought for their own mental processing to be complete.’

Teams of ‘knowledge enterers’ in Cyc’s offices, in academic institutions
and in MCC’s industrial partners feed the database with information. But
this is still not the kind of information you would find in an encyclopedia,
dictionary or almanac, says Lenat. Cyc does not know that London is the
capital of Britain or that John Major is Prime Minister, for example. But
it does know what a capital city is and what it means to be a prime minister.
Populating the database with facts will happen at a later stage.

The end of the project will have come when Cyc begins to read and learn
for itself. Cyc is already moving away from having all its knowledge entered
manually by humans to ‘scanning’ and absorbing on line texts, extracting
new knowledge to incorporate into its own database. This process began about
a year ago and now Cyc gets almost half its information this way. The Cyc
researchers are becoming less like brain surgeons and more like tutors,
says Lenat. They need only explain sentences such as the one about time
and fruit flies when asked. Sometimes they ask Cyc to paraphrase knowledge
to ensure that the database has really ‘understood’ the data.

Searching for surfboards

The next stage is to add Cyc to applications that people want to use.
Besides the obvious applications in natural language interfaces that enable
computer users to use everyday written language to communicate with their
machines and in speech recognition, Lenat suggests a plethora of ways to
use Cyc. These range from ‘smart’ spreadsheets, databases and image retrieval
systems to automatic brokering of share dealing. In a smart spreadsheet
or database, for example, Cyc would perform what Lenat calls ‘sanity checking’
– checking the rows and columns or fields for data that violate common sense.
This could include data where an employee appears to earn far too much,
or far too little, or where a person had listed themselves as their contact
in an emergency. In a smart image retrieval system for a database of many
thousands of photographs, Cyc could help match a query such as: ‘Show me
some people in danger of getting skin cancer’ to a caption of an image such
as: ‘Boys with surfboards’.

The success or failure of Cyc will only be decided by how well it does
in applications. Although MCC’s shareholders have yet to produce a Cyc-based
application, the amount of time and money they have spent on the project
suggests they have confidence in Cyc’s commercial potential.

Cyc, however, does have its critics. Hubert Dreyfus, professor of philosophy
at the University of California, Berkeley, and author of the provocative
critique of machine intelligence research, What Computers Can’t Do, has
most recently focused his attention on Cyc. Dreyfus describes Cyc as the
last defence of the AI dream of producing broad, flexible human intelligence.
Dreyfus believes Cyc is likely to fail for the same reason previous AI projects
failed. It is based, says Dreyfus, on the flawed belief that the mind is
a symbol manipulator and that a machine can be endowed with intelligence
simply by enabling it to manipulate symbols effectively.

The arguments Dreyfus uses have been called ‘situatedness’ – intelligence
has to be in a physical body, which, in turn, is situated in a particular
culture in a particular time and place. Humans bring their situated experience
to bear on understanding language and reasoning about the world. For example,
we know what the pronoun ‘it’ refers to in the sentence ‘Mary saw a dog
in the window and wanted it’ – not because we consult a database of common
facts but because we understand the emotions Mary feels when she sees the
dog, says Dreyfus. Similarly, we would know that ‘it’ referred to the window
if the sentence ended ‘. . . and she pressed her nose up against it’.

Lenat rejects this argument because to him it is almost indistinguishable
from the much quoted religious view that intelligence requires a soul. But
he acknowledges that a true machine intelligence of the future will be
a somewhat hybrid system which includes among other things a common-sense
knowledge base, programs that rely on statistical analysis and situated
elements. ‘If they (the situated elements) add power, why exclude them?
We’re after power, not the vindication of one ideology or another,’ says
Lenat. But even critics like Dreyfus admit that Lenat has had the courage
to face one of the hardest and most demanding challenges of AI.

Meanwhile, Marvin Minsky said recently that every country should have
a Cyc-type project and that one would be started at the AI laboratories
at the Massachusetts Institute of Technology where he is director. This
sparks an angry reaction from Lenat. Minsky, like all AI pioneers since
the mid-1950s, has known that a common-sense knowledge base was necessary
if machine intelligence were ever to be viable. They all could have done
it, but ducked it, says Lenat.

There is in fact no direct competition to Cyc. ‘If I felt that any other
group was doing anything like Cyc, or would do in the future, I would love
to hand the reins over to them,’ says Lenat. But he recognises that in another
sense Minsky has a point. ‘Common sense is slightly different in different
cultures, in different age groups, in different political ideologies,’
says Lenat. Local additions to Cyc will be required which build in local
beliefs and facts about people, places and activities – the sort of assumed
knowledge you need to complete a local crossword, says Lenat.

Ultimately, Lenat sees Cyc-type databases as a fundamental layer of
software which will underpin all intelligent computing in the future. ‘Today
you wouldn’t dream of buying a computer that doesn’t have an operating system,’
says Lenat. ‘In 20 years you might not dream of buying a computer that does
not have common sense underlying every application.’

Clive Davidson is a freelance journalist.