Diners Club Party. Ambiguity Tests Matter
As we've seen before with things that we do unconsciously, the impression
of simplicity that we get from the effortlessness we experience can be deceptive.
The complexity of what actually goes on often becomes apparent when we attempt
to make these inner workings explicit as precise instructions to a computer.
Take something like: The man in the hat on the hill was happy. Are
we talking about a man who's wearing a hat, being on a hill? Or a man being
in a hat that a hill is wearing? If it just seems . . . "obvious,"
how do you explain why to a computer? Simply telling it that hills don't
wear hats doesn't work, because having dutifully noted that fact it would be
none the wiser when confronted by a man (or hill) wearing a coat; yet surely
we don't operate by having a separate rule to cover every possible eventuality.
We're up against world knowledge again, and that's exactly where computers start
The difficulty isn't so much that so many words have multiple definitions,
as can be seen in any dictionary, and happily take on roles as noun, verb, adjective
or whatever else suits the speaker's fancy. For example, "log" can
mean a piece of tree, a device for measuring a ship's speed, a type of record,
to cut down or harvest trees, to rock to and fro, an abbreviation for logarithm,
or a Hebrew measure of liquids; "turn" in my dictionary has 32 senses
listed. If it were just a question of keeping track of the temporary ambiguities
of such phrases as the Chinese puzzle solver or a talk about books
by Mr. Rider until later information ruled out the incorrect possibilities,
a computer would have no problem coping--although the programming would be a
lot more complex. They perform perfectly well with convoluted but unambiguous
constructions in artificial languages. However, when precision is relaxed to
a degree typical of natural language, the results can be unresolvable by any
rules. Very often, they're equally ambiguous to people:
Her son has grown another foot.
I shot the woman with the Polaroid
Humans, however, zoom in so quickly on the meaning that makes sense in
the given context that they're rarely aware even that any alternative was possible.
This ability to home in on the obvious sometimes blinds us to meanings perfectly
legitimate within the grammar of the language, which the computer, meticulously
applying its rules in the way it was told to, doesn't see anything "obviously"
odd about at all. A classical example is from a parser written at Harvard in
the 1960s, which was presented with the sentence Time flies like an arrow.
Just five words, their one meaning pretty plain, yes? But to the surprise of
the programmers, the computer found five!
Time proceeds speedily, in the way that an arrow proceeds.
Measure the speed of flies in the way you measure the speed of an arrow.
Measure the speed of flies in the way an arrow measures the speed of flies.
Measure the speed of flies that are like an arrow.
Time flies--a kind of fly--are fond of an arrow.
So how do humans jump straight to the intended meanings without bogging
down over the impossibly implausible (interestingly reminiscent of the way our
visual systems don't waste time trying to construct logically consistent but
physically impossible objects)? One possibility is that our brains do in fact
work in something like the way computers do, but unconsciously--i.e. storing
large numbers of possible meanings all awaiting corroborating evidence and delivering
only the final, confirmed one to consciousness. The other is that we make some
kind of snap judgment based on experience, and just go with it until we're either
vindicated or run into a wall. If these strategies have a vaguely familiar ring
about them, they should. They're called breadth-first and depth-first searches.
For individual-word ambiguities, it seems we use something close to breadth-first.
It is an experimentally established fact that immediately after we hear a word,
words closely related to it are identified more quickly than neutral ones. If
sky is flashed on a screen, for example, words like blue and
cloud evoke faster responses than, say, peg. It's as if the
mental dictionary were stored in the form of a network in which related meanings
are clustered together, and activating a particular item leaves that area "primed"
to expect further accesses. When a word has, say, two possible meanings, one
of which makes no sense at all in the context being dealt with, both possibilities
are nevertheless unconsciously recognized, and memory is primed ready to deliver
close associations of both of them. Take the sentence He found several spiders,
roaches, and other bugs in a corner of the room. As would be expected
from the above, a person listening to this retrieves words like "ant"
and "insect" more rapidly immediately after hearing bug.
It also turns out that words like "spy" are primed too, which are
related to "bug" used in the sense of a surveillance device but which
have no connection with the above context. This effect last only for milliseconds
and leaves no awareness that it was ever there, but has been confirmed from
psychological experiments. What this gives us is a word lookup is that is fast
but not very selective, producing a lot of nonsense candidates that are then
At the more complex level of phrases and sentences, however, we don't
compute every possible branch of the pa sing tree. A clue to this is the quite
sensible ambiguities that we simply don't see because we've already rejected
them as irrelevant. Newspaper headlines provide a rich source for collectors:
KILLER SENTENCED TO DIE FOR SECOND TIME
The report gives totals of the number of students broken down by
sex, marital status, and age.
In fact, it sometimes happens that so insistent are we to see meaning
that makes sense, we fail to find the only tree that is consistent
with the grammatical structure.
The man who hunts ducks out on weekends.
Our first tendency is to blink, back up and look again, and then dismiss
it as an error. But in fact it is perfectly grammatical, as:
The man who fishes is available all week. The man who hunts ducks
out at weekends.
What happened the first time was that following a depth-first approach
our parser took duck to be (most plausibly) the object of hunts
and plunged on with that assumption, only to come to a perplexed halt later
when the pieces wouldn't fit. Another example:
Glass windows are made of has blemishes.
makes perfect sense when seen in:
The glass that crystal goblets are made of is perfect, but glass
windows are made of has blemishes.
What we see and hear is determined by our expectations as shaped by a
lifetime of living in the world and getting to know how the parts of it typically
relate to each other. One wonders what kind of an edifice the Blocks World stacking
program would attempt to build, given the statements:
The ball is on the table.
John is on the ball.
The boss is on John's back.
The detective is on the boss's tail.
The commissioner is on the detective's side.
All in all, comprehending language comes through as a daunting prospect.
As with anything else, you have to start somewhere.