Launched on 18 May, the long-term goal ofWolfram Alpha- a natural language-based search tool
which claims to offer an alternative approach to Google - "is to
make all systematic knowledge immediately computable and accessible
to everyone".
Its creators say, "When computers were young, people assumed
that they'd be able to ask a computer any factual question and have
it compute the answer." I'm sure they would have like to have added
"and I'm happy to say that"
If Wolfram Alpha's goal is to deliver such a system, it would
seem that it's the tool many of us have been dreaming of.
In my
previous article I tried - as have many - to say something
about how the system appears to function, and in this - the round
up - I'll take that a little further, and tell you why I'm a little
disappointed at the moment - yet hopeful too.
Last time I suggested that Wolfram Alpha works by a) feeding
your query through a linguistic-parser, b) taking that output, and
applying further rules and manipulation, and, c) assessing deep-web
data to provide - hopefully - an answer.
One of the hardest things about Wolfram Alpha is using it. That
is to say, getting your query "correct"; or at least in a form that
Wolfram Alpha can understand and use.
Understand and use
To "understand" means that the linguistic-parser (all the parts:
syntactic, semantic, etc) has made some initial "sense" of your
words; whereas "use" implies that this resultant sense is able to
be further understood and manipulated in order to be mapped to
Wolfram Alpha's deep-web data sources.
Important point: to get a successful result, you must realise
that although Wolfram Alpha might like your query - in other words,
all but one of the criteria above were satisfied - it might still
be short on relevant data. However, that just requires adding more
structured data sources. Worth bearing in mind though.
Getting a successful result also means you have not
misunderstood Wolfram Alpha's function altogether Wolfram Alpha is
not Google.
Wolfram Alpha has some useful (yet minimal) guidance here which
highlights the differences:
If you enter the first-parts of these examples into Wolfram
Alpha, you will mostly get "proper answers", whereas you won't with
the second bits. For instance, "highest mountain" doesn't produce
any output. With Google, either part results in hits, but that's
what you'd expect. However, don't expect to get 'proper answers' -
in fact, be surprised if you do.
So one hurdle in using Wolfram Alpha effectively is to get to
grips with how to create, and ask a good question.
I can think of one easy way to improve learning this skill. Give
users the chance to rate their Wolfram Alpha answers: "Yes, that's
exactly what I wanted"; "WTF!"; "Oh no - not that 'Wolfram Alpha
isn't sure what to do with your input' message again!" Or perhaps
just a simple 1 to 10 score.
Then, as a learning aid, Wolfram Alpha could display lists of
previous queries that scored highly - what better way to learn than
to see what has previously worked.
Anyway, let's have a look at an actual attempt to use Wolfram
Alpha; one where Wolfram Alpha produces data that is at least
semi-relevant.
The problem I wanted answering was to do with tossing a coin
multiple times - yet with a twist in the tail.
So to start with I entered "coin toss". Wolfram Alpha came back
with some basic probability stuff on tossing a coin - like the
probability of seeing 12 heads and eight tails given 20 tosses.
However, the question I wanted answering was quite a bit more
complex and concerns tossing a coin and observing a sequence in the
results. Here's the question:
Q. Given a series of consecutive coin tosses, is it more/less
likely that I would see the sequence HTT before I saw the sequence
HTH - or is it the same?
Just to cut to the chase here; the answer is that the average
number of tosses needed to see HTH is 10, whereas it's eight for
HTT (if you don't believe me, try it. Or just e-mail me for an
explanation).
Ok, so on entering that straight into Wolfram Alpha I got what
you might expect:
"Wolfram Alpha isn't sure what to do with your input."
And that's pretty reasonable, isn't it? For one thing, I
actually re-typed that question sentence about 10 times before I
was pretty sure that I had made the problem pretty clear, and all
linguists know that I could re-word that question an infinite
number of times, yet still have it make just about as much (or
better) sense. So, the "WolframAlpha isn't sure what to do with
your input" message might be a linguistic failure to
"understand".
However, maybe it did well with that - after all, I gave it a
fair few: "coins", "tosses", "likely", "sequence", "T", "H" - and
Wolfram Alpha does know some stuff about tossing coins.
Try as I might I couldn't get any further on this problem (so
you are going to have to try it for yourselves), and this might be
simply because Wolfram Alpha does not have enough relevant data to
draw upon yet. Alternatively, as I hope I have made clear already,
it may not be able to get enough clues about what I'm asking from
its linguistic analysis of inputs, or it is unable to take adequate
clues, and then manipulate them in such a way as to make querying
its data sources work in any way viable. Who knows?
Perhaps another improvement would be to have Wolfram Alpha tell
you more about its "I'm not sure" message? What exactly was the
problem?
In the end I found myself muttering, "Come on, it's only a
particular case of conditional probability," and so I found myself
entering "conditional probability".
It seems as though I'll have to wait, and, in my gut, I think
I'll probably have to wait a fairly long time before I can enter
something along the lines of my question and see something
pertinent come back.
OK, so it was a tough ask. I'll also admit that I don't profess
to know all the ins and outs of how to interact with Wolfram Alpha.
However, if Wolfram Alpha is hoping to be a natural language
computational knowledge engine, which I assume is the ultimate
goal, it has quite some way to go - that is if you don't want to
know how high Everest is in terms of Golden Gate bridges!
Peet Morrisstudied software engineering,
computational linguistics and statistics at the University of
Oxford. He is currently a researcher in the Department of
Experimental Psychology, and a college lecturer in statistics at St
Hilda's College.