Below is the text Barney sent us of his talk, which he said he was going to post in his blog, but I couldn't find it online. (no editing has been done)
Bringing NLP
to Market: 30 years in the making
By Ron Kaplan and Barney Pell
For Danny Bobrow’s Festshriff, March
26, 2008
Barney gave the talk on behalf of
himself and Ron Kaplan, who was in Turkey at the time of this event.
Ron’s story:
I first met Danny almost 40 years
ago, in 1969, when I was a first year graduate student and he was heading up
the AI Center at BBN. I was hired to develop an English grammar within
the Augmented Transition Network system that Bill Woods had created. And
it was my understanding of the benefits and disadvantages of that grammar and
framework that eventually led to many of the concepts and insights that
underlie today's practical NLP technologies (hierarchical attribute value
structures, Lexical Functional Grammar...). So Danny was there at the
beginning.
In fact, he was there before the
beginning. Most people in AI are aware of the Student program that he
developed for his thesis, one of the milestones in the early history of the
field. But what has been mostly forgotten is a seminal paper from the
late 60's that he wrote with Bruce Fraser. It was a very brief paper, but it
outlined the basic ideas of network grammars that evolved into Wood's ATN which
evolved into LFG which evolved into...(Powerset)
This early example illustrates one
of Danny's wonderful qualities, his ability to cut to the core of complex
problems and to find the essence of a solution. I've seen this many times
over the years, how he is able to abstract from the details of a situation and figure
out what really matters and what doesn't, what two problems have in common and
how they differ, which directions are promising and which will dead end.
This has always impressed me, but I have to say that it has also annoyed
me at times. I can remember several occasions when I was struggling to overcome
a really hard obstacle. I would have a whiteboard full of diagrams and
formulas, things crossed out and overwritten, marked up in different
colors--basically, totally confused. Danny would come by, pop his head
in, take a quick look, and say something like: "Oh yes, the answer is
5". And he would be exactly right: the answer was 5.
Damn! The annoying thing was that he could come up with the solution
without experiencing the pain of the problem itself--and I needed someone to
share the pain just as much as I needed the answer.
Barney: I got to see this
first-hand, and it’s funny how many people make the same observation. I have also had some people, including Ron,
make a similar observation about me. So
if I may be annoying at times, at least I have in Danny a good role model that
shows this is ok… at least if you’re right.
Ron: Danny left BBN for Parc in the
early 70's (and in the early days of Parc), and he set about creating a
Language Understanding group. I was still a graduate student (finishing
my thesis "any day now"), and I was lucky that he invited me to
interview and to become a member of the group (I arrived about 2 years after I
accepted the job, still about to finish my thesis "any day now"--he
pushed (he's very good at that) and I did finally finish!). He also attracted
other key people to work in and with the group--Martin Kay, Terry Winograd, and
later Don Norman, Richard Fikes, Ira Goldstein, and others.
Those were great years. We
were excited about integrating existing technologies and inventing new ones to
build a real language-understanding system--we were smart, well-funded, great
computing infrastructure, how hard could it be? We pulled everything
together in what was (and what still may be) a state-of-the-art dialog system,
called GUS (Genial Understander System). If you wanted to take a trip
from Palo Alto to San Diego, GUS was the reservation agent you wanted to talk
to--best fares and most convenient schedule on that route!
Barney: It is interesting to note
Ron’s statement that a system built 30 years ago was, and still may be, a state
of the art NLP system. For how many
fields is that even possible? I worked
at SRI in NLP in the 80’s and had the fortune to be part of another state of
the art NLP system at its time, and which might still be a state of the art NLP
system today. I saw the potential of NLP
to change the user experience, but I also saw that it was going to take a long
time before it was really ready. I decided to go work on other things for a
while (say, 20 years), and come back when hopefully the technology could be
ready for broad deployment. Somewhat
surprisingly, that’s pretty much what happened.
Ron: That was our first attempt at
integration. What we realized is that there was more work to be done on
the individual components, so we decided to back away from integration until we
got better individual pieces. Martin and I worked on syntax--and LFG and
the concepts of unification grammar came out of that--while Danny and Terry
focussed on knowledge representation and processing, producing one of the
earliest knowledge-representation systems. We decided we would make another run
at a full-fledged system as soon as we locked down the pieces.
And that's what we did. The
only thing, as everyone knows, is that it took a little bit longer to get the
pieces done than we had thought. And there were shifts and changes: I
forked off a separate group, Parc's Natural Language Theory and Technology
area, to work on the language side, Danny continued on the knowledge side with
Mark Stefik and others. And oh, we all worked on building the Lisp system
and launching Xerox' AI lisp machine business.
We made a lot of progress.
Every now and then, over the years, Danny and I would get together for a
brief status check--are the pieces done, are we ready to integrate? The
answers were always No and No, but Soon and Soon. This went on for 25
years or so--and some in management might have heard the answers as Never and
Never.
Barney: Most people in the field
gave up during this period. During the nineties, there was an AI winter, and
really a Semantic winter, during which most people just stopped working on
semantics. Instead, statistical
approaches emerged and a generation of people entered the field who did not
know or care about deeper knowledge or linguistics. Many people now think these approaches were
tried and failed. But the trying was still a process and since the project
hadn’t been finished, the conclusions of its failure mid-way through were
premature. This room contains perhaps 1/3 of all the people who stuck around
for the long haul. Luckily the people here did not give up.
Ron: But we were actually advancing,
and in fact, 7 or 8 years ago, the answer changed. We were ready--we
agreed that finally the integration experiment was worth running again, there
was a good chance that we could create something useful, and not simply learn
that more work was needed.
And that's what we did. We
defined a common project for the knowledge and language groups (extracting
knowledge from textual documents, with an eye towards detecting entailment and
contradiction). It was challenging and fun, required a lot of strategic
thinking (Danny is great at that), required picking and organizing the right
kinds of expertise (Danny was always great at that too).
And how did (at least this chapter)
end? Powerset.
Barney: I knew Danny from his days
as an advisor at Nasa, where he helped us think about human interaction with
intelligent systems. I ran into him
again at a time when I was developing plans for a natural language search
engine, at the AAAI conference in July, 2005. Danny told me the story
about 35 years of concluding PARC NLP wasn’t yet ready, and then finally
concluding it was ready. You almost never hear about researchers working on
something for such a long period of time at all. Much less, having the humility
and perspective to say that this is going to be really hard and take a long
time, and despite progress it continues to require more fundamental work. It then seems completely unheard of for such
researchers to finally come back at say it now is ready. Knowing that this story comes from the PARC
team, given the pedigree and history in NLP, I followed up immediately. Danny connected me with Ron, and for better or
for worse (remains to be seen, in short order), the combined linguistic and
semantic technologies seemed just right for the search application.
Ron: The story is still unfolding,
and Danny is still in the thick of it. 40 years ago, or even 10 years
ago, or 5, it wasn't clear what the killer app for natural language technology
would be. But we all believed that somewhere, somehow we would find a pony. And
it wasn't clear how we would structure things to go after the application when
it became apparent. We didn't predict Powerset or search, but here we
are, just about to launch a broad-scale search engine. An engine based on the
deep NL and KR technology that emerged from all those years of research--what a
ride! And what a great collaboration!
And search may be just the beginning
of the end. In the back of our minds and with Powerset success, there is still
the possibility of building that Genial Understander System, what we would now
call a Conversational User Interface (a CUI).
GUS2--how hard could it be?
Let's get started...
No comments:
Post a Comment