NLP plans for Shen

[Antti Ylikoski]

My Master of Science thesis, and the doctoral work, were in general

Artificial Intelligence.

I have some plans as to Natural Language Processing work with Shen, I

have previously written about these ones in the group.

Following are some notes as to this work:

1) I have been experimenting with the Common LISP code in Paul Graham's

book On Lisp, an Augmented Transition Networks parser with nondeterminism

with Common LISP.  This could be somewhat improved, and translated

into Shen.

2) I have done some experimenting with P H Winston's Augmented Transition

Trees, in his books LISP, and Artificial Intelligence.  Conversion

into Shen would be pretty straightforward.

3) The most promising road is grammars with Prolog, and Shen Prolog.

I have done some work with Definite Clause Grammars with the standard

Prolog, conversion into Shen straightforward.

4) Steven Tanimoto's book on AI has a Context Free Languages parser, I

wrote it and tested it with Common LISP.

5) The Russell--Norvig book has content about PCFG, Probabilistic

Context Free Grammars, I did some work concerning the PCFGs and the Shen

ADTs.

Then:

What would the applications of Shen NLP be like?

Natural language interface for a database system?  Or, for some web pages?

Or, at the other end of the complexity spectrum, (limited) full scale

text understanding?

yours, Dr AJY

Helsinki, Finland, the EU

[Mark Tarver]

In terms of performance and ease of use Shen-YACC would be the best choice for NLP.

Mark

[Antti Ylikoski]

I attach to this entry a file that was in the 'net in the public
domain.  Implementing that with the Shen-YACC would be, so I
understand, rather straightforward.

The Jurafsky--Martin is a well-known reference in the NLP field.

AJY

[Mark Tarver]

Yes; Shen-YACC would cope with that easily.   The problem with NLP is the number of rules needed to process English.  Geoff Sampson at Linguistic in Leeds reckoned that you needed about 20,000 CF rules to parse English comfortably.   Apart from that, parsing is 1/2 the story.  You need to transduce the English into some formal language that the computer can understand and you need to be able to reason with it.

Mark 

[Antti Ylikoski]

Yes, thanks for your comment showing very high level expertise!

Indeed, the last part of your comment is why I was asking the
question, what would be the practical applications of the Shen NLP.

I mean, creating a natural language interface for a database system; or,
for some queries from some www pages; or for a system to manage large
private networks; or something analogous to those ones, would require only
a relatively small fraction of those ~ 20'000 CFL rules that you mention.

For that purpose, the simple sample CFL grammar from the Jurafsky--Martin,
or some even simpler a-fraction-of-English grammar would be sufficient.

As to the conversion of the NL into a computer representation, that
point is completely up to the application.  The app might require
FOPL; or perhaps database commands; or maybe commands to manage a
large private subnetwork, with specific software built to manage the
said private network.

The above mentioned kinds of applications was what I had in mind.

Someone might want to build a program for the comprehension of full
scale English, but that would be a large project.  Clearly doable with
Shen, though.

yours, AJY
Helsinki,Finland, the EU

[fuzzy wozzy]

once upon a time a non-native speaker of english said that english is a very simple language,
compared to other languages, that english is very good for business and science but lacks
when it comes to expressing things that are non-business, non-scientific, in a sort of what you
don't know you don't know, you don't know, sort of way, ok am paraphrasing, but...

at the time it seemed a shocking thing to say and hear, but this was once upon a time, and
what seems shocking now is how slow the progress of computerized nlp seems to be, in light of
all the advances made in all other scientific and computing fields, especially... all the functional,
logical, object orientical progress has not yet come up with something that could match up to the elisa
program, that while people knew it was just a computer faking to be  human, they were engrossed and
immersed and became attached to the program's ability to mimick human linguistics especially in the
aspect of human emotional empathy,

if elisa fascinated programmers and users alike in the 70's, some 40+ years later, how much progress
based on that excellent start could we have expected by now, if the progress was allowed to be made
in a normal evolutionary fashion? it's possible that some great progress has been made, but has not been
made available to the public as yet, but one can't help but wonder whether the whole nlp computing
complexity is not as complex as people make it out to be, or at least a bit simpler than people believe it to be

[Antti Ylikoski]

Dear fuzzy wozzy:

Joseph Weizenbaum's ELIZA was a so called micro world.

Its domain of discourse was very limited, and its vocabulary (lexicon)

also was very limited, and therefore the "Rogerian psychiatrist

session" could be handled with very simple techniques.

See

http://www.aaai.org/home.html

the members' magazine of the AAAI (the Association for the Advancement

of Artificial Intelligence) has discussed some modern NLP systems.

One of those examples was a competition where the task was to read and

understand some chapters from an English language chemistry textbook,

and thereafter answer some questions from it.  The winner, IIRC, was

from Stanford University, and the run took I think two days of CPU

time in a modern LISP machine.  But the computer program was able to

read and understand the chemistry textbook.

Also, see the story of the Watson program by the IBM:

http://www.ibm.com/watson/what-is-watson.html

https://en.wikipedia.org/wiki/Watson_(computer)

yours, AJY