RE: ] Re: [CG] Language Games and Semantic Domains
In an earlier post you mentioned WordNet Domains and gave a url there for
wndomainds.fbk.eu which I visited. Is this WNDomains project based on the
Masterman approach in a recognized way? A lot of the concepts you described
in your paper http://www.jfsowa.com/pubs/mmb_rev.htm and in your post below.
MM's approach, as you described it, seems a lot more appropriate to the
realities of automated language analysis than the usual syntactic theories.
Her theories of the thesaurus of word groups, each word associated with a
fan of edges to context labels, even seem to have been at least slightly
mechanically structured for computation. She may have inspired the Quillian
who wrote about "spreading activation trees" of concept nodes.
But even with WNDomains, progress in implementing MM's mechanistic approach
seems stifled and unfulfilled. So much more could have been done in all
those years since 1980. Have there been any modern outgrowths or
continuations of MM's empirical semantic focus you can point to or name?
Rich AT EnglishLogicKernel DOT com
From: John F. Sowa [mailto:sowa@...]
Sent: Saturday, January 02, 2010 11:29 AM
Subject: [CG:] Re: [CG] Language Games and Semantic Domains
I received some offline comments about the previous posting
on this subject.
> People involved in NLP are, for the most part, not concerned
> with philosophy.
I certainly agree. For the most part, they just follow the fads,
which are generated by people who are concerned with philosophy.
> The members of the Vienna Circle really weren't philosophers.
Their PhDs were in math & science. That puts them in the same league
as Descartes, Leibniz, and Kant. (Kant, by the way, taught Newtonian
mechanics and is credited with the hypothesis that the planets formed
from a disk-shaped cloud of dust around the sun.) The VC followed
Ernst Mach, who was an experimental physicist (best known for the
Mach number). But their writings are put in the same category with
Plato the mathematician, Aristotle the biologist, Peirce the chemist,
Whitehead the mathematician, and Wittgenstein the engineer.
> Montague and Chomsky at least provided some guidance
> in implementing basic semantic interpretation.
But the basic parsing technology was developed in the computer
field, independently of Chomsky and Montague. John Backus was
the leader of the IBM project that produced the first FORTRAN
compiler in 1957 -- the same year of Chomsky's first book.
Backus and Naur developed BNF for the definition of Algol,
independently of Chomsky, and Peter Lucas implemented it with
the first version of a recursive decent parser.
Long before Montague, the computer scientists developed methods
of syntax-directed compilation (one-to-one association of grammar
rules and semantic rules). Bill Woods applied that technique
to English for his PhD dissertation of 1967. Montague didn't
publish his two famous papers that defined "Montague grammar"
until 1970. By that time, Bill was working at BBN, where he
implemented his approach in an English query system about moon
rocks -- long before anyone implemented Montague grammar.
And by the way, the three-way distinction of syntax, semantics,
and pragmatics was based on Peirce's triad of grammar, logic,
and rhetoric (which he based on the medieval Trivium). In the
1930s, Charles Morris replaced Peirce's terms with the currently
popular syntax, semantics, and pragmatics.
But there were earlier systems for NLP in the 1950s, long before
Chomsky. Silvio Ceccato published a paper about his "correlation
nets" for semantics in 1956, which he implemented on an IBM 650 --
a vacuum-tube machine with a drum memory. Following is the table
of contents for the 1961 International Conference on Machine
Translation of Languages and Applied Language Analysis:
There were several papers on network notations by Silvio Ceccato,
by David Hays, and by Margaret Masterman. After the table of
contents is a copy of Masterman's paper, in which she has the
first published use of the term "semantic net." At the end of
this note is the URL of a review I wrote of Masterman's
collected papers and some excerpts from it.
Michael Halliday was one of the cofounders of CLRU with Masterman,
and his work is a more tightly integrated combination of syntax,
semantics, and pragmatics than anything by Chomsky or Montague.
Some of Halliday's early work formed the basis for Terry Winograd's
SHRDLU system, which he implemented for his PhD dissertation (1971),
which also antedated any implementation of Montague grammar.
Halliday's later work stimulated Rhetorical Structure Theory
(RST), which is a widely used approach to pragmatics.
As I said, the bits and pieces of technology are fine. The
major problem is the question of how to put them together.
Both Masterman and Halliday, who were at Cambridge with
Wittgenstein, had found a better way. But it was swamped by
the far noisier and more misguided groups who followed the
Frege-Russell-Carnap-Quine direction. Montague and Chomsky
were infected with that virus, and they passed it along.
> We now have reasonable parsers, both structural and dependency
> oriented, and both statistical and symbolic, means of converting
> the results to partial semantic interpretations (partial in the
> sense of not accounting for everything), ways of dealing with
> lexical ambiguity, and highly active research in filling out
> the blanks (large scale discourse analysis, sentiment analysis, ..).
Yes, indeed. Those are excellent pieces of technology, none of
which require anything by Chomsky or Montague. The major question
is how to put them together. Masterman and Halliday, who were
influenced by Wittgenstein's later work, had found a good way.
But it was swamped by the far noisier and more misguided groups who
followed the Frege-Russell-Carnap-Quine direction. Montague and
Chomsky were infected with that virus, and they passed it along.
The dependency parsers were based on Tesnière. And Chomsky had
a strong negative influence on the use of statistics in NLP.
Peirce, by the way, is credited with having made some important
updates and extensions to Laplace's theory of probability.
In summary, ideas have consequences. Unfortunately for AI and NLP,
the bad ideas had much more hype behind them.
Excerpts from http://www.jfsowa.com/pubs/mmb_rev.htm
Margaret Masterman was one of six students in Wittgenstein's course of
1933-34 whose notes were compiled as The Blue Book (Wittgenstein 1958).
In the late 1950s, she founded the Cambridge Language Research Unit
(CLRU) as a discussion group, which evolved into one of the pioneering
centers of research in computational linguistics.
As a student of Wittgenstein, Masterman was also deeply concerned about
the foundations of theoretical linguistics. Around the same time that
Chomsky was developing his syntactic theories and Montague was
advocating a logic-based alternative, she was proposing a
"Neo-Wittgensteinian" view, whose organizing principle was a thesaurus
of words classified according to the "language games" in which they are
used. Although no single paper in the book formulates a succinct summary
that could be called a theory, the following principles are discussed
* Focus on semantics, not syntax, as the foundation for language:
"I want to pick up the relevant basic-situation-referring habits
of a language in preference to its grammar" (p. 200).
* Recognition that ambiguity is a consequence of the flexibility and
extensibility of natural language and not a defect that can be
eliminated by switching to a purified language of logic.
* Context-dependent classification scheme with three kinds of
structures: a thesaurus with multiple groups of words organized
by areas of use, a fan radiating from each word in the thesaurus
to the area in which it occurs, and dynamically generated
combinations of fans for the word tokens of a text.
* Emphasis on images as a language-independent foundation for meaning
with a small number (about 50 to 100) of combining elements
represented by ideographs or monosyllables, such as IN, UP, MUCH,
THING, STUFF, MAN, BEAST, PLANT, DO.
* Recognition that analogy and metaphor are fundamental to the
creation of novel uses of language in every field, especially in
the most advanced areas of science.
Unlike the a priori formalisms of Chomsky or Montague, this approach is
based on data about actual language use. In the commentary, Wilks noted
that Masterman's work contained "the germ of what was later to be called
EBMT or example-based translation (Nagao 1989), which is now perhaps the
most productive current approach to MT world-wide, and I have heard
Professor Nagao refer to [her] in this connection in a lecture" (p. 279).
As a whole, the book presents a cognitive view of language that has
strong similarities to the Cognitive Linguistics by Croft and Cruse
(2004). Croft's radical construction grammar, Cruse's dynamic construal
of meaning, and Lakoff and Johnson's work on metaphor (1980) are
compatible with and to some extent anticipated in Masterman's papers.
The multiplicities of context-dependent word senses discussed in the
first paper of the book could be aptly characterized by the term
'microsense', which was coined by Cruse (2000). Although most of the
papers are forty years old or older, the goal of implementing the ideas
in a computable form has forced a greater attention to detail and
precision than is found in some of the more recent work on cognitive
The age and origin of most of the papers as unpublished memos is evident
in their rather disorganized structure, but the book contains many
intriguing insights that still seem fresh today. Among them are her
penetrating criticisms of Chomsky's fixation on syntax:
MM> My quarrel with [the Chomsky school] is not that they are
> abstracting from the facts. How could it be? For I myself in this
> paper am proposing a far more drastic abstraction from the facts.
> It is that they are abstracting from the wrong facts because they
> are abstracting from the syntactic facts, that is, from that very
> superficial and highly redundant part of language that children,
> aphasics, people in a hurry, and colloquial speakers, quite
> rightly, drop. (p. 266)
As an alternative, she discussed the writings of the phoneticist Peter
Guberina (1954), who had worked in a school for the deaf:
MM> A large part of Guberina's daily life is spent in developing
> electronic techniques for helping the deaf to speak. This means
> that, for him, what is being talked about that is, the actual
> subject of any piece of discourse, and the linguistic elements
> that carry it is vastly more important than what is said about
> it. If the deaf man can once pick up the subject of conversation,
> three-quarters of this problem is solved, even if he cannot hear
> all that is said about it. If, on the other hand, he clearly hears
> some one thing that is clearly said about some basic subject of
> discourse, while the actual subject of discourse remains unknown
> to him, very little of the deaf man's problem is solved; he has
> only heard one thing. (p. 228)
In summary, she said that "human communication consists of patterns
of semantic interactions between ascertainably cognate subjects of
discourse." By cognate subjects, she meant ones that originate from the
same or similar language games and are grouped in the same area of a
thesaurus. The semantic patterns led to the templates of Wilks' own
theory of preference semantics, and they are closely related to the
chunks, frames, scripts, and schemata of other systems.
To unsubscribe, e-mail: cg-unsubscribe@...
For additional commands, e-mail: cg-help@...
To unsubscribe, e-mail: cg-unsubscribe@...
For additional commands, e-mail: cg-help@...