LINGUIST List 3.759

Thu 08 Oct 1992

Disc: Unification vs Constraint-Based Grammars

Editor for this issue: <>

Directory

1. , Unification theories et al.
2. John S. Coleman, 3.756 Unification vs Constraint-Based Grammars
3. John S. Coleman, 3.756 Unification vs Constraint-Based Grammars
4. Mark Johnson, Re: 3.756 Unification vs Constraint-Based Grammars

In answer to Joe Stemberger's question about `Postalian best theory,'
Postal wrote a paper during the `Linguistic Wars' (between Generative
Semantics and Interpretive Semantics) arguing that a theory that
only used one kind of rule (Global rules, which were super-transformational
rules that could look both forwards and backwards) was better than
a theory that used different kinds of rules (say, transformations
AND constraints). He called the paper `The Best Theory' (I think
he was trying to score one off Chomsky's defenders, who had coined
`Standard Theory' for the Aspects model). The reference is `in
_Goals of Linguistic Theory_, edited by Stanley Peters, pp. 131-170,
Englewood Cliffs: Prentice Hall. 1972.
Geoff Nathan <GA3662SIUCVMB.SIU.EDU>

I don't agree with Henry Thompson's claim that the term ``Unification''
> although often used to gloss a
> procedurally-flavoured tutorial, had no place in the definition of the
> grammar formalism as such.
In fact, I regard this sentiment as a) historically incorrect, and
b) sociologically inadvisable, and c) terminologically imperialistic ;-)

First, the history. Although Kay (1979) might have been the first or one
of the first to employ the term ``unification'' in the context of
grammatical theory, most linguists, I suspect, first encountered
it in the GPSG literature. Cf. Gazdar et al. 1985, for example, pp. 26-27:

	The present theory of features makes heavy use of notions of
	*extension* and *unification* ...

	Our concept of *unification* is essentially identical to that
	of Kay (1979), and is closely analogous to the operation of
	union on sets except that, as in the case of extension, the
	resulting set must be a function. Unification is undefined
	for sets containing feature specifications that contradict
	each other.

	(10)	Let K be a set of categories. The *unification* of K
		(|_|K) is the smallest category which is an extension of
		every member of K, if such a category exists, otherwise,
		the unification of K is undefined.

According to this description, it seems to me that unification is no
more "procedurally-flavoured" than, say, set union, or equality. Just
as we should not confuse set union or equality with procedures which
might be employed to compute the union of some sets or whether or not
two quantities are equal, so we should not confuse the unification of
a set of categories (which is a category) with procedures for computing
such a category.

Second, the sociology. For some reason which is a mystery to me, but which
may have something to do with Stu Shieber's book, the term
``unification-based'' has attached itself to a class of theories
which uncontroversially include GPSG, HPSG, LFG, Functional Unification
Grammar, Unification Categorial Grammar, and various other frameworks
of the same general flavour. The term is used to distinguish this class
from historically prior theories, such as unreconstructed transformational
grammar, and contemporary "competitor" approaches, i.e. GB grammar.
Since the practitioners of "unification-based" grammars do indeed wish
to dissassociate themselves from TG etc., and since the label
"unification-based" serves this purpose very well, it seems
foolhardy to me to get rid of the term which is most widely-known
and used, despite any reservations (however well-founded, perhaps)
that Henry and others may have about it at times. Trying to get
rid of the label will most likely be unsuccessful anyway.

``Constraint-based'' is altogether too broad for my liking. Henry
says:
> I'll leave it to the
> proponents to suggest new cover terms to distinguish the soi-disant
> constraint-based theories from GB and its descendants.
tacitly recognising that ``constraint-based'' may be too close to
including GB for some people's liking.

So my vote is to stick with, nay, embrace the term ``unification-based''.
It does no injustice to the frameworks to which it is attributed,
and is a useful uniting epithet for an area which seems to delight
in schism and proliferation of names.

--- John Coleman

	An Appeal for Unification (terminological and sociological)
	-----------------------------------------------------------

I don't agree with Henry Thompson's claim that the term ``Unification''
> although often used to gloss a
> procedurally-flavoured tutorial, had no place in the definition of the
> grammar formalism as such.
In fact, I regard this sentiment as a) historically incorrect, and
b) sociologically inadvisable, and c) terminologically imperialistic ;-)

First, the history. Although Kay (1979) might have been the first or one
of the first to employ the term ``unification'' in the context of
grammatical theory, most linguists, I suspect, first encountered
it in the GPSG literature. Cf. Gazdar et al. 1985, for example, pp. 26-27:

	The present theory of features makes heavy use of notions of
	*extension* and *unification* ...

	Our concept of *unification* is essentially identical to that
	of Kay (1979), and is closely analogous to the operation of
	union on sets except that, as in the case of extension, the
	resulting set must be a function. Unification is undefined
	for sets containing feature specifications that contradict
	each other.

	(10)	Let K be a set of categories. The *unification* of K
		(|_|K) is the smallest category which is an extension of
		every member of K, if such a category exists, otherwise,
		the unification of K is undefined.

According to this description, it seems to me that unification is no
more "procedurally-flavoured" than, say, set union, or equality. Just
as we should not confuse set union or equality with procedures which
might be employed to compute the union of some sets or whether or not
two quantities are equal, so we should not confuse the unification of
a set of categories (which is a category) with procedures for computing
such a category.

Second, the sociology. For some reason which is a mystery to me, but which
may have something to do with Stu Shieber's book, the term
``unification-based'' has attached itself to a class of theories
which uncontroversially include GPSG, HPSG, LFG, Functional Unification
Grammar, Unification Categorial Grammar, and various other frameworks
of the same general flavour. The term is used to distinguish this class
>from historically prior theories, such as unreconstructed transformational
grammar, and contemporary "competitor" approaches, i.e. GB grammar.
Since the practitioners of "unification-based" grammars do indeed wish
to dissassociate themselves from TG etc., and since the label
"unification-based" serves this purpose very well, it seems
foolhardy to me to get rid of the term which is most widely-known
and used, despite any reservations (however well-founded, perhaps)
that Henry and others may have about it at times. Trying to get
rid of the label will most likely be unsuccessful anyway.

``Constraint-based'' is altogether too broad for my liking. Henry
says:
> I'll leave it to the
> proponents to suggest new cover terms to distinguish the soi-disant
> constraint-based theories from GB and its descendants.
tacitly recognising that ``constraint-based'' may be too close to
including GB for some people's liking.

So my vote is to stick with, nay, embrace the term ``unification-based''.
It does no injustice to the frameworks to which it is attributed,
and is a useful uniting epithet for an area which seems to delight
in schism and proliferation of names.

--- John Coleman

Message 4: Re: 3.756 Unification vs Constraint-Based Grammars

Well, I guess my opinion is summed up by the topic of the seminar I'll
be teaching here at Brown next semester ``CG 251: Constraint-based
Natural Language Understanding'', which will focus on GB parsing!

In my opinion the major difference between the so-called ``unification-
based'' approaches and the ``principles and parameters'' (GB) approach
lies in their methodologies. (Of course I agree with Henry Thompson
that the name ``unification-based'' is a bad one - but that's the one
I'll use here). The ``unification-based'' approaches
usually focus on particular kinds of constraints (say, equational
constraints between feature values) and explore to what extent natural
language can be described in terms of them. Examples of this kind of
work include Pollard and Sag's ``Information-based Syntax and Semantics''
CSLI Lecture Notes Series, and the more theoretical Computer-Science
work done by Bob Carpenter ``The Logic of Typed Feature Structures'',
Cambridge Tracts in Theoretical Computer Science. I have my own line
on this type of approach - see my paper ``Features and Formulae'' in
Computational Linguistics 17.2 1991 and the references cited therein
(and email me if you're interested in some of my more recent papers).

The ``principles and parameters'' methodology regards the nature of
constraints determining natural language as an empirical issue. Rather
than fixing on a language in which constraints are to be expressed in
a priori, they regard the identification and informal characterization
of the constraints as the key issue. Mathematical precision is not
regarded as a virtue in itself. This does not mean that they theory
cannot be formalized: Ed Stabler's new book from MIT Press shows
that important components of Chomsky's Barriers theory can be expresed
in FOL, for example.

I think that the parsing problem for theories of both the ``unification-
based'' and ``principles and parameters'' kinds can be naturally seen
as a kind of simultaneous constraint satisfaction. These sorts of
techniques can be easily implemented in a logic programming setting
(as has already been pointed out); see the papers in Brown and Koch
``Natural language understanding and logic programming III'' for a
variety of applications of such techniques.

Mark Johnson mjcs.brown.edu