LINGUIST List 3.774

Wed 14 Oct 1992

Disc: Unification vs Constraint-Based Grammars

Editor for this issue: <>


Directory

  1. Hans Uszkoreit, Re: 3.756 Unification vs Constraint-Based Grammars
  2. Avery Andrews, `unification-based' vs `principles & parameters-based' grammars
  3. , Re: 3.759 Unification vs Constraint-Based Grammars

Message 1: Re: 3.756 Unification vs Constraint-Based Grammars

Date: Sun, 11 Oct 1992 15:50:45 Re: 3.756 Unification vs Constraint-Based Grammars
From: Hans Uszkoreit <uszkoreitcoli.uni-sb.de>
Subject: Re: 3.756 Unification vs Constraint-Based Grammars

"Unification-based grammars" vs. "constraint-based grammars" vs.
"information-based grammars" -- I received numerous questions concerning
the taxonomic issue when I gave talks and tutorials on formalisms of this
paradigm. Personally I do not care very much what the terminological
outcome of the discussion might be as long as the proposed term is
pronouncable and fits on one line. Most of the existing paradigms in
grammatical theory are known under names that are of little help to the
uninitiated. But the discussion can serve the purpose to create awareness
of some common misconceptions. This, I take it, was Henry Thompson's
motivation for formulating his reply.

Lets start with the broadest term "information-based grammars". Which
grammar isn't? No wonder that the term didn't catch on! I also agree with
Avery Andrews that "constraint-based grammar" is much too broad since it
may easily refer to grammars employing totally different constraint
systems. (But I could imagine using "constrained-based" if nobody else
besides Avery Andrews files a claim.)

So what about "unification-based grammar"? The term might still be too
broad
since it usually excludes the so-called "logic grammars" (DCG,
Extraposition Grammar, Gapping Grammar) that are PROLOG-based declarative
grammars and thus use unification. But then, feature-unification grammars
are based on some feature LOGIC and might therefore be called logic
grammars as well.

Now Henry points out that a class of grammar formalisms shouldn't be named
after some type of algorithm used to implement them.

First, let's ask what do these grammars have in common? They use
conjunctions of features (i.e. conjunctions of attribute-value pairs) as
partial representations of linguistic objects. They also use some notion of
equality to express the fact that two descriptions describe the same
object. In addition they need a binary operation on representations that
merges the information of two representations just in case they refer to
the same object. (Often this operation in the feature logic is also
correctly referred to by the more generic term "greatest lower bound" or
"least upper bound", depending on the underlying lattice structure.)

If the term "unification" is just applied to the algorithm implementing the
operation for merging feature structures or feature structure descriptions,
then I fully agree with Henry. However, if the term is (also) used for the
implemented operation in the underlying feature logic, then the basis for
Henry's objection disappears. Indeed, in the literature on feature logics
the operation is usually called "unification." Unification is the only
logical operation that is utilized by ALL grammar models that are called
"unification grammars" -- even if some of their proponents do not refer to
the operation explicitely.

Thus I would claim that in daily practice the term unification is actually
used for both the operation and the algorithm. Since computational
linguists almost always say "unification algorithm" when they refer to the
actual algorithm, it is obvious that they know the difference.

Of course one might say that the term "unification-BASED" is too strong. So
why not simply stick with the underspecified, old, short and simple term
"UNIFICATION GRAMMAR." Everybody knows this expression. It's a partial
description encoding one important feature of the grammar models. It can
easily be unified with other partial descriptions to yield terms such as
Functional Unification Grammar, Lexical Unification Grammar, Categorial
Unification Grammar etc.

In the beginning I said that I do not really care what the terminological
result of the discussion might be as long as the proposed term is short and
intuitive. But if I soon have to refer to publications on
"complex-feature-constraint-oriented grammar formalisms with equality",
I'll retaliate. I might start hair-splitting discussions on the terms
"categorial grammar", "transformational grammar", "word grammar",
"dependency grammar", "construction grammar", etc.

...or on slightly more reasonable questions such as: Is "Lexical Functional
Grammar" more lexical than dependency grammar, word grammar, categorial
grammar or HPSG. (Is it more functional?) Will Head-Driven Phrase Stucture
Grammar soon be neither? Which grammar theories are really principle-based
and why?

Hans Uszkoreit

P.S. Rich Hilliard, if you would like to catch up on connections between
CLP and unification grammar, please look at papers on the subject by
Smolka, Hohfeld and Smolka, Damas and Varile.
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue

Message 2: `unification-based' vs `principles & parameters-based' grammars

Date: Mon, 12 Oct 92 16:19:12 ES`unification-based' vs `principles & parameters-based' grammars
From: Avery Andrews <andalingdurras.anu.edu.au>
Subject: `unification-based' vs `principles & parameters-based' grammars

It strikes me that this selection from Mark Johnson's recent message
(Linguist 3-759) might be read as a strong endorsement of the `principles
& parameters' over the `unification-based one':

 >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.

On the other hand it takes a fair amount of work to figure out whether
some novel formal idea is even prima-facie appropriate for natural language
description, and even more to make a case that it is actually better
than some other possibilities. I tried to do this for unification vs.
movement in my 1990 NLLT paper on morphological blocking, but even if
this particular demonstration is taken as 100% successful it wouldn't
be enough to establish the case for movement vs. unification in general.

So I believe that large-scale and reasonably determined efforts to
get somewhere with specific formal notions ought to have a place in
(on ?) the enterprise.

Avery.Andrewsanu.edu.au
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue

Message 3: Re: 3.759 Unification vs Constraint-Based Grammars

Date: Mon, 12 Oct 1992 12:10:22 Re: 3.759 Unification vs Constraint-Based Grammars
From: <petersmack.uit.no>
Subject: Re: 3.759 Unification vs Constraint-Based Grammars

Mark Johnson (mjcs.brown.edu) characterizes the difference between HPSG*
and GB** as follows:

>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.

>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.

But isn't the insistence in GB work on characterizing constraints in terms
of structural configurations an instance of "focusing on a particular kind
of constraint"? E.g. it is assumed in most current GB work that a noun
phrase (asymmetrically) c-commands the anaphor it binds. The many apparent
counterexamples to this have led not to an abandonment of c-command as an
interesting relation (a notable exception would seem to be the recent work
on anaphora by Reinhart and Reuland), but instead to (1) much more
complicated conditions, all referring crucially to c-command or m-command,
and (2) complement structures of a highly counter-intuitive nature, such as
the double-object cluster most recently detailed by Larson.

It seems that configurational relations such as c-command and the spec-head
relation are not considered "an empirical issue", but are assumed as
guiding principles. Another example of a guiding principle in GB would
seem to be the assumption that branching is maximally binary, rarely
justified with any sort of argument (here Kayne is an exception). HPSG, it
should be noted, does not make any reference to c-command or m-command and
does not assume that branching is maximally binary (though there could be
said to be an analogue to the spec-head relation).

* Johnson mentions specifically the version of HPSG presented in Pollard
and Sag 1987, but I think the characterization would be intended to carry
over to the newer volume by the same authors as well.

** I persist in referring to the framework as GB for the sake of generality
(I mean to include work from the eighties, before the term 'Principles and
Parameters' was coined) and familiarity and out of old-fashionedness (I
still think of PP as meaning 'prepositional phrase').

References:
Kayne, Richard 1984: _Connectedness and Binary Branching_, Foris.
Larson, Richard 1988: 'On the Double Object Construction', LI 19.3:335-391.
Pollard, Carl and Ivan Sag 1987: _Information-Based Syntax and Semantics_,
 CSLI Lecture Notes Neries No. 13.
Pollard, Carl and Ivan Sag (to appear): _Head-Driven Phrase Structure
Grammar_,
 to be published jointly by Univ. Chicago Press and CSLI.
Reinhart, Tanya, and Eric Reuland 1991: 'Reflexivity', ms. Tel
Aviv/Groningen;
 to appear in LI? (I have seen it cited as such)

=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=
Peter Svenonius, UC Santa Cruz Linguistics and
ISL, Univ. i Tromsoe, N-9037 Tromsoe, Norway tlf:+47 (83) 44759
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue