nl-kr-request@CS.ROCHESTER.EDU (NL-KR Moderator Brad Miller) (03/11/88)
NL-KR Digest (3/10/88 23:10:41) Volume 4 Number 24
Today's Topics:
First Person Plural - Inclusive/Exclusive?
Re: First Person Plural - Inclusive/Exclusive?
Re: Ambiguity in 1st person plural
Lecturing job in NLP in Cambridge
re:what are grammars (for)?
What are grammars (for)?
Re: Study of typing errors
RE: What is a grammar (for)?
Filters (was 're: uses of Grammar')
Re: Becoming CAI literate (ergative note)
Language & Cognition Seminar
Submissions: NL-KR@CS.ROCHESTER.EDU
Requests, policy: NL-KR-REQUEST@CS.ROCHESTER.EDU
----------------------------------------------------------------------
Date: Mon, 29 Feb 88 18:20 EST
From: nehaniv@chex.berkeley.edu
Subject: First Person Plural - Inclusive/Exclusive?
In all this talk of first person plural, have we been careful to check what
we mean by inclusive and exclusive "we". It seems we have not: Mandarin
Chinese, cited as a first case by many netters, does have the distinction, but
only in the following sense: ZAMEN is the "inclusive" we in that it may
only include (and must include?) all
those present and no others.
WOMEN is neither inclusive nor exclusive, but
rather an unmarked first person plural.
So the distinction which exists in Mandarin is not inclusive/exclusive but
rather [+/- "inclusive"], where "inclusive" means restricted to those present.
This distinction is clearly different and unequivalent to saying that ZAMEN
must include and WOMEN must exclude.
Now is this the type of inclusion everyone has been talking about or were we
dealing with different phenomena for different languages?
C. Lev Nehaniv
------------------------------
Date: Mon, 29 Feb 88 18:59 EST
From: Tom Slone <potency@violet.berkeley.edu>
Subject: Re: First Person Plural - Inclusive/Exclusive?
In article <7281@agate.BERKELEY.EDU> nehaniv@chex.berkeley.edu (C. Lev Nehaniv) writes:
>So the distinction which exists in Mandarin is not inclusive/exclusive but
>rather [+/- "inclusive"], where "inclusive" means restricted to those present.
In Indonesian/Malay, "kami" is exclusive (we/us) and "kita" is inclusive (we/us).
There is not an equivalent to the Mandarin "women" in Indonesian.
potency@violet.berkeley.edu
{decvax|hplabs|ihnp4|decwrl|nbires| \
sdcsvax|tektronix|ulysses}!ucbvax!violet!potency
------------------------------
Date: Wed, 2 Mar 88 04:37 EST
From: "J. A. \"Biep\" Durieux" <biep@cs.vu.nl>
Subject: Re: Ambiguity in 1st person plural
In article <3804@whuts.UUCP> jsw@whuts.UUCP (WHALEN) writes:
[Spanish: nos vs. nosotros]
[French: nous vs. nous autres]
I am not a native speaker of English, but isn't the same difference
implied in the English "we" vs. "the rest of us" ?
--
Biep. (biep@cs.vu.nl via mcvax)
Relative ethics amounts to saying that the only mis-
take Hitler made was one of timing: at the time the
majority unfortunately didn't subscribe his ideas.
------------------------------
Date: Wed, 2 Mar 88 08:38 EST
From: Bran Boguraev <bkb%computer-lab.cambridge.ac.uk@NSS.Cs.Ucl.AC.UK>
Subject: Lecturing job in NLP in Cambridge
LECTURER
NATURAL LANGUAGE PROCESSING
COMPUTER LABORATORY, UNIVERSITY OF CAMBRIDGE
This new post, for which a formal announcement will appear
shortly, will further expand the established natural language
group in the Laboratory whose senior members are Karen Sparck
Jones, Steve Pulman and Bran Boguraev. The Laboratory and
Department of Engineering are jointly responsible for a one-year
M Phil in Computer Speech and Language Processing, and the
primary teaching duties of the new post will be related to this
course, though we are looking for someone who would also be able
to contribute to the computer science teaching of the Laboratory
and to develop relationships with other areas of computer science
research here.
The natural language group's research includes both
theoretical and practical work on language tools, syntactic and
semantic procesing, and meaning representation and inference, and
on specific task systems; the group has several funded projects
and is actively collaborating with other University departments
and with SRI International's Cambridge Computer Science Research
Centre.
We are seeking applicants with particular experience in
discourse processing, generation, translation, or speech, but
welcome applications from specialists in other areas of natural
language processing.
Further details from the Secretary to the Appointments
Committee, Computer Laboratory, New Museums Site, Pembroke
Street, Cambridge CB2 3QG.
------------------------------
Date: Wed, 2 Mar 88 10:52 EST
From: Stephan Busemann <BUSEMANN%DB0TUI11.BITNET@CUNYVM.CUNY.EDU>
Subject: re:what are grammars (for)?
In Reply to:
Date: Thu, 25 Feb 88 12:13 EST
From: John Nerbonne <nerbonne%hpljan@hplabs.HP.COM>
Subject: uses of grammar etc.
(in NL-KR Digest 4 (22), 03/01/88)
In Reply to:
Date: Tue, 16 Feb 88 18:13 EST
From: Bruce E. Nevin <bnevin@cch.bbn.com>
Subject: uses of grammar etc
(in NL-KR Digest 4 (20), 2/23/88)
BN> [...] many formalisms that have developed from the Generative
BN> paradigm but do not give themselves the Generative trademark: GPSG,
BN> HDPSG, unification grammars, categorial grammars, and so on and on.
BN> These formalisms have in common that they overgenerate, that is, they
BN> produce structures that do not occur in natural language and must be
BN> pruned or filtered out.
JN>But to speak of a formalism as (over)generating is a category
JN>error. The formalism doesn't generate at all--over-generate or
JN>under-generate or otherwise. It just allows the formulation
JN>of rules and grammars whose job is to generate.
I think both kinds of overgeneration do exist. GPSG is a good example
(in its 1985 version: Gazdar, Klein, Pullum, Sag: Generalized Phrase
Structure Grammar, Basil Blackwell, Oxford, Great Britain). The formalism
is declaratively defined without explicitly describing how the components
are supposed to work together. When you try to implement it, however,
you MUST specify the 'generative device' that correlates syntactic structures
with NL sentences. In the GPSG case, this means (list is not complete):
- to generate every possible extension for every category in an Immediate
Dominance (ID) rule,
- to determine which of them are legal (by virtue of Feature Cooccurrence
Restrictions (FCRs)),
- to generate all possible projections of ID rules with the legal categories
thereby creating every possible order of daughters,
- to determine which of them do not violate the different Feature
Instantiation Principles (FIPS) and the Linear Precedence (LP) statements.
As it turns out, GPSG-85 requires a vast overgeneration of categories and
projections of ID rules. This is a property of the formalism; it is
clear that a careless formulation of the grammar may cause further trouble.
JN>Because filters are suspiciously powerful devices, they ARE NOT USED
JN>(outside of MIT's Government and Binding theory). (A final
JN>qualification about unification grammars might be worthwhile:
JN>unification is a technique for combining feature complexes, and
JN>it certainly isn't incompatible with the use of filters. But
JN>unification-based theories such as LFG don't use filters.)
As should be clear from the above list, FCRs, FIPs and LP statements
in GPSG are nothing else but filters.
(As a consequence, GPSG-85, as it is described in the above-cited book,
can thus not be implemented on a real computer (i.e. with physical limits).
A different view has to be adopted instead, e.g. using FCRs, FIPs etc.
as constructive devices.)
Stephan Busemann
(busemann@db0tui11.bitnet)
------------------------------
Date: Wed, 2 Mar 88 11:55 EST
From: Rick Wojcik <rwojcik@bcsaic.UUCP>
Subject: What are grammars (for)?
Arild Hestvik (2/25) writes:
AH>
AH> be of no use in understanding natural language. We want the grammar to tell
AH> us *why* a string is grammatical or ungrammatical, i.e. the grammar should
AH> give a structural description (an analysis) of both well-formed AND
AH> ill-formed expressions.
One can use a generative grammar to give structural analyses to parts of
ill-formed strings. This is one possible use of a chart parser in NLP
systems. But having a lot of well-formed pieces does not tell you how
to render the expression interpretable. There are so many reasons why a
string could be ungrammatical that there is really no hope of building
an automated string-repair device into a grammar. Given the way in
which we currently define grammars, it is self-contradictory to talk
about grammars that analyze ill-formedness. You would have to develop a
concept of well-formed ill-formedness.
John Nerbonne (2/25) writes:
JN> A grammar that is generative in Chomsky's sense (a complete
JN> characterization of what is in the language) is of value to, say NL
JN> Understanding systems, since it allows one to arrive more surely at
JN> the genuinely available analyses of NL input. It is a partial
JN> solution to the problem of multiple parses in NLU. The problem
JN> doesn't thereby go away, but is reduced. Harris's work sounds to
Generative grammars do not reduce the problem of multiple parses. You
don't get multiple parses in the first place without grammars. In fact,
generative linguistic theory is not designed to explain how grammars are
used in language understanding. It is left up to the psychologist or
computer scientist to address the issue. The fact is that natural
languages are not "understood" in the same way that computers "understand"
artificial languages. Generative grammar would make a lot more sense if
they were. Natural speech/writing can be highly 'degenerate' (ill-formed),
unlike computer programs. Yet humans still manage to understand ill-formed
language. In fact, ill-formed readings are sometimes preferred over
well-formed ones, as the "Horse raced past the barn fell" sentence
illustrates. So Martin Kay's question about what grammars are for still
remains pertinent and largely unanswered.
--
Rick Wojcik csnet: rwojcik@boeing.com
uucp: {uw-june uw-beaver!ssc-vax}!bcsaic!rwojcik
address: P.O. Box 24346, MS 7L-64, Seattle, WA 98124-0346
phone: 206-865-3844
------------------------------
Date: Wed, 2 Mar 88 15:22 EST
From: Paul G. Chapin <pchapin@note.nsf.gov>
Subject: Re: Study of typing errors
Ken Laws mentions the interest of studying typing errors. It may be worth
mentioning that investigation of typing errors was what originally led the
great experimental psychologist Karl Lashley to the identification of "The
Problem of Serial Order in Behavior", a paper cited by Chomsky in his famous
review of B. F. Skinner's book _Verbal Behavior_ as providing evidence demol-
ishing the strict version of behaviorist psychology. Lashley categorized
the kinds of typing errors observed, and noted that many of them were of the
type Ken Laws describes, where a particular letter is out of sequence. A
strict S-R description of a motor behavior like typing is at a loss to account
for this phenomenon. Misplaced notes in rapid piano playing are another
example adduced by Lashley.
Just a bit of history now; hard to believe what a hot argument it was in
the late 50's.
Paul Chapin
NSF
------------------------------
Date: Fri, 4 Mar 88 11:17 EST
From: rutgers!gpu.utcs.toronto.edu!bnr-vpa!nrcaer!crcmar!patrick
Subject: RE: What is a grammar (for)?
Martin Taylor (mmt@dciem.UUCP) asks an interesting question:
>A question that is simple on the surface, but I suspect not so simple in
>implication: "What is a grammar, and what is a grammar for?"
An provocative discussion on this question may be found in an article
(and commentary) by Edward Stabler, Jr. in THE BEHAVIORAL AND BRAIN
SCIENCES, 1983, 6, 391-421. Here is the abstract to that paper:
Abstract: Noam Chomsky and other linguists and psychologists have
suggested that human linguistic behavior is somehow governed by a
mental representation of a transformational grammar. Challenges to
this controversial claim have often been met by invoking an explicitly
computational perspective: It makes perfect sense to suppose that a
grammar could be represented in the memory of a computational device
and that this grammar could govern the device's use of a language.
This paper urges, however, that the claim that humans are such a device
is unsupported and that it seems unlikely that linguists and
psychologists really want to claim any such thing. Evidence for the
linguists' original claim is drawn from three main sources: the
explanation of language comprehension and other linguistic abilities;
evidence for formal properties of the rules of the grammar; and the
explanation of language acquisition. It is argued in this paper that
none of these sources provides support for the view that the grammar
governs language processing in something like the way a program governs
the operation of a programmed machine. The computational approach, on
the contrary, suggests ways in which linguistic abilities can be
explained without the attribution of an explicit representation of
rules governing linguistic behavior.
I believe that Dr. Stabler can be reached at "stabler@uwocsd.uwo.cdn".
--
Andrew Patrick, Ph.D.
INTERNET: patrick@crcmar.uucp
UUCP: ... uunet!mnetor!utzoo!dciem!nrcaer!crcmar!patrick
BITNET: patrick%crcmar@UTORGPU
PHONE: (613) 990-4675
CANADA POST: Division of Behavioral Research, Communications Research Center,
P.O. Box 11490, Station 'H', Ottawa, ON, CANADA K2H 8S2
------------------------------
Date: Fri, 4 Mar 88 19:36 EST
From: John Nerbonne <nerbonne%hpljan@hplabs.HP.COM>
Subject: Filters (was 're: uses of Grammar')
In reply to:
Date: Wed, 02 Mar 88 10:52:41 MEZ
From: Stephan Busemann <BUSEMANN%DB0TUI11.BITNET@CUNYVM.CUNY.EDU>
Subject: re:what are grammars (for)?
1. Filters
SB> JN>Because filters are suspiciously powerful devices, they ARE NOT USED
SB> JN>(outside of MIT's Government and Binding theory). (A final
SB> JN>qualification about unification grammars might be worthwhile:
SB> JN>unification is a technique for combining feature complexes, and
SB> JN>it certainly isn't incompatible with the use of filters. But
SB> JN>unification-based theories such as LFG don't use filters.)
SB> [...] FCRs [Feature Cooccurrence Restrictions], FIPs [Feature
SB> Instantiation Principles] and LP [Linear Precedence] statements
SB> in GPSG are nothing else but filters.
A filter in grammar is a device that allows an otherwise legitimate
derivation to abort. You run the rules, get a structure, then check
it against the filter. The filter throws some things out. Here's
an example of Chomsky's:
* [_{Sbar} +/- WH NP-trace ... ]
The plus or minus WH indicates a complementizer like 'that' or 'whether',
and the Sbar (intended as subscript) indicates that this applies to a
subordinate clause. The asterisk indicates that these structures are
ill-formed. Using this filter, you may have a grammar that defines
an analysis tree for:
Who did say that left?
But if this has a substructure like this:
[_{Sbar} "that" NP-trace "left"]
then you get to filter the derivation out. It doesn't count.
My claim above is that GPSG and most other syntactic theories don't use
these, pace BN (Bruce Nevin), who I was replying to:
BN> [...] many formalisms that have developed from the Generative
BN> paradigm but do not give themselves the Generative trademark: GPSG,
BN> HDPSG, unification grammars, categorial grammars, and so on and on.
BN> These formalisms have in common that they overgenerate, that is, they
BN> produce structures that do not occur in natural language and must be
BN> pruned or filtered out.
2. GPSG and "Filters"
The concept of "generation" in GPSG is twofold due to its scheme of first
generating a grammar, then having the grammar generate the object language.
Schematically: Metagrammar ==> Grammar ==> Language
1 2
Let's call the generation at 1 "metageneration" and the generation at 2
just "generation", since (2) corresponds to the generation done in
theories without a level of metagrammar.
To emphasize my first point: there is NO use of syntactic filters that
allow some generations (at level 2) to abort.
SB says that Feature Cooccurrence Restrictions etc. are nothing but filters,
but the analogy is poor. A FCR simply restricts the categories in a language
so that we know e.g. that NP[subcat -] is an available category and
NP[subcat +] is not. In a standard CFG specification, this is achieved
via the provision of nonterminal symbols, not via filters. Similarly, it
seems to me for linear precedence principles and feature instantiation
principles.
In any case, the GPSG concepts refer to metagrammatical devices used in
specifying grammars, not ones used directly for language generation.
Yes, I see SB's parallel--both filters and the GPSG devices work to
constrain the rest of the grammar system they're part of, but techniques
are worth distinguishing. For one thing, the GPSG devices always
create CF grammars. I don't know of theoretical results on filtered
grammars. For another, I can imagine useful computational analogues
for FCR's etc. which might be used in compiling--as opposed to running
--grammars (in a parser). The analogue to a filter (abort) is clear, just
not particularly useful.
3. Filters and the Uses of Grammar
I think BN (Bruce Nevin) is right when he points out that the usefulness
of grammar depends on the sort of grammar, and that the use of filters
detracts from usefulness. I just objected to where he thought they
were used. Metagrammatical devices, on the other hand, needn't be attached
to any such disadvantages--for one thing, you can always just use them
and then put them away, making whatever use of the grammar you choose.
--John Nerbonne
nerbonne@hplabs.hp.com
------------------------------
Date: Tue, 8 Mar 88 12:25 EST
From: Rick Wojcik <ssc-vax!bcsaic!rwojcik@beaver.cs.washington.edu>
Subject: Re: Becoming CAI literate (ergative note)
gilbert@hci.hw.ac.uk (Gilbert Cockton) writes:
GC> Many Indo-European languages suffer from the pathological antinomy
GC> between subject and object. Either X effects Y, or Y effects X. Thus
GC> ...
GC> Basque has an 'ergative' case, which has been characterised as
GC> carrying the role of a fully co-operating, active object. In Basque,
GC> 'to teach' takes the ergative. In this sense of teaching, a computer
GC> could only teach a child IF it was capable of co-operative
GC> interaction. Watch a good classroom teacher and you will see that
I don't think that there is a real semantic difference between ergative and
accusative languages, as your note suggests. Both language types have subjects
and direct objects. The only real difference is in transitive sentences, where
an accusative language marks some relationship (e.g. verbal agreement) between
the verb and subject. An ergative language marks a relationship between the
verb and direct object. In fact, the pattern of an active transitive sentence
in an ergative language is similar to a passive sentence in an accusative
language. The superficial resemblance derives from the fact that ergative
sentence patterns often originate from passives historically. But this does
not mean that speakers of ergative languages have a different conception of
'activeness' of subjects and objects than do speakers of accusative languages.
--
Rick Wojcik csnet: rwojcik@boeing.com
uucp: {uw-june uw-beaver!ssc-vax}!bcsaic!rwojcik
address: P.O. Box 24346, MS 7L-64, Seattle, WA 98124-0346
phone: 206-865-3844
------------------------------
Date: Thu, 10 Mar 88 09:49 EST
From: Dori Wells <DWELLS@G.BBN.COM>
Subject: Language & Cognition Seminar
BBN Science Development Program
Language & Cognition Seminar Series
Parsing as Deduction: Using Knowledge of Language
Mark Johnson
Cognitive Science Center
MIT
BBN Laboratories Inc.
10 Moulton Street
Large Conference Room, 2nd Floor
10:30 a.m., Friday, March 18, 1988
Abstract: In this talk I show how natural language parsing can be viewed as
a specialized deductive process, and the parser itself as a highly
specialized theorem-prover. Knowledge of the language and phonetic
information about the utterance to be parsed function as "axioms",
the computation proceeds by the application of "inference rules", and
the output of the parser are "theorems" that following from its inputs.
I explain the "Parsing as Deduction" approach using a series of model
deductive parsers for Government and Binding Theory, all of which use
the same knowledge of language, but differ as to how they put this knowledge
to use. The knowledge of language used by these parsers is highly
modular, and refers to the four levels of representation of GB
theory, viz. D-structure, S-structure, PF and LF.
------------------------------
End of NL-KR Digest
*******************