LINGUIST List 11.2023

Sat Sep 23 2000

Review: Iwanska & Shapiro: NLP & Knowlege Rep.

Editor for this issue: Andrew Carnie <carnielinguistlist.org>


What follows is another discussion note contributed to our Book Discussion Forum. We expect these discussions to be informal and interactive; and the author of the book discussed is cordially invited to join in. If you are interested in leading a book discussion, look for books announced on LINGUIST as "available for discussion." (This means that the publisher has sent us a review copy.) Then contact Andrew Carnie at carnielinguistlist.org

Directory

  1. Philipp Strazny, Iwanska review

Message 1: Iwanska review

Date: Tue, 12 Sep 2000 22:17:41 -0500
From: Philipp Strazny <straznyyahoo.com>
Subject: Iwanska review

Iwanska, Lucja M., and Shapiro, Stuart C. (eds.) (2000) Natural Language 
Processing and Knowledge Representation. Cambridge: MIT Press. 459 pages,
$40 (amazon.com)


Reviewer: Philipp Strazny, University of Wisconsin-Madison

Synopsis:

This book is a collection of eleven papers on computational
linguistics, specifically on problems related with semantic
aspects. The papers are divided into two sections, the
first (papers 1 through 5) concerned with more formal
issues of semantic representation and inference, the second
(papers 6 through 11) with large-scale applications and
extensions for automatic acquisition of semantic
information. Both sections are introduced by an overview
article, in which the editors highlight what they perceive
to be the common theoretical claims.

The papers are followed by two appendices with introductions
to and definitions of propositional and predicate logic, set
theory, functions and relations, boolean algebra,
generalized quantifiers, etc. A third appendix provides
collections of natural language constructions that prove
problematic for NLP system design.

Here a brief description of the individual papers:

1. "Natural Language Is a Powerful Knowledge Representation
System: The UNO Model" by Lucja M. Iwanska.
 Describes a model (UNO) in which all semantic
information is encoded in a set-theoretic attribute-value
notation of the form
 x == { a, b, ..., n },
where x stands for a type to be defined and a, b, ..., n
stand for attributes with values. This notation is used for
both information storage and information processing, such as
inference.
2. "Natural Language Syntax and First Order Inference"by
David A. McAllester and Robert Givan.
 Presents a quantifier-free fragment of a Montagovian
logic that allows for efficient computation of inference.
3. "Issues in the Representation of Real Texts: The Design
of KRISP" by David D. McDonald.
 Describes KRISP, a semantic network of units that
express semantic objects or entities endowed with sets of
properties. Units have a similar structure to those used in
UNO, but attribute values are expressed as pointers to
objects in the network.
4. "Episodic Logic Meets Little Red Riding Hood - A
Comprehensive Natural Representation for Language
Understanding" by Lenhart K. Schubert and Chung Hee Hwang.
 A model (EPILOG) using quantified expressions to
implement ideas borrowed from event theory (Carlson 1982,
Chierchia 1985) and Discourse Representation Theory (Kamp
1982). The resulting "Episodic Logic" is used for tracking
information presented in narratives such as folk tales.
5. SNePS: A Logic for Natural Language Understanding and
Commonsense Reasoning" by Stuart C. Shapiro.
 Presents a logical notation (SNePS) to overcome
problems encountered by standard first-order predicate
logic.
6. "A Multi-Level Approach to Interlingual Machine
Translation: Defining the Interface between Representational
Languages" by Bonnie J. Dorr and Clare R. Voss.
 Presents a machine translation system (PRINCITRAN) that
maps expressions from one language to expressions from
another via a common semantic representation, modelled after
lexical-conceptual structure from Jackendoff (1983, 1990,
1991).
7. "Uniform Natural (Language) Spatio-Temporal Logic:
Reasoning about Absolute and Relative Space and Time" by
Lucja M. Iwanska.
 Describes an extension module for the UNO system,
designed to handle spatial and temporal information.
8. "Mixed Depth Representations for Dialog Processing" by
Susan W. McRoy, Syed S. Ali, and Susan M. Haller.
 Presents a model (B2) employing a network of phrase
structure rules, the elements of which encode discourse
(e.g. UTTERANCE), semantic (e.g. RELATION_IS), theta-
theoretic (e.g. AGENT) or syntactic information (e.g. NP).
9. "Enriching the WordNet Taxonomy with Contextual
Knowledge Acquired from Text" by Sanda M. Harabagiu and Dan
I. Moldovan.
 Proposes a path-finding algorithm for a lexical-
semantic network (WordNet) that can be used to extend
network links in response to input, thus implementing a form
of knowledge acquisition.
10. "Fully Automatic Acquisition of Taxonomic Knowledge
from Large Corpora of Texts: Limited-Syntax Knowledge
Representation System based on Natural Language" by Lucja
M. Iwanska, Naveen Mata, and Kellyn Kruger.
 Presents ad-hoc heuristics that can be used for
automatic information extraction from free text.
11. "A Computational Theory of Vocabulary Acquisition" by
William J. Rapaport and Karen Ehrlich.
 Presents a computational system using the SNePS logic
that draws contextual inferences to determine the meaning of
unknown words.

Critical Evaluation:

 The title of the book attracted me because it seems to
promise an integrated account of syntactic processing and
semantic representation: "Natural language processing"
(NLP) is often used as synonymous to (syntactic) parsing and
generating. However, I was disappointed when I found that
the focus here is really only on the representation of
semantic information in computational models and its
manipulation and acquisition, while parsing and generating
are mentioned only in passing.
 The promise of the title is reinforced in the subtitle
("Language for Knowledge and Knowledge for language") and
in the section overviews, where the editors propose that
"natural language is essentially the language of human
thought".
 A computational implementation to bolster this
philosophical claim would be very interesting. As a formal
linguist, however, I believe that the book falls short on
this - its most substantial - claim, due to two problems:
 The first problem is the assumption of a dichotomy
between first-order logic and natural language. On the basis
of this assumption any notation that reminds of surface
syntax can be presented as being qualitatively different
from first-order logic. However, one can think of surface
syntax and first-order logic as representing the extreme
points of a continuum with many potential notations falling
in between:

 SS < LCS < FOPL

(SS = surface syntax, LCS = lexical-conceptual structure,
FOPL = first-order predicate logic)

One may well argue about whether or not one particular
system is closer to surface syntax than another, but all
notations - including FOPL - are certainly "informed by
natural language".
 The editors believe that the attempt to integrate
aspects of modern linguistic theory into computational
theories of language represents "an exciting and, still,
quite controversial new research direction" (xiii).
However, one should not forget that classical logic was the
first attempt of assigning an underlying structure to
natural language expressions (this is, of course, a
reinterpretation in modern terms), and the philosophy of
language always remained informed by surface natural
language phenomena. Thus, it is not necessary to create a
bridge between logic and "natural language". One only
needs to decide how far one would like to proceed on the
existing one.
 Even within the tradition of artificial intelligence in
general and computational linguistics in particular, it is
easy to find precedents for the purportedly "new" interest
in natural language phenomena. To name just the major
figures, the inventors of the Turing machine and the
Chomsky-hierarchy were certainly not unaware of or
uninterested in the intricacies of surface natural language.
 If one thus accepts that all logical notations -
including first-order predicate logic - are informed by
natural language and differ only insofar as some may be
somewhat closer to surface syntax than others, it is hard to
maintain that a theoretical choice between them can
represent a qualitative step towards deciphering "the
language of thought".
 The second problem is that the system architects do not
follow through with the goal expressed by the editors to
have "all information and knowledge processing ... take
place at the natural language level" (xiv). The inference
heuristics for spatial and temporal expressions presented in
chapter 7 (the UNO model) depend on translating them first
into numerical values (265), i.e. into a decidedly
nonlinguistic format. Like UNO, the EPILOG model utilizes
"specialists", i.e. modules with predefined (sometimes
numerical) heuristics (155), designed to efficiently handle
certain types of information. KRISP uses predefined date
formats that are presumably handled by hard-coded heuristic
functions (93). The point here is not that such heuristics
are illegal, but that space, time and possibly other domains
of knowledge are handled with nonlinguistic algorithms. This
may well be unavoidable and find justification in our
cognitive makeup, but by using "specialists" the system
architecture transcends the domain of natural language per
se, contrary to the editor's claim.
 Setting aside the hyperbole about the "newness" of
the approach, the book represents a positive change in
attitude of computational linguistics towards formal
linguistics. The two disciplines are, of course, closely
related, and they should be mutually informative, but the
communication between them has been suboptimal in recent
years. To return the favor, formal linguists should keep
computability and/or efficiency in mind when they draw up
their theoretical models. A question one may then ask is
whether this particular book is a good entry point to
computational issues for formal linguists. Unfortunately, I
have to say that I don't think so.
 It is not clear to me who the target audience for this
book is. Computational linguists will presumably already be
familiar enough with the primary literature that they do not
need this sort of overview. The appendices appear to be
aimed at readers without a background in semantics. However,
few of such readers will be able to digest this thoroughly
semantic (and formal) treatment of language. Semanticists
may find it interesting to see some important theories "in
action" and to have their representations tested and
adjusted for efficiency, but they should not expect too much
feedback. Syntacticians will be disappointed by the lack of
detail concerning parsing and generating and puzzled by
statements like "we use a parser with a linguistically-
based grammar" (284). Linguists familiar with typological
diversity in word order will wonder how one can make claims
like "the system that comes closest to natural language as
knowledge representation and reasoning is ...Montagovian
syntax for first order logic"(3) just because of an
accidental similarity between a logical expression and the
word order in simple subject-verb-object sentences in
English (70). Readers in general may wonder why they should
read through the details of UNO, KRISP, and EPILOG, when
they overlap in many theoretical choices. I think it would
have been better to pick one system and explore the details
all the way from the parsing/generating front-end to the
semantic knowledge base.
 I should stress that most of the individual papers are
well written, highly informative and theoretically quite
sophisticated. Readers who read the book as a collection of
papers may thus find it stimulating. However, the editors
try hard to make it appear as a coherent work, which
justifies the above criticism leveled against the book as a
whole. I would have preferred if the editors had played a
much more active role by molding the different papers into a
single text with clearly demarcated topic sections.


References

Carlson, G.N. 1982. Generic Terms and Generic Sentences.
Journal of Philosophical Logic 11(2): 145-181.

Chierchia, G. 1985. Formal Semantics and the Grammar of
Predication. Linguistic Inquiry 16(3):417-443

Jackendoff, R. 1983. Semantics and Cognition. Cambridge,
Mass.: MIT Press.

Jackendoff, R. 1990. Semantic Structures. Cambridge, Mass.:
MIT Press.

Jackendoff, R. 1991. Parts and Boundaries. In Lexical and
Conceptual Semantics, eds. B. Levin and S. Pinker.
Cambridge, Mass.: Blackwell.

Kamp, H. 1981. A Theory of Truth and Semantic
Representation. In Formal Methods in the Study of Language,
eds. J. Groenendijk, T. Janssen, and M. Stokhof, 277-320.
Amsterdam, The Netherlands: University of Amsterdam.

About the reviewer

Philipp Strazny is currently working on a Ph.D. dissertation
in Linguistics at the University of Wisconsin-Madison. The
dissertation project involves a natural-language interface
for computer applications that aims to integrate formal
linguistic and practical computational requirements with
neurobiologically plausible data structures.







Philipp Strazny
4025 Winnemac Ave
Madison, WI 53711
USA

Tel./Fax: 001-(608) 231-9678
straznyyahoo.com 
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue