By Prof. Gerard Sabah
Langage et Cognition Group
LIMSI - CNRS
BP 133
91403 Orsay cedex
FRANCE
Published in the new journal of "Cognitive Processing".
Hausser, Roland 1999, Foundations of Computational Linguistics,
Springer-Verlag, Berlin.
Roland Hausser's book addresses the mechanisms of Natural Language
communication with speaking robots of the future. Although it focuses
on language, it also tries to take into account a few multi-modal
aspects of communication . The book has a very strict organisation :
four parts (language theory, grammar theory, morphology and syntax,
semantics and pragmatics), each made up of six chapters, each of
which, in turn, contains five sections.
An Introduction first presents the "SLIM" linguistic theory (meaning
Surface compositional Linear Internal Matching), stating its four
basic principles: surface compositionality (methodological principle),
derivational order's strict linearity relative to time (empirical
principle), utterance interpretation and production analysed as
cognitive processes (ontological principle), reference modelled in
terms of matching an utterance's meaning with context (functional
principle).
The first part (language theory) develops a view of language
illustrated by the description of a robot (Curious by name). This
theory (SLIM) relies upon cognitive semantic primitives, a sign
theory, along with their functional integration within generative and
interpretative processes.
More precisely, Chapter 1 immediately positions the topic within the
Human-Computer Interaction (HCI) paradigm, viewing computational
linguistics as basically pluri-disciplinary. The author distinguishes
two branches of AI (the classic, formal one, and the robotics-based
one, the difference being that, for the latter, the environment
undergoes perpetual, unpredictable changes). Two kinds of Artificial
Intelligence yield two kinds of HCI, to which he adds a third case of
communication, within the Virtual Reality framework. Following the
classic distinction between various levels (phonology, morphology,
lexicon, syntax, semantics and pragmatics), the author, who somehow
mixes computer processing and Internet aspects, presents the
innovative features of present-day publications (from CD-ROM to SGML).
Taking as examples the search in indexing mechanisms in data bases,
along with automatic translation (presently limited to on-line help,
crude translation or restricted language), Chapter 2 presents the
essentials of Natural Language Processing, and underlines phenomena
which call for linguistic solutions at the various aforementioned
levels, reminding the reader about the alternative between "smart" and
"solid" solutions (which reminds us of Tomita's distinction between
"interesting" and "useful" problems).Footnote
Chapter 3, starting from the connections between understanding and
communication on the one hand, and between understanding and
perception on the other, leads to a proposal for an "iconic"
representation of objects, which, although it is somewhat crude, seems
to be efficient in the current field. However, I found that the role
of lexical categories is treated in an over-simplified manner, and the
limits of this kind of representation could have been more seriously
explored (raises such questions as: can all sorts of objects be
described in this way? How does one deal with more or less abstract
entities?).
The reference phenomenon is then considered, using a mechanism of best
match between representations of known objects in the world and
representation from linguistic descriptions. This means that one
elaborates on relations between literal meaning, compositionality and
pragmatics aspects. Here, other language theories are presented,
(mainly Chomsky's and Grice's), but the author's own theory is also
compared with the traditional semantic theories: those of Frege,
Pierce' Saussure, Bühler, and of Shannon and Weaver, detailing their
formal and methodological foundations, and showing their connections
with the proposed one. This presentation remains very
"reference-oriented", and the study of conditions for success in HCI
could have gone deeper...
Project "CURIOUS" (a virtual robot which continually observes and
analyses its changing environment), presented in Chapter 4, relies
upon the use of a formal language, whose propositions can match states
of the world one-to-one. It can be noted here that such a language
does not necessarily exist, and we regret that so crucial a point is
not more thoroughly examined within a cognitive vision on language. As
to the various features of CURIOUS, we also wonder if they can be
extended to more complex worlds; a positive answer is in no way
certain. Nonetheless, the projects seem to be a good validation for
the proposed theory.
While dealing, in a rather cursive way, with Information Theory,
Chapter 5 puts more emphasis on pragmatic aspects (here seen as a
search for connections between an interpretation and a coherent
sub-context), making some fundamental principles of the domain
explicit.
Chapter 6 is of a more semiotic tendency. It comes back to the
question of reference and delivers another set of fundamental
principles of pragmatics, via some aspects of pronominal reference,
and of relations between icons and symbols.
The second part (grammar theory) is about the theory of formal
grammars and its methodological, mathematical and computational role
in the description of natural languages. Chapter 7 states empirical
and mathematical problems in the description of languages, as they
arise from questions of substitution and continuation, by developing
basic notions of generative and categorial grammars. As a tentative
remedy, the author proposes Left-Associative Grammars (LAG), which
means grammars with a left-associative derivation order, modelling the
time-linear essence of natural languages. Here, even if the rest of
the book argues well for this choice, I find that it could have been
provided with better credentials. Moreover, definitions (even short
ones) of grammars should be proposed in the first lines of the
chapter, which would make understanding easier for beginners.
The essential distinction is following: for generative and categorial
grammars, derivations rely upon the concept of substitutions, while in
LAGs, they depend on possible continuations. Chapter 8 compares the
respective complexities of those grammars, using examples from
artificial languages. Chapter 9 develops basic notions of syntactic
parsing, and its relations with the current grammars. Chapter 10 gives
a more formal view of LAGs and illustrates them with examples from
artificial languages, along with some English ones (mainly about
discontinuous components and non-grammatical input).
The next two chapters detail the LAG hierarchy: A-LAG (covering all
recursive languages), B-LAG (all context-sensitive languages), C-LAG
(all regular non contextual languages and some context-sensitive ones)
- the latter being, in their turn, divided in three subclasses. In
this hierarchy, natural languages are in a low-complexity class,
allowing for parsing in linear time. Using considerations about
ambiguities and their respective complexity, this hierarchy is
compared with that of phrase structure grammar, and the author shows
that they are non-equivalent, being indeed "orthogonal", and that a
LAG yields "better" parses of ambiguities (from my point of view, it
is rather a case of better representations of ambiguities, not
"parses", but this is open to discussion).
The third part (morphology and syntax), does focus on morphologic and
syntactic aspects of natural languages. It uses many examples in
English and German, and elaborate comparisons between them both
(particularly on German's free order as opposed to English's
constrained one), to show possible morphological processes, along with
grammars one can think of implementing. Starting from small examples
(inflexions, neologisms, allomorphs in English and German), the author
systematically augments them, showing more and more intricate phrases,
and derivations that are applied in each case.
This chapter argues for the possibility of using a unified formalism
for morphology, lexicon and syntax within the LAG framework. The
author could perhaps have developed a comprehensive comparison with
more traditional approaches (particularly Winograd-like procedural
parsing, or the declarative method illustrated by Pitrat - the storage
mechanism having, by the way, a close relationship with the latter
one).
Chapter 15 is about a possible use of that morphological analysis in
corpus processing and the connected distributional analyses; Chapter
16 deals with the basic concepts of parsing (valence, agreement and
word order), and to these traditional ones, it adds the higher
left-association principle. This is followed by a detailed analysis of
constraints on word order, freer in German, fixed in English (I for
one would haxe exchanged both paragraphs, as I find English more
understandable).
Chapter 17 provides us with various concrete examples of English
processing, and Chapter 18 does the same for German, comparing the
examples to the English ones. Again, I felt some inconvenience due to
the order of exposition: Paragraph 18.1 presenting general thoughts
about standard parsing processes would, in my opinion, better find its
place in Chapter 16, as well as Paragraph 18.3, developing elaborate
and interesting parallels between English and German verb
positions. This minor flaw in the text order notwithstanding, these
various elements are excellent illustrations of the observed
phenomena.
Finally, the fourth part (semantics and pragmatics) deals (as you
would expect) with semantics and pragmatics. Chapter 19 develops the
fundamental contrasts between three kinds of semantics: of formal
logic languages, of programming languages and
of natural languages. It explores at some depth a possibility of applying logic semantics to natural languages. Intensional context, propositional attitudes and the vagueness phenomenon make it visible that different semantics rely upon different ontologies (which is illustrated by various classical paradoxes, stated in Chapter 20, and by the difference between absolute and contingent propositions in Chapter 21).
The author then delivers a rather philosophical consideration of how a
semantic interpretation in natural languages leads to increasing
complexity, and tells us how this can be avoided in the SLIM
theory. The last two chapters show detailed examples of
representations at various levels of the "SLIM machine", respectively
in the receiving and in the generating stance. Here again, the
approach is well illustrated.
Viewed from the realm of Cognitive Science, this book, which aims to
span many disciplines, fails at it in some significant
regards. Nothing is said in terms of neurobiology of language, nothing
again on learning, on linguistic knowledge acquisition, and very
little on "error processing" (non-grammatical input, typos, syntax
errors) which does represent a fundamental aspect of HCI. Still, we
have here quite a substantive book, whose short chapters make for
pleasant reading and come complete with relevant exercises. As it
totals 534 pages in its present state, we can hardly blame it for not
taking every topic into account! Despite our minor criticisms, stated
in the above paragraphs, this book is a rich source of learning. It
clearly describes the basic elements of automatic natural language
processing, along with fundamental issues of the domain. On such
bases, it proposes an original and interesting grammatical theory, and
shows possible applications of it.
The mere length of the present analysis (even if it is somewhat critical at places) speaks in itself for the interest I felt in reading this book!
Gérard Sabah
----------------------
Footnote: The former have no obvious solution, call for the
development of sophisticated theories, or make it possible to exhibit
general linguistic principles, and can be treated by a small number of
general rules (e. g. relativisation, causation, group movement). In
contrast, "useful" problems are those for which obvious solutions
exist, or those for which no general principle can be exhibited; one
just has to add specific rules to solve each one of those problems
(punctuation, expression of dates, idioms). The latter kind of problem
is difficult to anticipate, and requires very large amounts of
knowledge. Note that it is just as essential to solve both kinds of
problems if one wants running applications! However, if one is mainly
interested in theoretical aspects, it will be possible to neglect
problems of the second kind, for which simple solutions potentially
exist, but would be costly to implement. Moreover, the attitude of
some linguists, who are constructing a complete catalogue of all
figures of discourse (Gross, Mel'cuk) implies that all problems are of
the second kind.