Burghard B. Rieger
From Computational Linguistics to Computing with
Words.
For the past decades, the concept of
symbolic representation together with the computer metaphor appeared to offer
an adequate framework to deal with cognitive processes scientifically. Formally
grounded by logical calculi and implemented as algorithms
operating on representational structures,
cognition is considered a form of information processing in the cognitive
sciences. Thus, computational linguistics (CL) as part of cognitive
theory identified the complex of language understanding as a modular system
of subsystems of information processing which could be modeled
accordingly. The alliance of logics and
linguistics, mediated mainly by (language) philosophy in the past, and by (discrete)
mathematics since the first half of the last century, has long been (and partly
still is) dominating in what way and terms natural languages and their
functioning should be explicated and how their processing could be modeled. In
replicating (and in parts also
supplementing) semiotically motivated strata of systematic sign description and
analysis, different levels of modular aggregation of information – external
and/or internal to a processing system – have been distinguished in cognitive
models of language understanding. They partly correspond to and partly cut
across the syntactics-semantics-pragmatics distinction in the semiotic relatedness of signs,
the utterance-discourse-corpus
levels of performative language
analysis, and the hierarchy of morpho-phonological,
syntax-sentencial and lexico-semantic
descriptions in structural
models of linguistics. It is ironic, however, that the dramatic increase of
computational power and symbol manipulation means has changed the fundamentals
of many scientific disciplines, creating even new ones, but has left
linguistically oriented disciplines, even new ones, adhere to the lore of
seemingly well grounded and traditionally dignified concepts in describing
natural language structures and their functions.
1. Cognition
Cognitive approaches tend to model
mind/brain processes based upon the evaluation of (in parts linguistic) data
generated in more or less sophisticated experiments of human thought/thinking
and understanding. For a linguist, however, more immediate results of
cognitively most relevant mind/brain activities – not the experimentally reduced segments
of them – come to mind as being easily accessible in form and structure of natural language
discourse which is abundantly available now in machine readable form. Other
than what premises of theoretical linguistics and main stream computational
linguistics suggest, some processes of language understanding require and might
well be modeled along observable but as yet unexploited traces of meaning constitution
in natural language text corpora which speakers/writers and hearers/readers
have enacted in situations of communicative language use[1].
This enactment is tied to representational or semiotic functions, based upon
regularities of entity usages[2] which not only generate observable
structures, but also serve in turn to allow these functions being activated to
modify these structures simultaneously. Thus, the complexities and dynamics of
natural languages themselves may be taken as a salient paradigm for information
granulation both, in its fuzzy as well as crisp modes of
structural representation and functional processing. It appears that the
conception of fuzzy and/or crisp granularity – once the process-result
ambiguity and its cognitive-linguistic ambivalence is solved as addressed by a facet of the symbol-matter problem
[Pattee89] – lends itself easily to a unifying view of how natural language
understanding or meaning constitution may be arrived at as a
computational process on structural entities adequately identified.
2.
Computational Linguistics
The
way structural linguists used to and still categorize (segment and classify) observable natural language
phenomena as tokens (like phones, morphs, utterances, etc.) to
constitute abstract linguistic entities as types (like phonemes,
morphemes, phrase, etc.) can be shown to be based on the very processes of granular meaning
constitution, however imperfectly. What may procedurally be derived either as soft
linguistic categories or fuzzy granules
represented by vectors, distributions, or
fuzzy sets for
(numerical) computation, has so far been (over-) generalized and abstracted to
form crisp categories represented by signs (symbols) for string
manipulation. Certainly, linear aggregation of these symbols serve to
understand and control one type of observable natural language phenomena as
part of aggregational string
formation or formal grammar. Its core concepts of
well-formedness (syntax) and truth-function (semantics) were made
explicit by way of specifying conditions of formal correctness and derivational compositionality.
Their symbolic representations in the form of productions or rewrite rules
– allowing for recursive application and generative string formation – not only
constituted a wealth of symbol aggregation systems (formal languages)
but were also employed to model comparable properties (in processes) of natural
language string formation.
However,
whereas formal rules whose application would allow to specify generative
properties and truth-functional constraints in formal constructs like sentence
and proposition, other properties of natural language expressions which
are communicationally more relevant, like e.g. making sense by having
specific meaning in situational contexts which are to be understood,
tended to be abstracted away. The process of understanding natural language
discourse below and above formal sentence reconstruction is, by and large,
still in
want
of principled analysis, formal representation, computational simulation, and
procedural realization. This is, were CL is to head in near future in order to
assist people not only in producing huge amounts of texts, but also in skim the
of verbal information worldwide without the need to read the texts concerned.
In
order to follow this line of structural, functional, simulative, and enactive
modeling of machine understanding, traits of traditional approaches to
cognition and natural language processing as forwarded by linguistics proper (LP), computational linguistics (CL), and language
processing in artificial
intelligence (AI) research will be reviewed in order to identify points of
departure from which to advance our understanding of how natural languages
function the way they do. Some presuppositions will have to be revised to
understand how communicative employment of languages is not only based upon (use)
but also establishes (usage) structural constraints which may be made
explicit by assumptions motivated by systems theory and/or by empirically
testable hypotheses derived thereof
[Rieger01]. It will be argued that the introspective assessment and
judgment of any speakers' own language faculty on linguistic functions and the
correctness of singular sentences or phrasal structures as conceived by an ideal
speaker's/hearer's internal language (IL) is not at all sufficient, let alone
superior to modern means of empirical investigation of masses of
natural
language discourse or external language (EL) [3]
as being produced by real speakers/writers in situations of
intended/successful communication. Instead, some of the inadequacies of CL models of natural
language processing that competence oriented linguistics have inspired
so far, will hopefully be revealed to be due to unwarranted abstractions from
relevant characteristics (e.g. contextuality, vagueness, variability,
adaptivity, openness etc., to name only the most salient) of processes of
natural language communication. Other than these idealizations which
purportedly allow immediate access to
cognitively relevant entities, we shall argue for an empirically
controlled understanding of functional sign constitution which does not readily
abstract from the emergent structures which models of a more semiotic cognitive information processing (SCIP) may
bring about. It is hoped to collect and produce some evidence that the traces
of such processing can not only be identified, but that these identification
procedures may also be employed to systematically (re)construct fuzzy
information granulation procedurally.
3.
Computing with Words
The
notion of computing with words
(CW) hinges crucially on the employment of natural language expressions. These
are considered to provide not only the representational structures of what can
semantically be meant but also the operational means of what can cognitively be
understood by processing these structures. They allow for decomposition of
wholes into their constituents or parts (granulation), or conversely,
for composition and integration of parts into wholes (organization), and
for the association of signs with meaningss (causation).
According
to Zadeh's early introduction of the notion of
granularity [Zadeh79] and his recent elaboration of that concept
[Zadeh97] as theory of fuzzy
information granulation (TFIG),
human cognition may be understood as based upon and structured by
processes of granulation, organization and causation. These are
meant to specify different types of mind/brain activities which can be
characterized as being computational in nature and hence to be modeled
mathematically and/or procedurally. Although this characterization suggests
different modes of these mind/brain activities to be distinguished sharply
both, as the enactment of processes and as the results which these processes
produce – a distinction that will have to be drawn and obeyed more clearly –
there is the need for yet another discrimination to be made in order to clarify
what CL will have to deal with in future. It is tied to the first one and
concerns the way mind/brain activities are made accessible by techniques,
models, or disciplines in different ways.
In fuzzy
linguistic (FL) models of computational semiotics (CS), the
situatedness of natural language communication is considered conditional. This
requirement is met by corpora of pragmatically
homogeneous texts (PHT) which assemble language material which is
situationally constrained by a number of variables (like e.g. communicative
media or domain,
register,
topic, author, etc.) whose values (like newspaper, report, economy, etc.)
define the PHT profile the incorporated texts will satisfy. There is good
reason to assume that such PHT
collections
realize the organized structure of natural language discourse, i.e.
integrating parts into wholes (sign formation), as well as the causative functions, i.e. associating
causes with effects (meaning constitution). The inherent structuredness of a
PHT corpus gives rise to the
multi-resolutional
representations of meaning functions which may be explored in order to model
(crisp and fuzzy) SCIP granules.
Based upon the empirically well founded observation
and
rigorous mathematical description of universal regularities in natural language
discourse, these regularities can be shown to structure and constitute
(different levels of) processes and/or their representational results when made
to operate on pragmatically homogeneous texts of either performed or intended
communicative interaction in actual situations. Only such a performance
oriented semiotic approach will give a chance to formally reconstruct and
model procedurally both, the significance of entities and the meanings
of signs as a function of a
first and second order semiotic embedding relation of situations (or contexts) and of language games (or cotexts) which corresponds to the
two-level actualization of cognitive processes in language understanding
[Rieger02]
.
References
Pattee,
H. (1989): Simulations, Realizations, and Theories of Life. In: Langton (ed): Artificial
Life, [SFI Studies in the Science of Complexity VI], Reading, MA (Addison
Wesley), pp. 63-77.
Rieger, B. B. (2001):
Computing Granular Word Meanings. A fuzzy linguistic approach in
Computational Semiotics. In: Wang (ed): Computing with Words, [Wiley
Series on Intelligent Systems 3]. New York (J. Wiley & Sons), pp.147-208.
Rieger, B. B. (2002): Semiotic Cognitive Information
Processing: Learning to Understand Discourse. A systemic model of meaning
constitution. In: Stamatescu et.al. (eds): Perspectives on Adaptivity and
Learning. Heidelberg/ Berlin/ New York (Springer), [in print].
Zadeh, L. A. (1979): Fuzzy Sets and Information
Granularity. In: Gupta/Ragade/Yager (eds):
Advances in Fuzzy Set Theory and its
Application. Amsterdam (North Holland), pp. 3-18.
Zadeh, L. A. (1997): Toward a Theory of Fuzzy
Information Granulation and ist Centrality in Human Reasoning and Fuzzy Logic. Fuzzy
Sets and Systems, 90(3): 111-127.
[1] In the sense of employment and act of
being used under certain conditions and to a particular communicative purpose.
[2] In the sense of customary practice and manner of using which may establish and modify rules and standards.
[3] Borrowing the term not “to refer to any other notion of language … never characterized in any coherent way” (Chomsky, personal communication), but understood – and diverging from Chomsky – to cover all phenomena of observable language performance.