Linguistics As1
-
Upload
zoltanparades -
Category
Documents
-
view
226 -
download
0
description
Transcript of Linguistics As1
-
How can we dene the meaning of a word?
October 22, 2014
1 Introduction
Despite the fact that the mental lexicon appears to be organized by meaning (at
least partially) [Marslen-Wilson, 1994, Meyer and Schvaneveldt, 1971], there is
no consensus on what word meaning... means [Braun and Sider, 2007]. Many
dierent, pairwise incompatible theories of meaning were proposed as solutions
during the last century.
To provide a basis for comparing these dierent theories, I will begin with
an overview of the explanatory work a theory of meaning is expected to do.
After establishing the comparison criteria, I will present three competing
theories (the classical, prototype and exemplar theories) by summarizing their
claims and evaluating their explanatory performance.
In the conclusion, I will single out exemplar theory as the most promising
candidate, and attempt to identify possible future directions.
2 Comparison criteria
The quality of a scientic theory is a function of its explanatory power. This
section provides a short review of several inuential linguistics studies. In the
1
-
sequel, I will compare the competing theories by their ability to account for the
phenomena revealed by these experimental results.
For detailed explanations of the experiments, the reader is referred to the
respective original papers.
2.1 Typicality
In 1973, Rips et al.[Rips and Smith, 1973] demonstrated the existence of a typ-
icality gradient: people classied some species of birds as more typical than
others, and this typicality was shown to correlate with the verication time of
sentences involving these species. The usual example: robins are rated more
typical than eagles, and verifying the truth of Aneagleisabird takes more time
than that of A robin is a bird .
The results have been replicated for various categories including clothing,
vegetables and furniture, as well as dierent sentence processing tasks [Rosch.E., 1975].
Somewhat surprisingly, even abstract number-theoretic categories (e.g. prime
numbers) display typicality eects
[Armstrong and Gleitman, 1983].
2.2 Fuzzy Boundaries
Some categories have fuzzy boundaries. On the level of a single individual,
one thinks of where a cup ends or a bowl begins (which may depend on its
contents!) [Labov, 1973], on the interpersonal level, more pressing issues, e.g.
the controversial boundary between fetus and human being in the abortion
debates.
Cognitive anthropoligsts were not the rst to notice the issue of fuzzy degree
of membership (despite what some sources claim! [CogLinWiki, 2014]). Indeed,
while Aristotle armed the existence of essential properties, he regarded mean-
2
-
ing as inherently fuzzy and ambigous, at least on the interpersonal level; so have
most philosophers since then [Tierney, 2002, Peirce, 1902].
The most important study is due to Labov [Labov, 1973]: no set of attributes
(in their case, opposing sentences) is sucient to unambigously separate bowls
from cups. However, a careful analysis of the data led to the creation of an
accurate linear classier
1
.
2.3 Knowledge Eect
Background knowledge also aects categorization, as well as determination of
meaning. Best exemplied by the recenty study of Lin et al. [Lin and Murphy, 1997]:
the same items in the synthetic category tut get wildly dierent typicality scores
depending on whether tuts are introduced as agricultural or venatic tools.
2.4 Mental Lexicon
An important goal for a theory of meaning (at least w.r.t. linguistics) is ex-
plaining the organization of the mental lexicon. Several studies indicate that
the mental lexicon is organized semantically. The most important experimental
idea of this eld is priming: the observation that detecting or recalling a stim-
ulus biases the processing speed of subsequent stimuli [Marslen-Wilson, 1994,
Meyer and Schvaneveldt, 1971]. An example, certainly semantic: dog is a prime
for wolf [Matsukawa, 2005]. The verication time eects in typicality are also
instances of priming.
Another phenomenon that requires explanation is that humans can acquire
new vocabularly, and that newly learned words take their place among the
others. Related to that, one should also mention compositionality: the common
1
attributes are assigned weights, the sum of all weights (hence linearity) is proportional to
predicted degree of membership
3
-
sense observation that the meaning of a compound depends on the meaning of
its constitutents[Kamp and Partee, 1995].
3 Theories of Meaning
3.1 The Classical (Aristotelian) Theory
3.1.1 Description
A misnomer, since the theory is neither classical nor related to Aristotle. In its
clearest formulation, given by Labov
2
[Labov, 1973], the classical theory holds
that meaning is
discrete: concepts are clearly separated from each other by formal or functional
properties,
invariant: all occurrences of a concept are the same,
conjunctively-dened: all essential properties must be present together for a
concept to be recognized,
composed-of-atomic-primitives: categories cannot be subdivided indenitely;
i.e. every concept satises a unique checklist of criteria.
A typical example would be bachelor, supposedly satisfying the unique checklist
{unmarried,male}. Which is, of course, nonsense: what about a four yearold boy, or the Pope? Wittgenstein once remarked that no-one ever produced
a working example of a denitional analysis, and not for the lack of trying
[Wittgenstein, 1953]!
2
amusingly in the form of a classical checklist!
4
-
3.1.2 Evaluation
The classical theory has little to say about typicality. By the way, this is not
very surprising, since refuting this theory was an overarching goal of Rosch's
typicality studies [Rosch.E., 1975]. Indeed, if verifying X is a bird involves a
unique checklist, then it should take the same time for any bird X, but this
is not what happens [Rips and Smith, 1973]. This also shows that the mental
lexicon is not organized according to classical meaning. Needless to say, unique
checklists also oppose fuzzy membership.
What about compositionality? One observes that adjectives don't commute
with hyponyms: hamsters are animals, but large hamsters are usually not
large animals. This is at odds with a unique checklist for the word large3.
3.2 Prototype Theory
3.2.1 Description
In the face of the numerous unexplained experimental phenomena, classical
theory becomes untenable. Rosch's study provides the foundation for the most
popular family of alternatives, the so-called prototype theories.
The dening
4
feature is the presence of prototypes, i.e. summary representa-
tions of concepts. The common variant (identied by Hampton [Hampton, 1979],
and clearly inspired byWittgenstein's idea of family resemblances [Kamp and Partee, 1995])
holds that a prototype is akin to a list of related features, weighted by impor-
tance. Unlike the classical checklists, these feature lists don't care about dis-
creteness, conjunctiveness or even consistency - e.g. a prototype for the word
dog will include the attributes short hair and long hair, as well as beige
3
Some advocates of the classical theory react by claiming that large is polysemous. Sincepositing enough polysemy can account for any empirical observation, one would prefer a
classical explanation for why this systematic polysemy occurs in the rst place.
4
Another essentialist notion! Should I say prototypical?
5
-
and black, presumably with dierent weights!
3.2.2 Evaluation
Prototype theory is consistent with typicality
5
eects: an instance is more typ-
ical than another if it has a higher weight of prototypical features. It is still
unclear why X isabird is easier to verify for typical birds than for atypical ones,
but at least one does not see an outright contradiction.
The situation is even better with fuzzy boundaries. The weighted lists pro-
vide a very natural framework for that. Armstrong observes that words with def-
initional meanings (e.g. odd) still display typicality eects [Armstrong and Gleitman, 1983].
However, this causes no actual problems, as long as one distinguishes typicality
6
from fuzzyness
7
. After all, an eagle, while atypical, is still unambigously a bird!
On the other hand, it seems that prototype theory cannot account for the
semantic structure of the mental lexicon. Compositionality is the usual culprit.
Armstrong observes that if you combine, say foolish and bird into the phrase
foolish bird, it is no longer a xed matter... which foolish elements and which
bird elements are intended to be combined ([Armstrong and Gleitman, 1983]
p. 272), and there are too many phrasal categories to be mentally represented in
terms of prototypes [Armstrong and Gleitman, 1983, Kamp and Partee, 1995].
Barsalou's experiment is denitive: people can make up categories on-the-y
(e.g. clothing to take on a vacation), and even these ad-hoc categories, not
described by a single word, produce typicality eects [Barsalou, 1983]. Such
categories could not have been stable mental representations.
5
its raison d'tre
6
goodness of exemplar
7
degree of membership
6
-
3.3 Exemplar Theory
3.3.1 Description
Another reaction to the failures of the classical theory, this one due to Medin
[Medin and Schaer, 1978]. The exemplar theory rejects summary representa-
tions, opting instead for memories, e.g. the concept of dog being represented
by the set of all remembered dogs. These memories are supposed to provide a
basis for classication whenever a new object is encountered.
Exemplar theory is a radical departure from the previous theories. On one
hand, exemplars avoid the drawbacks of the classical and prototype theories, on
the other hand, exemplar theory faces its own new, unique challanges.
3.3.2 Evaluation
Exemplar theory is the natural setting for typicality phenomena. Typical items
are the ones that are the most similar to the remembered category members.
This explains why e.g. a robin is a more representative bird than an eagle or
an ostrich. Typical items are easier to categorize because they are similar to
more category members [Medin and Schaer, 1978].
Exemplars convincingly account for fuzzy boundaries: a newly encountered
object can easily be similar to two or more remembered categories (e.g. both
tocups and to bowls).
Mental lexicon eects work so-so: priming is relegated to the realm of general
memory, and one gets an eective
8
explanation for compositionality.
The main diculty is with explaining knowledge eects: the study on tuts
comes to mind [Lin and Murphy, 1997]. How can the typicality score of certain
items be decided before any exemplars are encountered? Here, one reaches the
explanatory limits of the theory [Nosofsky, 2011].
8
if not unproblematic; can there really be a set of remembered foolish things?
7
-
Theory Typicality Fuzzy Boundaries Knowledge Eect Mental Lexicon eects
Classical + Prototype + + + Exemplar + + +
Table 1: Comparison of the theories of meaning
4 Conclusion
As Table 1 shows, there is no denitive end to the meaning debate. At least
the failure of the classical model should be clear. Its successors, the prototype
and exemplar theories, are much more powerful, but they still have signicant
explanatory blind spots.
The question of the mental lexicon is fundamental, yet Barsalou's work shows
that prototype theory will not be able to account for that [Barsalou, 1983].
The strongest candidate is therefore exemplar theory, which at least doesn't
contradict the experimental evidence [Nosofsky, 2011]. One could hope for a
further generalized exemplar theory, which solves the remaining issues. But are
there new paradigms emerging?
There's a lot of future directions to explore. Exemplar theory presumably
needs to be integrated with theories of memory formation, which should provide
evidence for the soundness of the concept. The case for prototypes is strong as
well: while pure prototype theory cannot work, a pluralist approach could still
maintain the word meaning has both prototype and exemplar features.
Either way, we're far from done: while we have come a long way since the
days of the classical, a truly satisfactory theory should correctly predict the
results of these experiments, but it should also speak about several phenomena
that we couldn't take into account
9
in this comparison: basic level categories,
superordinates, induction, continuous valuations etc. Until we get there, the
nature of meaning remains a mystery.
9
mostly because the present theories are agnostic or silent on these issues
8
-
References
[Armstrong and Gleitman, 1983] Armstrong, S. L., G. L. R. and Gleitman, H.
(1983). What some concepts might not be. Cognition, 13:263308.
[Barsalou, 1983] Barsalou, L. (1983). Ad hoc categories. Memory & Cognition,
11:211227.
[Braun and Sider, 2007] Braun, D. and Sider, T. (2007). Vague, so untrue.
Nous, 41:133156.
[CogLinWiki, 2014] CogLinWiki (2014). Aristotelian model of categorization.
retrieved 17 Oct 2014.
[Hampton, 1979] Hampton, J. A. (1979). Polymorphous concepts in semantic
memory. Journal of Verbal Learning and Verbal Behavior, 18:441461.
[Kamp and Partee, 1995] Kamp, H. and Partee, B. (1995). Prototype theory
and compositionality. Cognition, 57(129-191).
[Labov, 1973] Labov, W. (1973). J.N. Bailey and R. W. Shuy (eds.) New Ways
of Analysing Variation in English, chapter The boundaries of words and their
meanings. Georgetown University Press.
[Lin and Murphy, 1997] Lin, E. L. and Murphy, G. L. (1997). The eects of
background knowledge on object categorization and part detection. Journal
of Experimental Psychology: Human Perception and Performance, 23:1153
1169.
[Marslen-Wilson, 1994] Marslen-Wilson, W. e. a. (1994). Morphology and
meaning in the english mental lexicon. Psychological Review, 215:1+.
9
-
[Matsukawa, 2005] Matsukawa, J. e. a. (2005). Conceptual versus perceptual
priming in incomplete picture identication. Journal of Psycholinguistic Re-
search, 6.
[Medin and Schaer, 1978] Medin, D. L. and Schaer, M. M. (1978). Context
theory of classication learning. Psychological Review, 85:207238.
[Meyer and Schvaneveldt, 1971] Meyer, D. E. and Schvaneveldt, R. W. (1971).
Facilitation in recognizing pairs of words: Evidence of a dependence between
retrieval operations. Journal of Experimental Psychology, 90:227234.
[Nosofsky, 2011] Nosofsky, R. e. a. (2011). The generalized context model: An
exemplar model of classication. Formal Approaches to Categorization, 1:18
39.
[Peirce, 1902] Peirce, C. (1902). Vague. - entry in Dictionary of Philosophy and
Psychology, J.M. Baldwin (ed.).
[Rips and Smith, 1973] Rips, L. J., S. E. J. and Smith, E. E. (1973). Semantic
distance and the verication of semantic relations. Journal of Verbal Learning
and Verbal Behavior, 12:120.
[Rosch.E., 1975] Rosch.E. (1975). Cognitive representations of semantic cate-
gories. Journal of Experimental Psychology: General, 104:192233.
[Tierney, 2002] Tierney, R. (2002). Aristotle's theory of language and meaning
review. The Journal of Philosophy, 99:203209.
[Wittgenstein, 1953] Wittgenstein, L. (1953). Philosophical Investigations.
Blackwell.
10