Академический Документы
Профессиональный Документы
Культура Документы
Let me tell you a little story. Once upon a time there was a young man
whose dream was to reduce the world to pure logic. The young man was very
intelligent, and succeeded indeed. Once he concluded his work, he took a
step backward to admire it. It was marvelous. A world purged of imperfec-
tion and indeterminacy. Infinite acres of glittering ice extended as far as the
horizon. Thus the intelligent man looked a t the world he had created, and
decided to explore it. He took a step forward and fell flat on his back. He had
forgotten friction, you see. The ice was smooth, leveled and immaculate, but
it could not be walked on. Thus the intelligent young man sat down and
cried bitter tears. But as he grew old and wise, he came to understand that
roughness and ambiguity are not imperfections. They are what moves the
world. He wanted to run and dance. The words and things scattered on the
ground were all ruined and dim and ambiguous, and the old wise man saw
that was the way things are. But he kept longing for the ice, where every-
thing was beaming and absolute and inflexible. Even though he had come to
appreciate the idea of a rough ground, he could not convince himself to live
there. Thus, he was now marooned between ground and ice, and in neither
place he felt a t home. And this was the cause of all his pain. (Mi & Jarman,
1993)
Keynes describes the pain of a man who lost the illusions of his youth with-
out learning to enjoy the multiform beauty of reality. In language, vagueness,
indeterminacy, and ambiguity are not defects or imperfections; they are some
of the facets with which the human symbolic system manifests itself, equally
capable of expressing the inflexible laws of logic, the roughness of ground, and
the impalpable substance of poetry.
It took psycholinguistics more than two decades to overcome Ludwig
Wittgenstein’slonging for glacial perfection and fully appreciate the power and
The present research was supported by Ministry of University, Scientific, and Technology
Research (MURST) Fondi ex 40% and Fondi 60%. We thank Corinna Michelin for collecting some
of the data. We are also grateful to David Balota, Andrea Bubka, Dave Gorfein, and Greg Simpson
for their helpful comments.
11
http://dx.doi.org/10.1037/10459-001
On the Consequences of Meaning Selection: Perspectives on Resolving Lexical
Ambiguity, edited by D. S. Gorfein
Copyright © 2001 American Psychological Association. All rights reserved.
12 TABOSSI AND SBISA
flexibility of the human cognitive system deriving from its ability to cope effl-
ciently with the “imperfections” of language. Several years before the cognitive
era, Miller (1951) noticed that apart from close class words, the more frequent a
word is, the more likely it is to be ambiguous, and he concluded that it is easier
for people to deal with few ambiguous words than with many unambiguous
ones. Unfortunately, for a long time the psycholinguistic community underesti-
mated Miller’s insightful remark and considered lexical ambiguity as a special
phenomenon, an anomaly in a system mostly free from deficiencies of this sort
Copyright American Psychological Association. Not for further distribution.
is the focus of this chapter, the findings have proved so difficult to reconcile in a
coherent theoretical frame to justify the conclusion that the literature “does not
paint a very encouraging picture. . . . All the models of ambiguity processing
have found support from one or more studies, a n d . . . there is no simple way of
classifylng studies methodologically in any way that sheds light on the discrep-
ancies among results” (Simpson, 1994, p. 367). In the remainder of the chapter,
we take up this challenge and explore whether discrepancies among studies
can be explained. The available data, once reconsidered, may provide the em-
pirical basis for a coherent theory of lexical ambiguity processing. The chapter
is organized as follows. We first set the scene, introducing the main points of
interest and the main theoretical views. We then discuss two issues that are the
most likely to explain-separately or in interaction-the discrepancies in
the data, namely, the experimental paradigms and the types of contexts used
in the different studies. Finally, we draw some conclusions, focusing on how, in
the domain of spoken word recognition, work on lexical ambiguity resolution
complements our knowledge of lexical processing.
lution. However, dominance is a lexical phenomenon, and its effects are equally
compatible with the modular and the interactive architecture of the mind. It is
perhaps not surprising, therefore, that compared with sentential context ef-
fects, dominance effects have not received much attention. The same applies to
single-word context effects. A long tradition of priming studies has shown that
naming and lexical decision to a target word are facilitated when this is pre-
ceded by a semantically related word (Meyer, Schvaneveldt, & Ruddy, 1975).
Single-word contexts are also known to constrain the initial access of an am-
Copyright American Psychological Association. Not for further distribution.
biguous word (Schvaneveldt, Meyer, & Becker, 1976). But, once again, these
effects are lexical in nature, and their occurrence cannot discriminate between
the two competing views of the cognitive system.
With respect to the effects of sentential contexts, however, the two views
make different predictions. Whereas evidence showing multiple access to
the meanings of an ambiguous word regardless of its context of occurrence
would lend support to the modular hypothesis, evidence of a selective, context-
dependent access would corroborate the constraint-based hypothesis. In prac-
tice, testing these predictions has proved very hard, and a large body of contra-
dictory results has come to constitute the empirical basis for alternative theories,
each compatible with “its own data” and incompatible with most of the others.
Traditionally, there are three classes of theories: the exhaustive model, the
ordered search model, and the interactive model. The exhaustive model is per-
haps the best known (Onifer & Swinney, 1981; Swinney, 1979), according to
which lexical access is an autonomous subprocess of the language comprehen-
sion system. All the meanings of an ambiguous word are initially accessed, and
context intervenes in selecting the appropriate meaning only at a later stage. A
large number of studies, conducted with a variety of different paradigms, cor-
roborates this hypothesis (Blunter & Sommer, 1988;Cairns & Hsu, 1980;Conrad,
1974; Kintsch & Mross, 1985; Seidenberg, Tanenhaus, Leiman, & Bienkowski,
1982; Swinney, 1979; Tanenhaus, Leiman, & Seidenberg, 1979; Till, Mross, &
Kintsch, 1988).
In its strongest version, the hypothesis holds that access is also insensitive
to dominance effects. One of the most compelling pieces of evidence in support
of this claim comes from a study conducted by Onifer and Swinney (1981)using
the cross-modal priming paradigm (Swinney, Onifer, Prather, & Hirshkowitz,
1979). They had their participants listen to sentences such as
Morris, & Rayner, 1988; Rayner & Frazier, 1989; Simpson & Burgess, 1985;
Tabossi, Colombo, & Job, 1987).Hogaboam and Perfetti (1975)were the first to
stress the importance of this factor. They asked the participants to decide whether
the last word in a sentence was ambiguous. The sentences actually containing
an ambiguous item biased its dominant or its subordinate meaning. The results
showed that people were faster at detecting a homograph when the sentence
biased its subordinate rather than its dominant meaning. To account for the
results, Hogaboam and Perfetti proposed the ordered search model. Like the
Copyright American Psychological Association. Not for further distribution.
The water in the bay was so calm that it seemed to be in a pond, rather than
in the sea. (Pond water is still.)
Zn the shop, the smith molded with ease the bar of tin, and made it into a little
soldier. (Tin is soft.)
16 TABOSSI AND SBISA
high salience and subordinate: ACTORS; low salience and dominant: HOLD;
low salience and subordinate: ACTION). The target word was printed in differ-
ent colors, and the participants’ task was to name the ink color as quickly as
possible. The findings showed context effects from the beginning of the process
(at 0 ms), and Paul et al. interpreted them as evidence for the context-sensitive
hypothesis. In doing so, however, they overlooked the fact that the expected
interference effects of appropriate contexts over inappropriate contexts were
found, but only for the targets related to the subordinate meaning (844 ms vs.
800 ms). No evidence of such a difference was observed for the dominant re-
lated targets (838 ms vs. 828 ms), suggesting that context did have an effect on
initial access, but it was modulated by dominance. In a somewhat different
vein, Seidenberg et al. (1982)had initially interpreted their selective results as
due to intralexical factors and hence in agreement with the autonomous model.
Subsequently, however, the findings have been reinterpreted as evidence in
favor of the interactive view (MacDonald, Pearlmutter, & Seidenberg, 1994).
Likewise, data taken in support of the autonomous integration model (Rayner
& Frazier, 1989)have later been reconsidered in light of the hybrid reordered
access model (Dopkins et al., 1992). However, there are discrepancies in the
ambiguity resolution studies that are in need of an appropriate interpretation.
Two factors stand up as the best candidates to explain these discrepancies,sin-
gularly or in interaction: the different methodologies and the nature of the
sentential contexts used in the various studies. Let us consider them in turn.
Lexical processing is fast. Shortly after the initial access to an ambiguous word
has occurred, evidence of multiple activation is no longer found, suggesting that
ambiguity resolution is accomplished very quickly and is over soon after the
ambiguous word has been presented to the listenerheader (Kintsch & Mross,
1985; Lucas, 1987;Onifer & Swinney, 1981).For this reason, it is of fundamen-
tal importance that the process of access be tapped on-line, while it is taking
place, and appropriate experimental paradigms must be used to this end. In
principle, the best paradigm is one which, in addition to capturing access pro-
cesses on-line, relies on a task that does not interfere with the process under
investigation and does not allow participants to develop response strategies.
Moreover, it should give as much information as possible on the availability of
the different meanings of an ambiguous word.
18 TABOSSI AND SBISA
Various techniques have been devised over the years to match these requi-
sites. In the phoneme-monitoring paradigm, people monitor for a word begin-
ning with a specified phoneme in the sentence to which they are listening. Foss
(1970) first used this methodology with ambiguous words and found that pho-
neme monitoring took longer when the target-bearing word was preceded by an
ambiguous than an unambiguous word. As an on-line technique, phoneme moni-
toring initially enjoyed a large popularity (Cairns & Hsu, 1980; Cairns &
Kamerman, 1975; Foss & Jenkins, 1973; Swinney & Hakes, 1976). However,
Copyright American Psychological Association. Not for further distribution.
one problem with this methodology is that it is a “load task”: There is no direct
evidence of whether one or more meanings of an ambiguity have been accessed.
This can only be inferred comparing the phoneme detection times after an am-
biguous word and after a control (Simpson, 1994). Moreover, careful analysis of
the characteristics of the task showed that phoneme monitoring is influenced
by many factors (Cutler & Norris, 1979;Ferreira & Anes, 1994;Holmes & Forster,
1970; Shields, McHugh, & Martin, 1974). Detecting the initial phoneme of a
target word, for example, takes longer the longer and less predictable is the
vowel following the initial phoneme (Foss & Gernsbacher, 1983; Swinney &
Prather, 1980). It also takes longer when the target is preceded by a low-
frequency word (Foss, 1969), by a long word (Mehler, Segui, & Carey, 1987), or
by a word beginning with a similar phoneme (Dell & Newman, 1980; Foss &
Gernsbacher, 1983).
An interesting alternative to phoneme monitoring was offered by the cross-
modal paradigm (Swinney et al., 1979). The paradigm relies on the semantic
priming effect (Schvaneveldt & Meyer, 1973):Performing a lexical decision or a
naming task on a visual target word (e.g., CAT) is faster when the target is
preceded by a semantically related word (e.g., DOG) than by an unrelated one
(e.g., CHAIR). In the cross-modal paradigm, participants listen to sentences
containing an ambiguous word. At the offset of the ambiguous word (the prime),
they are shown a visual target word on which they perform either a lexical
decision or a naming task. In the experimental trials, the visual target is se-
mantically related to one meaning of the ambiguity or is a control word. Re-
sponse times to the visual target following the ambiguous word in different
contexts are taken to indicate which of the meanings of the ambiguity is avail-
able.
Although the cross-modal paradigm is admittedly questionable in its eco-
logical validity, it has several desirable characteristics. It provides a direct indi-
cation of whether a meaning of an ambiguous word has been accessed, and
according to Swinney (1979, p. 6481, “the task reflects the access of the auditory
(priming) words, without drawing attention to the relationship involved.”Also,
the acoustic presentation of the prime allows an accurate control of the interval
between the offset of the prime and the onset of the target (interstimulus inter-
val; ISI). Finally, the paradigm has been the object of intense scrutiny, and a
substantial amount of research devoted to test some or other of its features
suggests its reliability (Burgess, Tanenhaus, & Seidenberg, 1989; Peterson &
Simpson, 1989).
Accurate control of the temporal relations between prime and target is usu-
ally more complex when the prime is presented in the visual modality. Among
the various reading techniques used to investigate access to ambiguous words,
METHODOLOGICAL ISSUES 19
the most likely to reflect initial access is probably eye tracking (Pollatsek &
Rayner, 1990). Participants read sentences containing an ambiguous word, and
the time spent on the ambiguous item in different types of context is measured.
Alternatively, the times spent on an ambiguous word and an unambiguous con-
trol are compared.
One advantage of this methodology is that reading is a normal activity, and
even though the experimental setting required to record eye position is rather
unnatural, no disrupting task is given to the participants. However, adequacy
Copyright American Psychological Association. Not for further distribution.
In the lexical ambiguity literature, a biasing context is one that clearly sup-
ports one meaning of an ambiguous word occurring in it. The bias can be ob-
tained in different ways. For example, an ambiguous word in one of its mean-
ings (e.g., card deck) may be a very likely completion of a prior sentence fragment,
for example, Thegamblerpulled an ace from the bottom of the -(Van Petten
& Kutas, 1987). Syntactic information may also provide a very efficient bias
when the alternative meanings of an ambiguous word belong to different syn-
tactic categories (Tanenhaus et al., 1979).
Although intuition suggests that the various types of context may differ as
to how effectively they influence access, the issue has not received much atten-
tion. We have already mentioned the studies conducted by the proponents of
hybrid models (Dopkins et al., 1992; Tabossi, 1988a;Tabossi et al., 1987).Dif-
ferent types of context have also been explored by Carpenter and Daneman
(1981) and Seidenberg et al. (1982). Simpson (1981) manipulated strength of
context in a cross-modal lexical decision study. He found that when an ambigu-
ous word (e.g., count) with a dominant and a subordinate meaning occurred in a
neutral context (e.g., He had trouble keeping track of the count), the dominant
meaning was accessed first. However, when the sentential bias was weak, ac-
cess was selective if the sentence biased the dominant meaning (e.g., The musi-
cian kept losing track of the count) but was exhaustive otherwise (e.g., The king
METHODOLOGICAL ISSUES 21
kept losing track of the count). Finally, if the ambiguous word occurred in a
context that strongly biased one of its meaning (e.g., The dog wasn't included in
the final count and The vampire was disguised as a handsome count), only that
meaning was accessed.
These studies used various methodologies: cross-modal associated with lexi-
cal decision and with naming, reading, and eye-tracking. As for the nature of
the contexts, Tabossi and colleagues (Tabossi, 1988a;Tabossi et al., 1987) con-
trasted constrainingversus nonconstraining contexts, Dopkins et al. (1992)used
Copyright American Psychological Association. Not for further distribution.
positive versus negative biases, and Seidenberg et al. (1982) looked at the ef-
fects of associated versus related words. Similarly, Carpenter and Daneman
(1981) manipulated context strength by including in the sentences words re-
lated to one meaning of the ambiguity. These differences notwithstanding, all
of the studies indicated that different contexts may produce different effects on
access, thus challenging the modular, context-insensitiveview.
From a strictly modular perspective, some of the studies suggesting differ-
ential effects of context may be accounted for methodologically. Others, in par-
ticular those introducing words related to one meaning of the ambiguity in the
biasing contexts, may be seen as showing intralexical effects. But, how does one
reconcile a strictly modular view with results such as those obtained by Tabossi
and colleagues, who used a seemingly sound paradigm and did not rely on lexi-
cal associations to manipulate sentential biases?
Even though the mechanisms explicitly used to produce the biases were
not intended to be lexical, the individual words in the sentential contexts, par-
ticularly the constraining ones, were inevitably related to the biased meaning
of the ambiguities. Those words could affect access to the ambiguous words,
producing selectiveeffectsthat could, after all, be lexical (Cutler, 1995). It might
be argued that the occurrence of an associate is not sufficient t o yield selective
access to one meaning of an ambiguous word (Onifer & Swinney, 1981; Swin-
ney, 1979). In fact, lexical priming effects do not occur automatically (Foss &
Ross, 1983; Vu et al., 19981, and many studies that compared priming effects
produced by individual words in isolation, in coherent contexts, and in scrambled
contexts invariably obtained different results depending on the context of oc-
currence of the prime (Auble & Franks, 1983; OSeaghdha, 1989; Simpson,
Peterson, Casteel, & Burgess, 1989).
A further possibility is that the individual words in the context may di-
rectly prime a target, speeding up responses to it (Simpson & Krueger, 1991).
According to this hypothesis, access is context insensitive, and in a priming
paradigm, the ambiguous word will effectively prime the related targets re-
gardless of context. However, if another word in the context is related t o a tar-
get, that word may also prime that target, contributing to speed up responses to
it.
One difficulty with this hypothesis is that lexical priming is short-lived,
and unless the relevant word immediately precedes the ambiguity, it is most
unlikely that it can exert an effect on the target (Neely, 1991). Moreover, the
hypothesis predicts that responses to the target related to both meanings of the
ambiguity are faster than responses to unrelated controls. Yet, in the studies in
which we observed selective effects, responses to subordinate related, inappro-
priate targets were never found to be reliably faster than responses to controls.
22 TABOSSI AND SBISA
biguity or unrelated to it. The same results were obtained when the related
targets denoted a salient feature of one meaning of the ambiguous word and
when they were associated to it. Accordingly, in the first two experiments, we
examined the possibility that the constraining contexts used in that study con-
tained lexical items that could prime the dominant meaning of the ambiguous
word and lead to the selective access of that meaning. In Experiment 1, we
constructed a sequence of words for each of the ambiguous items used in Tabossi
(1988a). The sequence ended with the ambiguity (e.g., stagno: pondtin) and
began with a word that was selected from among the content words preceding
the ambiguity in the dominant constraining context (e.g., water, bay, calm, and
seem) as the most highly associated with it (e.g., water, henceforth: associate
word). In addition, each experimental sequence contained a number of new
unrelated items that equaled the number of content words occurring between
the associate and ambiguous word in the constraining context. Four content
words, for instance, separated water from pond in the dominant constraining
context; therefore, we included four new, unrelated words in the sequence.
Paired with each experimental sequence, there was a set of three words to
use as visual targets. The targets were the same associate words and unrelated
matched controls as in Tabossi (1988a).Experimental materials are illustrated
in the following example:
Participants were faster at deciding that a target was a word when it was
related to the dominant (M = 570 ms, SD = 62) or the subordinate (M = 570 ms,
SD = 84)meaning of the ambiguous word than when it was an unrelated con-
trol (M = 612 ms, SD = 100).Response latencies did not reliably differ from each
other.
Taken together, these findings suggest that in Tabossi (1988a) the selec-
tive effects observed after the constraining contexts were not produced by indi-
vidual words associated with the dominant meaning of the ambiguity and cor-
roborate the hypothesis of a genuine effect of context. If individual words in the
constraining contexts did not prime the dominant meanings of the ambiguities
in Tabossi (1988a),it is still possible that they primed the visual targets related
to those meanings, speedingup responses to them. We have already pointed out
some of the difficulties faced by this hypothesis. We might also add that the
selective effects produced by contexts constraining the dominant meaning of an
ambiguity have been obtained with two sets of targets-associate and feature-
denoting words-and it is difficult to see how the same words could equally well
prime two different sets of targets. Irrespective of these observations, we de-
vised two additional experiments to test whether indeed there were words in
the constraining contexts that primed the dominant-related visual targets. Ac-
cordingly, in Experiment 3, we selected from each of the constraining sentences
used in Tabossi (1988a)the word most strongly associated to the dominant re-
lated associated target, yielding a total of 9 pairs (Tabossi’s pairs: e.g., water-
FROG). In addition, we selected 9 new pairs in which prime and target were
strongly associated (new pairs: e.g., salt-PEPPER) and 18 filler pairs in which
the prime was a word and the target a legal pseudoword. We then constructed
two experimental lists, each of which contained all the filler pairs and all the
experimental materials. However, for the experimental materials, half of the
targets were preceded by their associated primes in one list and were preceded
by unrelated primes in the other list. The reverse was true for the remaining
half of the experimental targets. Participants were randomly assigned to one
list. They listened to a prime word immediately followed by a visual target, on
which they performed a lexical decision task.
The results indicate that with the new pairs participants were faster in the
associated (M = 510 ms, SD = 32) than in the unrelated (M = 560 ms, SD = 56)
condition. However, no such difference was observed in Tabossi’s pairs (related
24 TABOSSI AND SBISA
M = 539 ms, SD = 46; unrelatedM = 546 ms, SD = 29). Comparable results were
obtained in Experiment 4. In this experiment, the rationale, design, and par-
ticipants’ task were the same as in Experiment 3. Materials, however, were
different. In Tabossi’s pairs, the targets were the feature-denoting targets used
in Tabossi (1988a)and the primes were the words in the constraining contexts
most highly associated with them (e.g., calm-STILL). A new set of associated
pairs was also selected to match the frequency and length characteristics of the
feature targets and their primes (e.g., blood-RED). Participants’ responses were
faster in the associated (M = 551 ms, SD = 41) than in the unrelated (M = 630
Copyright American Psychological Association. Not for further distribution.
ms, SD = 61) condition for the new pairs, whereas no such difference was ob-
tained in Tabossi’s pairs (related M = 616 ms, SD = 55; unrelated M = 634 ms,
SD = 87). Taken together, Experiments 3 and 4 rule out the possibility that the
findings in Tabossi (1988a)may result from individual items in the prior con-
text directly priming the dominant related targets.
The study we conducted intended to provide empirical evidence in response
to two of the most common criticisms raised to a context-dependentinterpreta-
tion of selective results by those who strongly favor a completely autonomous
view of lexical access. Our findings suggest that in Tabossi (1988a),an intralexical
interpretation of the data is not likely. More generally, there are findings in the
literature that suggest effects of contexts in the early phases of lexical access
and cannot be reduced to lexical phenomena.
Conclusion
Over the years, the notion that access may not be an entirely autonomous pro-
cess has gathered increasing consensus, and the growing body of evidence run-
ning against a strictly modular hypothesis of access, although not always incon-
trovertible, has led some who were originally in favor of radically autonomous
views to reconsider the issue with a more flexible attitude. The result of this
change of perspective, along with the undisputable force of the evidence sug-
gesting a high degree of independence of lexical processing, has allowed most
researchers to converge on a hybrid view according to which dominance and
context both influence the process of lexical access, whose outcome depends
crucially on the relative strength of these factors.
Another important change that has taken place in the field of lexical ambi-
guity resolution is that early models implicitly assumed, as did most contempo-
rary models of lexical processes, that access is a discrete phenomenon that oc-
curs at one point during the processing of an ambiguous word. Subsequently,
however, access to the lexicon has come to be viewed as a continuous event that
is accomplished when semantic information about a word is sufficiently acti-
vated to be integrated, if appropriate,into prior discourse (Marslen-Wilson,1987).
Current models of lexical ambiguity resolution are accordingly expressed in
terms of time course of activation of the different meanings of an ambiguous
word. Simpson and Burgess (19851, for instance, presented their participants
with an ambiguous word followed by a target that was related either to the
dominant or to the subordinate meaning of the preceding ambiguity, or else
METHODOLOGICAL ISSUES 25
Many years ago, psycholinguists believed that ambiguous words were some
kind of anomaly in a system in which most words were unambiguous. They
later discovered that ambiguity is in fact a very common phenomenon that there-
fore needs to be explained. They still maintained, however, that different mecha-
nisms might be required for the processing of ambiguous and unambiguous
words. This attitude, which occasionally was explicitly expressed (Potter &
Faulconer, 19791, more often manifested itself in the division of labor: Research
on lexical ambiguity resolution and research on unambiguous word recognition
Copyright American Psychological Association. Not for further distribution.
were conducted for the most part independently from one another, and the pos-
sibility of integration was only a promise: