Hitting the Right Pitch: A Meta-Analysis of Effect of Sentence Context on Lexical Access

Mark F. St.John

artment of Cognitive Science, UCS

I would like to thank Patricia Carpenter for her many suggestions made 
throughout the course of this project.  I would also like to thank Jay 
McClelland and Norman Vinson for their useful suggestions and Ernest Mross 
for the extra statistics he provided from his experiments.

Send correspondence to Mark St. John, Department of Cognitive Science, 
UCSD, La Jolla, CA 92092-0515.  E-mail: mstjohn@ucsd.edu

     A long standing controversy in Psycholinguistics is whether sentence 
context affects the lexical access process.  The revealing case is whether 
sentence context affects the access of the meanings of ambiguous words.  
According to one account, termed Independent Access, each meaning of an 
ambiguous word is initially accessed and equally activated.  After a short 
delay, the preceding sentence context is applied to prune the inappropriate 
meanings, leaving only the contextually appropriate meaning active.  A number 
of studies (Swinney, 1979; Onifer & Swinney, 1981; Seidenberg, Tanenhaus, 
Leiman, & Bienkowski, 1982; Kintsch & Mross, 1985; Till, Mross, & Kintsch, 
1988; Blutner & Sommer, 1988) have found this pattern of results.

     These studies used a priming method to assess the relative degree of 
activation of each meaning of the ambiguous words.  The idea behind the 
priming method is that subjects can respond to a target word more quickly if 
the context preceding the target word "primes" that word.  For example, 
subjects can make a lexical decision (decide if a letter string spells a 
word) to "nurse" more quickly if it is preceded by a semantically related 
word like "doctor" (Meyer & Schvaneveldt, 1976).  The explanation of this
phenomenon is that the word "doctor" activates the concept of doctor which 
in turn activates semantically related concepts like nurse.  By observing 
which target words are facilitated in the lexical decision task, the priming 
method can be used to determine which concepts are activated when a word is

     The priming method can be used to study the influence of sentence 
context on lexical access.  To determine if a sentence context can affect which 
meanings are initially activated by an ambiguous word, a sentence containing 
an ambiguous word is used as the context for the lexical decision.  Once the 
sentence is presented, subjects respond to a target word related to one of the 
meanings of the ambiguous word.  For example, the sentence might be, "The 
janitor found several roaches, insects, and other bugs..."  The target word 
would be "ant" or "spy" or an unrelated control word.  If both the contextually 
appropriate and inappropriate meanings are facilitated, then the Independent 
Access account holds.  If, however, the contextually appropriate target word is 
primed more than the inappropriate target word (in this case, "ant" rather than 
"spy"), then the sentence context affected the processing of the ambiguous 
word "bugs".

     The delay between the presentation of the ambiguous word and the 
target can be varied to examine the time-course of processing of the 
ambiguous word.  With no delay, the meanings immediately activated by the 
ambiguous word, and possibly the sentence, can be observed.  With increasing 
delays, the activations of meanings at later points in processing can be 
observed (cf. Till, Mross, & Kintsch, 1988).

     The studies mentioned above found that both meanings of the 
ambiguous word were significantly primed when there was no delay between 
the ambiguous word and the target word.  The amount of priming for the 
contextually appropriate and inappropriate meanings were not significantly 
different.  These studies, therefore, showed that the sentence context did not 
affect the lexical access processing of the ambiguous word and supported the 
Independent Access account.  

     This empirical result has important implications for theories of language 
processing.  If lexical access if unaffected by the sentence context, then
lexical access can be viewed as a processing "module" that works independently 
for each word processed (Fodor, 1983; Kintsch & Mross, 1985).  A higher level 
sentence processing module would then take the set of meanings accessed 
and prune the inappropriate ones.

An Alternative Account

     A different account of sentence processing, however, predicts a 
different pattern of results and is also supported by a number of studies 
(Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Kintsch & Mross, 
1985; Van Petten & Kutas, 1987).  This account, termed Biased Activation, 
predicts that the sentence context does prime the contextually appropriate 
meaning of the ambiguous word.  The ambiguous word itself works to activate 
each meaning, but the sentence context works to activate the contextually 
appropriate meaning more.  As activation increases, the meanings compete, 
and the edge given the appropriate meaning increases.  The inappropriate 
meanings are suppressed, and the appropriate meaning is fully activated.  
Immediately after presenting the ambiguous word, therefore, all meanings are 
active, but the contextually appropriate meaning is more active.

     The Biased Activation account supports an interactive view of 
sentence processing (McClelland, 1987).  The effect of sentence context is 
interpreted as demonstrating the interaction between lexical access and the 
developing interpretation of the sentence.  As the sentence is processed, 
activation feeds back to the lexical access process.  Concepts related to the 
sentence interpretation, including the contextually appropriate meaning of the 
ambiguous word, are primed.  As processing continues, the feedback loop 
between processing levels leads to the full activation of the appropriate 
meaning and full suppression of the inappropriate meanings.

     Figure 1 depicts the time-course of activation predicted by each view.  
Note that both views make the same prediction for meaning activations after a 
delay.  It is in the immediate processing of the ambiguous word where the two 
views differ.


       Independent Access
     |      |--------    approp.
     |     |   |
  ACT|    |    |
     |   |      |
     |  |       |        inapprop.
     ambig.   target

       Biased Activation
     |      |--------    approp.
     |     |  
  ACT|    ||-- |
     |   ||     |
     |  ||      |        inapprop.
     ambig.   target

Figure Captions

Figure 1.  The time-course of activation of the contextually appropriate 
and inappropriate meanings of an ambiguous word for Independent Access and 
Biased Activation.  The horizontal axis marks the onset of the ambiguous 
word and the target.  The theories' predictions only differ during the early 
stages of processing in the relative levels of activation of each meaning.


     Since both views are supported by empirical observation, how might 
the controversy be resolved?  Part of the answer lies in the design of the 
studies.  Each study used Independent Access as its null hypothesis.  The null 
hypothesis, then, predicts no difference between the contextually appropriate 
and inappropriate meanings of an ambiguous word.  Failure to achieve a 
significant difference, then, was taken to affirm Independent Access.  A null 
result, however, is only weak evidence for a hypothesis.  By looking across 
studies, the evidence for each hypothesis can be assessed with greater 
power.   Meta-analysis (Wolf, 1986) provides a way of systematically 
combining experimental results across different studies.  Using a meta-
analysis, then, the effect of sentence context on lexical access can be 
assessed more powerfully.

     For a meta-analysis to be successful, the experiments included in the 
analysis must be chosen carefully.  The analysis presented here was limited to 
priming studies and to studies that used semantic sentence contexts.  While 
methods than priming have been used (See Simpson, 1984, for a thorough 
review), the priming method is most appropriate because it examines the 
activations of meanings.

     Experiments that examined the effect of syntactic context are excluded 
from this meta-analysis for two reasons.  First syntactic information is 
sufficiently different from semantic information that combining the two might 
obscure the observation of effects due to either.  Secondly, meta-analysis 
requires a reasonably large number of tests of a hypothesis, but few priming 
studies have examined the effect of syntax.

     An examination of the results of the chosen experiments reveals a 
consistent effect of sentence context.  In nearly every study, the contextually 
appropriate meaning shows greater facilitation immediately following the 
ambiguous word, though the effect is rarely significant.  The meta-analysis 
shows the effect to be significant across experiments thereby supporting the 
Biased Activation account of lexical access.  This result lends support to the 
interactive account of language processing.


     The psychological literature was searched for experimental tests of the 
Independent Access hypothesis.  Psychological Abstracts was searched for 
relevant articles and dissertations, and Social Science Citations was searched 
for references to two early studies: Swinney (1979) and Onifer and Swinney 
(1981).  Seven studies were found that used priming, semantic contexts, and 
tested the hypothesis after no delay between the ambiguous word and the 
target word (footnote 1).  Among these studies, 18 separate tests of the 
hypothesis were reported.

     A common problem for meta-analysis is that studies that do not find 
significant effects may not be published.  Hence, a meta-analysis is typically 
biased toward positive effects.  This is known at the "file drawer problem".  
However, this bias may not operate here.  In the literature on lexical access, 
Independent Access, the null hypothesis, has been a well established result, 
so null effects can be reported with confidence as replications of this result. It is unlikely, therefore, that a nonsignificant effect would preclude 

     The procedure followed in each experiment was similar.  In most 
studies, the subject listened to sentences presented over a pair of head 
phones.  Following the ambiguous word, a visual lexical decision or naming 
task would occur.  For lexical decision, a word or nonword string was 
displayed on a screen in front of the subject.  The subject responded yes if 
it were a word and no if it were not a word.  For the naming task, a word was 
displayed, and the subject read the word aloud.  The time taken to respond in 
either task was recorded.  

     In three experiments, however, both the sentences and the targets 
were presented visually (Kintsch & Mross, 1985: experiments 2 and 3; Van 
Petten & Kutas, 1987: experiment 1).  Each sentence was displayed one word 
at a time in the center of the screen.  Following the ambiguous word, a target 
was presented in the same location, flanked by rows of asterisks.  The 
presentation rate for the sentence contexts varied between experiments.  The 
effect of this procedural manipulation will be discussed later.

     The materials were less similar across studies.  The materials typically 
consisted of a set of individual sentences ending in an ambiguous word and a 
set of target words for each sentence: one related to each meaning of the 
ambiguous word, and perhaps several control words.  Other studies presented 
the sentences in paragraph form, or embedded the ambiguous word in the 
middle of a sentence.

     The length and frequency of targets words, important factors 
themselves in lexical access (Just & Carpenter, 1980), were controlled in 
either of two ways.  Some studies matched the length and frequency of the 
words representing to each meaning of the ambiguous word.  Analysis then 
consisted of comparing the reaction time to the target for one meaning against 
the time for the other meaning.  Other studies developed a control word, 
matched for length and frequency, for each target.  Analysis then consisted of 
finding the differences between target words and their controls, and then 
comparing the amounts of facilitation for each target word.

      Depending upon the study, the frequencies of the meanings of the 
ambiguous words were either controlled to be the same or manipulated.  If the 
frequency were manipulated, the effect of sentence context, averaged over 
frequency, is analyzed here.

     A meta-analysis provides a way to combine the results of independent 
tests of a hypothesis.  One method, the Fisher Combined Test (Fisher, 1948), 
combines the probability of the type I error (p-value) associated with each 
experiment into a single chi-square statistic using the formula below.
               X^2 = -2 * (sigma)  log   p				(1)
                                 i    e   i  

     The p-value for the chi-square statistic is then computed using 2n 
degrees of freedom, where n is the number of tests of the hypothesis.

     The nature of the Fisher Combined Test formula is to give 
exponentially greater weight to more significant values of p (the smaller
the p, the exponentially larger is the negative of its logarithm).  Each
experiment also contributes two degrees of freedom to the analysis.  For 
chi-square tests, the more degrees of freedom, the higher the chi-square value 
must be to reach significance.  Therefore, experiments with high significance 
(small p-values) add more to the size of the estimated chi-square value than
they take away through the degrees of freedom they contribute.

     Four experiments reported "F<1" rather than the exact F or p for the 
critical comparison (though in fact each experiment demonstrates a small trend 
toward greater activation of the contextually appropriate meaning).  It is 
important to include these studies since they test the hypothesis and show no 
significant effect.  These studies were assigned a very conservative p-value of 
p = 0.9.

    The effect size, d, describes the size of the experimental effect in 
standard units.  It can also be calculated for each experiment and averaged to 
obtain a mean effect size.  The effect size is calculated from the F or t 
statistic.  Unfortunately, this means that when the inappropriate meaning is 
faster, d remains positive, it just gets very small.  In the three tests where 
this difficulty occurs, an effect size of -0.1 standard units was used instead.  
    Again, the four studies without statistics had to be assigned an effect 
size; they were assigned: d = 0.1.

    The results of each test of the Independent Access hypothesis are 
shown in Table 1.  The Fisher Combined Test shows that words corresponding
to the contextually appropriate meaning of the ambiguous words are responded 
to significantly faster than words corresponding to the inappropriate meanings: 
X^2(36) = 58.68, p = .01.  The contextually appropriate meanings of ambiguous 
words are significantly more active, and consequently, sentence contests 
influence lexical access.

Table 1
Significance Level and Effect Size

  p	  d	study				experiment	method
.15     .23	Swinney, 1979			1		lexical decision
.06	.38		"			2		lexical decision
.61	.18	Onifer & Swinney, 1981		1		lexical decision
 --*	---		"			2		lexical decision
.92     -.10	Kintsch & Mross, 1985		1		lexical decision
.62	-.10		"			2		lexical decision
.02	1.01		"			3		lexical decision
.82	-.10	Till, Mross, & Kintsch, 1988	1		lexical decision
.14	.32		"			2		lexical decision
.57	.15		"			3		lexical decision
.09	.45	Blutner & Sommer, 1988		1		lexical decision
---	---		"			1		lexical decision
.05	.42	Seidenberg, et al., 1982	2		naming
---	---		"			3		naming
.01	1.85		"			4		naming
---	---		"			4		naming
.15	.69		"			5		naming
.01	3.13	Van Petten & Kutas, 1987	1		naming

	X^2 (36) = 58.68,  p = .01	average effect size (d) = .50

*The hyphen indicates an experiment reporting a nonsignificant effect, but 
providing no detailed statistics.

    The mean effect size, in standard units, is 0.50.  The effect size itself 
has a standard deviation of 0.81.  The effect's small size can be attributed 
speculatively to a number of factors.  One cause may be the high speed of the 
lexical decision and naming tasks.  Response times ranged between 600 and 
850 milliseconds.  During this brief moment, there is little time for the 
sentence context to have an impact.  Essentially, the tasks may be so fast as 
to be insensitive to some influences.

    Another plausible cause of the small effect size is that the sentence 
context is rather weak in some studies.  Compare the weak context of Kintsch 
& Mross (1985), "...The children liked best the huge wood stove in the kitchen 
which was made out of iron..." with the strong context of Van Petten & Kutas 
(1987), "The gambler pulled an ace from the bottom of the deck."  Carpenter 
and Daneman (1981) have shown that if a context is weak and one meaning of 
an ambiguous word is far more frequent than the other meaning, subjects may 
actually misinterpret the ambiguous word by choosing the more frequent, but 
contextually inappropriate, meaning.  Since a weak context may not affect the 
final interpretation of an ambiguous word, it may not immediately affect the 
relative levels of activation of its meanings.  Weak contexts, then, may not 
allow even the possibility for observing differences in activation.

    Variance in the effect size may have arisen from variability in word 
length and frequency.  Though control words to individual experimental words 
are matched for these variables, length and frequency across control words 
and experimental words are not typically well controlled.  Each factor adds a 
little variability until the small effect of sentence context is obscured.

    Finally, the robustness of the meta-analysis result should be 
determined.  One measure of robustness is the number of tests demonstrating 
equal facilitation of both meanings that would be required to make the meta-
analysis nonsignificant.  These hypothetical no-effect studies were defined as 
p = .5, since in a t-test, completely overlapping distributions would have this 
p-value.  Eight no-effect studies were required to drive the estimated chi-
square to nonsignificance (p > .05).

    A second measure of robustness is to determine how important each 
existing test is to the meta-analysis result.  The biggest contributors to the 
estimated chi-square are the four significant tests, particularly the two 
p < .01 tests.  Any one of the four tests may be eliminated without making 
the meta- analysis nonsignificant.  This result alleviates some concern that 
the meta- analysis result might rest on some experimental fluke in a single 
experiment.  Along this same line, it should be noted that the four significant tests come from three different laboratories.


    The contextually appropriate meanings of ambiguous words are 
significantly more active than the inappropriate meanings immediately after 
the ambiguous word's presentation.  This empirical support for the Biased 
Activation account of lexical access is often overlooked because many 
experiments produce statistically nonsignificant results that are individually 
taken to affirm the Independent Access account.  But Independent Access, 
which predicts all meanings to be equally active, is the null hypothesis in 
each experiment.  Consequently, a nonsignificant result is only weak evidence 
for Independent Access.  On the other hand, nearly every experiment shows at 
least a trend toward greater activation of the appropriate meaning.  In cases 
such as this one, multiple replication can draw the effect out of the variance.

     The next step is to consider what this result implies about theories of 
the lexical access process: Modularity or Interactive Activation.  The more 
straight-forward implication is Interactive Activation: top-down influences 
from the sentence context affect the access process from its beginning.  There 
are, however, alternative explanations of the Biased Activation result.

Alternative Explanations
     *Post-lexical access processing*  A common complaint of priming studies 
is that they reflect post-lexical access processes, or integration, as well as 
access.  The claim is that the response task, particularly the lexical decision 
task, is slow enough that there is time for integrative processes to boost the 
activation of the contextually appropriate target before a response is computed 
(e.g. Seidenberg, Waters, Sanders, & Langer, 1984).  If this claim is correct, 
then there still may be a lexical access module producing Independent Access, 
but post-lexical access processes obscure it.

     To eliminate the post-lexical access processes, a faster task, such as 
naming, can be used so that there is less time for these influences to develop. 
A look at Table 1, however, indicates that significant context effects have 
been found with both lexical decision and naming.  If anything, naming shows
the effect more strongly.  The post-lexical access processing hypothesis, then,
is not borne out by the data.      

     *Semantic associations*  Another rival explanation (Seidenberg, 
Tanenhaus, Leiman, & Bienkowski, 1982) is that facilitation of the 
contextually appropriate target occurs only when the context contains 
concepts semantically associated to the target.  For example, the concepts 
'gambler','deal', and 'cards' are semantic associates.  Given the sentence,
"The sly gambler dealt the hand from the bottom of the deck" and the target
word "cards", subjects may respond quickly to "cards" not because of top-down 
influences from the sentence context, but because it is semantically associated 
to other words in the sentence.

     The semantic association explanation grants that the context affects 
lexical access, but contends that it is not the sentence context per se, but 
simple semantic associations between the words found in a sentence.  These 
simple associations might well be computed within the lexical access module, 
thereby saving the modular theory of processing and denying any interactive, 
top-down influences.

     Of course, semantic associations are a perfectly acceptable, and large, 
aspect of a sentence context.  On the other hand, a sentence is not simply a 
string of words.  Its meaning is much more specific than the union of semantic 
associations.  Does this more specific meaning of the sentence context affect 
lexical access?

     Seidenberg et al. (1982) addressed this question by removing as many 
of the individually associated words from the sentence contexts as possible, 
creating what they termed "pragmatic contexts" (experiment 3).  They were 
left with contexts like, "bought a spade", and "played a spade".  They argued 
that pragmatics made one meaning more sensible in each context, but that 
there were only weak semantic associations.  These pragmatic contexts 
produced no significant context effect (F < 1).  The contextually appropriate 
targets were very slightly better primed (18 milliseconds of facilitation for
the appropriate targets vs. 15 milliseconds for the inappropriate targets).

     The problem with this study is that in removing all the semantic 
associations, Seidenberg et al. removed most of the meaning from the 
sentence contexts thereby weakened them.  It is possible that the difference 
between pragmatic contexts and contexts that contain semantic associations 
is not a difference in the kind of information, but in the quantity of
information each applies to the meaning of the ambiguity.

     A better approach is to actually test whether the more specific meaning 
of a sentence is a better prime than the associations of its individual words.  
Along this line, Tabossi (1988b) tested whether a sentence context, as a 
whole, would deferentially prime specific features of a concept.  Subjects
heard sentences ending in nouns.  Immediately after each sentence, subjects
were shown a target word to name aloud.  The target word either described a 
contextually appropriate feature of the noun or an inappropriate feature.  "The 
girl was pricked by a rose" makes "thorns" contextually appropriate, but does 
not make "scent" appropriate.  Though the experiment does not deal with 
ambiguous words, it does address a very similar situation.  Tabossi found that 
subjects were significantly faster to name the contextually appropriate 
feature.  The inappropriate feature was not significantly primed.

     As a control for a possible semantic association between "pricked" and 
"thorns", the study included a control condition which retained all but the
final noun: "The girl was pricked by the wasp."  If associations between 
individual words are responsible for the context effect, then the control
sentence should also demonstrate the effect, and prime "thorns".  Tabossi 
found no priming for the control sentences.  These results suggest that it is 
in fact the sentence context per se that influences lexical access.  

     Of course, this result again confirms a null hypothesis, so it actually 
provides only weak evidence against the semantic association explanation.  A 
second problem with Tabossi's result is the possibility that "wasp" 
sufficiently changes the meaning of "pricked" that it no longer primes "thorns".  
     This explanation, however, assumes rather complicated processing inside
the lexical access module and further degrades the distinction between lexical 
access and sentence integration.

     More research on the effects of semantic associations and sentence 
contexts may clarify these issues.  It seems likely, though, that both factors 
influence lexical access.  To understand the influence of the sentence context 
on lexical access, then, we should consider contributions by both the semantic 
associations between words and the specific meaning of the sentence as a 

Additional Influences on Lexical Access
     Biased Activation fits comfortably with our understanding of other 
influences on lexical access.  Sentence context, meaning frequency, sentence 
presentation rate, and attention all appear to influence the access of word 

     *Frequency*  The frequencies of the alternative meanings of ambiguous 
words affect the speed and strength of their access.  Simpson (1981) found 
that, in the absence of a biasing sentence context, the dominant meaning was 
more strongly activated than the subordinate meanings 120 milliseconds after 
the offset of the ambiguous word.  Unfortunately, this dominance effect has not 
been tested at shorter delays.  Carpenter and Daneman (1981) found that the 
more frequent was the dominant meaning of an ambiguous word, the less time 
subjects spent reading it.  

     *Sentence context*  The influence of sentence context also varies 
according to its strength, as discussed above.  Seidenberg et al. (1982) found 
small nonsignificant context effects in their weak "pragmatic" condition 
(experiment 3), but large significant effects in their strong semantic
condition (experiments 2 and 4).  Carpenter and Daneman (1981) pitted sentence 
context against meaning frequency for words like "sewer", which have a large 
disparity in meaning frequencies (plumbing >> tailor).  In a weak sentence 
context that supported the subordinate meaning, subjects accessed the 
inappropriate, dominate meaning.  Much stronger sentence contexts elicited 
the contextually appropriate, subordinate meaning.

     Kawamoto's (1986) model of lexical access illustrates these effects.  
The model is a single layer neural network that settles into a representation 
of the orthography and meaning of a word.  The sizes of weights represent the 
strengths of associations between elements of the representation.  For 
ambiguous words, the orthographic units are associated with two sets of 
semantic units.  When a word is read into the network, orthographic unis are 
activated.  Activation then spreads on each settling cycle to activate 
associated meaning units.  In this way, the model is able to simulate the time-
course of lexical access.

    The effect of meaning frequency is caused by the strengths of weights.  
Large weights produce strong and fast settling to a meaning.  When the 
frequencies of two meanings of an ambiguous word differ, the dominant 
meaning acquires activation faster than the subordinate meaning and then 
suppresses the subordinate meaning.

    Sentence context is incorporated by slightly activating the semantic 
units when the orthography is read in.  This activation biases the model to 
settle into that meaning.  However, if the inappropriate meaning is strongly 
dominant, the stronger weights from the orthography to that meaning will lead 
to an over-riding of the initial semantic bias.  This process reproduces 
Carpenter and Daneman's inconsistency results.

    *Presentation rate* The rate of presentation of the sentence also affects 
lexical access by affecting comprehension and thereby any feedback from the 
sentence context.  Most studies presented the sentences aurally, thereby 
giving an average of 350 milliseconds to each word.  Kintsch and Mross (1985: 
experiment 2), however, presented the sentences visually, one word at a time 
in the center of a computer screen.  They allotted only 190 miliseconds to each 
word.  This Rapid Serial Visual Presentation method has been shown to 
degrade comprehension even at speeds as slow as 240 milliseconds per word 
(Potter, Kroll, & Harris, 1980).  At this fast presentation rate, Kintsch and 
Mross found a 20 millisecond advantage for the inappropriate targets.  

    At the other extreme of sentence presenation rates, Van Petten and 
Kutas (1987) allotted 900 milliseconds to each word.  They found a large and 
significant effect of sentence context.  Part of this large effect was probably 
due to the strengths of their sentence contexts, but another part may have 
been due to the amount of time subjects had to develop and apply the sentence 

    The effect of a fast presentation rate could be produced by Kawamoto's 
model simply by reducing the size of the sentence context bias.  The reducted 
initial activation of the meaning units will then make a smaller impact on the 
access process.

     *Attention*  The focus of attention can also influence lexical access.  
Blutner and Sommer (1988) varied whether the ambiguous word occurred 
inside or outside the attentional focus of the sentence.  Focus was 
manipulated by first presenting a question that focused attention on a 
particular part of the sentence:

	Which scenery disappointed the visitors?
	The scenery with the mast disappointed the visitors in the gallery.
                                                            (footnote 2)

The lexical decision immediately following the ambiguous word was faster when 
the ambiguous word was in the sentence focus.  Attention did not affect the
contents of the lexical access, though there is a trend toward a larger 
context effect in the focus condition.  Rather, attention significantly affects
the speed of the process.  This effect could be modelled by universally reducingthe size of the weights in the network to simulate processing outside the focus of attention (cf. Cohen, Dunbar, & McClelland, 1990).  The smaller weights will 
slow and weaken processing.


     Small effects in any domain are difficult to verify.  Because they are 
small, they are occasionally obscured by variance.  This situation can create a 
body of conflicting results: some experiments find an effect while others do 
not.  But nonsignificant trends toward a significant difference should not be 
overlooked.  They may be caused by a small effect swimming in a sea of 
variance.  By aggregating across studies even subtle effects can be recovered.  
Such is the case with the effect of sentence context on the activation of 
ambiguous word meanings.  Nearly every experiment testing the Independent 
Access hypothesis has at least found a trend toward Biased Activation.  
Individually, this trend means nothing, but overall, it constitutes a strong 
affirmation of Biased Activation.  This result places sentence context among of 
set of factors, frequency, presentation rate, and attention, that influence 
lexical access.  Finally, this result favors an interactive activation account 
of lexical access over a modular account.


Blutner, R. & Sommer, R.  (1988).  Sentence processing and lexical access: 
The influence of the focus-identifying task.  Journal of Memory and 
Language, 27, 359-367.

Carpenter, P. A. & Daneman, M.  (1981).  Lexical retrieval and error recovery 
in reading:A model based on eye fixations.  Journal of Verbal Learning 
and Verbal Behavior, 20, 137-160.

Cohen, J. D., Dunbar, K., & McClelland, J. L.  (1990).  On the control of 
automatic processes: A parallel distributed processing account of the 
stroop effect.  Psychological Review, 97, 332-361.

Fisher, R. A. (1948).  Combining independent tests of significance.  American 
Statistician, 2, 30.

Fodor, J. A.  (1983).  The modularity of mind: An essay on faculty psychology.  
Cambridge, MA: MIT Press.

Just, M. A. & Carpenter, P. A. (1980).  A theory of reading: From eye 
fixations to comprehension.  Psychological Review, 87, 329-354.

Kawamoto, A. H.  (1988).  Distributed representations of ambiguous words 
and their resolution in a connectionist network.  In S. L. Small, G. W. 

Cottrell, and M. K. Tanenhaus (Eds.), Lexical ambiguity resolution: 
Perspective from psycholinguistics, neuropsychology, and artificial 
intelligence.  San Mateo, CA: Morgan Kaufmann.

Kintsch, W. & Mross, E. (1985).  Context effects in word identification.  
Journal of Memory and Language, 24, 336-349.

McClelland, J. L. (1987).  The case for interactionism in language processing.  
In M. Coltheart (Ed.), Attention and performance XII: The psychology 
of reading.  Hillsdale, NJ: Lawrence Erlbaum Associates.

Meyer, D. E. & Schvaneveldt, R. W. (1976).  Meaning, memory structure, and 
mental processes.  In C. Cofer (Ed.), The structure of human memory.  
San Francisco: Freeman.

Onifer, W. & Swinney, D. A.  (1981).  Accessing lexical ambiguities during 
sentence comprehension: Effects of frequency on meaning and 
contextual bias.  Memory and Cognition, 9, 225-236.

Potter, M. C., Kroll, J., & Harris, C.  (1980).  Comprehension and memory in 
rapid sequential reading.  In R. Nickerson (Ed.), Attention and 
performance VIII.  Hillsdale, NJ: Erlbaum.

Seidenberg, M. S., Tanenhaus, M. J., Leiman, J. M., & Bienkowski, M.  
(1982).  Automatic access of the meanings of ambiguous words in 
context: Some limitations of knowledge-based processing.  Cognitive 
Psychology, 14, 538-559.

Seidenberg, M. S., Waters, G. S., Sanders, M., & Langer, P.  (1984).  Pre- 
and postlexical loci of contextual effects on word recognition.  Memory 
and Cognition, 12, 315-328.

Simpson, G. B.  (1981).  Meaning dominance and semantic context in the 
processing of lexical ambiguity.  Journal of Verbal Learning and Verbal 
Behavior, 20, 120-136.

Simpson, G. B.  (1984).  Lexical ambiguity and its role in models of word 
recognition.  Psychological Bulletin, 96, 316-340.

Swinney, D. A.  (1979).  Lexical access during sentence comprehension: 
(Re)consideration of context effects.  Journal of Verbal Learning and 
Verbal Behavior, 18, 645-659.

Tabossi, P.  (1988a).  Accessing lexical ambiguity in different types of 
sentential contexts.  Journal of Memory and Language, 27, 324-340.

Tabossi, P.  (1988b).  Effects of context on the immediate interpretation of 
unambiguous nouns.  Journal of Experimental Psychology: Learning, 
Memory, and Cognition, 14, 153-162.

Till, R. E., Mross, E. F., & Kintsch, W.  (1988).  Time course of priming for 
associate and inference words in a discourse context.  Memory and 
Cognition, 16, 283-298.

Wolf, F. M.  (1986).  Meta-analysis: Quantitative methods for research 
synthesis.  Sage University Paper Series on Quantitative Application in 
the Social Sciences, series no. 07-001.  Beverly Hills, CA: Sage 


An additional study was found (Tabossi, 1988a), but it confounds the 
frequency of the meanings of the ambiguous words with the bias of the 
sentence context.  It could therefore not be used in the analysis.

In the original German, "mast" is ambiguous.

[CRL Newsletter Home Page]
[CRL Home Page]
Center for Research in Language
CRL Newsletter
Article 5-6