Journal of Memory and Language 46,245–266 (2002)doi:10.1006/jmla.2001.2810, available online at http://www.academicpress.com on
Making Sense of Semantic Ambiguity: Semantic Competition in Lexical Access
Jennifer Rodd
MRC Cognition and Brain Sciences Unit, Cambridge, United Kingdom
Gareth Gaskell
Department of Psychology, University of York, York, United Kingdom
William Marslen-Wilson
MRC Cognition and Brain Sciences Unit, Cambridge, United Kingdom
There have been several reports in the literature of faster visual lexical decisions to words that are semanticallyrelatedthis as-sesest that,antic
M
to Jthe the
ARodperibrid
ambiguous. All current models of this ambiguity advantage assume that it is the presence of multiple unmeanings that produce this benefit. A set of three lexical decision experiments reported here challenge sumption. We contrast the ambiguity seen in words like bark, which have multiple unrelated meanings, with wordthat have multiple related word senses (e.g., twist). In all three experiments we find that while multiple word sensdo produce faster responses, ambiguity between multiple meanings delays recognition. These results suggwhile competition between the multiple meanings of ambiguous words delays their recognition, the rich semrepresentations associated with words with many senses facilitate their recognition.© 2002 Elsevier Science (USA)
Key Words:lexical ambiguity; polysemy; distributed semantic representations.
any words are semantically ambiguous, anda u o
;d
itm
th
a coil or spiral, to operate by turning, to alter
c-d;
be-
enta-ty-
n-
e- to86;word
can refer to more than one concept. For exple, bark can refer either to a part of a tree orthe sound made by a dog. To understand swords, we must select one of these differentterpretations, normally on the basis of the ctext in which the word occurs.
Words can be ambiguous in different waysword like bark has two semantically unrelatemeanings, which seem to share the same wrand spoken form purely by chance. More comon than this type of accidental ambiguity is systematic ambiguity between related wosenses. For example, the word twist has a rangeof dictionary definitions including to make into
This research was funded by a MRC Studentship awa
24
ttho-o-me
the, a
. Rodd. Some of the research in this paper was reporteTwenty-first and Twenty-second Annual ConferencesCognitive Science Society (August 1999 and 2000).ddress correspondence and reprint requests to Jennd, Centre for Speech and Language, Department of mental Psychology, University of Cambridge, Camge, CB3 3EB, UK. E-mail: [email protected]
0749-596X/02 $35.005
m-toch
in-n-
a
ten-
erd
the shape of, to misconstrue the meaning of, towrench or sprain, and to squirm or writhe. Themeaning of this word varies systematically acording to the context in which the word is usefor example, there are important differences tween what it means to twist an anklecomparedwith to twist the truth. However, although themeaning of the word is ambiguous betwethese different interpretations, the interpretions are closely related to each other both emologically and semantically; this is quite ulike the ambiguity for a word like bark.
The linguistic literature makes a distinction btween these two types of ambiguity and refersthem as homonymy and polysemy (Cruse, 19Lyons, 1977, 1981). Homonyms, such as the tmeanings of bark, are considered to be differenwords that, by chance, share the same orgraphic and phonological form. Specifically, hmographs are different words that share the saorthographic form, while homophones share same phonological form. On the other hand
edd at
of
iferEx--
k.
© 2002 Elsevier Science (USA)All rights reserved.
D
asnth
rn.vti
tn
h n
ee
tws
nogiprge gn
hne
ue
at
meanings; the experiments reported here investi-
hat anin,kii-u-w-ible
inne
ndr- re-
calre
on,-the
u-ingon-nd
geion.gende).t istheis-ent
text,us
ualam-usdi-en anney,
246 RODD, GASKELL, AN
polysemous word like twist is considered to be single word that has more than one sense. Dethis linguistic distinction between homonymy apolysemy, psychologists have often used the terms interchangeably (see Klein and Murp(2001) for a discussion of this issue).
All standard dictionaries respect this distintion; lexicographers decide whether differeusages of the same orthography should cospond to different lexical entries, or differesenses within a single entry. Many criteria (eetymological, semantic, and syntactic) habeen suggested to operationalize this distincbetween senses and meanings. However, generally agreed that while the polysemy/hmonymy distinction appears easy to formulait can be difficult to apply with consistency areliability; people often disagree about whethtwo usages of a word are sufficiently related tthey should be considered to be senses of aysemous word rather than homonyms (Lyo1977, 1981; Kilgarriff, 1992). While there manot always be a clear distinction between thtwo types of ambiguity, it is important to rmember that words that are described as semtically ambiguous can vary between these extremes and that our mental representationthese two types of words are likely to be vedifferent.
Semantic ambiguity is very common in laguage, and our ability to understand ambiguwords is an important property of our languaprocessing system. Evidence about how ambity affects human language performance can vide important constraints on models of worecognition. In particular, models of word reconition have been required to accommodate dence that visual lexical decisions are fasterambiguous words. In this paper, we evaluateevidence of an ambiguity advantage in the liof the distinction between word meanings aword senses. In particular, we argue that wearlier studies in the literature show that semaambiguity can produce a processing advantagis not clear whether this is caused by ambigbetween multiple word meanings or betwemultiple word senses. Despite this, all currentcounts of the ambiguity advantage assume
the advantage is produced by multiple, unrelaMARSLEN-WILSON
t
pitedwoy
c-ntre-t
g.,e
iont iso-e,derat
pol-s,
yse-an-o ofry
-use
gu-ro-d-vi-forthehtd
iletic, it
itync-
hat
gate whether this assumption is correct.
THE AMBIGUITY ADVANTAGE
The ambiguity advantageis the finding thatvisual lexical decisions are faster for words tare semantically ambiguous. Early reports ofambiguity advantage came from RubensteGarfield, Millikan (1970) and Jastrzembs(1981), who found faster visual lexical decsions for ambiguous words than for unambigous words matched for overall frequency. Hoever, Gernsbacher (1984) discussed a possconfound between ambiguity and familiarity these experiments; words with more than omeaning are typically more familiar. She founo effect of ambiguity over and above familiaity. Since then, however, several papers haveported an ambiguity advantage in visual lexidecision experiments using stimuli that wecontrolled for familiarity (e.g., Azuma & VanOrden, 1997; Borowsky & Masson, 1996; Hin& Lupker, 1996; Kellas, Ferraro, & Simpso1988; Millis & Button, 1989; Pexman & Lupker, 1999). Although these studies vary in robustness of the effects they report, their cumlative weight has had the effect of establishthe ambiguity advantage as an important cstraint on theories of lexical representation alexical access.
Interestingly, a robust ambiguity advantahas only been observed using lexical decisFor word naming the ambiguity disadvantahas been very unreliable (see Borowsky aMasson (1996) for a discussion of this issuFurther, on a range of other tasks in which inecessary to disambiguate the meaning of ambiguous word, there is a clear ambiguity dadvantage. For example, when eye-movemmeasures are used for reading words in conand if the context is neutral and the ambiguoword has two meanings of approximately eqfrequency, then there is a disadvantage for biguous words compared with unambiguowords (see Rayner (1998) for a review). Adtionally, priming studies have shown that, evin an inappropriate context, both meanings ofambiguous word seem to be accessed (Swin
ed1979; Onifer & Swinney, 1981). Therefore, forA
coeue ire
atu
ne
he-
me
oiost
r
o
a
amn pfeo
istok
ce,alutedre-ss atic
icoci-on-o-
thetet of
rep-nd itro-ity of
avede-can
oor-ndam-, onate
-e itar-or-arnnd
pro-nticeenrenta-for
an-is-
thatan-
. In
SEMANTIC
an ambiguous word presented in a sentencetext it appears that both meanings of the ware initially activated and that this produclonger reading times for these words. This sgests that the ambiguity advantage may emonly in situations where it is not necessary totegrate the meaning of the word into a cohesemantic representation of a sentence. Forreason, many of the early accounts of the amguity advantage assumed that it is a presemeffect, and that participants are performing task without disambiguating the ambiguoword.
One interpretation of the ambiguity advatage has been that ambiguous words benfrom having multiple entries within the lexiconFor example, Kellas et al. (1988) assume twords are represented by individual nodwithin an inhibitory lexical network. They suggest that while the multiple nodes of an abiguous word do not inhibit each other, thboth act independently to inhibit all other competing entries, and this increased inhibitioncompetitors produces the faster recognittimes. A related account (e.g., Jastrzemb1981) assumes that the benefit arises frompresence of noise or probabilistic activation; bcause ambiguous words are assumed to hmultiple entries, they benefit from having mothan one competitor in the race for recognitioOn average, by a particular point in time, onethese competitors is more likely to have reachthe threshold for recognition than a word thhas only one entry in the race.
These accounts of the ambiguity advantpredict that the effect will be seen for any abiguous words whose meanings are sufficieunrelated that they correspond to separatetries in the mental lexicon; they make no scific predictions about what should happen words with multiple senses, as it is not clwhether related word senses would correspto separate entries within the network.
An alternative view of word recognition that words compete to activate a representaof their meaning. Several recent models of bspoken and visual word recognition have ta
this approach (Gaskell & Marslen-Wilson1997; Hinton & Shallice, 1991; Joordens &MBIGUITY 247
on-rdsg-rgen-nt
thisbi-ntiches
-fit
.ats
-y-fn
ki,hee-aveen.f
edt
ge-
tlyen-e-orarnd
ionth
en
Besner, 1994; Plaut, 1997; Plaut & Shalli1993). Rather than including localist lexicrepresentations, these models use distriblexical representations; each word is repsented as a unique pattern of activation acroset of orthographic/phonological and semanunits.
Within models of this type, the orthographpattern of an ambiguous word must be assated with multiple semantic patterns correspding to its different meanings. When the orthgraphic pattern is presented to the network,network will try to simultaneously instantiathe word’s two meanings across the same sesemantic units. These competing semantic resentations will interfere with each other, athis interference is likely to increase the timetakes for a stable pattern of activation to be pduced. At first glance, therefore, the ambiguadvantage is inconsistent with the predictionsthese models.
In response to this inconsistency, there hbeen several attempts to show, with varying grees of success, that this class of model show an advantage for ambiguous words. Jdens and Besner (1994) and Borowsky aMasson (1996) both suggest that because biguous words have more than one meaningaverage the randomly determined initial stwill be closer to a valid finishing state for ambiguous words, and this could reduce the timtakes for the network to settle. Kawamoto, Frar, and Kello (1994) suggested that if an errcorrecting learning algorithm was used to lethe mapping from orthography to semantics athen to compensate for the increased error duced by the ambiguous words in the semaunits, stronger connections are formed betwthe orthographic units. If lexical decisions amade on the basis of orthographic represetions, then this could improve performance ambiguous words.
These accounts of how the ambiguity advtage might arise from a model incorporating dtributed semantic representations all predict the effect should be strongest when the meings of the ambiguous words are unrelated
,the proximity advantage account of Joordensand Besner (1994) and Borowsky and MassonD
gaa h
tdb
bb eeseu
pwio o
awsard
a
eb
The-
ic-ts’ean-
denrastrds;ithed
offorly
This
-tici-his
eseelyrer-asythlese-
een
w-e- ahatth-fi-
248 RODD, GASKELL, AN
(1996), the benefit from having two meaninwill be maximal when the states of semantic tivation corresponding to the two meanings furthest apart, i.e., when the meanings aremantically unrelated. Similarly, according to tKawamoto et al. (1994), the ambiguity advatage is driven by the error produced during learning of the meanings of ambiguous worTherefore, the effect of ambiguity should greatest when this error is maximal, i.e., whthe meanings of the ambiguous words are higunrelated.
In summary, all current accounts of the amguity advantage assume that it is ambiguity tween unrelated meanings that produces thebiguity advantage. None of these modexplicitly predict what the effect of multiplword senses should be. For those modelwhich the benefit arises because of the presof multiple localist lexical entries for ambiguowords, the presence of a benefit for words wmultiple senses will depend on whether multisenses are represented as separate entries the network. Those models that involve distruted semantic representations predict that wwith multiple senses may show a processingvantage, but that this should be reduced cpared with words with multiple meanings.
In the following section we analyze in detthe stimuli used in previous studies that shorobust ambiguity advantage. This may help udetermine whether the assumption that the biguity advantage reflects a benefit for wothat have unrelated meanings is correct, andetermine whether multiple word senses malso play a role. In particular, we look in detat the stimuli used by Millis and Button (1989Azuma and Van Orden (1997), and Borowsand Masson (1996). These are three represtive studies which show robust effects of am
u-nd
en-ple,m-
aregiven in the Wordsmyth dictionary (t(46) 5 4.4,
1This particular dictionary was chosen because it reliablyseparates semantically unrelated meanings into distinct lexi-cal entries, but unlike some other dictionaries it does not re-quire that senses within an entry have the same syntacticclass. This reflects the intuitions of participants that mean-ings from different syntactic classes can be highly related
guity.
WORD SENSES AND WORD MEANINGS
As mentioned earlier, lexicographers rotinely distinguish between word meanings aword senses when they structure dictionarytries. These dictionary entries provide a simyet reliable way to classify words as being a
biguous between multiple meanings or betweMARSLEN-WILSON
sc-rese-en-hes.e
enhly
i-e-
am-ls
inncesithleithin
b-rdsad-m-
il a tom-
ds toayil),kynta-i-
multiple senses. We have used the entries inOnline Wordsmyth English Dictionary–Thesaurus (Parks, Ray, & Bland, 1998).1 As we re-port later, the classifications made in this dtionary correspond closely to participanjudgements about the relatedness of the mings of ambiguous words.
Looking first at the stimuli used by Millis anButton (1989) and Azuma and Van Ord(1997), neither study makes the direct contbetween ambiguous and unambiguous wowords with many meanings are compared wwords with few meanings. Words were assignto these groups by counting the numbermeanings that participants could provide each word. Crucially, both studies count highrelated word senses as separate meanings.can be demonstrated by example.
Millis and Button (1989) use tell as an example of a word that has many meanings. Parpants produced up to four meanings for tword. These were to inform, to explain, to un-derstand, and to relate in detail. Although thereare clearly important differences between thfour definitions, these differences are relativsubtle; all four definitions relate to a single comeaning of the word, to do with providing infomation. All these definitions are included senses within a single entry in the Wordsmdictionary. This is just one of several exampof high-ambiguity items that are ambiguous btween multiple word senses rather than betwunrelated word meanings.
We compared the groups of high- and loambiguity words in the two experiments rported by Millis and Button (1989) that foundsignificant ambiguity advantage, and found tthey do not differ in their number of Wordsmyentries (t(46) 5 .5, p . .6) (see Table 1). In contrast, the two groups of words did differ signicantly in the total number of senses they
en(Azuma, 1996).
ro
-nigr
a an
r
unieho v adoage
hisigh-areity
seforelyhec-s.9)aes
ant
yusra,ho noan-
d byerhe
nses
Ambiguous 1.8 12.1 8.8
p , .001), and in the number of dictionasenses of the dominant meaning of each w(t(46) 5 4.0, p , .001).2 Therefore, the highambiguity words used by Millis and Butto(1989) have more senses than the low-ambity words, but crucially, they do not have mounrelated meanings. This suggests that thebiguity advantage seen in this study shouldinterpreted as a benefit for words that have mrelated word senses, and not a benefit for ulated meanings.
Azuma and Van Orden (1997) also compawords with few (2–4) and many (6–10) meaings. Again, items were assigned to these groon the basis of the total number of meaniprovided by participants, and highly related dtionary senses were counted as separate mings. A different group of participants rated trelatedness of these meanings. For those wclassed as having unrelated meanings, therea benefit for those with many meanings othose with few meanings. However, it cannotassumed that these words only have unrelmeanings. The relatedness measure useAzuma and Van Orden (1997) was derived frthe relatedness of the words’ dominant meings with each of its subordinate meaninTherefore (as noted by the authors), this m
ure contains no information about the relate sefecte-
on be-igu-m-the
2For those words with only one entry in the dictionary, tdominant meaning was simply this single meaning. Fthose words with multiple entries in the dictionary, the doinant meaning was determined by asking a group of 38 pticipants to provide associates for each word and then seing the meaning for which the higher proportion oassociates were related. This procedure was used by Twet al. (1994) to produce dominance measures for ambiguwords.
yrd
u-em-
benyre-
edn-ps
gsc-an-erdswaserbeted bymn-s.as-d-
ness of the different subordinate meanings. Tmeans that we cannot be certain that the hambiguity words have more meanings that semantically unrelated than the low-ambiguwords.
Analysis of the dictionary entries for thestimuli shows a similar pattern to that seen the Millis and Button (1989) stimuli. First, thtwo groups of words did not differ significantin their number of dictionary entries. In fact thigh-ambiguity words have slightly fewer ditionary entries than the low-ambiguity wordSecond, as with the Millis and Button (198stimuli, the high-ambiguity words did have significantly higher total number of senswithin these entries (t(33) 5 4.6, p , .001) anda higher number of senses for the dominmeaning of each word (t(33) 5 3.2, p , .005)(see Table 1).
Finally, let us look at the stimuli used bBorowsky and Masson (1996). Their ambiguoand unambiguous words were taken from FeJoordens, Balota, Ferraro, Besner (1992), wasked participants to rate whether a word hadmeaning, one meaning, or more than one meing. This is the same procedure that was useKellas et al. (1988) and Hino and Lupk(1996). We chose to look in detail at tBorowsky and Masson (1996) stimuli becautheir result was the one of these where the efof ambiguity was statistically significant and bcause they used the largest set of words.
The stimuli used by Borowsky and Mass(1996) appear to provide a clear comparisontween words that people consider to be ambous and unambiguous. An analysis of the nuber of senses and meanings given in
heor
m-ar-
lect-f
illeyous
SEMANTIC AMBIGUITY 249
TABLE 1
Mean Number of Dictionary Entries and Senses for Stimuli
Stimulus group Dictionary entries Total senses Dominant meaning se
Millis & Button (1989) Few meanings 1.2 6.9 6.4Many meanings 1.3 12.8 11.5
Azuma & Van Orden (1997) Few meanings 1.9 9.6 7.2Many meanings 1.7 18.6 13.6
Borowsky & Masson (1996) Unambiguous 1.0 6.8 6.6
Wordsmyth Dictionary for the 128 words used
D
o
ii
g
ss
io
h
ie
h
l
htutehce
en
e
t isillif-
na-m-
s is
iplecialn-na-t it
ess-nte
ardssess, Tose
er
-esstor-
redre
ke-wre
in.1.p-
ave
oin
250 RODD, GASKELL, AN
in their experiment revealed that the two grouof words do differ significantly in their numbeof dictionary entries (t(126) 5 7.3, p , .001);the ambiguous words had, on average, mmeanings than the unambiguous words. Hoever the two groups of words also differed snificantly in their total number of senses withthese entries (t(126) 5 5.6, p , .001) and in thenumber of senses that the dominant meanineach word was given in the dictionary (t(126) 52.2, p , .05); again the high-ambiguity wordhad more word senses (see Table 1). It is poble that, as with Millis and Button (1989) anAzuma and Van Orden (1997), the ambiguadvantage shown by Borowsky and Mass(1996) may reflect an advantage for words wlarge clusters of related word senses.
In all three of these studies, the high-ambigity stimuli have more related word senses ththe low-ambiguity words. In contrast, only onof these studies showed a difference in the nuber of dictionary entries. This is surprising. TMillis and Button (1989) and Azuma and VaOrden (1997) studies defined high-ambiguwords as those for which participants could gerate many definitions. Therefore we might haexpected these words to be ambiguous botterms of number of senses and number of meings, and yet they seem to differ only in thenumber of word sense. Why is there a biastheir stimuli toward ambiguity between multipmeanings rather than multiple senses?
We believe that this bias reflects the fact tmultiple senses are simply more frequent in language than multiple meanings. This is sported by an analysis of the 4930 entries in Wordsmyth dictionary that have word-form frquencies of greater than 10 per million in tCELEX lexical database (Baayen, Piepenbro& Van Rijn, 1993). While only 7.4% of thesword-forms correspond to more than one enin the dictionary, 84% of the entries have mothan one sense. Further, 37% of the entries hfive or more senses. These figures show hcommon the systematic ambiguity betweword senses is, compared with the accideambiguity between unrelated meanings. Thefore, when words are selected for an experim
as being ambiguous, without a distinction bMARSLEN-WILSON
psr
rew-g-n
of
si-dtyn
tween word senses and word meanings, ilikely that a high proportion of these words wshow the more common ambiguity between dferent word senses. Importantly, this explation suggests that any experiment looking at abiguity without explicitly making the distinctionbetween word meanings and word senselikely to be influenced by this bias.
Overall, these analyses suggest that multsenses, and not multiple meanings, were cruin producing the ambiguity advantage. In cotrast, as described above, all current explations of the ambiguity advantage assume thais unrelated meanings that produce the procing benefit. We explore the potentially differe
ith
u-anem-entyn-ve inan-ir ine
athep-he-ek,
tryreaveowntalre-nt
effects of different types of ambiguity in ththree experiments reported below.
EXPERIMENT 1
In this first experiment, lexical decisions forlarge set of ambiguous and unambiguous woare analyzed using multiple regression analyto determine the effects of multiple meaningmultiple senses, and meaning relatedness.the extent that different effects emerge for thefactors, this would provide the basis for furthinvestigations.
Method
Participants. The participants were 25 members of the MRC Cognition and Brain SciencUnit subject panel. All had English as their firlanguage and had normal or corrected-to-nmal vision.
Stimuli and design. One hundred twenty-fouambiguous words were selected to be includin the experiment. One hundred thirteen wetaken from the Twilley, Dixon, Taylor, and Clar(1994) homograph norms. While most of the slected words had only two meanings, a fewords with three meanings were included whethe third meaning had a meaning probabilitythe Twilley et al. (1994) norms of less than 0The remaining 11 ambiguous words did not apear in the norms, but were considered to hsimilar properties.
For half these ambiguous words, the twmeanings correspond to separate entries
e-the Wordsmyth dictionary, and are therefore
A
ti
ouryi
ni-s
ish
d b
te
eontiRc
a 2og
ntwr
lee l
kerdso-
&
venef-ate 7-usve inar-as
re-
rtic-re-theas
ex-ave
hatngs, allthanersesureend
ing; aci-tesgfore-an-
SEMANTIC
ambiguousbetween two meanings accordingthe criteria used by lexicographers. The remaing ambiguous words were judged to have mthan one meaning by Twilley et al. (1994), bthese meanings corresponded to a single entthe Wordsmyth dictionary. For these words itnot clear whether their different interpretatioshould be classed as meanings or senses; theclusion will allow us to look for effects of the relatedness of the meanings of ambiguous word
Sixty unambiguous words were includedthe experiment.3 Only three of the unambiguouwords had more than one entry in tWordsmyth dictionary (frog, bus, prayer), thesecond entry for these three was considerebe sufficiently obscure that the words couldconsidered unambiguous.
All the stimuli were pretested for concreness and familiarity, variables that are knowninfluence visual lexical decisions. Relatednratings and dominance ratings were also tained for the ambiguous words. These ratiwere made by four separate groups of parpants who were either members of the MCognition and Brain Sciences Unit subjepanel or students at Cambridge University. Eof these ratings was made by a minimum ofparticipants. The three variables were rated 7-point scale as used by Gilhooly and Lo(1980).
For the concreteness ratings, participamade separate ratings for each of the meanings of the ambiguous words. The woappeared in a rating booklet together withword associate. This associate made it cwhich meaning of the word was to be ratFor each word, an associate was selectedeach of its two meanings (for examp
bark–dog and bark–tree). Word associates w eng, a
retedeibu-eding
3The imbalance between the number of unambiguowords and ambiguous words reflects the fact that this expiment was also designed to investigate the effects of the ative concreteness of the two meanings using a design slar to that used by Rubenstein et al. (1970). However tanalyses of the differences between the groups showedeffects of relative concreteness, but simply revealed a meffect of concreteness and so are not reported here. The proportion of ambiguous words also increases the setwords used in the analysis of meaning relatedness.
MBIGUITY 251
on-retin
ssr in-
.n
e
toe
-tossb-gsci-Ct
ch4
n aie
tso
dsaar
d.foreere
also given for the unambiguous words to mathe procedure consistent for the two wotypes. All the associates were taken from asciation norms (Twilley et al., 1994; Moss Older, 1996).
For the relatedness ratings, raters were gieach ambiguous word, together with short dinitions of its two meanings, and asked to rhow related the two meanings were on apoint scale. Eleven additional ambiguowords that according to Azuma (1996) hahighly related meanings were also includedthe booklet. These were included to help pticipants use the whole range of ratings, most of the ambiguous words had highly unlated meanings.
The mean relatedness rating across all paipants and items was 2.64. This low value flects the fact that participants saw many of pairs as completely unrelated; a rating of 1 wused more than any other rating. This is pected as these words were all selected to hmeanings that are sufficiently unrelated tthey should be considered separate meaniand not senses within a single meaning. Inanalyses of the response time data, rather using the mean relatedness ratings, the invof these values was used; this made the meamore sensitive to small changes at the lower of the scale.
The dominance scores were derived usthe procedure used by Twilley et al. (1994)group of 38 participants provided word assoates for each word, and then the associawere classified in terms of which meaninthey were related. The dominance score each word was the proportion of valid rsponses that correspond to its dominant meing. For example, a score of 1 would be givto a word with a highly dominant meaninwhile a balanced homograph would havescore of 0.5.
The summary statistics of the stimuli agiven in Table 2; the words themselves are lisin Appendix A. The nonword stimuli werpseudohomophones and had a similar distrtion of lengths to the word stimuli. We decidto use pseudohomophones following the find
user-rel-imi-he noainhigh of
of Azuma and Van Orden (1997), who found aD
hoil
ecmmvoin
ticndwstreA
alultio5sd orthh,
oizis
achn-
ed re-vid- not
hisre- byer-eseed.
dif-s-
rdses-
rdxi-ere
re-
is.ty;orehisy any
senses. All the other predictor variables exceptword length also accounted for unique variabil-
Neighborhood 3.55 4.84
significant effect of ambiguity using pseudomophone nonword foils but not when the fowere word-like nonwords.4
Procedure. All the stimulus items werpseudorandomly divided into four lists, suthat each list contained approximately the sanumber of words of each stimulus type. Soitems were then swapped between lists, to ahaving any ambiguous word occurring withthe same list as an item that might bias parpants toward one of its meanings. Participawere presented with the four lists in a pseurandom order such that each possible orderseen by at least one participant. Within the lithe order in which stimulus items were psented was randomized for each participant.the participants saw all of the stimulus materi
For each of the word and nonword stimthe participants were presented with a fixapoint in the center of a computer screen for ms, followed by the stimulus item. Their tawas to decide whether each item was a wora nonword; recognition was signaled with dominant hand, nonrecognition with the othand. As soon as the participant respondedword was replaced with a new fixation point.
A practice run, consisting of 30 items nused in the analysis, was given to familiarparticipants with the task. Each of the four l
was presented in a separate block of trials. P4We have repeated Experiment 1 using word-like nowords. Consistent with the findings of Azuma and VaOrden (1997) and Pexman and Lupker (1999) we foundsignificant, but reduced, effect of ambiguity that was cosistent with these results. The effect of relatedness wassignificant.
MARSLEN-WILSON
-s
heeid
i-tso-ass,-ll
s.i,n
00k
ticipants were given a short break after eblock. Each block began with five stimuli not icluded in the analysis.
Results
The data from two participants were removfrom the analysis because they had meansponse latencies greater than 1000 ms. Indiual responses longer than 1200 ms were alsoincluded in the analysis; for the word data tmeant that 1.2% of the data points were moved from the analysis. As recommendedRatcliff (1993), all analyses were also pformed on the inverse response times; for thanalyses, all correct responses were includThese analyses are only reported where theyfered in significance levels from the untranformed data.
Multiple Regression Analyses
The response time data for all 184 wowere entered in a simultaneous multiple regrsion analysis. Ambiguity, number of wosenses, word frequency, familiarity, length, lecal neighborhood, and mean concreteness wall included as predictors. A summary of the gression analysis can be seen in Table 3.
There are two crucial results in this analysFirst, there was a significant effect of ambiguiambiguous words were responded to mslowly than unambiguous words. Second, tambiguity disadvantage was accompanied bsignificant benefit for words that have ma
252 RODD, GASKELL, AN
TABLE 2
Experiment 1: Descriptive Statistics for Stimuli
Unambiguous Ambiguous
N 60 124Length 5.07 5.00Log frequency 5.26 5.49Familiarity 3.91 3.98Word senses 2.83 7.43Concreteness, meaning 1 5.13 5.45Concreteness, meaning 2 — 3.91Mean concreteness 5.13 4.68
eerthe
tetsar-
n-na
n-
TABLE 3
Experiment 1: Summary of Regression Analysisfor All Words
Predictor variable Standardized coefficient t
Ambiguity .18 2.7**Word senses 2.17 22.4*Frequency 2.29 23.2**Familiarity 2.26 22.9**Length .13 1.6Neighborhood .34 4.3***Concreteness 2.30 25.0***
notNote. df 5 177, (*)p , .1, *p , .05, **p , .01, ***p ,
.001.
an
on e
orrees
cded
o
ho- se-o-byess3.4.hatcy,t issi-
venp-
4.4un-rre- thewoent byersrdries. weeyisen the
ted onity.bledixad-
s inn
s ofN-
Note. df 5 117, (*)p , .1, *p , .05, **p , .01, ***p ,
.001.
ity in response times (the effect of length wmarginal in the analysis of the inverse respotimes (p , .1).
A second simultaneous multiple regressianalysis was then carried out on the respotimes for the 124 ambiguous words in orderlook for an effect of the relatedness of thmeanings. Dominance scores were also entein this analysis. A summary of the regressianalysis can be seen in Table 4. As in the eamultiple regression analysis of this data, fquency, lexical neighborhood and concretenaccounted for unique variability in respontimes. The effects of word length and numberword senses were marginal, and familiarity dnot account for any unique variance in thsmaller set of words. The effect of dominanwas also not significant. Importantly, relateness did account for unique variability in rsponse times; ambiguous words were responto faster when their meanings were judged tosemantically related.
Analyses of Variance
To provide further evidence for the effects
Nonhomographs 43 5.43
MBIGUITY 253
sse
nse
toirredn
lier-sseofidise--ed
be
f
were also performed. From the set of 124 mographs, two sets of 27 homographs werelected, containing related and unrelated homgraphs respectively. They were selected using only those homographs with a relatednscore of either less than 1.9 or greater than A few homographs were then removed so tthe two groups were matched for frequenmean concreteness, length and familiarity. Iworth noting that even the homographs clasfied as having related meanings were not giparticularly high relatedness ratings (see Apendix B); the mean relatedness score wason the 7-point scale. Twenty-three of the 27 related homographs had two meanings cosponding to separate Wordsmyth entries, forrelated homographs only three words had tentries. This shows a high level of agreembetween the relatedness judgements madeparticipants and the decisions of lexicographabout whether different usages of a woshould be classed as separate dictionary entIn a separate analysis, not reported here,grouped the words according to whether thhave one or two entries in Wordsmyth; thshowed a very similar pattern of results to whthe relatedness rating were used to classifywords.
A group of 43 nonhomographs was selecto be matched to the two homograph groupsfrequency, concreteness, length, and familiarThe properties of these words are given in Ta5; the words themselves are given in AppenB. Although the groups were not matched in vance for neighborhood size (N; Colheart, Dav-elaar, Jonasson, & Besner, 1977) the wordthe three groups did not significantly differ othis variable; F2(2,94) 5 1.38, p 5 .2.
The response times for these three groupwords were submitted to separate ANOVA/A
SEMANTIC A
TABLE 4
Experiment 1: Summary of Regression Analysis forAmbiguous Words
Predictor variable Standardized coefficient t
Frequency 2.45 23.9***Familiarity 2.10 2.8Length .18 1.9(*)Neighborhood .30 3.0**Concreteness 2.21 22.8**Word senses 2.14 21.7(*)Relatedness .17 2.3*Dominance 2.10 21.4
as
ood
ambiguity and relatedness, ANOVA/ANCOVAsCOVA analyses, with items and participants
TABLE 5
Experiment 1: Descriptive Statistics for Groups of Related and Unrelated Homographs
Group N Relatedness Log frequency Concreteness Length Familiarity Neighborh
Unrelated 27 1.37 5.46 4.85 5.04 4.03 5.04Related 27 4.39 5.44 4.80 5.00 3.94 4.52
4.84 5.02 3.99 3.16
D
im
o
ea
ur
uath
-rsm
n-tsa
rthg
hg
o
d athtlyinityam-e-ism-levi-
uldlelly,hatist ite
tesan
Orden (1997). The implications of this result
oftsi-lesultf
hatalte
he1,hed
254 RODD, GASKELL, AN
the random variables. The mean response tare given in Table 6.
In the participants analysis, the effect group was significant; F1(2,44) 5 4.79, p , .05.In the items analysis, using the log-transformfrequency, familiarity, mean concreteness, length as covariates, the effect of group wmarginal; F2(2,90) 5 2.88, p , .07. Multiplecomparisons were made between the individgroups, using the Newman–Keuls proceduResponses to the group of nonhomographs wfaster than to the group of homographs with related meanings; this difference was significin the participants analysis and marginal in items analysis; q1(3,44) 5 4.15, p , .05,q2(3,90) 5 3.22, p , .07. The related homographs were significantly faster than the unlated homographs in the participants analyagain, the difference was marginal in the iteanalysis; q1(2,44) 5 3.28, p , .05, q2(2,90) 52.53, p , .08. The difference between the nohomographs and the homographs with relameanings was nonsignificant in both analy(p . .5). The error data showed no significeffect of group in either analysis; F1(2,44) , 1,F2(2,92) , 1.
These results confirm the findings of the gression analysis; homographs with relameanings are responded to more quickly thomographs with highly unrelated meaninFurther, they show that homographs are sponded to more slowly than matched nonmographs only when their meanings are judto be unrelated.
Discussion
Three interesting results have emerged fr
this experiment. First, the analysis of the re)
Nonwords 636 155 7.14
MARSLEN-WILSON
es
f
dndas
ale.eren-nte
e-is;s
sponse times for this set of 184 words showesignificant ambiguity disadvantage; words wione meaning were responded to significanfaster than words with two meanings. This iscontrast with previous reports of an ambiguadvantage, and suggests that recognition ofbiguous words is delayed by competition btween their different meanings. Second, thdisadvantage for multiple meanings was accopanied by an advantage for words with multipsenses. This confirms our suggestion that preous reports of an ambiguity advantage shobe interpreted as an advantage for multipsenses rather than multiple meanings. Finathe significant effect of relatedness shows tthe disadvantage for semantic ambiguitymodulated by meaning relatedness, such thais maximal when the different meanings of thword are semantically unrelated; this replicathe effect of relatedness seen by Azuma and V
-esstor-
TABLE 6
Experiment 1: Mean Lexical Decision Times, AnalysisUsing Relatedness Ratings
RT (ms)
Ambiguity Relatedness Mean SD Error (%
Unambiguous 556 133 4.25Ambiguous Unrelated 577 136 4.83Ambiguous Related 561 134 3.54
m-edes nt
e-edans.re-o-ed
m-
will be discussed in the General Discussion.
EXPERIMENT 2
Experiment 1 suggests that the two typessemantic ambiguity have very different effecon lexical decision performance. While multple meanings delay recognition, multipsenses produce a processing benefit. This reis clearly controversial; all existing models othe ambiguity advantage have assumed tmultiple meanings produce faster visual lexicdecisions. Experiment 2 attempts to replicathe contrasting effects of ambiguity seen in tmultiple regression analysis of Experiment using a factorial design to directly compare teffects of lexical ambiguity and multiple worsenses.
Method
Participants. The participants were 25 members of the MRC Cognition and Brain SciencUnit subject panel. All had English as their firlanguage, and had normal or corrected-to-nmal vision.
Stimuli and design. The word stimuli were se-lected to conform to a 2 3 2 factorial designwhere the two factors were ambiguity and nu
ber of senses. Words were classed as being un-A
hniee
o ar a
he
ebeikn
er
fothn
othag
ncricuod
terfi-
zed ino-ord- of
o-n-rdsereanylistardtederetici-s
uli,tion500 asre-ts
ters ases.
Real words were signaled with the dominant
SEMANTIC
ambiguous if they had only one entry in tWordsmyth Dictionary (Parks et al., 1998) aas ambiguous if they had two or more entrTwo measures of the number of senses wused. These were the total number of wsenses listed in the Wordsmyth dictionary forthe entries for that word, and the total numbesenses given in the WordNet lexical datab(Fellbaum, 1998).
Thirty-two words were selected to fill eaccell of the factorial design such that the numbof word meanings was matched across each lof number of word senses, and the total numof word senses was matched across each levthe number of word meanings. Therefore, unlExperiment 1, the numbers of ambiguous aunambiguous words used in this experimwere equal. Of the words used in this expement, 16% were also used in Experiment 1.
The four groups of words were matched CELEX frequency (log- transformed), leng(number of letters), number of syllables, cocreteness and familiarity. Concreteness scwere obtained from a rating pretest in which words were rated on a 7-point scale by 25 pticipants who were members of the MRC Conition and Brain Sciences Unit subject paand who did not participate in the lexical desion experiment. The familiarity ratings wemade on a similar 7-point scale by 23 partpants from the same population. The growere not explicitly matched for neighborho
size; however, the number of words in CELENeighbors 6.03
MBIGUITY 255
eds.re
rdllofse
rvelerl ofed
nti-
r
-reser--
eli-ei-ps
that differed from each word by only one let(N; Coltheart et al., 1977) did not differ signicantly between the groups (p . .3).
The properties of the words are summariin Table 7; the words themselves are listedAppendix B. The nonword stimuli were pseudhomophones. They were chosen to be as wlike as possible and had a similar distributionlengths to the word stimuli.
Procedure. The stimulus items were pseudrandomly divided into four lists; each list cotained approximately the same number of wofrom each stimulus group. Some items wthen swapped between lists, to avoid havingambiguous word occurring within the sameas an item that might bias participants towone of its meanings. Participants were presenwith the four lists in a random order. Within thlists, the order in which stimulus items wepresented was also randomized for each parpant. All participants saw all of the stimulumaterials.
For each of the word and nonword stimthe participants were presented with a fixapoint in the centre of a computer screen for ms, followed by the stimulus item. As soonthe participant responded, the word was placed with a new fixation point. Participanwere told to decide whether each string of letwas a real English word, and to respondquickly as possible without making mistak
Xhand, nonwords with the other hand.
es
TABLE 7
Experiment 2: Descriptive Statistics for Stimuli
Ambiguous Ambiguous Unambiguous Unambiguousfew senses many senses few senses many sens
Example Pupil Slip Cage MaskN 32 32 32 32Wordsmyth meanings 2.03 2.09 1.00 1.00Wordsmyth senses 5.19 14.22 5.25 14.41WordNet senses 4.88 11.84 5.00 11.19Frequency 5.40 5.43 5.43 5.50Concreteness 5.19 5.07 5.06 5.05Familiarity 4.11 4.24 4.17 4.24Length 4.47 4.41 4.47 4.53Syllables 1.19 1.09 1.16 1.09
7.78 5.91 6.25
D
nizsPan
e a gth1al
ho
teae
o
be
as
refewto 6ghi-
e-ar-
he
two
fectade
thehed
nyr er-tesper-wasrdse
thekind of advantage for words with multiple mean-
256 RODD, GASKELL, AN
A practice session, consisting of 64 items used in the analysis, was given to familiarparticipants with the task. Each of the four liwas presented in a separate block of trials. ticipants were given a short break after eblock. Each block began with 10 stimuli not icluded in the analysis.
Results
The data from two participants were removfrom the analysis because of error ratesgreater than 10%. Incorrect responses wereincluded in the analysis. The overall error rfor responses was 3.6% (ranging from 0.87.7% for each participant). Responses lonthan 1200 ms were also not included in analysis; for the word data this meant that 1.of the data points were removed from the ansis. As with Experiment 1, all analyses were aperformed on the inverse response times; these analyses all correct responses werecluded. These analyses are reported only wthey differ in significance from the analysis the untransformed data.
Mean values were calculated separaacross participants and items. The participmeans were subjected to ANOVA, and the itmeans were subjected to ANCOVA. The meresponse times are given in Table 8.
The ANCOVA revealed significant effects familiarity (F(1,121) 5 11.4, p , .001) andmarginal effects of length (F(1,121) 5 3.65, p , .06) and frequency (F(1,121) 5 2.72, p 5.1). The effects of concreteness and neighhood were nonsignificant (p . .5), and so thes
variables were removed from the ANCOVA.a
)
Nonwords 659 143 3.92
MARSLEN-WILSON
otetsar-ch-
dofnottetoere%ly-soforin-enf
lyntman
f
The main effect of the number of senses wsignificant (F1(1,22) 5 14.6, p , .001; F2(1,121)5 4.42, p , .05); words with many senses weresponded to 14 ms faster than words with senses. Ambiguous words were responded ms slower than unambiguous words, althouthis effect of ambiguity was not significant in ether analysis (F1(1,22) 5 2.9, p . .1, F2(1,121)5 1.3, p . .2). In the analysis of inverse rsponse times, the effect of ambiguity was mginal in the participants analysis (F1(1,22) 5 3.8,p , .07), but was again not significant in titems analysis (F2(1,121) 5 1.5, p . .2). Therewas no significant interaction between these variables in either analysis (p . .2).
The error data also showed a significant efof the number of senses; fewer errors were mto words with many senses (F1(1,22) 5 12.2, p , .005; F2(1,121) 5 5.19, p , .05). In theerror data neither the effect of ambiguity nor interaction between the two variables reacsignificance (all p . .4).
Discussion
This experiment shows that words with masenses are responded to faster and with fewerors than words with few senses. This replicathe significant word senses benefit seen in Eximent 1. This advantage for multiple senses shown alongside a small disadvantage for wowith multiple meanings. Although this differencwas not significant, there was no indication of
rdrenttyad-ter
therlsome
inof
TABLE 8
Experiment 2: Mean Response Times (RT) and PercentError Rates
RT (ms)
Ambiguity Senses Mean SD Error (%
Ambiguous Few 587 143 4.08Ambiguous Many 578 135 1.77Unambiguous Few 586 141 2.99Unambiguous Many 567 129 1.63
ordor-ings that has previously been reported.
EXPERIMENT 3
Experiments 1 and 2 have shown that wosenses and word meanings have very diffeeffects on visual lexical decisions. Ambiguibetween multiple meanings produces a disvantage, while multiple senses produce fasresponses. This experiment investigates whethese contrasting effects of ambiguity are apresent in the auditory domain, using the safactorial design as Experiment 2.
We argue that the ambiguity effects seenExperiments 1 and 2 reflect the influence amodal semantic representations on visual w
ge
recognition. If this is the case, then it is of inter-
A
enwolennosu)toe
isinkeaee
itdpm
o oeh
cd
ayohro
dif-
at-istheeri-edsefor im-g- of
ntn-ut
u-sh
-
enteri- fill of inno-
forerord inessheds
ed inedilar
lio-iteme to
SEMANTIC
est to see whether the same pattern emergthe auditory domain. It is possible that semainformation plays a similar role across the tdomains, but it is also possible that the tempcharacteristics of speech may reduce the rosemantic information on spoken word recogtion compared with visual word recognitioHowever, we do know that the semantic infmation relating to spoken words is accesearly on, prior to the word becoming uniq(Marslen-Wilson, 1987; Zwitserlood, 1989and this makes it at least possible that audilexical decision will show an influence of smantic ambiguity.
A study by Holley-Wilcox (1977) (cited in &McCusker, Hillinger, Bias, 1981) supports thidea that it is possible to detect effects ofmantic ambiguity using auditory lexical decsion. They found that auditory lexical decisiowere significantly slower for homophones liplane and plain, which although sharing thsame phonology are spelled differently, thfor nonhomophones. This result can beplained by assuming that competition betwethe different meanings of the homophonesslowing the recognition. This is consistent wthe ambiguity disadvantage seen the visualmain in Experiment 1 and suggests that comtition between the different meanings of abiguous words does play a role in the auditodomain.
However, a possible problem with using hmophones that are not homographs to looksemantic ambiguity effects is that there mayinterference between their different orthgraphic representations. Although it may seunlikely that interference between orthograprepresentations should affect an auditory tathis idea is supported by Ziegler and Ferra(1998). They found slower auditory lexical desions for words whose rimes could be spellemore than one way (e.g., sleep). This raises thepossibility that orthographic interference mhave contributed to the finding of HolleWilcox (1977) and makes it preferable to avsuch items in any experiment designed to sthe effects of semantic ambiguity. TherefoExperiment 3 uses the same ambiguous w
as the visual experiment, words like bark thatMBIGUITY 257
s inticoral ofi-.r-ede,ry-
se--se
nx-nisho-e--
ry
-forbe-micsk,ndi- in
y-idowe,rds
share both orthography and phonology, and fer only in their meanings.
As well as allowing us to investigate the ptern of ambiguity in the auditory domain, thexperiment also allows us to check whether pseudohomophone nonwords used in Expment 2 were crucial to obtaining the observpattern of results. It is not yet clear how thenonwords affect lexical processing, and so us to argue that these ambiguity effects haveportant implications for models of word reconition, they should be shown in the absencepseudohomophones.
However, the primary aim of this experimeis to try and replicate the ambiguity disadvatage, which was significant in Experiment 1, bnot in Experiment 2.
Method
Participants. The participants were 26 stdents at Cambridge University. All had Englias their first language.
Stimuli and design. The word stimuli were selected to conform to the same 2 3 2 factorial de-sign as in Experiment 2. Seventy-seven percof the words selected were also used in Expment 2. Twenty-three words were selected toeach cell of the factorial design; the numberwords in each cell is smaller than that usedExperiment 2 because of the additional phological constraints used to match the groups.
The four groups of words were matched CELEX frequency (log-transformed), numbof phonemes, the phoneme at which the wbecomes unique, actual length of the wordsms, concreteness and familiarity. Concretenand familiarity scores were taken from tpretest described in Experiment 2. All worhad only one syllable.
The properties of the words are summarizin Table 9; the words themselves are listedAppendix C. The nonword stimuli were creatto be as word-like as possible, and had a simdistribution of lengths to the word stimuli.
Procedure. The organisation of the stimuinto four blocks of trials followed the same prcedure as Experiment 2. The onset of each was 1000 ms after the participants’ respons
the preceding item. If the participant did not re-thtogitl
h
eade1thnr
yem
teteath
o
e
ned
er
reith
lso
-ousari- but
re-rorsughb-sis;
igu- al-heifi-
s ses
Uniqueness 3.70 3.87 3.78 3.74
ge
spond within 3000 ms of the onset of a word,next item was presented. Participants were to decide whether each sound was a real Enword and to respond as quickly as possible wout making mistakes. Real words were signawith the dominant hand, nonwords with tother hand.
Results
The data from four participants were removfrom the analysis because of error rates of grethan 10%. Incorrect responses were not incluin the analysis. The overall error rate for rsponses was 5.8%. Responses longer thanms were also not included in the analysis; forword data this meant that 2% of the data poiwere removed from the analysis. As with Expements 1 and 2, all analyses were also performon the inverse response times; for these analall correct responses were included. Thanalyses did not differ in significance levels frothe untransformed data and so are not repor
Mean values were calculated separaacross participants and items. The participmeans were subjected to an ANOVA, and item means were subjected to an ANCOVA. Tmean response times are given in Table 10.
The ANCOVA revealed significant effects familiarity (F(1,86) 5 4.6, p , 0.05) and length(F(1,86) 5 236, p , 0.001). Concreteness, frquency, number of phonemes, and uniquenpoint were not significant predictors of respotimes (p . .2); these variables were not includ
in the ANCOVA.eld
lishh-ede
dtered-500etsi-edsesse
There was a significant effect of the numbof senses (F1(1,21) 5 20.7, p , .001; F2(1,86)5 6.6, p , .05). Words with many senses weresponded to 33 msec faster than words wfew senses. The effect of ambiguity was asignificant (F1(1,21) 5 22.4, p , .001; F2(1,86)5 4.7, p , .005). Ambiguous words were responded to 29 msec slower than unambiguwords. The interaction between these two vables was significant in the subjects analysisnot in the items analysis F1(1,21) 5 16.5, p ,.001; F2(1,86) 5 2.3, p . .1.
The error data showed a similar pattern ofsults to the response time data. Fewer erwere made to words with many senses, althothis difference was only significant in the sujects analysis and marginal in the items analy(F1(1,21) 5 10.5, p , .005; F2(1,86) 5 2.7, p ,.1). Fewer errors were also made to unambous words compared with ambiguous words,though this difference was only marginal in t
258 RODD, GASKELL, AND MARSLEN-WILSON
TABLE 9
Experiment 3: Descriptive Statistics for Stimuli
Ambiguous Ambiguous Unambiguous Unambiguoufew senses many senses few senses many sen
N 23 23 23 23Wordsmyth meanings 2.04 2.13 1.00 1.00Wordsymth senses 5.43 13.61 3.59 14.00WordNet senses 5.00 11.43 4.43 10.17Frequency 5.30 5.34 5.42 5.43Concreteness 5.11 5.01 5.08 4.99Familiarity 4.17 4.30 4.31 4.33Length 610 602 601 605Phonemes 3.43 3.43 3.52 3.56
d.lynt
hee
f
-essse
subjects analysis and did not approach sign
TABLE 10
Experiment 3: Mean Response Times (RT) and PercentaError Rates
RT (ms)
Ambiguity Senses Mean SD Error (%)
Ambiguous Few 986 176 8.3Ambiguous Many 935 174 4.3Unambiguous Few 939 167 5.7Unambiguous Many 924 186 3.6
Nonwords 1031 173 6.0A
t
s ilehlttaeuo
an-om-al-be
re-forons&aut,ar-the
uldgu-nerro-
nalntic
SEMANTIC
cance in the items analysis; (F1(1,21) 5 4.2, p , .06; F2(1,86) 5 0.7, p . .4). The interactionbetween the two variables was not significaneither analysis (p . .4).
Discussion
This experiment has shown that the patternambiguity effects in the auditory domain is esentially the same as in the visual domain;advantage for words with many senses coexwith a disadvantage for words with multipmeanings. A second important feature of texperiment is that it shows the effects of muple senses and multiple meanings without use of pseudohomophones. This demonstrthat these nonwords are not necessary to sepattern of results seen in Experiment 2 and sgests that the ambiguity effects we have dem
so
th
g
utsnsti
aidn
uch
re-chnticageeti-ownast theand
ac-rdshinowad-ousti-ler-
om-plee,rds
ansm-
un-e-ce
strated are pervasive in word recognition.
GENERAL DISCUSSION
The results of these three experiments repsent an important challenge to accepted viewhow semantic ambiguity affects recognition isolated words. Previous reports of an ambiguadvantage have been interpreted as showingthere is a processing advantage for words have multiple meanings. A range of models hbeen put forward to explain how this advantamight arise.
Our analyses of the stimuli used in threethe clearest demonstrations of this effect sgested to us that the accepted interpretamight be incorrect and that related word senand not unrelated meanings might be respoble for this processing advantage. The resultthe three experiments reported here support view. In all three experiments we found a signicant benefit for words that have many senscompared with words with few senses. In cotrast, ambiguity between unrelated meaninconsistently produced a processing disadvtage; this ambiguity disadvantage was signcant in Experiments 1 and 3. We now consithe implications for models of word recognitio
The Ambiguity Disadvantage
We have already discussed how models
word recognition have tried to explain the apMBIGUITY 259
in
of-
ansts
isi-hetes theg-n-
re- off
itythatat
ase
ofg-ionessi- ofhisf-es,n-gsn-
fi-er.
of
parent advantage for words with multiple meings, but our data suggest that they must accmodate exactly the reverse effect. This chlenge is less problematic than might expected.
The ambiguity disadvantage is a natural pdiction of models in which words compete the activation of semantic representati(Gaskell & Marslen-Wilson, 1997; Hinton Shallice, 1991; Joordens & Besner, 1994; Pl1997; Plaut & Shallice, 1993). As discussed elier, in these models interference between different meanings of ambiguous words wodelay their recognition relative to an unambious word. As noted by Joordens & Bes(1994), an ambiguity advantage can be pduced by these models only if an additiomechanism is present to overcome this semacompetition. Our results suggest that no smechanism is required.
The ambiguity disadvantage reported heremoves a major hurdle for models in whiwords compete to activate distributed semarepresentations. The ambiguity disadvantnaturally emerges from the semantic comption present in such models and has been shin a model of this type where a simple lemean square algorithm was used to learnmapping between distributed orthographic semantic representations (Rodd, 2000).
This new pattern of results can also becommodated by those models in which wocompete to activate abstract word nodes wita lexical network. Earlier, we discussed hthese models could produce an ambiguityvantage by assuming either that ambiguwords are more efficient at inhibiting competors, or that they benefit from having multipcompetitors in the race for recognition. Inteestingly, these models can just as easily accmodate a disadvantage for words with multimeanings. As in all experiments of this typthe ambiguous words and unambiguous wowere matched on total frequency. This methat the frequency of each meaning of the abiguous words is on average half that of theambiguous word. This difference in the frquency of the word meanings could produ
-faster lexical decisions for the unambiguousD
s-at
rd-
rg
ewetdeteve
th adulorh
nn
toaes rcathin
efteewentnisinep
er,Inisesor-onheeret itro-bi-
thispat-t ofncethereser-ge,in algo-enen-
re-fect, thatticord
ndrdsithun-
ul-van-pleceptfer-igu-be-tiver-
toob- to that
260 RODD, GASKELL, AN
words. Further, if lateral inhibition were preent between all word nodes within the lexicnetwork, including the nodes correspondingthe different meanings of an ambiguous wothis would act to slow the recognition of ambiguous words.
Therefore, it appears that both classesmodels considered here can accommodateambiguity disadvantage. The question thatmains is whether the ambiguity disadvantashould simply be explained in terms of an effof frequency of word meanings, or whether can claim that it provides evidence of comption between the different meanings of worWe suggested earlier that looking for an effof meaning relatedness might help us to demine the mechanisms by which any obserambiguity effects might arise.
In Experiment 1 we found that, at least in visual domain, the ambiguity disadvantagemodulated by the relatedness of the two meings of the ambiguous words; within the worthat we classified as ambiguous between mple meanings, there was a benefit for those wwhose meanings were moderately related. Tsuggests that the ambiguity disadvantage cabe explained solely as the results of a frequebias; this account cannot allow semantic facto modulate the size of the ambiguity disadvtage. Similarly, the relatedness effect suggthat the effect cannot be explained entirely assulting from lateral inhibition between abstraword nodes; if the effect was entirely presemtic, there would be no mechanism by which semantic relationship between the two meanof a word could play a role.
The only way to explain the relatednessfect in a nonsemantic way is to assume thasufficient number of the words that we classas ambiguous between different meanings, win fact ambiguous between multiple senses;think that this is unlikely. Therefore, we believthat the modulation of the ambiguity disadvatage by meaning relatedness is evidence ofactive involvement of semantic representatioin the process of lexical competition. Thisconsistent with models of word recognitionwhich words compete to activate semantic r
resentations (Gaskell & Marslen-Wilson, 1997MARSLEN-WILSON
lo,
ofthee-e
ctei-s.ctr-d
eisn-sti-dsis
notcyrsn-tse-tn-egs
-adree
-hes
-
Hinton & Shallice, 1991; Joordens & Besn1994; Plaut, 1997; Plaut & Shallice, 1993).these models, the ambiguity disadvantage arbecause of the difficulty in mapping a singlethographic or phonological pattern of activatito multiple patterns of semantic activation. Tdifferent possible semantic patterns interfwith each other, and the additional time thatakes for this competition to be resolved pduces the ambiguity disadvantage. If the amguity disadvantage is indeed caused byinterference between competing semanticterns, then we would expect to see an effecmeaning relatedness. The level of interfereis related to the degree of overlap betweentwo patterns, such that any semantic featushared by the two patterns will reduce the intference. As with the ambiguity disadvantathis relatedness effect has been simulatedmodel where a simple least mean square arithm was used to learn the mapping betwedistributed orthographic and semantic represtations (Rodd, 2000).
In summary, the ambiguity disadvantage ported here, together with the relatedness efcan most easily be interpreted as evidencecompetition to activate a distributed semanrepresentation is a fundamental part of the wrecognition process.
The Sense Advantage
All three experiments reported here fouthat lexical decision times are faster for wowith many dictionary senses than for words wonly a few senses. This result is somewhat coterintuitive. Given that ambiguity between mtiple meanings produces a processing disadtage, why should ambiguity between multisenses produce the reverse effect? If we acthat the ambiguity disadvantage reflects interence between the different meanings of ambous words, then although the interference tween different senses would be reduced relato words with multiple meanings, this interfeence would surely slow recognition relativeunambiguous words. The result is equally prlematic for models in which words competeactivate abstract word nodes. If we assume
;different word senses correspond to differentA
ly
eogu
eeeeyennoroenu
s nwlai
ddthnir
be
nsn
oanitaeaetathr
srahar
tur
ben-r-uity
ar-
aimc-re-
ssedon-en-te-nlyil-did
easfi- ofn-il-hich
ef- int ady in aatf a ofwe ofons ap-r-
en-ay
rec-m-th
ore
rdac-e-o
SEMANTIC
word nodes, then we would expect multipsenses to delay recognition in the same wamultiple meanings. Alternatively, if we assumthat multiple senses correspond to a single lcal node, we would expect them to be recnized as quickly but not faster than unambious words.
This idea that multiple senses might be pected to show a similar, but possibly reducambiguity disadvantage is reinforced by recevidence from Klein and Murphy (2001). Thembedded polysemous words in two differphrases which biased the reader interpretatioeither the same or different senses of the wthey found that sense consistency aided bmemory and comprehension. From this thconcluded that there are separate represetions for the multiple senses of polysemowords. If the representations of the meaningdifferent word senses are sufficiently indepeent to produce this pattern of results, then would expect the interaction between multipsenses to delay recognition in a similar, though reduced, way to that seen for words wmultiple meanings. This suggests that an ational mechanism is necessary to explain word sense advantage, and that this mechawould need to be sufficiently strong to ovecome any effects of semantic competition tween different word senses.
One possible explanation of the word sebenefit is that words with many senses awords with few senses differ in the amountsemantic information contained in their semtic representation. In other words, a word wmany senses can be considered to be semcally rich. This is essentially the same argumthat Plaut and Shallice (1993) put forward to count for the processing benefit of concrwords over abstract words. In their computional account of the concreteness effect, difference between abstract and concrete wois reflected in the number of semantic featurea distributed semantic representation; abstwords were given fewer semantic features tconcrete words. These additional features pduce more stable representations, which in lead to faster settling times for words with mo
semantic features. Such an account of the worMBIGUITY 261
eas
exi-g--
x-d,nt
t tod;thyta-sofd-e
el-thi-e
sm--
edf-
hnti-ntc-te-eds inctn
o-rne
sense effect would need to assume that this efit for semantic richness is sufficient to ovecome any disadvantage caused by the ambigof these words.
This is related to the Schwanenflugel, Hnishfeger, and Stowe (1988) context availabilityaccount of the concreteness effect. They clthat contextual information about words is neessary for the integration that occurs in comphension, and that concrete words are procemore easily because of the ease with which ctextual information can be accessed. Schwanflugel et al. (1988) report evidence that concreness has an effect on lexical decision time owhen it is confounded with contextual availabity; when contextual availability, frequency, anfamiliarity were partialled out, concreteness dnot significantly predict response times, whercontextual availability accounted for a signicant proportion of the variance independentfrequency and familiarity. It is possible that cotextual information may be more readily avaable for words that have many senses and ware used in a wider range of contexts.
A third possible explanation of the sense fect is that it is a direct result of using a taskwhich words are presented in isolation withousentential or semantic context. As alreastated, words with many senses can be usedwide variety of contexts, and it is possible ththis experience results in the development orelatively context independent representationthe word. On the other hand, words with fesenses are used in a far more restricted rangcontexts and so may develop representatiwhose access is more dependent on the wordpearing in the appropriate context. This diffeence in the extent to which the lexical represtation of the words are context independent mbe important when participants are asked to ognize the words without any context. Presuably this task is more difficult for the words wifewer senses whose representations are mcontext dependent.
A final possible explanation is that the wosense benefit reflects differences in the attrtor basins that develop within a distributed smantic network. As noted by Kawamot
ds(1993), the different senses of a word corre-D
oo
att
nothdeists
ectptodac
gha
icthc
ortnothe
ses
on-og-n-theon-ata.lay athatus
ene-entlysis
leelytos-
theat
entredre-ityted-
ness suggests that the effect must, at least in
262 RODD, GASKELL, AN
spond to a set of highly correlated patternssemantic activation, and these senses will clectively create a broad, shallow basin oftraction, containing more than one stable stacorresponding to each different sense. Itplausible that under certain conditions, settliinto the correct attractor may be quicker fsuch a broad attractors, compared withsteep, narrow attractor basins that wouldvelop for words with only few senses. Thcould potentially explain the opposite effecof the two types of ambiguity; while multiplmeanings correspond to separate attrabasins, multiple senses correspond to multistable states within a single, broad attracbasin. This hypothesis needs to be assesseperforming the appropriate network simultions to determine the conditions under whisuch a pattern of effects might arise.
Further, this hypothesis can be extendedsuggest that the word sense benefit miemerge only in lexical decision, and not inrange of other tasks. In lexical decision, partipants may be able to respond correctly onbasis of the general familiarity produced by acessing a representation of the word’s cmeaning, and it is not necessary for themdisambiguate between a word’s differesenses. In terms of attractor structure, it is psible that the response is made as soon as
pattern of semantic activation has entered thbroad attractor basin corresponding to thword’s meaning, but before the activation hasettled into a stable state corresponding toparticular sense of a word. This explanation othe word sense benefit predicts that, if we loofor an effect of word senses on a task which requires the retrieval of a particular sense of thword, then the sense benefit should be elimnated and possibly reversed due to the needMARSLEN-WILSON
fl--e,isgre-
orlerby
-h
tot
-e-eots-
disambiguate between individual word sen(cf. Klein & Murphy, 2001).
CONCLUSION
The data reported here require us to recsider how semantic ambiguity affects the recnition of isolated words. While we do not cosider these data to be inconsistent with existing ambiguity advantage data, they do ctradict the accepted interpretation of these dWe have found that multiple meanings derecognition, while multiple senses produceprocessing advantage. We therefore claim the ambiguity advantage reported in previostudies should be interpreted as showing a bfit for words with many senses; this is consistboth with the data reported here and our anaof the stimuli from previous studies.
Our claim that ambiguity between multipmeanings can slow lexical decisions is entirnew. Yet it can apparently be incorporated inmost current models of word recognition by asuming that there is competition betweendifferent meanings of ambiguous words or ththere is an advantage for the more frequmeanings of unambiguous words compawith ambiguous words matched on overall fquency. However, the finding that the ambigudisadvantage is modulated by meaning rela
eesafk-ei-to
part, be due to competition to activate a distrib-uted semantic representation. It is less clearhow the word sense benefit should be inter-preted, and further work is required to deter-mine the cause of this intriguing effect.Nonetheless, the overriding message from thisseries of experiments is that the word recogni-tion process is intimately tied in with the com-petitive process by which the stored meaningsof words are retrieved.
SEMANTIC AMBIGUITY 263
APPENDIX A
Experiment 1 Stimuli
Ambiguous words Unambiguous words
admit advance affair arms bus feearticle badger bark batter baby funblind bonnet bowl boxer clay sanebridge broke bulb cabinet coal growcalf can cane case frog seekchance charm chest china goat itemclog company craft cricket lung taskdeed degree dense digit hill votedry express feet fence tent warnfirm fling free glare lake poetglass grain hamper horn tiger aloneinterest jumper kid kind apple fraudlast late lean left bible griefletter lie like limp brain dozenlobby marble march maroon cider unitemight nail net novel cigar urbanodd organ palm panel glove thiefpark patient peer picket hotel throwpine pitcher poach poker lorry amusepole pride pupil ram metal brutalrare rate reflect refrain ocean miseryruler sack safe sage river prayerscrap screen seal season cattle terrorsecond sense sentence shed forest wintersign spade speaker spell weapon dollarstable staff stag stalk rabbit travelstamp staple static stern throat destroystore strand straw swallow custard kingdomswear temple tend tense diamond citizenterm toast trial trust
uniform vent watch yardes
APPENDIX B
Experiment 2 Stimulus Groups
Ambiguous Unambiguous
Few senses Many senses Few senses Many sens
ash angle ant beltcalf bark bandage bendchap blow bet bitecricket boil bone burncuff bowl bulk dipfleet bust cage drainfudge clip cake featherhide clutch carton flashlime compound crew griploaf duck crude hammerloom flush deaf hangmint fold farm hookmole gag feast load
novel gum foam loopses
264 RODD, GASKELL, AND MARSLEN-WILSON
Ambiguous Unambiguous
Few senses Many senses Few senses Many sen
page hail harsh maskpen jam heap nestpine jar hinge pinchpoach lap hurdle rollport lean join saddleprune lock lump scanpupil pitch path shaderare scale profit slicerash seal request sliderifle slip rust smashstable spell silk sour
stern stall slim spins
stunt stem slot steamtend strain snake swaytense strand soap threadtoast stud spy treadutter swallow stain whipyard tap trot wire
APPENDIX C
Experiment 3 Stimulus Groups
Ambiguous Unambiguous
Many senses Few senses Many senses Few sense
bark calf belt antboil chap bite boneclutch fleet bounce crudeduck fudge chill farmfit hide dip feastflush loaf drain fogfly mint hook grinfold mole kick growlfret page loop guessgag pen mask harshgum pine nest hingehail poach shade loudjam port slide rustjar prune smash shirtlean rare snap silklock rash soak sip
seal sage spin slotslip stern steam snowspell stunt sway soapstall tend thread spystem tense tread stainstick toast wheel taskstud yard wire winchr
t
)
-
b
n
,-r
tc
st
ni
o
htc
rt
o-
.heet-
i--
n
n
u
-of
f
d th
l-
e
e
a-
SEMANTIC A
REFERENCES
Azuma, T. (1996). Familiarity and relatedness of womeanings: Ratings for 110 homographs.Behavior Re-search Methods, Instruments and Computers, 28,109–124.
Azuma, T., & Van Orden, G. C. (1997). Why safe is betthan fast: The relatedness of a word’s meanings affelexical decision times. Journal of Memory and Lan-guage, 36,484–504.
Baayen, R. H., Piepenbrock, R., & Van Rijn, H. (1993The CELEX lexical database[CD-ROM] Philadelphia,PA: Linguistic Data Consortium, University of Pennsylvania.
Borowsky, R., & Masson, M. E. J. (1996). Semantic amguity effects in word identification. Journal of Experi-mental Psychology: Learning Memory and Cognitio,22,63–85.
Coltheart, M., Davelaar, E., Jonasson, J. T., & Besner,(1997). Access to the internal lexicon. In S. Dorn(Ed.), Attention and performance VI(pp. 535–555).Hillsdale, NJ: Erlbaum.
Cruse, A. D. (1986). Lexical semantics. Cambridge, Eng-land: Cambridge University Press.
Fellbaum, C. (Ed.) (1998). Wordnet: An electronic lexicaldatabase. Cambridge, MA: MIT Press.
Fera, P., Joordens, S., Balota, D. A., Ferraro, F. R.Besner, D. (1992). Ambiguity in meaning and phonology: Effects on naming. Paper presented at the 33annual meeting of the Psychonomic Society.
Gaskell, M. G., & Marslen-Wilson, W. D. (1997). Integraing form and meaning: A distributed model of speeperception. Language and Cognitive Processes, 12,613–656.
Gernsbacher, M. A. (1984). Resolving 20 years of incosistent interactions between lexical familiarity and othography, concreteness, and polysemy. Journal ofExperimental Psychology: General, 113,254–281.
Gilhooly, K., & Logie, R. (1980). Methods and designAge of acquisition, imagery, concreteness, familiariand ambiguity measures for 1,944 words. BehaviorResearch Methods & Instrumentation, 12,395–427.
Hino, Y., & Lupker, S. J. (1996). Effects of polysemy ilexical decision and naming—An alternative to lexcal access accounts.Journal of Experimental Psy-chology: Human Perception and Performance, 22,1331–1356.
Hinton, G. E., & Shallice, T. (1991). Lesioning an attractnetwork: Investigations of acquired dyslexia. Psycho-logical Review, 98,74–95.
Holley-Wilcox, P. (1977). The effect of homophony witauditory presentation of stimuli. Paper presented athe meeting of the Midwestern Psychological Assoation, Chicago, May 1977.
Jastrzembski, J. E. (1981). Multiple meanings, numberelated meanings, frequency of occurrence, and lexicon. Cognitive Psychology, 13,278–305.
Joordens, S., & Besner, D. (1994). When banking
meaning is not (yet) money in the bank—Explorations in connectionist modeling. Journal of Experi-MBIGUITY 265
d
ercts
.
i-
D.ic
&
d
-h
n-r-
:y,
-
r
i-
ofhe
on
mental Psychology: Learning Memory and Cognition,20,1051–1062.
Kawamoto, A. H. (1993). Nonlinear dynamics in the reslution of lexical ambiguity: A parallel distributed pro-cessing account. Journal of Memory and Language,32,474–516.
Kawamoto, A. H., Farrar, W. T., & Kello, C. T. (1994)When two meanings are better than one: Modeling tambiguity advantage using a recurrent distributed nwork. Journal of Experimental Psychology: HumanPerception and Performance, 20,1233–1247.
Kellas, G., Ferraro, F. R., & Simpson, G. B. (1988). Lexcal ambiguity and the timecourse of attentional allocation in word recognition. Journal of ExperimentalPsychology: Human Perception and Performance, 14,601–609.
Kilgarriff, A. (1992). Polysemy. Unpublished doctoral dis-sertation, University of Sussex.
Klein, D. E., & Murphy, G. L. (2001). The representatioof polysemous words. Journal of Memory and Lan-guage, 45,259–282.
Lyons, J. (1977). Semantics. Cambridge, England: Cam-bridge University Press.
Lyons, J. (1981). Language and linguistics. Cambridge,England: Cambridge University Press.
Marslen-Wilson, W. D. (1987). Functional parallelism ispoken word recognition. Cognition, 25,71–102.
McCusker, L. X., Hillinger, M. L., & Bias, R. G. (1981).Phonological recoding and reading. Psychological Bul-letin, 89,217–245.
Millis, M. L., & Button, S. B. (1989). The effect of poly-semy on lexical decision time: now you see it, now yodon’t. Memory & Cognition, 17,141–147.
Moss, H., & Older, L. (1996). Birkbeck word associationnorms. Hove: Psychology Press.
Onifer, W., & Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension—Effects frequency of meaning and contextual bias. Memory &Cognition, 9, 1–39.
Parks, R., Ray, J., & Bland, S. (1998). Wordsmyth Englishdictionary—Thesaurus. [ONLINE]. Available: http://www.wordsmyth.net/ [1999, February 1], University oChicago.
Pexman, P. M., & Lupker, S. J. (1999). Ambiguity anvisual word recognition: Can feedback explain bohomophone and polysemy effects? Canadian Journalof Experimental Psychology, 53,323–334.
Plaut, D. C. (1997). Structure and function in the lexicasystem: Insights from distributed models of word reading and lexical decision.Language and CognitiveProcesses, 12,765–805.
Plaut, D. C., & Shallice, T. (1993). Deep dyslexia: A casstudy of connectionist neuropsychology. CognitiveNeuropsychology, 10,377–500.
Ratcliff, R. (1993). Methods for dealing with reaction-timoutliers. Psychological Bulletin, 114,510–532.
Rayner, K. (1998). Eye movements in reading and inform
- tion processing: 20 Years of research. PsychologicalBulletin, 124,372–422.
266 RODD, GASKELL, AND MARSLEN-WILSON
Rodd, J. M. (2000). Semantic representation and lexicalcompetition: Evidence from ambiguity. Unpublished
o
b-
o
Twilley, L. C., Dixon, P., Taylor, D., & Clark, K. (1994).University of Alberta norms of relative meaning fre-
thedi-
se-
doctoral dissertation, University of Cambridge.Rubenstein, H., Garfield, L., & Millikan, J. A. (1970). H
mographic entries in the internal lexicon. Journal ofVerbal Learning and Verbal Behavior, 9, 487–494.
Schwanenflugel, P. J., Harnishfeger, K. K., & Stowe, R.(1988). Context availability and lexical decision for astract and concrete words. Journal of Memory and Language, 27,499–520.
Swinney, D. A. (1979). Lexical access during sentence cprehension: (Re)consideration of context effects. Jour-nal of Verbal Learning and Verbal Behavior, 18,645–659.
-
W.-
m-
quency for 566 homographs. Memory & Cognition, 22,111–126.
Ziegler, J. C., & Ferrand, L. (1998). Orthography shapesperception of speech: The consistency effect in autory word recognition. Psychonomic Bulletin & Re-view, 5, 683–689.
Zwitserlood, P. (1989). The locus of effects of sentential–mantic context in spoken-word processing. Cognition,32,25–64.
(Received October 27, 1999)(Revision received March 7, 2001)