=Paper=
{{Paper
|id=Vol-1347/paper11
|storemode=property
|title=Modeling lexical effects in language production: where have we gone wrong?
|pdfUrl=https://ceur-ws.org/Vol-1347/paper11.pdf
|volume=Vol-1347
|dblpUrl=https://dblp.org/rec/conf/networds/ZhaoM15
}}
==Modeling lexical effects in language production: where have we gone wrong?==
https://ceur-ws.org/Vol-1347/paper11.pdf
Modeling Lexical Effects in Language Production: Where Have We
Gone Wrong?
Ting Zhao Victoria A. Murphy
Department of Education Department of Education
University of Oxford University of Oxford
ting.zhao@education.ox.ac.uk victoria.murphy@education.ox.ac.uk
1 Introduction 2 Lexical characteristics that contribute
to the speed of spoken production
Words have their own conceptual representations,
semantic properties, and physical forms. These This study considers three lexical layers (i.e.
lexical characteristics not only set words apart as Meaning, Form, and Usage), each of which is
a distinct item in the lexical repertoire but also underpinned by its own manifest indicators. The
provide valuable insight into the processes and lexical variables under examination all have been
mechanisms of language production. found to significantly influence the speed of lex-
Over the past decades there has been a large ical processing, as will be briefly reviewed below.
body of research examining how word meaning, Meaning. (1) Word concreteness (WC): A
form, and usage directly affect the speed of main difference between concrete and abstract
monolingual speakers’ production (e.g. Alario et words lies in the existence of sensorimotor at-
al., 2004; Barry, Morrison, & Ellis, 1997; Bates tributes of the former. A number of studies have
et al., 2003; Bonin, Chalard, Méot, & Fayol, revealed that concrete words exhibit preferential
2002). Of note, almost all these studies have processing relative to abstract words (e.g. De
failed to accommodate the fact that word usage, Groot, 1992; Jin, 1990; Schwanenflugel & Akin,
given it is a behavioral outcome (Zevin & 1994). (2) Word typicality (WT): The degree of a
Seidenberg, 2002, 2004), likely mediates the re- lexical item’s typicality depends upon how many
lationship between meaning/form and spoken attributes that it shares with other members of the
production. Moreover, lexical characteristics same category. Typical items are usually pro-
have been predominantly examined as discrete cessed more accurately and faster relative to
variables in the literature, but in fact, some of atypical items in a range of tasks (e.g. Bjorklund
them may correspond to the same layer of lan- & Thompson, 1983; Jerger & Damian, 2005;
guage production or the same aspect of lexical Southgate & Meints, 2000). (3) Semantic neigh-
knowledge. Additionally, little work has been borhood density (SND): Words with high SND
done on children’s emerging bilingual lexical are characterized by having a great deal of se-
representations, especially those learning an L2 mantic neighbors and low semantic distance,
within input-limited contexts, possibly due to the whereas low-SND words typically have few se-
fact that this population has only recently begun mantic neighbors and high semantic distance.
to receive focused attention in the research field. The superiority of high SND over low SND
In order to delineate the exact manner in words for processing has been observed in lexi-
which lexical effects come into play, the present cal decision, word naming, and semantic catego-
study used structural equation modeling to per- rization (e.g. Buchanan, Westbury, & Burgess,
form a simultaneous test of the complex relation- 2001; Siakaluk, Buchanan, & Westbury, 2003;
ships among a variety of lexical variables and to Yates, Locker, & Simpson, 2003). (4) Number of
assess their direct, indirect and total effects on related senses (NoRS): Many words are polyse-
L2 lexical processing efficiency. Furthermore, mous in terms of having several different but
attempts were also made to estimate and then to related senses. Compared to monosemous words,
compare three types of hypothesized models, in polysemous words exhibit preferential pro-
which the lexical relationships were specified cessing in a variety of tasks (e.g. Beretta,
differently with respect to spoken production in Fiorentino, & Poeppel, 2005; Klepousniotou &
L2 learners. Baum, 2007; Lichacz, Herdman, Lefevre, &
Baird, 1999).
Copyright © by the paper’s authors. Copying permitted for private and academic purposes.
In Vito Pirrelli, Claudia Marzi, Marcello Ferro (eds.): Word Structure and Word Usage. Proceedings of the NetWordS Final
Conference, Pisa, March 30-April 1, 2015, published at http://ceur-ws.org
51
Form. Word length can be measured ortho- orded using the Audacity software, and then
graphically (i.e. NoL: number of letters) or pho- manually calculated for analysis.
nologically (i.e. NoP: number of phonemes and Norms of lexical variables. The values of WC,
NoS: number of syllables). The presence of WT, and SWF were rated by the participants on
length effects has been reported in several previ- Likert scales. The values of other lexical varia-
ous studies (e.g. Alario et al., 2004; Cuetos, Ellis, bles were obtained from psycholinguistics data-
& Alvarez, 1999; De Groot, Borgwaldt, Bos, & bases such as the Irvine Phonotactic Online Dic-
Van den Eijnden, 2002) although the predictive tionary (Vaden, Halpin, & Hickok, 2009) and the
power of each specific measure varies across Wordmine2 (Durda & Buchanan, 2006).
research contexts possibly due to their examina-
tion of different languages (Bates et al., 2003). 3.2 Analytical strategies
Usage. Usage is represented by subjective Structural equation modeling (SEM), which
word frequency (SWF) and /or age of acquisition combines path analysis, confirmatory factor
(AoA), both of which have been observed to sig- analysis, and analysis of structural models, was
nificantly affect the speed of spoken production used to estimate the goodness-of-fit of three
in such a way that individuals take less effort to types of hypothesized models. This analytical
access high-frequency and early-acquired words strategy, as an extension of multiple regression,
relative to low-frequency and late-acquired ones enables researchers to estimate not only the di-
(e.g. Balota, Cortese, Sergent-Marshall, Spieler, rect effects but also indirect effects that one vari-
& Yap, 2004; Barry et al., 1997; Morrison, Ellis, able has upon another. Moreover, SEM can be
& Quinlan, 1992). AoA effects interact with fre- used to measure the proportion of variance ex-
quency effects in such a way that the former is plained by the models proposed in the present
partly dependent on the latter (Brysbaert & study so as to hold general implications for the
Ghyselinck, 2006). lexical processing system as a whole, although it
should be acknowledged that this type of analy-
3 Methodology and analytical strategies sis might lack a specific focus on certain varia-
3.1 Methodology bles through purposeful manipulation of experi-
mental materials. Additionally, latent variables
Participants. Thirty-nine 5th grade children (aged are formed to manifest different dimensions that
10-11 years) and 94 undergraduates (aged 17-20 are underpinned by their own indicators. In so
years) were recruited. All had Chinese as their doing, the present study moves away from the
native language and English as their second. The examination of each lexical variable to that of
child sample had been learning English as a for- specified constructs and structural relations be-
eign language for about 2.5 years, and the adult tween constructs, thus a better understanding of
sample for approximately 10 years. the nature of lexical characteristics can be gained
Stimuli. The experiment consisted of two at a more macro level.
blocks of stimulus words and one block of filler Conducting SEM typically involves six steps
words. Each block had 35 (in the child group) / (Kline, 2011): model specification, model identi-
66 (in the adult group) valid trials. The stimuli fication, select good measures, model estimation,
were selected from ten semantic categories in model evaluation and modification, and inter-
almost equal numbers. They were all presented preting and reporting results. Moreover, as rec-
in the same format over the course of the exper- ommended by Kline (2011), SEM was conducted
iments. in two steps in the present study, that is, the
Procedures. The participants were tested indi- measurement models were validated in terms of
vidually in a quiet room. They performed picture convergent validity, discriminant validity, and
naming in L2 (English) and then L1 (Chinese)- reliability before the structural models proceeded
to-L2 (English) translation. As a stimulus ap- to be estimated. One last thing to note is that the
peared on the screen, the participants were asked data entered for analysis were lexical items. The
to produce the L2 word as rapidly and accurately stimulus size in the adult group was considered
as possible. The SuperLab software (Cedrus sufficiently large for performing SEM analysis.
Corporation, 2007) generated stimulus presenta- In order to reduce the complexity of the hypothe-
tions. Response latencies (RLs), defined as the sized model specifying children’s L2 lexical pro-
duration between the presentation of a stimulus cessing, composite variables rather than latent
and the initiation of a vocal response, were rec- variables were constructed to decrease the num-
52
ber of stimulus words required for this type of cies. Similar results held for adults’ L1-to-L2
analysis. translation (see Appendix C for details).
Three competing models were hypothesized
and estimated to determine which one best fitted
the data. The first model concerns only the direct
relationship between the lexical variables, and
picture naming and translation latencies. The
second model identifies word usage as a media-
tor and examines the indirect effects of meaning
and form variables on the recorded RLs. The
third model considers both direct and indirect
effects of word meaning and form on the out-
come variable. To illustrate, an example of these
three types of hypothesized models that specify Model 1
the possible relationships between lexical varia-
bles and the speed of adults’ picture naming is
presented in Appendix A.
The goodness of model fit was estimated ac-
cording to six types of indices, including model
𝝌𝟐 , CFI, RMSEA, AGFI, GFI, and NFI. A rule
of thumb is that an RMSEA below .08 indicates
reasonable fit, and values greater than .90 for the
CFI, AGFI, GFI, and NFI suggest close approx- Model 2
imate fit. SEM was run using IBM SPSS AMOS
v.20.
It should be noted that, before performing
SEM analysis, the whole RL data set was
screened for incorrect and omitted responses,
outliers (low cut-off: below 350ms, high cut-off:
3 SDs), and those participants and stimulus items
with an exceptionally high error rate. As conven-
tionally done, RLs were then averaged to gener-
ate a summary score for each lexical item, and Model 3
these values were entered into final SEM analy-
sis. Figure 1: SEM results: Picture naming in adults
4 Results As regards children’s picture naming, the re-
sults presented in Appendix B shows that Model
The model-fit indices of the three models under 3 reached a better model fit than Models 1 and 2.
examination across two types of productive tasks Moreover, Figure 2 indicates that Model 3 (38%)
in both populations are presented in Appendix B.
explained more variance in naming speed than
Comparatively, the child and adult data could Model 1 (36%) and Model 2 (24%). In addition,
best be modeled by the third model where word word usage, as represented by age of acquisition,
meaning and form not only make direct but also together with word typicality were found to sig-
indirect contribution to the RLs. nificantly and directly predict the naming speed
Take picture naming in adults as an example in Model 3. Similar results were observed with
(see SEM results in Appendix B and Figure 1), it children’s L1-to-L2 translation (see Appendix C
is clear that Model 3 achieved a much better for details).
model fit than Model 1, and Model 3 explained
more variance in naming latencies (59%) than
Model 1 (45%) and Model 2 (51%). Additionally,
among all the lexical variables included in Model
3, only word usage was found to make a signifi-
cant and direct contribution to the naming laten-
Model 1
53
whether or not the same results still hold for oth-
er L2 learner types, particularly those whose L1s
are not Sino-Tibetan languages, as well as for
monolingual speakers needs to be further inves-
tigated. Importantly, examining these issues
would allow us to gain a better understanding of
Model 2 the nature of lexical characteristics by addressing
the issue of whether lexical effects are language-
dependent or universal across languages. Second,
not all the variance can be explained the included
lexical variables, partly due to the fact that it
seems implausible to cover every possible fea-
ture of a lexical item because of theoretical and
Model 3 practical considerations. Third, given the use of a
non-experimental design, it would be difficult to
Figure 2: Path analysis results: Picture naming in make unequivocal explanations of causality
children among the variables of interest.
To conclude, the model that considers both di-
Taken together, these results indicate that rect and indirect effects of meaning and form on
word usage does not exist independently of other L2 lexical processing efficiency may be superior
lexical variables but rather mediates the impact to those that do not. As also observed in our
of meaning and form on L2 children’s and adults’ study, word usage does play a mediating role in
productive performance. In comparison, the indi- lexical processing, in part reflecting that ‘only in
rect effects of meaning and form on L2 lexical the stream of thought and life do words have
processing efficiency were found to be more no- meanings’ (Wittgenstein, 1967, p.31).
ticeable with adults relative to with children.
References
5 Discussion and conclusions
Alario, F-Xavier, Ferrand, L., Laganaro, M.,
The present study uses SEM as a methodological New, B., Frauenfelder, U.H., & Segui, J.
improvement to investigate the relationships be- (2004). Predictors of picture naming speed.
tween a range of lexical variables and L2 lexical Behavior Research Methods, 36(1): 140-155.
processing efficiency in both children and adults. Balota, D.A., Cortese, Michael J, Sergent-
A comparison of the three different types of Marshall, Susan D, Spieler, Daniel H, & Yap,
models indicates that word meaning and form MelvinJ. (2004). Visual word recognition of
makes not only direct but also indirect contribu- single-syllable words. Journal of
tion to the speed of L2 lexical processing, and Experimental Psychology: General, 133: 283-
word usage likely mediates the extent to which 316.
meaning and form influence the processing out- Barry, C., Morrison, C.M., & Ellis, A.W. (1997).
come. Furthermore, a comparison between chil- Naming the Snodgrass and Vanderwart
dren and adults suggests that the importance of pictures: Effects of age of acquisition,
word usage tends to increase with age. frequency, and name agreement. The
A note of caution thus should be raised when Quarterly Journal of Experimental
interpreting the results of previous studies where Psychology, 50(3): 560-585.
the mediating effects of word usage have not Bates, Elizabeth, D’Amico, Simona, Jacobsen,
been adequately addressed. Accordingly, future Thomas, Székely, Anna, Andonova, Elena,
research modeling lexical effects would be well Devescovi, Antonella, . . . Pléh, Csaba. (2003).
advised to consider the indirect effect that word Timed picture naming in seven languages.
meaning and form have on L2 learners’ produc- Psychonomic Bulletin & Review, 10(2): 344-
tive performance via usage. 380.
Although this study provides new insights into Beretta, A., Fiorentino, R., & Poeppel, D. (2005).
how lexical variables are related to each other, The effects of homonymy and polysemy on
there are several limitations that should be lexical access: an MEG study. Cognitive
acknowledged. First, since this research focuses Brain Research, 24(1): 57-65.
only on L2 learners within input-limited contexts,
54
Bjorklund, David F, & Thompson, B.E. (1983). Kline, Rex B. (2011). Principles and Practice of
Category typicality effects in children’s Structural Equation Modeling (3rd ed.). New
memory performance: Qualitative and York: Guilford Press.
quantitative differences in the processing of Lichacz, F.M., Herdman, C.M., Lefevre, J.A., &
category information. Journal of Baird, B. (1999). Polysemy effects in word
Experimental Child Psychology, 35(2): 329- naming. Canadian Journal of Experimental
344. Psychology, 53(2): 189-193.
Bonin, P., Chalard, M., Méot, A., & Fayol, M. Morrison, C.M., Ellis, A.W., & Quinlan, P.T.
(2002). The determinants of spoken and (1992). Age of acquisition, not word
written picture naming latencies. British frequency, affects object naming, not object
Journal of Psychology, 93(1): 89-114. recognition. Memory & Cognition, 20(6):
Brysbaert, Marc, & Ghyselinck, Mandy. (2006). 705-714.
The effect of age of acquisition: Partly Schwanenflugel, P.J., & Akin, Carolyn E. (1994).
frequency related, partly frequency Developmental trends in lexical decisions for
independent. Visual Cognition, 13(7-8): 992- abstract and concrete words. Reading
1011. Research Quarterly, 29(3): 251-264.
Buchanan, L., Westbury, C., & Burgess, C. Siakaluk, P.D., Buchanan, L., & Westbury, C.
(2001). Characterizing semantic space: (2003). The effect of semantic distance in
neighborhood effects in word recognition. yes/no and go/no-go semantic categorization
Psychonomic Bulletin & Review, 8 (3): 531- tasks. Memory & Cognition, 31(1): 100-113.
544. Southgate, V., & Meints, K. (2000). Typicality,
Cedrus Corporation. (2007). SuperLab 4.5. San naming, and category membership in young
Pedro, CA. children. Cognitive Linguistics, 11(1/2): 5-16.
Cuetos, Fernando, Ellis, A.W., & Alvarez, B. Vaden, KI, Halpin, HR, & Hickok, GS. (2009).
(1999). Naming times for the Snodgrass and Irvine Phonotactic Online Dictionary, Version
Vanderwart pictures in Spanish. Behavior 2.0. Retrieved January 30, 2013 from
Research Methods, 31(4): 650-658. http://www.iphod.com/search/V2ListWords.html.
De Groot, A.M.B. (1992). Determinants of word Yates, Mark, Locker, Lawrence, & Simpson,
translation. Journal of Experimental Greg B. (2003). Semantic and phonological
Psychology: Learning, Memory, and influences on the processing of words and
Cognition, 18(5): 1001-1018. pseudohomophones. Memory & Cognition,
De Groot, A.M.B., Borgwaldt, Susanne, Bos, 31(6): 856-866.
Mieke, & Van den Eijnden, Ellen. (2002). Zevin, J.D., & Seidenberg, M.S. (2002). Age of
Lexical decision and word naming in acquisition effects in word reading and other
bilinguals: Language effects and task effects. tasks. Journal of Memory and Language,
Journal of Memory and Language, 47(1): 91- 47(1): 1-29.
124. Zevin, J.D., & Seidenberg, M.S. (2004). Age-of-
Durda, K., & Buchanan, L. . (2006). acquisition effects in reading aloud: Tests of
WordMine2. Retrieved December 1, 2012. cumulative frequency and frequency
from http://www.wordmine2.org trajectory. Memory & Cognition, 32(1): 31-38.
Jerger, S., & Damian, M.F. (2005). What’s in a
name? Typicality and relatedness effects in
children. Journal of Experimental Child
Psychology, 92(1): 46-75.
Jin, Y.S. (1990). Effects of concreteness on
cross-language priming in lexical decisions.
Perceptual and Motor Skills, 70(3): 1139-
1154.
Klepousniotou, E., & Baum, S.R. (2007).
Disambiguating the ambiguity advantage
effect in word recognition: An advantage for
polysemous but not homonymous words.
Journal of Neurolinguistics, 20(1): 1-24.
55
Appendix A. An example of the hypothesized models
Adult picture naming
Model 1 Model 2 Model 3
Appendix B. Fit indices for the hypothesized models
𝜒 ! (𝑝) df CFI RMSEA AGFI GFI NFI
Picture Model 1 173.83 (.00) 27 .81 .17 .74 .87 .79
naming Model 2 52.16 (.00) 27 .97 .07 .90 .95 .94
Adults Model 3 45.96 (.00) 22 .97 .08 .90 .96 .94
Chinese- Model 1 169.46 (.00) 27 .81 .16 .75 .88 .79
English Model 2 45.69 (.01) 27 .98 .06 .91 .96 .94
translation Model 3 41.87 (.01) 22 .97 .07 .90 .96 .95
Picture Model 1 28.17 (.00) 11 .85 .27 .82 .93 .79
naming Model 2 28.52 (.00) 11 .88 .12 .83 .94 .79
Chil- Model 3 5.67 (.46) 6 1.00 .00 .93 .99 .96
dren Chinese- Model 1 28.17 (.00) 11 .81 .12 .82 .93 .74
English Model 2 23.09 (.02) 11 .86 .10 .87 .95 .79
translation Model 3 5.67 (.46) 6 1.00 .00 .93 .99 .95
Appendix C. SEM results of the hypothesized models
Adults:
Picture naming
Model 1 Model 2 Model 3
56
L1-to-L2 translation
Model 1 Model 2 Model 3
Children:
Picture naming
Model 1 Model 2 Model 3
L1-to-L2 translation
Model 1 Model 2 Model 3
57