=Paper=
{{Paper
|id=Vol-1347/paper07
|storemode=property
|title=Effects of processing complexity in perception and production. The case of English comparative alternation
|pdfUrl=https://ceur-ws.org/Vol-1347/paper07.pdf
|volume=Vol-1347
|dblpUrl=https://dblp.org/rec/conf/networds/Kunter15
}}
==Effects of processing complexity in perception and production. The case of English comparative alternation==
https://ceur-ws.org/Vol-1347/paper07.pdf
Effects of processing complexity in perception and production. The case of
English comparative alternation
Gero Kunter
English Language and Linguistics
Heinrich-Heine-Universität Düsseldorf
gero.kunter@uni-duesseldorf.de
Abstract For instance, an adjective that is morphologically
complex is assumed to be also cognitively more
This paper discusses the effect of pro- complex than a simplex adjectives, and in order to
cessing complexity on the English com- compensate for this increased cognitive complex-
parative alternation. The reported exper- ity, speakers may prefer the analytic comparative
iments show a processing advantage of the over the synthetic alternative.
synthetic comparative in perception, but a
preference of the analytic comparative in Yet, there is only little psycholinguistic research
sentence production if the base adjective is that investigated this assumed processing advan-
cognitively complex. These results imply tage of analytic forms. A notable exception is
that perceptual complexity and complex- Boyd (2007, ch. 2) who conducted a self-paced
ity in production have diverging effects on reading experiment to investigate processing dif-
the English comparative alternation. More ferences between synthetic and analytic compara-
generally, the paper calls for a fine-grained tives. Indeed, he reports shorter reaction times for
look at the role of processing complexity the sentences containing analytic comparatives,
in areas of morphosyntactic variation. but due to the experimental design, this evidence
is only indirect and allows for alternative interpre-
1 Introduction tations. As yet, then, there is only limited empiri-
cal evidence for the assumption that analytic com-
Most English comparatives are formed using ei- paratives are easier to process than synthetic com-
ther a synthetic form (e.g. easier) or an analytic paratives. In addition, as pointed out by Mondorf
form (e.g. more important). While most adjec- (2014, 201), it is still an unresolved issue whether
tives clearly prefer either the synthetic or the an- more-support is a response to increased processing
alytic comparative, there is a considerable num- loads in production or in perception.
ber of adjectives which frequently take both forms,
e.g. more friendly vs. friendlier. The decision This paper addresses these two issues. First,
for either form is influenced by several phonologi- it presents the results from a perception experi-
cal, morphological, syntactic and semantic factors. ment which tested whether analytic comparatives
For example, the probability of analytic compara- are indeed easier to process for listeners. Con-
tives increases with the number of morphemes in trary to this hypothesis, the reaction times show
the adjective base. It is also higher if the com- that analytic comparatives have a processing dis-
parative is in predicative than in attributive posi- advantage in perception. Then, a production ex-
tion, and it decreases with an increasing compara- periment is discussed which elicited spoken sen-
tive/positive ratio (see Szmrecsanyi 2005, Hilpert tences containing a comparative construction. The
2008 and Mondorf 2009 for detailed discussions). analysis reveals that the processing complexity is
Mondorf (2009) argues that these factors are all a significant predictor of the comparative alterna-
part of a more general, audience-oriented com- tion: with increasing complexity of the base adjec-
pensatory mechanism called more-support: if the tive, the probability of analytic comparatives in-
cognitive complexity of the adjectival base or its creases. Thus, the paper argues that speakers and
environment increases, speakers prefer the ana- listeners process the English comparative variants
lytic comparatives, because they have a processing differently, and that it is the speaker who benefits
advantage over the corresponding synthetic form. from a compensatory use of more comparatives.
Copyright c 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
32
2 Comparative variation in perception reaction time
−0.50
−0.55 Analytic
2.1 Method
Analytic
−0.60
−0.65 Analytic
Synthetic
31 native speakers of Canadian English partici- −0.70 Synthetic
pated in an auditory decision task in which they −0.75
−0.80
Synthetic
had to decide whether the acoustic stimuli was an −4 0 2 −0.5 0.5 1.0 −1 0 1 2
existing English form. The set of stimuli contained Preceding RT PLD20
(residualized)
No. of phonemes
(residualized)
the analytic and synthetic comparative form for 60 −0.50
−0.55
adjective types with at least 5 attestations for both −0.60
Analytic
Synthetic
Analytic
forms in the Corpus of Contemporary American −0.65
Synthetic
English (Davies 2008-). The stimuli were pro- −0.70
−0.75
duced by a male speaker of Canadian English with −0.80
phonetic training. He was instructed to produce 2 4 6 8 10 1 3 5 7
Synthetic frequency Analytic frequency
the stimuli in citation form with a single accent on
the primary stressed syllable of the base adjective
in both types of stimuli. Accordingly, more was Figure 2: Partial effects of significant interactions
produced stressed, but unaccented. of Class on reaction times
Alongside the 2 × 60 = 120 synthetic and ana-
lytic comparatives, the set of stimuli also included
360 distractors. Some of the distractors combined The density estimate suggests that reaction
more with non-existing words, others combined times are, in general, higher for analytic compar-
the adjective bases with the illegal suffix -ic. In atives than for synthetic comparatives. This vi-
addition, the set of distractor items contained non- sual interpretation is supported by a linear mixed-
existing words ending in -er as well as existing effects regression model with reaction times as
words and complex words. Examples of the test the dependent variable (in order to fulfill the lin-
stimuli are given in (1a), and distractor examples earity assumption of the linear model, the reac-
are given in (1b). tion times were power-transformed with λ = -1.52,
see Box and Cox 1964). The main predictor was
(1) a. colder, happier, yellower the factor Class (with values Synthetic and Ana-
more cold, more wealthy, more yellow lytic). Additional predictors addressed several in-
fluences that may be expected affect the reaction
b. ∗coldic, more ∗gorsty, ∗rilker times: the subject-specific variables Handedness,
on wire, chasting Sex, and Age, the experimental variables Trial
number and Reaction time in previous trial, (Pre-
2.2 Results ceding RT, see Baayen and Milin 2010 for a dis-
Figure 1 displays the density estimate for the dis- cussion), phonological variables (Metrical struc-
tribution of reaction times. The solid and the ture of base, residualized Number of phonems),
dashed lines correspond to the results for synthetic and the lexical variables Number of phonological
and analytic stimuli, respectively. neighbours, Mean RT of base adjective, residual-
ized Phonological Levenshtein distance (PLD20,
all three from Balota et al. 2007), Age of acqui-
0.0020
Synthetic
Analytic sition (from Kuperman et al. 2012), Frequencies
of base, Analytic comparative, Synthetic com-
Density
0.0010
parative (from COCA), Inflectional entropy (cf.
Moscoso del Prado Martı́n et al. 2004). With the
exception of the three Subject predictors, the ini-
0.0000
tial model contained interactions between Class
1000 1500 2000 2500 and the other predictors. Finally, random inter-
Reaction times times in ms cepts were included for the factors Subject and
Adjective base.
Figure 1: Density estimate of reaction times in After removal of insignificant predictors, the fi-
perception experiment nal model reports significant interactions between
33
stimulus Class and Preceding RT, PLD20, Number 3.2 Reaction times
of phonemes, Synthetic frequency, and Analytic
In order to be able to investigate the effect of the
frequency. Figure 2 displays the partial effects
processing complexity of the base adjective on the
for these interactions. The vertical axis shows the
preferred comparative variant, the same 41 speak-
transformed reaction times; higher values corre-
ers first participated in a visual lexical decision
spond to longer reaction times.
task that gathered reaction times for the 60 target
In agreement with figure 1, the partial effects
adjectives, as well as 150 other existing and non-
reveal significantly lower estimates for the syn-
existing distractor items. The participants were
thetic stimuli (solid lines) than for the analytic
not informed about the purpose of this task, and
stimuli (dashed lines). This is true even in the
there were at least 14 days for each participant be-
most adverse conditions (e.g. in cases in which
tween the lexical decision task and the production
the synthetic comparative of a comparative is at-
experiment. The reaction times obtained in this
tested only very rarely in a linguistic corpus, left
task were pooled for each adjective, and the me-
edge of lower right panel in figure 2).
dian was calculated.
3 Comparative variation in production
3.3 Results
3.1 Method
For most of the adjectives, the completion task
41 native speakers of Canadian English partici- was successful in obtaining comparative responses
pated individually in a spoken sentence comple- from the 41 speakers. However, two participants
tion task. The task used the same set of 60 ad- produced hardly any comparative in the task, and
jectives as in the perception experiment above, but were therefore excluded from the data set. 6 out
none of the participants in the production exper- of the 60 adjectives were excluded because the
iment had also participated in the previous task. responses contained almost exclusively synthetic
Participants were first shown a context sentence or analytic comparatives, or because the context
containing the adjective in the positive. After a key sentence did not elicit a considerable number of
press, an incomplete target sentence containing a comparative responses. 747 out of the remain-
blank and one or more target words appeared also ing 39 × 54 = 2106 responses contained a syn-
on the screen. The participants were instructed to thetic comparative (35 %), 843 contained an ana-
use the target words to fill the blank in the sen- lytic comparative (40 %). The remaining 516 re-
tence. If necessary, they could also use additional sponses (25 %) did not contain a comparative con-
words to complete the sentence. The sentences struction, and were discarded. There was notable
were constructed in such a way that a comparative variation between the two variants both across and
construction was the most likely target for comple- within items, which indicates that English compar-
tion, but participants were not explicitly instructed ative variation is indeed a highly non-deterministic
to use comparatives. The structure of the incom- field that is apparently affected by both speaker-
plete sentences was the same in all trials. The dependent and adjective-dependent factors.
subject was a simple noun phrase, followed by a Logistic general additive mixed-effects models
copula verb. The blank to be filled followed in (cf. Wood 2006) were used to investigate the re-
predicative position. This design ensured that the lation between the median RTs and the individual
context-dependent factors reported in the literature responses. These models have the advantage of re-
such as the increased probability of analytic com- vealing statistically significant effects of the inde-
paratives in predicative position were held con- pendent variable on the dependent even if the rela-
stant for all adjectives. Example (3) shows the ex- tion between them is not a linear one. For instance,
perimental trial for the target adjective wealthy. there could a threshold in the reaction times up to
which speakers strongly prefer the synthetic com-
(2) The duke is wealthy.
parative, but beyond which they shift to analytic
Yet, the king is .
comparatives in a nearly categorical way. In such
WEALTHY
a case, a linear model might fail to detect this non-
The experiment also contained 105 distractor linear effect of RTs on the responses.
trials that had a similar structure, but which did Two models were fitted: a null model which
not contain adjectives as the target words. contained only a random effect for speaker, and
34
a model with an additional smooth term for the 4 Discussion and conclusion
effect of the median RTs. If processing complex-
ity has a notable effect on speaker responses, the The results from the first experiment show that
smooth term should turn out to be statistically sig- synthetic comparatives have a clear perceptual
nificant, and the predictive accuracy of the model processing advantage over the analytic correspon-
should improve by the addition of the term. As dents. Even in conditions in which the morpho-
table 1 shows, this is indeed the case. While the logical form is particularly difficult to process, the
null model has a total predictive accuracy of about average reaction time is still faster than that for
69 %, the addition of the smooth term for median the phrasal variants. This finding makes it rather
RTs increases the accuracy by 5.6 %. There is a unlikely that the use of analytic comparatives in
larger increase of predictive accuracy for analytic cognitively demanding environments benefits the
responses than for synthetic responses (7.1 % vs. listener. Yet, the findings from the production ex-
3.9 %). periment reveal a significant relation between the
selected comparative form and the processing dif-
Synthetic Analytic Total ficulty of the adjective in question. For cognitively
Null model 515 580 1095 more complex adjectives which take longer to pro-
(68.9%) (68.8%) (68.9%) cess, the analytic comparative is preferred, sug-
Model with 544 640 1184 gesting that speakers resort to the phrasal alterna-
smooth term (72.8%) (75.9%) (74.5%) tive if processing demands are relatively high.
One aspect to keep in mind is that lexical de-
Table 1: Correctly predicted responses in the sen- cision tasks like those used above to collect reac-
tence completion task. tion times have a strong focus on form process-
ing, while they are less informative about func-
Figure 3 illustrates the contribution of the tional processing (see Yap et al. 2011 for a dis-
smooth term to the model. The vertical position cussion). Even if the perception experiment has
of the regression line indicates the predicted prob- shown that the analytic form is more difficult to
ability of analytic responses for the median RTs process for listeners, the higher explicitness of the
shown on the horizontal axis. The shaded area more comparative may still make the comparative
indicates the 95 % confidence band. As the fig- function more accessible for listeners than the -
ure shows, the relation between processing com- er comparative, which is also suggested by Mon-
plexity and comparative preference is indeed non- dorf (2009, 6). The experiments reported here do
linear: speakers strongly prefer the synthetic com- not address this issue of the comparative alterna-
parative for adjectives with very low RTs, but tend tion, but looking at functional accessibility offers
to favor the analytic comparative for adjectives a promising venue of future research.
with RTs larger than 600 ms. In sum, the produc- To conclude, the results imply that speakers and
tion experiment shows that the processing com- listeners process analytic and synthetic compar-
plexity of the base adjective has an effect on the atives differently: while the morphological form
preference of analytic comparatives by speakers. is easier to process for listeners, the phrasal form
has benefits for the speaker. More generally, these
1.0
findings also contribute toward our understanding
0.8 of morphosyntactic exponence. It is frequently ar-
P (analytic)
0.6 gued (e.g. in McWhorter 2001) that analytic forms
0.4 are less complex than synthetic forms, with conse-
0.2 quences for fields such as the structure of contact
0.0 languages or the diachronic development of a lan-
guage. This paper is one of the few that explicitly
500 600 700 800 900 address the processing efficiency of grammatical
Median reaction time (ms) variants where one form is morphological and the
other syntactic in nature. The findings suggest that
Figure 3: Effect of median reaction time on the the discussion of the alledged complexity of syn-
probability of analytic responses. thetic forms may also need to take into account
different demands of speakers and listeners.
35
Acknowledgments Benedikt Szmrecsanyi. 2005. Language users as crea-
tures of habit: A corpus-based analysis of persis-
This work was supported by the Deutsche For- tence in spoken English. Corpus Linguistics and
schungsgemeinschaft (grant KU 2896/1-1). I wish Linguistic Theory, 1(1):113–150.
to thank Ben Tucker (University of Alberta, Ed- Simon N. Wood. 2006. Generalized Additive Mod-
monton) for making available to me the facilities els. An introduction with R. Chapman & Hall/CRC,
of the Alberta Phonetics Laboratory for the exper- Boca Raton, FL.
iments reported in this paper. Melvin J. Yap, Sarah E. Tan, Penny M. Pexman, and
Ian S. Hargreaves. 2011. Is more always bet-
ter? effects of semantic richness on lexical decision,
References speeded pronunciation, and semantic classification.
Psychonomic Bulletin & Review, 18(4):742–750.
R. Harald Baayen and Petar Milin. 2010. Analyzing
reaction times. International Journal of Psycholog-
ical Research, 3(2):12–28.
David A. Balota, Melvin J. Yap, Michael J. Cortese,
Keith A. Hutchison, Brett Kessler, Bjorn Loftis,
James H. Neely, Douglas L. Nelson, Greg B. Simp-
son, and Rebecca Treiman. 2007. The En-
glish Lexicon Project. Behavior Research Methods,
39(3):445–459.
Jeremy Boyd. 2007. Comparatively speaking. A psy-
cholinguistic study of optionality in grammar. Ph.D.
thesis, University of California, San Diego.
George E. P. Box and David R. Cox. 1964. An analysis
of transformations. Journal of the Royal Statistical
Society. Series B, 26(2):211–252.
Mark Davies. 2008–. The Corpus of Con-
temporary American English (COCA): 450 mil-
lion words, 1990-present. Available online at
http://corpus.byu.edu/coca/.
Martin Hilpert. 2008. The English comparative. lan-
guage structure and language use. English Lan-
guage and Linguistics, 12(3):395–417.
Victor Kuperman, Hans Stadthagen-Gonzalez, and
Marc Brysbaert. 2012. Age-of-acquisition ratings
for 30,000 English words. Behavior Research Meth-
ods, 44(4):978–990.
John H. McWhorter. 2001. The world’s simplest
grammars are creole grammars. Linguistic Typol-
ogy, 5:125–166.
Britta Mondorf. 2009. More support for more-support.
John Benjamins, Amsterdam.
Britta Mondorf. 2014. Apparently competing
motivations in morpho-syntactic variation. In
Brian MacWhinney, Andrej Malchukov, and Edith
Moravcsik, editors, Competing motivations in gram-
mar and usage, pages 209–228. Oxford University
Press, Oxford.
Fermı́n Moscoso del Prado Martı́n, Aleksandar Kostić,
and R. Harald Baayen. 2004. Putting the bits
together. An information theoretical perspective on
morphological processing. Cognition, 94(1):1–18.
36