Mark Dingemanse & Bill Thompson (this version: 2019-12-03)
This code notebook provides a fully reproducible workflow for the paper Playful iconicity: Structural markedness underlies the relation between funniness and iconicity. To increase readability, not all code chunks present in the .Rmd source are shown in the output. A separate code notebook has the supplementary analyses.
Primary data sources:
- iconicity ratings: Perry, Lynn K. et al. Iconicity in the Speech of Children and Adults. Developmental Science. doi:10.1111/desc.12572
- funniness ratings: Engelthaler, Tomas, and Thomas T. Hills. 2017. Humor Norms for 4,997 English Words. Behavior Research Methods, July, 1-9. doi:10.3758/s13428-017-0930-6
We use these ratings in our analyses, but we also feed them to our imputation method, which regresses the human ratings against semantic vectors in order to generate imputed ratings for an additional 63.680 words.
Secondary data sources:
- number of morphemes: Balota, D. A., Yap, M. J., Hutchison, K. A., Cortese, M. J., Kessler, B., Loftis, B., … Treiman, R. (2007). The English Lexicon Project. Behavior Research Methods, 39(3), 445–459. doi: 10.3758/BF03193014
- word frequency: Brysbaert, M., New, B., & Keuleers, E. (2012). Adding part-of-speech information to the SUBTLEX-US word frequencies. Behavior Research Methods, 44(4), 991–997. doi: 10.3758/s13428-012-0190-4 (for word frequency)
- lexical decision times: Keuleers, E., Lacey, P., Rastle, K., & Brysbaert, M. (2012). The British Lexicon Project: Lexical decision data for 28,730 monosyllabic and disyllabic English words. Behavior Research Methods, 44(1), 287-304. doi: 10.3758/s13428-011-0118-4
- phonotactic measures: Vaden, K.I., Halpin, H.R., Hickok, G.S. (2009). Irvine Phonotactic Online Dictionary, Version 2.0. [Data file]. Available from http://www.iphod.com.
Secondary data sources used in supplementary analyses:
- valence, arousal and dominance: Warriner, A.B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal, and dominance for 13,915 English lemmas. Behavior Research Methods, 45, 1191-1207
- age of acquisition: Kuperman, V., Stadthagen-Gonzalez, H., & Brysbaert, M. (2012). Age-of-acquisition ratings for 30,000 English words. Behavior Research Methods, 44(4), 978-990. doi: 10.3758/s13428-012-0210-4
After collating these data sources we add a range of summary variables, mainly for easy plotting and subset selection.
words <- words %>%
mutate(fun_perc = ntile(fun,10),
fun_resid_perc = ntile(fun_resid,10),
ico_perc = ntile(ico,10),
diff_rank = fun_perc + ico_perc,
ico_imputed_perc = ntile(ico_imputed,10),
fun_imputed_perc = ntile(fun_imputed,10),
fun_imputed_resid_perc = ntile(fun_imputed_resid,10),
diff_rank_setB = fun_perc + ico_imputed_perc,
diff_rank_setC = fun_imputed_perc + ico_imputed_perc,
diff_rank_setD = fun_imputed_perc + ico_perc,
logletterfreq_perc = ntile(logletterfreq,10),
dens_perc = ntile(unsDENS,10),
biphone_perc = ntile(unsBPAV,10),
triphone_perc = ntile(unsTPAV,10),
posprob_perc = ntile(unsPOSPAV,10),
valence_perc = ntile(valence,10))
We have 4.996 words rated for funniness, 2.945 rated for iconicity, and 1.419 in the intersection (set A). We have 3.577 words with human funniness ratings and imputed iconicity ratings (set B). We have imputed data for a total of 70.202 words, and we’re venturing outside the realm of rated words for 63.680 of them (set C).
(We also have 1.526 words with human iconicity ratings and imputed funniness ratings in set D, the mirror image of set B; this is not used in the paper but reported on in Supplementary Analyses below.)
set |
n |
---|---|
A |
1419 |
B |
3577 |
C |
63680 |
D |
1526 |
## # A tibble: 3 x 9
## word ico fun ico_perc fun_perc ico_imputed fun_imputed
## <chr> <dbl> <dbl> <int> <int> <dbl> <dbl>
## 1 wigg~ 2.6 3.52 10 10 3.37 3.39
## 2 wobb~ 2.4 3.15 9 10 3.06 3.11
## 3 wagg~ NA NA NA NA 2.37 3.42
## # ... with 2 more variables: ico_imputed_perc <int>,
## # fun_imputed_perc <int>
The most important columns in the data are shown below for set A. Sets B
and C feature ico_imputed
and fun_imputed
instead of or in addition
to the human ratings. The field diff_rank
is the sum of fun
and
ico
deciles for a given word: a word with diff_rank
2 occurs in the
first decile (lowest 10%) of both funniness and iconicity ratings, and a
word with diff_rank
20 occurs in the 10th decile (highest 10%) of
both.
Structure of the data
word |
ico |
fun |
logletterfreq |
logfreq |
rt |
nmorph |
diff_rank |
---|---|---|---|---|---|---|---|
flop |
3.142857 |
3.031250 |
-3.223260 |
2.075547 |
587.9189 |
1 |
20 |
whine |
2.666667 |
2.833333 |
-2.706352 |
1.924279 |
588.6667 |
1 |
19 |
slip |
2.615385 |
2.586207 |
-2.978875 |
3.120903 |
546.2000 |
1 |
18 |
sigh |
2.800000 |
2.535714 |
-2.941718 |
2.240549 |
577.4595 |
1 |
17 |
must |
1.500000 |
2.636364 |
-2.952673 |
4.552206 |
569.8056 |
1 |
16 |
frog |
2.181818 |
2.440000 |
-3.097211 |
2.781037 |
533.2051 |
1 |
15 |
moose |
0.300000 |
3.103448 |
-2.622790 |
2.451786 |
550.5263 |
1 |
14 |
stretch |
2.400000 |
2.187500 |
-2.616122 |
2.874482 |
567.4615 |
1 |
13 |
block |
2.428571 |
2.153846 |
-3.366796 |
3.315551 |
537.0000 |
1 |
12 |
lark |
-0.900000 |
3.025641 |
-3.000745 |
1.924279 |
607.9375 |
1 |
11 |
For a quick impression of the main findings, this section reproduces the figures from the paper.
Figure 3: Funniness and iconicity
Figure 4: Highest rated words
Figure 5: Structural markedness
Engelthaler & Hills report frequency as the strongest correlate with funniness (less frequent words are rated as more funny), and lexical decision RT as the second strongest (words with slower RTs are rated as more funny). By way of sanity check let’s replicate their analysis.
Raw correlations hover around 28%, as reported (without corrections or controls) in their paper. A linear model with funniness as dependent variable and frequency and RT as predictors shows a role for both, though frequency accounts for a much larger portion of the variance (15%) than rt (0.6%).
To what extent do frequency and RT predict funniness?
Model m0: lm(formula = fun ~ logfreq + rt, data = words %>% drop_na(fun))
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
78.329 |
454.096 |
0 |
0.083 |
rt |
1 |
17.315 |
100.380 |
0 |
0.020 |
Residuals |
4993 |
861.264 |
If frequency and RT explain some of the variance in funniness ratings, how much is left for iconicity? We’ll do this analysis on the core set of 1419 words for which we have funniness and iconicity ratings.
Turns out that the magnitude estimate of iconicity is about half that of frequency, and with positive sign instead of a negative one (higher funniness ratings go with higher iconicity ratings). The effect of iconicity ratings is much larger than RT, the second most important correlate reported by Engelthaler & Hill.
Model m1.1: lm(formula = fun ~ logfreq + rt, data = words %>% filter(set == , “A”))
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
36.143 |
247.813 |
0.000 |
0.149 |
rt |
1 |
1.249 |
8.562 |
0.003 |
0.006 |
Residuals |
1416 |
206.519 |
Model m1.2: lm(formula = fun ~ logfreq + rt + ico, data = words %>% filter(set == , “A”))
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
36.143 |
258.779 |
0.000 |
0.155 |
rt |
1 |
1.249 |
8.941 |
0.003 |
0.006 |
ico |
1 |
8.891 |
63.661 |
0.000 |
0.043 |
Residuals |
1415 |
197.628 |
model comparison of m1.1 and m1.2
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
1416 |
206.5194 |
||||
1415 |
197.6281 |
1 |
8.891332 |
63.66118 |
0 |
Partial correlations show 20.6% covariance between funniness and iconicity, partialing out log frequency as a mediator. This shows the effects of iconicity and funniness are not reducible to frequency alone.
funniness and iconicity controlling for word frequency
estimate |
p.value |
statistic |
n |
gp |
Method |
---|---|---|---|---|---|
0.2064276 |
0 |
7.938811 |
1419 |
1 |
pearson |
Example words
Both high: zigzag, squeak, chirp, pop, clunk, moo, clang, oink, zoom, smooch, babble, squawk, thud, gush, fluff, flop, waddle, giggle, tinkle, ooze
Both low: silent, statement, poor, cellar, incest, window, lie, coffin, platform, address, slave, wait, year, case
High funniness, low iconicity: belly, buttocks, beaver, chipmunk, turkey, bra, hippo, chimp, blonde, penis, pun, dingo, trombone, deuce, lark, gander, magpie, tongue, giraffe, hoe
High iconicity, low funniness: click, roar, crash, chime, scratch, swift, sunshine, low, break, clash, shoot, airplane, dread
N.B. controlling for frequency in these lists (by using fun_resid
instead of fun
) does not make a difference in ranking, so not done
here and elsewhere.
What about compound nouns among high iconicity words? From eyeballing, it seems to be about 10% in a set of the highest rated 200 nouns. Many probable examples can be found by looking at highly rated nouns with multiple morphemes: zigzag, buzzer, skateboard, sunshine, zipper, freezer, snowball, juggler, airplane, bedroom, goldfish, seaweed, lipstick, mixer, corkscrew, doorknob, killer, moonlight, tummy, kingdom, razor, singer, ashtray, fireworks, pliers, racer, uproar (zigzag, one of the few reduplicative words in English, is included here because the Balota et al. database lists it as having 2 morphemes).
Here we study the link between funniness ratings and imputed iconicity ratings.
Compared to model m2.1 with just log frequency and lexical decision time as predictors, model m2.2 including imputed iconicity as predictor provides a significantly better fit and explains a larger portion of the variance.
Model m2.1: lm(formula = fun ~ logfreq + rt, data = words.setB)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
39.528 |
218.214 |
0 |
0.058 |
rt |
1 |
20.487 |
113.100 |
0 |
0.031 |
Residuals |
3574 |
647.400 |
Model m2.2: lm(formula = fun ~ logfreq + rt + ico_imputed, data = words.setB)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
39.528 |
245.736 |
0 |
0.064 |
rt |
1 |
20.487 |
127.365 |
0 |
0.034 |
ico_imputed |
1 |
72.669 |
451.769 |
0 |
0.112 |
Residuals |
3573 |
574.731 |
model comparison
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
3574 |
647.3997 |
||||
3573 |
574.7308 |
1 |
72.66885 |
451.7694 |
0 |
A partial correlations analysis shows that imputed iconicity values correlate with funniness ratings at at least the same level as actual iconicity ratings, controlling for frequency (r = 0.32, p < 0.0001).
Example words
High imputed funniness and high imputed iconicity: swish, chug, bop, gobble, smack, blip, whack, oomph, poke, wallop, funk, chuckle, quickie, wriggle, quiver, scamp, burp, hooky, oodles, weasel
Low imputed funniness and low imputed iconicity: subject, ransom, libel, bible, siege, hospice, conduct, arsenic, clothing, negro, mosque, typhoid, request, expense, author, length, anthrax, mandate, plaintiff, hostage
High funniness and low imputed iconicity: heifer, dinghy, cuckold, nudist, sheepdog, oddball, spam, harlot, getup, rickshaw, sac, kiwi, whorehouse, soiree, condom, plaything, croquet, charade, fiver, loch
Low funniness and high imputed iconicity: shudder, scrape, taps, fright, heartbeat, puncture, choke, tremor, biceps, glimpse, disgust, doom, stir, dent, scold, bully, reign, blister, check, horror
What about analysable compounds among high iconicity nouns? Here too about 10%, with examples like heartbeat, mouthful, handshake, bellboy, comeback, catchphrase.
Model 3.1: lm(formula = fun_imputed ~ logfreq + rt, data = words.setC)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
92.862 |
1018.093 |
0 |
0.050 |
rt |
1 |
13.422 |
147.150 |
0 |
0.008 |
Residuals |
19204 |
1751.629 |
Model 3.2: lm(formula = fun_imputed ~ logfreq + rt + ico_imputed, data = words.setC)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
92.862 |
1258.529 |
0 |
0.062 |
rt |
1 |
13.422 |
181.901 |
0 |
0.009 |
ico_imputed |
1 |
334.714 |
4536.279 |
0 |
0.191 |
Residuals |
19203 |
1416.915 |
model comparison
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
19204 |
1751.629 |
||||
19203 |
1416.915 |
1 |
334.7145 |
4536.279 |
0 |
Partial correlations show that imputed iconicity and imputed funniness share 43% covariance not explained by word frequency.
imputed funniness and imputed iconicity controlling for word frequency
estimate |
p.value |
statistic |
n |
gp |
Method |
---|---|---|---|---|---|
0.4274166 |
0 |
119.302 |
63680 |
1 |
pearson |
Example words
High imputed funniness and high imputed iconicity: whoosh, whirr, whooshing, brr, argh, chomp, whir, swoosh, brrr, zaps, squeaks, whirring, squelchy, gulps, smacking, growls, clanks, squish, whoo, clop
Low imputed funniness and low imputed iconicity: apr, dei, covenants, palestinians, covenant, clothier, palestinian, variant, mitochondria, israelis, serb, sufferers, herein, isotope, duration, ciudad, appellant, palestine, alexandria, infantrymen
High imputed funniness and low imputed iconicity: pigs, monkeys, herr, raja, franz, lulu, von, beau, caviar, penguins, elves, virgins, lesbians, fez, amuse, hawaiian, hens, salami, perverts, gertrude
Low imputed funniness and high imputed iconicity: slashes, gunshots, footstep, cries, footsteps, fade, froze, cr, swelter, crushing, piercing, shoots, breathing, sobs, tremors, strokes, choking, slammed, shocked, ng
What about compound nouns here? In the top 200 nouns we can spot ~5 (shockwave, doodlebug, flashbulb, backflip, footstep) but that is of course a tiny tail end of a much larger dataset than the earlier two.
A better way is to sample 200 random nouns from a proportionate slice of the data, i.e. 200 * 17.8 = 3560 top nouns in imputed iconicity. In this subset we find at least 30 non-iconic analysable compounds: fireworm, deadbolt, footstep, pockmark, uppercut, woodwork, biotech, notepad, spellbinder, henchmen, quicksands, blowgun, heartbreaks, moonbeams, sketchpad, et cetera.
words.setC %>%
filter(ico_imputed_perc > 9,
POS == "Noun") %>%
arrange(-ico_imputed) %>%
slice(1:200) %>%
dplyr::select(word) %>% unlist %>% unname()
set.seed(1983)
words.setC %>%
filter(ico_imputed_perc > 9,
POS == "Noun") %>%
arrange(-ico_imputed) %>%
slice(1:3560) %>%
sample_n(200) %>%
dplyr::select(word) %>% unlist %>% unname()
Mean iconicity and mean funniness are higher for lower log letter frequency quantiles:
Mean funniness and iconicity by log letter frequency quantiles
logletterfreq_perc |
mean_ico |
mean_fun |
---|---|---|
1 |
1.2562724 |
2.510892 |
2 |
1.0972947 |
2.434144 |
3 |
0.9435569 |
2.339590 |
4 |
0.7677072 |
2.313565 |
5 |
0.6163793 |
2.323666 |
6 |
0.7206575 |
2.286704 |
7 |
0.7950753 |
2.361308 |
8 |
0.8434129 |
2.284869 |
9 |
0.7531960 |
2.249879 |
10 |
0.5100479 |
2.273432 |
High-iconicity high-funniness words tend to have lower log letter frequencies:
Log letter frequency percentiles for upper quantiles of funniness + iconicity
word |
fun |
ico |
diff_rank |
logletterfreq_perc |
---|---|---|---|---|
zigzag |
3.113636 |
4.300000 |
20 |
1 |
squeak |
3.230769 |
4.230769 |
20 |
2 |
chirp |
3.000000 |
4.142857 |
20 |
1 |
buzzer |
2.833333 |
4.090909 |
19 |
1 |
pop |
3.294118 |
4.076923 |
20 |
1 |
bleep |
2.931818 |
3.928571 |
19 |
6 |
clunk |
3.344828 |
3.928571 |
20 |
1 |
moo |
3.700000 |
3.882353 |
20 |
4 |
clang |
3.200000 |
3.857143 |
20 |
2 |
boom |
2.829268 |
3.846154 |
19 |
1 |
bang |
2.843750 |
3.833333 |
19 |
1 |
murmur |
2.812500 |
3.833333 |
19 |
1 |
whirl |
2.911765 |
3.818182 |
19 |
2 |
crunch |
2.857143 |
3.785714 |
19 |
1 |
rip |
2.827586 |
3.736842 |
19 |
2 |
sludge |
2.875000 |
3.700000 |
19 |
2 |
ping |
2.875000 |
3.636364 |
19 |
1 |
oink |
3.871795 |
3.615385 |
20 |
3 |
zoom |
3.043478 |
3.600000 |
20 |
1 |
smooch |
3.333333 |
3.600000 |
20 |
3 |
Model comparison with funniness as the DV and log letter frequency as an additional predictor shows that a model including log letter frequency provides a significantly better fit.
Model m4.1: lm(formula = fun ~ logfreq + rt + ico, data = words %>% filter(set == , “A”))
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
36.143 |
258.779 |
0.000 |
0.155 |
rt |
1 |
1.249 |
8.941 |
0.003 |
0.006 |
ico |
1 |
8.891 |
63.661 |
0.000 |
0.043 |
Residuals |
1415 |
197.628 |
Model m4.2: lm(formula = fun ~ logfreq + rt + ico + logletterfreq, data = words %>% , filter(set == “A”))
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
36.143 |
265.179 |
0.000 |
0.158 |
rt |
1 |
1.249 |
9.162 |
0.003 |
0.006 |
ico |
1 |
8.891 |
65.236 |
0.000 |
0.044 |
logletterfreq |
1 |
4.906 |
35.994 |
0.000 |
0.025 |
Residuals |
1414 |
192.722 |
model comparison
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
1415 |
197.6281 |
||||
1414 |
192.7222 |
1 |
4.905856 |
35.9942 |
0 |
Partial correlations show that funniness rating and log letter frequency have a covariance of -15.7% controlling for iconicity, and that iconicity and log letter frequency have a covariance of -16.3% controlling for funniness ratings (all p < 0.0001 correcting for multiple comparisons).
funniness and log letter frequency controlling for iconicity
estimate |
p.value |
statistic |
n |
gp |
Method |
---|---|---|---|---|---|
-0.157001 |
0 |
-5.982098 |
1419 |
1 |
pearson |
iconicity and log letter frequency controlling for funniness
estimate |
p.value |
statistic |
n |
gp |
Method |
---|---|---|---|---|---|
-0.1634579 |
0 |
-6.234739 |
1419 |
1 |
pearson |
Model comparison for combined funniness and iconicity scores suggests that having log letter frequency as a predictor significantly improves fit over and above word frequency and lexical decision time.
Model m5.1: lm(formula = funico ~ logfreq + rt, data = words.setA)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
420.245 |
206.078 |
0.0 |
0.127 |
rt |
1 |
5.516 |
2.705 |
0.1 |
0.002 |
Residuals |
1416 |
2887.579 |
Model m5.2: lm(formula = funico ~ logfreq + rt + logletterfreq, data = words.setA)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
420.245 |
219.963 |
0.00 |
0.135 |
rt |
1 |
5.516 |
2.887 |
0.09 |
0.002 |
logletterfreq |
1 |
184.189 |
96.407 |
0.00 |
0.064 |
Residuals |
1415 |
2703.390 |
model comparison
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
1416 |
2887.579 |
||||
1415 |
2703.390 |
1 |
184.1887 |
96.40745 |
0 |
We carry out a qualitative analysis of the 80 highest ranked words (top deciles for funniness+iconicity) to see if there are formal cues of foregrounding and structural markedness that can help predict funniness and iconicity ratings. Then we find these cues in the larger dataset and see if the patterns hold up.
This analysis reveals the following sets of complex onsets, codas, and verbal diminutive suffixes that are likely structural cues of markedness (given here in the form of regular expressions):
- onsets:
^(bl|cl|cr|dr|fl|sc|sl|sn|sp|spl|sw|tr|pr|sq)
- codas:
(nch|mp|nk|rt|rl|rr|sh|wk)$
- verbal suffix:
[b-df-hj-np-tv-xz]le)$
" (i.e., look for -le after a consonant)
We tag these cues across the whole dataset (looking for the -le suffix only in verbs because words like mutable, unnameable, scalable, manacle are not the same phenomenon) in order to see how they relate to funniness and iconicity.
Model the contribution of markedness relative to logletter frequency. Model comparison shows that a model including the measure of cumulative markedness as predictor provides a significantly better fit (F = 52.78, p < 0.0001) and explains a larger portion of the variance (adjusted R2 = 0.21 vs. 0.18) than a model with just word frequency, lexical decision time and log letter frequency.
Model m5.3: lm(formula = funico ~ logfreq + rt + logletterfreq + cumulative, , data = words.setA)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
420.245 |
228.013 |
0.000 |
0.139 |
rt |
1 |
5.516 |
2.993 |
0.084 |
0.002 |
logletterfreq |
1 |
184.189 |
99.936 |
0.000 |
0.066 |
cumulative |
1 |
97.283 |
52.783 |
0.000 |
0.036 |
Residuals |
1414 |
2606.107 |
Model comparison of m5.2 and m5.3
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
1415 |
2703.390 |
||||
1414 |
2606.107 |
1 |
97.28312 |
52.78307 |
0 |
Now we trace cumulative markedness in the imputed portions of the dataset, and do the same model comparison as above.
First have a look at a random sample of top imputed words and their markedness:
Cumulative markedness in a random sample of words from the highest quantile of imputed iconicity
word |
ico_imputed_perc |
ico_imputed |
cumulative |
---|---|---|---|
brr |
10 |
4.065481 |
1 |
squish |
10 |
3.570914 |
2 |
clunks |
10 |
2.965891 |
1 |
scamp |
10 |
2.397420 |
2 |
spank |
10 |
2.360551 |
2 |
squoosh |
10 |
2.312993 |
2 |
crunk |
10 |
2.165342 |
2 |
sw |
10 |
1.898677 |
1 |
flipping |
10 |
1.875252 |
1 |
flatly |
10 |
1.819130 |
1 |
crumping |
10 |
1.737702 |
1 |
flourish |
10 |
1.722582 |
2 |
crispy |
10 |
1.721382 |
1 |
snappish |
10 |
1.612547 |
2 |
flush |
10 |
1.598260 |
2 |
scrumptious |
10 |
1.491476 |
1 |
blank |
10 |
1.435437 |
2 |
tramp |
10 |
1.426469 |
2 |
speakeasy |
10 |
1.393087 |
1 |
scornfully |
10 |
1.366685 |
1 |
And at a random sample of words from lower quadrants and their markedness:
Cumulative markedness in a random sample of words from lower quantiles of imputed iconicity
word |
ico_imputed_perc |
ico_imputed |
cumulative |
---|---|---|---|
spoilsport |
7 |
0.7898544 |
2 |
draughted |
7 |
0.7022719 |
1 |
drank |
6 |
0.6164557 |
2 |
sweetfish |
6 |
0.5390918 |
2 |
protectress |
6 |
0.5275526 |
1 |
transmits |
5 |
0.4734109 |
1 |
schrank |
5 |
0.4676592 |
2 |
blamed |
5 |
0.4105656 |
1 |
flatfish |
5 |
0.3870658 |
2 |
crystallised |
5 |
0.3525224 |
1 |
trench |
4 |
0.2827599 |
2 |
preshrunk |
3 |
0.1581131 |
2 |
spectroscopic |
3 |
0.1460491 |
1 |
splendours |
3 |
0.0495967 |
1 |
prelaunch |
3 |
0.0355877 |
2 |
spearfish |
2 |
0.0134083 |
2 |
flemish |
2 |
-0.0374782 |
2 |
flamingos |
2 |
-0.0452134 |
1 |
cryptography |
2 |
-0.1193397 |
1 |
triangulation |
2 |
-0.1590728 |
1 |
Looks like random samples of 20 high-complexity words always feature a majority of high iconicity words:
Imputed ratings for 20 random words high in cumulative markedness
word |
ico_imputed_perc |
ico_imputed |
fun_imputed |
cumulative |
---|---|---|---|---|
squoosh |
10 |
2.3129932 |
2.945588 |
2 |
squirt |
10 |
2.1139378 |
3.302116 |
2 |
crump |
10 |
1.7575309 |
3.072162 |
2 |
snaffle |
10 |
1.6898653 |
3.005494 |
2 |
crank |
10 |
1.6032802 |
2.772123 |
2 |
flush |
10 |
1.5982598 |
2.668945 |
2 |
clump |
10 |
1.5840648 |
2.744887 |
2 |
spangle |
10 |
1.5792685 |
3.046803 |
2 |
scribble |
10 |
1.5335878 |
2.832759 |
2 |
swink |
10 |
1.5061947 |
3.015406 |
2 |
tramp |
10 |
1.4264688 |
2.899633 |
2 |
slapdash |
9 |
1.3323017 |
2.544342 |
2 |
prank |
8 |
0.9101965 |
3.091282 |
2 |
crawfish |
8 |
0.8564163 |
2.726631 |
2 |
sweetheart |
8 |
0.8273857 |
2.711589 |
2 |
spinsterish |
6 |
0.5475747 |
2.329725 |
2 |
scrimp |
5 |
0.4658493 |
2.534127 |
2 |
flatfish |
5 |
0.3870658 |
2.628965 |
2 |
prelaunch |
3 |
0.0355877 |
2.304523 |
2 |
scottish |
1 |
-0.2946582 |
2.556597 |
2 |
Let’s have a closer look at subsets. First quadrants, then deciles.
Markedness cues across quartiles of imputed iconicity
target_perc |
n |
onset |
coda |
verbdim |
complexity |
---|---|---|---|---|---|
1 |
15920 |
0.0639447 |
0.0060302 |
0.0003769 |
0.0703518 |
2 |
15920 |
0.0731784 |
0.0076005 |
0.0009422 |
0.0817211 |
3 |
15920 |
0.0923367 |
0.0097362 |
0.0009422 |
0.1030151 |
4 |
15920 |
0.1583543 |
0.0155151 |
0.0049623 |
0.1788317 |
Markedness cues across deciles of imputed iconicity
target_perc |
n |
onset |
coda |
verbdim |
complexity |
---|---|---|---|---|---|
1 |
6368 |
0.0565327 |
0.0061244 |
0.0003141 |
0.0629711 |
2 |
6368 |
0.0684673 |
0.0045540 |
0.0004711 |
0.0734925 |
3 |
6368 |
0.0714510 |
0.0084799 |
0.0001570 |
0.0800879 |
4 |
6368 |
0.0675251 |
0.0069095 |
0.0010992 |
0.0755339 |
5 |
6368 |
0.0788317 |
0.0080088 |
0.0012563 |
0.0880967 |
6 |
6368 |
0.0819724 |
0.0078518 |
0.0006281 |
0.0904523 |
7 |
6368 |
0.0978329 |
0.0105214 |
0.0010992 |
0.1094535 |
8 |
6368 |
0.1103957 |
0.0111495 |
0.0025126 |
0.1240578 |
9 |
6368 |
0.1350503 |
0.0144472 |
0.0028266 |
0.1523241 |
10 |
6368 |
0.2014761 |
0.0191583 |
0.0076947 |
0.2283291 |
Comparison of models with combined imputed funniness and iconicity as a dependent variable shows that a linear model including cumulative markedness as predictor provides a significantly better fit (F1,19230 = 337.3, p < 0.0001) and explains a little bit more the variance (adjusted R2 = 0.124 vs. 0.109) than a model with just word frequency, lexical decision time and log letter frequency.
Model m5.4: lm(formula = funico_imputed ~ logfreq + rt + logletterfreq, data = words.setC)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
1025.303 |
361.774 |
0.000 |
0.018 |
rt |
1 |
4.778 |
1.686 |
0.194 |
0.000 |
logletterfreq |
1 |
5608.765 |
1979.029 |
0.000 |
0.093 |
Residuals |
19203 |
54423.210 |
Model m5.5: lm(formula = funico_imputed ~ logfreq + rt + logletterfreq + , cumulative, data = words.setC)
predictor |
df |
SS |
(F) |
(p) |
partial (\eta^2) |
---|---|---|---|---|---|
logfreq |
1 |
1025.303 |
368.110 |
0.00 |
0.019 |
rt |
1 |
4.778 |
1.716 |
0.19 |
0.000 |
logletterfreq |
1 |
5608.765 |
2013.687 |
0.00 |
0.095 |
cumulative |
1 |
939.486 |
337.299 |
0.00 |
0.017 |
Residuals |
19202 |
53483.724 |
model comparison
Res.Df |
RSS |
Df |
Sum of Sq |
F |
Pr(>F) |
---|---|---|---|---|---|
19203 |
54423.21 |
||||
19202 |
53483.72 |
1 |
939.4858 |
337.299 |
0 |
Thanks for your interest. Also see the separate code notebook with�supplementary analyses.
If you find this useful, consider checking out the following resources that have been helpful in preparing this Rmarkdown document:
- Two of my own past projects (remember, the person most grateful for
your well-documented past code is future you):
- Expressiveness and grammatical integration (by Mark Dingemanse)
- Coloured vowels: open data and code (by Mark Dingemanse & Christine Cuskley)
- Formatting ANOVA tables in R (by Rose Hartman, Understanding Data)
- Iconicity in the speech of children and adults (by Bodo Winter)
- English letter frequencies
And of course have a look at the paper itself — latest preprint here: Playful iconicity