Czech swarms: experimental evidence
15-11-2018, UCL London
Mojmír Dočekal
Intro
1/40
Basic pattern
• swarm constructions (examples from Hoeksema (2009)):
(1) a. Termites are swarming in my kitchen. [A-Subject]
b. My kitchen is swarming with termites. [L-Subject]
2/40
Outline
1) Swarm constructions
a) basic properties
b) two theories
c) alleged PPI behaviour
2) Experiment on Czech swarms
3) Results: extreme degree construction
Slides: https://bit.ly/2QIHag2
3/40
Data
German swarms (Hoeksema 2009, ex.(6))
(2) a. Ameisen
Ants
wimmeln
swarm
in
in
der
the
Küche.
kitchen
‘Ants are swarming in the kitchen’
b. Die
the
Küche
kitchen
wimmelt
swarms
von
with
Ameisen.
ants
‘The kitchen is swarming with ants’
c. Es
it
wimmelt
swarms
von
with
Ameisen
ants
in
in
der
the
Küche.
kitchen
‘The kitchen is swarming with ants.’
4/40
Czech swarms:
(3) a. Na
on
té
the
louce
meadow
bzučely
swarmed.3PL
včely.
bees.PL.NOM
‘The bees swarmed on the meadow.’
b. Ta
the
louka
meadow
bzučela
swarmed.3SG
včelami.
bees.PL.INSTR
‘The meadow was swarming with bees.’
c. Na
on
té
the
louce
meadow
to
it
bzučelo
swarmed.3SG
včelami.
bees.PL.INSTR
‘The meadow was swarming with bees.’
5/40
Dowty (2000) 5 classes of swarms:
1) small local movements (repeated): crawl, drip, buble, dance,
foam, rumble, pulsate, ...
2) animal (and other) sounds (repetitive): hum, buzz, whistle,
resonate, echo, ...
3) kinds of light emission: beam, blaze, flame, glow, glitter, ...
4) smells and tastes: smell, taste, reek, ...
5) degree of occupancy/abundance: brim, teem, be rampant, ...
6/40
Two analyses
1) Dowty’s dynamic texture hypothesis
• locations described by predicates: small, frequently repeated
events
• transfer of events to locations
(4) My kitchen is swarming with termites.
• many events → many subregions (+ many agents)
• texture perception
• many agents: cannot be counted
7/40
• analogical to transfer of telicity in accomplishments
• nicely explains the constraint on objects: indefinites (mostly
bare plurals and mass nouns: swarming with insect)
• Dowty (2000, 123):
(5) a. The room swarmed with mosquitoes.
b. The room swarmed with a hundred mosquitoes.
c. ??The room swarmed with seventy-three mosquitoes.
d. My philodendron is crawling with dozens of snails.
e. ??My philodendron is crawling with fifty-seven snails.
8/40
• A-construction is basic/default: “purely compositional”,
“semantically unmarked”
• L-construction: semantically potent/marked
• Dowty: similar to middle alternations, conative alternation
• the basic construction (A-construction for swarms) is not
distinctive/marked
• analogically to middle alternations, etc.:
(6) a. Janet broke the vase.
b. Crystal vases break easily.
9/40
2) Hoeksema’s analysis
• impressive empirical evidence (Dutch, English, corpora, ...)
• against Dowty’s texture hypothesis: not all subjects are strictly
locative
• but swarms always express a high degree:
(7) a. Q: Was John angry?
b. A: He was foaming with fury.
(8) a. Q: Was the crowd loud?
b. A: The walls were vibrating with their cheers.
10/40
• Hoeksema: swarms = causative degree constructions
• “the object of with causes the subject to exhibit a high degree
of some property by completely affecting it”
• evidence: compatibility with high degree adverbs,
incompatibility with low-degree adv:
(9) a. The book is literally littered with typos.
b. The yard was absolutely lousy with vermin.
c. ??The book is somewhat littered with typos.
d. ??The yard was a bit lousy with vermin.
11/40
Note
• similar ideas for extreme adjectives: Morzycki (2012)
• adjectives as fantastic, fabulous, awesome, ... allow only
modifiers of extreme degree:
(10) a. It was an absolutely/??somewhat fantastic concert.
b. It was an utterly/??a bit awesome dinner.
12/40
• not easy to find decisive evidence (some experimental below)
• Hoeksema claims for (11):
(11) The book is littered with typos.
• high degree doesn’t predict some typo to be on every page
• Dowty: completely affected (every page) but doesn’t predict
high degree
• both approaches: only L-constructions have real swarm
properties (restrictions on type of objects, high degree,
markedness, ...)
13/40
Relation to PPIs
• Hoeksema (2018) lists swarms (among extreme adjectives,
some idioms, etc.) as high degree predicates: special sub-type
of PPIs
• Hoeksema (2018): some corpus evidence (Dutch, English) for
the L-construction
• Hoeksema (2009): swarms avoid negation and DE contexts
(similar observations for extreme degree adjectives: Morzycki
(2012))
• shares “high degree” property with Krifka’s TONS of money
14/40
The question behind the experiment:
(12) To what extent are swarm constructions PPIs?
15/40
Corpus evidence from Czech
• Czech national corpus (120 748 715 positions)
• 2 of 3 prototypical swarm predicates show signs of PPI-hood
16/40
bzučet ‘buzz’ vs. zpívat ‘sing’
# [lemma="bzučet"&tag="V.........A.*"] ... 356
# [lemma="bzučet"&tag="V.........N.*"] ... 7
# [lemma="zpívat"&tag="V.........A.*"] ... 7741
# [lemma="zpívat"&tag="V.........N.*"] ... 253
challenge.df = matrix(c(356,7741,7,253), nrow = 2)
colnames(challenge.df) = c("pos","neg")
rownames(challenge.df) = c("bzučet","zpívat")
fisher.test(challenge.df)
##
## Fisher's Exact Test for Count Data
##
## data: challenge.df
## p-value = 0.2167
## alternative hypothesis: true odds ratio is not equal to 1
## 95 percent confidence interval:
## 0.7866306 4.2080621
## sample estimates:
## odds ratio
## 1.66214
17/40
hemžit se ‘swarm’ vs. pohybovat se ‘move’
# [lemma="hemžit"&tag="V.........A.*"] ... 823
# [lemma="hemžit"&tag="V.........N.*"] ... 5
# [lemma="pohybovat"&tag="V.........A.*"] ... 12795
# [lemma="pohybovat"&tag="V.........N.*"] ... 227
challenge.df = matrix(c(823,12795,5,227), nrow = 2)
colnames(challenge.df) = c("pos","neg")
rownames(challenge.df) = c("hemžit","pohybovat")
fisher.test(challenge.df)
##
## Fisher's Exact Test for Count Data
##
## data: challenge.df
## p-value = 0.01096
## alternative hypothesis: true odds ratio is not equal to 1
## 95 percent confidence interval:
## 1.227920 9.104574
## sample estimates:
## odds ratio
## 2.920073
18/40
třást se ‘tremble’ vs. hýbat ‘move’
# [lemma="třást"&tag="V.........A.*"] ... 3368
# [lemma="třást"&tag="V.........N.*"] ... 136
# [lemma="hýbat"&tag="V.........A.*"] ... 1786
# [lemma="hýbat"&tag="V.........N.*"] ... 918
challenge.df = matrix(c(3368,1786,136,918), nrow = 2)
colnames(challenge.df) = c("pos","neg")
rownames(challenge.df) = c("hemžit","pohybovat")
fisher.test(challenge.df)
##
## Fisher's Exact Test for Count Data
##
## data: challenge.df
## p-value < 2.2e-16
## alternative hypothesis: true odds ratio is not equal to 1
## 95 percent confidence interval:
## 10.51357 15.48580
## sample estimates:
## odds ratio
## 12.72324
19/40
• some properties resemble Krifka’s TONS of money
• but other properties are different:
1) swarms are lexically derived: many gaps (no Czech swarm
construction for whistle, flame, taste, ...)
2) i/ani are totally grammaticalized (pure formal constraints, no
lexical gaps)
20/40
3) Czech swarms are probably not emphatic (no or weaker positive
(=¬¬) bias in questions vs. minimizers):
(13) You neg-have ani one drop? minimizers
a. Yes, I had one beer.
b. ??No, I didn’t have even radler.
(14) Neg-swarms.3SG it policemen.INST.PL? swarms
a. ?No, neg-BE n-word policeman.
b. Yes, be.3PL there many policemen.
21/40
Experiment
22/40
The experiment on swarms
• joint work with Iveta Šafratová
• acceptability task: 5-point Likert scale: 1-worst, 5-best
• 4x2 conditions design
• 32 items (+32 fillers)
• Latin-square design, IBEX farm
• 50 subjects, all passed fillers
23/40
Czech data (example item from the experiment)
Reference level condition
(15) Ta
the
louka
meadow
bzučela
swarmed.3SG
včelami.
bees.PL.INSTR
The meadow was swarming with bees. L-construction
(16) Na
on
té
the
louce
meadow
bzučely
swarmed.3PL
včely.
bees.PL.NOM
The bees swarmed on the meadow. A-construction
24/40
Degree condition
(17) Ta
the
louka
meadow
trochu
slightly
bzučela
swarmed.3SG
včelami.
bees.PL.INSTR
(18) Na
on
té
the
louce
meadow
trochu
slightly
bzučely
swarmed.3PL
včely.
bees.PL.NOM
25/40
Negation condition
(19) Ta
the
louka
meadow
nebzučela
neg-swarmed.3SG
včelami.
bees.PL.INSTR
(20) Na
on
té
the
louce
meadow
nebzučely
neg-swarmed.3PL
včely.
bees.PL.NOM
26/40
Rescuing condition
(21) Jestli
if
to
it
dnes
today
na
on
louce
meadow
nebzučí
neg-swarm.3SG
včelami,
bees.PL.INSTR,
tak
then
zítra
. . .
bude.
(22) Jestli
if
dnes
today
na
on
louce
meadow
nebzučí
neg-swarm.3PL
včely,
bees.PL.NOM,
tak
then
zítra
. . .
budou.
27/40
> ddply(data_part_1, .(Condition), summarise, Means = mean(Answer, na.rm=TRUE))
Condition Means
1 Deg-Inst 2.927083
2 Deg-Nom 3.541667
3 Neg-Inst 3.739583
4 Neg-Nom 4.239583
5 Ref-Inst 3.744792
6 Ref-Nom 4.619792
7 Resc-Inst 3.614583
8 Resc-Nom 4.078125
> ddply(data_part_1, .(Condition), summarise, Medians = median(Answer,na.rm=TRUE))
Condition Medians
1 Deg-Inst 3
2 Deg-Nom 4
3 Neg-Inst 4
4 Neg-Nom 5
5 Ref-Inst 4
6 Ref-Nom 5
7 Resc-Inst 4
8 Resc-Nom 5
>
28/40
q
q q
q
2.8
3.2
3.6
4.0
4.4
Deg Neg Ref Resc
Condition
Answer
Nom
q No
Yes
Figure 1: Error-bars, Experiment
29/40
Linear model for the experiment
> summary(m1)
Linear mixed model fit by REML. t-tests use Satterthwaite's method ['lmerModLmerTest']
Formula: as.numeric(Answer) ~ Condition * Nom + (1 | Subj) + (1 | Item)
Data: data_part_1
REML criterion at convergence: 4917.1
Scaled residuals:
Min 1Q Median 3Q Max
-3.5572 -0.5778 0.1593 0.7057 2.3157
Random effects:
Groups Name Variance Std.Dev.
Subj (Intercept) 0.1993 0.4464
Item (Intercept) 0.1897 0.4356
Residual 1.2932 1.1372
Number of obs: 1536, groups: Subj, 46; Item, 32
30/40
Fixed effects:
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 3.733e+00 1.306e-01 1.383e+02 28.587 < 2e-16 ***
ConditionDeg-Inst -8.175e-01 1.163e-01 1.453e+03 -7.030 3.16e-12 ***
ConditionDeg-Nom -1.808e-01 1.164e-01 1.454e+03 -1.554 0.12034
ConditionNeg-Inst 6.586e-03 1.165e-01 1.454e+03 0.057 0.95491
ConditionNeg-Nom 5.242e-01 1.164e-01 1.454e+03 4.505 7.18e-06 ***
ConditionRef-Nom 9.123e-01 1.166e-01 1.455e+03 7.824 9.78e-15 ***
ConditionResc-Inst -1.158e-01 1.163e-01 1.453e+03 -0.996 0.31958
ConditionResc-Nom 3.551e-01 1.166e-01 1.455e+03 3.046 0.00236 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
(Intr) CndD-I CndD-N CndN-I CndN-N CndtnRf-N CndR-I
CndtnDg-Ins -0.445
CondtnDg-Nm -0.446 0.498
CndtnNg-Ins -0.446 0.501 0.498
CondtnNg-Nm -0.446 0.500 0.501 0.498
CondtnRf-Nm -0.447 0.499 0.502 0.502 0.500
CndtnRsc-In -0.445 0.498 0.500 0.501 0.498 0.501
CndtnRsc-Nm -0.447 0.501 0.500 0.502 0.502 0.503 0.499
fit warnings:
fixed-effect model matrix is rank deficient so dropping 8 columns / coefficients
31/40
Summary of the experiment
• reference level condition: Ref-Inst
• against PPI-status: no significant difference against Neg-Inst
• rescuing doesn’t work: Resc-Inst even worse
• the construction is sensitive to degrees: Deg-Inst significantly
worse
• Nominative conditions are always better (default)
• degree construction with no real PPI behaviour (against
Hoeksema (2009) and Morzycki (2012))
32/40
Another model
• some hint of PPI-hood: Ref-Nom and Neg-Nom are significantly
different:
> data_part_1$Condition <- relevel(data_part_1$Condition, ref="Neg-Nom")
> m1a <- lmer(as.numeric(Answer) ~ Condition * Nom + (1|Subj) + (1|Item), data=data_part_1)
fixed-effect model matrix is rank deficient so dropping 8 columns / coefficients
> summary(m1a)
Linear mixed model fit by REML. t-tests use Satterthwaite's method ['lmerModLmerTest']
Formula: as.numeric(Answer) ~ Condition * Nom + (1 | Subj) + (1 | Item)
Data: data_part_1
...
Fixed effects:
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 4.2567 0.1306 138.2625 32.601 < 2e-16 ***
ConditionNeg-Inst -0.5176 0.1166 1455.3944 -4.439 9.73e-06 ***
ConditionRef-Inst -0.5242 0.1164 1453.7258 -4.505 7.18e-06 ***
ConditionDeg-Inst -1.3417 0.1164 1453.7258 -11.531 < 2e-16 ***
ConditionDeg-Nom -0.7050 0.1163 1453.2580 -6.063 1.71e-09 ***
ConditionRef-Nom 0.3881 0.1165 1454.4890 3.332 0.000883 ***
ConditionResc-Inst -0.6399 0.1166 1455.3944 -5.488 4.78e-08 ***
ConditionResc-Nom -0.1690 0.1163 1453.2580 -1.453 0.146332
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
33/40
Linguistic interpretation
• PPI-status very improbable
• reasons:
1) Krifka: PPI need emphatic/focus alternatives which are ordered
logically (or by likelihood)
– even [¬ ... ONE ...]: alternatives {1,2,3,...}
– results of the experiment: no logical ordering (INST >d
NOM/INST >d NOM) between nominative and instrumental
– no likelihood ordering
– covert Emph.Assert/even wouldn’t produce scalar
presupposition
34/40
2) the nature of swarm-alternatives is different than in case of
i/ani
3) interesting counter-evidence to Dowty’s claim:
– A-construction (Nominative) is semantically marked even if
default
– clearly high degree constructions in both L/A realizations
– if the mapping (to degrees) happens, then in both L/A
realizations
– problematic even for Hoeksema’s analysis
35/40
Research question:
(23) If both nominative and instrumental are high-degree
constructions, what is the difference between them?
36/40
37/40
Thanks!
38/40
Appendix
• similar case pattern as in spray-load alternations:
• total affectedness ≈ Instrumental
• Instrumental: high degree, not prone for low degree modifiers
(trochu ‘a bit’)
(24) Petr
Petr
naložil
loaded.3SG
seno
hay.ACC
na
on
vůz.
truck.ACC
‘Petr loaded the hay on the truck.’ (CHECK)
(25) Petr
Petr
naložil
loaded
vůz
truck.ACC
senem.
hay.INST
‘Petr loaded the truck with hay.’
39/40
References I
Dowty, David. 2000. “The Garden Swarms with Bees’ and the Fallacy of’argument
Alternation’.”
Hoeksema, Jack. 2009. “The Swarm Alternation Revisited.” Theory and Evidence in
Semantics, no. 189: 53.
———. 2018. “Positive Polarity Predicates.” Linguistics 56 (2): 361–400.
Morzycki, Marcin. 2012. “Adjectival Extremeness: Degree Modification and Contextually
Restricted Scales.” Natural Language & Linguistic Theory 30 (2): 567–609.
40/40