Severe Storm Warnings for Four-Story Homeowners

Towards a Processing Model of Bracketing Paradoxes

Anna Prysłopska and Titus von der Malsburg

University of Stuttgart

2024-06-11

Why are some phrases ambiguous or preferentially interpreted as a bracketing paradox?

steiniger Fahrradweg stony bike.path
→ only a canonical reading

königliche Hochzeitsfeier royal wedding.celebration
→ both canonical and bracketing paradox readings

saurer Kirschbaum sour cherry.tree
→ only a bracketing paradox reading

Why are some bracketing paradoxes odd and others unremarkable?

schwere Unwetterwarnung severe storm.warning
→ warning for severe weather vs. severe warning for storms

? rustikale Wurstplatte rustic sausage.plate
→ a plate of rustic sausage vs. rustic plate with sausage(s)

?? dreiköpfger Familienvater three-headed family.father
→ the father of a family of 3 vs. Cerberus

?? vierstöckiger Hausbesitzer four-story home.owner
→ the owner of a 4-story home vs. gigantic owner of a home

??? fette Hennengasse fat hen.alley
→ the alley of the fat hen vs. the fat alley of the hen(s)

How do we model these phrases?

Do they violate the principle of compositionality? Is this an illusion? A garden-path effect? A metaphor? What are their components? What makes them possible and passable?

How are they processed?

There is barely any psycholinguistic research on this phenomenon!

Experiment 1

Does semantic compatibility between the adjective and the individual nouns affect the acceptability of the entire phrase?

psychologische
psychological

Beratungs-
counseling

stelle
center

AN₁N₂ Psychologische Beratungsstelle
psychological counseling.center

AN₁ Psychologische Beratung
psychological counseling

AN₂ Psychologische Stelle
psychological center

204 items (from literature and newspapers) in three conditions in 3 Latin square design lists

🚫 AN₁N₂ constructions where the adjective didn’t match either noun → unlikely to be used.

Predictions

Strict grammaticality and compositionality: the ratings for the AN₂ should be predicative of the AN₁N₂ rating.
→ N₂ is the head of the compound

Adjective 🩷 N₂ → canonical
good ratings, straightforward and compositional

Adjective 🩷 N₁ → (possibly) bracketing paradox
❔ ratings: bracketing paradoxes are a spectrum from unassuming to bizarre

N₁ 🩷 Adjective 🩷 N₂ → ambiguous, maybe some bracketing paradox potential

Predictions

The adjective’s morphosyntactic features necessarily match those of N₂

The adjective’s morphosyntactic features also match N₁ → bracketing paradox potential

königliche_fem/pl	Hochzeits_fem	feier_fem	✔️
schwere_fem/pl	Unwetter_neut/pl	warnung_fem	✔️
saurer_masc	Kirsch_fem	baum_masc	❌

Analysis

The ratings across scales were highly correlated (r ≥ 0.95)

→ We used the mean and scaled it to [0, 1]

Analysis

Annotated the morphosyntactic agreement between the adjective and N₁

✔️psychologische Beratungsstelle
✔️königliche Hochzeitsfeier
✔️rustikale Wurstplatte
✔️schwere Unwetterwarnung
❌saurer Kirschbaum
❌steiniger Fahrradweg
❌dreiköpfiger Familienvater
❌vierstöckiger Hausbesitzer

Analysis

AN₁N₂ ratings as a function of AN₁ and AN₂ ratings, morphosyntactic match between the adjective and N₁, and their 2-way interactions

saurer Kirschbaum_rating ~saure Kirsche_rating vs. saurer Baum_rating vs. A-N₁ morphosyntactic match

brm(
  bf(AN1N2rating ~ (AN1rating + AN2rating + AN1match)^2,
     phi   ~ (AN1rating + AN2rating + AN1match)^2),
  data    = experiment1,
  family  = Beta(),
  iter    = 3000,
  warmup  = 1000,
  chains  = 4,
  cores   = 4,
  backend = "cmdstanr")

Results

AN₁ rating vs. AN₂ rating

Results

AN₁N₂ rating vs. AN₂ rating

Adjective 🩷 N₂ semantically
→ positive effect on AN₁N₂ rating ✔️

Results

AN₁N₂ vs. AN₁ × AN₂ ratings

Adjective 🩷 N₂ semantically
→ positive effect on AN₁N₂ rating ✔️

Adjective 🩷 N₁ semantically
→ positive effect on AN₁N₂ rating ❗

N₁ 🩷 Adjective 🩷 N₂ semantically
→ negative effect AN₁N₂ rating ❗

❔ conflict between the possible attachment sites

Results

A+N₁ Morphosyntatic match

Adjective 🩷 N₂ semantically
→ positive effect on AN₁N₂ rating ✔

Adjective 🩷 N₁ semantically
→ positive effect on AN₁N₂ rating ❗

N₁ 🩷 Adjective 🩷 N₂ semantically
→ negative effect AN₁N₂ rating ❗

N₁ 🩷 Adjective 🩷 N₂ morphosyntactically
→ no impact on paradox potential ❌

Results

	Est.	Est. error	95% CrI
Intercept	−4.05	0.77	−5.61, −2.60
AN₁ rating	3.37	0.89	1.68, 5.17
AN₂ rating	6.33	0.90	4.61, 8.13
morph. match	−0.06	0.58	−1.21, 1.09
AN₁ rating × AN₂ rating	−4.01	1.08	−6.16, −1.92
AN₁ rating × morph. match	0.14	0.48	−0.80, 1.11
AN₂ rating × morph. match	−0.12	0.51	−1.10, 0.88

Discussion

Though both nouns have a positive influence on the compound’s acceptability, their effects are not strictly additive.

→ neidischer Zickenkrieg jealous bitch.war❔

In the absence of a suitable (N₂) head noun candidate, the first noun (N₁) becomes an attractive target for the adjective.

The comprehensibility, naturalness and stylistic form ratings were highly correlated, indicating that they record similar concepts.

There was no evidence for an effect of morphosyntactic match between adjective and N₁.

Experiment 2

Which noun in a compound is modified by the adjective?

Predictions

Grammar: the adjective is always the attribute of N₂

When N₂ is chosen more often than N₁ → canonical

When N₁ is chosen more often than N₂ → bracketing paradox

When N₁ is chosen as often as N₂ → ambiguous

When ? is chosen often → 🚫 interpretation?

When adjective 🩷 N₁ morphosyntactically → adjective attaches more easily to N₁

The semantic compatibility ratings from Experiment 1 should align with the selected attachment site.

Analysis

Rate of N₁ attachment as a function of the corresponding AN₁ and AN₂ ratings from Experiment 1, morphosyntactic match between the adjective and N₁, and their 2-way interactions

saurer Kirschbaum_attachment ~ saure Kirsche_rating vs. saurer Baum_rating vs. A-N₁ morphosyntactic match

brm(
  bf(N1attachment ~ (AN1rating + AN2rating + AN1match)^2,
     phi   ~ (AN1rating + AN2rating + AN1match)^2),
  data    = experiment2,
  family  = Beta(),
  iter    = 3000,
  warmup  = 1000,
  chains  = 4,
  cores   = 4,
  backend = "cmdstanr")

Results

Attachment preference

Grammatically: the adjective is always the attribute of N₂ ❌

N₂ was chosen more often than N₁ overall → grammatical ✔️

N₁ was sometimes chosen over N₂ → bracketing paradox ✔️

~~When ? is chosen often → 🚫 interpretation?~~

Results

Attachment preference

The prime bracketing paradox examples from linguistic literature weren’t overwhelmingly N₁ ❗

Abramov, B. (1992). Nochmals zur “reitenden Artilleriekaserne”. In R. Grosse, G. Lerchner, & M. Sehröder (Eds.), Beiträge zur Phraseologie, pp. 133–139. Bergmann, R. (1980). Verregnete Feriengefahr und Deutsche Sprachwissenschaft. Sprachwissenschaft, 5(3), 234–265. Burkhardt, A. (1999). Gut erhaltene Knochenfunde beim Urmenschen. Sprachreport, 15(2), 2–10. Maienborn, C. (2020). Wider die Klammerparadoxie. Zeitschrift für Sprachwissenschaft, 39(2), 149–200.

How much N₁ attachment makes a bracketing paradox❔

Results

N₁ attachment vs. AN₂ rating

Grammatically: the adjective is always the attribute of N₂ ❌

N₂ was chosen more often than N₁ overall → grammatical ✔️

N₁ was sometimes chosen over N₂ → bracketing paradox ✔️

~~When ? is chosen often → 🚫 interpretation?~~

The compatibility ratings from Experiment 1 align with the selected attachment site. ✔️

High adjective 🩷 N₂ semantic match reduced N₁ attachments.

Results

N₁ attachment vs. AN₁ × AN₂ rating

Grammatically: the adjective is always the attribute of N₂ ❌

N₂ was chosen more often than N₁ overall → grammatical ✔️

N₁ was sometimes chosen over N₂ → bracketing paradox ✔️

~~When ? is chosen often → 🚫 interpretation?~~

The compatibility ratings from Experiment 1 align with the selected attachment site. ✔️

High adjective 🩷 N₂ semantic match reduced N₁ attachments

High adjective 🩷 N₁ semantic match increased N₁ attachments

Adjective 🩷 N₁ match had a bigger effect when
adjective 🩷 N₂ match was poor.

Results

A+N₁ Morphosyntactic match

Results

	Est.	Est. error	95% CrI
Intercept	−1.95	0.75	−3.42, −0.50
AN₁ rating	4.00	0.85	2.32, 5.69
AN₂ rating	−1.92	0.82	−3.53, −0.29
morph. match	−0.24	0.65	−1.51, 1.04
AN₁ rating × AN₂ rating	−1.25	0.97	−3.17, 0.65
AN₁ rating × morph. match	0.31	0.56	−0.76, 1.43
AN₂ rating × morph. match	0.61	0.57	−0.51, 1.70

Discussion

Participants did not opt out of interpreting the phrases.

Most phrases were judged in accordance with grammar → N₂.

N₁ was a viable attachment site for a large group of phrases.

There were few unequivocal bracketing paradoxes.

The participants disagreed on the interpretation of a substantial group of phrases.

The results are in line with Experiment 1.

There was no evidence for an effect of morphosyntactic match between adjective and N₁.

Conclusion

Acceptability ratings and attachments seem to be determined largely by semantic factors.

Strong semantic cues can override grammatical constraints.

Compositional processing can be suspended to fulfill communicative goals.

Future work

Väter mit einem Kind und solche mit mehreren Kindern erhalten unterschiedliche staatliche Unterstützungen.
Fathers with one child and those with several children receive different levels of state support.

Männer mit drei Köpfen haben oft Schwierigkeiten, einen Job zu finden, um ihre Familie zu unterstützen.
Men with three heads often find it difficult to find a job to support their family.

Der dreiköpfige Familienvater bekommt deutlich weniger Geld.
The three headed family father receives significantly less money.

Eine Hochzeit kann ein prächtiges Ereignis oder eine bescheidene Feierlichkeit sein.
A wedding can be a magnificent event or a modest celebration.

Wenn zwei Adelige heiraten, dann wird die Hochzeit zu einem glamourösen Ereignis.
When two aristocrats marry, the wedding becomes a glamorous event.

Eine königliche Hochzeitsfeier ist grandios und atemberaubenden.
A royal wedding celebration is grandiose and breathtaking.

Severe Storm Warnings for Four-Story Homeowners

Why are some phrases ambiguous or preferentially interpreted as a bracketing paradox?

Why are some bracketing paradoxes odd and others unremarkable?

How do we model these phrases?

How are they processed?

Experiment 1

Predictions

Predictions

Analysis

Analysis

Analysis

Results

AN1 rating vs. AN2 rating

Results

AN1N2 rating vs. AN2 rating

Results

AN1N2 vs. AN1 × AN2 ratings

Results

A+N1 Morphosyntatic match

Results

Discussion

Experiment 2

Predictions

Analysis

Results

Attachment preference

Results

Attachment preference

Results

N1 attachment vs. AN2 rating

Results

N1 attachment vs. AN1 × AN2 rating

Results

A+N1 Morphosyntactic match

Results

Discussion

Conclusion

Acceptability ratings and attachments seem to be determined largely by semantic factors.

Strong semantic cues can override grammatical constraints.

Compositional processing can be suspended to fulfill communicative goals.

Future work

Thank you

AN₁ rating vs. AN₂ rating

AN₁N₂ rating vs. AN₂ rating

AN₁N₂ vs. AN₁ × AN₂ ratings

A+N₁ Morphosyntatic match

N₁ attachment vs. AN₂ rating

N₁ attachment vs. AN₁ × AN₂ rating

A+N₁ Morphosyntactic match