Bridge: Backgrounded Islands → Island Classification #
@cite{lu-pan-degen-2025}
Connects the formal backgroundedness model (Theories/Semantics/Focus) to the
shared island infrastructure in Phenomena/FillerGap/Islands/Data.lean and
to the lexical infrastructure in Core/Lexical/LevinClass.lean.
Layer connections #
Core/Lexical/LevinClass → mannerSpec = true for MoS verbs (§37.3)
↓
Theories/Focus/BackgroundedIslands → mannerSpec ↔ hasMannerWeight → island
↓
Phenomena/FillerGap/Islands/Data → ConstraintType.mannerOfSpeaking = discourse/weak
The MoS island is classified as weak (ameliorable) and discourse-sourced, and we derive both properties from the formal model. The derivation chain runs from Levin's meaning components through QUD-determined backgroundedness to extraction predictions, with no stipulation.
§1. Island Source Classification #
The paper's core contribution is the double dissociation between discourse-
sourced MoS islands and syntactically-sourced traditional islands. The MoS
source is imported from MannerOfSpeaking.mosIslandSources (derived from
the experimental evidence there). The traditional island classification
is the baseline consensus view: these islands arise from structural
constraints on movement (subjacency, PIC, Relativized Minimality).
Traditional islands (wh, CNPC, adjunct, coordinate, subject, sentential subject) are syntactically sourced. This is the baseline consensus against which the paper shows MoS islands are categorically different.
Note: @cite{hofmeister-sag-2010} argue that some of these (CNPC, wh-islands) have processing sources. That alternative classification is formalized in their study file, not here.
Equations
Instances For
§2. Levin Class → Manner Weight Bridge #
@cite{levin-1993} §37 classifies communication verbs into three subclasses:
- §37.7 say verbs (say, report, announce):
mannerSpec = false - §37.2 tell verbs (tell, inform, notify):
mannerSpec = false - §37.3 manner-of-speaking verbs (whisper, shout, mumble):
mannerSpec = true
The mannerSpec meaning component is exactly the property that drives the
MoS island effect: it indicates whether the verb's root specifies manner,
which determines whether manner alternatives are activated, which determines
QUD selection, which determines complement backgroundedness.
This section connects the Levin class infrastructure to the backgroundedness model, making the island prediction derivable from lexical classification by construction.
Map Levin class manner specification to BackgroundedIslands manner weight.
A verb with mannerSpec = true has lexical manner weight; one without has none.
(Compositional manner weight from adverbs is not captured by Levin classes.)
Equations
Instances For
MoS verbs (§37.3) have manner weight by Levin classification.
Bridge verbs (§37.7) lack manner weight by Levin classification.
Full derivation from Levin class to island prediction: the Levin
mannerSpec feature determines manner weight, which determines the default
QUD, which determines complement backgroundedness, which determines
extraction acceptability.
This makes the MoS island prediction a consequence of lexical classification, not an independent stipulation.
The Levin distinction between §37.3 and §37.7 is exclusively about manner. All other meaning components are identical (both are communication verbs with no change-of-state, contact, motion, causation, or instrument specification). Manner specification is the ONLY lexical feature that distinguishes them.
§3. Cross-Theory Predictions #
Different island theories make different predictions about which manipulations should affect which island types. The backgroundedness account uniquely predicts that discourse manipulations (prosodic focus, manner adverb addition) affect MoS islands but not structural islands.
A manipulation and the theories' predictions about its effect.
- manipulation : String
- affectsStructuralIslands : Bool
- affectsMoSIslands : Bool
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Predictions of the backgroundedness account vs. structural accounts.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Discourse and structural island types respond to DIFFERENT manipulations. This is the core empirical prediction that distinguishes the two account types.
§4. D-Linking Prediction #
The backgroundedness account predicts that D-linking (which-N vs bare wh) should NOT ameliorate MoS islands, because D-linking changes filler complexity (processing-relevant) but does not change the QUD or information structure.
This contrasts with structural weak islands (wh-islands), where D-linking DOES ameliorate. The dissociation is a testable prediction that distinguishes discourse-sourced from syntax/processing-sourced islands.
D-linking does not change QUD: D-linking modifies the filler's referential properties but does not affect which dimension of the communication event is foregrounded. The manner QUD remains active regardless of filler complexity.
Differential amelioration prediction: D-linking ameliorates structural weak islands but NOT MoS islands, while prosodic focus ameliorates MoS islands but NOT structural islands. This double dissociation is the core prediction separating discourse from syntax/processing accounts.
§5. Per-Verb Backgroundedness–Acceptability Correlation #
@cite{lu-pan-degen-2025} Experiment 2b (Figure 13) shows a negative correlation between per-verb backgroundedness proportion and extraction acceptability across the 13 verbs (12 MoS + say; β = −0.44, p = 0.014; MoS-only: β = −0.38, p = 0.076, marginally significant).
The formal model predicts this: verbs whose manner component is more salient activate the manner QUD more strongly, producing stronger default backgroundedness and therefore worse extraction.
Per-verb backgroundedness predicts acceptability: verbs that background
their complements more strongly also show more degraded extraction.
The model derives this from manner salience → QUD strength → backgroundedness.
The conceptually-right substrate for "backgroundedness" is
Core.Discourse.AtIssuenessDegree, not BinaryGivenness (which
orders by salience, given > new); future work could rephrase
complementStatus over AtIssuenessDegree directly.
§6. Fragment Verb → Island Prediction Pipeline #
Each MoS verb in Fragments/English/Predicates/Verbal.lean has
levinClass := some .mannerOfSpeaking, and each bridge verb has a non-MoS
Levin class. Per-verb verification theorems connect Fragment entries to island
predictions: changing a Fragment entry's levinClass field breaks exactly one
theorem, making the dependency explicit and auditable.
The derivation chain per verb:
Fragment entry → .levinClass = some .mannerOfSpeaking
→ levinClassToMannerWeight = true
→ hasMannerWeight = true
→ defaultDimension = .manner
→ complementStatus = .given
→ extraction degraded
Does a Fragment verb entry predict an island effect?
Derived from the verb's Levin class via levinClassToMannerWeight.
Equations
- Phenomena.FillerGap.Studies.LuPanDegen2025.fragmentPredictsIsland v = match v.levinClass with | some lc => Phenomena.FillerGap.Studies.LuPanDegen2025.levinClassToMannerWeight lc | none => false
Instances For
MoS verbs: all predict islands #
These 15 verbs have levinClass := some .mannerOfSpeaking in the Fragment.
The per-verb theorems cover both the 12 experimental stimuli from
@cite{lu-pan-degen-2025} (whisper, murmur, shout, scream, mumble, mutter,
shriek, yell, groan — 9 of 12 overlap with Fragment inventory) and 6
additional MoS verbs in the Fragment (cry, grumble, hiss, sigh, whimper, snap).
Three experimental verbs (stammer, whine, moan) are not yet in the Fragment.
Bridge verbs: no island prediction #
say and tell are bridge verbs (Levin §37.7 and §37.2 respectively). They lack manner specification and therefore do not background their complements by default.
Gradient predictions for Fragment verbs #
Using the gradient at-issueness model (§15 of BackgroundedIslands), Fragment MoS verbs have strictly lower complement at-issueness than bridge verbs. This connects Fragment entries → Levin class → manner weight source → gradient at-issueness in a single derivation chain.
Fragment MoS verbs map to lexical manner weight source, yielding the lowest complement at-issueness (maximally backgrounded). Bridge verbs map to none, yielding the highest (fully at-issue).
§7. Experimental Data → Formal Model Connection #
The experimental data in Islands/MannerOfSpeaking.lean records per-experiment
acceptability and backgroundedness values. Here we connect these empirical
observations to the formal model's predictions, closing the loop between
raw data and theoretical derivation.
The key connection: the formal model predicts that backgroundedness causes
extraction degradation (complementStatus .given → .rank = 0).
The experimental data confirms this directionally: higher backgroundedness
proportions consistently co-occur with lower acceptability ratings.
Experimental data matches formal model direction: the formal model predicts that backgrounded complements have degraded extraction. Experiments 1, 2b, and 3b all show the predicted anti-correlation: higher backgroundedness → lower acceptability.
Say+adverb replicates formal model prediction: adding manner weight compositionally (say + adverb) degrades extraction without changing syntax. This is exactly what the formal model predicts: manner weight → backgroundedness → island, regardless of whether the weight is lexical or compositional.
§8. Cross-Theory Comparison Across Manipulations #
This section integrates @cite{lu-pan-degen-2025}'s findings with @cite{hofmeister-sag-2010}'s processing manipulations and @cite{sag-2010}'s grammar-based island typology, comparing how three account types (competence, processing, discourse) score against the empirical data.
The key empirical claim of @cite{lu-pan-degen-2025}: discourse and processing accounts cover disjoint sets of manipulations. Together they explain the full range; neither suffices alone.
A nonstructural manipulation that changes island acceptability without altering the island configuration. Each account makes a prediction about whether the manipulation affects acceptability.
- description : String
- competencePredictsDifference : Bool
Does any competence theory predict an acceptability difference?
- processingPredictsDifference : Bool
Does the processing account predict a difference?
- discoursePredictsDifference : Bool
Does the discourse/backgroundedness account predict a difference?
- differenceObserved : Bool
Is a difference actually observed?
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
@cite{hofmeister-sag-2010} manipulations #
Filler complexity in CNPC (which-N vs bare wh — same island structure).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Filler complexity in wh-islands (which-N vs bare wh — same island structure).
Equations
- One or more equations did not get rendered due to their size.
Instances For
NP type in CNPC (definite vs indefinite — same CNPC configuration).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Filler complexity in adjunct islands (complex vs simple temporal adjunct).
Equations
- One or more equations did not get rendered due to their size.
Instances For
@cite{lu-pan-degen-2025} MoS manipulations #
Prosodic focus on embedded object in MoS islands. Focus changes information structure without changing syntax or processing load.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Say + manner adverb creates an island. Adding an adverb doesn't change CP structure but adds manner weight.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Verb-frame frequency in MoS islands: not significant in any experiment.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Processing correctly predicts the observed (non-)difference.
Equations
Instances For
Competence correctly predicts the observed (non-)difference.
Equations
Instances For
Discourse correctly predicts the observed (non-)difference.
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Processing scores 4/7: correct on all four H&S manipulations, incorrect on the three MoS manipulations (predicts effect or null incorrectly).
Competence scores 1/7 — only the frequency null result, where it correctly predicts no effect for the wrong reason.
Discourse scores 3/7: correct on prosodic focus, say+adverb, and the frequency null. Misses the four H&S effects, which are processing, not discourse.
Processing and discourse are perfectly complementary: for every manipulation, exactly one of the two accounts is correct (XOR). They have full coverage (together 7/7) with zero overlap.
§9. Connection to @cite{sag-2010}'s Construction-Based Islands #
@cite{sag-2010}'s F-G typology classifies which constructions are
grammar-based islands (those with [GAP ⟨⟩] on the mother).
@cite{hofmeister-sag-2010}'s findings explain within-island gradient
effects. @cite{lu-pan-degen-2025}'s MoS islands are a third mechanism.
Together the three accounts cover disjoint islands.
@cite{sag-2010}'s two island constructions are a proper subset of all F-G types. The non-island types (interrogative, relative, the-clause) freely permit extraction.
@cite{sag-2010}'s grammar-based islands (topicalization, exclamatives) are disjoint from @cite{hofmeister-sag-2010}'s processing-based islands (CNPC, wh-islands, adjuncts) and from @cite{lu-pan-degen-2025}'s discourse-based islands (MoS). The three accounts cover different cases under different mechanisms.
MoS islands are discourse-sourced and so distinct from the syntactic baseline assumed for traditional islands.