Bakay, Akkuş & Dillon (2026): hierarchical relations guide memory retrieval #
[BAD26] (JML 148, 104747): three visual-world experiments showing that
c-command between NPs within a single clause guides antecedent retrieval for the
Turkish reciprocal birbirleri, deconfounded from clause-mateness, case marking,
subjecthood, and linear order/recency — cues the paper argues are "plausibly
construed as item-level features". This file derives the retrieval predictions
from the Turkish.Anaphors fragment (its plurality requirement yields the number
cue), Principle A (the c-command and clause-mate cues), and
BarkerPullum1990.cCommand on tree addresses; each experiment's target advantage
is an instance of dominance. The retrieval model is the ordinal core of
[LV05] spreading activation — a weighted count of cue matches;
base-level decay, fan, mismatch penalty, and noise are not modeled.
Main definitions #
CueSource,Cue— cues tagged relational / item-level / positional, the paper's deconfound (clause-mateness is item-level, e.g. a clause index).birbirleriCues— the cue bundle birbirleri generates, from Principle A plus the fragment's plurality requirement.weightedActivation— activation as a weighted cue-match count.privileged— the rival representational account: direct access by structural position ([McE06]; [Obe02]; gated retrieval, [DMSP13]), defined on the tree configuration, not on cue matches.
Main results #
dominance— pointwise dominance of cue-match vectors gives strictly higher activation under every positive weighting.exp1_target_retrieved,exp1_target_retrieved_mismatch,exp2_io_retrieved— the target advantages (Exp 1: subject vs. possessor, both GEN; Exp 2: indirect object vs. adjunct NP, both DAT), for any positive relational weight.exp1_target_privileged,exp1_distractor_not_privileged— the same contrast under privileged access.
Implementation notes #
The paper's General Discussion opposes differentially weighted cues to a
privileged representation; both predict the target advantage and diverge only on
early interference from feature-matching distractors, which needs the unmodeled
fan/mismatch machinery — the paper finds limited, inconsistent number
interference and leaves the distinction open. The relational cue is realized as
a dynamically assigned item feature approximating c-command ([Kus13], in the
paper's summary). Not yet formalized: the c-command vs. coargumenthood
alternative the paper leaves open ([PS94]) — Binding.SimpleClause
cannot represent the possessor/IO configurations used here — and the monotone
activation-to-looks linking to Processing.VisualWorld observables.
Cue-based retrieval: ordinal core #
Source of a retrieval cue, following the paper's deconfound: relational
information holds between the retrieval site and the candidate (c-command);
itemLevel features are stored with the candidate (number, case, clause
index); positional cues track linear order/recency.
Instances For
Equations
- BakayEtAl2026.instDecidableEqCueSource x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- One or more equations did not get rendered due to their size.
Equations
- BakayEtAl2026.instReprCueSource = { reprPrec := BakayEtAl2026.instReprCueSource.repr }
Equations
- One or more equations did not get rendered due to their size.
Instances For
A retrieval cue: a required feature tagged with its source.
- source : CueSource
- feature : F
Instances For
Equations
- BakayEtAl2026.instReprCue = { reprPrec := BakayEtAl2026.instReprCue.repr }
Equations
- One or more equations did not get rendered due to their size.
Instances For
Number of cues from source s that a memory item's feature bundle matches.
Equations
- BakayEtAl2026.matchCount feats cues s = List.countP (fun (c : BakayEtAl2026.Cue F) => decide (c.source = s ∧ c.feature ∈ feats)) cues
Instances For
Activation of an item as a weighted count of cue matches.
Equations
- BakayEtAl2026.weightedActivation w feats cues = ∑ s : BakayEtAl2026.CueSource, w s * BakayEtAl2026.matchCount feats cues s
Instances For
If a's cue-match vector pointwise dominates b's, strictly at some source
carrying positive weight, then a out-activates b under every such
weighting — Core.Optimization.sum_mul_lt_sum_mul on cue-match vectors.
Retrieval cues for birbirleri #
Equations
- BakayEtAl2026.instDecidableEqFeature x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- BakayEtAl2026.instReprFeature = { reprPrec := BakayEtAl2026.instReprFeature.repr }
Equations
- One or more equations did not get rendered due to their size.
Instances For
Item-level number cue, generated exactly when the fragment's anaphor type imposes a plurality requirement on its antecedent.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Retrieval cues generated on encountering birbirleri: Principle A supplies the relational c-command cue and the clause-mate cue; the fragment's plurality requirement supplies the number cue.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Experiment 1: subject targets vs. possessor distractors #
Target = embedded subject (c-commanding clause-mate, GEN, plural). Distractor = possessor inside the subject NP (clause-mate, GEN, plural or singular, not c-commanding). Same clause, same case, and — in the Match condition — same number: only c-command distinguishes them.
CP_emb
/ \
NP_subj VP_emb
/ \ / \
NP_poss N' anaph V
Equations
Instances For
Instances For
Instances For
The embedded subject c-commands the anaphor.
The possessor does not c-command the anaphor.
Target subject (kameramanlar 'cameramen').
Equations
Instances For
Possessor distractor, Match condition (plural yönetmenler 'directors').
Equations
Instances For
Possessor distractor, Mismatch condition (singular yönetmen 'director').
Equations
Instances For
In the Match condition only the relational cue distinguishes target from distractor: item-level (and positional) match counts tie.
The target out-activates the Match distractor — the hardest case, where item-level cues do not distinguish them — for any weighting with positive relational weight.
In the Mismatch condition the distractor also loses the number cue, so the target advantage holds a fortiori.
Experiment 2: indirect-object targets vs. adjunct distractors #
Target = c-commanding indirect object (DAT). Distractor = NP inside a postpositional adjunct (DAT, e.g. göre 'according to'), not c-commanding. Extends the advantage to non-subject c-commanders, ruling out a composite subject-of-the-current-clause item-level cue.
IO condition: Distractor condition:
CP_emb CP_emb
/ \ / \
NP_subj VP NP_subj VP
/ \ / \
NP_IO V' PP_adj V'
/ \ / \ / \
anaph V NP_dist P anaph V
Instances For
The indirect object c-commands the anaphor.
The adjunct-internal distractor does not c-command the anaphor.
Indirect-object target: c-commanding clause-mate, plural, DAT.
Equations
Instances For
Adjunct-internal distractor: clause-mate, plural, DAT, not c-commanding.
Equations
Instances For
The indirect object out-activates the adjunct distractor for any weighting with positive relational weight. Experiment 3 is the paper's pre-registered, high-powered replication of the Experiment 1–2 contrasts; it introduces no new configuration.
Privileged representation #
The representational account grants c-commanding items a temporary association
with a privileged store — access by structural position, not cue matching — so
privilege is defined on the tree configuration, not on matchCount.
An NP position is privileged at a retrieval site iff it c-commands it: the region of direct access holds the current c-commanders.
Equations
- BakayEtAl2026.privileged np anaphor = (BarkerPullum1990.cCommand np anaphor = true)
Instances For
The Experiment 1 target is in the region of direct access.
The Experiment 1 distractor is not, whatever its feature match.