Bakay, Akkuş & Dillon (2026) #
@cite{bakay-etal-2026}
Hierarchical relations guide memory retrieval in sentence comprehension: Evidence from a local anaphor in Turkish. Journal of Memory and Language 148, 104747.
Summary #
Three visual-world experiments show that c-command relations between NPs within a single clause guide antecedent retrieval for the Turkish reciprocal birbirleri from the earliest moments of processing — above and beyond clause-mateness, case marking, subjecthood, and linear order/recency.
Design Innovation #
Prior studies confounded c-command with clause membership and/or case marking (Figure 1 Venn diagram: 45 experiments from 28 studies). Bakay et al. isolate c-command by placing target and distractor in the same clause with the same case marking, varying only whether the NP c-commands the anaphor.
Formalization #
We derive the retrieval predictions from three independently formalized components:
- Fragment (
Fragments.Turkish.Anaphors): birbirleri is a reciprocal requiring a plural antecedent - Theory (Principle A): reciprocals require a local c-commanding clause-mate antecedent
- Processing (
Processing.CueBasedRetrieval): retrieval cues from (1) + (2) feed a weighted activation model; structural cues (c-command, clause-mate) predict the target advantage
C-command is verified computationally on tree addresses using the
binary-branching cCommand from Phenomena.Anaphora.Compare.
Features relevant to antecedent retrieval for birbirleri. Divided into structural (relational) and item-level (intrinsic).
- cCommanding : Feature
Structural: c-commands the anaphor
- clauseMate : Feature
Structural: in the same local clause as the anaphor
- plural : Feature
Item-level: plural number
- singular : Feature
Item-level: singular number
- genCase : Feature
Item-level: genitive case -(n)In
- datCase : Feature
Item-level: dative case -(y)A
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
The retrieval cues that processing birbirleri generates are derived from two independent sources:
1. **Principle A** (syntactic theory): reciprocals must be bound by
a c-commanding clause-mate → structural cues
2. **Fragment property** (`requiresPluralAntecedent`): reciprocals
need a plural antecedent → item-level cue
birbirleri is a reciprocal (from the fragment)
Reciprocals require plural antecedents (from the fragment)
Retrieval cues generated when processing birbirleri.
- Structural cues from Principle A (c-command + clause-mate)
- Item-level cue from the fragment's plurality requirement
Equations
- One or more equations did not get rendered due to their size.
Instances For
Experiment 1: Subject vs. possessor #
Target = embedded subject (c-commanding, clause-mate, GEN, plural).
Distractor = possessor within the subject NP (clause-mate, GEN, plural,
but does NOT c-command the anaphor).
Both NPs are in the same clause, have the same case (GEN), and can be
plural — the **only** distinguishing feature is c-command.
Simplified tree for the embedded clause:
```
CP_emb
/ \
NP_subj VP_emb
/ \ / \
NP_poss N' anaph V
(dist) (head)
```
The target subject c-commands the anaphor.
The distractor possessor does NOT c-command the anaphor.
Target item: embedded subject (cameramen). Features: c-commanding, clause-mate, plural, genitive case.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Distractor item: possessor NP (director(s)), Match condition. Features: clause-mate, plural, genitive case — but NOT c-commanding. Same clause, same case, same number as target.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Distractor item: possessor NP (director), Mismatch condition. Singular distractor — does not match the reciprocal's number.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Experiment 1 prediction: target is retrieved over distractor in the Match condition — the hardest case, where item-level cues don't distinguish target from distractor. Holds for any positive structural weight.
Target also wins in the Mismatch condition. The distractor is singular, so it matches fewer cues on both the structural and item-level dimensions — a fortiori advantage for the target.
Experiment 2: IO vs. adjunct distractor #
Target = c-commanding indirect object (IO) with DAT case.
Distractor = non-c-commanding adjunct NP with DAT case.
The IO is an argument sister to V', so it c-commands the anaphor.
The distractor is inside a PP adjunct, so it does not.
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
```
The adjunct distractor does NOT c-command the anaphor.
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
Experiment 2 prediction: the IO is retrieved over the adjunct distractor, extending the structural advantage to non-subject c-commanding positions.
Key finding #
The structural advantage holds across all experiments:
- Exp 1: subjects over non-c-commanding possessors (GEN case)
- Exp 2: IOs over non-c-commanding adjuncts (DAT case)
- Exp 3: pre-registered replication combining Exp 1–2 conditions
In all cases, target and distractor share clause, case marking, and
(in the Match condition) number. The **only** distinguishing feature
is c-command — and the target is immediately retrieved.
This is captured by the structural advantage theorem: under *any*
retrieval model where structural cues carry positive weight, the
c-commanding item has higher activation.
The structural advantage is independent of the specific weight assignment: it holds for any positive structural weight, regardless of item-level and positional weights.
This formalizes the paper's claim that "hierarchical relational information guides antecedent retrieval above and beyond other sources of structural information and linear order."
Both the weighted activation model and the privileged-access model predict the target advantage: the target is privileged (matches all structural cues), while the distractor is not.
Under the privileged-access model, target and distractor have different accessibility status: the target is directly accessible, the distractor requires search.