Emotional Context Changes Segment Memory, Predictions Determine Its Strength

Technical Summary: Emotional Context Changes Segment Memory, Predictions Determine Its Strength

Problem and Motivation
The ability to divide continuous experience into discrete, meaningful units is known as event segmentation. While event segmentation theory posits that changes in stable contexts drive the formation of event boundaries, the specific interplay between emotional context changes and predictive processing remains underexplored. Previous research has established that emotional transitions (e.g., from neutral to negative) can induce temporal dilation in memory, suggesting a segmentation effect. However, it is unclear whether these effects are driven solely by the objective change in context or if they are modulated by the observer's moment-to-moment predictions. Specifically, does a prediction that aligns with an upcoming emotional transition weaken the segmentation effect, while a mismatched prediction amplify it? Furthermore, the role of individual differences, such as trait anxiety, in shaping these predictive trajectories and subsequent memory structures requires investigation.

Methodology
The study employed a modified event segmentation paradigm with 90 participants (mean age 22.13 years) who encoded sequences of neutral and emotional (high-arousal negative) images. The experimental design consisted of 12 blocks, each containing six events of variable lengths (5, 6, or 7 images). Crucially, the number and order of emotional events within blocks were pseudorandomized to create uncertainty regarding whether and when an emotional context change would occur.

During the encoding phase (Prediction Task), participants viewed images one at a time and provided moment-to-moment ratings on a 4-point scale indicating how likely they believed a negative image would appear soon. This allowed for the collection of dynamic prediction data. Following a distraction task, participants completed a Temporal Memory Task. In this task, they judged the temporal distance between image pairs (rated 1–4) and their temporal order (which appeared first). The image pairs were systematically selected to represent:

1. Within-event pairs: Images from the same event (neutral-neutral or emotional-emotional).
2. Across-event pairs: Images spanning a transition (neutral-neutral or neutral-to-emotional).

Affective judgments (valence and arousal) were collected post-encoding. Individual differences were measured using the State-Trait Anxiety Inventory-Trait Subscale (STAI-T) and the Revised Life Orientation Test (LOT-R). Statistical analyses utilized linear and generalized linear mixed-effects models (LMMs/GLMMs) in R, controlling for semantic similarity between image pairs using deep neural network features (VGG-16 and ResNet-50).

Key Results

1. Affective Validation: Emotional images were rated as significantly more unpleasant and arousing than neutral images. Trait anxiety did not moderate these subjective ratings.
2. Prediction Dynamics: Participants' predictions of negativity were sensitive to event type and position.
   • During neutral events, predictions of negativity increased linearly as the event progressed, reflecting growing uncertainty about an upcoming emotional transition. This trend was significantly steeper for participants with high trait anxiety.
   • During emotional events, predictions followed an inverted-U pattern: they initially increased and then decreased, suggesting participants anticipated a return to neutrality as the emotional event unfolded.
   • Prediction response times followed a U-shaped pattern across events (slower at onset and offset), indicative of event model updating and growing uncertainty.
3. Segmentation via Temporal Distance:
   • Context Changes: Neutral-to-emotional transitions produced significant temporal dilation (participants judged across-event pairs as farther apart in time than within-event pairs). In contrast, neutral-to-neutral transitions did not produce significant temporal dilation, indicating that stable context does not trigger segmentation in this paradigm.
   • Prediction Modulation: The magnitude of temporal dilation was modulated by predictions. When participants predicted high negativity prior to a neutral-to-emotional transition (i.e., the change was compatible with their prediction), the temporal dilation effect was reduced. Conversely, incompatible predictions amplified the segmentation effect.
   • Prediction Alone: Prediction errors without an accompanying emotional context change (e.g., predicting negativity during a neutral-to-neutral transition) did not induce temporal dilation.
4. Temporal Order Accuracy: Contrary to typical segmentation findings where boundaries impair order memory, neutral-to-emotional transitions resulted in higher temporal order accuracy compared to neutral-to-neutral transitions. This suggests that emotional salience may support order memory, potentially overriding boundary-driven impairment. Prediction compatibility did not significantly modulate order accuracy.
5. Individual Differences: While trait anxiety influenced the trajectory of predictions (e.g., higher anticipation of emotional transitions in neutral events for high-anxiety individuals), it did not significantly modulate the segmentation effects on temporal memory outcomes.

Significance and Claims
The authors claim that these findings reconcile context-based and prediction-based accounts of event segmentation. The study demonstrates that emotional context changes are the primary driver of memory segmentation, as evidenced by temporal dilation. However, predictions determine the strength (magnitude) of this segmentation. When a context change is anticipated, the segmentation effect is attenuated, likely because the upcoming context is actively represented and binds with the current one, reducing the perceived temporal distance. When the change is unexpected, the segmentation effect is amplified.

Crucially, the paper argues that predictions alone are insufficient to generate event boundaries; a prediction error must be accompanied by an actual change in the external context to trigger segmentation. The dissociation between temporal distance (segmentation) and temporal order accuracy (enhanced by emotional salience) suggests that these measures capture distinct aspects of episodic memory organization. The study concludes that the structure of episodic memory is shaped by the dynamic interplay between the contextual shifts we encounter and the predictions we carry into them.