Activation–Integration Model in Reading

• Activation–Integration Model is a mathematical framework that couples memory-retrieval dynamics with eye-movement control in reading.
• It integrates LV05 memory processing with SWIFT’s saccade targeting to quantitatively predict fixation patterns and memory-driven regressions.
• The approach uses Bayesian inference and MCMC methods to validate parameter roles and capture similarity-based interference effects in sentence comprehension.

The Activation–Integration Model is a mathematical framework for understanding the coupling of memory-based sentence processing and dynamical models of eye-movement control during reading. Conceived as a core component of the Sentence-Processing and Eye-Movement Activation-Coupled Model (SEAM), it formally integrates the memory retrieval dynamics from the Lewis & Vasishth (2005; "LV05") cue-based parsing model with the SWIFT model of eye-movement control, thereby enabling the quantitative prediction of both standard fixation patterns and memory-driven regressions that arise from similarity-based interference in reading (Rabe et al., 2023).

1. System Architecture and Subsystem Coupling

The Activation–Integration Model is instantiated in SEAM as two tightly coupled subsystems:

• Post-lexical working-memory module (LV05-based): Upon lexical identification, each word is encoded as an ACT-R “chunk” with a time-varying activation. On certain trigger words (typically verbs), a cue-based retrieval is initiated in which previously stored chunks compete. This subsystem formalizes similarity-based interference (fan effects) through the competitive dynamics among competing chunks.
• SWIFT eye-movement module: This component maintains a spatially organized field of discrete "word activation" counters $n_j(t)$ for all visible words, supporting parallel lexical and post-lexical processing via stochastic transition rates $w_j(t)$. Saccade programming is governed by a three-stage process (labile $\rightarrow$ non-labile $\rightarrow$ execution) and is largely decoupled from the processing timers, with the next saccade target stochastically chosen using a Luce-choice rule over normalized activations.

Coupling Mechanism ("Activation–Integration"): Each lexical processing completion in SWIFT triggers word encoding in the LV05 memory module. On retrieval triggers, LV05 chunk activations $A'_k(t)$ directly modulate SWIFT’s post-lexical transition rates $w'_k(t)$, causing reactivation of partially processed words proportional to their cue-match, and increasing the probability of regression saccades to previously read words (Rabe et al., 2023).

2. Mathematical Formalism

The model combines LV05 memory dynamics and SWIFT saccade targeting through explicit equations:

• Memory Activation (LV05):
  • $A_k(t) = S_k(t) + P_k(t) + B_k(t)$,
  • where $S_k$ encodes fan-effect (interference), $P_k$ penalizes feature mismatches, $B_k$ captures base-level decay.
  • $A'_k(t)\sim \mathrm{Logistic}(A_k(t), ans)$ introduces activation noise.
  • Predicted retrieval latency: $t^{retr}_k(t) = F \exp[-A'_k(t)]$; the chunk with maximal $A'_k(t)$ is retrieved.
• SWIFT Saccade Targeting:
  • Normalized activation: $a_j(t) = n_j(t)/N_a$.
  • Selection probability: $$\pi_j(t) = a_j(t)^\gamma / \sum_\ell a_\ell(t)^\gamma$$.
  • Baseline transition rates:
  • $w_j(t) = \left{ \begin{array}{ll} \alpha \Lambda_j(t) &amp; \text{lexical}\</li> <li>\max[\alpha \Lambda_j(t) proc, \omega] &amp; \text{post-lexical}\</li> <li>0 &amp; \text{after completion} \end{array}\right.$
• Coupling (Activation–Integration):
  • During post-lexical retrieval, $w_j(t)$ is replaced by $w'_j(t)$:
  • For the retrieval trigger ($j = m$), decay halts if $n_m(t) < \mu_2 N_m$.
  • For a retrieval candidate ($j \neq m$): $w'_j(t) = \mu_3 N_m [F \exp(-A'_j(t))]^{-1}$, ensuring the expected time to threshold matches LV05 retrieval latency.
  • After retrieval: standard post-lexical decline resumes.

The effect of this coupling is that, during retrieval, earlier words with high memory activation see their SWIFT activation counters $n_j(t)$ re-incremented, increasing $\pi_j(t)$ and the likelihood of regressive saccades, thus capturing memory-driven eye movement phenomena.

3. Parameter Roles and Cognitive Correlates

Key parameters and their cognitive interpretations include:

Parameter	Description	Cognitive Role
$S_{max}$	Max associative boost in $S_k$	Memory strength ceiling before interference
$\phi_{kl}(t)$	Count of competing chunks on cue $l$	Formalizes similarity-based interference
$p$	Penalty for feature mismatch	Models semantic/syntactic mismatches
$d$	Decay rate of base-level activation	Time-based memory decay
$ans$	Activation noise parameter	Stochastic retrieval variability
$F$	Latency scaling factor	Maps activation to retrieval time
$\alpha, \Lambda$	Baseline processing rate and modulating factors	Lexical/post-lexical difficulty, attention, etc.
$\mu_2, \mu_3$	Trigger/candidate thresholds	Gating post-lexical persistence/reactivation
$N_a, \gamma$	SWIFT normalizer and choice exponent	Saccade target scaling and stochasticity

This explicit mapping enables parameter estimation and cognitive interpretation in empirical datasets (Rabe et al., 2023).

4. Core Algorithmic Dynamics and Saccade Programming

Processing unfolds in a parallel, stochastic loop. In SWIFT, all visible words increment or decrement $n_j(t)$ based on $w_j(t)$ or $w'_j(t)$, with a global timer initiating labile saccade preparation. During the labile stage, a new saccade target can be selected according to $\pi_j(t)$. The Activation–Integration Scheme ensures:

• The activation of the retrieval-triggering word $m$ cannot fall below $\mu_2 N_m$ during retrieval, prolonging its status as potential saccade target.
• Retrieval candidates $j$ have their $n_j(t)$ counters reactivated towards $\mu_3 N_m$, reflecting memory-guided attentional shifts.
• This modulation of $n_j(t)$ directly increases the regression probability to earlier words during critical linguistic events, consistent with similarity-based interference signatures.

5. Statistical Inference and Model Evaluation

Bayesian inference for the Activation–Integration Model is accomplished via Markov Chain Monte Carlo (MCMC) using a DREAM$_{zs}$ sampler with multiple chains and differential evolution for propositional moves in log-parameter space. Data likelihood is factored into spatial (target choice via $\pi_j(t)$) and temporal (fixation duration) components. Model identifiability is established through profile log-likelihood analysis: for simulated datasets with fixed parameter vectors $\theta^*$, one parameter $\theta_k$ is systematically varied to verify a peaked profile log-likelihood at the generating value. Parameter recovery is demonstrated by simulating data from $\theta^*$ values, refitting, and checking that posterior CrIs contain the true value with high probability (Rabe et al., 2023).

6. Empirical Performance and Comparative Results

Empirical validation, using stimuli that manipulate animacy and subjecthood in distractors (e.g., "It turned out that the attorney whose secretary ... frequently complained about the salary"), reveals that SEAM’s Activation–Integration Model can account for nuanced effects in first-pass regression probability (FPR) and regression-path duration (RPD) in regions both preceding and at the critical verb:

• SWIFT alone (no memory component) fails to produce condition effects in FPR/RPD.
• SEAM, leveraging memory-saccade coupling, reproduces additive animacy × subjecthood interactions in both measures (see Figures 8–10 in (Rabe et al., 2023)).
• Analysis of word-activation fields demonstrates that, although distractor activation increases during retrieval, the probability of excessive regressions is constrained, avoiding over-prediction.
• Cross-validation shows that SEAM produces substantially lower absolute prediction errors compared to SWIFT for both FPR and RPD, even with parameters fitted on only 70% of available data.

7. Scope, Implications, and Future Directions

The Activation–Integration Model instantiated in SEAM constitutes the first fully implemented integration of comprehensive eye-movement control and sentence-processing models. It allows quantitative, process-level predictions of the interaction between lexical/post-lexical memory retrieval costs and moment-by-moment attentional shifts during reading. This unified approach supports the investigation of how similarity-based interference and working-memory constraints directly manifest in the production of regressive saccades, addressing limitations of previous models that treated these subsystems independently.

A plausible implication is that such tight memory-saccade coupling is required to capture the full spectrum of observed human reading behavior, including nuanced condition effects observable in sentence comprehension paradigms. Going forward, the model's generalizability will need to be assessed across diverse reading benchmarks and its parameters further constrained by large-scale empirical datasets (Rabe et al., 2023).