How lexical knowledge becomes decision evidence: Distributional representations and frequency modulate drift in lexical decision

How does lexical knowledge enter the decision process? We develop a framework that links learned lexical representations directly to the evidence accumulation process, allowing explicit tests of how lexical information shapes evidence strength (drift) in lexical decision. Words and nonwords are represented with word embeddings (FastText or BERT) and passed through a compact classifier; its outputs are mapped, via psychologically interpretable functions, to evidence strength, with an additional term capturing the independent influence of lexical frequency. Hierarchical Bayesian models were fitted to the English Lexicon Project and compared using standard information criteria; posterior predictive checks and a parameter-recovery study assessed adequacy and identifiability. Across eight integrated models, diffusion-based variants (one diffusion process per response) outperformed otherwise identical ballistic race variants, and explicitly modeling frequency reliably improved fit. The best integrated model (FastText + diffusion + frequency) made more accurate held-out predictions than the benchmark models and matched--sometimes slightly surpassed--their ability to reproduce key features of the RT distribution. These results suggest that distributional lexical information, augmented by frequency, is best captured as modulation of evidence quality within standard accumulation dynamics. The approach preserves interpretability, reproduces canonical RT/accuracy patterns, and can be extended to other recognition decisions by swapping encoders or the accumulation module without changing the estimation workflow.