Probabilistic modeling of the semantic fluency task with extended Markov networks

Semantic verbal fluency tasks (SFT) provide a window into the structure and dynamics of the mental lexicon by eliciting word sequences guided primarily by semantic associations. We propose a probabilistic framework that models SFT as censored random walks on semantic networks, extended with pseudo-nodes to account for local and global jumps. This representation enables the integration of associative retrieval and sudden resets, capturing both clustering and switching processes. To assess model quality, we define a suite of complementary metrics-global likelihood, frequency likelihood, and bigram relevance-that evaluate not only overall fit but also the distributional properties of word associations. Using a dataset of 677 lists in the category "clothing," we benchmark existing techniques against our proposed BIGRAM-CN model, which combines statistical constraints with empirical frequencies of word-to-word transitions. Results show that BIGRAM-CN avoids overfitting, generalizes across training and test data, and synthesizes realistic lists more accurately than prior approaches. This work advances computational models of lexical retrieval and offers practical tools for comparing populations, categories, and cognitive profiles in both linguistic and psychological research.