Probabilistic modeling of the semantic fluency task with extended Markov networks

Area of Science:

• Computational linguistics
• Cognitive psychology
• Psycholinguistics

Background:

• Semantic verbal fluency tasks (SFT) reveal mental lexicon structure through word sequences based on semantic associations.
• Existing models struggle to capture both associative retrieval and sudden resets (switching) in word generation.

Purpose of the Study:

• To propose a novel probabilistic framework for modeling SFT.
• To integrate associative retrieval and reset mechanisms within a unified computational model.
• To develop metrics for evaluating model fit and distributional properties of word associations.

Main Methods:

• Modeled SFT as censored random walks on semantic networks with pseudo-nodes for jumps.
• Introduced the BIGRAM-CN model combining statistical constraints and empirical word transition frequencies.
• Defined global likelihood, frequency likelihood, and bigram relevance metrics for model assessment.

Main Results:

• The BIGRAM-CN model demonstrated superior performance in synthesizing realistic word lists compared to existing methods.
• The model effectively avoids overfitting and generalizes well across training and test datasets.
• BIGRAM-CN accurately captures clustering and switching processes observed in human SFT.

Conclusions:

• The proposed probabilistic framework offers an advanced computational model for lexical retrieval.
• BIGRAM-CN provides a robust tool for analyzing semantic verbal fluency data.
• This research facilitates comparisons of populations, categories, and cognitive profiles in linguistic and psychological studies.