Semantic Fluency: Models and Mechanisms

• Semantic fluency tasks are experimental paradigms requiring rapid generation of category exemplars to elucidate the structure of human semantic memory.
• They reveal patch-switch dynamics with measurable inter-item retrieval times, supporting both dual-cue foraging and random-walk models.
• Recent studies integrate quantitative computational models and LLMs, highlighting the balance between deterministic search and stochastic fluctuations in retrieval.

Semantic fluency tasks are experimental paradigms probing the organization and retrieval mechanisms of human semantic memory by requiring participants to rapidly generate as many exemplars of a given semantic category (e.g., “animals”) as possible within a fixed time window. Analysis of response sequences—especially their temporal structure and the order of produced items—provides insight into both the structure of the underlying semantic space and the cognitive processes mediating search and selection. Contemporary work formalizes semantic fluency as a sequence generation problem embedded in a networked semantic space, governed by a combination of local associative structure and global category-level regularities, with recent emphasis on linking human data to quantitative computational models, including LLMs and stochastic network search (Heineman et al., 2024, Nematzadeh et al., 2016, 0903.4132, Miscevic et al., 2017).

1. Theoretical Foundations: Competing Accounts of Semantic Retrieval

Two canonical computational accounts structure current debates on semantic fluency. The first is the optimal foraging or cue-switching model of Hills et al. (2012), which conceives retrieval as a dual-cue process: a local cue exploits the current semantic “patch” by promoting transitions to proximate exemplars, while a global cue monitors overall retrieval utility, triggering patch switches following the Marginal Value Theorem (MVT) of foraging theory. Within-patch exploitation and between-patch exploration are formalized as separate cognitive mechanisms.

The second, as advanced by Abbott et al. (2015), posits that a single-process random walk over a densely clustered semantic network suffices. Here, patch transitions (i.e., cluster switches) are emergent properties of network topology rather than explicit strategic control—semantic transitions are uniformly sampled among nearest neighbors, and switches occur when the search path traverses sparse inter-patch connectors.

Both models successfully recapitulate the empirical signature of patch switches: an inter-item retrieval time (IRT) spike immediately prior to cluster exit and a subsequent drop at patch entry, consistent with the MVT profile (Heineman et al., 2024, Miscevic et al., 2017). However, these apparently distinct mechanistic stories can generate indistinguishable event-level predictions in classic analysis frameworks (e.g., IRT distributions, patch switch counts).

2. Formal Models: Mathematical Structure and Transition Probability Definitions

In the dual-cue (Hills) account, the transition probability from exemplar $x_n$ to $x_{n+1}$ is defined as

$$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$

where $P_{\mathrm{local}}$ is operationalized as the cosine similarity between GloVe embeddings for local associative transitions, and $P_{\mathrm{global}}$ is the empirical frequency (from corpora or human data) reflecting global typicality.

Abbott's random-walk model is mathematically constrained by $\beta_g=0$ and a uniform neighbor transition:

$$P(x_{n+1} \mid x_n) = \frac{\mathbb{1}[x_n \to x_{n+1}]}{\sum_{y \in A(n)} \mathbb{1}[x_n \to y]}$$

where $A(n)$ is the set of adjacent nodes to $x_n$.

Heineman et al. (Heineman et al., 2024) introduce a third cue reflecting empirical subcategory transition probabilities $P(c_{n+1} \mid c_{n})$, motivated by the need to capture nuanced human patch-switching biases that neither local nor global cues explain alone:

$x_{n+1}$0

This formalization supports direct adjustment of model sensitivity to both micro-level associative and macro-level categorical structure.

3. Semantic Networks: Representation Construction and Search Algorithms

Semantic networks underlying fluency tasks are constructed from distributional representations of words, often drawing on cross-situational learning from naturalistic data (Nematzadeh et al., 2016, Miscevic et al., 2017). Nodes represent category exemplars (e.g., animal names), while edges encode pairwise cosine similarity in feature space. For robust modeling, a lower similarity threshold for the category cue (e.g., “animal”) ensures global connectivity.

Random-walk algorithms on these networks provide a plausible mechanism for sequence generation. Transition matrices can be unweighted (uniform over neighbors) or weighted (normalized cosine similarities). Empirical studies demonstrate that with a sufficiently rich, small-world network structure—with high clustering and short characteristic path lengths—a simple random walk reproduces the human IRT profiles and patch dynamics observed in semantic fluency (Nematzadeh et al., 2016, Miscevic et al., 2017). The addition of global jump mechanisms—so-called switcher-random walk (SRW) models—allows explicit manipulation of the exploration–exploitation balance (0903.4132), with the overall transition kernel:

$x_{n+1}$1

where $x_{n+1}$2 is the local walk, $x_{n+1}$3 enables global jumps, and $x_{n+1}$4 controls the switching propensity.

4. Empirical Evaluation and Sequence-Level Metrics

Traditional evaluation centers on patch-switch prediction and IRT analysis, but these approaches underdetermine mechanistic distinctions: both foraging and random-walk models can produce similar spike patterns before patch switches (Heineman et al., 2024).

Full-sequence evaluation requires more granular metrics. Heineman et al. propose a BLEU-style $x_{n+1}$5-gram overlap metric adapted for semantic fluency:

• Exemplar BLEU: $x_{n+1}$6-gram overlap computed on animal names between generated and human reference sequences.
• Category BLEU: $x_{n+1}$7-gram overlap on subcategory label sequences.

For a candidate sequence and a bank of human reference runs, BLEU is computed as:

$x_{n+1}$8

with clipped $x_{n+1}$9-gram precision for $$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$0-grams and brevity penalty BP. This metric assesses not just event-level statistics (e.g., patch switch points) but also the global structure, diversity, and plausibility of entire runs.

Empirical results indicate that models relying solely on random walks or naive cue-switching yield low $$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$1-gram overlap scores. Augmenting transition probabilities with explicit subcategory bias or empirically calibrated global cue weights substantially improves alignment with human sequences (Heineman et al., 2024).

5. LLMs and Hybrid Mechanisms

Heineman et al. extend the evaluation of semantic fluency to neural LLMs, specifically using Llama 2 Chat 7B. The LLM is prompted with a natural-language description of the category fluency task and generates sequences autoregressively, with log probabilities and entropy recorded per step. Analysis shows that next-token entropy spikes align with human patch switches, paralleling foraging-threshold dynamics.

Out-of-the-box LLM performance is poor in matching human run structure, but the addition of a global cue (by upweighting empirically frequent exemplars) boosts both Exemplar BLEU and Category BLEU, outperforming classical cue-switching models. However, balancing between local and global cues ($$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$2) is critical to recover human-like switch rates and patch lengths—a single cue alone is insufficient.

6. Deterministic Versus Stochastic Search and Their Consequences

The choice of search strategy is found to be decisive, independent of the underlying transition model. Deterministic strategies such as greedy or low-width beam search deliver the most human-like exemplar orderings (highest BLEU scores $$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$3–$$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$4), emphasizing the importance of purpose-driven retrieval. Introducing stochasticity through temperature-based sampling or random walk degenerates performance, with sequences becoming less coherent and misaligned with human production statistics.

This highlights that, while a structured semantic space and an appropriate mixture of local/global cues are necessary, realization of human-like retrieval additionally requires a deterministic search process capable of traversing patches and triggering switches in a manner consistent with category structure (Heineman et al., 2024).

7. Summary of Modeling Implications in Semantic Fluency

Research converges on several key observations:

• Patch switch prediction is not unique to any single class of models—random walks, foraging-based cue switch, and LLM entropy-based models all generate the telltale IRT spike pattern.
• Full sequence generation and $$P(x_{n+1}\mid x_n)\;\propto\; [P_{\mathrm{local}}(x_{n+1}\mid x_n)]^{\beta_l} \times [P_{\mathrm{global}}(x_{n+1}\mid \mathrm{“animal”})]^{\beta_g}$$5-gram overlap metrics more stringently constrain model fit to human data than patch-level diagnostics.
• Structured semantic networks with strong small-world properties, explicit subcategory structure, and appropriately parameterized cue mixture (local/global and subcategory-level) are necessary for human-like sequence generation.
• Deterministic or beam search procedures outperform stochastic walks in capturing the exploitation–exploration balance revealed in human data.
• LLMs require global retrieval biases and balanced cue influence to align with empirical output; otherwise, they fail to emulate either patch structure or switching dynamics.

A plausible implication is that cognitively plausible memory search in semantic fluency requires (i) explicit representation of category structure in the semantic space, (ii) dual-cue mechanisms for balancing local exploitation with global exploration, and (iii) deterministic, targeted search methods rather than purely stochastic exploration (Heineman et al., 2024, Miscevic et al., 2017, Nematzadeh et al., 2016, 0903.4132). These components collectively provide a refined framework for modeling, analyzing, and evaluating both human and machine performance in category fluency paradigms.