Word Association Tests

Abstract

Word Association Tests, which the National Library of Medicine's Medical Subject Headings classifies under personality tests, present a person with a stimulus word and record the first word it brings to mind, treating the response and its latency as evidence about mental organisation. This article traces the method from Francis Galton's timed self-experiments, through Carl Jung's use of blocked and delayed responses as a complex indicator, to Grace Kent and Aaron Rosanoff's frequency tables, which turned a clinical probe into a standardised test with population norms. It then follows the method's migration into cognitive science, where response frequencies became data on semantic memory, drove the discovery of semantic priming, and were rebuilt at scale as small-world semantic networks. Three demonstrations model response-frequency norming, semantic priming through spreading activation, and the network structure recovered from association data.

Keywords: word association test, free association, semantic networks

A word association test asks a respondent to hear or read a stimulus word and to say the first other word that comes to mind. The instruction is trivial; what makes it a scientific instrument is the interpretation placed on the answer. The response reveals which word, out of the whole vocabulary, was most strongly linked to the cue in that person's mind, and the time taken to produce it indexes how readily that link was traversed. Over almost a century and a half the same simple procedure has been read three different ways: as a window onto individual temperament and hidden conflict, as a standardised clinical sign scored against population norms, and as a direct measurement of the associative structure of semantic memory. This article follows that sequence, from Galton's stopwatch to the million-response network norms of the present day.

What a Word Association Test Is

A word association test is any procedure that presents a stimulus, or cue, word and records the word a respondent produces in reply, usually the first that comes to mind. Two design choices define the family. In a discrete (or single-word) test the respondent gives one response per cue and then moves on; in a continuous (or continued) test the respondent keeps producing associates to the same cue for a fixed interval, yielding a chain of responses. Cutting across that distinction is the difference between a free association test, in which any word is acceptable, and a controlled test, in which the response is constrained, for instance to an antonym or a superordinate category. The unconstrained, discrete free association test, one response to each of a list of cues, is the classical form and the one Kent and Rosanoff standardised.

Because MeSH files the method beneath personality tests at tree position F04.711.647.905, its historical home is the assessment of the individual: what a person's associations, and the disturbances in them, reveal about temperament, conflict, or pathology. But the very feature that makes an association informative about a person, that the response is drawn from that person's own web of learned connections, also makes it informative about the web itself. This double reading, association as a sign of the individual and association as a measurement of shared semantic structure, runs through the method's whole history and is the reason a nineteenth-century clinical curiosity became a workhorse of twenty-first-century memory research. Figure 1 sets out the three successive readings the same response has carried.

Galton's Invention and Jung's Complex Indicator

The method began with Francis Galton, who in 1879 built the first systematic association experiment on himself. He prepared a list of cue words, exposed them one at a time while hidden from view, and used a chronograph to time how long each took to summon an idea, recording the associations that arose and later analysing their sources in his own history (Galton, 1879). Two of his findings set the template for everything that followed: that associations are lawful and repeatable rather than random, many recurring across separate trials, and that their latencies vary systematically, so that the time to associate is as informative as the association itself. Galton also noticed how often his associations reached back into early life, an observation that pointed the method toward the study of the individual past.

Carl Jung, working at the Burgholzli Clinic in Zurich, turned that pointer into a clinical instrument. In his association method a standard list of cue words was read aloud and the patient's response and its reaction time were recorded; a response that was unusually delayed, repeated the stimulus, failed to appear, or was accompanied by a disturbance was treated as a complex indicator, a sign that the cue had touched an emotionally charged cluster of ideas the patient would rather not confront (Jung, 1910). The reaction-time delay was central: an idea that is defended against takes measurably longer to associate to, so the stopwatch, not the content alone, exposed the complex. Jung's use of the method made two lasting contributions, one substantive, the idea that measurable response disturbances index hidden affect, and one methodological, the insistence that latency be recorded alongside content. The technique entered clinical psychology's standard armamentarium, where it was later catalogued among the projective and diagnostic instruments whose rationale is that unstructured responses expose material structured interviews miss (Rapaport et al., 1968).

The Kent-Rosanoff Test and the Birth of Norms

Jung's method diagnosed an individual by comparing a response against clinical intuition. Grace Kent and Aaron Rosanoff replaced the intuition with a table. Working at Kings Park State Hospital, they administered a fixed list of 100 common, emotionally neutral stimulus words to 1,000 normal subjects and tabulated every response, producing the first large frequency tables of what people ordinarily say to a given cue (Kent & Rosanoff, 1910). With those tables an individual's answer could be scored objectively: a response given by many of the normative sample was common, one given by few or none was individual (idiosyncratic), and the count of common responses across the list became the commonality score, a single number expressing how conventional a person's associations were. Kent and Rosanoff reasoned that psychiatric patients would give more idiosyncratic responses, and the frequency table let that hypothesis be tested rather than merely asserted.

The Kent-Rosanoff Free Association Test was a decisive methodological step: it converted association from a qualitative sign into a standardised, norm-referenced measurement, the same move that separates a clinical impression from a psychometric test. Its frequency tables also had a second life the authors did not foresee. A table of how often each response follows each cue is, read the other way, a map of the strength of association between words in the average mind, and that map is exactly the raw material cognitive psychology would later need. The first demonstration builds a small response-frequency table and computes a commonality score from it, showing how the same tabulation serves both the clinical and the structural reading.

From Clinical Sign to Semantic Structure

By the middle of the twentieth century the response frequencies Kent and Rosanoff had gathered to distinguish patients from normals were being reused to study the normal mind's organisation. Weston Bousfield showed that when people freely recall a list of words drawn from a few categories but presented in scrambled order, they nonetheless recall them in category clusters, grouping animals with animals and names with names, direct evidence that the words were connected in memory along associative and semantic lines rather than stored as an unrelated list (Bousfield, 1953). James Deese pushed the logic further, using association norms to predict errors: he showed that a list of words all strongly associated to one absent word would reliably provoke intrusions of that missing word in recall, so that the associative structure of a list determined not only what was remembered but what was falsely remembered (Deese, 1959). Deese's demonstration became the basis of the modern Deese-Roediger-McDermott false-memory paradigm, in which studying a list of associates reliably produces confident recall and recognition of the unpresented word they converge on (Roediger & McDermott, 1995).

The decisive bridge to cognitive psychology was the discovery of semantic priming. David Meyer and Roger Schvaneveldt had people judge whether letter strings were real words and found that a word was recognised faster when the preceding word was associatively or semantically related, nurse judged more quickly after doctor than after bread, showing that reading one word automatically makes its associates more available (Meyer & Schvaneveldt, 1971). Priming gave the association a mechanism: activation released at one node of a semantic network spreads along its links to neighbouring nodes, pre-activating them so that a related word is processed more efficiently, the account Allan Collins and Elizabeth Loftus formalised as a spreading-activation theory of semantic processing (Collins & Loftus, 1975). In this reading a word association test is a direct probe of that network, the response to a cue being simply the neighbour that received the most spreading activation. The second demonstration models that process, letting activation spread from a cue to associates of varying strength and reading off the reaction-time facilitation that priming produces.

Association Norms at Scale

If association responses measure the links of a semantic network, then collecting enough of them should reconstruct the network itself, and the last three decades have been an exercise in doing exactly that. Douglas Nelson and colleagues assembled the University of South Florida norms, free association responses from more than six thousand participants to more than five thousand cues, for years the standard association database in cognitive research (Nelson et al., 2004). Mark Steyvers and Joshua Tenenbaum then analysed such norms as graphs and found that the resulting semantic network has small-world structure, most word pairs separated by only a few associative steps despite sparse connectivity, together with a heavy-tailed degree distribution in which a few hub words have very many connections, properties they linked to a model of how semantic networks grow (Steyvers & Tenenbaum, 2005).

Simon De Deyne and Gert Storms brought the same network lens to a new, larger collection and showed that association graphs carry rich semantic information, predicting similarity judgements and category structure from graph distances alone (De Deyne & Storms, 2008). A key methodological refinement followed: asking each participant for three successive responses to a cue rather than one, a continued rather than discrete task, yields denser, more reliable networks that explain lexical and semantic behaviour better than single-response data, because the second and third responses expose weaker links the first response hides (De Deyne et al., 2013). This network programme was consolidated as a general approach to representing the mental lexicon (De Deyne et al., 2016) and realised at scale in the Small World of Words project, whose English norms cover more than twelve thousand cue words from hundreds of thousands of participants (De Deyne et al., 2019), with parallel norms now built for other languages including Rioplatense Spanish (Cabana et al., 2024). The third demonstration contrasts the sparse network recovered from single responses with the denser one recovered from continued association, reproducing the qualitative gain the continued task provides.

Current Directions

The most active use of word association today is as ground truth for models of the mental lexicon. Because the Small World of Words norms record, for thousands of cues, exactly which words hundreds of thousands of people produce, they serve as a benchmark against which distributional models of meaning, from co-occurrence counts to the embeddings inside large language models, can be tested: a model that captures human meaning should reproduce human association strengths (De Deyne et al., 2019). Reviews of semantic memory now treat association networks as one of the two great families of semantic representation, set beside distributional models derived from text, and a central question is how far each explains behaviour the other cannot (Kumar, 2021). The extension of the norming programme to further languages, including Rioplatense Spanish, is turning that comparison cross-linguistic, asking which features of the association network are universal and which are shaped by a particular language and culture (Cabana et al., 2024).

A second, more sceptical current re-examines the method's founding assumption, that the response is genuinely the first word to come to mind. When respondents were asked to report their actual first thought and then compare it with the word they wrote, a nontrivial fraction acknowledged editing the response, suppressing a first association that seemed embarrassing, taboo, or too personal in favour of a safer one (Playfoot et al., 2018). The finding does not undo the method, whose norms average over exactly such noise, but it recovers, in modern experimental form, the very phenomenon Jung built his method on: that what a person declines to say to a cue can be as informative as what they say.

Worked Example

The first demonstration makes the commonality score concrete. It shows a small normative table for a cue word, listing each recorded response and the number of the 100 normative respondents who gave it. Selecting a response as the one this person produced reads its normative frequency straight from the table: a response given by 68 of 100 people scores as highly common, one given by 2 of 100 as idiosyncratic. Summing a respondent's frequencies across several cues yields the commonality score Kent and Rosanoff used, so that a person who consistently gives the modal response scores high and a person who consistently gives rare responses scores low. The demonstration encodes the definition, not an estimate: the score is nothing more than the sum of the chosen responses' normative counts.

The second demonstration turns semantic priming into arithmetic. Recognising a word in a lexical decision task takes a baseline time when the preceding word is unrelated; a related prime shortens that time in proportion to how strongly the two are associated. Take a baseline unrelated-prime reaction time of 620 milliseconds and a maximum facilitation of 80 milliseconds for a perfectly associated prime. A prime of association strength 0.75 then produces a facilitation of 0.75 times 80, which is 60 milliseconds, so the primed reaction time is 620 minus 60, or 560 milliseconds, and the measured priming effect is 60 milliseconds. A weaker prime of strength 0.30 facilitates by 0.30 times 80, which is 24 milliseconds, giving a primed time of 596 milliseconds and a priming effect of only 24 milliseconds. The slider moves the association strength and reads the facilitation and the resulting reaction time straight off this rule, so that stronger associates always produce larger, faster priming.

The third demonstration counts the edges a task recovers. Imagine six cue words. In the discrete task each respondent gives one response per cue, so the six cues contribute six forward links, a mean out-degree of exactly one. In the continued task each respondent gives three responses per cue, so the same six cues contribute up to eighteen links, a mean out-degree of three, a threefold increase in the edges available to connect the network. Because a network with more edges has shorter paths between its nodes, the denser continued-task graph links words that the sparse discrete graph leaves unreachable, which is why continued association predicts semantic behaviour better. The demonstration toggles between the two modes and reports the edge count and mean out-degree for each, reproducing the density gain that motivated the continued task.

Discussion

The word association test is a rare instrument that has been genuinely useful under three incompatible theories of what it measures. For Jung it measured the individual unconscious, the delayed or blocked response betraying a defended complex. For Kent and Rosanoff it measured conformity of thought, the commonality score placing a person on a scale from conventional to idiosyncratic association. For cognitive science it measures neither the person nor their conventionality but the shared structure of semantic memory, the response revealing which neighbour in a semantic network lies closest to the cue. What allowed one procedure to serve all three is that a single association response carries information about all three at once: it is the respondent's own most available neighbour of the cue (individual), it is more or less the one most people give (normative), and it is by definition an edge in the semantic graph (structural). Table 1 sets the three readings side by side.

Read as a whole, the history of the word association test is a case study in how a measurement outlives its theory. The data Kent and Rosanoff gathered to sort patients from normals are, a century later, still analysed, now as edges in a semantic graph, because a record of what word follows what is theory-neutral: it can be re-read whenever a new theory tells us what association means. The current network programme, and its use of human association norms to benchmark machine models of meaning, is the latest such re-reading, and it is unlikely to be the last.

Common Misconceptions

That a word association test measures only personality or the unconscious.

That was its first use, but the same responses measure the shared structure of semantic memory, and most modern association data are collected to map that network rather than to assess an individual (Steyvers & Tenenbaum, 2005).

That the response given is always the genuine first word to come to mind.

When asked, a substantial minority of respondents admit editing out an embarrassing or taboo first association, so the recorded response is partly filtered, the very editing Jung's method was designed to detect (Playfoot et al., 2018).

That an idiosyncratic response is a sign of disorder.

The commonality score is a population statistic, not a diagnosis; rare responses are common in healthy, creative, and simply original respondents, and only extreme or clinically corroborated patterns carry diagnostic weight (Kent & Rosanoff, 1910).

That association strength is symmetric.

It is not: salt strongly evokes pepper more often than pepper evokes salt, so association norms are directed graphs and forward and backward strengths must be kept distinct (Nelson et al., 2004).

Glossary

Association norms.

Tabulated frequencies of the responses a population gives to each of a set of cue words, used both to score individuals and to estimate the strength of association between words.

Commonality score.

The sum, across a list of cues, of the normative frequencies of a person's responses; a high score marks conventional association, a low score idiosyncratic association.

Complex indicator.

In Jung's method, a disturbed response, delayed, absent, repeated, or agitated, taken as a sign that the cue word has touched an emotionally charged cluster of ideas.

Continued association test.

A task in which the respondent produces several successive associates to one cue, yielding denser networks than a single-response task by exposing weaker links.

Discrete association test.

A task in which the respondent gives exactly one response to each cue before moving to the next; the classical single-response form standardised by Kent and Rosanoff.

Forward association strength.

The proportion of respondents who give a particular response to a cue; a directed measure, since the strength from cue to response need not equal the strength in reverse.

Free association test.

A word association procedure in which any response is permitted, as opposed to a controlled test that constrains the response to a class such as an antonym or category.

Idiosyncratic response.

A response given by few or none of a normative sample, contrasted with a common response; also called an individual response.

Kent-Rosanoff Free Association Test.

The 100-word free association test standardised on 1,000 normal subjects in 1910, whose frequency tables introduced norm-referenced scoring of associations.

Lexical decision task.

A procedure in which a respondent judges as fast as possible whether a letter string is a real word, the reaction time serving as the standard measure of semantic priming.

Semantic network.

A representation of semantic memory as nodes (concepts or words) joined by links (associations), in which meaning is carried by a node's pattern of connections.

Semantic priming.

The finding that a word is recognised faster when preceded by an associated or semantically related word, taken as evidence that activation spreads between linked concepts.

Small-world network.

A network in which most nodes are reachable from one another in few steps despite sparse connectivity and high local clustering; the structure found in word association graphs.

Spreading activation.

The proposed mechanism of priming, in which activating one node of a semantic network releases activation that flows along its links to pre-activate neighbouring nodes.

Word association test.

A procedure that presents a stimulus word and records the respondent's associated word, and often its latency, as evidence about mental organisation.

Key Researchers

Marc Brysbaert. Experimental psychologist at Ghent University who builds large lexical databases and word-frequency norms and co-authored the Small World of Words English association norms. ORCID - Google Scholar

Simon De Deyne. Computational cognitive scientist at the University of Melbourne who leads the Small World of Words project, rebuilding association norms at scale as semantic networks. ORCID - Google Scholar

Francis Galton (1822-1911). Independent scholar in London who introduced the timed word-association experiment in 1879, establishing that associations are lawful, repeatable, and measurable by their latency. Wikipedia - Wikidata

Carl Gustav Jung (1875-1961). Psychiatrist at the Burgholzli Clinic, University of Zurich, who developed the association method as a complex indicator, reading blocked and delayed responses as signs of emotionally charged material. Wikipedia - Wikidata

Grace Helen Kent (1875-1973). Psychologist who co-developed the Kent-Rosanoff Free Association Test and its frequency tables, the first large standardised norms for word association. Wikipedia - Wikidata

Abhilasha A. Kumar. Cognitive psychologist at Bowdoin College who reviews the models that turn association data into theories of semantic memory. ORCID - Google Scholar

Aaron J. Rosanoff (1878-1943). Psychiatrist who, with Kent, supplied the clinical framing and the standardisation of the free association test on 1,000 normal subjects. Wikipedia - Wikidata

Mark Steyvers. Cognitive scientist at the University of California, Irvine who modelled word association data as small-world semantic networks and a model of semantic growth. ORCID - Google Scholar

Gert Storms. Experimental psychologist at KU Leuven who co-developed the Small World of Words norms and studies semantic concepts and categorization. ORCID - Google Scholar

Frequently Asked Questions

What is a word association test? It is a procedure that presents a stimulus word and records the first word the respondent produces in reply, together with how long the reply takes, treating both as evidence about how that person's memory is organised (Galton, 1879).

What did Jung use word association for? Jung read the response and its reaction time as a complex indicator: a delayed, blocked, or disturbed response signalled that the cue word had touched an emotionally charged, often unconscious cluster of ideas (Jung, 1910).

What is the Kent-Rosanoff test? It is a 100-word free association test standardised on 1,000 normal subjects in 1910, whose frequency tables allowed a person's responses to be scored as common or idiosyncratic against population norms (Kent & Rosanoff, 1910).

What is a commonality score? It is the sum of the normative frequencies of a person's responses across the test; a high score means the person tends to give the responses most other people give, a low score means their associations are unusual (Kent & Rosanoff, 1910).

How does word association relate to semantic memory? The response to a cue is the word most strongly linked to it in memory, so association responses map the links of a semantic network and can be assembled into a graph of the mental lexicon (Steyvers & Tenenbaum, 2005).

What is semantic priming and how is it connected? Semantic priming is the speeding of word recognition when a related word precedes it; it gives association a mechanism, spreading activation, in which reading a cue pre-activates its associates (Meyer & Schvaneveldt, 1971).

Why ask for several responses to one cue instead of one? Continued association, three responses per cue rather than one, exposes weaker links the first response hides, producing denser and more predictive semantic networks (De Deyne et al., 2013).

Are the largest modern association norms freely available? Yes. The Small World of Words English norms cover more than twelve thousand cues from hundreds of thousands of participants and are widely used to benchmark models of meaning (De Deyne et al., 2019).

References