Abstract:

The semantic search during creative thinking refers to the activation process of semantic information in long-term memory involved in creative activities. Influential theory has posited that the semantic activation process in free recall shows spreading activation within semantic network and is characterized by negative acceleration and clustering. Unlike the free recall, it is necessary to suppress the dominant response and to activate novel and distant information during creative thinking. Therefore, one might expect different semantic search processes during creative thinking, but such a hypothesis has not yet been directly tested. To explore the semantic search process during creative thinking, the present study described the quantitative dynamic characteristics of answer generation in a divergent thinking test using a series of parameters, such as cumulative response distribution and semantic similarity.
The experiment employed a within-subject design with the task type (novel V.S. normal) as the independent variable. The experiment included two versions of alternative uses task (AUT): novel and normal AUT. In the novel AUT, 39 participants (30 females; aged 18 to 22) were asked to report novel and valid uses for the daily-life items presented on the screen as many as possible, while in the normal AUT they were only asked to think of valid uses for objects as many as possible. During the experiment, participants completed two normal AUTs, followed by two novel AUTs. Each AUT lasted for three minutes. The novelty of responses and semantic similarity of responses were scored by participants themselves. The time function of the cumulative number of responses was fitted by the hyperbolic function, and clustering analysis was conducted based on the semantic similarity of responses.
The results showed that:
(1) The cumulative response distribution in the novel AUT condition was negatively accelerating similar to semantic search during free recall, but the search speed in the novel AUT condition was slower than that of the normal AUT condition (Figure 1).
Specifically, a nonlinear fit of $y~=~\frac{ax}{b+x}~$was performed on the answers time sequences generated by each subject in each AUT, and three answers sequences (two in the normal condition and one in the novel condition) were found to be non-convergent, which were removed in this section but retained in the semantic relationship analysis section. The average R² of the remaining 153 trials successfully converged fits was 0.96, the RMSE was 0.38, and the average fitting curve was shown in Figure 1.
The Kolmogorov-Smirnov method was used to test the normality of the fitted parameters a and b. It was found that they did not follow a normal distribution, so the median and the maximum values were used to represent the estimates of the fitted parameters. The Mann-Whitney U test was used to compare the differences of parameters a and b under different conditions (Table 1). The results showed that parameter a was not significantly different between the two conditions (p > 0.05), but parameter b was significantly higher in the novel condition than in the normal condition (p < 0.001).
(2) In the novel AUT condition, the semantic similarity between participants’ responses (i.e., the answers) and the items (i.e., the questions) was low and significantly lower than that in the normal AUT condition.
Specifically, we compared the semantic similarity between answers at each sequential position and questions in both conditions using an independent samples t-test. It was found that, except for position 6 and position 9, the semantic similarity between the answers and the questions in the normal condition was significantly higher than that in the novel condition (see Table 2).
A third-order polynomial function was used to fit the semantic similarity between the answers at each position and the questions, as shown in Figure 2.
(3) The responses in the novel AUT condition showed a significantly lower degree of clustering than that in the normal AUT condition. In the novel AUT condition, the semantic similarity between the clusterable and non-clusterable answers and the questions were low and not significantly different. Furthermore, there was no significant difference between the clusterable and non-clusterable answers in terms of novelty.
Specifically, according to the scoring rules of semantic similarity, the semantic distance (d) less than or equal to 0.5 was taken as the clustering criterion, and the clustering degree of the answers under different conditions was calculated separately. Since there was a significant difference in the number of answers between the normal and novel conditions, the percentage of clustered answers to the total number of answers was used to represent the degree of clustering. Independent sample t-tests were used to compare the differences in the degree of clustering between the conditions. The results showed that the degree of clustering was significantly higher in the normal condition than in the novel condition, t (154) = 4.72, p < 0.001, Cohen’s d = 0.76, 95% CI: [0.11, 0.27].
In addition, a 2 × 2 repeated-measures ANOVA was conducted with experimental condition (normal vs. novel condition) and clustering condition (clusterable vs. non-clusterable) as independent variables, and the mean novelty of answers and the mean semantic similarity between answers and questions as dependent variables (data from one subject who did not have clustered answers in the normal condition and seven subjects who did not have clustered answers in the novel condition were removed).
The ANOVA results showed that the main effects of both the experimental conditions and clustering conditions on the mean novelty of answers were significant (experimental condition: F (1, 30) = 59.56, p < 0.001, partial η² = 0.67; clustering condition: F (1, 30) = 27.03, p < 0.001, partial η² = 0.47). The interaction was significant, F (1, 30) = 4.88, p = 0.035, partial η² = 0.14. The results of the simple effects analysis showed that the novelty scores of the clusterable answers were significantly lower than those of the non-clusterable answers in the normal condition, F (1, 30) = 21.21, p < 0.001, partial η² = 0.41; while in the novel condition, there was no significant difference between the novelty scores of the two types of answers (p > 0.05) (see Table 3).
The main effect of both the experimental conditions and clustering conditions on the semantic similarity between answers and questions were also significant (experimental condition: F (1, 30) = 59.02, p < 0.001, partial η² = 0.66; clustering condition: F (1, 30) = 23.09, p < 0.001, partial η² = 0.44). The interaction was significant, F (1, 30) = 21.60, p < 0.001, partial η² = 0.42. The results of the simple effects analysis showed that the semantic similarity between the clusterable answers and questions was significantly higher than that between the non-clusterable answers and questions in the normal condition, F (1, 30) = 37.13, p < 0.001, partial η² = 0.55. In the novel condition, there was no significant difference between the clustered and non-clusterable answers in terms of semantic similarity to the questions (p > 0.05) (see Table 3).
These findings indicated that the semantic search during creative thinking was partly in line with spreading activation theory of semantic search in free call. But the search speed was relatively slower. Furthermore, the novelty requirement prompted the participants to break up the semantic restriction of the item at the initial search and avoid nearby search. The participants tended to generate few ideas in each semantic field. However, when it is far away from the item in the semantic field, individuals might generate clustering ideas.

Key words: creative thinking, semantic search, negative acceleration, clustering, semantic similarity