The semantic structure of accuracy in eyewitness testimony · Gustafsson PU, Sikström S, Lindholm T

Study 2

In this study, we aimed to replicate the findings in Study 1 with a new sample. That is, we (1) expected correct statements to occupy a different semantic space compared to incorrect statements. Moreover, we expected that incorrect and correct statements would differ semantically in relation to the communion, positive valence and dominance representations. Specifically, we hypothesized (2) that incorrect statements would be more closely related to communion and (3) positive valence, whereas (4) correct statements would be more closely related to dominance. Additionally, we wanted to explore potential linguistic differences between correct and incorrect statements further. We therefore added abstractness as an additional semantic representation. This idea was inspired from reality monitoring theory (Johnson and Raye, 1981), which postulates that people judge memories with greater sensory and temporal details as real, and judge memories with references to cognitive operations as imagined. Parallels can be drawn between correct statements and real memories on one hand, and incorrect statements and imagined memories on the other hand, as correct statements are recollections of that which did happen, whereas incorrect statements are recollections of that which did not happen. As cognitive operations reflect more higher-order functions (i.e., reflections on the memory itself) we expected (5) incorrect memories to be more closely related to an abstract semantic representation. This study has been preregistered2.

Method

Dataset

As the semantic analyses work better with greater datasets, this dataset contains combined data from two studies, both conducted in Swedish. The first half of the dataset comprised witnesses in Study 2 from Lindholm et al. (2018), specifically the participants that had viewed a mock crime of a stabbing attack (n = 10; Mage = 24.90, SD = 6.42; 50% men). The second half of the dataset comprised witnesses (n = 22; Mage = 24.50, SD = 4.97; 32% men) from Gustafsson et al. (2019), who had viewed the same mock crime video. All interviews went through the same transcription and coding procedure as described in Study 1 (see Lindholm et al., 2018; Gustafsson et al., 2019 for full descriptions of the procedures). This combined dataset yielded a total of 1,541 statements (75.47% correct).

The study was conducted in full in accordance with the ethical principles outlined in ALLEA (All European Academies) (2017), the ethical principles outlined on http://www.codex.vr.se/, and with the 1964 Helsinki declaration and its later amendments. The studies did not include factors that require ethical vetting according to Swedish legislation on research ethics Svensk författningssamling (2003). All participants had given informed consent to participate.

Data analysis

Data analyses were identical to those carried out in Study 1, with the addition of an analysis for the abstractness semantic representation. This abstractness representation was created with the same procedure as the semantic representations in Study 1, using the Swedish normwordlist (Waldhauser, 2022). Examples of abstract words were “time” and “plan” whereas examples of concrete words were “pizza” and “snake.”

Results and discussion

Accuracy and semantic spaces

To test the first hypothesis that correct and incorrect differed in semantic space, we again used t-tests for the BERT and LSA-models. The results were statistically significant both with the BERT analysis, t(1534) = −3.52, p < 0.001, d = 0.21, MSE = 1.07 (see Table 1), and with the LSA analysis, t(1534) = −8.29, p < 0.001, d = 0.48, MSE = 1.06 (see Table 1).

Semantic representations

As in Study 1, we investigated if correct and incorrect statements differed in relation to specific semantic representations (Hypotheses 2–5). Again, we used t-tests to examine the semantic scores for correct and incorrect statements in relation to each representation. Results are shown in Table 1. In line with hypotheses, the results showed that incorrect statements were more closely related to an abstractness dimension. Contrary to hypotheses, results showed no significant results for communion, positive valence, nor dominance.

Overall, we replicated the results of the first study, demonstrating that correct and incorrect statements differ in semantic content. This was shown both with the BERT analysis and the LSA analysis. However, regarding semantic representations and accuracy, we did not replicate the findings from Study 1 in which incorrect statements were more closely related to communion and correct statements more closely related to dominance (see Table 1). Instead we found that abstractness was higher for incorrect statements. This latter finding corroborates findings on reality monitoring (Hashtroudi et al., 1990; Stromwall and Granhag, 2005; Kensinger and Schacter, 2006; Sporer and Sharman, 2006), in which incorrect statements are uttered with greater “cognitive operations” than correct memories, which should entail more abstract ideas.