Get the Picture: Learning Referents in a Single-Day Context

Abstract

Associative learning provides a common substrate for studying learning and memory in language sciences. Studies on recognition memory tasks have found higher proportions of homonyms relative to non-homonyms, indicating a preference or optimum in word learning behavior. This study sought to identify the differences between associative learning for a new word form (pseudowords) and multiple referents (MR) as homonyms and single referents (SR) as non-homonyms. Encoding comprised stimuli for MR, SR, and another condition that comprised pseudowords alone (PW). After the learning phase, two tests were conducted for the participants to judge words from each condition (and an additional condition of untrained word forms), and pictorial referent figures. The results showed that proportions reflecting efficient learning favored SR over MR. The accuracy rates of the verbal working memory task were positively correlated with those of the SR condition in the first recognition task and for the MR condition across recognition tasks. These findings show that associative learning for pseudowords with an SR was more efficient than MR in a single-day learning context. Implications for future research on learning efficiency in vocabulary acquisition and multimedia learning are discussed.

Keywords:

• associative learning
• word learning
• multiple referents
• Japanese
• pictorial referents

1. Introduction

Individual differences in cognitive abilities and language experience are a part of a long research tradition in cognitive psychology and psycholinguistics. Conceptual integration has been broadly attempted with research on learner resources and self-regulated learning in educational psychology (Seufert 2020) and the optimization of processing through germane load in cognitive load theory and its related instructional implications (Paas and van Merriënboer 2020). Despite robust relationships between working memory and second language proficiency being observed beyond the effects of inhibitory control in the early stages of learning (Linck and Weiss 2011), links have been less theoretically established for the discipline of second language acquisition (SLA; Wen 2016; Wen and Schwieter 2022), with some research agenda exceptions (e.g., Suzuki, Nakata, and Dekeyser 2019). Meta-analyses summarized by Wen (2016) point to the need for both a sounder theoretical and methodological basis for SLA research, and mutual recognition of the unique features of SLA vis-à-vis L2 vocabulary, formulaic sequences, grammar, comprehension (reading and listening), production (writing and speaking), and their attendant processes (Wen 2016). As an approach that underlies first (L1) and second language (L2) research, associative learning focuses on the capacity of individuals to form associative links in memory (Wen 2016). A causal interaction between modal and associative brain regions for learning and word acquisition has also been observed (Vukovic et al. 2021). Vukovic et al. (2021) found that a short session on language learning focusing on action-related language induced rapid microstructural changes reflecting lexico-semantic processing and the formation of new memory circuits. To address the call for the integration of research needs along the nexus of working memory and language (Wen and Schwieter 2022) and to connect to empirical work on the sensorimotor perspective (e.g., Vukovic et al. 2021), an examination of associative learning for non-words or pseudowords and pictorial referents with reaction time methods would represent a common methodological framework of research paradigms recognized across disciplines in linguistics (Kaiser 2014) with clear practical implications (e.g., Clark and Paivio 1991; Yu and Trainin 2022). We seek to contribute to this common basis with a demonstration of efficient artificial learning through associative processing, ab initio, of pseudowords as paired associates (i.e., word-picture pairs) of written linguistic information (i.e., logogens) and perceptual representations of features (i.e., imagens) that we consider reflective of sensorimotor experiences (Paivio 1986; Pulvermüller 2003; Vihman 2022) with special attention to noun types.

2. Literature review

Psycholinguistic studies have reported that word forms are associated with single referents (SR; Breitenstein et al. 2005; Cornelissen et al. 2004; Cosper, Männel, and Mueller 2020; Cosper, Männel, and Mueller 2022; Grönholm et al. 2005; Havas et al. 2018; Hawkins, Astle, and Rastle 2015; Hawkins and Rastle 2016; Hultén et al. 2009; Kambara et al. 2013; Lee et al. 2003; Li, Fan, and Wang 2020; Miller et al. 2017; Takashima et al. 2014; Takashima et al. 2017; Vanek, Sóskuthy, and Majid 2021; Yan et al. 2021; Yang et al. 2021) or multiple words that are already associated with multiple referents (MR) from associative learning (e.g., linguistic conditioning; Ando and Kambara 2023; Cicero and Tryron 1989; Paivio 1964; Staats, Staats, and Heard 1959; Staats et al. 1959; Staats, Staats, and Heard 1961; Staats and Staats 1957; Staats and Staats 1958; Tryon and Cicero 1989). Based on these studies, the associative patterns of word forms and referents can be categorized into the following associative patterns. The first comprises a single word form and MR, which may be characterized by homonyms. The second comprises a single word form and referent, which may best be represented by non-homonyms. Learning homonyms has been shown to be more difficult than non-homonyms (e.g., Doherty 2004; Ellis 2008). Homonymic words complicate the vocabulary acquisition process as encounters with the word become divided (Parent et al. 2023) from the multiple referents and secondary meanings, lowering their salience compared to non-homonyms (Ellis 2008). Homonyms present a problem of ambiguity as their referents are separately represented (ambiguity effects; see Hino, Lupker, and Pexman 2002). Hino and colleagues reported that ambiguous words would be influenced in semantic categorization task associated with semantic processing (Hino, Lupker, and Pexman 2002). Owing to the implied use frequency preferences for these word types, this interpretation of the previous findings suggests that people may retrieve acquired homonymous words, which include MR, better than acquired non-homonymous words that would correspond to SR.

Another finding showed that the division of attention introduces a large disruptive effect for processing source information, even more than for processing item information in a single day of learning (Troyer et al. 1999). This aligns with research on the cognitive load theory, where learning from integrated sources of information imposes a lower extraneous form of cognitive load (Paas and van Merriënboer 2020), unless extraneous load is taken to some degree of excess. Stark and colleagues reported that participants (both amnesic patients and healthy) could recognize separate single items more accurately than associations between both (i.e., items and source memories; Stark, Bayley, and Squire 2002). Such participants could also recognize single items better than item pairs (i.e., items and sources; Turriziani et al. 2002). These findings suggest that participants can easily recognize individual over two items, and that performance vis-à-vis word recognition memory tasks depends on the number of items that the participants retain after a learning session, especially for learning that occurs in a single day, to which new (meaningless) stimuli would not be expected to include associations with acquired memories that have already been stored in long-term memory (semantic richness). Thus, participants may learn the associative pairs of a new word form and a SR more efficiently than MR owing to the effects of divided attention that emerge in the context of associative learning.

New word learning is associated with the verbal working memory capacity of new word forms (e.g., pseudowords). Gathercole and Baddeley (1990) showed that children with disordered language development could not repeat new word forms (pseudowords) better than younger children with matched verbal abilities. The performance of a verbal working memory task positively correlated with vocabulary scores (Gathercole and Baddeley 1989). For older individuals in another study, L2 word learning performance correlated with performance on a verbal working memory task (Service and Craik 1993). Working memory was observed as a moderate, language-general correlate of L2 reading comprehension in a meta-analysis of 59 studies (Jeon and Yamashita 2014), and in an example of L2 learners performing an elicited imitation test of oral proficiency, facilitative effects for phonological short-term memory capacity of semantically empty sequences of non-words were observed for less experienced learners relative to experienced ones who lacked these effects (Park et al. 2020). Thus, the accuracy rates of a verbal working memory task may be expected to correlate with those of associative pairs of new words and referents in recognition and retrieval tasks.

3. Purpose of the study

This study sought to examine whether associative learning for a new word form and SR is more efficient than MR. Meaningless word forms in a first language (L1; Japanese) were used as pseudowords and figures were used as pictorial referents. During the learning session, participants were exposed to three training conditions: associative pairs for a pseudoword with multiple (two) referents (MR) and for a pseudoword and a single referent (SR), and pseudowords only (PW). Studies have suggested that blocking rather than interleaved methods are better for learning (e.g., Carpenter and Mueller 2013; Goldstone 1996). Therefore, we opted for a block design for the MR, SR, and PW conditions in the learning task. After encoding, two types of tests were conducted. In the first test with retrieval tasks, participants determined whether presented words were categorized into MR, SR, PW, or untrained (UT) word form conditions. In the UT condition, participants did not directly check associated referents and words were not retained by the participants in the learning session. Words in the PW condition were retained. In real life, we learn each word with a context or referent. When we learn the word with multiple contexts or referents, the word would be categorized into MR in real life. Thus, we can imagine that the number of contexts or referents in real life would be associated with conditions in this retrieval task. Furthermore, in real life, the retrieval task would be associated with meaning recall, not meaning recognition. An applied linguistic study suggests that the meaning recall test would correlate with reading proficiency, compared to form recall and meaning recognition tests (McLean, Stewart, and Batty 2020). Therefore, we used both retrieval and recognition tasks in this study. In a second test with recognition tasks, participants chose the figure(s) to associate with the words presented in the three figure sets. After each judgment, the participants obtained the correct answer, which was provided as a form of feedback. They performed a verbal working memory task where they judged whether a target word form was present in the previously demonstrated five word forms.

Three predictions were made. First, the task performances (accuracies and response times) of the conditions (MR, SR, and PW) were expected to gradually improve in line with previous findings (e.g., Kambara et al. 2013; Liu et al. 2021). In the event that the participants misjudged items, those that were mistakenly judged as learned items were expected to increase from the first to the second retrieval tasks (e.g., for missed items in the MR condition, participants judged items as belonging to the SR or PW condition). This prediction was also expected to favor increases in task performance. Second, the task performance indices (accuracies and response times) of the SR condition would be better than those of the MR condition for both the recognition and retrieval tasks. Third, the accuracy rates of the verbal working memory task would positively correlate with those of recognition and retrieval tasks in the SR and MR conditions.

4. Materials and methods

4.1. Participants

A total of 30 university students (21 women; aged 18–23 years; M_age= 20.67; SD_age= 1.24) participated in this experiment. All participants were healthy and right-handed native Japanese speakers. We obtained informed consent from each participant before conducting the experiment. This experiment was approved by the ethical committee of the authors’ institution and was conducted in keeping with the Declaration of Helsinki.

4.2. Task procedures

We asked the participants to conduct four tasks: verbal working memory, learning, recognition, and retrieval (see Figures 1–4). The sequence began with verbal working memory, learning, retrieval, and recognition tasks, followed by the second and third recognition tasks, and ending with a second retrieval task (see Figure 5). We used SuperLab 4.5 and a Windows-based laptop for stimuli presentation and recording of key presses and reaction times as task performance.

Figure 1. Verbal working memory task.

Figure 2. Learning task.

Figure 3. Retrieval task.

Figure 4. Recognition task.

Figure 5. Experimental flow.

4.3. Verbal working memory task

All verbal stimuli were unfamiliar words collected in a Japanese psycholinguistic study (Koyanagi et al. 1960). We used unfamiliar words as pseudowords. Participants were instructed to rate their familiarity with each word on a 5-point scale with 1 being the most unfamiliar and 5 being the most familiar, and the familiarity rate ranged from 0.00 to 2.49 in line with Koyanagi et al. (1960). The word list is provided in Appendix A. We prepared 330 unfamiliar words for the verbal working memory task, 33 of which were selected for the practice rehearsal. For the former, we conducted 30 matching trials such that an unfamiliar target word was included in the 5 columns of 150 unfamiliar words. In the latter, we conducted 30 unmatching trials such that an unfamiliar target word was excluded from the first 5 columns of 180 unfamiliar words (Appendix A). Therefore, 60 trials were conducted in the verbal working memory task. We conducted 3 matching trials for the practice rehearsal of the task such that the unfamiliar target word was included into 5 columns of 15 unfamiliar words and 3 unmatching trials, where an unfamiliar target word was excluded from the first 5 columns of 18 unfamiliar words (Appendix B). Overall, six trials were conducted for the practice portion of the verbal working memory task. We randomized the trial order among the participants in both tasks. For the matching trials, the presentation number and order of the unfamiliar target word among the five was equally adjusted among the trials.

We revised a verbal working memory task used in prior studies (Kambara et al. 2017; Kambara et al. 2018), such that the participants judged whether the sixth word (a target unfamiliar word) was present among the previous five words by pushing one of two keys associated with the right index (presented) or middle fingers (not presented; Figure 1). Each unfamiliar word was displayed for 500 ms (Figure 1), followed by a fixation cross for 500 ms, starting from the first to the fifth unfamiliar word. A red colored fixation cross followed the fifth unfamiliar word for 2000 ms, and a post was displayed with the sixth unfamiliar word. If the participants judged that the sixth unfamiliar word was displayed among the five unfamiliar ones, they pressed a key associated with their right index finger. Otherwise, they pressed a key associated with their right middle finger. The sixth unfamiliar word was presented for 500 ms. After the sixth unfamiliar word, a fixation cross was displayed for 4500 ms. The participants conducted the practice rehearsal for the verbal working memory task before the full set of trials.

4.4. Learning, recognition, and retrieval tasks

We prepared 150 unfamiliar words and 90 meaningless figures for the learning, recognition, and retrieval tasks. The unfamiliar words (Koyanagi et al. 1960) and meaningless figures (Hirota 1982) were selected from two Japanese psychological databases. The familiarity rate for the selected unfamiliar words ranged from 0.00 to 0.99 in line with Koyanagi et al. (1960), who instructed their participants to evaluate familiarity values for each word form on a 5-point scale ranging from unfamiliar to familiar. The values of meaningfulness for the selected meaningless figures ranged from 1.41 to 1.86, in keeping with conventions established by Hirota (1982).

The 150 unfamiliar words were separated into 5 word lists (labeled A, B, C, D, and E). Each word list included 30 unfamiliar words. The A, B, and C word lists were used in the MR, SR, and PW conditions, whereas the D and E word lists were used in the UT condition. Further, 90 meaningless figures were categorized into 3 figure lists (labeled α, β, and γ). Each figure list contained 30 figures. We confirmed that there was no significant difference in meaningfulness for the figure lists by using between items analyses of variance (ANOVA; F(2, 87) = 0.08, n.s., f = 0.04) on js-STAR (https://www.kisnet.or.jp/nappa/software/star/). After randomization within each figure list, the α and β figure lists were collapsed into the MR condition, and the γ figure list was categorized as the SR condition. We made six patterns of word lists and conditions that were counterbalanced among the participants. All patterns of word lists and conditions are provided in Appendix C.

In the learning task, the participants pushed a key upon learning the associations between pairs of stimuli presented for each trial. Associative pairs comprised one word form and two pictures (MR), one word form and one picture (SR), and one word form only (PW) (Figure 2). Before the learning task, we explained to participants that: (1) we will present pairs of a word and meaning(s) as figure(s) (MR and SR) or words only (PW), (2) participants will be asked to push a key as soon as possible after they learn the presented pairs (MR and SR) or words only (PW). Each stimulus was presented for 2000 ms. After each stimulus, a fixation cross was presented for 2000 ms. Each condition was presented as a block in the learning task. The presentation order for each condition was counterbalanced among the participants. The order of stimuli was randomized in the task for each condition. Second, in the retrieval tasks, the participants judged the condition (MR, SR, PW, and UT) that the presented word form was associated with by pushing one of four keys associated with the right index, middle, ring, and little fingers, respectively (Figure 3). Each stimulus was presented for 2000 ms. After each stimulus, a fixation cross was presented for 2000 ms. The presentation order of each stimulus was randomized among the participants. Third, in the recognition tasks, the participants judged whether each presented word form was associated with one (SR condition) or two pictures (MR condition; Figure 4). The MR condition simultaneously presented three pairs of two pictures including two correct figures and a trained word form. In the SR condition, three choices that included one correct figure and a trained word form were presented simultaneously. Each stimulus was presented for 2000 ms. Participants pushed a key mapped to their right index, middle or ring fingers to respond. After each stimulus, a subsequent fixation cross was followed for 2000 ms. After the fixation cross, a correct answer was displayed for 2000 ms followed up by another fixation cross for 2000 ms (Figure 4). As in the retrieval tasks, each condition was presented as a block for each recognition task. The presentation order of each condition was counterbalanced among the participants. The order of the stimuli was randomized. For this procedure, the participants conducted an additional recognition task for the PW condition to control the number of stimulus presentations in the recognition tasks. In the additional recognition task for the PW condition, if the participants recognized (read) each word form in the PW condition, they pushed a key mapped to their right index finger. The duration of each stimulus depended on each response from each subject. Following each judgment, a fixation cross was followed for 2000 ms. The experimental flow is shown in Figure 5.

4.5. Analysis

We performed ANOVAs for the two factors of condition and time on the proportions and response times of correct trials for the recognition and retrieval tasks. We conducted these analyses using the Excel macro-based statistical software, HAD (https://norimune.net/had). The Greenhouse-Geisser correction was applied and Mauchly’s test of sphericity indicated a significant result (addressed in the Results section). To identify time-related differences within or between conditions, the Holm method was used for post-hoc multiple comparisons. We conducted paired t-tests to examine whether missed answers in MR, SR, and PW were judged as trained (e.g., SR and PW in the case of MR) or untrained words (e.g., UT across all conditions, including MR, SR, and PW) for each condition at each time (i.e., the number of retrieval tasks). As all the paired t-tests were applied to MR, SR, and PW at the first and second times, we corrected the alpha level of 0.05 by dividing it by 6 times of comparison per paired t-test. Finally, to account for potential relationships between both types of task performance, we conducted a Pearson’s correlation analysis on the accuracy rates of the verbal working memory task and of the SR and MR conditions in the recognition and retrieval tasks with SPSS. Two participants were excluded from the descriptive statistics of the verbal working memory task and the Pearson’s correlation analyses owing to the occurrence of program errors during the verbal working memory task.

5. Results

5.1. Recognition tasks

First, to examine the differences between the accuracy rates of the MR and SR conditions, we performed two-way ANOVA to analyze the conditions and times of recognition tasks as independent variables, and correct responses (accuracy rates) as dependent variables (Table 1). The results of the Greenhouse-Geisser correction were used to show whether Mauchly’s test of sphericity was statistically significant. The results showed that the main effect of conditions was significant (F(1, 29) = 14.91, p < .005, ${\eta }_p^2$ = .34). The main effect of time in the recognition tasks was also significant (Greenhouse-Geisser corrected: F(2, 58) = 35.03, p < .001, ${\eta }_p^2$ = .55). The interactions between conditions and times of recognition tasks was not significant (F(2, 58) = 0.40, n.s., ${\eta }_p^2$ = .01). For multiple comparisons between conditions, accuracy rates for the SR condition were significantly higher than those of the MR condition (t(29) = −3.86, p < .005, d = −0.36). As for multiple comparisons among a number of recognition tasks, accuracy rates for the second recognition tasks were significantly higher than those for the first ones (t(29) = −3.48, p < .005, d = −0.62), and for the third recognition tasks were significantly higher than those for the first (t(29) = −7.73, p < .001, d = −1.30) and second ones (t(29) = −6.70, p < .001, d = −0.51).

Table 1. Mean accuracy rates (%) and standard deviations in recognition tasks.

	MR		SR
	M	SD	M	SD
Time 1	52.89	11.57	56.64	12.74
Time 2	61.33	18.83	66.52	19.66
Time 3	69.85	18.60	76.45	18.04

Note. MR stands for multiple referents and SR for single referent. M refers to mean scores with SD standard deviations.

Second, to investigate the differences among the response times in the MR and SR conditions, we conducted two-way ANOVA to analyze the conditions and time of recognition tasks (the number of recognition tasks) as independent variables, and the response time as the dependent variable (Table 2). The results showed that the main effect of conditions was significant (F(1, 29) = 51.34, p < .001, ${\eta }_p^2$ = .64). The main effect of the number of times of the recognition task was also significant (F(2, 58) = 15.13, p < .001, ${\eta }_p^2$ = .34). The interactions between conditions and times of recognition tasks were significant (F(2, 58) = 6.10, p < .005, ${\eta }_p^2$ = .17).

Table 2. Mean response times (ms) and standard deviations in recognition tasks.

	MR		SR
	M	SD	M	SD
Time 1	3486.12	1227.87	2659.05	898.58
Time 2	2847.53	1347.14	2452.15	1067.31
Time 3	2772.46	1034.67	2182.52	717.51

Note. MR stands for multiple referents and SR for single referent. PW stands for new words and UT for untrained words. M refers to mean scores with SD standard deviations.

In the first recognition task, the response times in the SR condition were significantly faster than those in the MR condition (t(29) = 7.65, p < .001, d = 1.03), with Holm’s adjustments. In the second recognition task, response times in the SR condition were significantly faster than those in the MR condition (t(29) = 3.30, p < .005, d = 0.32). In the third recognition task, response times in the SR condition were significantly faster than those in the MR condition (t(29) = 5.74, p < .001, d = 0.65).

For multiple comparisons with the Holm method for time and recognition tasks in each condition, response times in the second recognition task were significantly faster than those in the first one in the MR condition (t(29) = 5.00, p < .001, d = 0.66), and response times in the third recognition task were significantly faster than those in the first one (t(29) = 4.66, p < .001, d = 0.62). There was no significant difference for response time between the second and third recognition tasks of the MR condition (t(29) = 0.47, n.s., d = 0.06). In the SR condition, response times for the third recognition task were significantly faster than those of the first one (t(29) = 4.29, p < .005, d = 0.58), and response times in the third recognition task were significantly faster than those in the second one (t(29) = 2.80, p < .05, d = 0.29). However, there was no significant difference between response times in the first and second recognition tasks in the SR condition (t(29) = 2.02, n.s., d = 0.21).

5.2. Retrieval tasks

Table 3 shows the results of two-way ANOVA for the accuracy rates of the MR, SR, PW, and UT conditions. The main effects of conditions were significant (Greenhouse-Geisser corrected: F(3, 87) = 13.64, p < .001, ${\eta }_p^2$ = .32), and the main effect of retrieval task time was significant (F(1, 29) = 46.57, p < .001, ${\eta }_p^2$ = .62). However, the interaction between conditions and times of retrieval tasks was not significant (F(3, 87) = 0.55, n.s., ${\eta }_p^2$ = .02). We found some differences among conditions from tests for multiple comparisons with the Holm method applied to the conditions for the retrieval tasks. First, the accuracy rates of the SR condition were significantly higher than those of the MR condition (t(29) = −2.77, p < .05, d = −0.68). Second, the accuracy rates of the PW condition were significantly higher than those of the MR condition (t(29) = −2.56, p < .05, d = −0.46). Third, the accuracy rates of the UT condition were significantly higher than those of the MR condition (t(29) = −6.28, p < .001, d = −1.26). Fourth, the accuracy rates of the UT condition were significantly higher than those of the SR (t(29) = −2.81, p < .05, d = −0.76) and PW (t(29) = −3.91, p < .005, d = −0.82) conditions. There was no significant difference between the accuracy rates of the SR and PW conditions (t(29) = 0.74, n.s., d = 0.14). The multiple comparisons with the Holm method for times of retrieval tasks showed that the accuracy rates of the second retrieval task were significantly higher than those of the first one (t(29) = −6.82, p < .001, d = −1.40).

Table 3. Mean accuracy rates (%) and standard deviations in retrieval tasks.

	MR		SR		PW		UT
	M	SD	M	SD	M	SD	M	SD
Time 1	22.03	10.72	31.00	10.83	26.35	13.04	43.11	20.64
Time 2	38.57	20.50	47.13	18.87	47.70	24.36	60.67	25.30

Note. MR stands for multiple referents and SR for single referent. PW stands for new words and UT for untrained words. M refers to mean scores, with SD standard deviations.

As for the results of two-way ANOVA for the response times in the MR, SR, PW, and UT conditions (Table 4), the main effect of conditions was not significant (Greenhouse-Geisser corrected: F(3, 87) = 0.41, n.s., ${\eta }_p^2$ = .01), but the main effect of times for retrieval tasks was significant (F(1, 29) = 11.36, p < .005, ${\eta }_p^2$ = .28). The interaction between conditions and number of retrieval tasks was not significant (Greenhouse-Geisser corrected: F(3, 87) = 1.04, n.s., ${\eta }_p^2$ = .04). From multiple comparisons with the Holm method applied to the number of retrieval tasks, response times in the second retrieval task were significantly faster than those in the first (t(29) = 3.37, p < .005, d = 0.56). For Holm-adjusted multiple comparisons of the times in each condition of the retrieval tasks, the response times in the second retrieval task were significantly faster than those in the first one in the SR and PW conditions (SR: t(29) = 2.21, p < .05, d = 0.41; PW: t(29) = 3.39, p < .005, d = 0.68). There was no significant difference between response times in the first and second retrieval tasks in the MR and UT conditions (MR condition: t(29) = 1.85, n.s., d = 0.42; UT condition: t(29) = 1.98, n.s., d = 0.43).

Table 4. Mean response times (ms) and standard deviations (SD) in retrieval tasks.

	MR		SR		PW		UT
	M	SD	M	SD	M	SD	M	SD
Time 1	2129.73	832.19	2291.54	812.17	2210.80	816.57	2388.18	2255.36
Time 2	1862.18	883.52	1954.06	803.72	1710.81	618.82	1634.63	884.94

Note. MR stands for multiple referents and SR for single referent. PW stands for new words and UT for untrained words. M refers to mean scores with SD standard deviations.

Finally, we analyzed the items that were missed in the MR, SR, and PW conditions to identify whether they were judged as trained items (e.g., a missed item of the MR condition was judged as an item in the SR or PW condition, and not as one in the UT condition) or an untrained item (e.g., a missed item of the MR condition was judged as an item in the UT condition, not as one in the SR or PW condition; see Table 5). We first calculated the trained and untrained rates of each condition (MR, SR, and PW) and performed paired t-tests on those rates. Considering that paired t-tests were conducted six times, we adjusted the p-value to p < .05/6 in order to avoid the inflation of the Type I error. Conditions in which all trials for each condition contained correct items (i.e., all the conditions did not involve any missed trials) were excluded as missing values. The paired t-tests for learned and unlearned rates of MR, SR, and PW conditions in the first retrieval task resulted in a learned rate that was significantly higher than the unlearned rate in the MR (t(29) = 3.53, p < .005, d = 1.30) and PW conditions (t(29) = 3.37, p < .005, d = 1.24), whereas no significant difference between the learned and unlearned rates was shown for the SR condition (t(29) = 2.23, n.s., d = 0.82). In the second retrieval task, the learned rate was significantly higher than the unlearned rate in the MR (t(29) = 7.08, p < .001, d = 2.60), SR (t(29) = 4.19, p < .001, d = 1.54) and PW conditions (t(29) = 3.67, p < .005, d = 1.35).

Table 5. Mean learned and unlearned rates (standard deviation) of miss items in retrieval tasks.

	MR		SR		PW
	Learned	Unlearned	Learned	Unlearned	Learned	Unlearned
Time 1	62.25	37.75	58.60	41.40	61.35	38.65
	(18.99)	(18.99)	(21.12)	(21.12)	(18.44)	(18.44)
Time 2	73.01	26.99	67.19	32.81	66.00	34.00
	(17.79)	(17.79)	(22.49)	(22.49)	(23.90)	(23.90)

Note. These rates show rates of miss as each learned or unlearned rates (standard deviations) wherein each number of learned or unlearned items is divided by each summation of the numbers of learned and unlearned items.

5.3. Correlations between the accuracy rates of the verbal working memory task and of the SR and MR conditions in the retrieval and recognition tasks

The mean accuracy rate (standard deviation) of the verbal working memory task was 90.65 (5.65). Pearson’s correlation analyses were conducted to detect potential relationships between the accuracy rates of the verbal working memory task and of the SR condition in the retrieval and recognition tasks (see Table 6). The results showed that the accuracy rates of the verbal working memory task were positively correlated with the SR condition in the first recognition task, whereas other correlations were not statistically significant. Another analysis of Pearson’s product-moment correlations for the same variables was conducted for the MR condition in the retrieval and recognition tasks (see Table 6). The results showed that the accuracy rates of the verbal working memory task were positively correlated with those of the MR condition in the first, second, and third recognition tasks, whereas the other correlations were not statistically significant. These findings are discussed in the following sections.

Table 6. Correlations between the accuracy rates of the verbal working memory (WM) task and SR and MR conditions in the recognition (Rec) and retrieval (Ret) tasks.

Trial (SR)	WM	Trial (MR)	WM
Rec 1	.42*	Rec 1	.44*
Rec 2	.33	Rec 2	.40*
Rec 3	.32	Rec 3	.44*
Ret 1	.35	Ret 1	.37
Ret 2	−.09	Ret 2	.24

Note. Asterisks indicate * p < .05.

6. Discussion

We investigated whether the associative pairs of pseudowords and SR are learned more efficiently than multiple referents (MR) in a single-day learning context. Two key findings emerged from the behavioral experiment. First, only the accuracies in the SR, MR, and pseudoword-only (PW) conditions increased, whereas response times decreased. Even though the participants miscategorized some items as trained ones, those items also increased from the first to the second retrieval tasks. Second, the accuracies and response times in the recognition tasks, and the accuracies in the retrieval tasks were better for the SR condition relative to the MR one, and there was no significant difference in response times between the SR and MR conditions in the retrieval tasks. The accuracy rates of the verbal working memory task positively correlated with the SR condition in the first recognition task and across recognition tasks in the MR condition. Taken together, these results show that associative learning for a pseudoword and SR contribute to efficient artificial language learning more than MR in the context of a single day of learning. Furthermore, the results show that the verbal working memory task contributes toward learning associations between new word forms and referents.

6.1. Effects on single-day learning

The task performance of all conditions (MR, SR, and PW conditions) learned by the participants increased in the retrieval and recognition tasks. The results are congruent with previous findings that encompassed a single day or more than two days (e.g., Breitenstein et al. 2005; Kambara et al. 2013; Lee et al. 2003). While previous studies used meaningful stimuli in an L1 or L2 (e.g., Lee et al. 2003), our study used meaningless stimuli as L1 stimuli (e.g., Kambara et al. 2013). Our findings suggest that associative learning for new L1 word forms and referents may be effective in single-day learning. The current results are therefore limited to single-day artificial learning in an L1-like context, but they may provide a baseline for future comparisons to L2 word-picture performance within that time span.

6.2. Associative learning of new L1 word forms and referents in a day

Associative learning for pseudowords and SR was more efficient than MR in our experiment that took place in a single day. These results appear consistent with findings relevant to accuracy costs to item and source memories in divided attention (e.g., Troyer et al. 1999) and findings relevant to item and associative learning (Stark, Bayley, and Squire 2002; Turriziani et al. 2002). For single pseudoword-referent associative pairs, participants may focus only on the two stimuli as sources of information. However, participants may have split their focus among the three stimuli for MR. In favor of theoretical framework integration, this may represent the extraneous load described in cognitive load theory (Paas and van Merriënboer 2020), or some property of the theory of competition in artificial learning (Buccola, Dautriche, and Chemla 2018; see Limitations and Directions for Future Research). Participants can more easily recognize single items over multiple ones (Stark, Bayley, and Squire 2002; Turriziani et al. 2002). The accuracy costs to source memories would also be greater than those related to items in a state of divided attention (Troyer et al. 1999). Therefore, the attentional load(s) distributed among MR may affect the associative learning of word-picture pairs, and therefore the recognition and retrieval of word forms learned in a single day.

The findings of this study were consistent with previous findings for homonym and non-homonym learning (Doherty 2004; Ellis 2008), and our hypothesized link for the ambiguous effects of words (Hino, Lupker, and Pexman 2002). Although the number of semantic features was generally found to promote the word recognition of acquired words (e.g., Pexman et al. 2007; 2008; Yap et al. 2012), response times were faster for unambiguous words than ambiguous words including homonyms in a semantic categorization task (Hino, Lupker, and Pexman 2002). Familiarity effects may have been involved in those experiments owing to the influence of word stimuli that may have been acquired before experimentation. This issue occurred for recognized words whose meaning was imagined easily when compared to words that were read (e.g., covertly pronounced; Davachi, Mitchell, and Wagner 2003). Witherby and Carpenter (2022) suggested that curiosity may mediate relationships between prior knowledge and new learning of domain-relevant knowledge, which would indicate potential influences of familiarity effects on learning. As we used pseudowords as non-words and pictorial referents as meaningless figures, we offer this difference in the chosen stimuli for the consistency with prior studies. However, some property of our stimuli may be implicated in the single-day task performance of recognition and retrieval tasks (Limitations and Directions for Future Research).

The findings of correlations between the accuracy rates of the verbal working memory task and those of the recognition tasks suggest that verbal working memory would be essential for learning both pseudowords (meaningless words) and words (meaningful words). Gathercole (2006) suggested that pseudoword repetition and word learning depend on the phonological storage of new word forms (pseudowords). As new word forms (pseudowords) are not associated with referents (meanings), participants do not remember these words for semantic knowledge in the current and previous studies (e.g., Gathercole and Baddeley 1989, 1990; Service and Craik 1993). Baddeley, Papagno, and Vallar (1988) suggested that the verbal working memory, or short-term phonological storage, is essential for learning unfamiliar (new or meaningless) words and artificial (meaningful) words. The current findings are consistent with previous ones for word learning and the verbal working memory task performance (Gathercole 2006; Gathercole and Baddeley 1989, 1990; Service and Craik 1993). Our findings do not show significant correlations between the accuracy rates of the verbal working memory task and those of the retrieval task. In the latter, the participants judged the conditions that the displayed words belonged to without checking referent(s) directly. In the recognition task, the participants judged the kind of figures that the displayed words were associated with by checking referent(s) directly. The current findings may be associated with the characteristics of the retrieval and recognition tasks. It may be the case that if participants can check the referent(s) associated with the learned words in the test phase directly, verbal working memory may contribute to task performance of associative learning. Studies must clarify these contributions to behavioral performance.

We interpreted our findings with a multimodal approach to meaning and with special attention to the properties of pseudowords (logogens) and referent presentation (imagens; Paivio 1986, 2007) as types of stimuli that could reflect sensorimotor interaction in ways that may have use frequency preferences for homonyms or non-homonyms. However, it is possible that our hypothesized link between word types and word learning efficiency may overstate the relationship between word learning and the mechanism of semantic density (or richness). This is especially the case for word learning in children, who may acquire non-homonyms more readily and whose outcomes may be affected by lexicality effects from the partial presentation of word forms (e.g., Cychosz et al. 2021), some relative familiarity of novel sound patterns from the possible developmental role of vocal production as a mediator to phonological memory (e.g., Vihman 2022), or language-specific influences on nonword repetition (Eikerling, Bloder, and Lorusso 2022). While we used adult participants and pseudowords validated by previous studies as possessing low familiarity, we acknowledge that there are nonetheless important factors such as age and properties of the stimuli that will require more empirical research for our context (see Limitations and Directions for Future Research). The UT condition of novel word forms in our study showed unexpected improvements in retrieval over time. We interpret this finding to mean that participants perceiving the untrained word forms were novel, similar to a previous study with a comparable design by Liu and colleagues (Liu et al. 2021). On the surface, this claim seems supported by the results of a study that manipulated visual repetition with and without semantic-associative contexts evaluated with brain signatures and drew findings in favor of a gradual, lexico-semantic-based processing of pseudowords (Bermúdez-Margaretto et al. 2018). Further studies may corroborate or integrate the findings of those brain signatures with the rapid microstructural changes reported by Vukovic and colleagues (Vukovic et al. 2021) for new memory circuits from short-term language learning; however, this is merely our conjectural synthesis. We recognize that our UT finding underscores the need to explore pseudoword properties and their evaluation in associative learning contexts further, as detailed in the next section.

6.3. Limitations and directions for future research

Our study has some limitations that are worth noting. First, this study only conducted tests for associative learning in a single-day experimental setting. Future experiments must identify the learning and decay effects of associative memories of word forms and multiple references over a period longer than two days (e.g., training periods in Breitenstein et al. 2007). For example, one study reported that people remember and retain new meanings of familiar words even after a week (Hulme, Barsky, and Rodd 2019). Second, this study only focused on associative learning of word forms in L1 (Japanese) and referents among Japanese participants. Therefore, future studies should investigate associative learning of word forms in a second language and referents in samples of participants with various background characteristics (e.g., pseudowords from languages with other kinds of orthographic systems), to account for factors such as age (e.g., for children, Tamura, Castles, and Nation 2017; for adults, Tamminen and Gareth Gaskell 2006), lexical competition effects (e.g., Benitez, Yurovsky, and Smith 2016), some property of pseudoword learning or pseudowords themselves (e.g., Bermúdez-Margaretto et al. 2018; Buccola, Dautriche, and Chemla 2018), bilingual or multilingual status and related nonword sensitivity differences (e.g., Eikerling, Bloder, and Lorusso 2022), or learner profiles (e.g., language learning motivation or proficiency variables, Dörnyei and Ryan 2015). Third, researchers may need to consider phonologically and orthographically symbolic effects in new words (e.g., Ando et al. 2021; Cuskley, Simner, and Kirby 2017; Kambara and Umemura 2021; Lin et al. 2021). Symbolic effects promote L1 and L2 word learning (Imai et al. 2008; Imai et al. 2015; Kantartzis, Imai, and Kita 2011). Fourth, psycholinguistic researchers need to consider developmental aspects. For example, children are more likely to find homonyms as they get older (Peters and Zaidel 1980). Additionally, student writing contains more lexical ambiguity than syntactic ambiguity (Demir 2020). Fifth, researchers should also study the aspects of homonyms with tools and approaches used in cognitive neuroscience as the lexical ambiguity of homonyms has been shown to affect brain activity in the left hemisphere (e.g., left inferior frontal gyrus; Rodd 2018). Finally, although participants simultaneously learned a word and two figures as the word referents in the MR condition, it might be the case that we do not encounter a word and two referents simultaneously in real life situations. Future researchers might account for this by improving the presentation order of the referents in the learning task. We believe that these findings provide opportunities for future research that experimentally manipulate sources of language stimuli as they relate to associative memory processes.

6.4. Implications for practical and applied research settings

Given the evidence of a possible learning efficiency that we observed for associative pairs with SR, researchers may need to consider the number of referents in the experiments for productive vocabulary acquisition and assessment. For example, our findings may have implications for experimental design choices like stimulus presentations of nonword repetition in phonological short-term memory tasks used in research to validate oral proficiency measures (e.g., Park et al. 2020). Our findings suggest that multimedia learning and technology-assisted L2 vocabulary learning contexts may consider manipulating the number of referents in their analysis of moderators (Yu and Trainin 2022) for efficiency, although future practical and applied research should draw conclusions on their effects on learning.

7. Conclusion

We investigated whether participants learn associative pairs of pseudowords derived from L1 Japanese and SR better than MR in a single-day learning context. The findings showed that only the accuracies for the SR, MR, and pseudoword-only (PW) conditions increased, whereas their response times decreased. Missed items that participants categorized as learned items increased from the first to the second retrieval tasks. Second, the accuracies and response times in the recognition tasks, and the accuracies in the retrieval tasks were better for the SR condition than the MR one, and there was no significant difference in response times between the SR and MR conditions in the retrieval tasks. The accuracy rates of the verbal working memory task correlated positively with the accuracy rates of the SR condition in the first recognition task and across all recognition tasks in the MR condition. These results suggest that associative learning of a pseudoword and a SR, rather than MR, contributes to efficient artificial language learning in a single-day learning context involving university-age Japanese adults.

Supplemental data and research materials

Supplemental data for this article can be accessed online at https://doi.org/10.1080/00437956.2023.2299070.

Acknowledgments

We are grateful to the editor and reviewers for their careful review and editorial handling of our paper. We would also like to thank the faculty members and students from the Department of Psychology at Hiroshima University, for their important suggestions and support.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The corresponding author was provided the KAKENHI Grants-in-Aid for Scientific Research (B). This study was also conducted as part of the School of Education Joint Research Project 2020 to 2023 at Hiroshima University and received research support from the School of Education.