Despite progress in second language (L2) instruction, vocabulary learning remains a task that people still accomplish by reading bilingual word lists and by repeating them until the words are memorized, as described by Oxford and Crookall (1990). One way to make learning faster and more resistant to decay has proven to be enrichment: novel words are accompanied by additional sensory information such as pictures or pantomimes. This is the principle on which mnemonic strategies have been based for centuries (Riesenberg et al., 2009). Shams and Seitz (2008) suggest that our brain is tuned to learn and to operate in multisensory environments. Therefore, multisensory learning can better simulate natural settings and lead to superior memory performance compared to unimodal learning (reading or listening). Better retrievability of enriched words can be accounted for by the Connectivity Model (Klimesch, 1987). It proposes that a concept is represented as a network of interconnected nodes – the features. The model assumes that the more the nodes are interconnected, the faster the information is processed and retrieved within the network. If a word in L2 (i.e., a string of phonemes) is enriched by a picture or other sensory information, the network grows larger. Complexity enhances information processing, and consequently memory and speed of retrieval (Klimesch, 1994).

In the free recall Italian test, we found a main effect of encoding condition, both by subjects and by items [F1(2,38) = 22.89; p < 0.01; η²= 0.54; F2(2,27) = 19.49; p < 0.01; η²= 0.59]. Specifically, items trained in the GE condition were better remembered than those trained in the PE and VE conditions in both by subjects and by items analyses (p < 0.01). However, the performance in the PE and VE conditions did not differ [by participants, p = 0.54; by items, p = 0.38]. In the free recall Vimmi test, the learning condition did not affect memory performance [F1(2,38) = 2.38; p = 0.8; F2(2,27) = 0.52; p = 0.6].

In the cued recall tests, the data evidenced similar patterns to those found in the free recall tests. In the Vimmi to Italian test, the encoding condition proved to have an impact on the number of recalled items only in the by subjects analysis [F1(2,38) = 5.47; p < 0.01; η²= 0.22; F2(2,27) = 1.91; p = 0.17]. Post hoc tests in the by subjects analysis revealed that the GE condition was more effective than PE (p = 0.03), whereas PE and VE did not differ (p = 1). The difference between GE and PE was almost significant (p = 0.05). The Italian to Vimmi test showed similar performance in all of the encoding conditions [F1(2,38) = 1.88; p = 0.17; F2(2,27) = 0.77; p = 0.47].

We checked the statistical power because of the modest sample size in the present study (N = 20). A post hoc power analysis conducted with the program G^∗Power (Faul et al., 2007) revealed that on the basis of the mean, i.e., the within-groups comparison, the effect size observed for the Vimmi to Italian cued recall test (=0.22) and our sample size (N = 20) reached statistical power at the recommended level, i.e., 0.80 (Cohen, 1988).

In the recognition test, RTs were affected by the encoding condition in the by subjects but not in the by items analysis [F1(2,38) = 3.8; p = 0.03; η²= 0.17; F2(2,27) = 2.09; p = 0.14]. Post hoc tests in the by subjects analysis revealed that participants were slower in recognizing the correct Italian translation if words had been encoded in PE than in the VE condition (p < 0.01). The RT of responses did not show significant differences between the PE and GE conditions (p = 0.27), nor between GE and VE (p = 1). The analyses of the error pattern revealed that there was a main effect of Encoding Condition [F1(2,38) = 3.96; p = 0.03; η²= 0.17; F2(2,27) = 5.6; p < 0.01; η²= 0.29]. Participants made less errors in recognizing words trained in the GE than in PE the condition (by subjects, p = 0.03; by items, p < 0.01), whereas the number of incorrect responses did not differ between the PE and VE conditions (by subjects, p = 0.25; by items, p = 0.09), nor between GE and VE (p = 1).

In the present study, we were interested in understanding whether enrichment through pictures and gestures impacts memory for abstract words. The choice to use this class of words was motivated by the desire to keep sensory-motor information in the representation of the L1 word low in order to observe the potential EE free from influences that might derive from the L1 word semantics. In our within-subjects learning protocol, we asked young adults to memorize 30 foreign abstract words. The words were trained according to three conditions: verbally (VE), by looking at a cartoon representing the word’s semantics (PE), by watching an actress performing a gesture and thereafter by performing the gesture themselves (GE). The analyses conducted both by subjects and by items did not always match across memory outcomes, indicating that the set of stimuli selected might have affected performances. Indeed, in the free recall Italian test did the results show better memory performance for abstract words memorized in association with a gesture, in both the analysis by subjects and by items; in the Cued recall Vimmi to Italian test the beneficial impact of gestures emerged only in the analysis by subjects, indicating that we cannot generalize the results to other sets of words. In the other memory tests, the performance did not differ based on the training condition. In the reaction time test, our hypothesis that gestures would support memory better than pictures was confirmed only partially. In the recognition test, RT values were higher for words encoded with picture enrichment compared to the verbal encoding alone, but again only in the analysis by subjects; however, both the analyses confirmed that accuracy was higher for words learned with gestures than for words encoded with pictures. Also in this case, the results only partially matched our hypothesis. Considering the potential random effect due to the items selection, we will only consider reliable the results that evidenced convergent effects in both the analyses.

Comparing this study with similar studies, i.e., conducted with the same corpus (Vimmi) and stimuli (audio, video files, pictures), population (young adults), the only striking difference resides in training time. It lasted 35 min and can be considered as short in comparison with similar studies. In studies with a robust EE for words (Macedonia and Knösche, 2011; Macedonia et al., 2011) subjects learned much more intensively, i.e., approximately 3 h daily for 5 days. In another combined experiment (behavioral and fMRI) with pictorial and gestural enrichment, gestures proved superior compared to pictures and the audiovisual baseline. This was the case for concrete (Mayer et al., 2015) and for abstract words (Mayer et al., 2017). In the last study, the training lasted about 12 h (subdivided in 4 days, with 3 h daily). Those studies included more items, however. These results seem to indicate that enrichment of a word in L2 with gestures needs longer and repeated training sessions in order to show EE and POE. We speculate that this might involve processes of neuroplasticity affecting procedural learning. As suggested by Macedonia and Mueller (2016), learning novel words with gestures engages procedural memory in addition to declarative memory. In that study, the authors could show that reading and hearing words that had been learned with gestures activate manifold structures of the motor system in the brain. Besides the motor cortices, the basal ganglia and the cerebellum – normally related to motion planning, sequencing and execution – participated in word recognition.

Traditionally, vocabulary learning is situated in the domain of declarative memory (Ullman et al., 1997; Squire and Dede, 2015). However, it is well-demonstrated that word representation in L1 – being experience-related – engages motor structures of the brain (Pulvermüller et al., 2005a,b) at least for those words that in their semantics include some kind of action. Thus we reason that the representation of novel items in L2, because of how they are learned, could also be stored in procedural memory. In other words, considering that the representation of a concept and its label (word) in the brain is an experience-related network, if the network has been created by engaging motor regions, the storage of the concept (word) must have to do with procedural (implicit) memory (Reber, 2013). In Doyon’s two-component model of motor learning (Doyon, 2008), we find immediate and slow motor skill learning. Slow motor learning needs consolidation, i.e., dynamic reorganization in motor networks including – besides the motor cortices – the basal ganglia and the cerebellum (Ungerleider et al., 2002). It could be the case that L2 word acquisition with gestures engages slow motor learning. Similarly, it is conceivable that neurobiological processes driving plasticity need “frequent” stimulus repetition in time in order to produce measurable behavioral effects. As described by Reber (2013), experience-driven changes in the adult sensory cortex occur “slowly.” An extremely changeable sensory cortex that continuously adapts and reorganizes by experience would be instable and lose abilities that have been acquired but that are not required often. According to Reber, practice would accrue implicit learning effects slowly – in our case EE – and make them inflexible. These apparently “negative” characteristics of the procedural system, i.e., slow speed of learning and inflexibility could be the key to better memory performance for L2 words learned with gestures. In fact, studies with massed training over 4 or 5 days show that significant better performance for gesture learning starts from day 2 to day 3 (Macedonia and Knösche, 2011; Macedonia et al., 2011; Mayer et al., 2015) and follow up tests reveal significantly better retrieval over time, i.e., slower information decay, compared to all other experimental conditions (Macedonia and Klimesch, 2014). Thus, procedural learning takes long(er) but information retrieval is superior in speed and resistance to decay.

A study with 30 Vimmi words and three repetitions per word (Macedonia et al., under revision) yielded poor behavioral results comparable to the present study. Participants lying in the fMRI scanner were asked to memorize the Vimmi items by reading, reading and hearing, or by reading, hearing and observing semantically related gestures. Words learned by observing the gestures were significantly better memorized than those learned in the other conditions, however, in the free recall in L1 only. Free recall in Vimmi and cued recalls did not show better performance for gesture observation training. Although participants were lying quietly in the scanner, the brain imaging revealed simulation activity in motor regions triggered by the gesture observation. Still, behavioral performance was low. We attribute this again to the low number of repetitions (3) and to slow motor learning.

Despite the “poor” training in the present study, gestural training enabled participants to perform the “easy” task, i.e., free recall in L1, better than the other training conditions. It is not new that enactment substantially improves performance on conceptually driven explicit-memory tests (Nyberg and Nilsson, 1995), i.e., free recall. We reason that during L2 word learning, the concept is stored first. Thereafter the label to the concept follows as documented in infancy for L1 (Yin and Csibra, 2015). In free recall, when learners are asked to retrieve the words in L2, it is possible that retrieval starts at the conceptual level. Learners know the concepts and label it in L1. However, naming the concepts in L2 requires retrievability of the L2 label. This is only possible after adequate learning and connected consolidation processes. In contrast to infants (Yin and Csibra, 2015) and depending on the age of the learner (Schatz et al., 2010), this process might need more stimulation than participants were provided in the present study. Hence, when learners retrieve the concept’s label, the word in L1 is an easy task compared to retrieving the word in L2. Results showing that retrieval in L1 comes before retrieval in L2 have been previously reported (Macedonia et al., 2011). They have been attributed to the duration of the training in relation to the difficulty of the task. These factors can modulate the effects of gesture-based training as previously addressed in a study by Macedonia et al. (2010) that investigated the impact of gestures on foreign word learning in high and low performers. The former took advantage of gestures only in the more difficult tasks (free recall in L2 and cued recall from L1 into L2), whereas the latter also did so in the easier ones as in the present study. In a similar study yielding poor behavioral results, Krönke et al. (2013) reasoned that gestures can promote learning only when the task can be considered difficult related to the individual’s learning abilities. In the reverse situation, when the task is “easy,” learners can use other cognitive strategies that mask the effect of gestures. Thus, it stands to reason that the impact of gestures on learning is a complex mechanism that emerges from the intersection of various variables, including the difficulty of the task and the duration of the training.

In the recognition task, we had hypothesized short reaction times for GE items as described in the literature (Spranger et al., 2008). Unexpectedly, we found no differences in the RTs among conditions. This result does not match the POE (Zimmer et al., 2000), in which lower reaction times would be expected. Though there might be an explanation for our RT results. Considering the short training and that the recognition test took place after a short break, it might also be the case that learning circuits were not consolidated. If we assume that by the participant performing a gesture while learning a novel word, the information is stored in both declarative and procedural memory, as advanced in Macedonia and Mueller (2016), it stands to reason that the sensorimotor networks need time to consolidate. To this process sleep would also contribute in order to reach full functional efficiency (Brashers-Krug et al., 1996; Stickgold, 2005). Furthermore, we expected that words encoded with gestures would be recognized more accurately than in the other conditions. Actually, participants made less recognition errors with words encoded with gestures than with words encoded with pictures. However, accuracy in recognition did not differ between words encoded with gestures and words encoded with no enrichment.

We hypothesized that besides gestures also pictures (PE) would have a positive impact on memory compared to pure verbal encoding (VE). However, our data did not provide evidence for this hypothesis, and consequently for the Dual Code Theory (Paivio and Csapo, 1969) and the LOP Framework (Craik and Lockhart, 1972). However, we do not question these theories. We speculate that in our study better retrieval effects might have occurred after more intensive training with a higher number of picture presentations.

Taken together, the present findings partially support our prediction: enacted gestures, compared to pictorial enrichment, are critical to enhancing learning. This perspective is coherent with Engelkamp’s approach (Engelkamp, 2001), which assigns the major role to a motor trace derived from the self-performed action. Our position agrees with approaches of embodied cognition (Glenberg, 2007; Barsalou, 2008). We deduce that even if abstract words are represented in the brain with less visual-sensory-motor features than concrete words, superimposing a motor component related to word meaning promotes learning. It is noteworthy that this is not merely a matter of adding information. Even motor enrichment per se is not as effective as gesture (see Macedonia et al., 2011; Mayer et al., 2017). This indicates that, in our cognitive system, gesture holds a specific privileged status compared to sensory and simple motor modalities. Furthermore, gestures, in addition to intrinsic motor features, are not independent of manifold perceptual and sensory features even if they represent abstract concepts (Vigliocco et al., 2014). During encoding, gestures enhance verbal memory performance because they also make sensory information connected to the concept available by simulation.

A limitation to this study is that the experimental protocol did not allow us to collect data about consolidated memories and about the decay of memories over time. We cannot provide information on whether gesture-based training is capable of making memory traces more stable than enrichment by pictures. Future research could address this issue by setting up multiple-day training with consolidation phases and successive follow-up tests.

Considering that this research might flow into educational practice, learning of vocabulary items by means of pictures and gestures is recommended particularly in low access-level instruction. Even a few repetitions of enriched L2 words lead to better memory performance and accuracy. However, massed and highly frequent practice by means of enrichment could be the key to exploiting EE and POE and leading to superior memory performance, faster recognition, and boosted retrieval.

CR and MM conceived the study. CR and EP set up the experimental protocol, collected and analyzed the data. CR and MM wrote the paper.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.