Modeling Form and Meaning Connections.

The process of building connections between form and meaning representations underpins lexical development, and describing this pattern of interconnections helps explain L2 vocabulary learning. Several models have described connections between L1 and L2 forms and meanings, such as the RHM, which merged the Word Association and Concept Mediation Models proposed by Potter, So, von Eckardt, and Feldman (Reference Potter, So, Eckardt and Feldman1984). The RHM was developed to account for the observation that during initial learning, L2-L1 translation is lexically mediated whereas L1-L2 translation is conceptually mediated, and further that L1-L2 translation is slower and more error prone than L2-L1 translation. This model described a pattern of interconnections that vary in direction and strength – whereas L1 forms are connected directly and strongly to corresponding meanings, L2 forms are (at least initially) only weakly connected to their corresponding meanings. Similarly, L1 words are weakly connected to L2 words, but L2 words are strongly connected to L1 words, and this asymmetry explains why L2-L1 translation is faster than L1-L2 translation (see Figure 1). Although recently some researchers have argued that the RHM should be replaced with newer connectionist models of bilingual word recognition (e.g., the BIA+ model of Dijkstra & van Heuven, Reference Dijkstra and van Heuven2002; see Kroll et al., Reference Kroll, van Hell, Tokowicz and Green2010), the present review contends that an adapted version of the RHM, the RHM-Repetition Elaboration Retrieval (RHM-RER; see Figure 2), is best suited to describe adult L2 vocabulary learning during explicit instruction in the laboratory.

FIGURE 1. The Revised Hierarchical Model (RHM; Kroll & Stewart, 1994)

FIGURE 2. The Revised Hierarchical Model – Repetition, Elaboration, Retrieval (RHM-RER)

The RHM-RER Model

This review will use the framework provided by the RHM in characterizing the pattern of interconnections between L1 and L2 form and meaning representations, but will place new emphasis on the processes by which these connections are established and strengthened. This reconceptualization allows us to understand contradictory patterns of results reported throughout the adult L2 vocabulary learning literature by bringing the process of creating and strengthening connections between representations to the forefront.

To this end, we propose a three-tiered model that describes a pattern of connections between L1 and L2 forms and meanings, whose strength changes with increased proficiency. The first tier describes how L1 and L2 lexical representations are conceptualized and interconnected, the second tier describes how semantic representations are conceptualized and interconnected, and the third tier describes the mechanistic procedures by which connections between and within these tiers are established and strengthened.

In the top tier, as in the RHM, L2 lexical representations are strongly connected to L1 lexical representations, whereas L1 lexical representations are weakly connected to L2 lexical representations. This asymmetry reflects that L2 forms are connected to L1 words before they are connected to meanings, whereas L1 words are connected to meanings before they are connected to L2 forms (Kroll & Stewart, Reference Kroll and Stewart1994). In the top and middle tiers, in contrast to the RHM, we conceptualize both form and meaning representations as a set of distributed features (e.g., de Groot, Reference de Groot1992; Laxén & Lavaur, Reference Laxén and Lavaur2010; van Hell & de Groot, Reference van Hell and de Groot1998). This enables us to demonstrate that neither lexical forms nor meanings completely overlap across languages. We depict orthography and phonology as separate sets of distributed representations because words may differ in how much between-language overlap there is in the phonological and orthographic levels. In the bottom tier, we expand the RHM to include the concepts of repetition, elaboration, and retrieval (RER), which we view as strengthening mechanisms for the associative connections between and within the first and second tiers. Throughout the review we demonstrate that better learning results from the use of more of these strengthening mechanisms as well as making connections across tiers rather than only within a tier. Through repeated coactivation of forms and meanings, both within and across tiers, learners encode novel words, and establish connections between forms and meanings. Retrieving a form or meaning from memory and engaging in elaborative processing strengthens the connections. At the end of each major section of the review we return to an RHM-RER guided conceptualization of L2 vocabulary learning to help understand the results discussed within that section.

Repetition Training without Retrieval Practice

In this section, we discuss studies that use repetition training without additional strategies such as retrieval practice or spacing (discussed in the next sections). A handful of studies contrast this type of repetition training with other training methods, and the majority report weak evidence for the effectiveness of repetition training (see Pressley et al., Reference Pressley, Levin and Delaney1982, for a review; also see Lawson & Hogben, Reference Lawson and Hogben1998; Moore & Surber, Reference Moore and Surber1992; and Sagarra & Alba, Reference Sagarra and Alba2006, for experimental reports).

Lawson and Hogben (Reference Lawson and Hogben1998) investigated the effectiveness of repetition training in a study in which they instructed English-speaking students to learn English-Italian word pairs by repeating a list (with no opportunities for retrieval practice) or using a variation of the keyword method. They reported that repetition training was significantly less effective than the keyword method on immediate and delayed L2-L1 written translation production tests. Furthermore, our examination of the original data revealed that on an immediate posttest, participants trained with repetition remembered on average fewer than 8 of the 32 nouns to which they were exposed, and by 10 days after the initial training this number had fallen to just more than 5 of the 32 nouns. Concrete words fared better than abstract words – on the immediate posttest participants in the repetition condition remembered an average of 5 of the 16 concrete words but fewer than 3 of the 16 abstract words, and similarly on the 10-day delayed posttest participants remembered an average of 4 of the 16 concrete words, but just more than 1 of the 16 abstract words. These results describe a case in which repetition training was relatively ineffective for both short-term learning and long-term retention of novel L2 vocabulary.

Similarly to Lawson and Hogben (Reference Lawson and Hogben1998), Sagarra and Alba (Reference Sagarra and Alba2006) trained low-to-intermediate proficiency L1 English-L2 Spanish learners on low-frequency Spanish nouns and compared the efficacy of repetition training, semantic mapping (i.e., diagramming relationships of an L2 target word and its L1 semantic associates), and the keyword method using a within-subjects design. In the repetition condition, participants silently and continuously read and copied words for 1 minute per pair, and were instructed not to think about links between target words and translations. In the keyword condition, participants saw an English-Spanish word pair, and were instructed to think of an additional English word that shared orthographic, phonological, or semantic overlap with the Spanish word, and to write down this keyword in their study booklet. Following training, participants completed immediate and 3-week delayed posttests, in which they were given L2 words and asked to match them to corresponding pictures. On both the immediate and 3-week delayed posttests participants recalled significantly more words when trained with the keyword method (93%; 70%) than words trained with repetition (74%; 48%) or than words trained with semantic mapping (49%; 19%).

How can we understand the results of this study in the context of the RHM-RER? The most successful learning was observed with the keyword condition. This condition required generating and describing keyword-target connections, which promoted meaning activation and strengthened connections between L2 forms and meanings. In the context of the RHM-RER, this condition activated both the form and meaning levels, as well as connections between the levels. The repetition condition resulted in less successful learning than the keyword method. This condition, which consisted of reading and copying words, favored the activation of orthographic forms over meanings, thereby strengthening the connection between L1 and L2 forms. In the context of the RHM-RER, this condition activated only the form level of the model. Finally, the semantic mapping condition was unsuccessful. This is possibly because this method required learners to generate and map L1 semantic associates of the L2 word. In the context of the RHM-RER, instead of strengthening connections between the L2 forms and meanings, this condition may have simply strengthened connections between the L1 and L2 forms, or even strengthened connections between the L2 form and semantically related L1 forms that would have been extraneous for the testing task. The transfer-appropriate processing model (Morris, Bransford, & Franks, Reference Morris, Bransford and Franks1977), provides a compelling explanation for why this particular type of semantic elaboration was ineffective. In brief, the transfer-appropriate processing model states that performance is strongest when training and testing tasks are congruent. In Sagarra and Alba’s (2006) semantic mapping condition, the training and testing tasks are incongruent – whereas the training task involved generating L1 semantic associates of the L2 word, the testing task required matching L2 words to pictures. This incongruency may explain the relative ineffectiveness of the semantic mapping condition.

Learner proficiency may also impact the effectiveness of repetition training. Moore and Surber (Reference Moore and Surber1992) trained L1 English speakers enrolled in first-, second-, or third-year German language instruction on low-frequency German words, using three training conditions: repetition, the keyword method, or learning from context sentences. Participants were tested on L2-L1 translation production immediately after training and after a 3-week delay. A 3 training condition (repetition, keyword, context sentences) × 3 year (first, second, third) × 2 time (immediate, delayed) ANOVA revealed no significant three-way interaction. However, the authors broke down the immediate testing data by level and condition, which revealed that whereas participants in their first or second year of German instruction who were assigned to the repetition condition recalled fewer correct translations (first = 67%, second = 62%) than participants in the keyword (76%, 63%) and context sentences conditions (74%, 76%), participants in their third year of German performed better in the repetition condition (82%) than the keyword (71%) or context sentences (65%) conditions. A similar pattern was found for the delayed posttest; participants in their first or second year of German instruction who were assigned to the repetition condition recalled fewer correct translations (first = 30%, second = 35%) than participants in the keyword (35%, 41%) and context sentences conditions (36%, 48%), but participants in their third year performed better in the repetition condition (48%) than participants in the keyword condition (37%) and approximately equivalently to participants in the context sentences condition (47%).

Similarly, van Hell and Candia Mahn (Reference van Hell and Candia Mahn1997) compared repetition training and the keyword method with experienced and inexperienced L2 learners. In Experiment 1, they recruited Dutch speakers with significant foreign-language experience in French, English, and German, but no experience with Spanish. They taught the participants Spanish-Dutch word pairs with either repetition training or the keyword method. Experiment 2 followed the same procedures as Experiment 1, but participants were native English speakers with no L2 learning experience. Experienced learners recalled target words more quickly and accurately in the repetition condition than the keyword condition, whereas inexperienced learners showed no difference between the two instructional methods.

One possible explanation for these findings is that, as discussed in the preceding text, repetition training encourages a strong connection between L1 and L2 forms, but does not encourage any connection between L2 forms and meanings, whereas learning from context sentences and the keyword method do encourage the learner to engage with meaning. As outlined by the RHM, L2 form-meaning mappings are particularly weak at low levels of proficiency. Therefore, in this study, learners in lower levels seem to have benefitted more from training methods that allowed them to focus on connecting novel L2 words to meanings, whereas the advanced learners, who may already be more adept at connecting L2 forms to meanings, may have benefitted more from the simpler procedure of repetition training.

In addition to learner characteristics, a number of word characteristics may also impact the efficacy of repetition training. Generally speaking, when words are easy to learn, repetition training may be sufficient to promote memory, but when words are difficult to learn repetition training may fail. For example, Bartolotti and Marian (Reference Bartolotti and Marian2014) and Bartolotti and Marian (Reference Bartolotti and Marian2017) demonstrated that wordlikeness (i.e., the orthographic and orthotactic similarity of novel words to L1 words) plays an important role in novel word learning. Novel word learning is more successful when there is more overlap between the word characteristics of novel words and L1 words. There is also evidence that cognate status and word concreteness affect novel word learning in similar ways as wordlikeness. For example, de Groot and Keijzer (Reference de Groot and Keijzer2000) trained native Dutch speakers on artificial words paired with Dutch words that varied in word frequency, concreteness, and cognate status. Learners attempted to memorize word pairs through repetition and were tested on L1-L2 and L2-L1 translation production. De Groot and Keijzer reported that cognates and words that were higher in concreteness were easier to learn than noncognates and words that were lower in concreteness, although word frequency did not have a substantial effect on learning. Similarly, Lotto and de Groot (Reference Lotto and de Groot1998) trained adult Dutch speakers on L2 Italian words paired with either Dutch translations or pictures. During training, participants encountered stimulus pairs three times, and were given an immediate production task in which they saw either an L1 word or a picture (depending on the condition to which they had been assigned), and were asked to produce the L2 word. Participants returned the next day, and completed the same test a second time. Results showed that generally cognates and high-frequency words were easier to learn than noncognates and low-frequency words. Finally, there is also evidence that phonotactic word characteristics impact novel word learning in important ways. Although it is outside of the scope of the present work to cover this issue in depth, the interested reader may see Thorn and Frankish (Reference Thorn and Frankish2005) for a discussion of the independent effects of biphone frequency and neighborhood size on novel word learning.

The RHM-RER represents the influence of lexical properties of words such as those described previously by including sets of distributed orthographic and phonological features in the top tier of the model (see Figure 2). This enables us to represent graded overlap between L1 and L2, as in the case of cognates and more word-like novel words. For example, when an L2 word is a cognate of an L1 word, the RHM-RER represents this with increased overlapping coactivation of L1 and L2 orthographic and phonological units. Thus, the RHM-RER views overlap in L1 and L2 word forms as another way of strengthening form representations and form-form connections.

In summary, repetition training can be an effective learning method for more-proficient learners and easier-to-learn words (i.e., cognates, high-frequency words, and when novel words appear word-like to the learner). However, for less-proficient learners and for words that are more difficult to learn, repetition training methods fare less well than methods that engage meaning representations. Nevertheless, repetition methods do seem to be effective in teaching form-form connections, and so the combined use of repetition training with methods that encourage semantic engagement might prove effective, even for less-proficient learners and words that are more difficult to learn. We examine that possibility in the next section.

Repetition Training with Retrieval Practice

Repeated exposure to word forms and meanings strengthens memory and leads to faster and more accurate retrieval upon subsequent encounters (e.g., Bolger et al., Reference Bolger, Balass, Landen and Perfetti2008). But, as demonstrated in the previous section, mere repetition is not always enough to promote successful learning. The following section concludes that repetition training is most successful when combined with retrieval practice. In support of this, we present several studies that have tested whether retrieval events strengthen memory for novel words beyond the effects of simple repetition.

What is retrieval practice, and why is it so powerful? Retrieval practice occurs when a learner is given a cue and retrieves a word form or meaning in response, thereby strengthening representations and the connection between them. This strengthening process is sometimes called the testing effect, which describes the finding that retrieving information under test conditions makes it easier to retrieve on subsequent occasions, and leads to more accurate retrieval than if it had simply been restudied (e.g., Bjork & Kroll, Reference Bjork and Kroll2015; Dunlosky et al., Reference Dunlosky, Rawson, Marsh, Nathan and Willingham2013; Joe, Reference Joe1998; Karpicke & Roediger, Reference Karpicke and Roediger2008; Kornell, Hays, & Bjork, Reference Kornell, Hays and Bjork2009; Pyc & Rawson, Reference Pyc and Rawson2010; Richland, Kornell, & Kao, Reference Richland, Kornell and Kao2009; Roediger & Karpicke, Reference Roediger and Karpicke2006; Schneider, Healy, & Bourne, Reference Schneider, Healy and Bourne2002). In the following section we briefly review studies that investigate the testing effect (hereafter referred to as retrieval practice) with L2 vocabulary learning, and then move on to studies that examine repetition training with and without retrieval practice. Although space does not permit a full discussion, we also point the interested reader to a related area of research concerning the use of the testing effect to elicit desirable difficulties in vocabulary learning, particularly Bjork and Kroll’s (Reference Bjork and Kroll2015) comprehensive review of this topic.^{Footnote 2}

Retrieval practice is an effective method of learning in general (e.g., Dunlosky et al., Reference Dunlosky, Rawson, Marsh, Nathan and Willingham2013), and robust evidence supports its generalization to learning novel L2 words. Two studies of retrieval practice during L2 vocabulary learning were conducted by Royer (Reference Royer1973) and Carrier and Pashler (Reference Carrier and Pashler1992). Royer trained participants on 20 Turkish-English word pairs using three different types of flashcards and instructions: (a) flashcards with an English-Turkish word pair on one side and the Turkish word on the other, with participants instructed to self-test until all words were mastered; (b) flashcards with an English-Turkish word pair on one side and nothing on the other side, with participants instructed to study words for an amount of time determined by yoking to a participant in the first condition; and (c) flashcards with an English-Turkish word pair on one side and nothing on the other side, with participants instructed to study until all words were mastered. Royer reported a significant advantage for the first condition – participants who were given flashcards that supported retrieval practice (i.e., had Turkish on the back) performed significantly better (96.5%) on an L2-L1 translation production test than those who were given flashcards that did not support retrieval practice (i.e., were blank on the back; 84.5%); these participants had an equivalent amount of time to study. However, the difference between the first and third conditions was not significant, demonstrating that although retrieval practice may speed up learning, approximately the same level of learning may eventually be reached by repetition training if enough time is allowed; the mean amount of time required for participants in the first versus third conditions to nearly master the 20 words was 13.8 versus 16.5 minutes. Therefore, retrieval practice seems to encourage faster and more accurate learning than simple repetition.

Similarly, Carrier and Pashler (Reference Carrier and Pashler1992) tested whether practice retrieving one word of a translation pair leads to better long-term recall than rehearsing translation pairs. They taught English-speaking undergraduates English-Siberian Yupik Eskimo word pairs. Following initial training in which a word pair was viewed for 20 seconds, two conditions were compared: a repetition condition in which both words were displayed on the screen for the duration of the trial, and a retrieval condition in which the L2 word was presented alone for half the trial, during which time participants were instructed to say the L1 word aloud, before the L1 word was added to the screen for the remainder of the trial. Words learned in the retrieval condition had a significant advantage over words learned in the repetition condition on both immediate and one-day delayed L2-L1 translation production trials.^{Footnote 3}

These studies demonstrate that retrieval practice is an effective tool to promote L2 vocabulary learning and can be used to enhance the effectiveness of repetition training (see also Barcroft, Reference Barcroft2007). However, the studies reviewed so far have tested only L2-L1 translation production, which limits the generalizability of their findings. L2-L1 translation production proceeds more rapidly than L1-L2 translation production and is less likely to engage conceptual processing (e.g., Kroll & Stewart, Reference Kroll and Stewart1994), and therefore testing in only one direction provides an incomplete picture of the effectiveness of training methods.

A study by Kang et al. (Reference Kang, Gollan and Pashler2013) partially addressed this shortcoming. Kang et al. trained native English speakers with no prior experience in Hebrew on Hebrew words using a training procedure in which learners heard Hebrew words while viewing the corresponding picture on a screen, either three (Exp. 1) or six (Exp. 2) times. Next learners encountered the target words in one of two conditions: (a) a retrieval condition, in which participants saw a picture and tried to produce a Hebrew word, and then after a pause heard the correct pronunciation; and (b) an imitation condition, in which participants saw a picture paired with an auditory pronunciation of the Hebrew word, and then were immediately asked to repeat the Hebrew word. Participants were given a production task in which they viewed a picture and orally produced the Hebrew word. On both immediate and two-day delayed L2 picture naming tests there was an advantage for words trained in the retrieval condition over the imitation condition. Although this task used pictures as cues, it demonstrated that the testing effect generalizes to L2 production, a task that involves conceptual mediation (e.g., Potter et al., Reference Potter, So, Eckardt and Feldman1984).

Although the studies reviewed in the preceding text did not report accuracy of retrieval attempts, there is evidence that retrieval events do not have to be successful to lead to better long-term memory, at least if corrective feedback is given after an incorrect attempt. Although somewhat counterintuitive, this has been demonstrated in L1 and L2 studies and is thought to occur because making an incorrect response might trigger a learner to engage more active learning processes and enhance encoding of corrective feedback (e.g., Kornell et al., Reference Kornell, Hays and Bjork2009; Richland et al., Reference Richland, Kornell and Kao2009). An example of this in L2 vocabulary learning comes from Potts and Shanks (Reference Potts and Shanks2014), who trained native English-speakers on rare English words (Exp. 1) or Basque words (Exp. 2 and 3). There were three conditions in all experiments: (a) translation pairs were shown together for 13 seconds (repetition condition); (b) a target word was shown by itself for 8 seconds while the participant guessed the meaning and then the correct translation was shown for 5 seconds (generation condition); or (c) a target word was shown by itself for 8 seconds while the participant attempted to choose the meaning from four alternatives and then the correct translation was shown for 5 seconds (choice condition). After training, participants completed a forced-choice recognition task in which they viewed a word and had to select the correct definition from four alternatives. Potts and Shanks compared accuracy for the three conditions and found that for all experiments the generation condition was significantly more accurate than the other conditions. To examine whether accuracy on the final test was affected by errors during learning, Potts and Shanks examined final test performance for items that were incorrectly generated or chosen during training (but note that all incorrect responses were followed by corrective feedback). Overall, errorful generation during learning led to better final testing accuracy than did incorrect choices or simple repetition during training, indicating that making mistakes did not harm performance. This may be due to the fact that participants received immediate feedback on whether generated translations or choices were correct, so connections between incorrect translation pairs did not have a chance to be strengthened, and in fact extra attention might have been directed to the correct translations when they were presented, strengthening these connections instead. These results demonstrate that corrective feedback is important to ensure the success of retrieval practice.

Thus far, we have reviewed studies of repetition training with and without retrieval. We now turn to a discussion of how these results can be understood in the context of the RHM-RER, which is especially interesting given the inconsistency in training efficacy between studies of repetition training with and without retrieval attempts. To review, the general pattern of results is that studies that used simple repetition without retrieval produced poorer outcomes than studies that used repetition training with retrieval practice. How can we understand this in the context of the RHM-RER?

First, we examine what happens when learners engage in repetition training without retrieval. In these studies, learners received L1-L2 word pairs, and were instructed to use repetition to learn the words. This repeated coactivation of L1 and L2 word forms uses one of the three strengthening mechanisms described by the RHM-RER (i.e., repetition), to strengthen the connections between word forms. However, this method does not use either of the other two strengthening mechanisms (i.e., elaboration or retrieval) and furthermore does not require the learner to access meaning representations. Thus, repetition training without retrieval can be understood as activation within only tier one of the RHM-RER, using only one of the three strengthening mechanisms.

In contrast, when learners engage in repetition with retrieval practice, they are using repetition to strengthen form-form connections as previously described, but in addition they are making use of a second strengthening mechanism (i.e., retrieval). The act of searching memory to retrieve a response leads to the formation of elaborative links between the correct response and associated concepts (e.g., the elaborative retrieval hypothesis; Carpenter, Reference Carpenter2009), which activates meaning representations. Thus, repetition training with retrieval can be understood as activation within and across the first and second tiers of the RHM-RER, using two of the three strengthening mechanisms. The benefit of using a greater number of strengthening mechanisms and the additional benefit of increasing activation within and across multiple tiers explains why repetition training is enhanced with the addition of retrieval practice.

Finally, although outside of the scope of this section to thoroughly review, it is worth noting an interesting avenue of inquiry that informs our understanding of the role of repetition learning with and without retrieval practice – the forward testing effect (Pastötter & Bäuml, Reference Pastötter and Bäuml2014; Yang, Potts, & Shanks, Reference Yang, Potts and Shanks2018). The classic testing effect can be thought of as a backward testing effect because the act of retrieving an already learned item from memory is known to enhance memory for already learned items. In contrast, recent research has suggested that retrieving previously studied items may also enhance learning and memory for novel related material. This effect suggests interesting future directions for the study of repetition and retrieval in L2 vocabulary learning, especially within the context of classroom studies, because most of the research on this effect thus far has been conducted in laboratories.

Spaced Repetition Training

Thus far, this review has discussed repetition training both with and without retrieval and demonstrated that the combination of repetition training and retrieval practice is effective for L2 vocabulary learning. In the following section we examine another method for successful repetition training – the spacing effect. The spacing effect states that learning proceeds more rapidly when repetitions of a word are spaced over longer periods (e.g., Benjamin & Tullis, Reference Benjamin and Tullis2010; Cepeda, Vul, Rohrer, Wixted, & Pashler, Reference Cepeda, Vul, Rohrer, Wixted and Pashler2008; Dempster, Reference Dempster1987; Seabrook, Brown, & Solity, Reference Seabrook, Brown and Solity2005). Expanding the spacing between repetitions of a target L2 word or between retrieval attempts significantly improves vocabulary learning outcomes (e.g., Atkinson, Reference Atkinson1972; Bahrick, Bahrick, Bahrick, & Bahrick, Reference Bahrick, Bahrick, Bahrick and Bahrick1993; Bahrick & Hall, Reference Bahrick and Hall2005; Cepeda et al., Reference Cepeda, Vul, Rohrer, Wixted and Pashler2008; Kang, Reference Kang2016; Pavlik & Anderson, Reference Pavlik and Anderson2005), but the optimal number of repetitions and the optimal spacing between the repetitions or sessions are the subject of ongoing inquiry. Additionally, different types of words may benefit from different spacing intervals. This section will briefly discuss the mechanisms that make spaced repetition effective for L2 vocabulary learning, review adult L2 vocabulary training studies of spaced repetition, and outline the training conditions that make spaced repetition more or less effective. Furthermore, this section will examine the effect of word pair characteristics and ask if words that are more difficult to learn (i.e., abstract words, noncognates, and low-frequency words) benefit from different spacing intervals than words that are less difficult to learn. Spacing is one area of the L2 vocabulary learning literature in which there are significant gaps in our knowledge because relatively few studies investigate spacing effects over periods longer than a few days, and even fewer investigate the effects of word characteristics.

We turn first to a discussion of the mechanisms of the spacing effect. One theory that describes the success of spaced repetition, the study-phase retrieval theory (e.g., Hintzman, Reference Hintzman2004, Reference Hintzman2010; Pavlik & Anderson, Reference Pavlik and Anderson2005, Thios & D’Agostino, Reference Thios and D’Agostino1976; Toppino, Phelan, & Gerbier, Reference Toppino, Phelan and Gerbier2018) states that when items that have not been seen for a while are repeated (i.e., spaced practice), they must be retrieved from memory and this process of retrieval strengthens associative connections.^{Footnote 4} In contrast, when encounters are presented without a gap between presentations (i.e., massed practice), they are still active in working memory and so there is no need to engage in retrieval. In this way spacing amplifies the effects of retrieval.

If spaced repetitions are more effective than massed repetitions, is there an optimal spacing between repetitions? One of the earliest and most comprehensive studies of the spacing effect in L2 vocabulary examined this question. Bahrick et al. (Reference Bahrick, Bahrick, Bahrick and Bahrick1993) directed a 9-year longitudinal study of retention of L2 vocabulary after training. They enrolled four participants with previous knowledge of either French or German. They selected 300 foreign language words that were unfamiliar to the participants and trained them in either 13 or 26 training sessions, with 14, 28, or 56 days between training sessions, and then tested retention at four intervals: 1, 2, 3, or 5 years. They reported that shorter spaces between sessions promoted better initial learning, but these differences leveled off over time, and on retention tests there was a large benefit of a longer gap between sessions – the percent of words recalled was highest for words trained with the longest intersession interval. Specifically, the benefits of spacing were greatest when the retrieval sessions were two months apart.

To date, Bahrick et al. (Reference Bahrick, Bahrick, Bahrick and Bahrick1993) is the only study that examines the effects of spacing of L2 vocabulary learning over such a long interval. However, shorter intervals have been successfully used, and these studies are informative when considering the applications of spacing in the laboratory or a classroom setting. For example, Pavlik and Anderson (Reference Pavlik and Anderson2005) recruited native English speakers and taught them Japanese-English word pairs using a paired-associate training procedure with interspersed testing trials, and manipulated the number of times participants were tested on a given word (one, two, four, or eight), the number of intervening trials between encounters (2, 14, or 98), and the retention interval (1 or 7 days). During training, accuracy was lower with a greater number of intervening trials, regardless of number of exposures. However, a comparison of training and posttest accuracy showed an interaction of spacing and forgetting, such that wider spacing in training resulted in less information forgotten in testing for both the 1- and 7-day delays. During testing, there were significant interactions of (a) repetition and spacing, such that the benefit of spacing went up with more initial exposures to a word, and (b) repetition and retention interval, such that the benefit of spaced practice was greater for the 7-day delay than the 1-day delay.

Although these studies give us a general idea of how to best implement spacing, it is possible that the optimal amount of time between retrieval attempts might not be the same for all words. Atkinson (Reference Atkinson1972) investigated the optimal way to order exposures and retrieval during L2 vocabulary learning, and whether this order differed depending on word characteristics. In this study, L1 English speakers learned German words using a paired associate learning paradigm with interspersed testing trials to provide retrieval practice and improve learning. Participants studied a list of English-German word pairs, and either they or the computer selected a test pair, leading to variation in both the number of intervening trials between encounters with a word, as well as the total number of encounters with a word. There were four conditions that determined the ways in which these variations were implemented: (a) students selected spacing of testing trials; (b) the spacing was selected randomly by the computer; (c) an algorithm^{Footnote 5} determined spacing, with equal difficulty assigned to all words; and (d) the same algorithm determined spacing, but varying levels of difficulty (i.e., expected error) were assigned to words. During training, performance was best for random spacing, followed by the self-selected order and the equal difficulty algorithm, and worst for the unequal difficulty algorithm. However, after either a 1-day or 7-day delay, the pattern of results was reversed – performance was best for the words trained with the unequal difficulty algorithm, second best for the self-selected spacing, and worst for the words trained with random spacing.

There are several interesting implications of this study. First, the success of the self-selected testing order shows that learners are aware of what they do and do not know during vocabulary learning, and they are able to use this information to create a spaced testing schedule that is more effective than a random order or an algorithm that assumes equal difficulty of vocabulary words. It is also interesting to note that the algorithm that assumes unequal difficulty of the vocabulary words performed the best during final testing. This suggests that word characteristics that make a word more or less difficult to learn (e.g., concreteness, frequency, cognate status, phonological familiarity) may influence the optimal spacing of encounters with a word. To the best of our knowledge there are no L2 vocabulary learning studies that investigate the effects of these word characteristics on the spacing effect, so this is an area in need of further investigation. Specifically, one could hypothesize that more difficult words (e.g., abstract words, lower frequency words, noncognates) might benefit more from spacing determined by learning algorithms rather than random spacing, whereas random spacing or self-selected spacing may be sufficient for less difficult words. Similarly, based on the predictions of the study-phase retrieval theory described in the preceding text, more difficult words might benefit from shorter spacing because overly long gaps might lead to failed retrieval. Conversely, less difficult words might benefit from longer spacing because these words are retained longer and if the spacing is too short the retrieval attempt will not engage effortful processing as well as longer spacing. Alternately, Sense, Behrens, Meijer, and van Rijn (Reference Sense, Behrens, Meijer and van Rijn2016) suggest that algorithms with parameters for individual rates of forgetting for different materials may outperform traditional approaches to determining optimal spacing.

Taken together, the results reviewed in this section demonstrate that the benefits of the spacing effect generalize to L2 vocabulary learning, and further that certain types of spacing may be more helpful for long-term retention of L2 vocabulary than others. For example, a greater number of days between encounters with a target word (Bahrick et al., Reference Bahrick, Bahrick, Bahrick and Bahrick1993), a greater number of intervening encounters with other words between encounters with a target word (Pavlik & Anderson, Reference Pavlik and Anderson2005), and varying spacing to according to the difficulty of the target word (Atkinson, Reference Atkinson1972) can all lead to better long-term learning outcomes. However, there are still many questions that remain to be answered in this area. First, it is not known if individual differences in cognitive skill interact with the spacing effect. For example, the spacing that would require a learner with higher working memory to engage in effortful retrieval may be different than the spacing that would require a learner with lower working memory to engage in effortful retrieval. Second, an interesting avenue for further research is whether words that are more difficult to learn might benefit from shorter or longer spacings than easier words.

Semantic Elaboration through Generation

One training method often used to encourage semantic processing is elaboration. Elaboration occurs when a learner accesses meaning representations and takes an action that strengthens these representations. Examples of strengthening actions include using a word in a context sentence, listing synonyms of a word, drawing a semantic map of associated words, or activating relevant background information and making connections between a word and semantically related concepts. However, whereas some studies report a benefit of semantic elaboration others have found that semantic elaboration has a negative effect on L2 vocabulary learning, a contradiction that we attempt to unravel in the following section.

We first examine a study that reported a benefit of enhancing semantic processing during L2 vocabulary learning. Coomber, Ramstad, and Sheets (Reference Coomber, Ramstad and Sheets1986) demonstrated that elaboration can be a powerful tool for promoting engagement with meaning representations. In this study, the authors taught artificial vocabulary words to undergraduates enrolled in an English-speaking university, using three training methods: definition matching, example matching, and sentence generation. In the definition condition, participants read along as an experimenter pronounced a list of novel vocabulary words and gave their definitions, and then participants practiced matching the word with the correct definition from a given list. This condition likely strengthened L1-L2 form connections because the definitions were short infinitive verb translations (e.g., to destroy, to create). In the example condition participants read a list of word-definition pairs, and then selected an example of how the word was used from a list of word-example pairs. This condition encouraged L2 form-meaning connections because matching a word to an example required the learner both to know the meaning of the novel word and to understand that certain examples did not fit with the meaning, but it did not require participants to engage in elaborative processing. Finally, in the sentence-composing condition participants read the list of word-definition pairs and then generated two grammatically correct and semantically meaningful sentences using the target word. This condition required participants to both retrieve a meaning representation as well as generate a response, which is a form of elaborative processing.

Following training, all participants were tested on accuracy of matching words with correct definitions, matching words with examples, or writing a semantically meaningful sentence containing the target word, and all items were tested with the three methods. Results revealed that words trained in the sentence-composing condition scored the highest on all three tests, and words trained in the definition condition scored the lowest on all three tests. Although words trained with examples scored higher than words trained with definitions, the differences were only marginally significant. The highest accuracy overall was for words trained with sentence writing and tested on definitions, whereas the lowest overall accuracy was for words trained with definitions and tested on sentence writing.

Coomber et al. (Reference Coomber, Ramstad and Sheets1986) explained the sentence-writing advantage in relation to the depth of processing hypothesis (Craik & Lockhart, Reference Craik and Lockhart1972) – sentence composing required more elaborative (i.e., deeper) processing than rehearsing definitions, and so resulted in better memory. This is one possibility, but another possibility is that the example condition did not require elaboration and the definition condition did not require either retrieval or elaboration, whereas the sentence generation condition required both retrieval and elaboration – participants had to recall the meaning of a word and then generate a context sentence for it. Semantic elaboration through generation is known to be a powerful tool for enhancing memory for words. Indeed, the generation effect describes the finding that participant-generated stimuli are remembered better than stimuli that participants simply read (Bertsch, Pesta, Wiscott, & McDaniel, Reference Bertsch, Pesta, Wiscott and McDaniel2007; Slamecka & Graf, Reference Slamecka and Graf1978). One theoretical explanation of the generation effect, the two-factor theory, posits that generation both strengthens connections between stimuli and responses (i.e., words and meanings), and also enhances activation of semantic features of the word (Hirshman & Bjork, Reference Hirshman and Bjork1988).

In the context of the RHM-RER, semantic elaboration through generation can be described as activation within and across the first and second tiers of the model, using two of the three strengthening mechanisms: retrieval and elaboration. When learners activate background knowledge related to a word, they retrieve and elaborate on meaning representations (tier 2) that strengthen the connections between the L2 form and meaning, instead of just the L2 and L1 words (tier 1). This activation of meaning features provides an additional pathway (i.e., a conceptually mediated pathway in addition to a lexically mediated pathway) that enables faster retrieval of the L2 word.

Semantic Elaboration through Thematic Grouping

In addition to encouraging conceptual access by means of generation, the ways in which words are grouped can impact the development of form-meaning connections. Clustering words according to a shared theme or schema in which words are part of some relational category (e.g., beach, sun, towel), but do not share semantic features (e.g., apple, orange, peach) may be more effective than either random presentation or clustering according to semantic features (e.g., Choi, Reference Choi2003; Elgort, Reference Elgort2011; Tinkham, Reference Tinkham1997; Tseng, Doppelt, & Tokowicz, Reference Tseng, Doppelt and Tokowicz2018). In the following section we review evidence for the benefits of thematic groupings and discuss possible mechanisms behind this effect (see Table 2 of the Supplementary Materials for a tabular overview of key studies in this section).

Tinkham (Reference Tinkham1997) conducted two experiments that compared thematically and semantically related groupings of English-artificial vocabulary word pairs. In this study, participants were trained with words in thematically related groupings (e.g., frog, hop, pond), thematically unrelated groupings (e.g., cloud, office, erase), semantically related groupings (e.g., tin, bronze, iron), and semantically unrelated groupings (e.g., paint, recipe, uncle), and tested on immediate L2-L1 production and 2-week delayed L1-L2 production. Semantically related sets took significantly more trials be learned than semantically unrelated sets, but thematically related sets took significantly fewer trials to be learned than thematically unrelated sets. No statistical comparisons of semantically and thematically related sets were conducted, but an examination of the raw data show that the thematically related words were learned in fewer trials than the semantically related words as shown by both immediate L2-L1 production and 2-week delayed L1-L2 production. Additionally, at the end of the study participants were interviewed and asked what words they found hardest or easiest to learn. The majority of participants reported that the thematically related words were easiest and almost all subjects reported that the semantically related words were the hardest to learn.^{Footnote 6}

Building on the work of Tinkham (Reference Tinkham1997), Tseng et al. (Reference Tseng, Doppelt and Tokowicz2018) examined the effect of teaching Arabic words and phrases to native English speakers using thematically related and unrelated groupings. The authors also included a manipulation of lexical quality by including or not including transliterations of the Arabic words (the Arabic script was never shown), and hypothesized that the inclusion of richer lexical representations (i.e., transliterations) might allow participants to encode and retrieve words more easily than with weaker lexical representations (i.e., no transliterations). Following training, participants completed a free recall task and an L1-L2 oral translation production task, and then returned for seven more sessions over a 4-week period. Results for translation production did not show an effect of grouping or transliterations. However, free recall data showed a significant benefit for transliterations over no transliterations, and a four-way interaction of transliteration, session, grouping, and working memory span. Participants with higher working memory spans were less affected than participants with lower working memory spans by grouping or the presence or absence of transliterations. Furthermore, although lower working memory span participants generally performed more poorly than higher working memory span participants, when lower span participants were given words in a thematic grouping with transliterations they outperformed their higher working memory span peers. This finding suggests that the benefits of thematic grouping may be influenced by individual differences (e.g., Tinkham, Reference Tinkham1993, Reference Tinkham1997). One additional factor that may explain some of the differences between this study and previous research is that whereas earlier studies examined words (and primarily concrete nouns) in isolation, this study included words spanning a number of different word classes as well as short phrases.

Taken together, the results of these studies indicate an advantage for thematic groupings, at least in some circumstances. Why might this be the case? Presenting words in thematic groupings may allow learners to take advantage of prior theme-related knowledge, and thereby engage in semantic elaboration during learning. As described previously, elaboration activates background knowledge related to a word, and this activation of additional meaning features provides an additional pathway that may later enable faster retrieval of the L2 word because multiple pathways for later retrieval were encoded during learning. Note that this explanation does not apply to semantic groupings, which will be discussed in the following section.

Negative Effects of Semantic Grouping

Whereas classroom approaches to L2 vocabulary instruction often present semantically related sets of words at the same time (e.g., Gairns & Redman, Reference Gairns and Redman1986; Marzano & Marzano, Reference Marzano and Marzano1988; Tinkham, Reference Tinkham1997), recent laboratory studies have shown that semantic groupings (i.e., words that have overlapping semantic features, such as cherry, apple, plum) may cause semantic interference and interfere with encoding and retrieval (Finkbeiner & Nicol, Reference Finkbeiner and Nicol2003; Kroll & Stewart, Reference Kroll and Stewart1994). There is evidence that semantic grouping are less effective than either random presentation or thematic groupings (e.g., Elgort, Reference Elgort2011; Tinkham, Reference Tinkham1997; Tseng et al., Reference Tseng, Doppelt and Tokowicz2018).^{Footnote 7} The following section reviews these studies (see Table 2 of the Supplementary Materials for a tabular overview of key studies in this section).

Similar to Tinkham (Reference Tinkham1997), Finkbeiner and Nicol (Reference Finkbeiner and Nicol2003) provided evidence that semantic groupings negatively impact vocabulary learning. In this study, monolingual English speakers learned novel English-like nonword-picture pairs and were tested on L2-L1 and L1-L2 translation production. Semantic grouping was manipulated in both training and testing, creating four conditions: (a) related training, related testing; (b) related training, unrelated testing; (c) unrelated training, unrelated testing; and (d) unrelated training, related testing. Thus, the authors were able to test both the effects of semantic grouping and also the effects of congruency of training and testing. They reported a strong negative effect of semantic grouping in training and testing for both directions of translation, and an effect of congruency such that the best performance was found for the congruent unrelated training, unrelated testing condition and the worst performance was found for the incongruent related training, unrelated testing condition.

Why might this be the case? Kroll and Stewart (Reference Kroll and Stewart1994) hypothesized that presenting words in semantic groupings causes a large number of shared meaning features to become active. When too many meaning features are active, difficulties arise in mapping these features to a single word form and thereby in selecting a single lexical entry for production, leading to competition and interference in encoding and retrieval. In the context of the RHM-RER, in contrast, thematic words do not share many semantic features, and so are less likely to cause competition due to overactivation of meaning representations. Nevertheless, thematically associated words derive some benefit over semantically and thematically unrelated words. This can be viewed as a benefit that arises from semantic elaboration during learning. Thematic associations allow learners to use prior knowledge to make connections between novel and existing words and meanings during learning, which can be conceptualized as one of the three strengthening mechanisms – elaboration.

Negative Effects of Semantic Elaboration

As described in the preceding text, the generation effect describes the finding that participants remember stimuli they generate better than stimuli that they simply read (e.g., Bertsch et al., Reference Bertsch, Pesta, Wiscott and McDaniel2007; Slamecka & Graf, Reference Slamecka and Graf1978), and this effect may underlie the benefits of semantic elaboration (e.g., Coomber et al., Reference Coomber, Ramstad and Sheets1986). However, there are a handful of studies that report a negative effect of semantic elaboration. In this section we review studies in which semantic elaboration produces effects that are inconsistent with the predictions of the RHM-RER.

Barcroft (Reference Barcroft2009) examined semantic elaboration through synonym generation during the early stages of L2 vocabulary learning. A sample of 114 native Spanish speakers in lower- or higher-intermediate English (L2) classes were assigned to incidental or intentional learning conditions and instructed to read an L2 passage with 10 low-frequency L2 words and their L1 translations. The incidental group was instructed to read only for meaning, whereas the intentional group was informed that there would be a vocabulary test afterward. In addition, half of each learning group was instructed to generate L1 synonyms for the target words, and the other half did not receive that instruction. At the end of the study participants completed an L1-L2 translation production task, followed by an L2-L1 translation production task. Mean recall scores were significantly higher for the intentional study group than the incidental study group, and higher for the groups that did not generate synonyms than the groups that did generate synonyms. This pattern was true for both the L1-L2 and L2-L1 translation tasks, although accuracy was higher for L2-L1 translation production. There was no interaction of study group and synonym generation, and the effects of these conditions did not differ for the higher versus lower proficiency groups.

At first glance this study appears to discourage the use of semantic elaboration using synonym generation. However, as the author pointed out, incongruence between the training and testing conditions may explain the negative effects of synonym generation on learning (i.e., transfer-appropriate processing). It may also be the case that the instruction to generate a synonym did not accomplish the goal of semantic elaboration but instead simply distracted participants from the task at hand and made it more difficult for the participants to form L1-L2 word form connections. Without testing an additional condition (e.g., letter counting), the possibility that any secondary task would have caused the same findings cannot be ruled out. Establishing mappings between L1 and L2 word forms is a critical step in early vocabulary learning, and focusing on semantic elaboration during initial exposures may have distracted learners and reduced their ability to learn a new form and its appropriate mapping.

A similar study by Barcroft (Reference Barcroft2002) compared semantic and structural elaboration in a group of low-intermediate proficiency English-Spanish learners. Words were trained in one of three within-subjects conditions: (a) no elaboration, for which participants were simply instructed to learn word pairs; (b) structural elaboration, for which participants were instructed to count the number of letters in the L2 word; and (c) semantic elaboration, for which participants were instructed to rate the words according to perceived pleasantness. During all conditions the L2 word and a picture representing meaning were displayed on the screen. Testing consisted of immediate L1 and L2 free recall tasks, and then L2 translation production. Barcroft reported that for Spanish free recall, the no elaboration and the structural elaboration sets were recalled significantly more accurately than the semantic elaboration set. For English free recall the semantic elaboration set was recalled significantly more accurately than the structural elaboration set, and the no elaboration set was recalled significantly more accurately than either of the elaboration sets. In other words, for a known language, semantic was better than structural elaboration, but for an unknown language, structural was better than semantic elaboration. In L2 translation production, no elaboration was significantly better than either type of elaboration, and structural elaboration was better than semantic elaboration, although this effect was only marginally significant.

The results of this study seem to support the idea that semantic elaboration can negatively impact L2 vocabulary learning, but several questions should be considered before accepting this conclusion. First, due to the use of picture-word pairs during training, it is possible that all three conditions activated meaning representations and helped strengthen form-meaning connections, which means that this study tested whether adding semantic or structural elaboration in addition to activating meaning representations confers an added benefit. Additionally, it is not clear whether the semantic elaboration condition (i.e., pleasantness rating) should be considered elaboration, at least not in the same way that a task like sentence generation promotes elaboration. Although perhaps sufficient to engage meaning representations (likely already engaged by pictures), providing a pleasantness judgment does not require one to transform or generate connections between related meaning features. In light of this, it seems possible that the disadvantage for semantic elaboration reported by Barcroft (Reference Barcroft2002) was due more to a failure to engage semantic elaboration than an actual disadvantage of this type of elaboration.

Overall, there is a relative paucity of research examining the effects of semantic elaboration during L2 vocabulary learning, and the findings seem to depend on the tasks used in training and testing. Additionally, all studies presented in this section used immediate testing, and none examined the effects of semantic elaboration after a delay. It seems possible, given research examining desirable difficulties (e.g., Bjork & Kroll, Reference Bjork and Kroll2015), that although the effects of semantic elaboration often initially appear to be negative, this pattern might reverse after a delay. Further research should examine this possibility.

Another possibility for incongruent effects of semantic elaboration is that the effectiveness of training methods may vary according to learner L2 experience. As reviewed in the preceding text in more detail, van Hell and Candia Mahn (Reference van Hell and Candia Mahn1997) compared the efficacy of repetition training and the keyword method for inexperienced and experienced L2 learners. They reported that experienced learners recalled target words more accurately and more quickly in the repetition condition than the keyword condition, whereas inexperienced learners showed no difference between the two instructional methods.

Additional evidence that the effectiveness of elaborative methods depends on the L2 experience of the learners comes from Moore and Surber (Reference Moore and Surber1992). As previously described, English-speaking students enrolled in beginning, intermediate, or advanced German classes completed training in one of three conditions: repetition training, learning meanings from context sentences, and learning meanings using the keyword method. After testing on L2-L1 translation production and L2 sentence completion, the authors reported an advantage of context or keyword training for lower and medium proficiency learners, but more words were recalled by higher proficiency learners on the L2 sentence completion task and an immediate L2-L1 translation production task when words were trained with repetition than with the keyword method or with context sentences. The authors suggested that advanced learners have developed successful strategies for learning, and attempting to shift those strategies in the laboratory may be detrimental, whereas for beginning learners who have not developed their own strategies it may be helpful.

Taken together, these studies demonstrate that methods of semantic elaboration, barring the handful of exceptions discussed previously, are effective at promoting long-term retention. Specifically, elaborative processing is beneficial for less advanced learners, but repetition training is more effective than elaborative methods for more advanced learners (Barcroft, Reference Barcroft2004; Coomber et al., Reference Coomber, Ramstad and Sheets1986; van Hell & Candia Mahn, Reference van Hell and Candia Mahn1997). This is consistent with studies that have demonstrated that bilinguals are more adept at novel word learning than monolinguals (e.g., Kaushanskaya & Marian, Reference Kaushanskaya and Marian2009; Papagno & Vallar, Reference Papagno and Vallar1995). In the context of the RHM-RER, semantic elaboration allows a user to activate background knowledge related to a target word. This enhanced activation strengthens meaning representations and allows learners to establish multiple routes from form to meaning, which aid in later retrieval. Additionally, the effectiveness of semantic elaboration methods depends on congruence between training and testing task demands (Morris et al., Reference Morris, Bransford and Franks1977). Studies of semantic elaboration that do not align with the predictions of the RHM-RER tended to use tasks that were too difficult for beginning learners or training tasks that were incongruent with testing tasks.