Image-schema-based-instruction enhanced L2 construction learning with the optimal balance between attention to form and meaning

2.1 Input flood and frequency distribution

Constructions, defined as learned pairings of form and meaning or function (Ellis and Wulff 2020), are the basic units of our linguistic knowledge (Goldberg 2006). The usage-based model of language learning holds that constructions must be learned by processing many exemplars in language input, and our language system (i.e., a network of constructions) emerges gradually on the basis of generalizing over item-specific constructional knowledge (Ellis 2006; Goldberg 2003, 2006; Tomasello 2000a, 2000b). In this regard, input flood, which is one of the most implicit form-focused instruction (Loewen 2018), can aid frequency-based construction learning by artificially increasing the frequency of a target construction. Furthermore, repetitive exposure to multiple exemplars seeded in the input can enhance the salience of a linguistic target (Han et al. 2008; Liu et al. 2021), assisting learners in noticing linguistic features within meaningful contexts (e.g., stories), and thereby facilitating the form-meaning association of a construction. Such theoretical motivation notwithstanding, a large body of empirical research into the effectiveness of input flood has found mixed results. Some studies found significant effects of input flood alone on L2 learning (e.g., Hernández 2011; Loewen and Inceoglu 2016; Reinders and Ellis 2009), but other studies failed to observe such benefits (e.g., Jahan and Kormos 2015; Loewen et al. 2009; Szudarski and Carter 2016). These findings suggest that input flood by itself might not be salient enough to draw learners’ attention to targeted linguistic structures (Loewen 2020). In this regard, several attempts have been made to investigate the effects of input flood in combination of other techniques such as textual enhancement and explicit instruction. Overall, the benefits of compound input flood over simple input flood or a control group have been reported (e.g., Issa and Morgan-Short 2019; Liu and Hao 2021; Rassaei 2015), with the caveat that overenhancing target constructions could be counterproductive (Han et al. 2008; White 1998).

Given the repetitive nature of input flood (Szudarski and Carter 2016), the inconsistent results of previous studies might be due to large variations in the number of occurrences of a target construction. For example, Hernández (2011) and Loewen and Inceoglu (2016), which found the positive effects of input flood, presented 20 and 28 times of target constructions each, whereas the number of encounters with target items were 5 and 12 times in Jahan and Kormos (2015) and Szudarski and Carter (2016), respectively. While no clear evidence has been obtained for the optimal number of exemplars for input flood (Wong 2005), a small group of studies have compared the effects of the input with different numbers of exemplars in a systematic way (e.g., Leow et al. 2003; Shook 1999). Most recently, Liu et al. (2021) found that providing contextual cues for inferring a word’s meaning had more profound effects on L2 lexical learning than manipulating the number of instances of the target item from 2 to 10. These results indicated that merely the higher frequency of a construction in input flood might not necessarily lead to better learning outcomes because learners become desensitized and pay less attention to the later encounters of the same target words (e.g., Winke 2013).

In tandem with the frequency of exposure to a construction, frequency distribution is another methodological factor that has remained underinvestigated in input flood (e.g., Denhovska et al. 2016). There is good evidence that type/token frequency distributions can play a substantial role in construction learning (Ellis and Ferreira-Junior 2009; Goldberg 2006), specifically facilitated by the presence of one prototypical member that accounts for the lion’s share of a particular construction. For example, Goldberg et al. (2004) examined extensive corpus data of children’s and mothers’ speech. The analysis of fifteen mothers’ speech to their 28-month-old children revealed that one prototypical verb appeared in a particular construction with very high frequency (e.g., the verb put accounting for almost 38 % tokens of caused-motion constructions) and this tendency was mirrored in their children’s speech. The authors concluded that such skewed input might produce a cognitive anchoring effect which may induce children to form a tentative category organized around the central member. This initially formed category, in turn, might serve as a stronger category representation for easier assimilation of other members to that category (Elio and Anderson 1984). Goldberg et al. (2007) further manipulated the order of presentation of instances. In the experiment, the skewed-first group was first provided with eight instances of one prototypical verb before those of four other verbs (i.e., 8-2-2-2-2) whereas the skewed-random group met the same eight instances later than the other instances (i.e., 2-2-2-8-2). It was found that the skewed-first group identified new instances of the target construction more accurately than the skewed-random group. Their results suggested that shared concrete similarity in the initial stage of construction learning allowed learners to form an abstract category that could account for most of the instances of the target construction. In a recent study, Zhang and Mai (2021) reported an additional facilitative effect of skewed-first input on the learning of English present and past counterfactual conditionals, which shared subtle differences in both form and meaning. Their findings highlighted that an initial low variance of the skewed input might reduce negative interference that arises when two similar constructions are learned together. With these benefits in mind, the present study manipulated frequency distribution in input flood to be skewed disproportionately towards a single prototypical verb of a target construction. It was hypothesized that such skewed input flood would facilitate frequency-based construction learning (e.g., extracting regularities in language usage).

2.2 Image-schema-based-instruction in L2 pedagogy

Mark Johnson, known for his early contributions to cognitive semantics, criticized the objectivist view of meaning, which assumed that the nature of meaning was independent of human experiences with their surrounding physical and social environment. Instead, Johnson argued that “meaning cannot be separated from the structures of our embodied perceptual interactions and movements” (Johnson 1989, p. 109). In a critical response to the objectivist view, cognitive semantics introduced image-schema in order to explain this embodiment of meaning (e.g., Lakoff and Johnson 1980).

Image-schema, as one of the key terms of cognitive linguistics, is “a relatively abstract conceptual representation that arises directly from our everyday interaction with and observation of the world around us” (Evans 2007, p. 106). To put it simply, image-schema is the abstract generalization of everyday experiences of the physical and social world. For instance, infants as young as six months can distinguish caused-motion from self-motion (Mandler 2004). Their cognitive system begins to develop the conceptual structure of caused-motion through repeated encounters with the spatio-physical events in which someone exerts force to move an entity from one place to another. The image-schema exemplified in Figure 1 emerged through generalization over perceived similarities of such recurring experiences. That is, it is our embodied experiences with the surrounding environment that give rise to such image-schema. Mandler (2004) further argued that such image-schema, in turn, could shape our cognition and provide fundamental frames of reference for linguistic concepts and meanings. For example, the caused-motion image-schema in Figure 1 serves as a conceptual tool for the interpretation of caused-motion constructions. In this respect, language is firmly grounded in our experience, and meanings are inseparable from the ways we perceive and categorize the real world around us (Ellis and Cadierno 2009). Therefore, meaning is embodied (Johnson 2007; Tyler 2012).

Figure 1:

Caused-motion image-schema (Mandler 1992, p. 595).

With many variations in terminology (e.g., mental imagery, schematic diagram, core-image, proto-scene, schema of a complete orienting basis of action), the idea to utilize image-schema for L2 pedagogy is not new, and a number of quasi-experimental studies have investigated the effectiveness of image-schema-based instruction on the learning of various types of constructions such as (a) prepositions (e.g., Hung et al. 2018; Tyler et al. 2011, Wong et al. 2018; Zhao et al. 2020), (b) verbs (e.g., Morimoto and Loewen 2007; Mueller and Tsushima 2019; Yamagata 2018; Verspoor and Lowie 2003), (c) modal verbs (e.g., Tyler et al. 2010), (d) phrasal verbs (e.g., Lee 2016; White 2012), (e) figurative idioms (e.g., Boers et al. 2004; Szczepaniak and Lew 2011), and (f) syntactic structures (e.g., Jacobsen 2016). The common principle of such image-schema-based instruction is to take advantage of image-schema as a vehicle to visualize an abstract core meaning that underlies a particular construction. The increased salience of the central meaning in this way is expected to help learners internalize embodied meaning which lays conceptual and semantic foundations of the construction.

While the potential benefits of image-schema-based instruction for L2 construction learning are generally acknowledged (Yu 2022), its main rationale can be justified by the following two theoretical assumptions. First, image-schema can facilitate deeper processing of the form-meaning association of constructions. The levels of processing model (Craik and Lockhart 1972) indicates that the greater cognitive efforts involved when information is processed will create richer memory representations, ultimately leading to better performance and retention than otherwise shallow processing. In the scene of L2 learning, deeper levels of processing refer to the simultaneous attention to form and meaning in that it presumably requires higher amount of cognitive involvement and elaboration (Leow 2015; Rice and Tokowicz 2020). However, since our limited attentional resources tend to give priority to processing input for meaning over form, making the concurrent association of form and meaning would be more effortful and demanding for L2 learners (VanPatten 2004). In this regard, image-schema-based instruction could save L2 learners the trouble of implicitly extracting the shared semantic structure of a construction from multiple instances in the input, the process of which might be challenging in EFL classrooms due to limited amounts of the input (Morimoto and Loewen 2007). As a result, such meaning salience can reduce cognitive load for semantic processing and simultaneously allow more attentional resources to be allocated for a higher level of form processing (i.e., accurate understanding of the underlying rule).

Second, the abstract and schematic nature of image-schema can help learners make inferences on novel usages (Tyler and Evans 2004; Mahpeykar and Tyler 2015), resulting in better transfer to similar linguistic contexts. Since image-schema is established through generalization of similar and recurring experiences, they are “vague and flexible enough to be adjustable to a range of different contexts” (Littlemore 2009, p. 130). Pedagogically, this also means that image-schema can present the shared core semantics in an intuitively appealing way, preventing L2 learners from being constrained to possibly inaccurate L1–L2 equivalents or metalanguage (Morimoto and Loewen 2007). Therefore, the inherent flexibility of image-schema can help L2 learners have a deeper understanding of the cognitive mechanism behind seemingly unmotivated meanings of constructions, and, in turn, better generalize to the novel usages in different contexts in systematic and consistent ways.

It has been pointed out that previous studies on image-schema-based instruction are “largely confined to the pedagogical exploitation of figurative thought, and this mostly pertains to polysemous words (especially prepositions and particles) and idiomatic expressions” (Boers and Lindstromberg 2006, p. 305). In this regard, future research is needed to see whether the potential benefits of image-schema could be transferred to more complex syntactic constructions (Boers and Lindstromberg 2008; Tyler 2012). To address this gap, argument structure constructions were chosen as the linguistic target for the study.

2.3 Argument structure constructions

Unlike a traditional view that constructions only refer to clause-level expressions (e.g., wh-construction), constructions in the usage-based approaches can vary in their size and complexity across all layers of language (Goldberg 2006). For instance, constructions can be morphemes (e.g., pre-, -ing), words (e.g., and, avocado, regular plural [N-s]), and idiomatic expressions (e.g., give the Devil his due, jog <someone’s> memory). Furthermore, abstract syntactic patterns by themselves, so-called argument structures, are also constructions. More specifically, the verb put in (1) is a predicate which typically takes three arguments: an agent (Jane), a theme (her toys), and a goal (into the box). Whereas the verb swim only requires an argument of someone who is swimming, it can also appear with the non-compositional meaning as in (2), roughly interpreted as ‘the dinosaur moved his friends to the mainland by means of swimming’.

Goldberg (1995, 2006 suggested that it was the argument structure that overrode the meaning of the verb swim and coerced it into having a specific sense. In other words, the shared syntactic pattern (i.e., Subject Verb Object Oblique_path/loc) underlying in (1) and (2) yields the caused-motion interpretation (i.e., X causes Y to move along or towards Z). This is possible because additional theme and goal arguments are imposed on the verbs that otherwise cannot convey such non-compositional meaning by themselves. These examples well demonstrate that argument structures are also one kind of constructions in that their syntactic patterns alone can carry an inherent meaning, independently of the actual verb with which they combine. In this particular case, the argument structure constructions exemplified in (1) and (2) are called caused-motion constructions, which are the linguistic target of the present study. Especially, caused-motion constructions like (2) were defined as non-typical because the verbs (e.g., swim, fly, drive) that appeared in the non-typical construction did not have a caused-motion sense by default, and must be integrated with its syntactic structure to gain such interpretation.

4.1 Participants

A total of 156 high school students (aged 15–16) in South Korea participated in the study. The students’ average English proficiency was low-to-intermediate, as measured with their English grades in Korea National College Scholastic Ability Test. The majority belonged to or below the fifth bands on the Stanine scale (i.e., the bottom 60 % of the nationwide testing population). Eight intact classes were randomly divided into (a) an input flood (IF) group (n = 37), (b) a metalanguage (MT) group (n = 38), (c) an image-schema (IS) group (n = 43), and (d) a control group (n = 38).

4.2 Target constructions

4.3 Learning materials

The IF group was asked to read four narrative stories (see Figure 2) in which multiple exemplars of the typical caused-motion constructions were embedded with boldfaced so that learners’ attention could be also given to form incidentally while processing meaning (e.g., Issa and Morgan-Short 2019; Winke 2013). Specifically, caused-motion constructions with the verb put appeared more frequently (26 times) than those with nine other verbs (5.8 times on average) on the assumption that a prototypical verb of a particular construction could play a path-breaking role in the initial detection of pattern similarities and construction category (Goldberg et al. 2004).

Figure 2:

Example of the narrative story.

Furthermore, 15 true-or-false comprehension questions were provided at the end of each story to encourage learners to spend enough time reading relevant caused-motion constructions. For example, learners should properly process the typographically highlighted sentence ‘ Mother Goat put her scissors, needle, and thread into her pocket ’ in the reading text to judge the truth of the statement ‘Mother Goat put her scissors, needle, and thread into her basket’. The instruction lasted about 1 h (i.e., 15 min for each story), and no explicit rule explanation on caused-motion constructions was provided to ensure, as much as possible, that learners remained unaware of what was being learned. The reading activities for the IF group were structured in such a way that caused-motion constructions could be learned implicitly during meaning-oriented activities (e.g., Godfroid 2016; Long 2017).

The MT group was opposite to the IF group in that the rules of caused-motion constructions were explicitly taught by using grammatical terminology (e.g., subjects, objects, and adverbials). For instance, learners in the MT group participated in decontextualized activities such as completing quizzes on the syntactic features of caused-motion constructions (e.g., how many nouns does a caused-motion construction have?) and putting words into a right order to match given L1 Korean sentences (see Figure 3). In this respect, the MT group received form-oriented instruction where the formal aspects of caused-motion constructions were more emphasized than meaning. The lesson lasted about 1 h.

Figure 3:

Example of the sentence-analyzing activity.

An instructional focus of the IS group was to balance between meaning- and form-oriented instructions. This goal was achieved by employing the image-schema that underlied caused-motion constructions. Unlike image-schema presented as static pictures in previous research, the image-schema in the present study was created with animation effects to highlight the dynamic nature of caused-motion. For example, learners in the IS group studied a typical caused-motion construction ‘Jane put apples into the basket’ by watching the apples depart from Jane to the basket moving along the arrow (see Figure 4). At the end of the animation, a color-coded word corresponding to each picture was shown one at a time so that learners’ attention could be paid to both form and meaning, which would promote deeper levels of processing (Craik and Lockhart 1972; Leow 2015; Rice and Tokowicz 2020). Replacing the explicit rule presentation in the MT group, the animated image-schema was presented as a part of 1-h activities such as explaining a given image-schema with learners’ own words and drawing the image-schema of caused-motion constructions.

Figure 4:

Example of the animated image-schema of caused-motion construction.

4.4 Korean-to-English translation test

A translation test was used to measure a controlled use of caused-motion constructions (e.g., Jahan and Kormos 2015; Wong et al. 2018) due to low English proficiency level of learners. Key content words were supplied as stimuli to elicit target responses. As for each item, the first Korean sentence served as a situational context and learners were asked to translate the second Korean sentence into English by using the words provided (see Figure 5). In the translation test, learners were required to (a) provide a proper preposition and (b) arrange words according to the syntactic pattern of caused-motion constructions.

Figure 5:

Example of the context-embedded Korean-to-English translation test.

The translation test consisted of 10 items for the typical and non-typical caused-motion constructions, respectively, and 10 filler items were included to mask the purpose of the test (see Supplementary Material A). One point was given for each correct answer. Inappropriate words arrangement or preposition selection were scored as an incorrect answer with 0 points. Errors not directly related with caused-motion constructions (e.g., spelling, tense, article) were not reflected in scoring. The translation test was conducted three times (i.e., pretest, immediate posttest, and delayed posttest) with a 15-min time limit for each. In order to prevent prior testing experience from providing extra language input that could affect a subsequent test, there was a two-week interval between each test. The items were presented in a randomized order for each testing phase. The experimental design of the study is summarized in Figure 6.

Figure 6:

Overview of the experimental design.

4.5 Analysis

Using the lme4 package in R, separate mixed-effects logistic models were built in order to investigate the effectiveness of three types of instruction on the learnability of the typical caused-motion constructions (RQ 1) and generalizability to the non-typical caused-motion constructions (RQ 2). Fixed effects included Group (i.e., IF, MT, IS, and control groups), Time (i.e., pretest, posttest 1 and posttest 2), and the interaction of Group by Time. Treatment coding was used for Group and Time with the control group and pretest as a baseline, respectively. By-participant random slopes for Time were included to allow the effect of Time to vary for each participant. By-verb random intercepts were also requested to take the idiosyncrasies of individual target verbs into account. The inclusion of the bobyqa optimizer led to successful convergence of this model. Cronbach’s α for the internal consistency of the translation test were 0.95 and 0.90 for the typical and non-typical caused-motion constructions, respectively.

5.1 The IS group showed the largest amount of improvement on the learning of the typical caused-motion constructions (RQ 1)

Figure 7 visualizes the results of the translation tests on the typical type of caused-motion constructions. Mean scores of individual participants were dotted around 95 % CIs of the group means across the testing times. It was revealed that all four groups showed improvement on two posttests in general, but they appeared to already be able to produce nearly half of the typical caused-motion constructions correctly at the time of pretest (see Supplementary Material B).

Figure 7:

Mean scores and 95 % CIs of the translation test on the typical caused-motion constructions.

The results of the mixed-effects logistic regression indicate that 74 % of the variance was explained by both fixed and random effects combined. It is also confirmed that the simple effects of posttests 1 and 2 on the control group were significant (β = 0.94, 95 % CI [0.12, 1.77], z = 2.24, p < 0.05; β = 1.08, 95 % CI [0.29, 1.88], z = 2.69, p < 0.01, respectively). In addition, the simple effects of the IS group on posttests 1 and 2 were significantly larger than the simple effect of the IS group on pretest (β = 2.51, 95 % CI [1.31, 3.71], z = 4.10, p < 0.001; β = 2.28, 95 % CI [1.14, 3.42], z = 3.91, p < 0.001, respectively) (see Supplementary Material C).

Figure 8 is a graphic illustration of the results in terms of the predicted probability of providing correct responses on the typical caused-motion constructions and 95 % CIs by group over time. A considerable degree of overlap among the CIs at the time of pretest suggests that there were no significant group differences in the pretest scores. However, the IS group significantly outperformed the control group on both posttests 1 and 2, as indicated by the non-overlapping CIs between the two groups. More specifically, the learners in the IS group were predicted to produce the typical caused-motion constructions most correctly with about 0.95 (95 % CI [0.85, 0.98]) chance on posttest 1 and 0.94 (95 % CI [0.85, 0.98]) chance on posttest 2. On the contrary, the IF and MT groups failed to perform better than the control group on two posttests. Although it is true that the predicted probability increased over time for all four groups, the varying widths of the CIs show the relative influence of each group on learnability. That is, the narrower CIs of the IS group on two posttests, which even do not overlap its own CI of pretest, could be interpreted as substantially reduced gaps between the individuals’ test scores in the IS group over time. To summarize, the image-schema-based instruction showed the largest amount of improvement on the learning of the typical caused-motion constructions and such benefits were maintained over two weeks.

Figure 8:

Predicted probability of correct responses and 95 % CIs on learnability.

5.2 The IS group showed the greatest degree of generalizability to the non-typical caused-motion constructions (RQ 2)

Figure 9 visualizes the results of the translation tests on the non-typical type of caused-motion constructions. Mean scores of individual participants were dotted around 95 % CIs of the group means across the testing times. It was revealed that the IS group scored highest on two posttests, being able to produce nearly half of the non-typical caused-motion constructions correctly. However, the rest three groups showed only a slight upward trend in general (see Supplementary Material D).

Figure 9:

Mean scores and 95 % CIs of the translation test on the non-typical caused-motion constructions.

The results of the mixed-effects logistic regression indicate that 60 % of the variance was explained by both fixed and random effects combined. It is also confirmed that the simple effects of the IS group on posttests 1 and 2 were significantly larger than the simple effect of the IS group on pretest (β = 2.14, 95 % CI [1.34, 2.94], z = 5.22, p < 0.001; β = 2.27, 95 % CI [1.42, 3.12], z = 5.24, p < 0.001, respectively) (see Supplementary Material E).

Figure 10 is a graphic illustration of the results in terms of the predicted probability of providing correct responses on the non-typical caused-motion constructions and 95 % CIs by group over time. A considerable degree of overlaps among the CIs at the time of pretest suggests that there were no significant group differences in the pretest scores. However, the IS group significantly outperformed the IF, MT, and control groups on both posttests 1 and 2, as indicated by its CIs that were slightly overlapped at the end or not overlapped at all by those of the rest three groups. More specifically, the learners in the IS group were predicted to produce the non-typical caused-motion constructions most correctly with about 0.56 (95 % CI [0.33, 0.76]) chance on posttest 1 and 0.64 (95 % CI [0.40, 0.82]) chance on posttest 2. However, its CIs with a wide range possibly indicate considerable individual variations in the test scores. On the contrary, the IF and MT groups failed to perform better than the control group on two posttests. To summarize, the image-schema-based instruction showed the greatest degree of generalizability to the non-typical caused-motion constructions on both posttests, but such benefits should be interpreted with caution due to its more widened CIs than those for the typical caused-motion constructions.

Figure 10:

Predicted probability of correct responses and 95 % CIs on generalizability.