Explicit and Implicit Knowledge and the Interface Issue

The constructs of IK and EK are central to SLA theory construction and evaluation. There is a consensus that SLA draws on both implicit and explicit learning mechanisms (Bley-Vroman, Reference Bley-Vroman1991; DeKeyser, Reference DeKeyser, Doughty and Long2003; N. C. Ellis, Reference Ellis2005; R. Ellis, Reference Ellis2004), which in turn may result in IK and/or EK (Williams, Reference Williams, Ritchie and Bhatia2009). Explicit knowledge is knowledge we are consciously aware of, whereas IK is the knowledge that we have but are not aware of (DeKeyser, Reference DeKeyser, Long and Doughty2009; Hulstijn, Reference Hulstijn2005; Williams, Reference Williams, Ritchie and Bhatia2009). These two kinds of knowledge also differ in the extent to which one can or cannot verbalize them (R. Ellis, Reference Ellis2004, Reference Ellis2005). Our conscious access to EK allows us to verbalize it; ^{Footnote 1} however, because IK is beyond awareness, it cannot be verbalized (DeKeyser, Reference DeKeyser, Long and Doughty2009).

It is currently believed that IK underlies the ability to use a L2 fluently and confidently; therefore, development of IK should be the ultimate goal of SLA (N. C. Ellis, Reference Ellis1993; R. Ellis, Reference Ellis2005; Hulstijn, Reference Hulstijn and Robinson2001). ^{Footnote 2} Although there is solid evidence showing that implicit and explicit representations are neurologically distinct (Paradis, Reference Paradis2009; Ullman, 2011), the interaction between the two and how they influence each other are still subjects of controversy. In SLA research, this controversy is referred to as the interface issue. Central to the interface issue is to what extent explicit learning and instruction impact implicit learning and the development of IK (N. C. Ellis, Reference Ellis2011). There are three positions with regard to the interface issue—the noninterface, strong-interface, and weak-interface positions.

The noninterface position is often associated with Krashen (e.g., Reference Krashen and Ellis1994), who contended that conscious learning about L2s and subconscious acquisition of L2s are two completely different phenomena that result in distinct sources of knowledge (EK and IK, respectively) with no interface between them. According to this position, subconscious acquisition dominates L2 performance, learning can never convert into acquisition, and conscious learning can only be used as a monitor (editor) for performance. Proponents of the noninterface position believe that EK and IK are located in different areas of the brain and are thus accessed by different processes (Paradis, Reference Paradis and Ellis1994). ^{Footnote 3}

The strong-interface position, which is usually associated with DeKeyser (e.g., Reference DeKeyser and Williams2007), offers an opposite view. His strong-interface position should be understood within the models of skill acquisition, such as ACT-R (Anderson & Lebiere, Reference Anderson and Lebiere1998). Within these models, a distinction is made between declarative and procedural knowledge. According to DeKeyser (Reference DeKeyser, Long and Doughty2009), “Declarative knowledge is knowledge THAT something is,” and procedural knowledge is “knowledge HOW to do something” (p. 121). According to skill-acquisition models, learners first develop a declarative encoding, whereby extensive practice is required to ensure that declarative knowledge leads to procedural knowledge. Further practice leads to automatized knowledge, which may not require any conscious processing (DeKeyser & Criado, Reference DeKeyser, Criado and Chapelle2012). This strong-interface position implies a causal relationship between declarative knowledge and proceduralization and automatization of the knowledge. In other words, EK forms a prerequisite for the generation of IK (Segalowitz & Hulstijn, Reference Segalowitz, Hulstijn, Kroll and de Groot2005).

The main claim of the weak-interface position is that EK does not have a causal relationship with IK and only triggers or speeds up the implicit learning process, which subsequently leads to the generation of IK. For example, N. C. Ellis (e.g., Reference Ellis2005, Reference Ellis2008) contended that EK contributes indirectly to the acquisition of IK by promoting some implicit learning processes. According to the author, EK can make relevant linguistic features salient, thus enabling learners to notice them and recognize the gap between the input and the linguistic knowledge they possess (N. C. Ellis, Reference Ellis1994). N. C. Ellis (Reference Ellis2008) further suggested that explicit and implicit processes work in tandem and that there is a dynamic interaction between them for consolidating IK.

The interface issue has been debated for decades and remains important to SLA research for theoretical and pedagogical reasons (Hulstijn, Reference Hulstijn2005). However, dealing with EK and IK constructs and testing rival interface hypotheses is challenging for several reasons. One of these challenges is the lack of reliable and valid measurement tools, which is particularly significant when attempting to measure IK. As both IK and EK sources are involved in L2 performance, it is almost impossible to devise pure behavioral measures of these two constructs (DeKeyser, Reference DeKeyser, Long and Doughty2009; R. Ellis, Reference Ellis2004, Reference Ellis2005). In addition, in constructing measures of IK, the operationalization of the concept of awareness and other challenges may be encountered. Consequently, researchers have been using a variety of imperfect measures (e.g., metalinguistic tests for EK and timed GJTs for IK). For this reason, rigorous validation studies on these measures are required.

Thus far, GJTs have been among the most extensively used measures in research on L2 acquisition. They have been subjected to several validation studies. However, due to several methodological limitations affecting previous studies, the construct validity of GJTs remains open to scrutiny. The following section highlights the limitations of previous studies and explains the motivation for the current project.

Extant Validity Research on GJTs

Previously, it was thought that GJTs provide a direct window into the learners’ linguistic competence. However, it is currently acknowledged that GJTs only provide a measure of linguistic performance (R. Ellis, Reference Ellis, Burmeister and Rounds1991). Many researchers have attempted to establish the type of linguistic knowledge—whether implicit, explicit, or a combination of both—learners draw on in their performance on GJTs. Earlier studies of this type yielded the conclusion that the nature of learners’ knowledge, whether explicit or implicit, affects their judgment on GJTs (Chraudron, Reference Chraudron1983). In other words, GJTs potentially lead participants to draw on both IK and EK, depending on what source of knowledge they mostly have at their disposal (R. Ellis, Reference Ellis2005).

Further research on GJTs aimed to investigate whether manipulating GJT designs leads to differential participant performance. For instance, it was hypothesized that if GJTs only ask participants to discriminate between well-formed and deviant sentences, it is possible that they evoke the use of pure intuition. On the other hand, the use of EK is very likely if GJTs require locating the error and even editing/correcting or describing the rule for judgment (R. Ellis, Reference Ellis, Burmeister and Rounds1991). In addition, it was hypothesized that the kind of knowledge GJTs prompt learners to rely on to make their judgment depends on time conditions (i.e., whether the test is timed or untimed). Time pressure may encourage test takers to respond on the basis of their IK, whereas unlimited time may allow them to rely on their EK (Bialystok, Reference Bialystok1979). Moreover, it has been hypothesized that the grammaticality of GJT sentences may induce differential performance because learners rely on IK and EK for judging grammatical and ungrammatical sentences, respectively (Gutiérrez, Reference Gutiérrez2013).

Several factor-analytic validity investigations tested these hypotheses. This series of factor-analytic studies commenced with the psychometric work of Rod Ellis. R. Ellis (Reference Ellis2005) conducted a psychometric study to design several measures of EK and IK and evaluated their construct validity with respect to EK and IK constructs. For operationalizing the two constructs and distinguishing between them, Ellis proposed seven criteria: degree of awareness, time available, focus of attention, systematicity and certainty, metalanguage, and learnability. He subsequently created five measures: namely, a timed GJT, an untimed GJT, an ON task, an EI task, and a MKT. Based on the seven criteria, Ellis predicted that the ON task, the EI task, and the timed GJT would tap into IK, whereas the untimed GJT and the MKT would evoke the use of EK.

R. Ellis (Reference Ellis2005) submitted his test battery data to exploratory factor analysis (EFA). The results yielded a two-factor structure that confirmed his predictions. More specifically, the ON task, the EI task, and the timed GJT loaded on the first factor, whereas the untimed GJT and the MKT loaded on the second. Ellis labeled the two factors IK and EK, respectively. However, as Isemonger (Reference Isemonger2007) explained, Ellis’s (2005) approach suffered from a few major flaws from a methodological and analytical perspective. As a result, the inferences and interpretations drawn were not well supported. For example, from a methodological and analytical point of view, because Ellis approached the factor analysis with an a priori hypothesis, the use of EFA was not acceptable. Instead, CFA should have been used because the prior hypothesis implied that the measures would measure the distinct constructs of EK and IK.

To overcome the analytical flaws of the approach employed by R. Ellis (Reference Ellis2005), R. Ellis and Loewen (Reference Ellis and Loewen2007) reanalyzed the data used in that study through a CFA. In the CFA, they tested the two-factor model from the original EFA against a decision and production model, as a rival model. This is also a two-factor model, with EI and ON tasks loading on the production, and the two GJTs and the MKT loading on the decision factor. The results yielded by this approach showed that only the first hypothesized model produced adequate fit. However, whereas the CFA approach is indeed a better option than EFA, the authors’ execution of the CFA could have been more thorough.

Confirmatory factor analysis is regarded as a process involving several stages—namely, (a) initial model specification, (b) parameter identification and estimation, (c) data-model fit assessment, (d) possible model modification, and (e) rival model identification (which may jeopardize causal inferences made from the original hypothesized model). However, the research conducted by R. Ellis and Loewen (Reference Ellis and Loewen2007) lacked a final CFA process—an adequate rival model identification. In addition, either alternative models were not specified in the GJT validation studies conducted to date or the tested alternative models were irrelevant to the main purpose of the study. When rival CFA models are not specified, conclusions about specific sets of measures can be highly compromised. As Isemonger (Reference Isemonger2007) explained, “It is important that rival models are tested because the fit of a particular model does not preclude the possibility that other untested models fit better” (p. 109).

The only rival model R. Ellis and Loewen (Reference Ellis and Loewen2007) tested was the decision and production model. The constructs in this rival model were, however, irrelevant to the main constructs of the study. One important rival model that could have provided a rebuttal to R. Ellis and Loewen’s claims is a one-factor model, which enables directly testing whether the measures can actually distinguish the two constructs. Had this model been tested and had it fit the data as well as their two-factor model, it would not be possible for the researchers to conclude that their measures tapped into two distinct constructs of IK and EK.

More recently, Kachinske and Vafaee (Reference Kachinske and Vafaee2014) set out to examine a one-factor model as the alternative model to R. Ellis and Loewen’s (Reference Ellis and Loewen2007) original model. They reanalyzed R. Ellis and Loewen’s data and found that the one-factor model, which also accounted for the method effect (by correlating the error terms between similar tasks), fit the data as well as the authors’ original model. As this finding suggests that the one-factor model is statistically as acceptable as R. Ellis and Loewen’s two-factor model, the construct validity of the latter could not be supported. Kachinske and Vafaee further pointed out similar flaws in the factor-analytic approach adopted in several subsequent GJT validation studies (i.e., Bowles, Reference Bowles2011; Gutiérrez, Reference Gutiérrez2013).

Bowles (Reference Bowles2011) created a battery of five tests of Spanish as a second language by closely following R. Ellis’s (Reference Ellis2005) guidelines. Among other analyses, Bowles conducted a CFA to examine the factorial structure of her test battery and reported results that concurred with those obtained by R. Ellis and Loewen (Reference Ellis and Loewen2007). However, Bowles did not examine any rival CFA models against her two-factor model. Furthermore, the fitted two-factor model indicated high correlation between the two hypothesized factors (r = .87). With such a high correlation between the factors, it is hard to claim that the two factors are distinct. This reemphasizes the importance of testing rival models. If a single-factor model had been tested, results could have shown a good model fit, and the interpretations of the study could have changed.

Gutiérrez (Reference Gutiérrez2013) employed a similar test battery for a different L2 population—that is, Spanish learners in the United States. Gutiérrez followed the guidelines proposed by R. Ellis (Reference Ellis2005) and created a battery of tests consisting of a timed GJT, an untimed GJT, and an MKT. The only test in the battery hypothesized to be a measure of IK was the timed GJT, whereas the untimed GJT and the MKT were considered measures of EK. The author simultaneously examined the role of time pressure on GJTs (timed and untimed) and the types of test items (grammatical and ungrammatical) in order to scrutinize GJTs as measures of IK and EK. It was hypothesized that, irrespective of time conditions, judging the grammatical sentences in GJTs taps into IK, and judging ungrammatical sentences engages EK. Gutiérrez conducted both an EFA and a CFA to test his hypotheses.

In the CFA, Gutiérrez (Reference Gutiérrez2013) tested two rival models. In the first, both the grammatical and ungrammatical sentences of the timed GJT loaded on the construct of IK, and the grammatical and ungrammatical sentences of the untimed GJT and the MKT loaded on the construct of EK. In the second model, the grammatical sentences of both the timed and untimed GJTs loaded on the construct of IK, and the ungrammatical sentences of both types of GJTs and the MKT loaded on the EK construct. The analyses yielded a better fit for the second model, implying that, regardless of time pressure, grammaticality of the stimulus is what distinguishes between the use of IK and EK in performing GJTs. However, Gutiérrez’s study was not free from limitations.

First, having different types of GJTs as the only measures of IK, and a MKT as the sole measure of EK, limited Gutiérrez’s ability to test important rival models. Second, bivariate correlations revealed statistically significant coefficients between the MKT and all types of GJTs, irrespective of stimulus type and time condition. Although the correlations between the MKT and the ungrammatical sentences of both types of GJTs were the greatest, the correlations between the MKT and the other measures cannot be ignored. Therefore, a one-factor model that accounted for method effects could have been a plausible rival model.

In summary, factor-analytic studies, including those conducted by Bowles (Reference Bowles2011), Ellis (Reference Ellis2005), Ellis and Loewen (Reference Ellis and Loewen2007), and Gutiérrez (Reference Gutiérrez2013), suffer from methodological issues, especially the failure to test rival models. This omission in the approach compromises any conclusions about the structural relations among measures and the factors in a model. Therefore, in future studies, several important rival models (e.g., one-factor models) should be tested.

Measuring IK through Online Processing

Psycholinguistic research has revealed how L2 learners process grammatical structures in real time (e.g., Clahsen & Felser, Reference Clahsen and Felser2006; Kaan, Reference Kaan2014). Online grammatical processing has often been examined through reaction time (RT) measures, such as self-paced reading tasks (e.g., Jiang, Reference Jiang2004; Roberts & Liszka, Reference Roberts and Liszka2013) and word-monitoring tasks (Granena, Reference Granena2013; Jiang, Hu, Lukyanchenko, & Cao, Reference Jiang, Hu, Lukyanchenko and Cao2010; Suzuki & DeKeyser, Reference Suzuki and DeKeyser2015). These tasks can examine whether L2 learners are sensitive to grammatical violations while they are reading or listening for comprehension. In the self-paced reading task, for instance, participants read a sentence containing a target grammatical structure word by word, and the RT to each word read is recorded. Researchers examine whether participants slow down to read the word(s) once they encounter a grammatical error. For instance, when participants read a sentence with a subject-verb agreement violation like “The boy in the room enjoy reading many books,” they will slow down when reading the word after the violation (i.e., reading), compared to their performance on the grammatical version of the same sentence (The boy in the room enjoys reading many books). By computing the RT difference between the grammatical and ungrammatical sentences, the sensitivity to grammatical violation can be estimated.

These online sentence-processing tasks are promising measures for tapping into IK (Suzuki & DeKeyser, Reference Suzuki and DeKeyser2015). First, the tasks can capture sensitivity to grammatical violation as the sentence unfolds in real time. This minimizes the possibility that participants access their linguistic knowledge consciously, that is, that they rely on EK (Paradis, Reference Paradis2009). Second, they can direct attention to meaning, as each sentence is followed by a comprehension question asking about the content of the sentence. This second feature contrasts with the design of GJTs because GJTs of any kind draw participants’ attention to form. When learners are instructed to decide whether a sentence is grammatical or ungrammatical, they inevitably pay attention to the form, which potentially invokes the use of EK. Although time pressure makes using EK harder, it does not rule out the possibility that EK is accessed. Therefore, the validity of GJTs can be further scrutinized if online processing measures that draw learners’ attention to meaning are included in the current study.

The present investigations hold promise for scrutinizing the validity of GJTs because recent evidence suggests that online processing measures tap into IK. Suzuki and DeKeyser (Reference Suzuki and DeKeyser2015) compared the word-monitoring task and the EI (a previously attested measure of IK) as measures of IK. The tasks employed in their study—along with a MKT as a measure of EK and a probabilistic serial reaction time (SRT) task, which served as a measure of aptitude for implicit learning—were administered to Japanese L2 learners who live in Japan. The study results showed that the word-monitoring performance was related to the SRT scores only, whereas the EI performance was associated with the MKT scores only. This pattern was found only among the learners who had lived in Japan longer than a certain number of years. This finding suggests that the word-monitoring task can serve as an implicit processing measure among learners with sufficient naturalistic L2 exposure.

The findings reported by Suzuki and DeKeyser also provide some implications for participant selection. Participants in the validation studies may need to have had sufficient naturalistic learning experience. For example, participants in the studies conducted by Gutiérrez (Reference Gutiérrez2013) and Zhang (Reference Zhang2015) were classroom learners with limited exposure in L2 environments. As noted earlier, behavioral measures potentially prompt learners to draw on both IK and EK, depending on what source of knowledge they (mostly) have at their disposal (R. Ellis, Reference Ellis2005). Therefore, more rigorous validation studies on GJTs should at least recruit participants that are typically exposed to naturalistic, as well as classroom-based, learning opportunities.