The explicit/implicit dichotomy has been a core concept in SLA (Andringa & Rebuschat, 2015; Ellis, 2009; Hulstijn, 2005) since the distinction between acquisition and learning was made (Krashen, 1982). Language acquisition refers to the process in which the learner acquires the target language subconsciously; language learning, on the other hand, is related to the learner knowing about the target language. Krashen and Terrell (1983) have proposed “the natural approach”, according to which the role of L2 instruction is to provide comprehensible input for language acquisition to take place while providing learners with metalinguistic knowledge is ineffective and can impede L2 development. Opponents of the natural approach, however, argue that providing L2 learners with comprehensible input alone may not suffice for desirable levels of attainment in the target language (Hammerly, 1987; Harley, 1989). Students receiving abundant L2 input may attain a high level of fluency in the target language, but continue to make grammatical errors. Moreover, communicative interaction, input processing, comprehensive output, and feedback may all have an important role to play in the language classroom, because they contribute to the probability of learners noticing and internalizing the linguistic structures (Gass et al., 1998; Lightbown & Spada, 1990; Pica, 1994; Swain, 1985). Meaningful input, then, mainly serves to create a context where intervention facilitates language acquisition. Therefore, L2 instruction should concentrate on providing enough opportunities for the “external manipulation of learners’ focal attention” (Norris & Ortega, 2000, p. 420).

However, how “explicit” instruction should be in order to promote noticing and foster L2 development has been repeatedly debated. Scholars provided several definitions to distinguish different styles of L2 instruction. Long (1991) categorizes instruction into three types: Focus on Meaning, Focus on Forms, and Focus on Form. A Focus on Meaning instruction is compatible with Krashen’s acquisition hypothesis, insisting that L2 acquisition takes place incidentally in the practice of learners being exposed to authentic and comprehensible input. In contrast, Focus on Forms deals with linguistic features in isolation; learners, with their attention directed to forms, learn the constructions externally and systematically. Finally, Focus on Form integrates several properties of the former two. Here, language is learnt in communicative contexts in which meaningful input is provided while the learners’ attention is directed to certain linguistic forms which may hinder L2 progress. This categorization provides a well-founded theoretical framework for studies on L2 instruction; however, in a real language classroom, it is impractical to solely follow one type of the three.

Spada (1997) refers to a variety of L2 instruction approaches as “form-focused instruction”, which occur in a meaning-based context with the attention drawn to linguistic forms “in either spontaneous or predetermined ways” (p. 73). In essence, Spada’s terminology includes both Focus on Forms and Focus on Form, and the two instructional styles are further distinguished as “isolated” and “integrated” form-focused instruction respectively (Spada & Lightbown, 2008). Form-focused instruction thus became an umbrella term through which a broader range of instructional activities can be included. For the convenience of ISLA studies, however, more concise terms – explicit versus implicit instruction – have been proposed and used. Explicit instruction requires focal attention to forms and provides relevant explanations of the grammatical rules whereas implicit instruction provides forms without discussing them further (Hulstijn, 2005). The concepts of L2 instruction have evolved in the past decades, together with an increase in empirical studies that attempted to compare the effectiveness of various instructional approaches.

Given the variations on the operationalization of instructional effectiveness in SLA studies (Norris & Ortega, 2000), it is not surprising that mixed results have been found. Norris and Ortega (2000), for instance, examined the effectiveness of explicit versus implicit L2 instruction by conducting a meta-analysis in which 77 research reports were selected after extensive evaluation. They classified the instructional treatments as being explicit or implicit and adopted Cohen’s d to calculate the effect size for comparison. The results show that explicit instruction displayed a clear advantage (d = 1.13) over implicit instruction (d = 0.54), with further confidence intervals qualifying the difference was valid. Nevertheless, the authors have noted that the findings should be interpreted with caution because of biased testing measures of the learning outcomes, brief interventions, and the methodological variance among the reviewed studies.

Following Norris and Ortega’s (2000) work, more meta-analyses were conducted to update the knowledge in ISLA. Spada and Tomita (2010) examined 30 empirical studies and aimed to compare the efficacy of explicit and implicit instruction on learning simple and complex L2 structures. Goo et al. (2015), on the other hand, investigated 34 studies published from 1993 to 2011 and paid special attention to the potential factors that might mediate instructional effects. Both studies endeavored to compare the effectiveness of explicit and implicit instruction and suggest that explicit instruction works better than implicit instruction. Yet, the conclusion seems premature because the empirical studies reviewed in these two meta-analyses were hampered by the same problems pointed out earlier by Norris and Ortega (2000): Measures favored explicit knowledge and the treatments were brief.

Such bias has been avoided in recent studies conducted in the Netherlands. In a three-year investigation, Rousse-Malpat et al. (2019) compared the effects of structure-based (i.e. explicit) instruction and dynamic usage-based (i.e. implicit) instruction on secondary learners’ acquisition of L2 French. The structure-based instruction included a traditional textbook with grammatical explanations, whereas the dynamic usage-based treatment avoided grammatical rules but used the Accelerative Integrated Method (“What is AIM”, n.d.) which provided storytelling, music, and gestures to facilitate understanding. Free-response written data were analyzed and the results show that usage-based instruction leads to higher linguistic complexity both in morphosyntax and lexicon. These findings correspond to an earlier and similar study conducted by Rousse-Malpat and Verspoor (2012). They conducted a two-year investigation to compare instructional effects of Focus on Meaning and Focus on Form through free-response oral data. The study shows that meaning-focused instruction helped the learners to develop equal accuracy, but higher fluency, in their L2 French oral production.

Another longitudinal study was carried out by Piggot (2019), who investigated the effects of delaying and reducing explicit grammar instruction to adolescent Dutch learners of English. In this study, two cohorts of students (N = 222 for the explicit group; N = 241 for the implicit group) were observed during their first two years of secondary education. The explicit group were taught with a communicative language teaching approach in which explicit rule explanation and grammar drills were included; the implicit group were taught by the same teachers with the same textbooks, but all pages with explicit rule explanation and drills had been eliminated for two academic years. Piggott’s findings are in line with the evidence documented in the abovementioned Rousse-Malpat and Verspoor (2012) and Rousse-Malpat et al. (2019) studies: Learners receiving implicit instruction show a more complex and fluent L2 performance than their counterparts do both in speaking and writing.

Thus far, recent longitudinal investigations seem to argue for advantageous effects of implicit instruction on promoting complex and fluent L2 writing performance. Moreover, in terms of overall writing quality and accuracy, long term investigations did not find explicit instruction to be more effective, which is at variance with the conclusions drawn from previous meta-syntheses (e.g., Norris & Ortega, 2000; Spada & Tomita, 2010). As suggested in these meta-syntheses, length of intervention and outcome measures might have largely contributed to the different findings in instructed SLA studies.

Regarding the length of intervention, Ortega and Iberri-Shea (2005) argue that as L2 development takes time, longitudinal research can tap into issues that cannot possibly be solved by short-term observations. For example, longitudinal studies can provide a lens to identify developmental characteristics at varying stages of L2 development, which facilitates learners in attaining desired learning outcomes across different contexts (Ortega & Byrnes, 2009). Ortega and Iberri-Shea (2005) further suggest that the length of observation should be scaled on institutional time. As far as measurement is concerned, learning achievement has often been expressed in terms of metalinguistic knowledge with constrained tests (e.g., grammaticality judgement tests). Metalinguistic knowledge can be learned quickly and lends itself well to short-term interventions. Constrained tests thus have been widely criticized for being biased and free-response tests have been called for (Norris & Ortega, 2000). In fact, empirical studies that have used free-response data did not find explicit instruction to be more effective than implicit instruction (Macaro & Masterman, 2006; Piggott, 2019; Rousse-Malpat et al., 2019, Rousse-Malpat & Verspoor, 2012). Taken together, long-term interventions and unconstrained tests are required in order to obtain more robust findings in L2 type-of-instruction research.

As L2 proficiency is a complex construct and multi-dimensional, different measures provide different yet equally meaningful insights into L2 development. Holistic human scoring presents a rounded view of learners’ language competence in that a human mind can weigh different strengths and weaknesses of a text simultaneously and has been considered very informative in L2 studies (Hou et al., 2016; Polio, 2001; Verspoor et al., 2012). Meanwhile, the multidimensionality of L2 performance can be investigated by looking at separate sub-components (Ellis & Barkhuizen, 2005; Skehan, 1998). For example, quantifiable measures tapping into the complexity, accuracy and fluency (CAF) of L2 performance have long been recognized and used in SLA research (Wolfe-Quintero et al., 1998). Complexity, from a linguistic perspective, concerns mostly syntactic and lexical constructions; accuracy concerns the degree of conformity to target language norms; fluency concerns the speed of L2 processing in given contexts (Abdel Latif, 2013; Bulté & Housen, 2014; Housen & Kuiken, 2009; Larsen-Freeman, 2009). CAF measures have been used both as descriptive guidelines for learners to assess their language proficiency and as indices to gauge L2 progress (Housen et al., 2012). In the current study, both holistic ratings and CAF measures were employed to compare the learners’ written production under explicit or implicit instructional conditions.

In terms of which specific CAF measures are reliable and suitable for our learners, references were taken from Verspoor et al. (2012). This is because our participants were similar to the L2 learners investigated in Verspoor et al. regarding the learners’ L1 background and educational context. Verspoor and colleagues contribute to a comprehensive evaluation of 64 measures to assess L2 writing development ranging from beginner to intermediate levels. Among these measures, we chose mean length of T-unit (MLT) and ratio of dependent clause per T-unit (DC/T) to investigate broad features of syntactic complexity (pp. 260–261), the ratio of Present Simple Tense (%PST) to measure verbal complexity (p. 256), and the Guiraud index (“a Type Token ratio adjusted for text length”, p. 252) to measure lexical richness, given that these measures demonstrated significant correlations to holistic writing quality.

Apart from the abovementioned broad measures, we especially endeavored to investigate writing accuracy. After two years of explicit versus implicit instructional treatment, we expected to see different competence in tense/aspect related verbal accuracy. Previous research has revealed that learners at low levels may exhibit different developmental patterns in their mastery of the contextual use of tense/aspect and the corresponding inflections of the form (Rousse-Malpat et al., 2019; Spoelman & Verspoor, 2010; Verspoor et al., 2012). We thus operationalized two measures to gauge tense/aspect related verbal errors, of which one looked at the contextual use of the tense/aspect (hereafter, verb use) and the other the inflectional form of the verb phrases (hereafter, verb form). In addition, we also investigated other types of errors, namely, lexicon, grammar, word order, spelling, mechanics and punctuation errors (see Verspoor er al., 2012, p. 253).

Lastly, we used text length to measure fluency. Writing fluency is a complex construct that is difficult to measure (Abdel Latif, 2013). An anonymous reviewer pointed out that text length might not be a sound measure of fluency unless the writing task explicitly asked the writer to write as many words as possible. Indeed, a writer might invest more time into coherence and cohesion, or word choice, instead of writing a longer text. However, as text length clustered together with words per minute in a factor analysis of argumentative essays by Oh (2006; reviewed in Norris & Ortega, 2009), we argue that text length can still be an informative variable under a timed test condition to gauge fluency (Jarvis et al., 2003). Compared to a shorter text, a longer text indicates that the writer is likely to have L2 resources at their disposal more readily and is better able to manipulate these resources for text-based communication (Skehan, 1998).

To sum up, the present study intends to revisit the efficacy debate of explicit versus implicit instruction by following intact classes and applying a free-response writing test to compare the long-term learning outcomes. The data used in this study was taken from a larger project which was previously reviewed (i.e. Piggott, 2019). Previously, it was found that after two years of intervention learners receiving implicit instruction showed steeper growth in writing complexity and fluency than their counterparts did both in speaking and writing. But the impact on accuracy, especially on the tense/aspect related competence in writing, remained unclear. This was of special interest given that tense/aspect development was placed at the center of the form-focused curricula and that a great deal of time and effort was spent in rule explanation and grammar drills in the explicit teaching condition (Piggott, 2019, p. 70). As our investigation required intensive manual coding of the written texts, we took a subset of the data to reduce the workload: the texts written by learners from the highest scholastic competence stream. The reason for choosing this subgroup was that more proficient L2 writers would use a wider range of linguistic features compared to less proficient ones, and thus we expected to see more variation in the written production produced by students from the highest scholastic stream.

The overarching research question is: After two years of intervention, does explicit instruction lead to better writing proficiency than implicit instruction in terms of overall writing quality, linguistic complexity, accuracy, and fluency? Based on previously reviewed meta-syntheses and longitudinal studies, we hypothesized that:

The participants involved in the present study were two cohorts of students aged 12 to 14 who entered the same secondary school in two successive years. Students who entered school in 2014 comprised the explicit group (N = 55), whereas the 2015 cohort formed the implicit group (N = 59).¹

The participants’ leaving reports for primary education showed that their English proficiency was approximately at A1 to A2 level according to the Common European Framework of Reference (CEFR) before entering secondary education, i.e. at the start of the current period of observation. At the beginning of the observation, a reading and a listening test were administered in class and scored by the English teachers in the participating school, the results of which were used to examine the participants’ English proficiency before the observation.

Both groups were instructed within a communicative language teaching framework using the same course books: More! (Puchta & Stranks, 2008) for the first year and Activate! (Boyd & Barraclough, 2010) for the second year. While the explicit group received instruction that contained both meaning and explanation of grammatical rules, the implicit group was instructed without rule explanation. For the implicit group, the grammar-oriented pages in the textbooks had been torn out and grammatical explanation, drills, and explicit feedback about grammatical rules were avoided both in class and after class. In other words, the instructional difference between the two groups lay in whether metalinguistic knowledge was provided. The groups had the same amount of classroom time (i.e. same amount of instruction) and were both taught in English only by the same teachers. While the explicit group invested time in learning grammatical rules, the implicit group spent time on meaning-focused textbook exercises (see Piggott, 2019, pp. 70–71 for more detail).

Towards the end of the observation, which was in 2016 for the explicit group and in 2017 for the implicit group, the participants were tested on their writing proficiency. The writing tests were organized as formal in-class examinations, in which the students were given 50 minutes to write a letter about their latest holiday experience. The required text length was no less than 140 words. The participants were allowed to write as long as they could within the given time. Dictionary use was not allowed. As tense/aspect had been the grammatical focus in the curriculum, the writing test was designed to elicit the use of varied tense/aspect forms. The exam instructions are given in Appendix 1. The hand-written texts produced by the learners were collected as raw data.

All the hand-written texts were transcribed by the first author and checked by the second author. The greeting and closing parts of the letters were deleted as they were standard formulae. In order to conceal private information (i.e., names and addresses) and to avoid overrating lexical richness, we replaced proper nouns with Name or Place wherever that was applicable (cf. Verspoor et al., 2012). Exceptions were country names (e.g., France) and names of capital cities (e.g., Paris) which were kept because they were part of the learning objectives in the curriculum. Finally, texts were converted by the transcription conventions of the Codes of the Human Analysis of Transcripts (CHAT); accordingly, redundant Dutch words and data that could not be recognized by CLAN (MacWhinney, 2014) were cleaned. Such cleaning is listed and explained in Appendix 2.

We employed six trained raters (two professors and four master students of Applied Linguistics) to holistically assess the text quality with a rubric used by Hou et al. (2016). This rubric included not only complexity, accuracy, and fluency, but also idiomaticity and coherence (the CAFIC rubric, see pp. 90–91). We chose this rubric because it could reflect L2 writing quality from a broader linguistic scope.

The training procedure for rating was similar to the one used by Verspoor et al. (2012) and the CAFIC rating scale was used for the holistic scoring. To allow raters to discover for themselves what characteristics contributed to weaker or stronger texts, twelve texts of varied writing quality were selected and used as samples in a trial session. The raters read three texts at a time and judged independently which text was the weakest and which the strongest in each set. Disagreements were discussed until rater agreement was reached. The raters continued to cooperate closely until the 12 samples were ordered according to quality and assigned a score from 1 (lowest) to 5 (highest). These samples served as the benchmarks for rating the remainder (see Appendix 3).

After the trial, the remainder of the samples were randomly divided into three piles and each pile was rated independently by two raters. A valid score was established when a consistent score was given by both raters. In case of discrepancy, a third rater was consulted in order to resolve the disagreement.

Several CAF measures were adopted to investigate the writing production from multiple dimensions (see 2.3 for the rationale behind the chosen measures). For complexity, we employed four measures: the Mean Length of T-unit (MLT), the Ratio of Dependent Clauses per T-unit (DC/T), the ratio of the presence of Present Simple Tense (%PST) against the tokens of all tense/aspect, and the Guiraud index. For accuracy, we counted the number of errors per 100 words: A high error ratio means low accuracy. We analyzed three types of error ratios, namely, tense/aspect related use and form error ratios and a non-verbal error ratio (i.e. errors irrelevant to tense/aspect). Regarding verb use/form errors, five types of tense/aspect which were used in the learners’ writings were included for analyses: Present Simple, Present Continuous, Present Perfect, Past Simple, and Future Simple. Regarding non-verbal errors, six types of errors were included: lexicon, grammar, word order, spelling, mechanics, and punctuation use (Verspoor et al., 2012). For fluency, text length was employed as a measure of writing fluency. The operationalization of the eight CAF measures is summarized in Table 1 below.

Except for text length which could be automatically counted in CLAN, most measures required manual annotation before useful statistics could be extracted and calculated. T-units, correct T-units, clauses, use of tense/aspect, and types of errors were manually annotated in CHAT format, automatically counted in CLAN, and then exported to Excel for calculation. For verb use errors, we coded the errors when the tense/aspect was overused. For example, when the Present Simple was used to talk about the past, it was coded as a Present Simple use error. For verb form errors, we coded the incorrect inflection of the verb phrases based on the writer’s intended tense/aspect choice. Coding examples are given in Appendix 4.

The first two authors coded the texts together. The coding conventions were first discussed and agreed upon by the coders; subsequently, a collaborative trial coding of ten texts was conducted to train the coders and resolve coding discrepancies between them. The remainder was divided and coded independently by the two coders. To ensure inter-coder reliability, all independent coding was examined by the other annotator and disagreements were discussed and resolved.

Due to the normality issue and the scale issue of some datasets, Mann-Whitney U tests were run to compare the two groups for the pre-intervention listening and reading scores, the holistic rating scores, and the three error ratios. A one-way multivariate analysis of covariance (MANCOVA) was run to compare the writing complexity between the two groups, with the pre-intervention listening and reading scores as covariates. A one-way analysis of covariance (ANCOVA) was run to compare the text length between the two groups, with the pre-intervention listening and reading scores as covariates. Because the Mann-Whitney U test was run six times, the α level was adjusted to .008 (0.05 divided by 6) for that test. All statistic tests were conducted with the statistic software SPSS version 26.