Application of our framework: How to learn a language

Our framework applied - example

Now, that we have reviewed the scientific literature in the first part, we can apply our framework of how a user can learn a language optimally:

The start

It is probable that before starting the learning journey, it might be advisable to take one introductory class to get the gist of the language. For some languages, this might be less needed, for example if a Spanish speaker wants to learn Portuguese. If an English speaker wants to learn Chinese however, its probably helpful to kick-start the learning journey by learning a bit about what the language is about. For mandarin Chinese, the user would learn about the basics, such as the tones, Hanzi signs, European pinyin letters, and maybe that the grammar is fundamentally different. For this, a class might be appropriate, however, the user might as well watch a lecture online.

From word schema to input with the language schema

After this, the user has to put in the long term work of language acquisition as thousands of words need to be learned. First, the user would learn all the word schemas by obtaining bilingual word pairs in the form of vocab flashcards. The user would then build and extend a personal word bank by constructing and automating his word schemas with the vocab flashcards. To keep the relevancy of the words and enable a fast transition to reading input, the user would sort the vocab flashcards by their frequency in the language. The user wouldn't spend any time learning the pronunciation, spelling, gender or conjugation of the words, instead his goal would be to build only a crude understanding of them. After he has learned some amount this way (maybe 300 or 600 words but this number is arbitrary), he can now finally start his transition into the language schema, or as Krashen calls it, the acquisition of L2, by supplementing his vocab learning with reading text. The point of this transition is determined by the point in time when the user feels ready (as in, the cognitive load is manageable) to try to understand text.

Best form of text input

With only such a limited amount of words, it will still be almost impossible for the user to really understand a text. This means he will struggle with reading a book, because the intrinsic cognitive load would still be far too high. Instead, he has to find text input that is as simple as possible, while still being as authentic as possible and interesting. As Krashen points out, the user should focus on the meaning, as opposed to the form. He adds that input, based on form, quickly loses its authenticity and so becomes uninteresting and it can be assumed, that this also extends to text meant for novices, that is artificially kept simple. The best type of reading input, that fits this need of natural authentic but also simple text, are news articles. Articles are widely available on the internet, where up-to-date articles of any topic can be accessed via their news sites. When comparing the characteristics of all news articles, articles can be selected that are very short, simple in their language, with many pictures and about simple topics.

The case for articles

While a book can be quite long, a short article of about 120 to 250 words will be appropriate for the user. With his limited word bank, he has to use high intensity focus to decipher the written text, similar to how someone, who just learned the alphabet, needs to manually read every letter. This high in cognitive load and taxing activity of reading a simple article will only be for a short time because of the limited text size. This is supported by simple text (no rare words), a simple meaning and extra lingual elements such as pictures.

With the first articles, that the user reads, he will initiate the construction of his language schema, as this is the first time that he sees his learned vocab (word schemas) used in a real world language use case. With this he will first see the gender, relationships and grammar of the words. That he has not learned things, such as verb conjugations, will not prevent him from understanding the meaning. He will see that a verb, he knows, is always slightly different when its together with a noun ('yo tengo' vs 'tu tienes' vs 'el tiene') - the same goes for the gender. While the initial manual deciphering of the text will have a high cognitive load and be tedious, CLT predicts, that as the user continues to read such beginner articles in small chunks, maybe two per day, he will continuously automate his language schema and thus it will be continuously more easy for him to read articles. At this point he will actively build and automate his language schema or, in Krashen's words, acquire the language.

The advanced user

Reading such articles over time will build up the user's language schema. When the user has automated reading simple articles, as in that they don't challenge him anymore (the example article could become boring), the user can move on to more advanced articles, and after those, even more advanced ones, and so on. This will also give more autonomy to the user, as he is less and less dependent on overly simple articles. Instead he can be more picky and read articles that are of direct interest to him. This will boost his motivation, as the motivational section has outlined, because autonomy is an important variable for intrinsic, as well as extrinsic motivation. Krashen also devotes a section to the importance of relevant and interesting input (Krashen, 1982, p. 60). Additionally, if the user has specific intrinsic interests in the culture the language stems from, he will have more incentive to read a bigger amount of articles. By doing this, the user can submerge into the culture by reading regional news and learning about local happenings and events. . At this stage the language schema is already advanced. The user reads text, that is of interest to him, and he doesn't struggle anymore in doing so. It is easier for him to learn new words because his word bank has reached a critical mass. It is big enough that he understands enough words that the context of new words helps with explaining them. In addition to that, learning five new words per article is far easier and less daunting than learning 15 or 25 new words. Now, as the user starts to understand more of the nuance of the language, he can slowly transition into adding more elements to his word schemas. This means that the user can now start to deliberately learn sub-elements, such as grammar and spelling and also use the language in the real world. This transition is, however, not clear cut as it depends on the user himself. He will know how advanced his language schema is, and how ready he feels to take on new challenges. At this stage the user has a very high level of autonomy. As Krashen points out, the job of a teaching system is to some degree over now because its role is to prepare the user to go into the real world and use the language as it is intended to be: "[the classroom's] goal is not to substitute for the outside world, but to bring students to the point where they can begin to use the outside world for further acquisition, [...]" (Krashen, 1982, p. 54, p.68).

Limitation and challenges

Manual synchronization and knowledge management

Unfortunately, this way of L2 acquisition still has a high extraneous cognitive load because the user has to get vocab flashcards and articles. He has to know when to learn new words, when to review and practice words and he might not find sufficiently simple articles immediately.

Furthermore, when starting with reading text, the user will inevitably not know many words, so he either has to ignore them, which makes understanding the meaning harder, thus increasing the cognitive load, or he has to manually look them up, each time he doesn't understand a word. He could also instead write down all the words he doesn't know, look them up, and then read the text again. He could study and deliberately learn such a word list. This was already presented as the highly desirable synergy of Deliberate learning and acquisition. While this is an effective way of learning and acquiring, it is not efficient because the extraneous cognitive load will be very high. The user spends with time tracking and making lists of words instead of actually learning them. As already mentioned, an example for extraneous cognitive load is, when the user has to use his mental energy for finding missing information, instead of using it directly for schema construction and automation. It also seems likely that when motivation is limited, the user might do this effective way of acquisition for a couple of times but then stop because it is too inconvenient. As a consequence, it can be noted that manually syncing Deliberate learning and text acquisition puts a strain on the user's motivation and his mental energy and therefore has its limitations. This means that the user would read text, and if he has the extra motivation and energy, he would sync his vocab learning with his article reading and achieve peak effectiveness in acquisition. If he doesn't have the energy, he would only achieve moderate effectiveness.

Can software reduce limitations? The state of E-Learning

A solution to solve this problem would be to introduce and use software. Software can reduce the extraneous cognitive load by automating all those tasks. Software can provide access to vocabulary and articles, software can asses how much a user knows and based on that, adapt to the user. It can also show the user how much he knows, and it can sync the knowledge between all learning activities. An accurate adaptation of software is also desirable in the sense, that software can estimate the cognitive load of the user. Salden states, for example, in regards to CLT, that when adapting the teaching material to a user, first the user skill has to be assessed, then a dynamic selection of the next learning tasks follows. Salden argues that at this step only at an expert level of this specific task (here the recall of a vocab), the next learning task should be given. When the expert level is not reached, the user should practice (eg. automate the schema) until he becomes an expert at the task (Salden, Paas & Van Merriënboer, 2006; Van Merriënboer & Sweller, 2005). To determine the possibilities and limitations of software for L2 acquisition, we can review the role of existing software:

In the introduction in his mobile app language learning review, Heil (Heil, Wu, Lee, Schmidt & Edu, 2016) states, that today the use of software for language learning has been accepted by users and is widespread, at least for mobile Smartphone apps. Many people use apps to learn a language and there is a large selection of commercial (and free) learning apps available on mainstream app stores such as the Apple App Store or the Android Google Play App Store (Google, 2019; Apple, 2019), as well as a significant market value that corresponds to this demand.

Heil states, while referring to Burston (Burston, 2015), that while in the scientific literature the idea of 'adapted learning' in using computers as personalized teaching machines has emerged, there has been a lack of research in this area (Heil, Wu, Lee, Schmidt & Edu, 2016, p. 33). In turn, while there is a wide amount of theoretical background knowledge about L2 acquisition, the knowledge offered by it, has not been utilized by most and even the best of those language software applications, Heil continues. This appears to be a first hint towards an answer to our initial question: When observing attempts of the digitalization of educational content, often times only a transfer occurs instead of a reorganization that capitalizes on the new possibilities for executing the recommendations of the theory.

I , as if the same disconnect between teachers and theory, that Krashen mentioned, extends also to app developers. Heil, however, distills this disconnect down to the finding that most apps only focus on vocab learning, instead of finding a balance between different aspects (Heil, Wu, Lee, Schmidt & Edu, 2016, p. 34). These aspects correspond mostly to the categories that Elgort described, and this argument by Heil is almost identical to Elgort's argument for a diverse balance of learning activities. The conclusion of the critique is the same: Learning vocab through a learning app has its advantages, and such an app would benefit from being supplemented by more than just vocab - namely by text input that can be read by the user. For all users, except intermediate to advanced users, this is likely to be enough, as only advanced learners benefit from nuanced word schema knowledge.

To get a better understanding of how a good language learning software app could function, we will now review a study that has used one of the most successful and well known learning apps, Duolingo, to compare its use for A1 (beginner) and B2 (intermediate) Spanish learners.

Case study Duolingo

What is Duolingo?

Duolingo is a free language learning software application that can be accessed via its website (web app) or via its mobile app for Smartphones. It offers multiple languages for speakers of many different native languages. According to a study, about 34 hours, on average, of time spent on Duolingo equals the first semester of a traditional language class (Vesselinov & Grego, 2012). Duolingo splits up its vocabulary into lessons. The lessons are organized as a 'tree' (named so by Duolingo). All lessons are not available at once, but instead have to be sequentially unlocked. A user has to complete a lesson in order to unlock the next lesson. The user can unlock all lessons up onto a certain benchmark by completing a placement test. Duolingo determines how well a user knows a word with its algorithms. It uses a Spaced Repetition System to decide when to review which words (Duolingo, 2019).

While Duolingo offers some explanation about the language (such as for grammar) and has forums for general topics and discussions for specific words and phrases, besides this, it mainly has a focus on learning vocab. This was also the main critique point when Krashen reviewed Duolingo: Duolingo, Krashen noted, only focuses on learning vocabulary, but not on language acquisition (Krashen, 2014).

Simple sentences - Duolingo's strong point

Duolingo's teaching methods for vocabulary resemble the traditional practices of writing, listening, speaking and multiple choice. However, one aspect that differentiates Duolingo is that it gives users short and simple sentences to translate. An example would be: 'the cat drinks water', which the user would translate to 'el gato bebe agua'. Duolingo of course also shows the correct answer and where mistakes were made. This technique of translating simple sentences might enable better retention and a good balance between learning new and the review of familiar words. CLT would confirm this: Automated word schemas are further automated and combined with new word schemas. These new word schemas are easier put into the language schema as the user can see in which context they belong. Moreover, CLT states that a high variability of problems can also benefit the user: "learners were [...] able to deal with, and profit from, the germane load imposed by high variability of practice problems" (Paas, Renkl & Sweller, 2004, p. 3). At the same time the sentences are kept artificially simple. The overly simple structure of sentences allows the user to quickly understand these basic structures. If the same sentence reappears, but with different words (eg. 'the man drinks tea'), the user can focus more on vocab recall rather than the sentence and its grammar structure. A simple grammar structure thus further helps the user transition into reading input text, as it reduces complexity and allows the user to get familiar before encountering complex real world sentence structures. Moreover, translating the whole sentence as opposed to filling out a blank (missing word in a given sentence) seems beneficial because the user cannot only review more words at once, he also doesn't need to mentally switch and think about what response is required for the blank spot of the sentence. Instead he can focus on writing the simple sentence and if he makes mistakes, he is notified and can attempt to improve on them with the next sentence or time the word appears. The sentence translation exercise thus directly benefits the transition from learning word schemas to language schema acquisition.

Mix of activities and violation of autonomy - Duolingo's weak point

The same can, however, not be said for Duolingo's other types of exercises of listening, speaking, filling blanks, multiple choice, and so on - for the same reasons that was already elaborated on when discussing Heil's and Elgort's expression for the need for a more balanced and diverse range of exercises.

Moreover, Duolingo's restriction on lesson availability forces users to complete previous lessons for a skill level that Duolingo mandates, even if they don't want to do so. As already mentioned, Salden stated in relation to CLT that a user should reach expert level at a task before being presented with the next task, so Duolingo is in line with the theoretical background. While this might stem from a good intention, it takes away the learner's autonomy which is, as already explained, detrimental for both intrinsic and extrinsic motivation.

Duolingo also offers a section called 'words' in which users can see all the words that they have encountered in the app along with a value on how well they know the word. A different color is shown depended on the skill for each lesson, and a small line chart depicting their earned learning points called 'XP' that they have earned during each of the last seven days. Besides this, Duolingo offers no further knowledge management or visualization tools for learned vocabulary.

Evaluating Duolingo Stories

Recently Duolingo added a Stories and a podcast section. The podcast section is the, to natural input, closest feature Duolingo offers, even though it is mostly audio but also has a transcript. However, the podcast has no real difference to other ordinary podcasts. While the transcript of these podcasts offers natural text, it is not integrated with their vocab learning in any way. It is just plain text independent of the other parts of the platform. That makes its use case identical to the wide selection of article texts on foreign news sites or blogs, with the only difference being that articles are generally not targeted at language learners, which can make them more authentic to the user. The stories offer simplified small stories that are written and spoken out in parallel. They also require the user to solve simple exercises. This partly responds to and averts Krashen's critique as it focuses on comprehensible input. From the view of CLT, this is also on par as it simplifies the complexity of real world text and heavily aids the user. Duolingo Stories thus further help with the transition into real world text. This also has its limitation, as Duolingo stories are artificial and simplified (with a similar character as stories in language textbooks). Therefore, they cannot fully replace the natural text input required for building the language schema. Additionally, it seems likely that Krashen would criticize them as they run the risk of focusing more on the form than on the meaning. As presented in the section about Krashen's theory, he suggests that a mix between grammar and meaning is difficult if not almost impossible because a story or a text built around conveying a from, structure or grammar quickly looses its authenticity. This can result in the user losing interest. This risk is amplified by the exercises integrated in Duolingo Stories and further amplified because the exercises are of different types. While my previous critique of mixing exercises applies, here the cognitive energy is further split. Translation of unfamiliar words already takes mental energy away from focusing on the meaning when reading a text (and with that building the language schema), and doing active exercises that have right and wrong answers increases this problem. However, if the stories are simple enough, and the exercises are integrated in a more streamlined way, meaning that they don't distract from focusing on the meaning (if that is possible), they might be a highly desirable tool for transition to reading natural text. Unfortunately, Duolingo doesn't offer the user a way to synchronize the stories with what the user is learning even though such a synergy between learning and acquisition would be desirable.

Student's interactions with Duolingo - two groups

In his study, which was conducted before the release of Duolingo Stories, Munday (Munday, 2016) compares two groups of his English speaking university students who learn Spanish in university and their mandatory complementary usage of Duolingo. The first group consists of complete Spanish beginners who are at the A1 level. The second group consists of advanced Spanish learners who were in an 'advanced intermediate course' at the B2 level. In response to Krashen's critique of Duolingo, Munday notes that he agrees that Duolingo is only for learning vocabulary and not for language acquisition. In contrast to that, Munday states that he doesn't view Duolingo as a class replacement but rather as a homework supplement. Thus, he only sees Duolingo as a tool to learn vocabulary and not a solution to all learning needs (Munday, 2016, p. 85, p. 89).

In his study, all students took the placement test from Duolingo and while all beginners where placed at the start of the lesson tree, the skill level of advanced B2 Group had a high variability which resulted in them being placed at different stages of the lesson tree. The students also had to complete five lessons a week as they got a higher grade if they completed the lessons spaced out over time instead of all at once. Their Duolingo performance influenced their final grade by 10% (Munday, 2016, p. 91). Munday wanted to incentivize his students to study in small chunks, and while most students had a positive attitude towards the use of Duolingo over traditional textbooks, these restrictions of five lessons per week were one of the student's biggest frustrations (Munday, 2016, p. 94). This appears to be a good example of the misunderstanding teachers have when using software such as Duolingo.

For grading language students, their end result or competence at a certain time is more relevant than how they achieved it. When designing the most optimal way of L2 acquisition, the method employed, in this case learning the material spaced out or all at once, does matter. When using software to help with learning, the design of the software should incorporate this behavior but it should not force it. And if the software allows for all words to be learned at once, then the teacher should accept this and not interfere. Instead, the software could be changed, as in this case the software could prioritize familiar words over new ones. In the end, the user's autonomy should be maintained. In practice this means to offer the user a 'golden path' but still allowing him to deviate from its path, to change the software settings to his own preferences and allow him to customize his learning experience. It is thus unsurprising that the student's main frustration came from the requirement to complete five lessons per week, as this interfered with their autonomy of time managing their learning efforts. In addition, students also complained that their required Duolingo lessons and their class lessons where not in sync (Munday, 2016, p. 96). This seems like a valid reason why one might require students to complete a certain number of lessons: So that the contents of those lessons can be used in class which would utilize the synergy of deliberately learning vocab and then using it in real world cases. This problem of out of sync offline to online education seems like a comparatively trivial problem to solve with a high benefit: Either offer custom lessons in the online software that correspond to the material in class or adjust to material in class to correspond to the material of the online software.

The results of Munday's study show that almost all beginner students liked and enjoyed using Duolingo for their homework. In contrast, the attitudes towards Duolingo were much more mixed for the advanced B2 students. This in line with previous conclusions: Beginners enjoy Duolingo to a high degree because Duolingo focuses on learning vocab. It is thus optimal for people who are just starting to learn a language and build a first vocabulary base. Duolingo helps the users build their words schema, and paired with their simple sentences, does so efficiently. As users get more advanced in their language schemas, they become more limited by Duolingo: Not only does Duolingo violate their autonomy, Duolingo only enables them to learn vocabulary, as Krashen notes (Krashen, 2014). It doesn't directly help them with language acquisition because it doesn't offer natural input, even with its stories, as they are artificial input and are just an aid for the transition to natural input. While Duolingo is a good tool for learning vocabulary (word schemas) it doesn't do anything beyond this, which might be why advanced users are less impressed.

Mobile vs desktop and handwriting vs typing

Something that the students in the study enjoyed was the mobile aspect of Duolingo. They could do their homework when and wherever they wanted and because of Duolingo's gamification elements, it didn't feel like homework but more like playing a game, the students reported (Munday, 2016, p. 94). And while the majority of time spent on Duolingo was with Duolingo's mobile app, the web based version still received a considerable amount of traffic (Munday, 2016, p. 95).

While the advantages of mobile independence seem promising, it is not clear if learning and acquisition primarily focused on mobile use is beneficial in the long term. To maximize the learning and acquisition output with a given input of time, motivation and cognitive energy, it seems that mobile may have disadvantages and a desktop approach might have benefits: First, while learning on the go, risk of distraction is high. And secondly, the interface of a Smartphone has its limitations because the screen is smaller and and usually no physical keyboard is available. These problems might be less significant, but they get countered by the advantages of a personal computer environment: Such an environment resembles a working station for deliberate, productive and focused work to a higher degree. Furthermore, a bigger screen enables a different user interface that makes it easier to do a given activity, but also helps to find and display additional information. One example of this is that up until recently, Duolingo made its forum and exercise discussion boards available only on their web page, but not on their mobile app - and Duolingo Stories is still not available on mobile. Finally, the physical keyboard allows for efficient input. While traditional handwriting is better for memory recall than typing, typing via keyboard still provides some benefit for memory recall (Smoker, Murphy & Rockwell, 2009). Moreover, keyboard typing is faster than handwriting: In a study that compared both, keyboard typists where slightly faster with 36 words per minute (wpm) and 1.6 errors, compared to 31 wpm and 0.1 errors of hand writers (Brown, 2017). Error frequency is higher for keyboard typing which indicates that when using typed translations for word recall, it is probable that spelling errors are not due to wrong recall but simply due to the input method. Another study found that with practice an average of 60 wpm was reached (Clarkson, Clawson, Lyons & Starner, 2005). In this study, only mini keyboards were used (found commonly on old Smartphones), so 60 wpm might be a low estimation. Faster input via keyboard is in line with the emphasis on quantity rather than quality when learning vocabs and also opens up the possibilities of data analysis because typed in answers can be saved. To enable this also for handwriting, a method would be needed such as taking a picture of the writing and then recognizing the letters via software - a process that seems unpractical.

The advantage of mobile seems more specific in that users can space out their learning more and increase their immersion through frequent use. Especially if a user is more advanced and thus more focused on schema automation instead of construction, it seems appropriate to do a light review on words or to read an article that only has a low cognitive load because the user is already familiar with most of its words.

Conclusion on the role of software

While Duolingo can be seen as one of the market leaders and innovators of language learning software, it also has many limitations such as a lack of natural text input. These limitations can be countered by one of the many other available language learning apps that might have a bigger focus on reading natural text. However, by doings this, the user loses synergy effects because he has to use different apps that don't share his activity on them. An automation of synergy between learning words and acquiring language through reading text is thus lost. Subsequently, it would be desirable, if Duolingo, or any other app, would combine these activities and adapt them according to the user behavior.

Our established framework to learn a language benefits from language learning software. The general market trends, as well as the example show that this is not only accepted and effective, but also shows the already existing potential as well as the future potential that could be achieved by synergising deliberate vocab learning with natural comprehensible text as input, all while offering the user a 'golden path' but at the same time maintaining his autonomy.