Discussion

In this paper we discussed the purpose and philosophy behind the development of LANGA, a game-based platform for L2 learning, and presented preliminary evidence for its effectiveness. The purpose of the study we present here was to test whether LANGA is an effective tool to learn an L2 vocabulary of 72 Spanish words, whether and how different training strategies lead to different behavioral performance and to what extent they engage brain systems typically involved in the processing of native language.
In this experiment, subjects learned 12 novel words/session, with each session lasting approximately 25 mins for a total of approximately 3 hours of training. Training was distributed over 6 days and, at the end, subjects demonstrated statistically significant improvement in performance as measured by a naming and forced-choice task. Interestingly, the acquisition of L2 vocabulary was accompanied by changes in the amplitude of the N400, which was observed after but not before learning Spanish words. This indicates that  stable neural representations had been successfully formed for the newly learned words.
Group-level analysis showed a significant decrease in accuracy only in the naming task over the period of no-exposure, while performance in the forced-choice task and N400 amplitude did not show statistically significant changes.
Nonetheless, a closer look at N400 amplitude over the different stages of the experiment (pre-training, post-training and two weeks follow-up) calculated in the time window from 300-500 msec post-stimulus onset from the electrode Cz. Barplot represents the mean across subjects, while the bars represent 95% CI. Note: vertical axis is reversed reveals the presence of an opposite trend in N400 amplitude changes induced by different training strategies between post-training and follow-up. In fact, compared to post-training, follow-up testing showed an increase in the N400 amplitude for words trained through rote memorization, and vice-versa for words trained through the inductive strategy. A possible interpretation for this trend is that items trained using the inferential strategy are more susceptible to forgetting, presumably because of a less efficient encoding during the training phase. Nonetheless, at present we can only speculate about the mechanisms characterizing the differential effect of training strategy, and further studies employing larger sample size and more complex statistical analysis to better model the high-dimensional spatio-temporal dynamics of brain activity are required to support our hypothesis.
In addition to it, investigation of mechanisms characterizing retention of the trained material should assess behavioral and neurophysiological performance after longer periods of no-exposure, in order to resemble the typical periods of little or non-use frequently encountered after classroom-based L2 learning. In this regard, two recent studies have investigated the effect of training strategy (explicit and implicit, whose principles inspired our rote and inferential training methods) on L2 grammar proficiency immediately after Morgan-Short 2012  and over a period of 3-6 months after the end of training Morgan-Short 2012. These studies showed that a relatively short training schedule (3 training/practice sessions over a period of 9 days) resulted in significant gains in L2 grammar proficiency after training, and that these gains were maintained after a relatively long period of no-exposure, independently of teaching strategy. Interestingly, despite no changes in performance were observed over this period, analysis of EEG/ERP highlighted changes in spatio-temporal patterns of brain activity wherein both implicit and explicit training promoted a shift towards brain mechanisms resembling those seen when processing L1 syntax. Also, this pattern was more evident for implicit than explicit strategy before and after the period of no-exposure, suggesting that the training strategies differently engage neural systems underpinning language processing. Given that our study focused on teaching vocabulary, it it possible that an explicit or implicit approach to learning results in different patterns of brain activity compared to the ones observed for the processing of syntax. Clearly, a better understanding of how a particular strategy influences the mechanisms underlying processing of different aspects of language will help us designing training methods that, in the long-term, will hopefully result in more effective learning.
Despite the promising results, with this study we were only able to gauge the efficacy of  LANGA on a restricted and relatively homogenous sample of participants, limits our ability to generalize our results to the general population. Since one of our overarching goals is to develop a platform that can optimally adapt to the unique abilities of any individual, we are currently moving from this stage to a deeper investigation of which factors (i.e. education, lifestyle, neurocognitive factors etc.) predict individuals' ability to learn and retain L2 material. To this end, a further avenue we want to better investigate is whether models that are based on behavioral outcomes can be improved by integrating real-time analysis of EEG activity to predict how  to adapt the training parameters such as, for example, the intensity of exposure of the items “at-risk” of forgetting. This is based on the assumption that there might be EEG features (including, but not limited to, the N400 component) that predict later retrieval/forgetting of specific words or grammatical structures. At present, we have little evidence to inform the design of such a brain-driven adaptive system, in particular because studies in the domain of L2 learning and processing generally focus on changes in brain activity from pre- to post-training, while patterns of ongoing activity measured over the course of learning have rarely been studied using high-temporal resolution neuroimaging such as EEG or MEG. Our approach presents several additional challenges to traditional methods, since features that are relevant to predict language outcomes from the online EEG activity might reflect complex neural dynamics in both spatio-temporal and spatio-frequency domains. This results in a high-dimensional problem that not only requires a large amount of data, but also the use of advanced statistical methods such as Generalized Additive Mixed Methods (GAMM), and powerful machine learning techniques such a conditional inference Forests (cForest).
At the same time, disposing of a web-based system represents a unique opportunity to conduct empirical studies that will farther our understanding of the mechanisms mentioned above and, in turn, inform the design of next generation of platforms for language learning. This aspect of the research project relies on the development of a User Database containing all the information related to subjects’ life, neurocognitive background and to their behavioral and neurophysiological data acquired over during the training; the wealth of data will be then mined through Big Data Analytics. Future studies will aslo implementat and test sistematically the effectiveness of novel forms of training including new content (i.e. word classes such as adjectives, prepositions etc.) and strategies (i.e. teaching high-frequency n-grams such as I don't know, or I want that).

In conclusion, in this paper we have described the core concepts behind the development of LANGA. This project is innovative in several aspects. First, the core principles behind the design of the training exercises and material have been largely inspired by scientific evidence on L2 learning and processing, and sistematically tested using well-established behavioral and neuroimaging methods. This is of particular relevance especially in a field were, despite the abundance of available products, there is still scant evidence about their effectiveness. Second, LANGA has been built from the ground-up to serve as both a research tool and, once reached a certain standard of quality, a commercially available product. This intertwinement of research-driven and market-oriented approach provides unique opportunities to innovate both the research and design sectors. In fact, the former is currently mostly charaterized by small scale studies and slow progress, while the latter oftentimes lacks personalities with the scientific skills required to design products that are not only estetically appealing, but also highly effective.


Acknowledgments

The authors wish to thank Copernicus Studios, Inc. for their assistance on this project, including the donation of software.. This research was supported by funding from that Natural Sciences and Engineering Research Council of Canada (NSERC) awarded to AJN. FU and KO were  supported by the RADIANT NSERC CREATE training grant.