An AI-based Learning Companion Promoting Lifelong Learning Opportunities for All

Education will be impacted, and possibly transformed, by AI. AI is already changing the knowledge and skills needed in our global and innovation centred world. But AI is also enabling innovative methods of teaching and learning [15]. This report offers a summary and some examples of how AI could support the task of learners and teachers around the world, reflecting on some of the social aspects of AI and the technical and pedagogical challenges of the field. AI in Education (AIEd) includes everything from AI-driven, personalized and conversational educative systems, intelligent agents, automatic scoring and assessment and learner-support chatbots, to AI-facilitated matching of learner to learner/teacher and collaborative learning, putting learners in full control of their learning [15]. The field itself generates its own research questions: What is the nature of knowledge, and how can it be represented and understood with AI? How can a learner be supported in their lifelong learning path? How can a AI system support through personalisation via different styles of learning and be inclusive? How can the AI system be built in a transparent way so that it allows learners to self-reflect on their path and knowledge acquired? How can the AI system automatically gather and filter educational resources available online and suitable for the learner? Thus, it proves as well a research tool to test hypotheses within the learning sciences. As a more specific example, AI can offer easy access to a complete individualization of learning in the form of your own lifelong learning companion. Particularly, formal evaluations have shown that intelligent tutoring systems, which use AI to support the tutoring process, produce similar learning gains as one-on-one human tutoring, which has the potential to increase student performance to around the 98 percentile in a standard classroom [25, 11, 2]. Additionally, intelligent tutors could effectively reduce by one-third to one-half the time required for learning [25], increase effectiveness by 30% as compared to traditional instruction [25, 14, 13], reduce the need for training support personnel by about 70% and operating costs by about 92%. These gains will enable great opportunities facilitating quality education in developing countries [18, 24].

We identify several unique technical challenges within the application of AI to education. Many of these challenges are well-aligned with some of the progress in the X5GON project. While this list of challenges may not be complete, we believe that addressing these is vital in order to build a new generation of technological tools for education and that creating a dialogue around these would bring us closer to setting the foundations needed for the future of AIEd:
Aim 1: Scalable and automatic mining/comprehension of educational materials: Automatic annotation and processing of educational materials has been a key challenge faced by different branches of the AIEd community since its inception. Many of the current applications rely simply on human labelling, but this is not scalable in practice. With groundbreaking ideas such as Open Education Resources emerging, the rate of production of novel educational resources has skyrocketed. However, learners can often be overwhelmed by all the available information, (specially as the number of choices increases), and could greatly benefit from an educational recommender. It is thus essential for these AIEd systems to understand the universal structure and direction of knowledge, identifying automatically knowledge prerequisites, topics covered and difficulty for educational materials, while at the same time filtering the materials by some metrics of quality assurance, all with the final goal of matching the most appropriate materials with the learners. This objective needs to be achieved across multiple languages and cultures, supporting automatic translations.
Aim 2: Understanding learners and personalisation: AIEd systems need to be aware of the state of the learners, which involves keeping track of their ever evolving knowledge, interest and learning goals. Furthermore, these systems need to bring novelty and serendipity to the learner, to keep the user growing effectively and expose them to new high quality material they may not have found otherwise. These systems also need to consider learning preferences/styles and the special needs of users.
Aim 3: Transparency, conversation and Human-AI collaboration: AIEd systems need to be transparent, to build trust and enable self-reflection in learners [8]. They need to keep the human in the loop to communicate the beliefs that the system holds about the user and allow the user to scrutinise the model parameters and the model’s perception of the learner [1]. Such communication should be done in a conversational and intuitive way for the learner. This could be done for example by building chatbots that can interact with learners, helping them revise and explain concepts within the learner’s knowledge gaps. Part of these aims not only apply to AI but more generally to the design of the user interfaces or the human-computer interaction. For example, by building visualisation of the materials in the user interface that allow the learner to understand/filter content more efficiently.
Aim 4: Extensive evaluation: The hypotheses and techniques developed within the AIEd field need to be evaluated thoroughly and extensively, for example through the use of massive A/B testing, user studies, qualitative and quantitative learner feedback and available datasets. However, few educational datasets are publicly available, partly because the data comes from proprietary platforms that often do not release data that could be crucial to leveraging these personalised systems in the future. This is also in part due to privacy concerns, but this significantly precludes the advancement of the field. The community needs to propel the release of publicly available datasets, while finding ways to maintain the privacy of the users.
Aim 5: Generative and planning tasks: AIEd systems need to conquer as well highly ambitious generative tasks, such as automatic generation of test questions and exercises, as well as personalised learning paths that entail a large planning component.
Aim 6: Data efficiency and scalability: Such systems need not to overload the user and be able to infer information from learners simply from interactions with the educational materials, rather than relying heavily on explicit feedback. Finally, such systems need to have a moderate computational complexity, in order to allow for thousands of learners concurrently.

Open Educational Resources (OERs), defined as teaching, learning and research material available in the public domain or published under an open license [23], are growing at a very fast pace. While OERs could be the richest collection of learning resources that will ever be available freely to all people, a challenge that remains to be addressed is to identify ways to enhance adoption, ease of use and discovery of these OERs. The project X5GON: Cross Modal, Cross Cultural, Cross Lingual, Cross Domain, and Cross Site Global OER Network is an initiative that attempts to develop various innovative and open technology elements that will converge the currently scattered OERs.

When considering the OER ecosystem, the rapid rate of creation of new resources [19], together with the imbalanced number of materials coming from different languages/cultures [17], the variable quality of resources [9], and the clustering of resources in multiple isolated silos [21] are some of the noteworthy challenges that should be surmounted. X5GON combines various elements such as content understanding, user modelling and personalisation, and quality assurance to create a unified network of OERs across the globe. The open technologies that are anticipated to be developed through X5GON envisages access to OERs focusing on the 5 Xs, in a 1) Cross Modal, 2) Cross Cultural, 3) Cross Lingual, 4) Cross Domain, and 5) Cross Cultural setting.

X5GON project has pursued several fronts to address the previously mentioned challenges. For example, gaining visibility into the majority of educational resources available in the world, as a prerequisite to unlocking their potential. X5GON is comprised of several core technologies:

X5GON Connect Service: Connecting OER repositories across the world into a single network. Within the network, the Connect Service [27] is able to identify how a user is navigating between OERs on the same or across other repositories, thus enabling user pattern analysis. [20, 22]

X5GON Discovery: AI-powered search engine [28] that enables learners to discover learning materials. It enables finding OERs by type, license, and language. In addition, the X5GON team partnered with Creative Commons and integrated their Creative Commons Catalog API [10], thus supporting searching through CC licensed images.

X5GON Recommend: Aligning learning materials with learner context by finding other OERs they are likely to be interested in. The recommendation plugin is a component that allows users to find materials that are related to the OER they are currently observing.

X5GON Translate: Providing high quality automatically generated translations of learning materials beyond borders. This is achieved through the use of the Media Transcriptions and Translation Platform [12], supporting several languages and focusing the development of high quality translation models for minority languages.

X5Learn: An intelligent learning platform that bridges the right educational resource to the right user at the right time within a futuristic user interface [3].

X5GON TrueLearn: A scalable, transparent educational recommender for lifelong learners [5, 4].

X5GON Moodle: A Moodle plug-in to use X5GON technologies to build learning activities.

X5GON Blind: A prototype learning environment for the blind and visually-impaired [26]. The environment leverages existing technologies adapted for the visually-impaired and integrates X5GON TrueLearn to provide recommendations [16].

X5GON Feed: Providing quality data to further research in open education. The public X5GON API [29] enables accessing OER metadata, transcriptions and translations, as well as the models and services described above.

X5GON datasets: X5GON has released several datasets that can be used to approach some of the previously mentioned technical challenges [6, 7]. As the project approaches its end more datasets will be released.

AI will impact the education industry and society greatly and broadly. However, there are enormous technical and pedagogical challenges ahead. This report has summarised some of these challenges and the foundations that the X5GON project has set towards these. While plenty remains to be done, we believe that building a dialogue around the future and challenges of AIEd will bring us closer to a next generation of personalised technological educative systems with great potential across the globe.

[1] Krisztian Balog, Filip Radlinski, and Shushan Arakelyan. Transparent, scrutable and explainable user models for personalized recommendation. In Proc. of the 42nd Int. ACM SIGIR Conf. on Research and Development in Information Retrieval (SIGIR ’19), 2019.

[2] Benjamin S. Bloom. The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6):4{16, 1984.

[3] Sahan Bulathwela, Stefan Kreitmayer, and Mara Perez-Ortiz. What’s in it for me? augmenting recommended learning resources with navigable annotations. In Proceedings of the 25th International Conference on Intelligent User Interfaces Companion, IUI ’20, 2020.

[4] Sahan Bulathwela, Maria Perez-Ortiz, Emine Yilmaz, and John Shawe-Taylor. Towards an integrative educational recommender for lifelong learners. In Proc. of the AAAI Conference on Artificial Intelligence, 2019.

[5] Sahan Bulathwela, Maria Perez-Ortiz, Emine Yilmaz, and John Shawe-Taylor. Truelearn: A family of bayesian algorithms to match lifelong learners to open educational resources. In Proc. of the AAAI Conference on Artificial Intelligence, 2019.

[6] Sahan Bulathwela, Maria Perez-Ortiz, Emine Yilmaz, and John Shawe-Taylor. Vlengagement: A dataset of scientific video lectures for evaluating population-based engagement, 2020.

[7] Sahan Bulathwela, Mara Perez-Ortiz, Aldo Lipani, Emine Yilmaz, and John Shawe-Taylor. Predicting engagement in video lectures. In Proc. Of the International Conference on Educational Data Mining, 2020.

[8] Susan Bull and Judy Kay. Metacognition and open learner models. In The 3rd workshop on meta-cognition and self-regulated learning in educational technologies, at ITS2008, pages 7{20, 2008.

[9] Anthony F. Camilleri, Ulf Daniel Ehlers, and Jan Pawlowski. State of the art review of quality issues related to open educational resources (OER), volume 52 S. – JRC Scientific and Policy Reports of Publications Office of the European Union 2014. 2014.

[10] Creative Commons. Creative commons catalog api. https://api.creativecommons.engineering/v1/, 2019. Accessed: 2020- 11-24.

[11] Albert Corbett. Cognitive computer tutors: Solving the two-sigma problem. In Mathias Bauer, Piotr J. Gmytrasiewicz, and Julita Vassileva, editors, User Modeling 2001, pages 137{147. Springer Berlin Heidelberg, 2001.

[12] Universitat Politecnica de Valencia. Mllp | media transcription and translation platform. https://ttp.mllp.upv.es/index.php?page=faq, 2020. Accessed: 2020-11-23.

[13] JD Fletcher. Intelligent training systems in the military. Defense applications of artificial intelligence: Progress and prospects. Lexington, MA: Lexington Books, 1988.

[14] JD Fletcher. Does this stuff work? some findings from applications of technology to education and training. In Proceedings of conference on teacher education and the use of technology based learning systems. Society for Applied Learning Technology Warrenton, 1996.

[15] W. Holmes, M. Bialik, and C. Fadel. Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Independently Published, 2019.

[16] Martin Molan, Sahan Bulathwela, and Orlic Davor. Accessibility recommendation system. In Proceedings of the Conference on Open Educational Resources, OER ’20, 2020.

[17] Erik Novak, Jasna Urbancic, and Miha Jenko. Preparing multi-modal data for natural language processing. In Proc. of Slovenian KDD Conf. on Data Mining and Data Warehouses (SiKDD), 2018.

[18] Benjamin D Nye. Intelligent tutoring systems by and for the developing world: A review of trends and approaches for educational technology in a global context. International Journal of Artificial Intelligence in Education, 25(2):177{203, 2015.

[19] Jan M. Pawlowski, Volker Zimmermann, and Imc Ag. Open content: A concept for the future of e-learning and knowledge management? 2007.

[20] Ayse Saliha Sunar, Erik Novak, and Dunja Mladenic. Preferences of users on cross-site oer recommendations: Stay or leave? In Slovenian KDDConference on Data Mining and Data Warehouses, SiKDD ’19, 2019.

[21] Ayse Saliha Sunar, Erik Novak, and Dunja Mladenic. Users’ learning pathways on cross-site open educational resources. In H. Chad Lane, Susan Zvacek, and James Uhomoibhi, editors, Proceedings of the 12th International Conference on Computer Supported Education, CSEDU 2020, Prague, Czech Republic, May 2-4, 2020, Volume 2, pages 84{95. SCITEPRESS, 2020.

[22] Ayse Saliha Sunar, Erik Novak, Jasna Urbancic, and Dunja Mladenic. Users’ learning pathways on cross-site open educational resources. In CSEDU (2), EDM ’20, pages 84{95, 2020.

[23] UNESCO. Open educational resources (oer). https://en.unesco.org/themes/building-knowledge-societies/oer, 2019. Accessed: 2020-04-01.

[24] Ricardo Vinuesa, Hossein Azizpour, Iolanda Leite, Madeline Balaam, Virginia Dignum, Sami Domisch, Anna Fellander, Simone Daniela Langhans, Max Tegmark, and Francesco Fuso Nerini. The role of artificial intelligence in achieving the sustainable development goals. Nature Communications, 11(1):1{10, 2020.

[25] Beverly Park Woolf. Building intelligent interactive tutors: Student- centered strategies for revolutionizing e-learning. Morgan Kaufmann, 2010.

[26] X5GON. X5gon blind. https://blind.x5gon.org/, 2019. Accessed: 2020-11-24.

[27] X5GON. X5gon connect. https://platform.x5gon.org/products/connect, 2019. Accessed: 2020-11-23.

[28] X5GON. X5gon discovery. https://discovery.x5gon.org/, 2019. Accessed: 2020-11-23.

[29] X5GON. X5gon feed. https://platform.x5gon.org/products/feed, 2019. Accessed: 2020-11-23.