Learners can then have adaptive access to interactive course materials and relevant services on various devices (e.g. mobile devices, computers and even Smart TVs) via a seemless frontend.

AI is one of the key drivers of adaptive learning. Analyzing usage data leads to a better understanding of the learning process and thus, optimized content, teaching and learning thereafter. For instance, in digital learning environments, real-time analyses of students’ interactions with learning objects provides vital information about content validity, teaching methods and personal learning habits. All user interactions (for example, with learning objects) are persisted as ADL xAPI or IMS CALIPER statements in a learning record store. To address the evolution of sustainable software architectures, FOKUS researchers developed the Common Learning Middleware (CLM), a field-proven enabler for interoperable microservice infrastructures. This service mediator is strongly recommended in order to manage AI services, data storage, content databases and adaptive user interfaces. 

Our Gen AI activities can be summarized as follows:

Document-based Chatbots and Semantic Search

Techniques like Retrieval Augmented Generation (RAG) split large documents into small chunks. These chunks are fed into an LLM with the goal of extracting the representation of a textual chunk in a way of how an LLM sees it: As a set of floating-point numbers. Once these representations are extracted, mathematical operations can be performed with them, like comparing the distance between the chunks and the numerical representation of user queries. This way, specific parts of documents can be strategically retrieved to expand the knowledge of an LLM by including these parts in prompts. Which allows for a low threshold interaction with concrete documents like learning material.

Structuring Unstructured Data with LLMs

While there are many detailed data specifications for specific data within Learning Technologies, the actual information concrete Learning Objects come with can be rather sparse. Additionally, most of the data companies deal with is unstructured, like internal memos, protocols etc. In both cases, LLMs can be utilized to generate structured outputs based on the actual unstructured documents. For example, free-text course descriptions can be annotated with actual skills specified within the ESCO (European Skills, Competences, Qualifications and Occupations) taxonomy and structured outputs that follow data specifications can be generated from natural language user queries to enhance content discovery.

Multi Modal Content Understanding

LLM-based applications can often be expanded by capabilities of dealing with content from other modalities than text. For example, Visual Language Models (VLMs) enable applications to reason with image and video data and thus to apply techniques like Retrieval Augmented Generation (RAG) to collections of visual data. This applies to models dealing with audio data as well, which can also be used as a backbone of powering Speech-To-Text/Text-To-Speech pipelines.

Thereby, the Recommender Systems follow a particular paradigm that is different from e-commerce or entertainment domains. These particular systems must consider the unique characteristics that educational settings comprise of, such as the learners’ motivation and needs, typical user behaviors and resulting activity data. As such, Learning Recommender Systems should enable more efficient and effective learning for its users and in turn, future learning activities in a positive manner. In other words, rather than being based on certain preferences, such systems should be based on the actual user’s learning needs.

To meet the unique requirements in the specialized paradigm, our Smart Learning Recommender System utilizes a novel knowledge-based filtering approach that combines the strengths of content-based and collaborative filtering techniques. The algorithms transfer multi-contextual activity data into time-dependent user models. The analyzed contexts may include interactions with and processing times of learning media, self-assessments, exercise performances, fullfiled knowledge prerequisites, lecture times and exam relevance. Additionally, more advanced models may consider the learning patterns of other learners as well as predict the individual forgetting effect after each learning iteration. The resulting scores represent the individual user’s need to learn specific learning objects at a particular point in time of the course. The recommendations are presented to the user as hints for learning that are well explained ( explainable AI) and easy to understand. In comparison to the evaluation results of other educational recommender systems (built with the same datasets), our knowledge-based filtering approach generates the most accurate and time-sensitive recommendations.