Developing analytic rubrics requires considerable preparation time for teachers, highlighting the necessity for adequate support mechanisms. This paper presents and evaluates a teacher-support tool designed to automate the generation of analytic rubrics for assessing student submissions and delivering actionable feedback. The tool incorporates a large language model as a core component, employing generalized prompt templates to enhance its versatility. The generated outputs are systematically compared against predefined teacher expectations to evaluate the tool's applicability and effectiveness in three real-world educational scenarios.

This paper introduces a tool for auto-generating criteria-oriented feedback in higher education. A model is trained based on 39 historical student submissions and teacher ratings. The tool combines natural language processing with large language model capacities to analyze contextual criteria, for which regression trees are trained. It aims to mimic teachers when providing feedback utilizing analytic rubrics. The evaluation results with 38 new submissions reveal that predictions are good, serving as a fruitful base to optimize the process when providing formative feedback.

Personalizing online courses has been the subject of exploration for decades. Central to this discourse is the contentious concept of learning styles, which has given rise to the meshing hypothesis. In this paper, that concept has been transferred from traditional learning styles to preferences for certain instructional methods. Especially gamified elements, competitions, and group work modes have been imitated to either be incorporated or excluded within an adaptive course besides conventional personalization features. A controlled experiment was conducted to evaluate the effect of the preference-based meshing hypothesis. The results revealed a statistically significant performance difference in the final, most challenging task (with$p=. 039$in the Welch test) for a specific setting. This study underscores the potential value of focusing on the meshing hypothesis, particularly concerning preferences for instructional methods, as a significant area for future research.

Computer-generated feedback can be created manifold. This paper compares two approaches for generating feedback: rule-based and prediction-based. Both approaches have several advantages and disadvantages, which are discussed in detail considering precision, recall, human effort for model creation, and explainability requirements.

Learning material is designed based on didactical concepts and methods. To create an item sequence, strategies are applied by rule-based descriptions. Merging them requires their combination, which is complex to apply. In this paper, strategy combination is applied using LSTMs. Samples are generated to train different LSTM models for specific strategies. Those models are then merged by averaging their weights. The approach gives teachers the controllability over the selection of an appropriate trade-off between different strategies, without the need to create complex rules by hand.

Online courses can be adapted to suit learner needs. Although it is known that learners are diverse, courses are often optimized using split tests to find an optimum that results in the best performance for the majority of participants. This is the best-practice approach, which is cost-efficient using well-defined statistical fundaments. However, learners are considered as one cohort, independently of subgroups, and their existence is seldom further analyzed. In this paper, we examine two versions of a 45min language learning online course, which cover the same learning content, but one version is enriched by simulations to create different settings of being observed, collaborating with a peer, or taking part in a competition. T-tests over all 157 users identify some tasks, which are optimum for the majority of learners. Nevertheless, such tasks must not be the best for everyone. If learners are split by preference levels, learner performances differ, but the results are in line with the literature, without statistically significance.

Many prediction tasks can be done based on users’ trace data. This paper explores divergent and convergent thinking as person-related attributes and predicts them based on features gathered in an online course. We use the logfile data of a short Moodle course, combined with an image test (IMT), the Alternate Uses Task (AUT), the Remote Associates Test (RAT), and creative self-efficacy (CSE). Our results show that originality and elaboration metrics can be predicted with an accuracy of ~.7 in cross-validation, whereby predicting fluency and RAT scores perform worst. CSE items can be predicted with an accuracy of ~.45. The best performing model is a Random Forest Tree, where the features were reduced using a Linear Discriminant Analysis in advance. The promising results can help to adjust online courses to the learners’ needs based on their creative performances.

Online courses have very high dropout rates worldwide. Learners are demotivated based on bad learning experiences. While some factors of online courses could be optimized for all learners, e.g. the quality, it is essential to note that a one-size-fits-all environment is not existing. Some learners are comfortable with certain methods while others may not. In the paper, we identified five learner preferences that can be used to adapt teaching methods in online courses. We provide a 10-item questionnaire, validate it based on exploratory factor analysis, and examine whether learners differ in preferences.

Online courses have very high dropout rates worldwide. Learners are demotivated based on bad learning experiences, lack of time, or motivation. While some factors of online courses could be optimized for all learners, e.g. the quality, it is essential to note that a one-size-fits-all solution is not sufficient. Learners have different preferences. They feel well with some methods while others may not. In this paper, we predict five learning preferences based on trace and performance data. The promising result shows that we can predict the values of our five preferences with acceptable accuracy up to.75, which can be used for further adaptions.

The combination of gaming and language learning is not new, but it has not been applied in open education adventure games. Although it is known that especially young people aim to play games, their combination with learning is mainly limited to simulations. This does not have a high impact on intrinsic motivation. This paper provides an engine to create adventure games with a virtual world based on images and a dialog-based storyline. Learners interact with virtual entities in plausible contexts using the foreign language only. If learners are unaware of a word’s meaning, cheating is possible, enhancing the user model. Based on that, learning material can be generated to practice unknown vocab by regularly interrupting the game. The generated micro-course needs to be finished before the game continues. We tested a game created using the engine and got wide acceptance.

In this paper, we use a Generative Adversarial Network (GAN) as a sequence generator for language learning online courses. Therefore, we cluster a very small dataset of manually created training samples to derive rules. Then, we train a GAN that can mimic rule-based sequences, where we use our derived rules to evaluate generated samples. We enhance our approach by a parameter that course creators can select deviations they want to have in new sequences without manual adjustments. The resulting sequences follow the core structure of the small sample set. Based on deviations of the generated new learning paths, new combinations of methods can be used that course creators did not previously have in mind. This opens up a new way to generate course sequences without the need to model many alternative learning paths for adaptions.

In language learning, getting corrective feedback for writing tasks is an essential didactical concept to improve learners' language skills. Although various tools for automatic correction do exist, open writing texts still need to be corrected manually by teachers to provide helpful feedback to learners. In this paper, we explore the usefulness of an auto-correction tool in the context of language learning. In the first step, we compare the corrections of 100 learner texts suggested by a correction tool with those done by human teachers and examine the differences. In a second step, we do a qualitative analysis, where we investigate the requirements that need to be tackled to make existing proofreading tools useful for language learning. The results reveal that the aim of enhancing texts by proofreading, in general, is quite different from the purpose of providing corrective feedback in language learning. Only one of four relevant errors (recall=.26) marked by human teachers is recorded correctly by the tool, whereas many expressions thought to be faulty by the tool are sometimes no errors at all (precision=.33). We provide and discuss the challenges that need to be addressed to adjust those tools for language learning.

Nowadays, online courses can be created in very efficient ways including interactive tasks using H5P. While it is easy to use this technology, the design of H5P has the disadvantage to be not useful for exams as answers can be found in the source code, which can be accessed without any effort. In this paper, we propose a technical concept to use interactive media in exams, created with H5P. Therefore, we provide insights into which parts need to be changed and how answers should be stored for grading.

The adaption of online courses according to the user knowledge is not new, but its application is limited. The major problem is the lack of missing open-source technology to personalize online courses with individual learning paths on a large scale. In this paper, we introduce our open-source framework that allows us to generate adaptive online courses in Moodle. We focus on language learning as an example implementation. Courses are generated based on a knowledge base and several XML templates that allow the generation of interactive tasks using H5P. The novelty of our approach is the application and combination of existing Moodle libraries to generate individual online courses, although they were not designed for that purpose. It is the first working example that supports individual learning paths on a large scale in Moodle. Besides, we provide technical details on how to overcome limiting problems by using Moodle as an LMS.