AUTHOR=Hromada Daniel D. , Kim Hyungjoong TITLE=Proof-of-concept of feasibility of human–machine peer learning for German noun vocabulary learning JOURNAL=Frontiers in Education VOLUME=8 YEAR=2023 URL=https://www.frontiersin.org/journals/education/articles/10.3389/feduc.2023.1063337 DOI=10.3389/feduc.2023.1063337 ISSN=2504-284X ABSTRACT=

The present study provides the first empiric evidence that the creation of human–machine peer learning (HMPL) couples can lead to an increase in the level of mastery of different competences in both humans and machines alike. The feasibility of the HMPL approach is demonstrated by means of Curriculum 1 whereby the human learner H gradually acquires a vocabulary of foreign language, while the artificial learner fine-tunes its ability to understand H's speech. The present study evaluated the feasibility of the HMPL approach in a proof-of-concept experiment that is composed of a pre-learn assessment, a mutual learning phase, and post-learn assessment components. Pre-learn assessment allowed us to estimate prior knowledge of foreign language learners by asking them to name visual cues corresponding to one among 100 German nouns. In a subsequent mutual learning phase, learners are asked to repeat the audio recording containing the label of a simultaneously presented word with the visual cue. After the mutual learning phase is over, the subjacent speech-to-text (STT) neural network fine-tunes its parameters and adapts itself to peculiar properties of H's voice. Finally, the exercise is terminated by the post-learn assessment phase. In both assessment phases, the number of mismatches between the expected answer and the answer provided by human and recognized by machine provides the metrics of the main evaluation. In the case of all six learners who participated in the proof-of-concept experiment, we observed an increase in the amount of matches between expected and predicted labels, which was caused both by an increase in human learner's vocabulary as well as by an increase in the recognition accuracy of machine's speech-to-text model. Therefore, the present study considers it reasonable to postulate that curricula could be drafted and deployed for different domains of expertise, whereby humans learn from AIs at the same time as AIs learn from humans.