Semantic matching of open texts to pre-scripted answers in dialogue-based learning

Abstract

Gamification is frequently employed in learning environments to enhance learner interactions and engagement. However, most games use pre-scripted dialogues and interactions with players, which limit their immersion and cognition. Our aim is to develop a semantic matching tool that enables users to introduce open text answers which are automatically associated with the most similar pre-scripted answer. A structured scenario written in Dutch was developed by experts for this communication experiment as a sequence of possible interactions within the environment. Semantic similarity scores computed with the SpaCy library were combined with string kernels, WordNet-based distances, and used as features in a neural network. Our experiments show that string kernels are the most predictive feature for determining the most probable pre-scripted answer, whereas neural networks obtain similar performance by combining multiple semantic similarity measures.