Towards Reducing Foreign Language Anxiety Using Level-Appropriate Embodied Conversational Agents

In plain English

A multi-agent embodied conversational agent generates English dialogue pitched at the learner's CEFR proficiency level, using a generate-evaluate-regenerate loop with a level classifier to keep its language simple enough. It worked at the difficulty-matching job: 87.4% of its sentences landed within one CEFR level of the learner's self-assessed proficiency, against 54.1% for an unsimplified agent. It did not demonstrably reduce foreign language anxiety in a small pilot with Japanese university students, which the authors attribute to sample size — a useful reminder that controlling an agent's output register is a much easier win than changing how a learner feels.