Using AI Techniques in a Serious Game for Socio-Moral Reasoning Development

Ange Tato, Roger Nkambou, Aude DufresneAAAI 2020 · DOI

Researchers from the Université du Québec à Montréal have developed LesDilemmes, a serious game designed to assess and foster Socio-Moral Reasoning (SMR) in teenagers. By leveraging a hybrid deep learning architecture that integrates psychological theory with neural networks, the project demonstrates how AI can automate complex cognitive assessments and drive adaptive learning experiences.

The Challenge of Automating Moral Assessment

Socio-Moral Reasoning (SMR) is the cognitive framework individuals use to navigate moral emotions and social conventions. Traditionally, assessing SMR is a labor-intensive manual process. Participants are presented with dilemmas, and their verbal justifications (verbatims) are recorded and scored by experts based on Kohlberg’s stages of moral development.

Automating this process faces two primary hurdles:

  1. Engagement: Standard computerized SMR tasks are often dry, failing to motivate younger learners.
  2. Data Scarcity: High-quality labeled datasets for moral justifications are small. Purely data-driven models (e.g., standard CNNs or LSTMs) struggle with these "small data" scenarios and often ignore established psychological theory.

Methodology: The Global Attentional Hybrid Model

The core innovation of LesDilemmes is its hybrid Learner Model, which combines expert psychological knowledge with deep learning architectures to evaluate player maturity.

The hybrid architecture integrates a priori expert knowledge with neural networks via an attention mechanism.

1. Hybrid SMR Predictor

Instead of relying solely on a neural network to learn from scratch, the authors built a "Global Attentional Hybrid Model." Expert knowledge—descriptions of the five SMR levels ranging from egocentric to universal-rights-based—was encoded using NLP techniques like Word Movers’ Distance (WMD). A "consultation" attention layer allows the neural network (CNN/LSTM) to weigh these expert descriptions while processing the player's actual justifications. This ensures the model's predictions remain constrained by established theory.

2. Emotional Tracking and Adaptation

The game uses the Facereader tool to monitor seven basic emotions and arousal levels via the player's webcam. The game then adapts based on both expert-defined rules (e.g., providing specific feedback for low SMR levels) and data-driven rules extracted via decision trees (e.g., adjusting environmental music based on arousal).

Key Results

The research evaluated the model's predictive power and the game's impact on a cohort of 70 participants (ages 8–19).

Model Performance

The final hybrid model achieved an accuracy of 75% (F1-score 0.73), significantly outperforming the "Expert-Only" baseline (40%) and the "CNN-only" model (53%). As shown in Figure 5, the hybrid approach (final-model) was particularly effective at bridging the gap for higher levels of reasoning (Levels 4 and 5), which are typically underrepresented in training data.

Comparison of model performance across SMR levels, showing the final hybrid model's superiority over baseline architectures.

Learning Gains and Immersion

Players in the AI-based adaptive version showed a highly significant increase in SMR maturity ($p = .001$), moving from an average pre-test score of 1.70 to a post-test score of 2.53. Furthermore, the adaptive version was rated as significantly more immersive ($p = .045$) than the static version, suggesting that AI-driven personalization directly enhances the educational experience.

Conclusion

By using an attention mechanism to let a model "consult" textbook-style descriptions of categories, the researchers successfully combined symbolic AI (rules/knowledge) with connectionist AI (neural networks). While limitations remain regarding speech-to-text transcription accuracy and game interactivity, the study proves that theory-infused AI can effectively automate complex psychological assessments and create more engaging, effective learning environments.

Summary updated

BibTeX
@inproceedings{tato2020using,
  author = {Ange Tato and Roger Nkambou and Aude Dufresne},
  title = {{Using AI Techniques in a Serious Game for Socio-Moral Reasoning Development}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i09.7066},
}
Figure