Automatic Detection of a Student's Affective States for Intelligent Teaching Systems.

Area of Science:

• Educational Technology
• Affective Computing
• Human-Computer Interaction

Background:

• Intelligent tutoring systems (ITS) aim to personalize learning experiences.
• Understanding learner emotional states is crucial for effective ITS adaptation.
• Automated affect detection can enhance ITS feedback and engagement.

Purpose of the Study:

• To determine if affect states can be automatically classified from ITS interactions.
• To predict the next user emotion based on AutoTutor's feedback patterns.
• To identify key predictors of learner emotional states within ITS.

Main Methods:

• Utilized data from AutoTutor, an ITS simulating human tutors.
• Applied data mining to identify frequent item sets predicting feedback/emotion sequences.
• Employed feature extraction techniques (multilayer-perceptron, naive Bayes) for classification.
• Analyzed 200 turns of interaction data from 34 participants.

Main Results:

• Dominant frequent item sets were identified, predicting subsequent student responses.
• 'Flow' and 'Frustration' demonstrated the highest classification accuracy among detected emotions.
• 'Flow' and 'Confusion' emerged as the most common frequent item sets.
• Affective state classification was achieved using machine learning techniques.

Conclusions:

• Automated classification of learner affect states is feasible within ITS.
• Predictive models can identify emotional shifts, enabling timely intervention.
• Findings contribute to the development of more responsive and adaptive intelligent tutoring systems.