UpStory: the uppsala storytelling dataset

Area of Science:

• Child psychology and developmental science
• Educational technology and machine learning
• Human-computer interaction and affective computing

Background:

• Rapport and friendship are crucial for social interactions and learning outcomes in children.
• Automating the analysis of child social dynamics requires specialized datasets.
• Existing datasets lack explicit rapport measures for child-child interactions.

Purpose of the Study:

• To introduce the Uppsala Storytelling dataset (UpStory), a novel resource for studying rapport in child-child interactions.
• To facilitate machine learning research on predicting rapport in educational settings.
• To provide a publicly available, anonymized dataset of naturalistic dyadic child interactions.

Main Methods:

• Collected audiovisual recordings of 35 pairs of children (aged 8-10) engaged in a collaborative storytelling task.
• Employed a within-subjects design manipulating rapport levels based on pre-existing friendships.
• Extracted per-frame head pose, body pose, and facial features for analysis.

Main Results:

• The UpStory dataset contains 3 hours and 40 minutes of interaction data.
• Anonymized data including pose and facial features is publicly available.
• Baseline models achieved 68% accuracy predicting rapport from one child's data and 70% from pair data.

Conclusions:

• The UpStory dataset is valuable for advancing research on automated rapport detection in children.
• The findings demonstrate the feasibility of predicting rapport using behavioral data.
• This resource supports the development of AI systems that can understand and foster positive social interactions in educational contexts.