Supplementary data for the paper 'Social robots in education: A meta-analysis of learning outcomes'

Previous meta-analyses have shown that social robots have a positive impact on learning. However, these analyses were often limited in scope or had problematic inclusion criteria, such as different control conditions. In this meta-analysis, we examined learning outcomes with more studies and a focus on the type of control conditions. Studies were included if they used a physical social robot for training cognitive skills. We retrieved a total of 147 studies, comprising 184 post-test effect sizes between robot and control group, and 377 pre-post effect sizes. Results from 79 studies with post-test effect sizes showed that social robots produced larger learning gains compared to no intervention (Mean <em>d</em> = 0.75). Additionally, participants in the robot group, on average, learned more than those in a control group with a human teacher (Mean <em>d</em> = 0.30), although there was considerable variability in the effect sizes, largely attributable to whether human and robot were co-teaching (<em>M</em> = 0.88) or whether the study involved a robot-only vs. human teacher comparison (<em>M</em> = -0.07). Pre-post effects are mostly greater than 0 (Mean <em>d</em> = 1.08), which can be explained because learning inevitably occurs with practice. A sentiment analysis using a large language model revealed that papers from outside Europe used more positive language when describing the robots. The conclusion drawn from the current meta-analysis is that the effect size does not stand on its own but is influenced by the way the robot is used and the control condition chosen.

过往的元分析（meta-analysis）均已证实，社交机器人（social robots）对学习具有积极影响。但此类元分析往往存在研究范围受限、纳入标准存在缺陷的问题——例如对照条件设置不统一。
本项元分析纳入了更多研究，并聚焦于对照条件的类型，以此对学习效果展开分析。纳入标准为：研究需采用实体社交机器人开展认知技能训练。
本研究最终检索得到147项相关研究，包含机器人组与对照组间的184个后测效应量，以及377个前后测效应量。针对包含后测效应量的79项研究的分析结果显示：相较于无干预组，社交机器人可带来更显著的学习增益（平均d值=0.75）。此外，机器人组受试者的平均学习表现也优于配备人类教师的对照组（平均d值=0.30）；不过效应量存在较大变异，这在很大程度上取决于教学模式：若为人类与机器人协同教学（平均M=0.88），或是仅对比纯机器人组与人类教师组（平均M=-0.07）。
前后测效应量大多大于0（平均d值=1.08），这一结果可通过练习必然会引发学习行为得到解释。本研究采用大语言模型（Large Language Model，LLM）开展情感分析，结果显示：欧洲以外地区的相关论文在描述社交机器人时，使用了更多积极表述。
本项元分析得出的结论为：效应量并非独立存在，其取值会受到机器人使用方式与对照条件选择的影响。