A semantic diagram tool is proposed in this study in order to structure collaborative problem solving (CPS) based on cognitive load theory (CLT). To investigate its effects on transaction cost and the deepening of user understandings, a comparative quasi-experiment was designed and conducted with 49 participants from a university in East China. Analysis of group dialogic acts and self-reported cognitive load showed that a semantic diagram tool can decrease transaction costs during CPS. It can thereby help collaborators invest less effort into those procedural conversations that are necessary for managing social interaction but which do not directly contribute to deepening collaborative learning in the task domain. Data analysis of pretested and posttested domain understanding suggested that learners achieved a greater depth of understanding after CPS supported by the semantic diagram tool. These findings are interpreted in terms of CLT. A semantic diagram tool can release the learner's limited working memory from procedural conversation, creating more working memory capacity for deepening problem understanding. Therefore, this study demonstrates one means whereby a semantic diagram tool functions as a promising technological support to structure CPS.