This thesis presents a framework for supporting interaction regulation through computational means. Regulation of collaborative problem-solving includes aspects related to the task as well as to the interaction itself. Task related aspects consist of establishing a strategy, planning actions and evaluating progress. Interaction regulation on the other hand refers to the organization of collaboration through communication rules as well as division of labor. These rules and strategies might be established at the outset of the collective activity, but they also need to be monitored and adapted as the interaction evolves. On a moment to moment time scale, regulating collaborative interaction consists of deciding “who does what" in addition to “what to do". We chose to describe these regulation processes as a negative feedback loop, a concept borrowed from control theory. Following this metaphor, interaction regulation is a four step process that starts with the collection of raw data about the participants’ behavior (e.g. verbal contributions, mouse clicks, messages). In the second step, raw data is aggregated into a set of psychologically and pedagogically meaningful indicators that constitute the current state of interaction (e.g. symmetry of participation, quality of knowledge sharing). In the third step, the current state is compared to a representation of a desired state (standard) of interaction. Then, if there is a discrepancy between these two states of interaction, remedial actions are proposed in the fourth step (e.g. encourage participation or ask participants to clarify their explanations). Computers may offer support for any or all of these four steps. Support for the first two steps might be provided by mirroring tools, which assist learners and teachers in the collection of data by providing them with graphical feedback about their interaction. Support for the second and third step might be provided by metacognitive tools, which assist learners’ or tutors’ diagnosis of the interaction through visualizations which also contain a normative aspect that represents the standards of productive interaction. Support for the fourth step might be provided by guiding systems, which propose remedial actions based on a computational assessment of the situation. Our experimental studies show that a representation of the desired state of interaction is critical for regulation. A mirroring tool did not substantively affect the behavior of subjects while a metacognitive tool led to increased participation in dialogue, including more precise planning. Subjects were able to use the standard provided by the metacognitive tool to judge the quality of their current interaction and to take remedial actions. Mirroring tools might still be effective means to provide feedback to a group of problem-solvers, given that the standards to judge interaction are defined through instructions or are part of the subjects’ mental model of productive interaction.