With the goal of incorporating the use of social network tools into the formal curriculum of public schools in Chile, Kelluwen developed a b-learning approach in which students work in groups and engage in peer review dynamics. To support such activities, the Kelluwen platform provides a complementary micro-blogging tool called Worklog. In this study, we analyze data collected from Worklog messages generated by 101 cohorts of students from different schools and different levels, who participated in Kelluwen activities. We first used machine learning techniques to discover three types of messages which coincide with three dialogic functions; referential, phatic, and emotional. We then analyze how different factors of the cohort and group setup relate to the presence of different dialogic functions in Worklog. Results show the positive effect of pairing groups of students from different schools, and the positive influence of teacher participation. Our work contributes to the discussion on the role of technological social tools on the implementation of complex b-learning activities in schools, as well as demonstrating how automatic and semi-automatic data analysis techniques can be used to better understand such complex interventions.