Text data pertaining to socio-technical networks often are analyzed separately from relational data, or are reduced to the fact and strength of the flow of information between nodes. Disregarding the content of text data for network analysis can limit our understanding of the effects of language use in networks. We present a computational and interdisciplinary methodology that addresses this limitation by combining theory from socio-linguistics with social network analysis and machine learning based text mining: we use network analysis to identify groups of individuals who assume the theoretically grounded roles of change agents and preservation agents. People in these roles differ in their motivation and capability to induce and adopt change in a network. Topic modeling is then constrained to the texts authored by people in these roles. We apply this methodology to a public dataset of about 55,000 research proposals that were granted funding. Our results suggest that the people per role differ in the research domains they work on and the strength of association with those domains that both roles are involved with, but are similar with respect to fulfilling the task or additional role of being a project manager.