There is limited research on how perceived peer network influences student collaboration in project-based instruction. Research based on interviews or content analyses may overlook the semantic structure of discourse. In this study, we combine content analysis with computational linguistics to explore the collaboration patterns of 22 first-year students during face-to-face group design (n = 7,514 conversational turns) in a project-based engineering course. We find that students who reported smaller peer network generally produced discourse that provided new information, but less cohesion, compared to students with perceived median and large peer networks. Overall, students with small networks also engaged in group planning, evaluation, and shared regulation less frequently compared to the other two groups. Findings have implications for adjusting group arrangement in design activities. The study also illustrates the potential of incorporating computational approaches to detecting discourse.