Conversation-based assessment systems allow for students to display evidence of their knowledge during natural language conversations with artificial agents. In the current study, 235 middle-school students from diverse backgrounds interacted with a conversation-based assessment system designed to measure scientific inquiry. There were two versions of the conversations where the initial question was manipulated to examine the effects of question framing on the quality of evidence of students' knowledge and skills elicited during the conversation. We analyzed the human input during these conversations post-hoc using LIWC to discover linguistic profiles of students that may be related to the type of question asked as well as overall task performance. Furthermore, we compared these linguistic profiles to human ratings as a validity check and to inform our interpretation. Results indicate four separate profiles determined by linguistic features that indeed align to human scores and performance in directions consistent with the effects of question framing. These results offer important implications for improved detection of types of student learners based on linguistic features that do not differ by diversity as well as for designing conversation-based assessments.