Academic opinion recognition aims to identify subjective expressions in scientific texts, particularly those conveyed through implicit reasoning or discourse-level cues. Existing approaches struggle with such subtle expressions, often relying on shallow lexical cues or flat classification, which neglect reasoning structures and limit interpretability. In contrast, we propose an explanation-driven distillation framework that captures academic subjectivity through linguistically grounded rationale templates. Drawing on linguistic theory, we summarize seven prototypical expression patterns and encode them into a three-step reasoning template, which guides large language models as teacher models to generate rationales. These rationales serve as the sole supervision signal for training a compact student model, from which labels are deterministically parsed. To enhance reasoning fidelity, we introduce a contrastive distillation strategy based on counterfactual rationales with inverted logic, encouraging the student model to distinguish valid from flawed explanations. Experiments demonstrate superior performance in opinion classification and explanation quality, with an F1-score of 89.10% and ROUGE-L of 74.25%. Ablation studies confirm that both rationale structure and the contrastive learning strategy are essential to achieve interpretability and robust opinion recognition.