Among the topics discussed in Social Media, some lead to controversy. A number of recent studies have focused on the problem of identifying controversy in social media mostly based on the analysis of textual content or rely on global network structure. Such approaches have strong limitations due to the difficulty of understanding natural language, and of investigating the global network structure. In this work we show that it is possible to detect controversy in social media by exploiting network motifs, that is, local patterns of user interaction. The proposed approach allows for a language-independent and fine-grained and efficient-to-compute analysis of user discussions and their evolution over time. The supervised model exploiting motif patterns can achieve 85% accuracy, with an improvement of 7% compared to baseline structural, propagation-based and temporal network features.