Detecting deception in real interrogations for criminal cases is critically important. Interrogation is composed of evidence-driven conversation that calls for a need for proper integration of context, where most prior works treat it as a sequence modeling task. In this work, we propose a context-constrained sentence modeling approach for deception detection. Specifically, we introduce the use of a global context label that is defined on multi-sentences, i.e., a context label is marked as deception if any of its sentences are deceptive. Then, by using a contextual integrator that aggregates predictions on local sentences for context label prediction, we improve deception detection by jointly optimizing global and local labels. Our approach significantly outperforms other models and achieves 76.38% and 73.15% in Unweighted Average Recall (UAR) at the local and global levels, respectively. We also conducted two analyses to further demonstrate the effectiveness of our approach.