Abstract: Definitional pattern and centroid word vector are the popular features for definitional question answering. The former determines how to describe the information, and the latter determines what topics should be concerned. Most current systems use sequence pattern and sequence centroid words extracted only by relative position to question target to identify definition sentences. In contrast to sequence knowledge, we propose dependency-based knowledge, including dependency pattern and dependency centroid word extracted by dependency relation to question target. Specifically, we use the improved ultraconservative online algorithm and margin infused relaxed algorithm (MIRA) for the task of candidate sentences ranking. Experiments have demonstrated that the improved MIRA can greatly increase the system performance in terms of progressive reduction of optimization constraints as training moves forward.