Most current systems apply flat pattern and flat centroid words, which are extracted only by relative position to question target, to identify definition sentences. In contrast to the flat knowledge, we propose dependency-based knowledge, including dependency pattern and dependency centroid word, which are extracted by dependency relation to question target. Specifically, we use the improved ultraconservative online algorithm, binary margin infused relaxed algorithm (MIRA), to estimate the weight of each dependency knowledge for the task of candidate sentences ranking. We demonstrate that the dependency-based knowledge is more effective than the flat knowledge. Meanwhile, we also show that our definitional question answering system outperforms the state-of-the-art systems on recent TREC data.