<p>Reducing semantic dilution and guiding answer candidates are crucial parts in visual question answering. Existing visual question answering systems lack effective Chinese embeddings and methods for narrowing the answer candidates based on question classification. To address these issues, we propose an architecture that integrate speech act detection module and hierarchical syntactic embedding module. The speech act detection module classifies question types and filters potential answers based on these types, thereby reducing the candidate answer space. The hierarchical syntactic transformer incorporates character, word, and phrase-level syntactic structures as input embeddings to the transformer. This hierarchical syntactic approach provides the transformer with richer contextual information from the question, thereby enhancing the effectiveness of embeddings in Chinese visual question answering systems. The proposed method is implemented and evaluated on the VQAv2 dataset. Experimental results demonstrate that our approach gains a modest but consistent 0.72% accuracy improvement on the English dataset. At the same time, a Chinese version VQAv2 dataset is constructed. The result in Chinese VQAv2 dataset show a 7.52% accuracy increase over the baseline model. The results validate the effectiveness of the proposed system, particularly in advancing Chinese visual question answering.</p>