Abstract: Commonsense question answer QA aims to make models simulate the way humans solve problems to predict answers. Due to the lack of common sense in questions expression process, it brings great challenges to the traditional machine learning methods. For the multi-background knowledge fusion problem, a random walk was used to obtain two-hop related entities of the answer entity in the common sense knowledge graph ConceptNet, which was fused with the given question and answer text, and the enhanced question was input into the pretraining model RoBERTa, using context-aware attention, which strengthened the semantic representation between the question and answers. Experimental results indicate that after effectively introducing external knowledge, CAARK model performs well on the CommonsenseQA dataset, providing a new paradigm for solving common sense question answering problems.