As a task in automatic question answering (QA), answer selection tasks require the model to select the top-ranked answer given a question and multiple candidate answers, and are important in applications such as dialog systems, and information retrieval. Current QA methods, however, exhibit limitations in addressing complex problems, particularly when the contextual interplay between questions and answers is disjointed. To mitigate this shortfall and augment the efficacy of QA, we introduce a novel framework that utilizes external knowledge fusion. By searching for background knowledge in an external knowledge base, employing a multi-attention mechanism to obtain the relevance scores of knowledges and text vector at the sequence and pooling levels, which are weighted summation with the cosine similarity scores to dynamically select relevant knowledge, the framework enhances and compresses the filtering vectors to prevent model overfitting. We do not require costly pre-training and show significant improvements in Mean Reciprocal Rank (MRR) and Mean Average Precision (MAP). We will name this framework KnFus-Net. This study demonstrates that the effective utilization of external knowledge can improve the accuracy and robustness of QA, providing new directions and ideas for future research.