<p>Aspect-level implicit sentiment analysis is a task that identifies the sentiment polarity of the aspect in a review with implicit sentiment. The existing works primarily classify implicit sentiment based on linguistic expression patterns. Meanwhile, they utilize deep learning methods to extract multi-level semantic information from text or fuse single-source knowledge. However, the ambiguity inherent in implicit sentiment makes it challenging for these methods to grasp deeper semantic meanings such as irony. In addition, large language models demonstrate robust contextual understanding and reasoning capabilities, offering new paths for analyzing model interpretability. Therefore, an implicit sentiment classification method based on model perspective is designed. Based on this, we propose a multi-source knowledge Fusion Networks with Adaptive Weighting and Multi-step Reading (FN-AWMR). On the one hand, we propose a dynamic multi-source knowledge fine-grained matching method, which matches fine-grained multi-source knowledge relevant to the context. On the other hand, we propose an adaptive multi-source knowledge weighting and multi-step reading module. It can adaptively weight knowledge at different granularities. In addition, to enhance the interpretability of the model, we design implicit sentiment dynamic prompt template, which drives the large language model to generate logically coherent and reasonable explanations. Rich experiments show that our model is significantly effective for aspect-level implicit sentiment analysis.</p>