With the continuous development of natural language processing (NLP), sentiment analysis has become increasingly critical, profoundly impacting daily social life. In aspect-level sentiment analysis (ALSA) based on deep learning, researchers primarily employ models such as LSTM networks, attention mechanisms, and Transformers to extract textual features for sentiment polarity classification. Although existing models achieve satisfactory performance, sentences often contain multiple aspects with distinct polarities, and linguistic complexity hinders models from accurately capturing semantic dependencies. Consequently, these models struggle to effectively extract aspect terms and their critical contextual information, limiting further improvements in accuracy. To address these challenges, we introduce a BERT-based Information Fusion Aspect-Level Sentiment Analysis Model (BIFNet). BIFNet first leverages the BERT model for text pre-training, then introduces a relational computing module (RCM) to enhance dependency modeling between word vectors, improving the extraction of aspect terms and their associated sentiment words. Additionally, BIFNet employs a feature fusion module (FFM) that integrates graph optimization and attention mechanisms for comprehensive vector fusion, enriching contextual information around aspect terms. Experiments on two benchmark datasets have substantiated that BIFNet surpasses existing models by more accurately capturing relationships between aspect terms and sentiment words, achieving superior performance in aspect-level sentiment analysis.