方面级情感分析从细粒度角度检测了给定方面术语的情感极性从挖掘文本方面的情感态度的角度给电商、消费者等群体提供决策辅助。现有方法对句法信息的不同依赖关系的平等建模，忽略了卷积网络中不同依赖关系和语义关系中的全局信息对方面词情感分类的影响。为了解决上述问题，考虑到图结构的灵活性、复杂性和辅助任务在基于全局语义信息捕捉方面情感、补全细粒度方面信息的优良表现，提出了包含加权关系卷积网络（weighted relational convolutional networks, WRCN）和辅助任务互补学习（complementary task, CL）两个任务的WRCN-CL （weight relational convolutional networks and complementary task, WRCN-CL）模型。该模型利用Bi-LSTM （bi-directional long short-term memory, Bi-LSTM）提取文本特征；之后分别输入WRCN和CL中；CL从全局角度寻找方面相关的语义信息以达到增强知识的效果，WRCN基于CL中的方面表示结合利用GCN （graph convolutional networks）深入挖掘基于位置信息和关系类型的句法信息，最后将融合全局信息和局部信息的特征输入到池化层，得到全面的信息表征以提高模型的分类效果。实验结果表明：相较于其他基线模型，WRCN-CL在三个公开数据集上准确率分别达到了83.49%、78.19%、75.89%，从而证明了本文模型能有效地解决方面级情感分析分类任务。

Aspect-level sentiment analysis detects the sentiment polarity of given aspect terms from a fine-grained perspective, providing decision support for e-commerce, consumers, and other groups by mining textual aspect sentiment. Different syntactic dependencies were treated equally in existing methods resulting in the influence of relation types in convolutional networks and the global information from semantic perspective being overlooked. To address these issues, considering the flexibility and complexity of graph structures, the excellent performance of auxiliary tasks in capturing aspect sentiment based on global semantic information and completing fine-grained aspect information, the model WRCN-CL (weight relational convolutional networks and complementary task) which incorporates two tasks: WRCN(weighted relational convolutional networks) and CL (complementary learning) was proposed. Specifically, Bi-LSTM (bidirectional long short-term memory network) was used to extract textual features, which were entered into WRCN and CL tasks separately. Aspect-related semantic information was collected from a global perspective to enhance knowledge, while the aspect representations from CL combined with GCN (graph convolutional networks) to deeply explore syntactic information based on positional and type-aware relational information in WRCN. The fused global and local features were then input into a pooling layer to obtain comprehensive information representation for improved classification performance. Experimental results demonstrate significant improvements with the accuracy of 83.49%、78.19%、75.89% on three public datasets compared to baseline models, proving the effectiveness of the proposed model in aspect-level sentiment analysis classification task.