Knowledge graph embedding aims to address the incompleteness of graphs through low-dimensional vector mapping, but existing methods face a trade-off between modeling simple and complex relationships: translational distance models excel at handling one-to-one relationships but struggle with complex ones, while neural network models capture higher-order interactions but suffer from parameter redundancy, degrading performance for simple relationships. In this paper, we propose MultiE, a model that integrates dual-path encoding and multi-dimensional adaptive convolution. Specifically, the model employs a dual-path encoding architecture: the main path concatenates head entities with relationship vectors to capture interaction features, while the auxiliary path processes head entity vectors independently to preserve ontological attributes, enabling collaborative modeling of simple and complex relationships. For feature extraction, a multi-dimensional convolutional group is designed: 1D convolution focuses on surface semantic matching, 2D convolution models deep complex relationships, 3D convolution captures higher-order inference semantics, and an adaptive selection mechanism computes optimal channel and dimension weights. Experimental results show that MultiE achieves improvements of 16.3% and 5.2% on complex and simple relationships respectively, indicating its balanced capability in modeling diverse relational patterns.