A multi-scale local attention mechanism for aspect extraction

Aspect extraction is a critical step in constructing knowledge graphs and involves extracting aspect information from unstructured text. Current methods typically employ attention-based techniques such as global or local attention mechanisms, each with significant limitations. Global mechanisms are prone to introducing noise, while local mechanisms face challenges in determining the optimal window size. To address these issues, we propose a novel aspect extraction approach utilizing a multi-scale local attention mechanism (MLA). This method leverages a pre-trained model to convert text into vector representations. Feature extraction is then performed with gated recurrent units, followed by representation learning at various window sizes through the MLA. Features are selected using max pooling and decoded by a fully connected neural network combined with a conditional random field to generate precise aspect labels. Experimental validation on the Zhejiang Cup e-commerce review mining dataset demonstrates that our proposed method outperforms existing models in aspect extraction performance.