Keele Research Repository

We introduce a neural method that is able to fuse concepts from a knowledge base with the context information for the task of grouping of aspect terms. Rather than only using context information, we use the corresponding concepts of aspect terms as additional information for aspect terms representation. We also introduce a location-based attention mechanism for accurately representing context features. As both the concept and the aspect term are same level features, i.e. aspect level features, we develop a model with gating mechanism to fuse them together. All of the above features are fed into a parallel metric learning network which has the ability to learn an easier grouping representation of samples. Experimental results demonstrate that our approach outperforms different baselines and model variants on five datasets.