Boosting aspect category detection with inference heuristics and knowledge enhancement

Abstract

Aspect category detection (ACD) aims to identify the aspect categories from reviewers’ expressed opinions in a given sentence, where one or multiple predefined aspect categories are mentioned explicitly or implicitly. With the boom of the pretrained language model, related studies have achieved significant improvements in ACD. However, the studies usually follow the canonical method of fine-tuning and neglect deeply mining internal knowledge or incorporating external knowledge, which could lead to suboptimal results. To address this issue, we propose a novel multilevel knowledge-aware ACD model by innovatively converting ACD to a binary sentence-pair classification task from the viewpoint of natural language inference, which is effective and consists of four key components. The model first expands the predefined aspect categories by introducing terms with high semantic similarity scores from commonsense knowledge bases. Next, the model generates synthetic premise-hypothesis sentence pairs based on the aspect categories and an inference heuristic template. Then, the training data are effectively augmented and used for fine-tuning the proposed model. Moreover, the model designs a pooling strategy to mine the rich syntactic and semantic knowledge encoded in the internal layers of BERT. Finally, the pooled low-dimensional representation is fed to a linear classifier to detect aspect categories. Experimental results on the SemEval-2014 and SemEval-2016 benchmark datasets achieve F1-scores of 92.75% and 83.58%, respectively, which demonstrate the superiority of our proposed model compared with some strong baselines.