Construction of Knowledge-Enhancing Prompt Templates for Aspect-Level Sentiment Analysis

Currently, aspect-level sentiment analysis models based on neural networks primarily involve discrete training of two subtasks: aspect term extraction and aspect polarity classification, or fine-tuning of pre-trained language models. These approaches neglect the interplay between the two subtasks, making it challenging to fully leverage the language model knowledge acquired during pre-training. This paper proposes an external knowledge-enhanced prompt template approach（KPT）, which constructs the vocabulary of the pre-trained language model into a K-dimensional tree. The K-nearest neighbor search algorithm is employed to search for optimal prompt words on this K-dimensional tree, thereby constructing an optimal prompt template. In this process, external knowledge is integrated into the vocabulary to enrich its semantic information. Features from the aspect term extraction task are utilized to further enhance the performance of the aspect polarity classification task. A multi-head attention mechanism is employed to enable interaction between the two tasks and integrate them, thereby improving the accuracy of sentiment polarity judgment by the external knowledge-enhanced prompt template. Experimental results on three public datasets, namely Lap14, Rest14, and Twitter, demonstrate that the proposed method outperforms existing models such as ASGCN, BiGCN, CDT, and others.