Improving aspect-based sentiment analysis with Knowledge-aware Dependency Graph Network

Abstract

Aspect-based sentiment analysis (ABSA) aims to mine multiple sentiment–target pairs contained in a review sentence. The main challenge of this task is how to extract the sentiment polarity of a specific sentiment item efficiently. Earlier research focused on recurrent neural networks (RNNs), which implicitly associate the sentiment items with sentiment polarities through an attention mechanism. However, due to the complexity of language and the fact that a sentence contains multiple sentiment pairs, these models often fail to capture sentiment pairs accurately. Most recent efforts have applied syntactic information, especially dependency information, to construct structured models (e.g., tree-based models or graph neural networks) for sentiment analysis. Although these structured models achieve better results, they ignore the domain knowledge related to the entities of the comment sentences. This domain knowledge (e.g., brand reputation, influence) significantly impacts the sentiment polarity. Hence, this paper proposes a Knowledge-aware Dependency Graph Network (KDGN) based on the dependency graph incorporating domain knowledge, dependency labels, and syntax path. Experimental results on the benchmarking datasets demonstrate that our KDGN significantly outperforms previous state-of-the-art methods on the ABSA task, further illustrating that the domain knowledge, dependency labels, and syntax path are crucial for the ABSA task.