LSOIT: Lexicon and Syntax Enhanced Opinion Induction Tree for Aspect-based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) is a subtask of sentiment classification, and the difficulty is how to capture the sentiment aspect and sentiment polarity pairs in a sentence. Early studies applied serialization models with attention mechanisms to mine sentiment information. These models are simple and effective, but cannot accurately capture sentiment pairs when encountering complex sentences. Recently, scholars have applied dependency information to construct various graph neural networks for the ABSA task. Compared with serialized models, these structured models demonstrate the graph neural network is powerful in capturing information, and further illustrates the syntactic information is effective for sentiment analysis. However, these syntactic models are usually influenced by syntactic parsers, especially for complex sentences. Hence, this paper builds a Lexicon and Syntax Enhanced Opinion Induction Tree for Aspect-based Sentiment Analysis (LSOIT). Specifically, inducing knowledge-aware opinion induction trees for each aspect word applied by reinforcement learning and attention mechanisms that integrate the lexicon knowledge (i.e. sememe knowledge) and syntax knowledge (i.e. phrase structures, and dependency relationships). Finally, we establish graph neural networks on knowledge-aware opinion induction tree for ABSA. Experimental results on four benchmarking datasets (i.e., Rest14, Laptop14, Twitter and MAMS) demonstrate that LSOIT significantly improves 2.21%, 0.47%, and 0.18% on Rest14, Laptop14, and MAMS comparing with state-of-the-art models, respectively. Ablation Study and Case Study manifest that external knowledge is useful, especially for datasets with standardized grammar rules.