Store, share and transfer: Learning and updating sentiment knowledge for aspect-based sentiment analysis

Previous studies have shown that incorporating sentiment knowledge (e.g., sentiment scores) is effective for aspect-based sentiment analysis (ABSA). However, sentiment knowledge is used to create static features, which cannot be propagated over an entire corpus. Unlike previous researchers, we designed a corpus-level sentiment knowledge fusion mechanism with storage, update, and sharing functions, which can help the model to better understand the sentiment information of various opinion words in the dataset. Specifically, we first constructed a dependency graph for each sentence and refined the weights of the edges by the relative distance between the aspect terms and context words. We then introduced two special sentiment knowledge nodes in the graph to establish connections with opinion words by leveraging external sentiment lexicons. We set these two nodes to be globally shared and updatable, which allowed the model to learn corpus-level and domain-specific sentiment knowledge. This knowledge can help the model to generate a better aspect representation that contains rich contextual information and sentiment knowledge. Extensive experiments were conducted on several public datasets, and the experimental results demonstrated the effectiveness of our method. We also analyzed the performance gains from using learned corpus-level sentiment knowledge to transfer across different datasets.