Is syntax structure modeling worth? Leveraging pattern-driven modeling to enable affordable sentiment dependency learning

Anonymous

Abstract: Is structure information modeling really worth in Aspect-based sentiment classification (ABSC)? Recent popular works tend to exploit syntactic information guiding sentiment dependency parsing, i.e., structure-based sentiment dependency learning. However, many works fall into the trap that confusing the concepts between syntax dependency and sentiment dependency. Besides, structure information (e.g., syntactic dependency tree) usually consumes expensive computational resources due to the extraction of the adjacent matrix. Instead, we believe the sentiment dependency mostly occurs between adjacent aspects. By proposing the sentiment patterns (SP) to boost the sentiment dependency learning, we introduce the Local dependency aggregating (Lena) to explore sentiment dependency in the text. Experiments show that Lena is more efficient than existing structure-based models without dependency matrix constructing and modeling expense. The performance on all five public ABSC datasets makes a big step compared to state-of-the-art models, and our work could inspire future research focusing on efficient local sentient dependency modeling.

Paper Type: long

Consent To Share Data: yes