Multiscale feature aggregation network for aspect sentiment triplet extraction

Aspect sentiment triplet extraction (ASTE) aims to extract all aspect terms with their corresponding opinion terms and sentiment polarity simultaneously from reviews. Recent work processed the ASTE task in an end-to-end manner, which fully utilized the interactive relations among tasks and modeled the interactive relations between words. However, span-level features have not been fully explored. To this end, we propose a novel multiscale feature aggregation network (MSFAN) for end-to-end aspect sentiment triplet extraction (E2E-ASTE), which extracts multiscale local feature representations and explores the deeper interactions between aspect terms and opinion terms. We also design a simple span-awareness representation selection mechanism (SRSM) to further obtain span-level word representations. Extensive experimental results indicate that our model significantly outperforms strong baselines and achieves state-of-the-art performance.