Single-Pass, Depth-Selective Reading for Multi-Aspect Sentiment Analysis

Researchers introduced DABS, a Transformer-based framework for multi-aspect sentiment analysis.

DABS encodes a sentence once into a reusable depth-ordered representation, letting each aspect query relevant tokens and layers without re-encoding.

On four benchmarks, it achieved competitive performance while cutting end-to-end computation by up to 60% in multi-aspect settings.

The approach addresses the efficiency-versus-expressiveness tradeoff in aspect-term sentiment analysis.