DABS: Single-Pass Depth-Selective Reading for Aspect Sentiment Analysis
The newly introduced framework, DABS (Depth-Adaptive Batch Substrate), tackles the challenge of balancing efficiency and expressiveness in Aspect-Term Sentiment Analysis (ATSA) for sentences with multiple aspects. Current models tend to either re-encode sentences for each aspect or rely on fixed deep representations, leading to inefficiencies. In contrast, DABS processes each sentence a single time into a reusable, depth-ordered substrate, allowing aspects to selectively access pertinent tokens and abstraction levels without the need for re-encoding. This approach separates shared encoding from the lightweight aspect-conditioned readout. Tests on four ATSA benchmarks indicate that DABS not only performs competitively but also cuts computation by as much as 60% in multi-aspect scenarios (M >= 2). The study highlights that Transformer depth represents a valuable, queryable resource.
Key facts
- DABS is a single-pass inference framework for ATSA.
- It encodes each sentence once to construct a reusable depth-ordered substrate.
- Each aspect queries the shared representation to selectively read relevant tokens and abstraction levels.
- It decouples shared sentence encoding from lightweight aspect-conditioned readout.
- Experiments on four ATSA benchmarks show competitive performance.
- End-to-end computation is reduced by up to 60% in multi-aspect settings (M >= 2).
- The paper argues Transformer depth is a costly, queryable resource.
- The approach addresses the efficiency-expressiveness tradeoff in multi-aspect sentences.
Entities
Institutions
- arXiv