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

2026-05-20

AI Summary

Existing aspect-level sentiment analysis methods often sacrifice efficiency and representational capacity due to redundant encoding or static representations. This work proposes the DABS framework, which, for the first time, treats the depth of Transformer layers as a queryable resource. After a single pass of sentence encoding, DABS employs a lightweight, aspect-conditional querying mechanism to adaptively retrieve relevant tokens from different levels of abstraction, thereby decoupling shared encoding from aspect-specific reading. The approach achieves competitive performance on four ATSA benchmarks while reducing end-to-end computational cost by up to 60%. Notably, it excels in handling complex linguistic phenomena such as negation and contrastive constructions.

Abstract

Aspect-Term Sentiment Analysis (ATSA) in multi-aspect sentences faces a fundamental tradeoff between efficiency and expressiveness. Existing models either re-encode the sentence for each aspect or rely on static use of deep representations, leading to redundant computation and limited adaptivity. We argue that Transformer depth is a costly, queryable resource, and propose DABS, a single-pass inference framework that encodes each sentence once to construct a reusable, depth-ordered substrate. Each aspect then queries this shared representation to selectively read relevant tokens and abstraction levels, without re-encoding. This decouples shared sentence encoding from lightweight, aspect-conditioned readout. Experiments on four ATSA benchmarks show that DABS achieves competitive performance while reducing end-to-end computation by up to 60% in multi-aspect settings (M >= 2). Further analyses indicate that adaptive depth querying is most beneficial for linguistically complex cases such as negation and contrast. Code is publicly available at https://github.com/panzhzh/acl-dabs

Problem

• Aspect-Term Sentiment Analysis
• multi-aspect sentences
• efficiency
• expressiveness
• redundant computation

Innovation

• single-pass inference
• depth-selective reading
• multi-aspect sentiment analysis
• adaptive querying
• efficient encoding