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.

Multimodal rumor detection with external evidence faces the challenge of dynamically assessing evidence of varying quality and relevance. We propose Dynamic Evidence Reliability Modeling (DERM), a framework that unifies evidence quality assessment and semantic relevance evaluation through information-theoretic principles. DERM introduces three synergistic components: (1) An information density gating mechanism that distinguishes high-quality evidence via cross-modal consistency; (2) A Semantic Alignment Module that adjusts evidence weights by jointly considering information density and claim-evidence similarity; (3) A multi-objective optimization with alignment regularization that promotes sparse attention while preserving semantic grounding. Experiments on Weibo-MR2 and Twitter-MR2 demonstrate DERM’s state-of-the-art performance. Our method achieves 91.4% and 87.8% accuracy, outperforming the strongest baseline by 0.9 and 0.5 points. Notably, DERM boosts rumor class F1 scores by 0.4 points on both datasets, reaching 88.2% and 75.5%. Visualizations further confirm our design: high-quality evidence receives nearly three times more attention than noisy counterparts, enhancing interpretability.