Multimodal information systems increasingly route generated visual content back through the same vision-language index that informed its production, so the output must remain retrievable by the queries it was meant to serve. When the scene contains entities at vastly different scales, existing language-guided generators condition on a single, globally pooled text embedding and quietly drop scale-specific concepts, breaking concept-query retrieval even when pixel fidelity is high. We formalise this failure as semantic collapse and propose CERES, a closed-loop multimodal indexing framework that builds a three-level semantic pyramid, mines implicit concepts via a co-occurrence-aware router, performs scale-routed cross-attention into a lightweight U-Net generator, and verifies coverage by re-indexing the generated image with the same frozen VLM. A continuously differentiable soft-Jaccard coverage objective returns dense gradients to the 0.39M-parameter generator under explicit non-degeneracy conditions, and coverage is verified by an independent DINOv2 linear probe trained only on external scene and object labels. On four pansharpening benchmarks across seven settings, CERES delivers the new state of the art with the largest gains where scale variation is most extreme. It also improves concept-query retrieval Recall@5 by +14.0 points and image-text mean reciprocal rank by 0.19 over the strongest baseline, showing that the closed loop preserves queryable content rather than self-referential feature consistency.
Guangyuan Dong, Chuang Liu, Yangchen Zeng et al.· 0 citations
Chinese English neural machine translation remains challenging due to substantial syntactic divergence, word-order variation, lexical ambiguity, and cross-lingual semantic mismatch. These challenges often lead to semantic omissions, over-translation, and weak source–target semantic alignment in transformer-based translation systems. Although transformer architecture has achieved remarkable progress, they frequently exhibit limitations in capturing complementary lexical and sentence-level semantic information while offering limited interpretability of the translation process. To address these challenges, this paper proposes DSF–MarianMT, a semantic fusion enhanced neural machine translation framework built upon MarianMT. The proposed framework integrates word-level and sentence-level semantic representations through a dynamic semantic fusion mechanism and employs a contrastive semantic learning objective to improve source–target semantic consistency during training. Experimental evaluation on a Chinese English translation dataset demonstrates the effectiveness of the proposed approach, achieving 36.5 BLEU, 60.2 chrF, 39.8 TER, and a COMET score of 0.78, outperforming the standard MarianMT baseline across all evaluation metrics. In addition to improved translation quality, interpretability analyses reveal reduced attention entropy, lower redundancies among attention heads, and stable token-level semantic learning. Furthermore, error analysis indicates fewer semantic omissions and over-translation errors, while consistent performance is maintained across sentences of varying lengths. These findings demonstrate that the proposed framework effectively enhances semantic representation learning and translation fidelity for Chinese–English neural machine translation.
