Evidence-Density Adaptive Retrieval for Few-Label Multimodal Entity Extraction

Abstract

The rapid proliferation of multimedia content across digital platforms has necessitated the development of advanced information extraction systems capable of processing heterogeneous data streams. Multimodal Entity Extraction represents a critical task in this domain, aiming to identify and categorize entities by simultaneously interpreting text and associated visual or auditory signals. However, existing architectures heavily rely on large-scale annotated datasets, rendering them impractical for specialized domains where labeling is prohibitively expensive. Furthermore, current retrieval-augmented approaches apply a static retrieval strategy regardless of the inherent informational completeness of the input, leading to computational inefficiency and semantic noise. To address these limitations, this paper introduces an Evidence-Density Adaptive Retrieval framework designed specifically for few-label Multimodal Entity Extraction. By conceptually quantifying the informational density of intrinsic multimodal evidence, the proposed system dynamically determines the necessity, scope, and modality of external knowledge retrieval. The architecture leverages a dual-stream encoder aligned through a contrastive objective, coupled with a density estimation module that computes the semantic saturation of the input. We evaluate the proposed methodology across multiple benchmark datasets under strictly constrained few-label settings. The empirical results demonstrate that our adaptive approach significantly outperforms static retrieval baselines in both extraction accuracy and computational efficiency, mitigating the introduction of extraneous noise while maximizing the utility of a limited demonstration memory bank.