摘要
Multimodal named entity recognition(MNER)and relation extraction(MRE)are key in social media analysis but face challenges like inefficient visual processing and non-optimal modality interaction.(1)Heavy visual embedding:the process of visual embedding is both time and computationally expensive due to the prerequisite extraction of explicit visual cues from the original image before input into the multimodal model.Consequently,these approaches cannot achieve efficient online reasoning;(2)suboptimal interaction handling:the prevalent method of managing interaction between different modalities typically relies on the alternation of self-attention and cross-attention mechanisms or excessive dependence on the gating mechanism.This explicit modeling method may fail to capture some nuanced relations between image and text,ultimately undermining the model’s capability to extract optimal information.To address these challenges,we introduce Implicit Modality Mining(IMM),a novel end-to-end framework for fine-grained image-text correlation without heavy visual embedders.IMM uses an Implicit Semantic Alignment module with a Transformer for cross-modal clues and an Insert-Activation module to effectively utilize these clues.Our approach achieves state-of-the-art performance on three datasets.
