Entity-Aware Bidirectional Attention-Based Gated Fusion for News Classification

초록

As the online news environment shifts toward combining diverse media such as text and images, research on multimodal AI-based news classification is underway to address the limitations of information loss in single-modality methods and improve classification accuracy by analyzing complex contexts. However, existing fusion techniques based on simple concatenation or unidirectional interaction struggle to preserve the core context in lengthy texts and fail to resolve the semantic discrepancies caused by irrelevant visual noise, ultimately leading to model bias and degraded classification performance. This paper proposes an entity-aware bidirectional attention-based gated fusion designed to maintain the key textual context and mitigate the bias issues induced by visual noise. The proposed architecture consists of a preprocessing stage that prevents context loss using NER and keywords extracted from unstructured text, a bidirectional cross-attention stage that aligns features through cross-modal referencing, and a gated fusion stage that suppresses noise amplification via modality-specific bias initialization. Experimental results using the New York Times N24News dataset demonstrate that the proposed model improves accuracy by up to 14.04% and F1-score by up to 15.73% compared to existing baseline models, validating its applicability for robust news classification even in environments characterized by semantic misalignment and noise.

키워드