GEODENSECONVNET: A BREGMAN DIVERGENCE-INSPIRED FRAMEWORK FOR ADAPTIVE FEATURE FUSION IN TEXT CLASSIFICATION

Abstract

Effective text classification requires models to synthesize both local syntactic patterns and global semantic dependencies. While Convolutional Neural Networks (CNNs) excel at ex- tracting hierarchical local features and Attention mechanisms are proficient at modeling long- range context, the optimal method for integrating these complementary paradigms remains an open research question. Most hybrid architectures rely on static fusion strategies, such as concatenation, or simple arithmetic combinations, which lack the capacity to adapt to the input’s specific linguistic properties. In this paper, we introduce GeoDenseConvNet, a novel hybrid architecture that employs a principled adaptive fusion mechanism derived from the fundamentals of information geometry. Our model comprises two parallel branches: a multi- scale CNN for capturing features at different granularities and a multi-head self-attention network for global context. The central contribution is a novel fusion mechanism derived from the optimization problem of finding a Bregman barycenter with the negative entropy generator. This principled approach yields an adaptive fusion rule based on a learned, element-wise

**geometric mean**, which naturally operates in the space of positive feature activation. This allows the model to fluidly arbitrate between local and global information in a theoretically- grounded manner. We conduct extensive experiments on three diverse benchmark datasets: IMDB, AG News, and EmoBank. Results demonstrate that GeoDenseConvNet consistently outperforms standard baselines and achieves competitive performance, highlighting the efficacy of our geometrically-inspired fusion strategy for building more robust and interpret- able text classification models.