Two-Channel News Title Classification Model Based on Ernie

News title categorization is a crucial domain in Natural Language Processing (NLP), marked by brevity, distinctiveness, and varied structures. To overcome the shortcomings of conventional neural networks that often fail to fully capture intrinsic feature information and to simultaneously extract both key global and detailed local features, we introduce a dual-channel network model named EDCGA, which is constructed upon the ERNIE pretraining framework. This model leverages ERNIE's capabilities to extract lexical, semantic, and contextual features, creating dynamic word vectors that are contextually integrated. Following this, the DPCNN component tackles the challenges posed by long-range text dependencies to extract comprehensive local features. The BiGRU-Attention channel's weights aid in the secondary extraction of global features, with the attention mechanism prioritizing the most significant elements. The combined local and global feature vectors are subsequently processed through a fully connected layer, and the Softmax function is applied to determine the final classification outcomes. Empirical evidence suggests that our proposed model outperforms traditional and single-channel neural network models in terms of accuracy, precision, and F1-score for news title classification. The model shows robust performance in classifying news titles in large-scale datasets.