Abstract
Text classification, a fundamental task in natural language processing (NLP), aims to categorize textual data into predefined labels. Traditional methods struggled with complex linguistic structures and semantic dependencies. The advent of deep learning, particularly recurrent neural networks (RNNs) and Transformer-based models, has significantly advanced the field by enabling nuanced feature extraction and context-aware predictions. Despite improvements, existing models exhibit limitations in balancing interpretability, computational efficiency, and long-range contextual understanding. This paper proposes the Dynamic Bidirectional Elman with Attention Network (DBEAN), which integrates bidirectional temporal modelling with self-attention mechanisms. DBEAN dynamically assigns weights to critical segments of input, improving contextual representation while maintaining computational efficiency.
Abstract (translated)
文本分类是自然语言处理（NLP）中的一个基本任务，旨在将文本数据归类到预定义的标签中。传统方法在处理复杂的语言结构和语义依赖方面遇到了困难。深度学习技术的发展，尤其是循环神经网络（RNNs）和基于Transformer的模型的应用，极大地推动了该领域的发展，使得细微特征提取和上下文感知预测成为可能。然而，现有的模型在可解释性、计算效率以及长程上下文理解之间存在难以平衡的问题。 本文提出了一种新的方法：动态双向Elman注意力网络（DBEAN），它将双向时间建模与自注意力机制相结合。DBEAN能够根据输入文本的重要部分动态分配权重，在提升上下文表示能力的同时保持较高的计算效率。
URL
https://arxiv.org/abs/2503.15469
