知识与财富的链接
金融文本多标签分类算法可以根据用户需求在海量金融资讯中实现信息检索。为进一步提升金融文本标签识别能力,建模金融文本多标签分类中标签之间的相关性,提出基于图深度学习的金融文本多标签分类算法。图深度学习通过深度网络学习局部和全局的图结构特征,可以刻画节点之间的复杂关系。通过建模标签关联实现标签之间的知识迁移,是构造具有强泛化能力算法的关键。所提算法结合标签之间的关联信息,采用基于双向门控循环网络和标签注意力机制得到的新闻文本对应不同标签的特征表示,通过图神经网络学习标签之间的复杂依赖关系。在真实数据集上的实验结果表明,显式建模标签之间的相关性能够极大地增强模型的泛化能力,在尾部标签上的性能提升尤其显著,相比CAML、BIGRU-LWAN和ZACNN算法,该算法在所有标签和尾部标签的宏观F1值上最高提升3.1%和6.9%。
Multi-label financial text classification can retrieve relevant information from massive financial news according to user needs.To further improve the performance of multi-label financial text classification, this study proposes an algorithm to model the correlation between labels based on graph deep learning.Graph deep learning can describe the complex relationships between nodes by learning local and global graph structure features through deep neural networks.Modeling the correlation between labels can realize knowledge transfer between labels, which is key to constructing an algorithm with strong generalization ability.Therefore, this study utilizes graph neural network to learn the complex dependency between labels based on statistical information along with feature representations extracted using the bi-directional gated recurrent network and label attention mechanism. Experimental results on real world datasets show that modeling label correlations can significantly improve the classification performance, especially on tail labels.Compared with CAML, BIGRU-LWAN and ZACNN algorithms, the proposed algorithm improves the macro F1 values of all labels and tail labels up to 3.1% and 6.9%.
