Two-stage three-way enhanced technique for ensemble learning in inclusive policy text classification
Two-stage three-way enhanced technique for ensemble learning in inclusive policy text classification
复制标题
DOI：
10.1016/j.ins.2020.08.051
复制
发表时间：
2021-02-08
影响因子：
8.1
通讯作者：
Yi, Bochun
中科院分区：
文献类型：
--
作者：
Liang, Decui;Yi, Bochun
随着社会经济的发展，中小企业在推动经济发展中发挥着至关重要的作用。中国多个地方政府正在开发政策推荐平台，以帮助中小企业更好地了解包容性政策。但这些线上平台从普惠政策文本中手工提取关键信息，耗时多，效率低。政策文本由若干段落组成，每一段落对应一个主题。当我们将段落划分为不同的主题时，存在文本错误分类的决策风险。因此，我们设计了基于两阶段的三向增强技术来自动将这些文本段落分类到预定义的类别中。在第一阶段，我们使用集成学习算法，构建一个集成卷积神经网络（CNN）模型，以确保文本分类结果的泛化能力和稳定性。同时，我们开发了一种新的权重确定方法，根据准确率和分类置信度来整合所有基分类器的预测结果。利用三因素决策方法，将分辨率较差的样本分配到边界区域进行二次分类，降低了决策风险。在第二阶段，为了对边界区域样本进行分类，提高整体分类结果，我们进一步利用传统的机器学习方法作为二级分类器。最后，我们开发了一些比较实验来验证我们提出的方法。实验结果表明，两阶段三向增强分类框架是有效的，并取得了较好的性能。该方法能够有效支持政策推荐平台的设计，为中小企业提供服务。(C)2020爱思唯尔公司All rights reserved.
With the development of the social economy, small and medium-sized enterprises (SMEs) play a vital role in promoting economic development. Multiple local governments in China are developing policy recommended platforms in order to help SMEs better understand the inclusive policy. However, these online platforms manually extract the key information from the inclusive policy texts, which takes a lot of time and causes low efficiency. The policy text is composed of some paragraphs and each paragraph corresponds to a topic. When we classify the paragraphs into different topics, there exists a decision risk of text misclassification. Therefore, we design two-stage based three-way enhanced technique to automatically classify these text paragraphs into the predefined categories. At the first stage, by using ensemble learning algorithms, we construct an ensemble convolution neural network (CNN) model in order to ensure the generalization ability and stability of text classification results. Meanwhile, we develop a new weight determination method to integrate the prediction results of all base classifiers according to the accuracy and classification confidence. With the help of three-way decisions (3WD), we assign the samples with poor resolution to the boundary area for secondary classification, which can reduce the decision risk. At the second stage, in order to classify the boundary region samples and improve the overall classification results, we further utilize traditional machine learning method as the secondary classifier. Finally, we develop some comparison experiments to verify our proposed method. The experimental results show that the two-stage three-way enhanced classification framework is valid and obtains a better performance. Our proposed method can effectively support the designment of policy recommended platforms and serve SMEs. (C) 2020 Elsevier Inc. All rights reserved.