中文科技政策文本分类：增强的TextCNN视角*

Research on Text Classification of Chinese S T Policies: Enhanced TextCNN Perspective

摘要:

近年尽管针对中文本文分类的研究成果不少，但基于深度学习对中文政策等长文本进行自动分类的研究还不多见。为此，借鉴和拓展传统的数据增强方法，提出集成新时代人民日报分词语料库（NEPD）、简单数据增强（EDA）算法、word2vec和文本卷积神经网络（TextCNN）的NEWT新型计算框架。实证部分，基于中国地方政府发布的科技政策文本进行了算法校验。实验结果显示在取词长度分别为500、750和1 000词的情况下，应用NEWT算法对中文科技政策文本进行分类的效果优于RCNN、Bi-LSTM和CapsNet等传统深度学习模型，F1值的平均提升比例超过13%。同时，NEWT在较短取词长度下能够实现全文输入的近似效果，可以部分改善传统深度学习模型在中文长文本自动分类任务中的计算效率。

Abstract:

In recent years, although there are many research outputs on the classification of Chinese text, there are still very few publications involving automatic classification of Chinese policy texts based on deep learning. Based on the current studies, a new computing framework-NEWT is proposed, which integrates NEPD (New Era People's Daily Segmented Corpus), EDA (Easy Data Augmentation), Word2Vec and TextCNN. In the empirical analysis, the text of science and technology policy of Chinese local government is extracted, and the classification experiment is conducted. The experimental results show that the NEWT algorithm is better than the traditional deep learning models such as RCNN, Bi-LSTM and CapsNet when the length of words is 500, 750 and 1 000, respectively. The average increase ratio of F1 value is more than 13%. At the same time, NEWT can achieve the approximate effect of full-text input under a relatively short word length, which can partially improve the computational efficiency of the traditional deep learning model in the task of automatic classification of Chinese long text.