【作者】 王晓明；

【导师】 李建平； 刘国平；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 深度学习作为目前人工智能领域主流研究方向之一,在各种分类任务中都取得了非常不错的效果,其中就包括中文文本分类任务。中文文本分类任务根据操作流程一般包括数据预处理、中文分词、向量表示以及文本分类等关键技术。本文主要研究基于深度学习的中文分词技术以及文本分类技术,并通过对这两种关键技术的改进,提升各自任务的效果。最后结合这两种改进的关键技术从整体上提升中文文本分类效果。本文的工作内容主要包括以下几个方面:1、结合BERT与GRU模型的优势,改进得到一种新的多标准中文分词模型。在传统多标准分词模型中只使用了Bi-LSTM模型,但是随着数据集的增大训练时间也会增加,因此通过使用更简单的Bi-GRU模型来加快模型的训练速度。同时为了从文本中提取出更丰富的语义特征,使用BERT预训练模型作为语义特征提取层加入到模型中,以此提升模型分词效果。根据上面两项改进得到一种新的多标准中文分词模型,并设置相关实验,从训练时间以及分词效果两个方面的提升证明改进的有效性。2、在短文分类模型中加入计算机视觉领域的混合域注意力模块,改进得到一种新的短文本分类模型。传统短文本分类中的卷积神经网络在提取特征时对所有特征一视同仁,为了增强模型对关键特征的提取能力,仿照计算机视觉中的做法,将混合域注意力模块加入短文本分类模型中。通过设置与原模型的对照实验,证明了加入混合域注意力模块对文本关键特征的提取确实很有帮助。3、在层次Attention模型中应用多通道机制,改进得到一种适合中文长文本数据的多通道层次Attention分类模型。使用层次Attention模型进行中文文本分类任务时,长文本增加了分词错误的概率继而产生信息的丢失,于是加入对字粒度文本表示的特征提取通道来补充词粒度文本表示带来的损失。同样设置与原模型的对照实验,通过分类效果的提升证明了加入字粒度文本表示后提取的特征更加全面。4、通过将上面中文分词技术以及文本分类技术的改进进行结合,设计并实现了一个简易的中文长短文本混合分类系统,并利用该系统测试上述两种关键技术的改进在结合后,整体上对中文文本分类效果的影响。更多还原

【Abstract】 As one of the mainstream research directions in the field of artificial intelligence,deep learning has achieved pretty good results in various classification tasks,including Chinese text classification task.The key technologies in Chinese text classification task generally include data preprocessing,Chinese word segmentation,vector representation,text classification and so on.The Chinese word segmentation technology and text classification technology based on deep learning are the focus of this study.In order to improve the performance of Chinese text classification tasks,this paper has improved the two key technologies.Finally,a simple Chinese text hybrid classification system is designed and implemented by combining the two improved technologies.The work of this paper mainly includes the following aspects:1、Combining the advantages of BERT and GRU models,a new multi-criteria Chinese word segmentation model is proposed.In the traditional multi-criteria word segmentation model,only the Bi-LSTM model is used.The training time will increase because of the bigger data set,so the simpler Bi-GRU model is used to speed up the training process of the model.At the same time,in order to extract richer semantic features from the text,BERT,a pre-trained model,is currently added to the model as a semantic feature extraction layer.According to the above two improvements,a new multi-standard Chinese word segmentation model is proposed,and a control experiment is set up.The improvement of the training time and the word segmentation effect proves the effectiveness of the two improvements.2、A mixed domain attention module in the field of computer vision was added to the short text classification model,and a new short text classification model was proposed.The convolutional neural network in traditional short text classification treats all features equally when extracting features.In order to enhance the model’s ability to extract key features,following the practice in computer vision,a mixed domain attention module is added to the short text classification model.By setting up a control experiment with the original model,it is proved that adding the mixed domain attention module is really helpful for extracting key features of the text.3、A multi-channel hierarchical Attention model was proposed by applying a multi-channel mechanism to the hierarchical Attention model.When using the hierarchical Attention model for Chinese text classification tasks,Chinese long text increase the probability of word segmentation errors and then cause the loss of information.Therefore,a feature extraction channel for character-level text representation is added to supplement the loss caused by word-level text representation.The paper set up a comparison experiment between the improved model and the original model,and the improvement of classification effect proves that the feature extracted through adding character-level text representation is more comprehensive.4、A simple long and short text classification system was designed and implemented by combining the above improvements in Chinese word segmentation technology and text classification technology in Chinese text classification tasks.更多还原