基于深度BLSTM和分类元数据的自定义情感分类

Customized Sentiment Classification Based on Deep BLSTM and Categorical Metadata

在传统的情感分类任务中,存在无法有效捕捉文本深层特征的问题,同时也存在不考虑如用户信息和产品信息等分类元数据而直接进行粗糙建模的问题.针对第一个问题,本文首先通过深度BLSTM(DBLSTM)来识别上下文词义联系和获取文本深层特征;其次利用自注意力机制网络层捕获文本中重要的特征.针对第二个问题,本文融合分类元数据自定义分类器,该分类器利用上下文感知注意力为分类元数据配制特定参数,这使得分类器可以参考文本中存在的不同分类元数据来对网络层提取到的特征做出综合评价分类.在Yelp2013、Yelp2014、IMDB等三个数据集上测试,实验结果显示,本文构建的模型与现有的多个基线情感分类模型相比效果均有一定的提高.

In the traditional sentiment classification task,there is a problem that the deep features of the text cannot be effectively captured,and there is also a problem of directly performing rough modeling without considering categorical metadata such as user information and product information.For the first problem,this paper firstly identifies the contextual meaning and the deep features of the text through deep BLSTM.Secondly,it uses the self-attention mechanism network layer to capture the important features in the text.For the second problem,This article combines categorical metadata custom classifier.The classifier uses context-aware attention to formulate specific parameters for categorical metadata,which enables the classifier to make a comprehensive evaluation classification of the features extracted by the network layer by referring to different categorical metadata existing in the text.Tested on three data sets,such as Yelp2013,Yelp2014,and IMDB,the experimental results show that the model constructed in this paper has a certain improvement compared with the existing multiple baseline sentiment classification models.