基于A-Capsule的多标签文本分类研究

Multi-label text classification based on A-Capsule

目前在进行多标签文本分类任务中,大部分算法忽视了句子中词语的关键程度以及句子中词与词之间的相互联系。提出一种A-Capsule(Attention-Capsule,注意力胶囊网络)模型,由Capsule网络和Attention机制结合而得到。使文本数据转换成机器可以识别的向量化形式,然后使Multi-Head-Attention机制对不同单词进行学习,以此来确定词语的重要程度,利用Capsule网络的动态路由算法提取文本中的局部空间特征信息,使用分类器进行分类。实验使用的数据集来自今日头条下的数据集,使用其中的新闻标题多标签数据集进行多组对比实验,依据实验结果得出,提出的组合模型在分类任务中,有较好的性能。

At present,in the task of multi-label text classification,most algorithms ignore the critical degree of words in sentences and the interrelation between words in sentences.An Attention-capsule(A-Capsule)model was proposed.The model was A combination of the Capsule network and the Attention mechanism.Then,the Multi-Head-Attention mechanism was used to learn different words to determine the importance of the words.The dynamic routing algorithm of the Capsule network was used to extract the local spatial feature information in the text,and the classifier was used for classification.The data set used in the experiment is from the data set under Toutiao,and the multi-label data set of the news title is used to carry out multiple groups of comparison experiments.According to the experimental results,the proposed combination model has good performance in the classification task.