基于自注意力的BiGRU和SVM的观点句识别

Opinion sentence recognition based on self-attention BiGRU and SVM

现有的观点句识别方法大多依赖于人工的特征选择,并且提取的数据稀疏。针对这些问题,提出一种基于自注意力双向门控循环单元(BiGRU)和支持向量机(SVM)相结合的方法。首先,将词向量输入到BiGRU中,引入自注意力机制,为BiGRU的隐藏层状态提供求和权重,使之与隐藏层状态相加权,将句子语义的不同方面分别提取到多个向量表示中,形成二维的句嵌入矩阵;然后,将矩阵转化成向量形式,输入到SVM分类器中输出分类结果。与SVM、LSTM和自注意力BiLSTM模型相比,该方法能够提取句子的关键特征,提高观点句识别的精确率。

Most of the existing opinion sentence recognition methods rely on manual feature selection,and the extracted data is sparse.Therefore,a method based on self-attention is proposed,which combines bidirectional gated recurrent unit(BiGRU)and support vector machine(SVM).Firstly,the word vector is input into BiGRU.The self-attention mechanism is introduced to provide for the hidden layer state of BiGRU and summation weight,which is weighted with the hidden layer state.Different aspects of the sentence are extracted into multiple vector representations to form a two-dimensional sentence embedding matrix.Then,the matrix is converted into vectors,and being input into the SVM classifier to output the classification result.Compared with SVM,LSTM and self-attention BiLSTM models,this method can extract key features representing sentences and improve the accuracy of opinion sentence recognition.