基于交互式叠加注意力网络的实体属性情感分类

当上下文中单词的情感对给定属性敏感时,仅通过注意力建模无法解决情感分类性能下降的问题。提出一种基于交互式叠加注意力(Attention-Over-Attention,AOA)网络的属性级情感分类模型。模型在词向量层用BERT代替传统静态词向量表示;用LSTM分别提取属性和上下文中单词的隐藏语义;用AOA网络计算属性和上下文中每个单词的注意力权重;将权重与对应的隐藏语义状态做点积分别得到属性和上下文的最终特征表示,拼接两个特征表示用来分类。研究并分析模型中词向量和属性单独建模对情感分类结果的影响。实验表明,该模型较其他LSTM结合注意力机制的模型在准确率和F1值上都有显著提高。

SENTIMENT CLASSIFICATION OF ENTITY ASPECTS BASED ON INTERACTIVE AOA NETWORK

When the sentiment of a word in the context is sensitive to a given aspect,only attention modeling can not solve the problem of degraded sentiment classification performance.An aspect level sentiment classification model based on interactive AOA network is proposed.BERT was used to replace the traditional static word vector representation in the embedding layer;LSTM was used to extract the hidden semantics of the words in the aspect and context respectively;AOA network was used to calculate the attention weight of each word in the aspect and context respectively;the weight and the corresponding hidden semantic state were dot product to obtain the final feature representation of the aspect and context respectively,and the two feature representations were spliced for classification.This paper studied and analyzed the influence of word vector and aspect modeling on sentiment classification results.The experiments show that the model significantly improves accuracy and F1 value compared with other LSTM models combined with attention mechanism.