融合多层次语义的网络评价方面抽取方法研究

Aspect Extraction Method of Online Reviews Based on Multi-level Semantic Fusion

针对现有的评价方面抽取方法无法充分利用评论文本中字词包含的复杂语义问题,提出了一种融合多层次语义的网络评价方面抽取模型。首先将卷积神经网络(CNN)训练的字符表示与Word2Vec预训练的词向量进行拼接,构建字词融合的特征表示,利用注意力机制对字词融合的特征表示进行重要程度标识,获得融合注意力的特征表示;构建由CNN和双向长短期记忆网络(BiLSTM)组成的混合神经网络,利用CNN的强学习能力提取字、词的局部语义特征信息,利用BiLSTM全局特征提取能力捕捉字、词之间长距离的上下文语义信息,实现多层次语义融合,最后利用条件随机场学习标签之间的约束条件,输出评论文本最优的序列标注结果,并以酒店评论文本为实验数据集,对所提模型与方法进行可行性和有效性验证。结果表明:本模型具有更好的评价方面抽取效果,可以为基于文本的评价研究与分析提供优质的数据源。

Aiming at the issues that the existing methods for aspect extraction cannot make full use of the complex semantics contained in the character-word,an aspect extraction model of online reviews based on multi-level semantic fusion was proposed.First,a feature representation of character-word fusion was constructed by concatenating the character representation trained by the convolutional neural network(CNN)with word vectors pretrained by Word2 Vec.The attention mechanism is used to mark the importance degree of the feature representation of character-word fusion.The feature representations of fusion attention is obtained.A hybrid neural network consisting of CNN and BiLSTM is constructed for realizing multi-level semantic fusion,which uses CNN’s strong learning ability to extract the local semantic feature of character-word,and BiLSTM’s global feature extraction ability to capture the long-distance contextual semantic information of character-word.Finally,the conditional random field is used to learn the constraints between the tags,and the optimal sequence annotation results of online reviews are obtained.Hotel reviews are used as the experimental data set to verify the feasibility and effectiveness of the model and method proposed in this paper.The results show that this model has better extraction effects and can provide high-quality data sources for text-based evaluation research and analysis.