面向商品中文评论的方面抽取研究

Research on Aspect Extraction for Commodity Review in Chinese

文本提出一种基于预训练语言模型双向Transformers编码表示(Bidirectional encoder representation from transformers,简称BERT)的商品中文评论方面抽取模型。首先,利用BERT模型对商品评论文本进行词嵌入;然后,利用BiGRU网络对词向量进行特征提取以获得文本特征,再通过注意力机制为每个字词赋予不同的关注度;接着,将文本特征和关注度融合得到新的文本特征;最后,将模型输出输入到CRF层中,抽取出评价对象的有关方面。该模型能够在编码阶段充分学习到词语的语义,实验结果表明,本文提出的方法提高了抽取准确度。

This paper proposes a BERT based extraction model of commodity review in Chinese.First,the BERT model is used to embed words in commodity review text.Then,the BiGRU network is used to extract the embedded word vectors to obtain text features.After that,new text features are obtained by combining the previous text features and attention.Finally,the model output is input into the CRF layer to extract the relevant aspects of the evaluation object.This model can fully learn the semantics of words in the embedding stage.Experimental results show that the proposed method improves the accuracy of extraction.