| Comparative text can convey information regarding commonalities and differences between products.Comparative opinion mining can not only help enterprises mine competitive intelligence,but also help enterprises identify risks and potential markets in the early stage.With the accumulation of network big data,product reviews of e-commerce platforms are often used as experimental data in the field of comparative opinion mining.However,accurately identifying the text conveying comparative opinions in network big data is not only a prerequisite for comparative opinion mining,but also an important task.However,the characteristics of low density of network big data and the diversity of comparative text make comparative text recognition a difficult task.Prior research generally extracts features by combining manually predefined keywords and rule-based method,and then uses these features to train machine learning models.However,these methods are time-and cost-consuming,present weak generalization capability and weak robustness,and are semantic-unaware.Although the deep learning model can effectively alleviate these problems,due to the diversity of Chinese comparative sentences,a single depth model cannot cover all types of comparative text.For different types of samples,different depth learning models can extract different information features,so they can complement each other.Therefore,we propose a novel neural network architecture named DRFA（Document Representation Fusion with Attention）for identifying comparative text from large amounts of product reviews on E-commerce platforms.Our developed method adopts the attention mechanism to fuse various representations derived by applying LSTM,CNN,and BERT.A series of experiments verified the effectiveness of fusing these different representations and the utility of the attention-based fusion strategy,in terms of improving predictive performance and enhancing robustness.This study has implications for both researchers and practitioners.Finally,The F1-score of the comparative sentence recognition model proposed in this paper reaches 81.76%and 60.85%in the two data sets.Compared with the traditional comparative sentence recognition methods and deep learning methods,the drfa model has better recognition effect and higher robustness.At the theoretical level,this study can provide reference for researchers in the field of comparative sentence recognition,and is beneficial to the follow-up research on comparative relation extraction;At the practical level,this research will be beneficial to assist consumers in decision-making,and assist enterprises in competitive analysis and product positioning. |
