Attribute-level Opinion Analysis For Products Based On Attention Mechanism

Nowadays,people are used to trading products on e-commerce platform.When consumers browse online,they will notice some characteristics of product attributes,so it is essential to analyze users' opinion on these attributes.Fine-grained opinion of users have a significant valuable reference for the potential customers and manufacturers.However,with the rapid growth of e-commerce,it is impracticable to analyze such massive data manually.The main work of this article focuses on the attribute level opinion mining of products,aiming to analyze more detailed and specific opinion on a specific aspect of products.To this end,a product-oriented attribute-level sentiment analysis method is proposed,and further studies the elimination of redundant attributes.The contents are as follows:(1)In order to fine-grained mining the opinion of the products,a hybrid attention model is proposed for attribute level sentiment analysis in this paper,which only utilizes the attention mechanisms rather than recurrent or convolutional structures.In this model,a self-attention mechanism and an aspect-attention mechanism are design for generating the semantic representation at the word and sentence levels respectively.Two auxiliary features of word location and part-of-speech also explored for the proposed models to enhance the semantic representation of sentence.Extensive experiments are conducted on three benchmark datasets for attribute level sentiment analysis.Experimental results show that the proposed models outperform the baseline methods on both efficiency and execution effectiveness.(2)Due to the fact that the attributes of product usually redundant and inconsistency,in the further research,the elimination of redundant information in product titles is studied.According to the characteristics that the commodity titles have weak syntax but strong semantics in structure,we proposed a self-attention based model,which can eliminate redundant information at semantic level via a binary classification method.And we also explored two improving strategies based on GRU and gating mechanism.A series of experiments are conducted on the e-commerce datasets with different redundancy constraints.The results show that the model we proposed can effectively eliminate the redundant information of product titles.