[extensive reading of papers] sentimental analysis of online reviews with a hierarchical attention network

Thesis title ： Online comment emotion analysis based on hierarchical attention network
Time ：2017
source ：ACM
Address of thesis ：

Abstract ：
   Images may not contain obvious emotional information , But the visual aspect of images can enhance the emotional information of text comments . Based on this idea , This paper presents a visual aspect of attention 、 The hierarchical attention network of sentence attention and self attention , Provide effective emotional analysis for online comments . Through this mechanism , We can model the interaction between words in sentences and between text and images . We are right. Yelp.com Experiments were carried out on the data set of online restaurant reviews , And compare our model with the existing 5 Three models are compared . The results show the superiority of our proposal .
Innovation points ：
  (1) We propose an online comment oriented hierarchical attention network , The network can simultaneously process images and texts in comments , To produce effective emotional polarity . Specially , We use self - attention as the basic coding layer , Capture interactions between words that are far away , To achieve information expansion .
  (2) We propose a dense layer to connect the document representation generated by visual attention and sentence attention , To avoid ignoring some important sentences in emotion analysis .
  (3) We are from Yelp.com A real restaurant review data set was tested , Our model is compared with various baselines . Experimental results show that , This method can improve the accuracy of emotion recognition , And it can be extended to other scenes where images reflect the main content of the text .

   The rest of this paper is structured as follows . The second part gives a brief literature review of the research progress in recent years . The third section describes the proposed model . The fourth section reports the experimental results , Section V summarizes the full text .