Text Sentiment Analysis Of User Online Reviews

In recent years,with the rise of internet penetration and the continuous increase of the scale of internet users,the internet-based industrial models have developed rapidly;Simultaneously the development of self-media has made everyone the subject of information dissemination,which has led to more and more people publishing their attitudes or sentiments about products and services on the internet,accumulating a large number of online reviews containing personal opinions.These online reviews are not only numerous and diverse in style,but also unstructured text data.It takes a lot of effort and time to process them manually.so that it is difficult for other users to understand the overall overview of the reviews.Therefore,it is of great significance to extract emotional information from online review data,judge people's attitudes toward products and services,and determine the focus and emotional tendency of user reviews.Based on this background,this paper takes user online reviews as the research object,and studies two subtasks in text sentiment analysis—aspect extraction task and aspect level sentiment classification task.Firstly,for aspect extraction task,this paper proposes an aspect extraction model based on two kinds of embedding representations and gated recurrent unit.On the basis of domain-general word embeddings representation,this method takes into account that the dataset from different domains have different semantic representations,thus the domain-specific embeddings is introduced to effectively solve the problem of insufficient semantic information extraction in reviews.In addition,the model further introduces an attention mechanism to increase the weight of words related to the aspect words in the reviews,thereby improving the performance of aspect extraction.Secondly,aiming at the problem of inaccurate modeling of context position information and aspect phrases in aspect level sentiment classification,a hierarchical neural network model is proposed for the specific aspect sentiment classification problem.First,the word vectors of each word in aspect phrase are added together to find the average value,which as the vector representation of the aspect,and then concatenate with the word vector of each word in review text,obtain the hidden representation of the sequence through the word-level and sentence-level long short term memory layer,respectively.At the same time,the position information of each context word in review is combined to calculate the attention weight of each word,and a better feature representation is obtained.Experimental results also prove that this method has improved performance compared to other methods.Finally,in order to solve the problem of ignoring the interaction between aspect phrase and contexts in the aspect level sentiment classification task and the modeling of aspectphrases,this paper proposes an interactive attention network model for aspect level sentiment classification.This method uses three bidirectional long short term memory networks to model the three parts of the left context,the aspect phrase and the following context,respectively,to obtain their respective hidden feature representations.Then,combined with the interactive attention mechanism,the feature representations of context under the influence of aspect phrase and the feature representations of aspect phrase under the influence of context are calculated,respectively.The experimental results also prove the effectiveness of this method in aspect level sentiment classification.