Intra-Attention And Inter-Attention For Aspect-Level Sentiment Classification

With the emergence and development of web2.0,users are free to express their opinions on various e-commerce platforms and social networks.When users express their views on an entity,in addition to giving an overall evaluation in the comments,they usually express their views on specific aspects of the entity.This aspect-oriented sentiment polarity can help other users make informed decisions and help businesses improve their services.Therefore,it has become an important topic to mine the sentiment polarity of user reviews for specific aspects from massive comments,and aspect-level sentiment classification has become a research hotspot in the field of Natural Language Processing.Based on the Sem Eval series of data,this paper constructs a new aspect-level sentiment classification framework based on deep learning networks to achieve the research goal of predicting the user's sentiment polarity for the specified aspects of comments.The work of this paper mainly includes the following three aspects:(1)Comment data preprocessing and acquisition of word vectors.We perform word segmentation on the comment data in the corpus.Because the numbers in the comments have little effect on sentiment classification,we remove the numbers in the comments and separate the punctuation from the words.In addition,because sentiment words in different domains may have different sentiment polarity,we use the Glo Ve algorithm to train a word vector with a dimension of 100 on the corpora of two specific domains and in order to solve the problem of vacancy of rare word vectors,we use Char-CNN to obtain the character-level vector representation of the word,so as to obtain a more accurate review vector representation.(2)This paper uses the idea of attention mechanism to design an aspect-level sentiment classification model for the review data containing aspects in the Sem Eval series dataset.Firstly,the word embedding obtained through the Glo Ve model and the character embedding obtained based on the Char-CNN model jointly build the network's input data——context embedding.Secondly,the intra-attention module is used to capture the related sentiment features between each word.Then,in the interattention module,aspect-sentiment pair attention is used to represent the sentiment features of the document under the specific aspect information and sentiment features,so as to obtain the final vector representation of the review text on a given aspect.Finally,aspect-level sentiment classification is carried out based on the final vector of the obtained comments.(3)We experiment the proposed model on the real comment corpus.First,we used Keras to build our network framework and set the parameters in the network.After determining the experimental parameter settings,this paper chooses five classic algorithms as comparative experiments,and conducts a lot of experiments on four data sets to compare the accuracy of sentiment classification between the proposed algorithm and the benchmark algorithm.Experiments on four real datasets show that our proposed model is superior to the benchmark algorithm in prediction accuracy.