Research On Affective Classification Of Commodity Comments Based On Deep Learning Of Attention

With the rapid popularization of e-commerce,online shopping has become a necessary channel for people to consume.The emotional polarity of online commodity comments is the most direct way to get feedback from customers on such commodities.Merchants can gain insight into customer experience,optimize products and adjust sales plans through analysis of comments.Compared with machine learning and rule learning methods,deep learning does not require artificial construction and extraction of detailed text rules and semantic features,but can extract text features from bottom to top,with significant resource saving effect.In the current task of classifying emotion of commodity review,the quantity and quality of comment text and emotion dictionary are always the difficulties in the research,and the existing models also have some problems such as poor classification effect.Aiming at the existing difficulties,the main research contents of this paper are as follows:(1)Use crawler technology to get real comments from Jingdong Mall,and process the comments to ensure quality.Now widely used corpus set text defect,the types of goods is uneven and other deficiencies,in order to ensure the accuracy and completeness of using corpus set,this article USES the crawler technology from jingdong mall for mobile phones,computers and other electronic product reviews as a data set,will be subject to product reviews star classified storage,convenient for subsequent use.Since the product review data obtained may contain duplicate comments and invalid comments,etc.,more than 20,000 valid corpus were obtained after rule de-duplication and manual verification,and three types of emotions were designed,namely positive,neutral and negative emotions.(2)Construct the Internet word dictionary and the electronic field emotion dictionary.Because the standardization of online comment text is usually weak,jieba,the common word segmentation tool used by NLP,has limitations in the field of its own dictionary,which cannot perfectly segment the comment text.A common online word dictionary needs to be built.Only when the two are used together can the text be properly segmented,providing a good foundation for the subsequent use of word text.Different words in the review text contain different degrees of emotional tendency.In order to accurately identify the overall emotional polarity of the review,emotional tendency should be given to words.At present,there are relatively few affective dictionaries of Chinese characters in the field of emotion classification.Therefore,the mainstream How Net and the Affective Dictionary of Taiwan University are selected as the foundation,and the affective dictionary of network words constructed based on the Internet dictionary is expanded to be the subsequent affective dictionary of electronic products.(3)Improve the structure of convolutional neural network.The convolutional layer is optimized and several dimensional convolution kernels are used to obtain multidimensional semantic features.In order to avoid losing a lot of useful features,k-Max pooling is adopted to improve the pooling operation of pooling layer.By adding skip layer to introduce original semantic features,the convolutional layer plays an information supplement role in extracting features.As a part of XL-GSATMSC algorithm,the improved convolutional neural network is used to fully learn the implicit text characteristics of commodity comments.(4)The joint word vector was constructed and applied to THE XL-GSATMSC model.The current natural language processing tasks are mostly use Word2 Vec and Glove to advance,training term vectors for Word2 Vec when word vector generation can't solve the problem of polysemy,this article selects the Word2 Vec combined with XLNet way to generate a word vector,guarantee term vectors can better reflect the text semantic relationship in space,and combining with the words in the dictionary of emotional words emotional polarity build contain a combination of semantics and the emotion word vector.Due to the defects of CNN and RNN in text emotion classification task,this paper constructs a new model named XL-GSATMSC.Because the cyclic neural network is unable to handle long statements and parallel operations,this paper uses gated loop unit combined with attention mechanism to extract long distance features,which can not only capture the long distance dependent features of text,but also reduce the low computing efficiency caused by the accumulated memory of gated loop unit.Then,the features already learned in the gated loop unit are extracted by MSCNN to extract the short-distance features,so as to fully learn the implicit text features in the commodity review.The classification results can be used in the user recommendation system to find potential customers for the product.