Research On Aspect Level Emotion Analysis Based On Deep Learning

In natural language processing,aspect level emotion analysis is a very important work.Most of the previous studies made an overall emotional judgment on a sentence in the text,ignoring that each sentence may contain different aspect words.With the development of deep learning,aspect level emotion analysis based on deep learning effectively solves this problem and improves the classification effect to a certain extent.However,in the current aspect level emotion research based on deep learning,there are some problems in the text data,such as single selection model,short sentences and fuzzy expression of emotion,so it is difficult to extract their features,and thus it is impossible to judge the emotional polarity of aspect words.Or there are problems such as only modeling the context of the text,and the pre training model cannot express polysemy.Therefore,the main research contents proposed in this dissertation are as follows:(1)The traditional aspect level emotion analysis only researches by generating context word vectors.However,some text comments have the problems of few words and vague expression of emotion,so the traditional methods cannot fully extract the feature information of sentences.In order to solve the above problems,this dissertation proposes an aspect level emotion analysis model based on emotional characteristics and attention mechanism.Firstly,the model splices the text word vector and the emotional feature word vector as input,which expands the amount of information that the word vector takes as input;Secondly,the convolutional neural network CNN and bidirectional short-and long-term memory network BiLSTM are combined to obtain local information and contextual semantic information of the text;Finally,the attention mechanism is introduced,focusing on words that have an important impact on the classification results,and the Softmax classifier is used for classification.Experiments show that this model is better than other comparison models in accuracy and F1 value.(2)The traditional pre training model generates static word vectors,so it can not express polysemy,and only considers the context modeling,ignoring the modeling of aspect words,so it can not learn the correlation between context and aspect well.In order to solve the above problems,this dissertation proposes a BERT aspect level emotion analysis model based on interactive attention mechanism.First,the model generates context and aspect word dynamic word vectors through BERT,which can effectively solve the problem that static word vectors cannot express polysemy words;Secondly,the text word vector and aspect word vector are used as the input of Bi directional Gated Recurrent Unit(BiGRU);Then,the interactive attention mechanism is added to learn the correlation between context and aspect words;Finally,the representation of aspect words and context is reconstructed and classified by Softmax classifier.Experiments show that this model is better than other comparison models in accuracy and F1 value.Figure[23] Table[10] Reference[75]...