Fine-grained Sentiment Analysis Technique On Large-scale Internet Texts

Sentiment analysis is an important task in natural language processing tasks,with broad application prospects and practical value.With the advent of the “Self-Media” era,more and more people are interested in publishing their opinions and emotion online,and with the arrival of an increasing number of online text data,the demand of that automatically handle these large-scale web texts is getting stronger.At present,sentiment analysis of large-scale network texts has the openness,freedom and non-standardity of the topic,which makes the traditional method more complicated to process.The object or attribute of the comment is more difficult to extract,the expression is more concealed,and the expression of the effective context needs more and more.Dealing with text lacks richer linguistic knowledge and many other issues.In view of the above problems,this paper is based on deep learning methods,supplemented by a variety of information fusion techniques,from aspects term extraction,aspect-based sentiment analysis and stance detection analysis.The aspect extraction work mainly solves the problem of recognition of emotional receptors in sentiment analysis,and finds the object of review or its attributes.The aspect level sentiment classification is the emotional orientation judgment of emotional receptors.The stance detection is based on the given sentences.The issuer's positional preference for the specified object or event.The main work is as follows:1.In the research of word extraction problem,we propose a tree-structured neural network model named KBTreeNet which integrates the knowledge base to extract the aspect words,and uses the tree structure information and the emotion existing in the sequence information.For the traditional deep learning,the syntactic structure information of the sentence cannot be used,the contact between the target word and its corresponding emotional word can not be established,and the semantic divergence existing in the common language task can not be solved.This paper proposes a model based on recurrent neural network.Each processing unit uses the vallina LSTM model.The whole model is expanded into a tree structure,which intuitively conforms to the syntactic tree structure.At the same time,the huge training advantage brought by the use of LSTM can more effectively and accurately grasp the information contained in the syntactic structure.In addition,by using the tree syntax structure information obtained above and the context sequence information,a calculation method similar to the residual network is introduced,and the two kinds of information are deeply integrated to provide more sufficient features for the subsequent classifier.In addition,the nouns and adjectives in the sentence are replaced by the action of the attention mechanism.The replacement content is the meaning of the words in the HowNet dictionary,which is more in line with the context,thus solving the problem of confusion.The experimental results show that the KBTreeNet model can effectively extract the aspect words in the comment sentences,.2.Based on the aspect-based sentiment classification algorithm,we propose a language model PBAL that combines multiple neural networks to model the sentiment semantics of sentences while considering the influence of aspects on the emotional representation in sentences.Aiming at the more prominent problem in this task: a given emotional word may consist of multiple words,which may not correctly represent the true meaning expressed by the phrase.The existing attention mechanism only pays attention to the information related to the emotional and aspect words.The omission of other important information,the relationship between the words and the source sentences is not easy to construct and so on.Firstly,for the universality of phrase representation in linguistic expression,this paper avoids the single word by using the convolutional layer in the convolutional neural network to extract the phrases that may be contained in the sentence and directly deal with the semantics from the phrase layer.The semantic bias caused by the addition.This article does not use the attention mechanism in the sentence modeling phase,but uses the entire information of the source statement as input to train the sentence representation.On the issue of the combination of related words and sentences,with the idea of solving the question and answer mechanism,a compound attention mechanism is used to tandem the intrinsic relationship between the two expressions.3.Based on the research of position-based classification algorithm,we propose a network model RSN based on attention mechanism.The model does not appear in the original statement for a given target and thus affects the judgment of the final prediction result.Non-critical information in the original statement The influence of modeling leads to the decline of the accuracy of sentiment analysis.This paper uses the original self-attention mechanism to calculate the correlation between external words and sentences to get the corresponding weights.According to the weighted summation,the sentence with the final external words is obtained.It is expressed in the self-attention mechanism to obtain the sentence representation of the word order;in addition,a hard attention mechanism is used to extract the key information in the sentence,and the reinforcement learning can provide an optimization scheme for the above discrete target problem.