Research On Fine-grained Sentiment Analysis Based On Sequence Generation Model

With the rapid development of Internet technology people are spending more and more time on the Internet everyday.Tremendous amount of information is generated and spread all the time,among which the commentary information is of great importance for commercial decision making and public opinion observation and monitoring by relevant government.Due to the fact that towards the same object the commentary text usually contains different aspects of emotions,the traditional coarse-grained emotional analysis method cannot make accurate deconstruction of document-level and sentence-level sentimental text.Based on the current demand situation this paper has developed an aspect level fine-grained sentiment analysis method for comment text focusing on the specific aspect of emotional inclination.From the comparable research on models that have better performance,we have found that by treating the sentence as a combination of word sequence and training text word vector based on its recurrent neural network learning sequence features,the word2 vec can more effectively find the inclination of text from the hidden state.By introducing attention mechanics,the information that is irrelevant with targeted vector in the hidden state can be filtered when synthesizing the contextual information of the implicit state sequence,thereby improve the classification accuracy of the model.However,there are some limitations.First is that word2 vc word vector cannot express the polysemy of the word.Second,the single-layer RNN model has long-term reliance problem that it cannot effectively learn the deep semantic information of long texts.Since this method is basically a single-objective classification method,it needs to repeatedly train multiple classifiers when operating multi-objective classification tasks that require fine-grained sentiment analysis.Besides,such method ignores the correlation between each aspects of emotions.Based on such conditions the ELMo vector is proposed to both replace the word2 vec and its fine-grained emotion classification method that generate model from aspect sentiment sequences.The main structure of this paper is as follows:(1)The replacement of the original word2 vec to the ELMo.By taking advantage of the structural advantages of the ELMo model,the word vector can be obtained from the linear combination in the hidden layer state and the context,thus the problem that the fixed word vector of word2 vec cannot express multiple meanings of words can be solved.(2)The transformation of fine-grained sentiment analysis into aspect sequence generation tasks through analyzing the characteristics of fine-grained sentiment analysis tasks.To generate models from designing aspect emotional sequence,the Enocoder uses the same Bi-LSTM network structure as ELMo does to effectively learn complex semantic text information by improving the additive attention mechanism and calculating the weighted coefficients using the aspect features as well as relative distances.When using RNN to learn tag correlation we can also get the output of the emotional sequences.This sequence generated model is an exploration of new ideas and new methods for fine-grained sentiment analysis.From the experiments the improved emotional sequence generation model is proved to work more accurately in fine-grained emotional classification for text than the existing Baseline model.