Research On Hierarchical Text Emotional Classification Based On Deep Learning

With the rapid popularization of Internet technology and the continuous emergence of various public opinion platforms,people gradually express their opinions and pour out their emotions on these network platforms.Therefore,a large number of network comment texts have been produced.These texts contain rich information.How to tap the value empowerment of text information has also become a research hotspot.Text emotion analysis in the field of natural language processing is a technology to mine the value of massive text,in which text emotion analysis based on deep learning technology is the focus of current research.Although the text emotion analysis model based on deep learning technology is emerging and widely used,there are still many technical points that can be improved.For example,the natural language pre training model includes static word embedding technology and dynamic word embedding technology.Among them,the static word embedding technology has the problems of insufficient word semantic representation and unable to solve polysemy,Then,the emergence of dynamic word embedding technology can solve the problems of static word embedding methods and obtain better text quantitative representation,however we still need to explore how to further use these representations to better complete relevant tasks;From the perspective of human language characteristics,the emotional tendency of a text is determined by some specific words in the text,or a sentence has multiple aspect words,and each aspect word has corresponding emotional words.Therefore,it is necessary to design a model to focus on the important words of the text and filter irrelevant information;Furthermore,the current mainstream time series model LSTM network has the problems of high computational complexity and can not be parallelized training.Therefore,it is necessary to find a time series model,which can not only meet the needs of tasks,but also better extract semantic information and has less computational complexity;In order to solve the above problems and further improve the overall performance,the main work and innovations of this thesis are as follows:(1)The time series model of simple recurrent unit(SRU)similar to LSTM network is used.It is spliced to show that it not only has the ability of fusion semantic coding similar to LSTM network,but also parallelized computing,which effectively reduces the computational complexity and training time,and conforms to the downstream attention mechanism calculation.(2)The attention mechanism is used to calculate the attention weight of each word,which is very consistent with the characteristics of human language.The overall emotional tendency of a sentence and the emotional tendency of a certain aspect of the sentence are largely determined by some words.The attention mechanism is used to better reconstruct the temporal information of the text and reduce the performance loss of text emotion analysis.(3)In order to carry out sentence level emotion analysis and aspect level emotion analysis tasks,further capture sentence semantics and pay attention to words related to the correct prediction results,two hierarchical models of text emotion analysis combined with BERTBi SRU-AT and BERT-Bi SRU-AMM are proposed to complete the task.Experiments on various public data sets show the effectiveness of the model,and the model can obtain higher accuracy.It is verified that the introduction of dynamic word embedding technology,two-way simple cycle model and attention mechanism can effectively improve the overall performance of the model.