Research And Implementation Of Aspect-based Sentiment Analysis Algorithm Based On Transfer Learning

In the past,due to the limited computing power,semantic analysis tasks can only obtain limited and rough text information in the training of task model.However,with the introduction of attention mechanism and pre training model,more abundant language and general knowledge can be obtained in the text representation stage.In the specific domain task scenario,through the transfer training of the pre training model,the knowledge in the specific domain can be obtained,which greatly enhances the downstream task effect of natural language processing.The traditional analysis based on text and sentence level can not meet the needs of text semantic analysis in reality.For the content of the text needs more fine-grained analysis,the emergence of aspect level semantic analysis is to obtain the attitude of specific aspects or terms in the sentence.From this point of view,there are a lot of researches on aspect level sentiment analysis tasks.Aspect level sentiment analysis task is the most basic aspect level semantic analysis task,whose goal is to judge the emotional polarity of the aspect mentioned in the sentence.The purpose of this paper is to get the best text representation model through transfer learning,and then to construct an aspect level sentiment classification neural network to analyze COVID-19 related tweets on twitter.The main work is as follows:1)Experiments are designed to compare the AEN classification network model and LCF classification network model.With the help of attention mechanism,these two network models can more accurately obtain the aspect information and related emotional tendency information in the text.With the basic version of the best model used for text representation,experiments are carried out on three benchmark task datasets,and the control experiment models td-lstm,tc-lstm and best-spc are set.The LSTM model is represented by GLo Ve model.The experimental results show that the effect of text representation model based on BERT is better than that of LSTM model based on Glo Ve,although LSTM model also focuses on the local context information of aspect words.On the Laptop and Restaurant datasets that include the evaluation of goods and services,the BERT-SPC model is the best one among the best.On Twitter's short text social media data set,the AEN and LCF models are the best,so we choose AEN and LCF as the model infrastructure of the final model.2)In order to enhance the text representation and obtain knowledge in a specific field,we use a large-scale corpus collected from Twitter to transfer and train the BERT model,and get the BERT model containing specific domain knowledge,so as to improve the performance of the classification model.According to the conclusion that the classification performance on social media of AEN and LCF model is better in the previous part of the experiment,we use COVID-BERT as the text representation pre-trainied model,and carry out the experiment on the benchmark classification task semeval4 task 4 to verify its classification performance.The experimental results show that the accuracy of AEN is73.12%.3)Based on the best model of the first two parts,cowid-aen-bert,this model is used in the speech analysis system of epidemic related.As a module algorithm of sentiment Tendency Classification,the model classifies the speech and puts the results into the database.According to the needs of display,it sorts out the speech in the back-end interface and returns it to the front-end.This paper also introduces the architecture design of the analysis system and the function design of each module,to clarify the role of the analysis model,and to verify the function of the system through the functional requirements and non functional requirements.