The Research And System Design Of Multi-Classification Of Text Sentiment Based On Pre-Trained Models

In the era of big data and artificial intelligence,whether machines can have emotion has become a hot topic.A real artificial intelligence system should not only have the ability of thinking and reasoning like human beings,but also be able to perceive and express emotions.As an important carrier of emotion perception and expression,language plays an important role in emotion analysis,and it usually exists in the form of text.Therefore,this paper will study emotion analysis from the perspective of text,and apply pre-trained model and multi classification technology to solve the key difficulties and pain points.This paper first introduces the research background and significance of text emotion multi classification,then expounds the implementation scheme status and existing problems of text emotion classification in industry.A text multi category emotion classification system("qianyu" system)is implemented,which provides an important reference for theoretical research and practical application.In terms of method design and system implementation.The main work of this paper includes:(1)It has realized 500000 sentence level data tagging of eight emotion categories.Based on the existing open source data set,we adopts the strategy of regular expression matching and cross verification,which significantly reduces the workload and labor cost of manual tagging.In the part of manual tagging,we has developed an easy-to-use tagging website;(2)Based on the open source pre-trained model,the strategy of pre-trained and fine-tuning is used to train the classifier,which can ensure that the deep neural network classifier obtained under the condition of less training data has high accuracy and generalization performance.Finally,the accuracy of text emotion eight classification reaches 94.4%;(3)The deployment phase of the model server relies on lightweight coroutine and dynamic invocation of model instances in different formats to realize a high-performance server API.With the characteristics of high concurrent access support,it can still greatly reduce the data return delay.On the server equipped with 1080 GPU,it can reach the throughput of 1127qps(1127 requests per second),and the maximum delay is less than 300 ms.The emotion classification system implemented in this paper has good economy,feasibility and practicability,can effectively promote the research of emotion classification technology,and provide an implementation scheme for the industrial implementation of emotion classification technology.