Research On Sentimental Analysis Of Online Comment Text Based On Deep Transfer Learning

With the rapid development of the Internet,people's consumption behavior has changed accordingly.Consumers can actively choose the commodities they prefer according to their needs.At present,all the mainstream Internet shopping platforms allow consumers to evaluate the goods and services they have purchased.The evaluation information not only shows consumers' subjective emotional attitude towards the goods and services,but also provides important reference for other consumers when making purchase decisions.At the same time,merchants can also collect consumers' comments and conduct emotional analysis,so as to have a deeper understanding of consumers'personalized needs and continuously improve and upgrade the products.It can be seen that comment information has become a precious resource.How to extract and utilize the emotional information is the focus and research direction of the mainstream.At present,the main research methods to solve the traditional affective analysis problems include:affective analysis based on affective dictionary and affective analysis based on machine learning.The core of the analysis method based on emotion dictionary is to construct a high quality manual annotated emotion dictionary.However,for different fields,the emotion words are different,and the same words may express completely different meanings in different fields.Therefore,it is necessary to build the corresponding emotion dictionaries for specific fields.At the same time,new words are emerging in various fields.To ensure the quality of emotional vocabulary,it is necessary to maintain and expand it constantly,which requires a lot of human resources and time costs.The emotional analysis method based on machine learning needs to adopt manually constructed text features and select features based on experience.In recent years,the development of this method has fallen into a stagnant stage,and the final results of the model are often affected by the data set or feature selection mode.Limited or semantic comparatively clear when the amount of data,the traditional although sentiment analysis method can obtain some good results,but face the now vast amounts of data and the diversity of semantic expression,traditional sentiment analysis method is used to solve such problems is no longer the reality,be badly in need of new and effective method was proposed to solve the above problems.In recent years,the emergence of deep learning and transfer learning can solve these problems more effectively.In this paper,an algorithm model combining deep learning and transfer learning is proposed to conduct emotional analysis on online review paper.The deep learning method is used to solve the feature engineering work required by the traditional machine learning emotion analysis.The deep learning method is used to automate this step and learn all features at one time without manual design.At the same time for traditional deep learning words in the process of the training vectors can't accurate representation of context in the field of problem,put forward a preliminary training Language for Text categorization fine-tuning Model(Universal Language Model Fine-tuning for Text Classification),first in the large-scale any unsigned massive Text data set on the work of training sequence generation Model,its purpose is to learn the characteristics of general Language and characters,the training Model can be predicted through the input Text sequences of it below.The pre-trained model in the source domain,including the overall network structure and some of its parameters,is then migrated to the target domain by migration learning.The parameters and weights of the network model transferred from the target domain data set were thawed layer by layer and combined training was carried out on them.The model was fine-tuned continuously to reduce the confusion of the model,improve the robustness of the model,and reduce the overfitting of the model.At the same time,a new model adapted to the task of the target domain was obtained.Finally,at the end of the model output layer,a full connection layer for text emotion classification is spliced to obtain an emotion classification probability,which can be divided into two categories:positive and negative.In this paper,three kinds of online comment data sets of different fields are selected and preprocessed,and the original data sets are divided into training set,verification set and test set.Then the affective classification model proposed in this paper is used for training on the training set and verification set,and experimental results are verified on the test set.Compared with other traditional machine learning and deep learning affective classification algorithms,experimental results show that the proposed method has achieved a significant improvement in classification accuracy.On the verification set,the migration learning can effectively improve the training effect of the model,especially on the small data set,and effectively solve the problem of data dependence.