Research On Method Of Short Text Sentiment Classification Based On Weak Supervision

Sentiment analysis is a computational study of the emotional attitudes of an entity in a text.Emotional classification is one of the main research areas.Traditional sentiment classification methods rely on artificially designed features,while the current more popular deep learning methods can automatically extract text features and perform well in solving short text sentiment classification problems.The deep learning method needs to provide enough monitoring data to train a good model.However,the supervision data needs a lot of manual labeling and is not easy to obtain.In order to solve this problem,a method of training deep learning sentiment classification model using large-scale weak supervised data is proposed,and the validity of the model and the feasibility of using the weak supervised data to replace the supervised data to train the deep learning model to some extent are proposed.The specific work is as follows:1.A short text sentiment classification method based on weak supervision model is proposed.Using the commentary information with scoring data on the Internet as the large-scale weak supervised training data set,a deep learning model was established,and the training model was adopted by the “weak supe rvised pre-training-supervised micro-adjustment” strategy.Experiments show that weakly supervised data can be used as an effective alternative when monitorin g data is lacking,and validates the validity of the model and the “weak-superv ision-micro-adjustment” strategy.2.To further reduce the influence of noise on the deep learning model in weak supervisory data,instead of using the weakly supervised data training ta rget prediction function,the ternary loss function pre-training model is used to prevent the deep learning model from over-fitting the noise.And try to use a s a new deep learning model,the experiment proves that using the triple loss function in the pre-training stage can effectively improve the model effect.