A Few-shot Sentiment Classification Method Based On User-Tagged Text

Sentiment classification is a hot topic in natural language processing.The early research focuses on lexicon and traditional machine learning,and the performance of these two methods is limited by the design difficulties of sentiment lexicon and feature engineering respectively.Therefore,most of the current research adopts the deep learning method.However,the performance of such methods depends on large-scale and high-quality manual annotation data,and the data annotation project is time-consuming and laborious.Fortunately,network users produce a large number of user-tagged texts.These data resources can be used as weakly-labeled data to train sentiment classifiers.However,there are noisy data in weakly-labeled data that are inconsistent with the real sentiment semantics of the text,which will cause serious negative impact on the model training.Therefore,the weakly-labeled data cannot be directly used for the deep model training.Although user-tagged text can provide a new data source for deep learning,the difficulty of obtaining a large number of high-quality manual annotation data is still one of the important factors limiting the performance of the model.Therefore,few-shot learning method becomes particularly important.Most of the current research aims to capture more feature information from the original text to increase the accuracy of the model in few-shot scenarios.However,such methods ignore the implicit guidance information in the sentiment label.In order to solve the above problems,this thesis first proposes the Weakly-supervised Anti-noise based on Contrastive Learning(WACL)sentiment classification framework WACL,which aims to alleviate the negative impact of noisy data by designing anti-noise strategies,so as to make full use of large-scale weakly-labeled data and improve the performance of model.Secondly,this thesis design a few-shot sentiment classification framework Fusing Label Feature(FLF).This framework first captures the supervision information contained in labels by designing a label feature learning method,and then integrates additional label information into the training process of the model to further improve the classification performance in the few-shot sentiment classification scenario.To sum up,the contributions of this thesis are as follows:In view of the negative impact of noisy data on model training,this thesis proposes the Weakly-supervised Anti-noise based on Contrastive Learning(WACL)sentiment classification framework,which has three stages: first,the contrastive learning strategy is used for pre-training on massive weakly-labeled data.This strategy can guide the model to learn the clear distribution pattern between sentiment categories and mitigate the harm of noisy data.Secondly,a simple but effective anti-noise strategy Dropping-layer is designed to remove the part of the model that is greatly affected by noisy data.Finally,we add a classification layer at the top of the remaining models and use manual annotation data to fine-tune the model parameters.This thesis demonstrates the effectiveness of WACL on three datasets,which can significantly improve the performance of deep model even in scenes with large noise ratio.Aiming at the issue that most deep learning methods ignore label information in few-shot scenarios,this thesis proposes a few-shot sentiment classification framework Fusing Label Feature(FLF).FLF has two steps: first,a multi-task label feature learning scheme is designed based on contrastive learning and prompt learning on the weakly-labeled dataset to train the label feature generator,which can explicitly learn the guidance information contained in the label while resisting hazards of noisy data.Finally,when training the sentiment classifier,the label feature generator is used to integrate the label feature vectors of small-scale training samples into the training process,so as to improve the classification performance of the model in few-shot scenarios.In this thesis,several sets of experiments are executed on four datasets.The results show that the performance of FLF in few-shot scenarios is significantly superior to other similar methods and has significant advantages.