Research On Aspect-level Sentiment Classification Based On Text Data Enhancement And Generation

Aspect based sentiment analysis(ABSA)is one of the most important research directions in the field of natural language processing(NLP),which can explore the sentiment tendency of a particular entity or attribute in text data,and thus more accurately reflect This approach can be used to explore the sentiment of a particular entity or attribute in text data,and thus more accurately reflect complex emotional features.As aspect-level sentiment analysis has a wide range of applications,some domains contain small data sizes and lack a complete and high-quality corpus,resulting in problems such as small samples and low resources in the available datasets.In addition,some domain aspect-level sentiment analysis datasets do not have strong contextual links between aspect words and context,which may lead to a lack of full use of global information.To address these problems,this study improves the accuracy of aspect-level sentiment classification models by constructing a multi-strategy text data enhancement scheme and a contextual text generation model.The main work of this paper focuses on three aspects:(1)A multi-strategy text data enhancement scheme is proposed,which firstly expands the data based on the word level by performing operations such as synonym replacement,word position exchange and partial word deletion for each document of the original corpus to obtain the word-enhanced data.Secondly,the original data is then expanded at the sentence level,including sentence regrouping,text cropping,grammar tree manipulation and back-translation,to form the sentence-level augmented data.The multi-strategy enhancement method combining the two methods was validated on the aspect-level sentiment classification domain dataset and compared with word-level corpus and sentence-level corpus on several classifier models.The final experimental results show that the multi-strategy text enhancement technique can effectively expand the amount of data and improve the classification effectiveness of the model.(2)A contextual text generation model is constructed based on the Bert(Bidirectional encoder representations from transformers)pre-trained language model and text filtering algorithm.The model first utilizes the Bert model to make full use of contextual information,while paying more attention to the internal relationships between sentences and effectively combining the relationships between sentences and tags to obtain a preliminary corpus of data enhancement.Secondly,a filtering algorithm is designed to filter the initial enhanced corpus,filtering out the data of low quality to form the final text-enhanced data.Finally,the final dataset was tested experimentally,and the results showed that the enhanced dataset improved the text quality and effectively enhanced the aspect-level sentiment classification effect of the model.(3)An aspect-level sentiment analysis system based on text data augmentation is designed and implemented.The system consists of four modules: data pre-processing,text data enhancement,sentiment polarity discrimination and aspect-level sentiment classification.Firstly,the system can perform a series of text data enhancement operations on the original data set to obtain the enhanced text,and then integrate the text enhancement tool and aspect-level sentiment classification to predict the effect of aspect-level sentiment classification after data enhancement.Finally,the system’s usefulness is demonstrated by applying it to text data from the online travel review domain to discriminate the sentiment polarity of the text.