A Short Text Augmentation Approach Based on Three-Order Semantic Graphs for Imbalanced Sentiment Multiclassification
Text augmentation techniques have been widely recognized for their ability to significantly enhance the performance of sentiment classification tasks,particularly when dealing with imbalanced datasets.However,when generating short text data for the minority class during text augmentation,it can result in overlapping emotions across different categories if the generated data fails to capture the complete semantic features of sentiment.To better understand and represent the emotional features of minority classes,this study introduces a third-order semantic graph data augmentation method specifically designed for imbalanced text sentiment multi-classification.The proposed method is based on the construction of a third-order semantic graph that models complex relationships between multiple words within short texts.The proposed method allows for the representation of a wide range of local sentiment semantics and is able to capture the dependencies between word nodes,offering a more nuanced understanding of emotional context in minority classes.By leveraging this intricate relational model,the third-order semantic graph enables a more comprehensive representation of sentiment,ensuring that the emotional characteristics of minority classes are more accurately reflected in the generated data.Once the third-order semantic graph is constructed,a data augmentation method based on this graph is applied to balance the distribution of sentiment categories in multi-class text datasets.This approach is designated to address the shortcomings of traditional text augmentation methods that often introduce noise and fail to adequately represent minority class sentiments by ensuring that the generated text data can capture the essential emotional features of the minority class,thus leading to improved classification performance across imbalanced datasets.Compared with traditional text augmentation methods,the proposed method in this paper can improve the minority evaluation indicators F1-measure and F2-measure by 5.75%and 9.65%,respectively,and the G-means value of balanced emotion recognition ability by 2.91%on the unbalanced Indonesian dataset.On the unbalanced Malay dataset,the proposed method also increases the minority evaluation indicators F1-measure and F3-measure by 2.45%and 4.81%,respectively,and the G-means value of balanced emotion recognition ability by 1.24%.In addition,compared with existing sentiment classification models based on short text augmentation oversampling models,traditional machine learning methods and deep network models,the proposed method achieves the highest Accuracy values on the publicly available imbalanced short text datasets in Indonesian,Malay,English,and Chinese.The experimental results also demonstrate that the proposed method provides a comprehensive and effective solution to the challenges posed by imbalanced sentiment classification.By integrating third-order semantic graph structures across different modalities,the proposed method effectively captures local emotional semantics and word dependencies.This not only improves the representation of minority class emotions but also significantly reduces the risk of introducing noise during data augmentation.In addition,traditional oversampling methods often introduce errors that can degrade classification performance,whereas the proposed method avoids these pitfalls by leveraging the detailed relational structure of the semantic graph.As a result,it can also achieve better multi-class classification performance on imbalanced short text sentiment classification tasks.
third-order semantic graphstext augmentationbalancing strategiesshort text sentiment classificationmotif
万方数据
维普
