Scholarlink

<p>In the context of aspect category detection tasks, a common challenge arises in the form of class imbalance. This situation manifests itself within a review dataset in which certain categories boast an abundance of comments, while others are represented sparsely. Previous research centered predominantly on improving category detection through an examination of the text feature space. However, researchers have grappled with the precise discernment of the semantic quality emanating from textual data, thereby intensifying the complexity of the detection of aspect categories in imbalanced environments. This study aims to address the challenge of class imbalance in aspect category detection by proposing a hybrid augmentation approach that combines XLNet-based text generation with noise perturbation. In detail, our study introduces a dataset-level solution that amalgamates XLNet text generation with noise perturbation techniques to construct a harmonized category dataset. Subsequently, this dataset undergoes fine-tuning using the BERT pre-trained model. More specifically, XLNet is employed to augment categories that possess a limited number of samples within the dataset. We feed samples from these categories into XLNet to enable the model to acquire their language patterns. Noise perturbation, which involves the introduction of minor random variations (i.e., noise) into the data, is leveraged to generate fresh samples. Additionally, we harness the BERT model for the extraction of text features and subsequently train the aspect category detection model based on these extracted features. Experimental findings, as demonstrated on the SemEval-2014 and 2016 datasets, support the effectiveness of our proposed model in enhancing the performance of the detection of aspects categories.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Harmonizing imbalanced review datasets for aspect category detection through XLNet augmentation and BERT fine-tuning

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Chuanjun Zhao,
  • Lu Kang,
  • Zhihe Yan,
  • Xuzhuang Sun

摘要

In the context of aspect category detection tasks, a common challenge arises in the form of class imbalance. This situation manifests itself within a review dataset in which certain categories boast an abundance of comments, while others are represented sparsely. Previous research centered predominantly on improving category detection through an examination of the text feature space. However, researchers have grappled with the precise discernment of the semantic quality emanating from textual data, thereby intensifying the complexity of the detection of aspect categories in imbalanced environments. This study aims to address the challenge of class imbalance in aspect category detection by proposing a hybrid augmentation approach that combines XLNet-based text generation with noise perturbation. In detail, our study introduces a dataset-level solution that amalgamates XLNet text generation with noise perturbation techniques to construct a harmonized category dataset. Subsequently, this dataset undergoes fine-tuning using the BERT pre-trained model. More specifically, XLNet is employed to augment categories that possess a limited number of samples within the dataset. We feed samples from these categories into XLNet to enable the model to acquire their language patterns. Noise perturbation, which involves the introduction of minor random variations (i.e., noise) into the data, is leveraged to generate fresh samples. Additionally, we harness the BERT model for the extraction of text features and subsequently train the aspect category detection model based on these extracted features. Experimental findings, as demonstrated on the SemEval-2014 and 2016 datasets, support the effectiveness of our proposed model in enhancing the performance of the detection of aspects categories.