王洁,李旭晖.基于关键特征增强的金融长文本事件分类[J].情报工程,2024,10(3):104-113 |
基于关键特征增强的金融长文本事件分类 |
Key-Features Enhanced Financial Long Text Event Classification |
|
DOI:10.3772/j.issn.2095-915X.2024.03.008 |
中文关键词: 事件分类;长文本分类;关键特征;特征增强;自注意力机制 |
英文关键词: Event Classification; Long Text Classification; Key Features; Feature Enhancement; Self-Attention Mechanism |
基金项目:国家自然科学基金重点项目“基于知识关联的金融大数据价值分析、发现及协同创造机制”(91646206);国家社科基金重大项目“文化遗产智慧数据资源建设与服务研究”(21&ZD334)。 |
作者 | 单位 | 王洁 | 武汉大学信息管理学院 武汉 430072 | 李旭晖 | 1. 武汉大学信息管理学院 武汉 430072;2. 武汉大学大数据研究院 武汉 430072 |
|
摘要点击次数: 586 |
全文下载次数: 842 |
中文摘要: |
[目的/意义]为了解决长文本模型输入长度限制问题,通过抽取事件关键句和事件关键词,对长文本进行关键特征增强,以提高模型的特征表示能力。[方法/过程]基于关键特征增强的模型,在原文的基础上利用TextRank算法抽取事件关键句,并利用TF-IDF算法抽取事件关键词,将二者作为关键特征对长文本进行特征增强,再利用BERT和Self-Attention模型进行特征的进一步提取,最后进行事件分类。[局限]模型仅在金融领域事件分类上进行实验,可以考虑在其他领域内也进行实验并进一步验证模型效果。[结果/结论]在金融长新闻事件分类数据集上,提出的模型准确率达到88.40%,比基准模型提升了2个以上的百分点,表明了模型的有效性。 |
英文摘要: |
[Objective/Significance] In order to address the issue of length limitations in long-text models, this study enhances the feature representation capability of the model by extracting event-related key sentences and keywords from long text. [Methods/Processes] The key-features enhanced model utilizes the TextRank algorithm to extract key event sentences and the TF-IDF algorithm to extract event keywords from the original text. These key features are used to enhance the long text, and further feature extraction is performed using BERT and Self-Attention models, followed by event classification. [Limitations] The model in this study was only tested on event classification in the financial domain. It is recommended to conduct further experiments and verify the effectiveness of the model in other domains as well. [Results/Conclusions] On the financial long news event classification dataset, the proposed model achieved an accuracy rate of 88.40%, outperforming other benchmark models by more than 2 percent, which demonstrates the superiority of the model. |
查看全文
查看/发表评论 下载PDF阅读器 |
关闭 |