| 邵德奇,关培培,石聪.基于 BERT+A-Softmax 的多分类模型构建与应用研究[J].情报工程,2022,8(2):051-061 |
| 基于 BERT+A-Softmax 的多分类模型构建与应用研究 |
| Research on Construction and Application of Multi-classification Model Based on BERT+A-Softmax |
| |
| DOI：10.3772/j.issn.2095-915X.2022.02.004 |
| 中文关键词: 科技新闻资讯；人工智能；自然语言处理；分类体系；BERT |
| 英文关键词: Sci-tech news and information; AI; NLP; classification system; BERT |
| 基金项目:国家重点研发计划项目“基于可信与共治的全媒体内容社会众创服务平台研发与运营示范”（2020yfb1406900）。 |
| 作者 | 单位 | | 邵德奇 | 科技日报社技术研发部 北京 100038 | | 关培培 | 科技日报社技术研发部 北京 100038 | | 石聪 | 科技日报社技术研发部 北京 100038 |
|
| 摘要点击次数: 1342 |
| 全文下载次数: 1202 |
| 中文摘要: |
| [ 目的 / 意义 ] 在信息分类领域中，通过传统的机器学习与深度学习的方法可以对大多数稿件进行分类并取得整体较高的准确率。但是这种方法没有对稿件文体进行区别，而现实生产环境中存在新闻类稿件多，通知报告类少等样本不均衡的情况，如果对文体不加以区分，会产生少样本文体类别准确率低的情况。[ 方法 / 过程 ] 本文提出一种可以区别文体的深度学习分类模型方法，该方法先根据稿件文体对稿件进行分类，再根据分类结果分别调用分类模型进行进一步分类，解决样本不均衡、小样本文体类别准确率低等问题。[ 结果 / 结论 ] 在公开的数据集上实验结果表示，相对于传统的分类模型，本文提出的多分类模型方法在性能上有了显著提高。 |
| 英文摘要: |
| [Objective/Significance] In the field of information classification, the traditional classification methods of machine
learning and deep learning can classify most manuscripts and achieve high overall accuracy. However, when using this method, there is no distinction between article’s style. In production environment, there are many news manuscripts and few notification reports. If it is mixed trained, it will cause low accuracy of few samples class. [Methods/Process] This paper proposes a deep learning classification model that can distinguish article’s style. This method first classifies the style of manuscripts according to the style characteristics, then calls the classification model for further classificationl, so as to solve the problems of unbalanced samples and low accuracy of few samples class. [ResultsConclusions] The experimental results on the public data set show that the performance of the multi classification model method proposed in this paper has been significantly improved compared with the traditional classification model. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
