| 梁冬* **,张程*,史骁*,谭文婷* **,吕存驰* **,赵晓芳* ** ***.基于对比学习增强句子语义的事件检测方法[J].高技术通讯(中文),2023,33(7):669~682 |
| 基于对比学习增强句子语义的事件检测方法 |
| Event detection based on enhanced sentence semantics via contrastive learning |
| |
| DOI：10. 3772/ j. issn. 1002-0470. 2023. 07. 001 |
| 中文关键词: 事件检测； 自监督对比学习； 监督对比学习； 语义增强； 自动调整权重 |
| 英文关键词: event detection, self-supervised contrastive learning, supervised contrastive learning, semantic enhancement, automatic weighting |
| 基金项目: |
| 作者 | 单位 | | 梁冬* ** | （*中国科学院计算技术研究所北京 100190）
（**中国科学院大学北京 100049）
（***中科苏州智能计算技术研究院苏州 215028） | | 张程* | | | 史骁* | | | 谭文婷* ** | | | 吕存驰* ** | | | 赵晓芳* ** *** | |
|
| 摘要点击次数: 905 |
| 全文下载次数: 727 |
| 中文摘要: |
| 事件检测旨在识别文本中提到的事件及其类型。基于触发词的事件检测方法需要额外的人工成本标注事件触发词。本文从无触发词的文本语义提取出发，提出了一种基于对比学习增强句子语义的事件检测方法。该方法首先在事件检测数据集上通过自监督学习对预训练的语言模型基于转换器的双向编码表示器(BERT)调优，提高语言模型的领域适应性。然后利用掩码（mask）操作和丢弃（dropout）操作构建自监督对比样例，增加监督对比样例，实现了自监督对比和监督对比2种句子语义增强的方法。此外在训练过程中自动调整对比损失和事件分类的交叉熵损失的权重，以降低人工调参的成本，同时提高模型收敛速度。在自动内容抽取（ACE）2005中英文语料上的实验结果表明，本文方法比先前无触发词事件检测方法取得更好的结果，与利用预训练BERT模型微调的事件检测方法相比也具有优势。 |
| 英文摘要: |
| Event detection aims to find the events and their types mentioned in the text. Previous work based on triggers requires additional labor to label event triggers. Considering that event detection mainly relies on the text semantics without triggers, an event detection method based on enhanced sentence semantics via contrastive learning is proposed. First, the pre-trained language model biderectional encoder representations from transformers (BERT) is fine-tuned with self-supervised learning for better domain adaptability. Then, a novel event detection method based on sentence semantic enhancement through contrastive learning is presented, using mask operation or dropout operation to construct the self-supervised contrastive samples and increase the supervised contrastive samples. Furthermore, the weight assignment between the cross-entropy loss and the contrastive loss is automatically adjusted during training to reduce the cost of manual parameter tuning and improve the convergence speed. Experimental results on automatic context extraction (ACE) 2005 Chinese and English corpus show that the proposed method is superior to the previous event detection methods without triggers, and outperforms the methods with triggers extracted by the pre-trained BERT model fine-tuning. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
