| 摘要: |
| 事件检测旨在识别非结构化文本中的事件触发词, 并将其分类为预定义的事件类别, 可用于知识图谱构建及舆情监控等. 然而, 其中的数据稀疏和不平衡问题严重影响了事件检测系统的性能和可用性. 现有大多数方法没有很好地解决这一问题, 这源于其将不同类别的事件独立看待, 并通过分类器或空间距离对触发词进行识别和分类. 尽管有研究考虑事件大类下子类的事件元素存在关联性, 采用多任务学习进行互增强, 但忽略了不同类别事件触发词之间的共享属性. 已有相关建模事件类别关系的工作需要大量的规则设计和数据标注, 导致作用域局限, 泛化性不强. 因此, 提出一种基于元属性的事件检测方法. 其旨在学习不同类别样本中包含的共享内在信息, 包括: (1) 构造触发词的特殊符号表示并通过表示向量的映射来提取触发词的类别无关语义; (2) 拼接触发词表示, 类别的样本语义表示和类别的标签语义表示, 输入一个可训练的相似度度量层, 从而建模关于触发词和事件类别的公用相似度度量. 通过学习以上两种信息以缓解数据稀疏和不平衡的影响. 此外, 将样本的类别无关语义集成到分类方法中, 并构建完整的融合模型. 在ACE2005和MAVEN数据集上通过不同程度稀疏和不平衡情景下的实验证明所提出方法的有效性, 并建立传统和少样本设置之间的联系. |
| 关键词: 事件检测 类别无关语义 度量学习 少样本学习 |
| DOI：10.13328/j.cnki.jos.007147 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62376192, 61976154) |
|
| Event Detection Based on Meta-attribute Learning |
|
HE Rui-Fang1,2, MA Jin-Song1,2, HUANG Xiao-Jia1,2, ZHANG Shi-Qi3, BAI Jie3
|
|
1.College of Intelligence and Computing, Tianjin University, Tianjin 300350, China;2.Tianjin Key Laboratory of Cognitive Computing and Applications, Tianjin 300350, China;3.The 54th Research Institute of CETC, Shijiazhuang 050081, China
|
| Abstract: |
| Event detection (ED) aims to detect event triggers in unstructured text and classify them into pre-defined event types, which can be applied to knowledge graph construction, public opinion monitoring, and so on. However, the data sparsity and imbalance severely impair the system’s performance and usability. Most existing methods cannot well address these issues. This is due to that during detection, they regard events of different types as independent and identify or classify them through classifiers or space-distance similarity. Some work considers the correlation between event elements under a broader category and employs multi-task learning for mutual enhancement; they overlook the shared properties of triggers with different event types. Research related to modeling event connections requires designing lots of rules and data annotation, which leads to limited applicability and weak generalizability. Therefore, this study proposes an event-detection method based on meta-attributes. It aims to learn the shared intrinsic information contained in samples across different event types, including (1) extracting type-agnostic semantics of triggers through semantic mapping from the representations of special symbols; (2) concatenating the semantic representations of triggers and samples in each event type as well as the label embedding, inputting them into a trainable similarity measurement layer, thereby modeling a public similarity metric related to triggers and event categories. By combining these representations into a measuring layer, the proposed method mitigates the effects of data sparsity and imbalance. Additionally, the full fusion model is constructed by integrating the type-agnostic semantic into the classification method. Experiments on ACE2005 and MAVEN datasets under various degrees of sparsity and imbalance, verify the effectiveness of the proposed method and build the connection between conventional and few-shot settings. |
| Key words: event detection type-agnostic semantics metric learning few-shot learning |
