基于软参数共享的事件联合抽取方法

事件抽取是项重要的信息抽取任务，旨在抽取文本中的事件信息。目前基于多任务学习的事件联合抽取方法大多基于硬参数共享，此类方法往往会导致跷跷板现象的出现，即一项任务的性能往往通过损害另一项任务的性能来提高。为了解决这一问题，提出了一种基于软参数共享的事件联合抽取方法，该方法明确地分离了共享参数和任务特定参数，并通过双层门控网络增强模型提取和筛选语义知识的能力，使模型能同时为两个任务学习到合适的特征表示，实现了更高效的信息共享和联合表示学习。在DuEE1.0公共数据集上进行了实验，使用准确率、召回率、F1值作为评价指标，并通过对比实验和消融实验验证了方法的有效性。对比基于硬参数共享的联合抽取模型事件识别任务F1值提高了2.0%，论元角色分类任务F1值提高了0.9%，有效地缓解了跷跷板现象的出现，验证了方法的有效性。

Joint event extraction method based on soft parameter sharing

Event extraction is an important information extraction task, which aims to extract event information from text. Most of the current event joint extraction methods based on multi-task learning are based on hard parameter sharing, which often leads to the seesaw phenomenon, in which the performance of one task tends to improve at the expense of the performance of another. In order to solve this problem, this paper proposed a method based on soft parameter sharing, this method clearly separated shared parameters and task-specific parameters, and enhanced the ability of model extraction and screening semantic knowledge through a double-layer gated network, so that the model could learn the appropriate feature representation for both tasks at the same time, and realized more efficient information sharing and jointed representation learning. This paper conducted experiments on the DuEE 1.0 public dataset, using accuracy, recall, and F1 values as evaluation indicators, and through the contrast experiment and the ablation experiments verify the effectiveness of the method. The F1 value of event recognition task is improved by 2.0%, and the F1 value of argument role classification task is improved by 0.9% compared with the joint extraction model based on hard parameter sharing, which effectively alleviated the emergence of seesaw phenomenon and verified the effectiveness of the method.