基于序列增强的事件主体抽取方法

Event Subject Extraction Method Based on Sequence Enhancement

为了解决在事件抽取中使用固定文本长度造成短句子填充过多,从而引发语义偏移的问题,提出一种基于序列增强的事件主体抽取方法。具体而言,首先,将固定长度文本通过预训练模型映射到1个稠密向量中;然后,将文本对应的稠密向量与自定义Mask层和SpatialDropout层按位相乘,得到编码输出;最后,将该输出连接BiGRU层以及Mask层,得到解码输出,并将其映射在MLP层中,得到最后的结果。该模型既能够避免预训练模型对文本表征过拟合的问题,也能够限制填充文本在语义上的过度表达。使用CCKS 2022所供金融领域事件主体作为数据集进行不同模型读取对比实验,所得实验数据表明,对填充文本加以负影响的增强序列比传统序列在事件主体识别的正确性、F_(1)值上皆显著提升。

In view of a solution of semantic deviation brought about by overfilling short sentences with fixed text length in event extraction,a sequence enhancement based event subject extraction method has thus been proposed.Specifically,an initial mapping of the fixed-length text is given to a dense vector through a pre-trained model.Subsequently,the dense vector corresponding to the text is bitwise multiplied by the custom Mask layer and SpatialDropout layer,thus obtaining the encoded output.Finally,the output is connected with BiGRU and Mask layers to get the decoded output,which is then mapped to an MLP layer to obtain the final result.This model can not only avoid the problem of overfitting the text representation in the pre-trained model,but also limit the semantic overexpression of thefilled text.By using thefinancialfield event subjects provided by CCKS 2022 as a dataset for different model reading comparative experiments,the experimental data obtained shows that the enhanced sequence with negative impact on filled text significantly improves the accuracy and F_(1) value of event subject recognition compared to traditional sequences.