融合实体信息和时序特征的问答式事件检测方法

Question answering-based event detection method fusing entity information and temporal feature

针对现有问答方法在处理触发词歧义性问题上的不足,提出一种融合实体信息和时序特征的问答式事件检测方法EDQA-EITF。构建一种基于RoBERTa的问答框架,增强模型的语义表示能力;通过在模型输入序列中显示地添加实体、实体类型等先验信息,进一步帮助模型根据句子的上下文语境对触发词进行分类;采用最小门控循环单元(minimal gated unit,MGU)和Transformer编码器对输入序列中的时序依赖关系进行建模,提升模型对于句子的语义关系、句法结构的阅读与理解能力。公共数据集上的实验结果表明,所提方法在进行事件检测时具有更优的性能,有效缓解了触发词的歧义性问题。

Aiming at the shortcomings of existing question answering methods on the problem of trigger ambiguity,a question answering-based event detection method fusing entity information and temporal feature named EDQA-EITF was proposed.A question answering framework based on RoBERTa was constructed to enhance the model’s semantic representation ability.The priori information such as entities and entity types was added to the model input sequence in a displayed way,to further help the model classify the triggers based on the contextual semantic environment of the sentence.The minimal gated unit(MGU)and Transformer encoder were used to model the temporal dependencies in the input sequence,which improved the model’s ability to read and understand the semantic relation and syntactic structure of sentence.Experimental results on public dataset show that the proposed method has better performance in event detection and effectively alleviates the problem of trigger ambiguity.