基于角色信息引导的多轮事件论元抽取

针对通用领域的事件论元抽取研究中角色信息利用不足和论元间缺少交互两个问题,提出角色信息引导的多轮事件论元抽取模型,用于增强文本的语义信息和论元之间的交互能力,提升事件论元抽取的性能。首先,为了更好地利用角色知识来引导论元的抽取,该模型根据角色定义构造角色知识,对角色信息和文本独立编码,并采用基于注意力机制的方法获取标签知识增强的文本表示,进而采用增强嵌入来预测各角色论元的起始和结束位置。同时,为了在抽取过程中充分利用事件论元之间的交互,受多轮对话模型的启发,设计一种多轮事件论元抽取算法。该算法参照“先易后难”的自然逻辑,每次选择预测概率最大,也是最容易确定的角色进行抽取。在论元抽取过程中,为了对论元之间的交互进行建模,模型引入历史嵌入,并在每一次预测结束后更新历史嵌入,帮助下一轮事件论元的抽取。实验结果表明,角色信息的引导和多轮抽取算法均有效地提升了论元抽取的性能,使得该模型的表现优于其他基线模型。

Multi-turn Event Argument Extraction Based onRole Information Guidance

Aiming at the two problems of insufficient utilization of role information and lack of interaction between arguments in general domain event argument extraction research, a role information-oriented multi-turn event argument extraction model is proposed to enhance the semantic information of texts and interactions between arguments. The interactive capability can improve the performance of event argument extraction. First, to better utilize role knowledge to guide argument extraction, the model builds role knowledge based on role definitions, independently encodes role information and text, and uses a method based on attention mechanism to obtain label-knowledge-enhanced representations. Then the augmented embeddings are used to predict whether or not each token is a start or end position for some category. At the same time, in order to make full use of the interaction between event arguments in the extraction process, inspired by the multi-turn dialogue model, this paper designs a multi-turn event argument extraction algorithm. The algorithm refers to the natural logic of “easiness to hardness”, and selects the character with the highest prediction probability, that is, the most predictable character, for extraction each time. In the process of argument extraction, in order to model the interaction between arguments, the model introduces historical embedding, and updates the historical embedding after each prediction to help the extraction of the next round of event arguments. The experimental results show that the guidance of role information and multi round extraction algorithm effectively improve the performance of argument extraction, and the method achieves state-of-the-art performance.