使用全局优化方法识别中文事件因果关系

Global optimization to recognize causal relations between events

分类器模型是目前识别因果关系的主要模型,该方法存在的问题是只考虑2个事件之间的关系,没有考虑同一文档中其他关联事件所包含的信息,识别结果往往存在逻辑矛盾。该文提出了一个中文事件因果关系识别的全局优化方法,该方法采用整数线性规划(integer linear programming,ILP)的推理方法,对基本逻辑关系、因果标志词、事件类型、论元信息进行有效约束,以文档为单位来优化因果关系识别。在该文标注语料上的实验结果表明:与分类器方法相比,该文提出的全局优化方法的F1值提升了5.54%。

Classifier-based models are widely used to identify causal relations between events. However, these models only consider the relationship between two specified events while ignoring related events. Thus, the results may have many logical contradictions. This paper presents a global optimization approach to recognize causal relations between events using an inference method based on integer linear programming （ILP）. This approach introduces various kinds of constraints, i. e. , a basic logical relationship, causal signal words, event types and argument information constraints to improve the performance. Tests on an annotated corpus show that this global optimization approach improves the F1 score by 5.54% compared with a classifier based model.