基于有监督关联聚类的中文共指消解  

Coreference Resolution with Supervised Correlation Clustering

作　　者：刘未鹏[1] 周俊生[2] 黄书剑[1] 陈家骏[1] 

机构地区：[1]南京大学计算机软件新技术国家重点实验室,南京210093 [2]南京师范大学计算机科学系,南京210097

出　　处：《计算机科学》2009年第9期182-185,共4页Computer Science

基　　金：国家自然科学基金项目(60673043);国家社科基金(07BYY0);江苏省高校自然科学基金(07KJB520057)资助

摘　　要：共指消解是文本信息处理中的一个重要问题。提出了一种有监督的关联聚类算法以实现对中文实体提及的共指消解。首先将共指消解过程看成图的关联聚类问题,从全局的角度实现对共指等价类的划分,而不是孤立地对每一对名词短语分别进行共指决策;然后给出了关联聚类的推导算法;最后设计了一种基于梯度下降的特征参数学习算法,使得训练出的特征参数能够较好拟合关联聚类的目标。在ACE中文语料上的实验结果显示,该算法优于传统的"分类-聚类"共指消解学习算法。Coreference resolution plays an important role in natural language processing. A supervised correlation clustering algorithm for coreference resolution was proposed. Firstly,coreference resolution was treated as a graph correlation clustering problem,which partition the coreference relation from the global view, rather to make pairwise coreference decisions independently of each other. Then, the inference algorithms for correlation clustering were presented. Finally, a learning algorithm based on gradient descent was proposed to make the features parameters be trained from the training corpus, so that the learned parameters can better fit the objective of the correlation clustering. The experimental results on the ACE Chinese corpus demonstrate that the proposed method achieves better performance, compared with the traditional approaches.

关 键 词：共指消解 关联聚类 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] X832[自动化与计算机技术—计算机科学与技术]