学术论文

      基于马尔可夫逻辑的中文零指代消解

      Chinese Zero Anaphora Resolution with Markov Logic

      摘要:
      中文零指代消解问题包括零指代项的识别和零指代项的消解2个相互关联的子任务.传统的方法在解决该问题时,往往不考虑2个子任务间的关联关系,比如识别出的零指代项必须被消解以及发生消解的必须是零指代项等约束.基于马尔可夫逻辑网络模型可以将零指代项的识别和零指代项的消解2个子任务融合在统一的机器学习框架下进行联合推断与联合学习,采用局部规则分别针对零指代项的识别和消解进行预测,采用全局规则描述这2个子任务间的关联关系.基于OntoNotes3.0的中文数据集上的实验结果显示,基于马尔可夫逻辑网络的联合学习模型相比于独立学习模型以及多个baseline方法能够获得更好的实验效果.
      Abstract:
      Chinese zero anaphora resolution includes tw o subtasks:zero pronoun detection and zero anaphora resolution ,which are correlated with each other .Zero pronoun detection means to recognize all the zero anaphors in a given text ,which mainly include null subject or null object ,and exist widely in Chinese ,Japanese and Italian .Zero anaphora resolution means to determine the antecedent for each recognized zero anaphor ,which has already appeared as a noun ,pronoun or common noun phrase before the detected zero anaphora in the previous text . Traditional methods to solve Chinese zero anaphora resolution problem generally employ some common‐used learning features to construct independent classifiers for zero pronoun detection and zero anaphora resolution ,but it cannot capture association relationship between these two subtasks ,e .g .recognized zero anaphora must be resolved or the one to be resolved must be zero anaphora and so on .In our method ,these two subtasks are combined into a unified machine learning framework with Markov logic to make joint inference and joint learning .We use local formulas to describe zero pronoun detection and zero anaphora resolution respectively ,and use global formulas to represent the association relationship between these two subtasks .We find that joint learning model which makes learning with inference can acquire more effective feature weights than independent learning model which just makes learning without inference .Experimental results on OntoNotes3 .0 Chinese dataset show that our joint learning model can achieve better results compared with independent learning model and other baseline methods .
      作者: 宋洋 [1] 王厚峰 [2]
      Author: Song Yang [1] Wang Houfeng [2]
      作者单位: 北京大学计算语言学研究所北京 100871 北京大学计算语言学研究所北京 100871; 计算语言学教育部重点实验室 北京大学 北京 100871
      刊 名: 计算机研究与发展 ISTICEIPKU
      年,卷(期): 2015, (9)
      分类号: TP301
      机标分类号: TU9 P20
      在线出版日期: 2015年10月8日
      基金项目: 国家“八六三”高技术研究发展计划基金项目(2015AA015402);国家自然科学基金项目(61370117,61333018);国家社会科学基金重大项目