融合WordNet的无监督语义分析研究  

Research on Unsupervised Semantic Parsing Joint with WordNet

作　　者：杨博[1] 蔡东风[1] 赵奇猛 杨华[2] 

机构地区：[1]沈阳航空航天大学知识工程研究中心,沈阳110136 [2]沈阳航空航天大学计算机学院,沈阳110136

出　　处：《小型微型计算机系统》2014年第2期368-373,共6页Journal of Chinese Computer Systems

基　　金：国家"十二五"科技支撑计划项目(2012BAH14F00)资助;国家自然科学基金项目(61073123)资助

摘　　要：应用机器学习方法处理机器阅读的相关任务是人工智能的长远目标,但通常需要大量的人工监督操作.研究一种无监督学习在机器阅读的一个主要任务-语义分析中的应用,这种无监督方法得益于统计关系学习统一框架-Markov逻辑网.鉴于该方法通过依存句法信息无法解析语义分析中普遍存在的反义词、词形变化等语言现象,该文融合WordNet进行改进,促进概念的抽取及合并,并将机器阅读的主要目标-问答作为评价手段,结果表明这种WordNet词典与无监督机器学习相结合的方法可更好地进行语义分析,并且问答正确率可提高至90.6%.Developing machine learning approaches for some relative tasks of machine reading is a long-standing goal of AI. Most of traditional approaches need a large amount of manual supervised work. This paper researches an unsupervised learning method for se- mantic parsing, which plays an important role in machine reading. This method benefits from a unifying framework of statistical rela- tional learning that called Markov Logic Networks. Since the method relied on dependency syntax information cannot resolve some linguistic phenomena such as antonym, synonymous variations etc. that existing commonly in semantic parsing, this paper improves the model by combining WordNet with it, which can help extract and merge concepts occurred in the corpus. We evaluate the improved method by the ultimate goal of machine reading-Q＆A, the results reveal that the combination of WordNet and unsupervised machine learning is better for semantic parsing, and the accuracy can advance to 90.6%.

关 键 词：MARKOV逻辑网 无监督学习 开放式信息抽取 WORDNET 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]