Recent years,the development of knowledge bases is very fast.They store large scale of entities and the relations between entities.However,most of the relations which have the same meanings are not in the same form.It is necessary to resolute the relations.For this purpose,this paper proposes an approach based on Markov cluster algorithm to cluster the relation with same meanings.Firstly,this paper calculates the semantic similarity between every two relations,and then it uses the relation similarity as weighted-edge to build a graph.Finally,this paper runs a Markov cluster algorithm on the graph and gets the result of relation clusters.Experiments show that the proposed approach has a higher purity than hierarchy cluster and k-means cluster.

关注非结构化文本中命名实体属性值的抽取问题.当前主流有监督属性值抽取方法仅使用局部特征,抽取效果有限,开展了利用文本全局特征改善属性值抽取的研究.通过适用于中文属性值抽取的全局特征,用局部特征以外的有价值信息提高抽取效果.据此,提出结合全局特征的感知机学习算法,该算法能够方便地融合文本全局特征,并将全局特征和局部特征统一结合到模型学习过程中,使模型具有更好的特征表示能力.实验结果表明,所提出方法的整体抽取效果高于仅使用局部特征的CRF模型和平均感知机模型.该方法适用于开放领域的属性值获取,具有较好的泛化能力.

在领域知识库的构建过程中,领域概念的识别是一项非常重要的步骤.当前基于统计方法仅按词频进行领域概念的识别,而一些较长的重要领域概念恰恰是低频词,因而对低频领域概念识别准确率不高.为了提高低频领域概念的识别准确率,本文提出了一种基于词向量的加权HITS算法.此方法首先将开放文本中的领域概念表示为词向量,然后使用加权HITS算法计算其领域相关度,最终筛选出领域相关度超过一定阈值的领域概念构建领域知识库.实验证明,本文提出的方法与现有方法相比,在领域概念识别的准确率和召回率方面有一定的改进.特别地,该方法将低频领域概念识别召回率提高了10％.

In order to exert potential commercial value and social value of subjectivity text in enterprise business intelli-gence and public opinion survey and so on,a novel feature selection algorithm based on sentiment topic model was pro-posed,which takes both opinion term and opinion co-occurrence term into consideration to help topic modeling,and then the conditional distributions of opinion term in positive topic and negative topic were effectively estimated.This method tries to measure the importance of opinion feature in sentiment orientation.SVM was used in the experimental stage for classification.The experiment result shows that the algorithm has a higher recognition ratio and offers practical capabilities for cross-domain.