Learning to generate entity summarization for entity view

Learning to generate entity summarization for entity view 龚赛赛 saisaigong@gmail. com 2013 -12 -09

Contents � Introduction � Related Work � Overview of Approach � Learning to generate entity summarization � Initial Evaluation � Conclusion

Introduction � Entity ◦ ◦ summarization is important Entity browsing Entity Searching Google knowledge graph … � We focus on getting a entity summarization for viewing entities ◦ Based on user feedback

Related Work � Mainly considering characteristics of properties ◦ Google. Query logs ◦ Sindice. Property popularity and value diversity � Leveraging ◦ ◦ value space of properties Cheng (2011). Informativeness and relatedness Thalhammer(2012). nearest neighbors Zhang (2012). Query based facet ranking Facet ranking � Few works considering relationship between entity types and properties

Overview of Approach � Given ◦ a set of classes as entity types, ◦ entity property set from entity descriptions, ◦ and a max summary size, Then select a size-limited subset of properties with related values to form a summary. � User feedback � ◦ +1 : property belongs to summary ◦ -1 : not belong �May be little � Modeled as a classification problem ◦ Using classification confidence to cut out a subset

Learning to generate entity summarization � Class dependent features ( maxc∈types. F(c, p) ) ◦ Co-occurrence of labels/localname of property and class using bing search num � 2*Search. Num(c, p)/(Search. Num(c)+Search. Num (p)) ◦ Conditional probability prob(p|c) �Doc. Num(p, c)/Doc. Num(c) ◦ Normalized pointwise mutual information �npmi(x, y)=-log(p(x, y)/p(x)p(y))/log(p(x, y)) �p(x, y)=Doc. Num(x, y)/total. Doc. Num ◦ Co-occurrence of labels/localname of property’s ranges/domains and class using bing search num

Learning to generate entity summarization � Class independent features ◦ Word num of URI label / localname ◦ The num of rdf documents describing the property ◦ Bing Search result num of property’s localname/ label ◦ Bing Search result num of the localname/label of property’s ranges/domains ◦ Number of property values ◦ Is datatype property or object property ◦ Is typed literal or not ◦ Length of property’s localname/label ◦ Max length of property value’s label/lexical. Form

Learning to generate entity summarization � Using non-linear SVM as model ◦ Gaussian kernel � Confidence: margins of points, i. e. distance to the hyperplane

Initial Evaluation � Construct training example from 30 entities of various classes ◦ 618 training examples as a total, with label +1, -1, and relevant feature values

Initial Evaluation � Some � summarization results http: //yago-knowledge. org/resource/Ronaldo (13) ◦ ◦ ◦ ◦ http: //yago-knowledge. org/resource/has. Height http: //yago-knowledge. org/resource/has. Gender http: //yago-knowledge. org/resource/plays. For http: //yago-knowledge. org/resource/was. Born. In http: //yago-knowledge. org/resource/has. Won. Prize http: //yago-knowledge. org/resource/was. Born. On. Date http: //yago-knowledge. org/resource/has. Wikipedia. Url ◦ ◦ ◦ http: //yago-knowledge. org/resource/extraction. Source http: //yago-knowledge. org/resource/links. To http: //yago-knowledge. org/resource/is. Affiliated. To http: //yago-knowledge. org/resource/has. Wikipedia. Article. Length http: //yago-knowledge. org/resource/is. Preferred. Meaning. Of

Initial Evaluation � http: //dbpedia. org/resource/Ronaldo ◦ ◦ ◦ (54) http: //dbpedia. org/property/youthyears http: //dbpedia. org/ontology/Person/height http: //dbpedia. org/property/youthclubs http: //dbpedia. org/property/fullname http: //dbpedia. org/property/nationalyears http: //dbpedia. org/property/fg http: //dbpedia. org/ontology/position http: //dbpedia. org/property/position http: //dbpedia. org/ontology/height http: //dbpedia. org/ontology/team

Conclusion � Contributions � A light-weight user feedback based learning algorithm for entity summarization �Consider different ordering of properties for different classes �Exanimate whether the relationship between classes and properties

Reference Cheng, G. , Tran, T. , Qu, Y. : RELIN: Relatedness and Informativeness-based Centrality for Entity Summarization. 2011 � Zhang, L. , Zhang, Y. , Chen, Y. : Summarizing Highly Structured Documents for Effective Search Interaction 2012 � Thalhammer, A. , Toma, I. , Roa-Valverde, A. J. , Fensel, D. : Leveraging Usage Data for Linked Data Movie Entity Summarization. 2012 �

Thank you for you attention !