1 Entity Ranking Using Wikipedia as a Pivot (CIKM 10’) Rianne Kaptein, Pavel Serdyukov, Arjen de Vries, Jaap Kamps 2010/12/14 Yu-wen,Hsu

2 Outline Introduction From Wikipedia Entities to Web Entities and back Entity Ranking on Wikipedia Entity Ranking on Web Conclusion

3 Introduction Entity ranking is the task of finding documents representing entities of a correct type that are relevant to a query.  presenting a ranked list of entities directly, rather than a list of web pages with relevant but also potentially redundant information about these entities.

4 Differs from document retrieval on at least three points:  i) returned documents have to represent an entity  ii) this entity should belong to a specified entity type  iii) to create a diverse result list an entity should only be returned once.

5 Main Goal To Rank Web entities  1. Associate target entity types with the query  2. Rank Wikipedia pages according to their similarity with the query and target entity types  3. Find web entities corresponding to the Wikipedia entities

6 Using Wikipedia as a pivot  entities: Wikipedia pages  the name of the entity: the title of the page  the content of the page: the representation of the entity  Each Wikipedia page is assigned to a number of categories: topical, type, and administrative categories.

7 From Wikipedia Entities to Web Entities and back From Web to Wikipedia  these repositories provide enough clues to find the corresponding entities on theWeb?  they contain enough entities that cover the complete range of entities needed to satisfy all kinds of information needs?

8 From Wikipedia to Web  Use External Link

9 Entity Ranking on Wikipedia * Entity Types Entity Type Assignment  exploit the existing Wikipedia categorization of documents  Pseudo-relevance feedback of the top retrieved documents we extract the categories that are most frequently assigned the top 10 results, and look at the 2 most frequently occurring categories belonging to these documents

10 *Entity Types -Scoring Entities estimate background probabilities smooth the probabilities of a term occurring in a category name with the background collection : the name of the category : the category : the query terms : the document : the entire Wikipedia document collection

11 Similarity between two categories The entity type score for a document in relation to a query topic Score Normalization

12 Entity Ranking on Wikipedia *Experimental Setup Data Set:  INEX: specific, ex countries, national parks..  TREC: people, organization, product Advantage: clear, few options, could be easily selected Disadvantage: cover a small part of all possible entity ranking queries  manually assigned more specific entity types

13 rerank the top 2,500 results of the baseline  Manually assigned (author)  Automatically assigned (PRF) evaluation  2009 TREC:P10 and  INEX:P10 and MAP INEX consisting of 79 topics INEX 2009 topics consisting of a selection of 55 topics from the topics. only count the so-called ‘primary’ pages

14

15 Entity Ranking on The Web We have three approaches for finding web pages associated with Wikipedia pages.  1. External links: the External links section of the Wikipedia page  2. Anchor text: Wikipedia page title as query retrieve pages from the anchor text index  3. Combined: not all Wikipedia pages have external links not all external links of Wikipedia pages are part of the Clueweb collection less than 3 webpages are found, we fill up the results to 3 pages using the top pages retrieved using anchor text

16

17 Conclusion Our experiments show that our wikipedia- as-a-pivot approach outperforms a baselines of full-text search. Both external links on Wikipedia pages, and searching an anchor text index of the web are effective approaches to find homepages for entities represented by Wikipedia pages.