Set Expansion with understanding

This poster was part of Set Expansion project

Transcript

1. Set  Expansion  with  Understanding  
   Dharmesh  Kakadia,  Avinash  Ram,  Ravishankar  Shravanam  
      Search  and  Informa7on  Extrac7on  Lab
      (SIEL)  
    Interna4onal  Ins4tute  of  Informa4on  Technology,  Hyderabad    
   Set Expansion
   •  Set expansion aims to expand the given set into a larger
   set by discovering other entities belonging to the
   conceptual set.
   •  Example, given { dollar, rupee, dinar } we want to identify
   that this set is about currency and identify all the other
   currencies.
   Motivation
   •  At the core of many IR applications.
   •  Most application also require the identification of the
   concept also.
   •  Problems are known to be hard, when approached
   individually.
   Set Expansion
   •  Web Search
   -  Used Bing API.
   -  Queries are of the AND form along with the concept.
   •  WebList Extraction
   -  Uses the html formatting like list and select tags and
   extracts the potential entities for expansion.
   -  Also extracts other information like, context text, title,
   heading, url etc.
   •  Candidate Identification
   •  Expansion Validation
   -  Based on the overlap between the candidate and the
   current set.
   -  Use of selected information from web pages instead of full
   document similarity.
   Architecture
   Concept Identification
   •  Concept prediction
   -  Approach based on accumulated frequency of overlapped
   information.
   -  URL, page title, context and heading of the WebList as
   potential indicators for the concept.
   •  Concept Validation
   -  Ensuring that the identified concept is general and not tied
   to any particular entity in the set.
   -  Overlap between the entities in the set and concept is used
   as a distance and its variation indicates the quality of the
   set and the concept.
   The combination
   -  The intuition behind using the concept in set expansion is
   the easily identifiable hints of concept for noisy lists.
   Features
   •  Language Independent.
   •  Domain Independent.
   •  No training required.
   •  Scalable
   Applications
   •  Semantic Set Instance generation.
   •  User recommendations.
   •  Computational advertisement.
   Evaluation
   Future Work
   •  Other approaches for WebList Extraction.
   •  Other models for validation.
   •  Influence based framework for set quality.
   References
   •  R. Wang and W. Cohen. Iterative set expansion of named
   entity using the web. In ICDM, 2008.
   •  Y. He and D. Xin. SEISA: set expansion by iterative
   similarity aggregation. In WWW, pages 427–436, 2011.
   candidate = max Wi *
   1
   Rank(i)
   ∑
   Query  Concept  
   Seed  Set  
   Size  
   Expanded  
   Set  Size  
   Concept  
   Iden4fied  
   False  
   en44es  
   Countries   4   236   countries   1  
   Car  
   Manufacturers  
   4   79   cars   0  
   Football  players   3   68   players   4  
   

   View Slide