Collection Fusion in Carrot2

1
Collection Fusion in Carrot2

• Mithun Sheshagiri

2
Acknowledgements

• Prof. Scott Cost
• Srikanth Kallurkar
• Hemali Majithia

3
Overview

• Collection Fusion Problem in IR
• Possible solutions
• Equal Distribution Assumption
• Comparable similarities
• Modeling Relevant Document Distribution
• Query Clustering
• Carrot2 System
• Query Routing in Carrot2

4
Overview

• Collection Fusion in Carrot2
• Future Work
• Conclusions
• References

5
The Collection Fusion Problem

• Centralized Indexing and Retrieval.
• Distributed IR Systems
• The Collection Fusion Problem
• Determining the number of documents that need to
  be retrieved from each sub-collection
• Interleaving the documents returned by each
  sub-collection

6
Possible Solutions

• Equal Distribution Assumption
• Assumes that relevant documents are distributed
  equally across all sub-collections
• Comparable similarities
• Documents in the final result are listed as
  though the similarities are normalized across
  sub-collections.
• Similarity values are dependant on
  sub-collections
• A rare but not so relevant document can have
  higher ranking

7
Possible Solutions

• Equal Distribution Assumption
• Assumes that relevant documents are distributed
  equally across all sub-collections
• Comparable similarities
• Documents in the final result are listed as
  though the similarities are normalized across
  sub-collections.
• Similarity values are dependant on
  sub-collections
• A rare but not so relevant document can have
  higher ranking

8
Possible Solutions

• Modeling Relevant Document Distribution
• The document distribution model is built using
  training queries.
• The document distribution for a query q is
  obtained by averaging the number of relevant
  documents retrieved by the k nearest queries.
• This is done for all sub-collections.
• These document distributions along with the total
  number of documents to be retrieved is passed to
  a maximization procedure.

9
Possible Solutions

• Modeling Relevant Document Distribution
• This maximization procedure calculates a cut-off
  value for each sub-collection.

10
Possible Solutions

• Query Clustering
• Query clusters are formed by grouping training
  queries which return some identical documents.
• A weight is assigned to each cluster.
• Weight is computed based on the number of
  relevant documents returned by the queries
  belonging to the cluster.
• The centroid of the query cluster is calculated
  by averaging the query vectors belonging to that
  query cluster.

11
Possible Solutions

• Query Clustering
• The cluster whose centroid is most similar to the
  user query is selected and its weight is
  returned.
• The set of weights returned by all the
  sub-collections are used to apportion the
  retrieved set.
• wi (N)
• ?wi
• wi Weight returned by the cluster
• N Number of documents in the final result

12
Carrot2 System

• Carrot2 is a agent based distributed IR system.
• Uses Jackal Communication Infrastructure
• KQML is used by agents for communication
• Agents interface with IR engine through a
  wrapper
• Wrapper provides functionality to index documents
  as well as metadata

13
Carrot2 System

• Metadata is a reduced representation of the
  sub-collection. (8-10)
• Metadata is a vector consisting of N-grams
  (terms) and the number of documents that contain
  it.
• On start-up an agent is allotted a
  sub-collection.
• Every agent has an associated metadata object.
  
• An agent also has access to a metadata pool.

14
Query Routing in Carrot2

• Query is submitted to a Query Manager.
• Query manager picks an agent from a list of
  agents returned by the Collection Manager.
• Every agent queries its metadata pool and makes a
  decision.
• Query its local collection.
• Forward the query.
• Combination of both.

15
Query Routing in Carrot2

• The process ends when
• There are no more agents that have not already
  received the query.
• The number of times the query has been forwarded
  has reached a threshold value.

16
Collection fusion in Carrot2

• An approach similar to query clustering.
• Query cluster Metadata object
• Representations of sub-collections
• Both have a weight/similarity which is an
  indication of the relevance of the documents in
  the sub-collection to the given query.
• The similarity values of the metadata objects can
  be used to apportion the total number of
  documents that need to be returned.

17
Collection Fusion in Carrot2

• Requirement for implementation
• Access to the metadata object of all
  participating sub-collections (C2 agents).
• Using the metadata pool of one agent when the
  metadata objects are distributed in broadcast
  mode. (Flooding strategy)
• A new agent which accesses the metadata objects
  of all participating agents.

18
Collection Fusion in Carrot2

• Similarity value is appended to the result
  returned by each agent.
• The interleaving can be done by rolling a C-faced
  die which is biased by the number of documents
  that are still to be picked from the original
  result set.

19
Future Work

• The suitability of the proposed technique to the
  C2 system should be experimentally verified.
• This technique makes use of existing entities and
  information, implementation can be done with
  minimal changes to the existing architecture.

20
Conclusion

• Combination of query clustering like approach
  along with probabilistic interleaving is a good
  candidate for collection fusion in C2
• Decentralized nature
• Use of existing entities
• Easy to implement
• Less prone to scalability issues.

21
References

• Ellen M. Voorhees, Narendra Gupta, and Ben
  JohnsonLaird. Learning collection fusion
  strategies.
• James P Callan, Zhihong Lu and Bruce Croft
  Searching Distributed Collections With Inference
  Networks.
• E. M. Voorhees, N. K. Gupta, and B.
  JohnsonLaird. The collection fusion problem.