Keyword Search

1
Keyword Search

• Su Zhan
• suzhan_at_fudan.edu.cn

2
Introduction

• What is keyword search
• Present works
• Effective keyword search in relational databases.
  
• BLINKS ranked keyword searches on graphs.

3
Effective keyword search in relational databases.

• Fang Liu
• Clement T. Yu
• Weiyi Meng
• Abdur Chowdhury

4
Outline

• Introduction
• Answer generation
• Background in IR ranking
• Novel ranking strategy for relational databases
• Experiment results
• Conclusion

5
Introduction
6
Introduction

• Suppose a user is looking for albums titled off
  the wall and he/she cannot remember the exact
  title.
• Select from Album B
• Where Contains(B.title, off wall, 1) gt 0
• Order by score(1) desc
• Or
• off wall

7
Introduction

• Query 1 off wall
• Query 2 lyrics how come by D12
• Query 3 album by D12 and Eminem
• Tuple Tree 1 b2
• Tuple Tree 2
• a1? ab1? b1? bs1? s1
• Tuple Tree 3
• a1? ab1? b1? ab2 ? a2

8
Introduction

• 3 key steps for processing a given keyword query
• Generate all candidate answers, each of which is
  a tuple tree by joining tuples from multiple
  tables.
• Then compute a single score for each answer. The
  scores should be defined in such a way so that
  the most relevant answers are ranked as high as
  possible.
• And finally return answers with semantics.
• This paper focuses on search effectiveness, that
  is, step (2).

9
Answer Generation

• Schema Graph
• Tuple Tree
• Keyword Query
• Answer
• Query Tuple Set RQ
• Free Tuple Set RF
• Answer Graph

10
(No Transcript)
11
Background in IR ranking

• Ranking Model in IR
• 11-point precision and recall
• Mean average precision
• Reciprocal rank

12
Background in IR ranking

• Ranking Model in IR

13
Novel ranking strategy for relational Database

• Let T be a tuple tree and D1, D2, , Dm be all
  text column values in T. We define each text
  column value Di as a document and T as a
  super-document. Then we can compute a similarity
  value between the query Q and the super-document
  T as shown in Formula 3 to rank tuple trees.
• Our focus is on weight(k,T)

14
Novel ranking strategy for relational Database

• Four Normalizations
• Tuple Tree Size Normalization
• Document Length Normalization
• Document Frequency Normalization
• Inter-Document Weight Normalization

15
Novel ranking strategy for relational Database

• (jojo leave lyrics)
• b3 and s3 score higher than (b3, bs3, s3)!
• Tuple Tree Size Normalization

16
Novel ranking strategy for relational Database

• (how come)
• Title and Lyrics score the same!
• Global average?
• Document Length Normalization

17
Novel ranking strategy for relational Database

• idf
• Document Frequency Normalization

18
Novel ranking strategy for relational Database

• A term tends to appear more frequently in a T
  with a larger size.
• Inter-Document Weight Normalization

19
Novel ranking strategy for relational Database

• Schema terms in query
• Value terms
• Schema terms
• Schema-based document frequency
• Assign the largest document frequency value among
  all terms to df
• Assign 1 to tf
• What if k is both value term and schema term?

20
Novel ranking strategy for relational Database

• Phrase-based Ranking
• Utilize phrase-based ranking to improve
  effectiveness.
• If a sub-query of Q, Pki,ki1,..kj, where iltj,
  appears in a document D, and ki-1 does not appear
  in an adjacent location to ki in this occurrence
  of P in D, and kj1 does not appear in an
  adjacent location to kj in this occurrence of P
  in D, then we define it as an occurrence of the
  phrase P in D.

21
Novel ranking strategy for relational Database
22
Novel ranking strategy for relational Database

• Suppose Q1, 2, 3, 4 and a document D in T is
  .. 1, 2, 3 .. 2, 3, 4 .. 2, 3, 4 .. 1, 2 .. 1
  ...
• 1, 2, 3 and 2, 3, 4 overlap
• Choose the phrase with the highest weight.

23
(No Transcript)
24
Novel ranking strategy for relational Database

• concept set(CQ)
• Phrase model
• a document that contains only some highly
  weighted terms
• Concept ranking model
• concept similarity value Sim(CQ, T).
• if (1) Sim(CQ, T1) gt sim(CQ,T2)
• or (2) Sim(CQ,T1)Sim(CQ,T2) and Sim(Q,T1) gt
  Sim(Q,T2), then T1 is ranked higher than T2.

25
Experiment results
26
Experiment results
27
Experiment results
28
Experiment results
29
Experiment results
30
Conclusion

• Four normalizations
• tuple tree size normalization
• document length normalization,
• document frequency normalization
• inter-document weight normalization.
• The results show that
• all the four new normalization factors are
  critical to search effectiveness
• phrase-based search and concept-based search
  improve effectiveness significantly
• our strategy is significantly better than related
  works and significantly outperforms Google.

31
BLINKS Ranked Keyword Searches on Graphs

• Hao Hey
• Haixun Wang
• Jun Yang
• Philip S. Yu