Effective XML Keyword Search with Relevance Oriented Ranking

1
Effective XML Keyword Search with Relevance
Oriented Ranking

• Zhifeng Bao, Tok Wang Ling, Bo Chen, Jiaheng Lu

2
Introduction

• XML Keyword search
• Inspired by IR style keyword search on the web
• Enables user to access information in XML
  database
• XML data modeled as a rooted, labeled tree
• Recent research efforts
• Efficiency
• Effectiveness

3
Effectiveness

• Capture users search intention
• Identify the target that user intends to search
  for
• Infer the predicate constraint that user intends
  to search via
• Result ranking
• Rank the query results according to their
  objective relevance to user search intention

4
State of the Art

• Search semantics design
• LCA (Lowest Common Ancestor)
• Node v is a LCA of keyword set Kw1, w2,,wk if
  the sub-tree rooted at v contains at least one
  occurrence of all keywords in K, after excluding
  the sub-elements that already contain all
  keywords in K
• SLCA (Smallest LCA)
• Node v is a SLCA of keyword set Kw1, w2,,wk
  if
• (1) v is a LCA of K
• (2) no proper descendant of v is LCA of K
• XSeek
• Infers the search intention based on the concept
  of objects and an analysis of the matching
  between keyword and data node

5
State of the Art (cont)

• Efficient result retrieval
• Designed based on a certain search semantics
• XKSearch, Multiway SLCA etc.
• Result ranking
• XRANK, XKSEarch, EASE
• They only consider
• Structural compactness of matching results
• Keyword proximity
• Similarity at node level

6
Problems Unaddressed

• Not address the user search intention adequately!
• Meaningfulness of query result
• SLCA is less meaningful in many cases
• Keyword Ambiguity Problems
• A keyword can appear both as an xml node type
  and as the text value of some other nodes
• A keyword can appear in the text values of
  different xml node types and carry different
  meanings

Neither SLCA nor Xseek can well address keyword
ambiguity
7
Meaningfulness
Problems

• Keyword query rock music
• Search intention find customers interested in
  rock music C3
• SLCA returns interest node of C3

8
Keyword Ambiguity
Problems

• Q customer, interest, art
• Ambiguity 1 customer, interest Ambiguity 2
  art
• Intention find customer whose interest is art
• less relevant or irrelevant result to be returned
  also --- C1,C3, B1s title

...
...
...
name
Oxford
customer
...
purchases
interests
name
ID
purchase
interest
C
2

street art
John Martin
9
Keyword Ambiguity (cont)
Problems

• Q customer, art
• art can be the value of interest node(C2, C4),
  name node(C3), or street node of customer(C1), or
  title node of book(B1)
• customer can be tag name of customer node, or
  (part of) value of title of(B1)

- How to rank C1 to C4 and B1?
10
Objectives Challenges

• Address the below as a single problem
• Search intention identification
• Query result retrieval
• Result ranking
• Extend original TFIDF from text database to XML
  database, while capture the hierarchical
  structure of XML data

• Challenges
• How to decide which sub-tree(s) with appropriate
  node types can capture user desired information
• How to return sub-trees of an appropriate size
  (i.e. contain enough but non-overwhelming
  information)
• How to rank those sub-trees by their relevance

11
Challenges

• Difficulty in applying TFIDF to XML
• XML DB carries semantic information while text DB
  contains pure text information. XML TFIDF must
  be aware of the underlying semantics.
• All contents of XML data are stored in leaf nodes
  only
• What is analogy of flat document in XML?
• Sub-tree classified according to its prefix path
• Normalization factor is not simply the size of
  sub-tree
• Structure of sub-trees may also infest the ranks

12
TFIDF Recap

• Rule 1 A keyword appearing in many documents
  should not be regarded as more important than a
  keyword appearing in a few. --- IDF
• Rule 2 A document with more occurrences of a
  query keyword should not be regarded as less
  important for that keyword than a document that
  has less. --- TF
• Rule 3 A normalization factor is needed to
  balance between long and short documents
• as Rule 2 discriminates against short documents
  which may have less chance to contain more
  occurrences of keywords.

13
Our Approach

• Extend IR-style keyword search techniques (like
  TFIDF) from text database to XML database, in
  order to capture the hierarchical structure of
  xml document
• by analyzing the knowledge of statistics of
  underlying XML data
• Major Contributions
• Identify users desired search-for node and
  search-via node(s) in a heuristic way
• Define XML TF (term frequency) and XML DF
  (document frequency)
• Confidence Formulas for search for/via candidates
• Define XML TFIDF Similarity
• Propose 3 guidelines specifically for xml keyword
  search
• Take keyword ambiguity problems into account
• Design a Keyword Search Engine XReal

14
Data Model

• Node type - Two nodes are of same node type if
  they share the same prefix path
• /storeDB/customers/customer/name vs.
• 
  /storeDB/books/book/publisher/name

• Value node text values contained in leaf node
• Structural node
• Single-valued node type, multi-valued node type
• Grouping type all its children are of same
  multi-valued type

storeDB
customers
books
...
...
book
customer
...
customer
...
ID
customer
ID
interests
publisher
title
name
authors
interests
...
...
...
ID
...
interest
C
3

name
name
interests
author
author
ID
interest
name
Art Smith
B
2

C
4

contact
address
...
rock music
Oxford
interest
book
C
1

Edward Martin
art
Rock Davis
customer
no
.
...
Sophia Jones
city
authors
...

1

purchases
street
title
ID
...
interests
name
ID
author
author
Mary Smith
B
1

purchase
interest
Art Street
fashion
John Williams
C
2

Art of Customer
Daniel Jones
Interest Care
street art
John Martin
15
XML TF and IDF

• XML DF (document frequency)
• The number of T-typed nodes that contain keyword
  k in their sub-trees in XML database.
• Granularity of similarity measurement is
  sub-trees of certain node type T
• XML TF (term frequency)
• The number of occurrences of a keyword k in a
  given value node a in XML database.

16
Infer the desired search-for node

• Guidelines A node type T is considered as a
  desired search for node if
• T is intuitively related to every query keyword
• XML nodes of type T should be informative enough
  to contain enough relevant information
• XML nodes of type T should be not overwhelming to
  contain too much irrelevant information
• Confidence of T as the search for node w.r.t.
  query q.
• product instead of sum is used to follow 1st
  guideline
• log part designed to follow 3rd guideline
• exponential part designed to follow 2nd guideline
• r is a decay factor in (0,1.

17
Infer the Search-Via Nodes

• Infer structural node to search via
• Structural node n is a good candidate if it is
  related to as many (but not necessarily all)
  keywords as possible
• Search via node type normally is not unique
• Infer individual value node to search via
• Statistics alone is not adequate to infer the
  likelihood of a value node as (part of) search
  via node
• Capture keyword co-occurrence

18
Capture keyword co-occurrence

• E.g. Q customer, name, rock, interest, art
• Easy to find name and interest have high
  confidence to be the search via nodes
• But hard to know rock is value of name or
  interest, art is
  value of interest or name
• How to differ customer C4 from C3?

19
Capture keyword co-occurrence

• Proximity factors for a value node v of type kt
  containing keyword k
• Given a query q and a certain value node v, if
  there are two keywords kt and k in q, s.t. kt
  matches the type of an ancestor node of v and k
  matches a keyword in v
• In-Query distance
• Distance between keyword k and node type kt in
  query q
• Favors kt appears before k
• Structural distance
• Depth distance between v and the nearest kt typed
  ancestor node of v
• Value-Type distance
• Max of the above two

20
Principles of XML keyword search

• Principle 1
• When searching for D-typed nodes via a
  single-valued type V, ideally only the values and
  structures nested in V-typed nodes can affect the
  relevance, regardless of the size of other typed
  nodes nested in D-typed nodes.
• However, TFIDF similarity in IR normalizes the
  relevance score of each document w.r.t. its size
  
• Principle 2 address keyword Ambiguity 2
• When searching for nodes of type D via a
  multi-valued type V, the relevance of a D-typed
  node which contains a query relevant V-typed
  node should not be affected (i.e. normalized) too
  much by other query-irrelevant V-typed nodes.
• Example query art - C4 should not be
  less relevant than C1

21
Principles of XML keyword search

• Principle 1 and 2
• Especially useful for interpreting pure keyword
  query - find search via node correctly
• Principle 3
• The order of keywords in a query is important to
  indicate the search intention
• Incorporate the search via confidence Cvia we
  defined before

22
XML TFIDF Similarity

• To calculate the similarity between the search
  for node and the query q
• Base case similarity between value node a and q
• Apply original TFIDF directly since a contains
  keywords only without any structure
• Recursive case similarity between structural
  node n and q
• Based on similarities of its children c and the
  confidence level of c as the node type to search
  via

TF
IDF
Normalization factor
23
XML TFIDF Similarity (cont.)

• Recursive Case
• Intuition 2. An internal node n is relevant to q,
  if n has a child c such that the type of c has
  high confidence to be a search via node w.r.t. q
  (i.e. large Cvia(Tc , q)), and c is highly
  relevant to q (i.e. large sim(q, c)).
• Intuition 3. An internal node n is more relevant
  to q if n has more query-relevant children when
  all others being equal.

Weighted sum of all ns childrens similarity and
their confidence to be the search via node
Overall weight of node n w.r.t query q which
essentially plays the role of a normalization
factor
24
Flowchart of answering a query

• Identify user search intention
• Compute the confidence of all possible candidate
  node types and choose desired search for node
  Tfor
• Relevance-oriented ranking
• Compute XML TFIDF similarity in a bottom-up
  approach from value nodes containing keywords up
  to nodes of type Tfor
• Return a ranked list of sub-trees rooted at nodes
  of type Tfor
• If more than one search for node type have
  comparable confidence, a ranked list for each
  search for node is returned

25
Experimental Result

• Data set
• DBLP, XMark, WSU, eBay
• Comparison
• Compare XReal with SLCA, Xseek
• Equipment
• Implement in Java
• Run on 3.6GHz pentium IV, 1 GB memory PC with
  Windows XP
• Berkeley DB java edition for storing keyword
  inverted lists and keyword frequency table

26
Search Effectiveness

• Accuracy in inferring the search for node
• Conducted by user survey
• Tested queries contain at least one of the two
  ambiguity problems
• Conclusion
• XReal works well, especially when the search for
  node is not given explicitly in the query

27
Search Effectiveness

• Result effectiveness
• Measured by precision, recall, F-measure
• Observations
• XReal achieves higher precision than SLCA and
  Xseek for queries that contain ambiguities
• XReal Performs as well as XSeek when queries have
  no ambiguity in XML data
• XReal Top-100 precision higher than overall
  precision
• F-measure also shows good overall effectiveness
  of both XReal and XSeek

28
Ranking Effectiveness

• Metrics
• Number of Top-1 answers that are relevant
• Reciprocal Rank (R-Rank)
• Mean Average Precision (MAP)

29
Efficiency Scalability

• Compare three adoptions of indices for XReal, and
  SLCA
• Dup
• Store only the dewey id and XML TF
• DupType
• Stores an extra node type (i.e. its prefix path)
• DupTypeNorm
• Stores an extra normalization factor Wa for value
  node

30
(No Transcript)
31
QA

• Thank You

32
(No Transcript)