Searching XML Documents via XML Fragments

1
Searching XML Documents via XML Fragments

• D. Camel, Y. S. Maarek, M. Mandelbrod, Y. Mass
  and A. Soffer

Presented by Hui Fang
2
Background(1) --- Data
Why is semi-structured data important?
Un-structured Data
Well-structured Data
Lack of flexibility Lack of extensibility
ltpapergt lttitlegt XIRQL lt/titlegt ltauthorgt N.Fuhr
lt/authorgt ltauthorgt K.Grobjohann lt/authorgt ltconfgt
SIGIR lt/confgt lt/papergt
Lack of the logical structure of a document.
3
XML in a nutshell
ltbook id25gt ltyeargt1997lt/yeargt ltauthorgt
Karen Sparck Jones lt/authorgt ltauthorgtPeter
Willett lt/authorgt ltpublishergtMorgan
Kaufmannlt/publishergt lttitlegt Readings in
Information Retrieval lt/titlegt lt/bookgt

• Hierarchical data format
• Nested element structure having a root
• Self describing data (tags), schema is attached
  to the data itself.

4
Background(2) --- Query
How to query semi-structured data (e.g. XML data)
?
5
Related Work

• DB-oriented approaches
• E.g. XML-QL, XQL, XQUERY

WHERE ltbookgt lttitlegtHarry
Potter lt/titlegt ltauthorgtalt/authorgt,
ltyeargt y lt/yeargt lt/bookgt in books.xml,
ygt2002 CONSTRUCT ltresultgt
ltauthorgttlt/authorgt lt/resultgt

• DBIR approaches
• E.g. XIRQL
• IR-oriented approaches
• E.g. this paper

6
Problem Refinement---CAS Search

• Document collection
• XML documents
• Each document is a hierarchical structure of
  nested elements
• Markup in the document mainly serves for exposing
  the logical structure of a document.
• Query
• content explicit references to the XML
  structure
• specifies the target element need to be returned

An example
Retrieval all articles from the years 1999-2000
and deal with works on nonmonotonic reasoning. Do
not retrieve articles that are calendar/call for
papers.
7
Approach

• Compare apple and apple
• Recall vector space models
• Both documents and queries are expressed in free
  text.
• Compare unstructured data to unstructured data
• This paper
• Search XML documents via XML fragments

8
Query---XML Fragments(1)

• Topic 1 Find all books about fishing
• ltbookgt fishing lt/bookgt
• Topic 2 Find all books having a title about
  search
• ltbookgt lttitlegt fishing lt/titlegt lt/bookgt

More intuitive More flexible
9
Query --- XML Fragment(2)

• Limited expressiveness
• E.g. Finding figures that describe the Corba
  architecture and the paragraphs that refer to
  those figures.

Requires a join operation between two elements
figures and paragraphs
10
Recall Text Retrieval Task

• Give a query
• According to the retrieval formula, compute the
  relevance score for each document
• Rank the documents according to relevance score.

11
Extending the Vector Space Model(1)

• Indexing unit
• E.g. (Harry Potter , /book/title)
• Can be matched with
• (Harry Potter ,/book)
• (Harry Potter ,/book/sec/title)
• Retrieval Formula

12
Extending the Vector Space Model(2)
If c is rare, idf(t,c) would be high in spite of
t being very common.
13
Evaluation

• Runs
• Partial-match
• Partial-match. merge-idf
• Partial-match.merge
• Fuzzy-match.merge-idf
• Flat (ignore context)

14
Result(1)

• Result for free-text-oriented topics
• An example topic
• ltyrgt1995,1996,1997,1998,1999lt/yrgt
• ltbdygtXML Electronic commerce lt/bdygt

15
Result(2)

• Result for context-oriented topics
• An example topic
• ltatlgt Content-Based retrieval of video
  databaseslt/atlgt

16
Summary

• Using XML fragments with an extended vector space
  model is promising.
• Use different solutions for different types of
  applications
• Something wrong?

17
Another Problem --- CO Search

• Document collection
• XML documents
• Query
• a set of keywords
• Task Find smallest element satisfying the query

Challenge rank the components instead of document
18
Possible Solutions
,where
Possible Method(1) treat each component as a
document.
Problem with this method XML components are
nested.
ltarticlegt t1 ltsecgt ltpgt t2lt/pgtlt/secgt lt/articlegt
19
Possible Solutions (Cont.)
,where
Possible Method(2) counting TF at the component
level computing N DF at the document level.
ltarticlegt ltsecgtt1lt/secgt ltsecgtt1lt/secgt ltsecgtt2lt/
secgt lt/articlegt
20
Proposed Solution

• Create a index for each component type
• Elements in each index are regarded as documents
• Keep N, DF,TF for the specific component type
• Can apply the regular vector space model on each
  index
• Given a query
• Run the query in parallel on each index
• Return one ranked list of results, one from each
  index
• Normalize the scores in each index into the range
  (0,1)
• Achieved by computing
• Merge the normalized results into a one ranked
  list of all components

Assume the set of potential components to be
returned must be known in advance. Assume no
nesting of the same component.
21
Conclusion

• Possible solutions to solve the following
  challenges.
• Challenge 1 (Information/Doc Unit) What is an
  appropriate information unit?
• Document may no longer be the most natural unit
• Components in a document may be more appropriate
• Challenge 2 (Query) What is an appropriate query
  language?
• Keyword (free text) query is no longer the only
  choice
• Constraints on the structures can be posed

22
References

• Retrieving the most relevant XML components, by
  Y. Mass, M. Mandelbrod. INEX03 workshop.
• Searching XML Documents via XML fragments, by D.
  Carmel, Y. S.Maarek, M. Mandelbrod, Y. Mass and
  A. Soffer. SIGIR03
• XIRQL A Query Language for Information Retrieval
  in XML Documents by N. Fuhr, K. Großjohann.
  SIGIR02