Mercure applied to XML retrieval

1
Mercure applied to XML retrieval
Karen SauvagnatSIG-RFI team, IRIT118 route de
Narbonne31062 Toulouse Cedex 4Francesauvagnat_at_
irit.fr
2
Plan

• Mercure system
• INEX approach
• General system architecture
• Indexing the INEX corpus and queries
• Retrieval
• Submitted runs
• Results
• Perspectives

3
Mercure system

• Mercure is based on a connectionist model
• The system has been constantly redesigned and
  updated to accommodate the information retrieval
  to users needs
• Frequent participation to the TREC evaluation
  initiative.

4
Main features of Mercure system

• It supports any text format (free text, markup
  text, )
• It takes into account term proximity
• The search engine

• Formatting extraction of the textual part of
  documents
• Indexing
• Searching. Two kinds of queries
• list of terms (free text queries)
• document
• Relevance feedback blind or driven by the user

• full-text information
  retrieval system
• full-text information retrieval
  system

5
INEX APPROACH

• General system architecture
• Indexing the INEX corpus and queries
• Retrieval
• Submitted runs
• Results

6
General System Architecture
7
Query indexing
8
Query examples

• CO query no 111
• lttitlegt natural language processing -
  programming language - modeling language
  human language lt/titlegt
• ltkeywordsgt natural language processing, human
  language technology, computational linguistics,
  speech processing, parsing natural language,
  natural language understanding, natural language
  interface lt/keywordsgt
• CAS query no 76
• lttitlegt//article.//yr2000 or .//yr1999)
  AND about(.,intelligent transportation
  system)//secabout(.,automationvehicle)
  lt/titlegt

9
Retrieval

• Searching
• using Mercure
• with or without using term proximity
• Results ordered list of 1000 documents for each
  query

10
Structured retrieval

• Two modules, which aim at filter the most
  specific and exhaustive elements of the documents
  returned by Mercure
• a content-oriented-module, dealing with queries
  composed of simple keywords terms (CO queries)
• a content-and-structure-oriented module,
  performing queries containing both explicit
  references to the XML structure and content
  constraints (CAS queries)

CAS queries
Ordered list of results (documents)
Formatting
Content-and-structure-oriented module
Content-oriented module
Indexing
Ordered list of results (documents/elements)
Ordered list of results (documents/elements)
Ordered list of results (documents)
11
Retrieval with CO queries the content-oriented
module

• Browses through documents returned by Mercure
• Elements types that can be retrieved are
  pre-specified manually, according to the DTD

ltarticlegt ltfnogtlt/fnogt ltdoigtlt/doigt ... ltabsgtltpgtSoft
ware based on nonlinear shell theory can
simulate 3D motions related to real
fabric-manufacturing processes.
...lt/pgtlt/absgt lt/fnogt ltbdygt ltsecgt ltpgtThe
automotive and aircraft industries (among many
others) have streamlined their design and
manufacturing processes by adopting CAD/CAM
approaches. lt/pgt ... lt/secgt ltsecgt
ltstgtLiterature reviewlt/stgt ltpgt The state of the
art in fabric drape simulation can be traced to
the work of Peirce, ... lt/pgt lt/secgt lt/bdygt lt/art
iclegt
ltarticlegt ltfnogtlt/fnogt ltdoigtlt/doigt ... ltabsgtltpgtSoft
ware based on nonlinear shell theory can
simulate 3D motions related to real
fabric-manufacturing processes.
...lt/pgtlt/absgt lt/fnogt ltbdygt ltsecgt ltpgtThe
automotive and aircraft industries (among many
others) have streamlined their design and
manufacturing processes by adopting CAD/CAM
approaches. lt/pgt ... lt/secgt ltsecgt
ltstgtLiterature reviewlt/stgt ltpgt The state of the
art in fabric drape simulation can be traced to
the work of Peirce, ... lt/pgt lt/secgt lt/bdygt lt/art
iclegt
ltarticlegt ltfnogtlt/fnogt ltdoigtlt/doigt ... ltabsgtltpgtSoft
ware based on nonlinear shell theory can
simulate 3D motions related to real
fabric-manufacturing processes.
...lt/pgtlt/absgt lt/fnogt ltbdygt ltsecgt ltpgtThe
automotive and aircraft industries (among many
others) have streamlined their design and
manufacturing processes by adopting CAD/CAM
approaches. lt/pgt ... lt/secgt ltsecgt
ltstgtLiterature reviewlt/stgt ltpgt The state of the
art in fabric drape simulation can be traced to
the work of Peirce, ... lt/pgt lt/secgt lt/bdygt lt/art
iclegt

• Searches occurrences of query terms in all
  retrievable elements
• Returns elements containing the greatest number
  of query terms
• If more than 2 elements are supposed to be
  relevant, the module returns the root element

12
Retrieval with CAS queries use of the
content-and-structure-oriented module

• Three parts used in the query
• target element
• content constraint
• year constraint

• Browses documents returned by Mercure
• Searches occurrences of query terms in all target
  elements
• Returns the target element containing the
  greatest number of query terms
• If the target elements do not contains any of the
  terms of the content constraint, the document is
  removed from the list of results
• Filters elements according to the article
  publication date (if there is a year constraint)

ltarticlegt ltfnogtlt/fnogt ltdoigtlt/doigt ... ltabsgtltpgtSoft
ware based on nonlinear shell theory can
simulate 3D motions related to real
fabric-manufacturing processes.
...lt/pgtlt/absgt lt/fnogt ltbdygt ltsecgt ltpgtThe
automotive and aircraft industries (among many
others) have streamlined their design and
manufacturing processes by adopting CAD/CAM
approaches. lt/pgt ... lt/secgt ltsecgt
ltstgtLiterature reviewlt/stgt ltpgt The state of the
art in fabric drape simulation can be traced to
the work of Peirce, ... lt/pgt lt/secgt lt/bdygt lt/art
iclegt
13
Submitted runs

• Aim of our experiments
• test whether a full-text information retrieval
  system can be easily adapted to structured
  retrieval
• measure the effect of term positions in the INEX
  query types
• 5 runs performed, using or not term proximity
• 2 for the CO task
• 2 for the SCAS task
• 1 for the VCAS task

14
Results
15
Discussion (1)

• Runs using term positions are definitely better
  than simple search for both query types
• For the CO task, elements that can be returned by
  the content-oriented module are preselected
  manually.
• statistics or aggregation methods may be used to
  find those elements automatically

16
Discussion (2)

• The content-and-structure-oriented module is not
  able to perform all the content and structural
  constraints it processes only content
  constraints on the target element and year
  constraints
• topic 90 //articleabout(./sec, (...)
  electronic commerce (...))//absabout(.,trust
  authentication)
• topic 84 //pabout(.,overview distributed
  query processing join)
• Query processing is relatively slow, because the
  modules have to browse all documents returned by
  Mercure

17
Perspectives

• An indexing model taking into account structural
  and content information of documents seems to be
  necessary
• The combination of different IR measures (IDF,
  IEF,..) could be tested

18
The fetch and browse model (1)

• Fetch
• Horizontal preselection of documents Di
  satisfying

19
The fetch and browse model (2)

• Browse
• Vertical selection of most specific units within
  the preselected documents (fetch)
• Recursive case
• Stop case

failed
20
Weights calculation formula for retrieval (1)

• Weight of a term i in a document j
• where
• hi parameters which depend on the documents
  collection
• tfi frequency of term ti in document dj
• N total number of documents
• ?l mean size of documents
• dlj size of document j
• ni number of documents which contains the term
  ti

21
Weights calculation formula for retrieval(2)

• Weight of a term i in a query u
• Query evaluation

Where qtfui is the frequency of term ti in query
u
22
Term proximity

• Modification of the ranking function
• Documents having close query terms compute a new
  input value
• Where
• a is a constant parameter so that
  . a is set to 4 for INEX2003 experiments
• proxi,i-1 is the number of terms separating the
  query terms ti and ti-1 in the window of a
  terms..