Identifying and Ranking Relevant Document Elements

1
Identifying and Ranking Relevant Document Elements

• Andrew Trotman
• Richard A. OKeefe
• Department of Computer Science
• University of Otago
• Dunedin, New Zealand

2
IR System Design

• Build a corpus tree
• Annotate the indexes
• Convert the topic into a Boolean expression
• Annotate the accumulators
• Resolve the query against the indexes
• Compute the coverage
• Choose the results

3
Document Trees

• Each well-formed document is a tree
• Attributes _at_ and instancing are included
• Trees are not unique

4
Build a Corpus Tree

• The superimposition of all document trees
• Built as encountered
• Label each node as encountered
• Matches the structure of no one document
• Has every path through every document

5
Annotate the Indexes

• Traditionally in inverted files
• ltd1,f1gtltd2,f2gt,,ltdn,fngt
• Now
• ltd1,p1,f1gt,ltd2,p2,f2gt,ltdn,pn,fngt
• or
• ltd1,p1,w1gt,ltd2,p2,w2gt,ltdn,pn,wngt
• Where
• pn is a node from the corpus tree
• wn is a term number used for phrase search

6
Fast Access Indexes
term1

• Sorted by pn
• Store d before w
• Only examine some pn
• Decompress only those dn needed

7
Nodes to Use

• Build a big list of tags and children nodes
• about (./p, term)
• about(./_at_c, term)
• about(./p2, term)
• about(./sec/p, term)
• Tag equivalence
• e.g. _at_c p2

8
Convert Topic to Boolean

• For each about clause
• Collect like terms
• Mandatory terms are ANDed
• Optional terms are ORed
• Exclusion terms are ANDed
• Build a sub-expression
• Mandatory AND ( OR Optional) NOT Exclusion
• Operators around about clause are preserved
• Boolean / BM25 hybrid Ranking

9
Resolve Query Against Indexes

• Build then walk the parse tree
• At leaves, load postings and process
• BM25 score added to each document accumulator
• Coverage added to each document accumulator
• Convert into bit-string one per leaf
• At nodes, Boolean operate
• Boolean operate the bit-strings
• Ignore the accumulators
• At root of parse tree
• Bit-string of relevant documents
• Accumulators with BM25 scores and coverage

10
Annotate the Accumulators

• Traditionally
• Per document
• One real-valued cell
• Now
• Per document
• One real-valued cell
• Per node in corpus tree
• One integer-valued cell
• As Always
• Accumulators are stored sparsely
• There is only one set of accumulators per query

11
Compute the Coverage

• For each posting
• document elements at or above the posting
• For a term / document pair
• superimposition of coverage of each posting
• Weighted Coverage for a document node
• The how many times a given node is covered by
  unique terms (how many unique search terms lie at
  or below the node)

12
Properties of Weighted Coverage

• The highest coverage
• Must be at the root of the document tree
• For all nodes with the highest coverage
• Lowest nodes are most information dense
• As weighting decreases
• specificity increases
• Likely that
• tags between two relevant tags are also relevant

13
Choose The Results

• Recall determined by Boolean expression
• Precision determined by
• Documents ranking
• BM25 score
• Node ranking
• Weighted coverage
• Use the lowest nodes with the highest coverage
• For SCAS
• Restrict weighted coverage to required nodes

14
INEX03 Results

• Did badly at inex_eval
• Prefers tags high in the tree
• Doesnt return overlapping results
• Creates a dependence of tags on documents
• Did well at inex_eval_ng
• Coverage related to density
• Relevant documents have relevant sections
• Did better at generalized than specific
• Tags returned in document order by BM25

15
Questions?