Structured Text Retrieval Models

1
Structured Text Retrieval Models
2
Str. Text Retrieval

• Text Retrieval retrieves documents based on index
  terms.
• Observation Documents have implicit structure.
• Regular text retrieval and indexing strategies
  lose the information available within the
  structure.
• Text Retrieval desired based on structure.
• e.g. All documents having George Bush in the
  caption of a photo.

3
Models for Str. Text Retrieval

• PAT Expressions
• Overlapped Lists
• Proximal Nodes
• List of References
• Tree-based
• Query Languages (SFQL,CCL)

4
Proximal Nodes

• By Gonzalo Navarro and Ricardo Baeza-Yates
• Based on hierarchical structure of documents
• Structure computation is static and all
  structural elements are defined. nodes
• Model attempts to define operators on these nodes
  based on their definition and content.
• Only nodes at a particular hierarchy are returned
  as results.

5
Proximal Nodes
Document
Chapter
Chapter
Section
Section
Section
6
Proximal Nodes

• Nodes are structural in nature, e.g. Chapter,
  Section, etc.
• Each node has a defined segment (Contiguous part
  of text)
• Operators are defined with respect to this model.
• Structure operators and Text operators.

7
Proximal Nodes

• Structure Operators
• Name
• Inclusion
• Positional Inclusion
• Distance operators
• Child/Parent operators
• Set Manipulation operators
• Text Operators
• Match

8
Retrieval on Evidence

• By Mounia Lalmas
• Based on documents made up of objects.
• Objects are modeled as independent entities and
  can be in different media, language or locations.
• Document indexing degree of uncertainty that
  the index term actually represents the object.
• Uncertainty must be captured to get better
  results.
• Use the Dempster-Shafer theory of evidence

9
Retrieval on Evidence

• Model takes into consideration disparity between
  indexing vocabularies.
• Aggregation of indexing vocabulary and also the
  aggregation of the uncertainty.
• Object o ? O and a type t ? T, the function type
  is defined as O ??(T)
• Aggregation is defined over objects and composite
  object types contain all the types of the
  contained objects

10
Retrieval on Evidence

• Indexing vocabulary is defined over a
  proposition-space. e.g. Wine (english,text),
  Blue(colour,feature)
• Sentence space defines that indexes in the same
  proposition space can be used together.
• Semantic between indexing vocabulary is
  maintained using the the notion of worlds.

11
Retrieval on Evidence

• Each type t has S, W, v, p
• St is the sentence space for a type
• W is the possible worlds associated with St
• vt is true, false over Wt x Pt
• ?t is true, false over Wt x St
• Logical and equivalence between sentences is
  built around the notion of their semantics being
  equivalent in all or most worlds.

12
Retrieval on Evidence

• However, the uncertainty of the representation
  remains.
• This is represented by the weighting function
  based on the Dempster Shafer model.
• These objects and their syntactic and semantic
  models are aggregated for the objects which
  contain them. E.g. A section containing sentences
  indexed by terms a,b,c,d.. Will be equivalent to
  sentences over the worlds also implying a,b,c,d

13
Comparisons

• Proximal Nodes is based on structured documents.
  It presents the matter clearly and provides
  approaches towards building a software
  architecture. It presents findings of conducted
  experiments.
• The Evidence paper tries to model heterogeneous
  documents, made up of different media, languages,
  etc. Overall the model is complex and no results
  are given to its implementation and performance.