Autonomous Ontology Extraction Using Hierarchical UNSO Hypercube Graph

1
Autonomous Ontology Extraction Using Hierarchical
UNSO Hypercube Graph

• Yosi Ben-Asher
• Shlomo Berkovsky
• Yaniv Eytani

2
Outline

• Data Integration
• UNSpecified Ontologies
• Hierarchical UNSO
• Autonomous Maintenance Operations
• Ontology Extraction

3
Data Integration Motivation

• Consider two E-Commerce ads
• Selling red BMW car, 2000, good condition,
  10,000 miles
• Wanna buy second hand sports car with leather
  seats
• Different descriptions
• Might refer to the same object

Smart mechanism for integration of heterogeneous
data is needed
4
Data Integration - Ontology

• Ontology captures the semantic relationships
  between objects
• Definition
• A shared formalization of a conceptualization of
  a domain

5
Semantic Data Management HuperCup
Schlosser et al.

• Implemented ontology-based
• data management over hypercube
• topology
• Objects are mapped to the hypercube graph
• using a predefined ontology.
• Efficient broadcast and search
• Achieved within a logarithmic in N (number of
  users) steps

6
Unspecified Ontology (UNSO)

• UNSO generalizes HyperCup by defining the notion
  of UNSpecified Ontology
• Ontology which is not specified
• i.e, no master ontology, each user specifies
  his own ontology
• UNSO acts as a classification method
• Similar objects that are mapped to the UNSO graph
  are in a close proximity
• i.e, in near vertices (nodes) in the hypercube
  topology

7
UNSO Unspecified Descriptions

• Each object is described as a vector of
  (propval) pairs
• Selling red BMW car, 2005, good condition, 10000
  miles
• becomes
• productcar, manufacturerBMW, colorred,
  conditiongood, production_year2005,
  mileage10000
• Each vector is projected on the UNSO graph

8
UNSO Mapping Description vectors

• Description are mapped using hashing
• Different props are hashed to different
  hypercube dimensions
• vals are hashed to numeric values within the
  respective dimension
• How to handle ambiguity?
• Similar patterns for objects descriptions assumed
• Common sense
• Zipfs law
• props and vals undergo simple standardization
  using lexical reference system (e.g., WordNet)
• e.g., car auto automobile

9
UNSO A Mapping Example

• Consider a simple ontology for cars domain
• Color dark0, bright1
• Gearbox automatic0, manual1
• Size small0, large1
• White manual Mini-Minor
• Is mapped to a vector (1,1,0)

gearbox
(1,1,1)
size
color
(0,0,0)
10
UNSO Multi-Layered Hypercube

• MLH-UNSO is comprised of nodes recursively
• containing hypercubes
• Description vectors are
• partitioned between different layers
• Example for 3-dimensional hypercubes
• Higher layer (0,1,1)
• Lower layer (0,0,1)

(0,1,1,0,0,1)
11
UNSO Pros and Cons

• No predefined ontology
• Unlimited range of props and vals
• Locality similar objects are mapped to adjacent
  locations
• All the props are of the same importance
• Un-weighted flat ontology
• Sparse and unbalanced distribution in the graph
• Queries processing complexity is fixed
• Inefficient processing of typical search queries

12
HUNSO Hierarchical UNSO

• Some props are more significant
• manufacturer, color, production_year
• seats_type, wheels_type
• Significance of props is assumed to be
  correlated with statistical frequency of
• Appearance in objects descriptions
• Appearance in search queries descriptions

13
Real-life data (130 E-Commerce ads)
14
HUNSO - Layers

• Statistical frequencies of the properties are
  collected over time
• For example during users interactions
• Significant properties are located in the higher
  layers, insignificant in the lower ones
• Advantages
• Variable-length search operation
• Fast processing of popular queries
• Allows for self-management and load-balancing
• Autonomous ontology extraction

15
HUNSO Autonomous Maintenance

• Ontology Evolves
• New types of objects, props and vals
• Old type become obsolete
• To maintain a dense ordered structure in the
  hypercube, we use three autonomous operations
• EXPAND convert a dense node into lower-layer
  hypercube
• SHRINK convert a sparse hypercube into
  higher-level node
• SWAP exchange between properties of 2
  adjacent layers
• Due to inconsistency of properties significance
  order

16
HUNSO EXPAND A Node

• Performed when
• A node gets overloaded we objects descriptions
• No single property dominates
• i.e. all props appear roughly with the same
  frequency
• Basic stages
• Choose K most frequent properties to form the
  lower-layer hypercube dimensions
• Remap the descriptions from the single node in
  the higher-layer hypercube to the nodes of the
  new lower-layer hypercube

17
HUNSO EXPAND (for K3)
After
Before
18
HUNSO SHRINK A Node

• Performed when
• A lower-layer hypercube is sparse
• Basic stages
• Remap the descriptions from the nodes of the
  lower-layer hypercube to a single node of
  higher-layer hypercube

19
HUNSO Splitting (example)
After
Before
20
HUNSO SWAP

• Performed when
• Lower hypercubes gets overloaded with objects
  descriptions
• props significance order is inconsistent between
  layers
• A prop in the lower layer appears more than a
  prop in the higher-layer
• Basic stages
• Find the dominating property in all lower-level
  hypercubes
• Find the least frequent property in higher-level
  hypercube
• Swap the respective dimensions between the layers
• Remap the descriptions from the nodes of the
  lower-layer hypercubes to the higher-layer
  hypercube and vice-versa

21
HUNSO SWAP (example)
22
Ontology Extraction

• Objects properties for each domain are ordered
  in an hierarchical structure
• Lower layers represents a specification of higher
  layers
• Consider red 2000 BMW vs. red 2000 BMW,
  leather seats, manual, good condition
• Domain ontology constantly evolves
• No data engineering by human experts is
  required
• Flexible ontologies, sensitive to dynamic changes
  in the objects

HUNSO automatically captures the commonalities in
the underlying objects descriptions
23
Future Work

• Using NLP tools for recognition of terms affinity
• Calculate weights and distances between the
  properties and particular values
• Ranking results
• Failed match, k nearest neighbors
• Using world facts from external knowledge
  repositories

24
Q A
Thank You!