GinsengA Guided Input Natural Language Search Engine for Querying Ontologies

1
GinsengA Guided Input Natural Language Search
Engine for Querying Ontologies

• Making the Semantic Web Accessible to the Casual
  User

2
Problems of the Semantic Web

• Computational complexity
• Expressiveness of DL
• Markup prisoners dilemma
• Formal logic is not simple
• only 1 of Web users use Boolean Logic (Spink et
  al. 2001)
• Spoerri (1993) showed
• People arent at ease at constructing Boolean
  queries? MIT students constructed wrong queries

How can we bridge the gap between the
(description) logic-based Semantic Web and
real-world users, who are at least ill at ease
and, oftentimes, unable to use formal logic
concepts?
3
Possible Answers

• Query by example
• But the Semantic Web is in NF2 (and not on 1NF)
• Graphical Query Languages
• Are also formal (but maybe more intuitive)
• Natural Language (Androutsopoulos et al. 1995)
• High computational complexity
• Difficult to adapt to new domains
• Offers no guidance
• Usually only process a part of English while
  suggesting full comprehension

4
GinsengDemo
5
Mooney Geo Ontology

• classes 19, properties 30, instances 1350

6
Ginseng User Guidance
7
GinsengArchitecture
SimpleQuestion Grammar
Full Grammar
Grammar Compiler
Incremental parser
Jena SPARQL Engine
Jena Ontology Model
8
GinsengGrammar

• (1) ltSTARTgt ltOQgt ? SELECT ltltOQgtgt
  WHERE (ltltOQgtgt)
• (2a) ltOQgt which ltsubjectgt ltverbgt
  ltltsubjectgtgt ltltsubject1gtgt ltltverbgtgt
• (2b) ltOQgt what ltsubjectgt ltverbgt
  ltltsubjectgtgt ltltsubject1gtgt ltltsubjectgtgt)
  (ltltsubject1gtgt ltltverbgtgt
• (3) ltsubjectgt state ?state
  ltRDFtypegt ltgeostategt (typeltgeostategt)
• (4) ltverbgt borders ltobjectgt -
  ltgeobordersgt ltltobjectgtgt (domainltgeostategt,
  rangeltgeostategt)
• (5) ltobjectgt New York City ?newyorkcity
  ltgeonewYorkCitygt (typeltgeocitygt,
  ltgeocapitalgt)
• (7) ltobjectgt Mississippi ?mississippi
  ltgeomississippiStategt (typeltgeostategt)

9
GinsengGrammar Details I

• Static Grammar
• 120 mostly empirically constructed rules
• E.g., rules (1), (2a), (2b)
• Dynamic Grammar
• Parses Grammar and turns major elements into
  grammar elements
• Class stateltNCcgt state?statelthttp//www.mo
  oney.net/geostategtltNCcgt states?statehttp/
  /www.mooney.net/geostate
• State InstanceltNIgt West Virginia
  ?westvirginia lthttp//www.mooney.net/g
  eowestVirginiagt

10
GinsengGrammar Details II

• PropertiesltowlObjectProperty
  rdfID"isMountainOf"gt ltrdfsdomain
  rdfresource"Mountain"/gt ltrdfsrange
  rdfresource"State"/gt ltginsengignore
  rdfvalue"id text"/gt ltginsengphrase
  rdfvalue"lies in"/gt ...lt/owlObjectProperty
  gt
• ?
• ltVCUgt lies in ltObjgt-ltgeoisMountainOfgt
  ltltObjgtgtltVCUgt lies in the state of ltObjgt-
  ltgeoisMountainOfgt ltltObjgtgtltVIUgt is
  in ltObjgt-ltgeoisMountainOfgt ltltObjgtgt

11
Evaluation

• Dataset NL- Benchmark (Tang and Mooney 2001)
• 3 domains (US-Geography, jobs, restaurants)
• 1770 Questions (generated by undergraduates and
  web-interface)
• Setup
• Query reformulations with 30 randomly chosen Qs
• Precision/recall
• SUS (Brook 1997)
• 30 subjects(each a pair of tools)

12
Evaluation Results

• Ginseng and SQL, Ginseng was
• faster than SQL queries (p 1.8 )
• rated a better integrated (t-test p 3.5)
• easier to learn (t-test p 0.97).
• Higher overall SUS score (but not significant)
• Retrieval Performance
• Parsed 67.1 of appropriate queries (40 overall)
• Precision 92.8
• Recall 98.4.

13
Limitation

• Learning cost
• Subject pool
• Size
• Population
• Benchmark with other tools
• PRECISE had comparable results (e.g., Popescu et
  al. 2003)
• NLP vs. Ginseng