Extracting World and Linguistic Knowledge from Wikipedia Simone - - PDF document
Extracting World and Linguistic Knowledge from Wikipedia Simone Paolo Ponzetto Michael Strube University of Heidelberg EML Research gGmbH Outline Introduction Deriving world knowledge from Wikipedia Leveraging linguistic knowledge
Simone Paolo Ponzetto Michael Strube University of Heidelberg EML Research gGmbH
Introduction Deriving world knowledge from Wikipedia Leveraging linguistic knowledge Applications Outlook and future work Conclusions
Introduction Deriving world knowledge from Wikipedia Leveraging linguistic knowledge Applications Outlook and future work Conclusions
The crisis at General Motors threatens to drag down Adam Opel, a storied German brand that GM bought 80 years ago, on the eve of the Great Depression. Many in the industry say Opel has a future only if it can get a temporary helping hand from the German government. But whether Chancellor Angela Merkel will make available the public financing needed to help release Opel from the clutches of General Motors now depends on a reluctant government, an influential automotive union that wants politicians to save jobs, and employees who yearn to re-establish Opel as an independent German company.
source: Herald Tribune Europe, March 6, 2009
What about a widely used resource like WordNet?
The crisis at General Motors threatens to drag down Adam Opel, a storied German brand that GM bought 80 years ago, on the eve of the Great Depression. Many in the industry say Opel has a future only if it can get a temporary helping hand from the German government. But whether Chancellor Angela Merkel will make available the public financing needed to help release Opel from the clutches of General Motors now depends on a reluctant government, an influential automotive union that wants politicians to save jobs, and employees who yearn to re-establish Opel as an independent German company.
source: Herald Tribune Europe, March 6, 2009
What about a widely used resource like WordNet? And Cyc?
The crisis at General Motors threatens to drag down Adam Opel, a storied German brand that GM bought 80 years ago, on the eve of the Great Depression. Many in the industry say Opel has a future only if it can get a temporary helping hand from the German government. But whether Chancellor Angela Merkel will make available the public financing needed to help release Opel from the clutches of General Motors now depends on a reluctant government, an influential automotive union that wants politicians to save jobs, and employees who yearn to re-establish Opel as an independent German company.
source: Herald Tribune Europe, March 6, 2009
What about a widely used resource like WordNet? And Cyc? Let’s check Wikipedia on that topic!
Introduction Deriving world knowledge from Wikipedia Leveraging linguistic knowledge Applications Outlook and future work Conclusions
project-specific domain knowledge bases: + very high quality – small domain – reusability – high cost
d d d
LTE−Lite−25
d d d d d d d has−hd−drive has−system−software uses−disk
Storage−Space
storage−space access−time has−cpu
Seagate−ST−3144 LTE−Lite−20
clock−frequency has−central−unit developed−by
Compaq Trackball
has−ball
Ball Notebook AT−Bus−HD−Drive Hard−Disk Capacity−MB−Pair Hard−Disk−Drive System−Software Computer−System Central−Unit Clock−MHz−Pair CPU Clock−Frequency Access−Time Workstation PC Time−MS−Pair
developed−by
Seagate
has−trackball
WordNet + pretty high quality + good coverage of everyday language + many languages – very high cost – sense proliferation – coverage in domains arbitrary
entity physical entity abstract entity thing thing
causal agent substance process abstraction change freshener horror jimdandy stinker whacker
Cyc + pretty high quality + good coverage of everyday language + common sense knowledge – very high cost – coverage in domains arbitrary – English only
Ontology Learning from Text + low cost + potentially domain independent – mostly only small domains – low quality
General Motors isa isa isa isa US car company car company German car company Opel belongs to
Manual approach
+
Knowledge is manually input by human experts
➠ it produces high-quality information −
Limited amount of human experts
➠ expensive and low scalability to cover all domains
Automatic approach
+
It requires minimal supervision on large amounts of data
➠ low cost and scalable −
Overall quality lower than humans
➠ unconstrained output, not necessarily ‘ontologized’
And Wikipedia? “one of the most fascinating developments of the Digital Age” “incredible example of open-source intellectual collaboration” “faith-based encyclopedia” “a joke at best”
And Wikipedia? “. . . an expert-led investigation carried out by Nature – the first to use peer review to compare Wikipedia and Britannica’s coverage of science . . . revealed numerous errors in both encyclopedias, but among 42 entries tested, the difference in accuracy was not particularly great: the average science entry in Wikipedia contained around four inaccuracies; Britannica about three.” (Nature 15. Dec. 2005)
And Wikipedia? + low cost + very good coverage, domain independent + very many languages + up to date – ?? We evaluate quality empirically!
we are after a “steak and lobster” combination . . .
manual approaches achieve high quality for a limited coverage automatic ones achieve large coverage for a lower quality ➠ use manually annotated semi-structured input ➠ develop lightweight methods to generate large-coverage,
high-quality structured output
Wikipedia is . . .
e.g. MediaWiki, running on top of a LAMP architecture
Wikipedia is . . . 1. domain independent
➠
it has a large coverage 2. up-to-date
➠
to process current information 3. multilingual
➠
to process information in many languages
category network
WikiRelate! (Strube & Ponzetto, 2006):
➠ use the category network as a semantic network . . . ➠ . . . to compute semantic relatedness
WikiRelate!, ESA and WLM leverage different features of Wikipedia
topics of articles
➠
task label the relations between categories as isa and notisa
transform a thematic categorization into a fully-fledged taxonomy
e.g. categories under WIKIPEDIA ADMINISTRATION
strings: MEDIAWIKI, TEMPLATE, USER, PORTAL, CATEGORIES,
ARTICLES, PAGES
still to be processed
ALBUMS BY ARTIST MILES DAVIS ALBUMS is−refined−by FRENCH CUISINE CUISINE BY NATIONALITY is−refined−by
network
relations to be analyzed
SCIENTISTS COMPUTER SCIENTISTS
BRITISH COMPUTER SCIENTISTS
head word (or lemma)
using morpha
ISLAMIC MYSTICISM ISLAM
head of one of the categories occurs in non-head position in the
MYSTICISM and ISLAM
67,437 notisa relations ➠ relatively ‘simple’ (→ baseline) ➠ still large coverage
COMPANIES LISTED ON NASDAQ MICROSOFT MICROSOFT (page)
isa same lexical head category titled as page
COMPUTER AND VIDEO GAME COMPANIES
[instance−of]
(Wikipedia pages) and classes (categories) can be labeled as instance-of (Suchanek et al., 2007)
➠ identifies 14,886 isa relations
CARBAMATES AMIDES ETHYL CARBAMATE (page)
isa redundant link [instance−of] [instance−of]
relations, e.g. ETHYL CARBAMATE
➠ identifies 16,523 isa relations
we are left with 147,929 unclassified relations . . .
STIMULANT CAFFEINE
isa NP2,? (such as|like|, especially) NP* NP1 pattern match:
STIMULANTS SUCH AS CAFFEINE
corpora to identify isa relations (Hearst, 1992; Caraballo, 1999)
(Berland & Charniak, 1999) indicate that the relation is not an isa relation ➠ notisa
➠ NP2,? (such as|like|, especially) NP* NP1
a stimulant such as caffeine
➠ NP1 NP* (and|or|,like) other NP2
caffeine and other stimulants
➠ NP2’s NP1
car’s engine
➠ NP2 with NP1
a car with an engine
Wikipedia itself (8× 108 words)
NP-chunking ➠ 15GB data
filter out 3,226 previously identified isa relations
transitive relation
isa
(previously found)
isa
(inferred) BORGHESE PAPAL FAMILIES ITALIAN NOBLE FAMILIES
same lexical head
lemma matches the head lemma of a previously identified isa supercategory
➠ propagate the isa relation to the sisters of the previously
identified isa supercategories
Wikipedia Wikipedia ResearchCyc WordNet (sem. network) (taxonomy)
# concepts 300,000 # nodes # synsets 117,659 # categories 337,522 209,919
# assertions 3,000,000 # edges # semantic pointers 285,348 # category links 743,140 335,128
(results on Wikipedia from March 2008, journal submission)
M&C R&G Lin (1998) .82
−
Li et al. (2006) .89
−
M&C R&G Bollegala et al. (2007) .81
−
M&C R&G ESA .73 .82 WLM .70 .64 WikiRelate! isa .87 .78
1.106 instances evaluated manually by three judges R P F random baseline 51.1 51.6 51.3 syntax (1-3) 17.0 95.4 28.9 connectivity (1-4, 6) 38.9 88.1 54.0 pattern-based (1-3, 5-6) 62.7 84.3 71.9 all (1-6) 69.5 81.6 75.0