際際滷

際際滷Share a Scribd company logo

Mining Historical Data for DBpedia via Temporal Tagging of Wikipedia Infoboxes

Download as PPT, PDF
Volha Bryl
Volha Bryl

Norman Weisenburger, Volha Bryl, Simone Paolo Ponzetto NLP & DBpedia 2014 Workshop @ ISWC 2014, Riva del Garda, Italy, October 20, 2014

1 of 27
Download to read offline
Mining Historical Data for DBpedia via Temporal Tagging of Wikipedia Infoboxes Norman Weisenburger, Volha Bryl, Simone Paolo Ponzetto Data and Web Science Research Group University of Mannheim Germany NLP & DBpedia @ ISWC, Riva del Garda, Italy, October 20, 2014
Outline 1. State of art: Temporally annotated data in DBpedia and LOD 2. Temporally annotated data extraction pipeline 3. Company Dataset Statistics Comparison with other KBs 1. Ongoing and future work Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 2
Why we need historical LOD Historical data == any data that is or can be temporally annotated population of a city, revenue of a company, current club for a football player Why we need such data Allows having a more precise description of an entity Enables LOD-based data mining for trend prediction Availability of temporally annotated data on the Web of Data Poor and scarce Examples can be found in Freebase, Wikidata, YAGO, Temporally annotated facts or not so frequently time series Some exceptionally good examples follow Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 3
Temporally annotated data: Examples Apple Inc. in Wikidata http.//www.wikidata.org/wiki/Q312 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 4
Temporally annotated data: Examples Apple Inc. in Freebase http.//www.freebase.com/m/0k8z Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 5
Temporally annotated data in DBpedia DBpedia's main source of knowledge are Wikipedia infoboxes Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Often only the latest value is present When new value is available, the old one is overwritten Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 6
Temporally annotated data in DBpedia DBpedia's main source of knowledge are Wikipedia infoboxes Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Often only the latest value is present When new value is available, the old one is overwritten Our focus: case 3, temporal annotation is a part of an attribute value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 7
Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated (1) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 8
Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Often lost during DBpedia data extraction E.g. no connection between populationTotal and populationAsOf properties (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 9
Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Ends up in DBpedia only if an intermediate node mapping is defined in the mapping wiki (2) (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 10
Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Ends up in DBpedia only if an intermediate node mapping is defined in the mapping wiki (2) (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 11
Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Annotation is lost during extraction In most cases value is regularly overwritten Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 12
Idea: go back in time Properties of interest Temporally annotated, annotation is a part of an attribute value Use case: Business and Financial Data (Companies) Key observations Attribute values are often temporally annotated If annotation is part of attribute value DBpedia extraction framework ignores it Attribute values are regularly overwritten by Wikipedia editors, but the trace remains in Wikipedia revision history DBpedia data extraction process is run on one (e.g. the latest) dump only Proposed solution Run extraction on (part of) revision history Add a temporal tagger to the process Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 14
Extraction pipeline 1. Select and download Wikipedia revisions 2. Extract temporal facts 3. Merge facts Code available at https.//github.com/normalerweise/mte Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 15
Extraction pipeline 1. Select and download Wikipedia revisions Select 4 revisions per year (1st, 2nd, 3rd quartile and the last revision) Use MediaWiki API to download the revisions Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 16
Extraction pipeline 2. Extract temporal facts Parse each infobox attribute twice For a value: Mapping Extractor of the DBpedia Extraction Framework For time validity (point or interval): HeidelTime HeidelTime is a multilingual cross-domain rule-based temporal tagger Developed at the University Of Heidelberg http.//dbs.ifi.uni-heidelberg.de/index.php?id=129 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 17
Extraction pipeline {{ Infobox company | name = Netflix, Inc. | revenue = US$4.37&nbsp; million (''FY 2013'') ... <Netflix, revenue, 4.37E9, usDollar, 2013, 610604061> 2. Extract temporal facts Parse each infobox attribute twice Revision ID For a value: Mapping Extractor of the DBpedia Extraction Framework For time validity (point or interval): HeidelTime HeidelTime is a multilingual cross-domain rule-based temporal tagger Developed at the University Of Heidelberg http.//dbs.ifi.uni-heidelberg.de/index.php?id=129 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 18
Extraction pipeline 3. Merge facts Group triples by subject, property, temporal validity, value In case of value conflicts, select the most frequent value In case of ties, select the most recent value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 19
Extraction pipeline <Netflix, operatingIncome, 1.28E8, usDollar, 2008, 352001234> <Netflix, operatingIncome, 1.92E8, usDollar, 2009, 387048342> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 439282478> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 426138580> <Netflix, operatingIncome, 1.28E8, usDollar, 2008, 352001234> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 439282478> 3. Merge facts Group triples by subject, property, temporal validity, value In case of value conflicts, select the most frequent value In case of ties, select the most recent value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 20
Data model Our choice for RDF representation Singleton property approach Vinh Nguyen, Olivier Bodenreider, and Amit Sheth. Dont like RDF reification? Making statements about statements using singleton property, WWW 2014 Motivation: performance in terms of #triples, query size and execution time Main idea: unique URI for each predicate instance <Netflix, revenue#uniqueId, 4.37E9> <revenue#uniqueId, singletonPropertyOf, revenue> <revenue#uniqueId, date, 2013> <revenue#uniqueId, sourceRevision, 610604061> Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 21
Company dataset Dataset available at http://tiny.cc/tmpcompany Started from DBpedia resources of type dbpedia-owl:Company and yago:Company108058098 51,214 companies, for 18,489 at least one fact is extracted for assets equity netIncome numberOfEmployees operatingIncome revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 22
Company dataset Dataset available at http://tiny.cc/tmpcompany 51,214 companies, for 18,489 at least one fact is extracted for assets, equity, netIncome, numberOfEmployees, operatingIncome, revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 23
Company dataset Dataset available at http://tiny.cc/tmpcompany 51,214 companies, for 18,489 at least one fact is extracted for assets, equity, netIncome, numberOfEmployees, operatingIncome, revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 24
Company dataset vs other KBs 10 random companies with well-maintained infoboxes Manually mapped ontology properties YAGO2 0 triples for these companies for hasNumberOfPeople and hasRevenue Our dataset Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 25
Company dataset vs other KBs 10 random companies with well-maintained infoboxes Manually mapped ontology properties YAGO2 0 triples for these companies for hasNumberOfPeople and hasRevenue Freebase 201 vs 58 triples Our dataset Freebase Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 26
Evaluation Evaluating the precision (preliminary, not in the paper) 100 random tuples, 2 properties, so far only one annotator 75% for numberOfEmployees and 78% for revenue Caused by parsing errors: DBpedia extraction framework is always tuned to work with the latest Wikipedia version After fixing some errors: 97% for numberOfEmployees and 92% for revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 27
Ongoing and future work Ongoing: extracting missing attributes from Wikipedia article texts Company dataset is used for distant supervision Anticipating some questions Yes, we tried the approach for another domain: American football Yes, making the data available through an endpoint is on our todo list Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 28

More Related Content

Mining Historical Data for DBpedia via Temporal Tagging of Wikipedia Infoboxes

  • 1. Mining Historical Data for DBpedia via Temporal Tagging of Wikipedia Infoboxes Norman Weisenburger, Volha Bryl, Simone Paolo Ponzetto Data and Web Science Research Group University of Mannheim Germany NLP & DBpedia @ ISWC, Riva del Garda, Italy, October 20, 2014
  • 2. Outline 1. State of art: Temporally annotated data in DBpedia and LOD 2. Temporally annotated data extraction pipeline 3. Company Dataset Statistics Comparison with other KBs 1. Ongoing and future work Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 2
  • 3. Why we need historical LOD Historical data == any data that is or can be temporally annotated population of a city, revenue of a company, current club for a football player Why we need such data Allows having a more precise description of an entity Enables LOD-based data mining for trend prediction Availability of temporally annotated data on the Web of Data Poor and scarce Examples can be found in Freebase, Wikidata, YAGO, Temporally annotated facts or not so frequently time series Some exceptionally good examples follow Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 3
  • 4. Temporally annotated data: Examples Apple Inc. in Wikidata http.//www.wikidata.org/wiki/Q312 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 4
  • 5. Temporally annotated data: Examples Apple Inc. in Freebase http.//www.freebase.com/m/0k8z Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 5
  • 6. Temporally annotated data in DBpedia DBpedia's main source of knowledge are Wikipedia infoboxes Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Often only the latest value is present When new value is available, the old one is overwritten Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 6
  • 7. Temporally annotated data in DBpedia DBpedia's main source of knowledge are Wikipedia infoboxes Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Often only the latest value is present When new value is available, the old one is overwritten Our focus: case 3, temporal annotation is a part of an attribute value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 7
  • 8. Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated (1) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 8
  • 9. Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Often lost during DBpedia data extraction E.g. no connection between populationTotal and populationAsOf properties (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 9
  • 10. Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Ends up in DBpedia only if an intermediate node mapping is defined in the mapping wiki (2) (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 10
  • 11. Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute Ends up in DBpedia only if an intermediate node mapping is defined in the mapping wiki (2) (2) Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 11
  • 12. Temporally annotated data in DBpedia Temporal (time-dependent) infobox attributes 1. Not at all temporally annotated 2. Temporally annotated, annotation is modeled as a separate attribute 3. Temporally annotated, annotation is a part of an attribute value Annotation is lost during extraction In most cases value is regularly overwritten Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 12
  • 13. Idea: go back in time Properties of interest Temporally annotated, annotation is a part of an attribute value Use case: Business and Financial Data (Companies) Key observations Attribute values are often temporally annotated If annotation is part of attribute value DBpedia extraction framework ignores it Attribute values are regularly overwritten by Wikipedia editors, but the trace remains in Wikipedia revision history DBpedia data extraction process is run on one (e.g. the latest) dump only Proposed solution Run extraction on (part of) revision history Add a temporal tagger to the process Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 14
  • 14. Extraction pipeline 1. Select and download Wikipedia revisions 2. Extract temporal facts 3. Merge facts Code available at https.//github.com/normalerweise/mte Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 15
  • 15. Extraction pipeline 1. Select and download Wikipedia revisions Select 4 revisions per year (1st, 2nd, 3rd quartile and the last revision) Use MediaWiki API to download the revisions Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 16
  • 16. Extraction pipeline 2. Extract temporal facts Parse each infobox attribute twice For a value: Mapping Extractor of the DBpedia Extraction Framework For time validity (point or interval): HeidelTime HeidelTime is a multilingual cross-domain rule-based temporal tagger Developed at the University Of Heidelberg http.//dbs.ifi.uni-heidelberg.de/index.php?id=129 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 17
  • 17. Extraction pipeline {{ Infobox company | name = Netflix, Inc. | revenue = US$4.37&nbsp; million (''FY 2013'') ... <Netflix, revenue, 4.37E9, usDollar, 2013, 610604061> 2. Extract temporal facts Parse each infobox attribute twice Revision ID For a value: Mapping Extractor of the DBpedia Extraction Framework For time validity (point or interval): HeidelTime HeidelTime is a multilingual cross-domain rule-based temporal tagger Developed at the University Of Heidelberg http.//dbs.ifi.uni-heidelberg.de/index.php?id=129 Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 18
  • 18. Extraction pipeline 3. Merge facts Group triples by subject, property, temporal validity, value In case of value conflicts, select the most frequent value In case of ties, select the most recent value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 19
  • 19. Extraction pipeline <Netflix, operatingIncome, 1.28E8, usDollar, 2008, 352001234> <Netflix, operatingIncome, 1.92E8, usDollar, 2009, 387048342> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 439282478> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 426138580> <Netflix, operatingIncome, 1.28E8, usDollar, 2008, 352001234> <Netflix, operatingIncome, 1.94E8, usDollar, 2009, 439282478> 3. Merge facts Group triples by subject, property, temporal validity, value In case of value conflicts, select the most frequent value In case of ties, select the most recent value Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 20
  • 20. Data model Our choice for RDF representation Singleton property approach Vinh Nguyen, Olivier Bodenreider, and Amit Sheth. Dont like RDF reification? Making statements about statements using singleton property, WWW 2014 Motivation: performance in terms of #triples, query size and execution time Main idea: unique URI for each predicate instance <Netflix, revenue#uniqueId, 4.37E9> <revenue#uniqueId, singletonPropertyOf, revenue> <revenue#uniqueId, date, 2013> <revenue#uniqueId, sourceRevision, 610604061> Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 21
  • 21. Company dataset Dataset available at http://tiny.cc/tmpcompany Started from DBpedia resources of type dbpedia-owl:Company and yago:Company108058098 51,214 companies, for 18,489 at least one fact is extracted for assets equity netIncome numberOfEmployees operatingIncome revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 22
  • 22. Company dataset Dataset available at http://tiny.cc/tmpcompany 51,214 companies, for 18,489 at least one fact is extracted for assets, equity, netIncome, numberOfEmployees, operatingIncome, revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 23
  • 23. Company dataset Dataset available at http://tiny.cc/tmpcompany 51,214 companies, for 18,489 at least one fact is extracted for assets, equity, netIncome, numberOfEmployees, operatingIncome, revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 24
  • 24. Company dataset vs other KBs 10 random companies with well-maintained infoboxes Manually mapped ontology properties YAGO2 0 triples for these companies for hasNumberOfPeople and hasRevenue Our dataset Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 25
  • 25. Company dataset vs other KBs 10 random companies with well-maintained infoboxes Manually mapped ontology properties YAGO2 0 triples for these companies for hasNumberOfPeople and hasRevenue Freebase 201 vs 58 triples Our dataset Freebase Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 26
  • 26. Evaluation Evaluating the precision (preliminary, not in the paper) 100 random tuples, 2 properties, so far only one annotator 75% for numberOfEmployees and 78% for revenue Caused by parsing errors: DBpedia extraction framework is always tuned to work with the latest Wikipedia version After fixing some errors: 97% for numberOfEmployees and 92% for revenue Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 27
  • 27. Ongoing and future work Ongoing: extracting missing attributes from Wikipedia article texts Company dataset is used for distant supervision Anticipating some questions Yes, we tried the approach for another domain: American football Yes, making the data available through an endpoint is on our todo list Mining Historical Data for DBpedia, Weisenburger, Bryl, Ponzetto 28

Editor's Notes

  • #15: Only the latest value is present in the infobox of interest?
  • #26: Comparison with DBpedia: http://dbpedia.org/resource/Apple Inc. Our dataset contains 45 temporal facts whereas DBpedia currently has one fact for reach relation, i.e. 6 triples
  • #27: Freebase lists EDGAR as one of its data sources. EDGAR is a database which contains information about publicly traded US companies operated by the United States Security and Exchange Commission.