Knowledge Acquisition from Music Digital Libraries

Knolwedge Acquisi0on from Music Digital Libraries

Sergio Oramas, Mohamed Sordo

IAML/IMS Congress 2015

•  Musicological Knowledge is hidden between the lines •  Machines don’t know how to read

Mo0va0on

•  Obtain knowledge automa0cally •  Make complex ques0ons •  Visualize the informa0on •  Improve naviga0on •  Share knowledge

Why Knowledge Acquisi0on?

4

Musical Libraries

Musical Libraries

digital recording, scan

Musical Libraries


OCR, manual transcrip0on

Musical Libraries



Informa0on Extrac0on, seman0c annota0on

Musical Libraries

Current DL




Musical Libraries

Current DL

Web search




Musical Libraries




Current DL

Web search

•  The Seman&c Web aims at conver0ng the current web, dominated by unstructured and semi-‐structured documents into a web of linked data.

•  Achievements useful for Digital Libraries –  Common framework for data representa0on and interconnec0on (RDF, ontologies)

–  Seman0c technologies to annotate texts (En0ty Linking) –  Language for complex queries (SPARQL)

Seman0c Web

Wikipedia and DBpedia

13

-‐  Digital Encyclopedia -‐  Unstructured -‐  Keyword search

-‐  Knowledge Base -‐  Structured -‐  Query search

14

15

•  Dbpedia example queries –  Composers born in Vienna in XVIII Century –  American jazz musicians that have wri_en songs recorded by RCA Records

•  Dbpedia graph applica0ons –  En0ty Relevance –  En0ty Similarity –  En0ty Recommenda0on

DBpedia

16

Google Knowledge Graph

17

•  Current naviga0on –  Indexes –  Keyword-‐based search

•  From searchable repositories to knowledge environments –  Complex queries –  Informa0on visualiza0on –  Recommenda0on –  Data Analy0cs –  Q&A

Music Digital Libraries


19


20

•  Encyclopedic dic0onary •  One of the largest reference works in Western music •  Ar0st biographies crawled from the Grove Music Online

–  16,707 biographies (1st paragrahph) –  From pre-‐medieval to contemporary

Dataset: The New Grove

•  What are the most relevant music schools? •  What are the ar0sts most similar to Schoenberg? •  Which are the most represented roles in the Grove? •  Is there a migra0on tendency in ar0sts? To which ci0es? •  What is the best city to die for a musician?


22


23

Anton Webern (b Vienna, 3 Dec 1883; d Mi_ersill,15 Sept 1945). Austrian composer and conductor. Webern, who was probably Schoenberg's first private pupil, and Alban Berg, who came to him a few weeks later […] Boulez and Stockhausen and other integral serialists of the Darmstadt School […]

Informa0on Extrac0on

24



25



26



27


Informa0on Extrac0on: En0ty Linking

28

Domain Knowledge Base

•  Webern, who was probably Schoenberg's first private pupil

Informa0on Extrac0on: Rela0on Extrac0on

29

•  Webern, who was probably Schoenberg's first private pupil

Informa0on Extrac0on: Rela0on Extrac0on

30

Webern Schoenberg pupil_of

Methodology

31

Knowledge Graph

32

•  16,707 biographies •  434 roles •  Graph

–  47,367 nodes –  274,333 edges

Knowledge Graph: Data Analy0cs

33

Role Amount composer 2618 teacher 1065 conductor 968 pianist 704 organist 676 singer 404 violinist 285 … musicologist 144 cri0c 133

34

Birth Year

Death Year

35

Birth Year

Death Year

Knowledge Graph: Data Analy0cs

36

Country Births Deaths Difference United States 2317 2094 -‐10% Italy 1616 1279 -‐21% Germany 1270 1292 2% France 991 1058 7% United Kingdom 882 877 -‐1%

City Births Deaths Difference London 322 507 57% Paris 304 720 137% New York 266 501 88% Vienna 177 292 65% Rome 159 256 61%

•  City –  Paris, London, Vienna, Rome, Venice, Berlin, Paris, New York

•  Venue –  Covent Garden Theatre, King's Theatre, Drury Lane, Carnegie Hall, Théâtre de la Monnaie, Stad_heater, Theatre Royal, …

•  Educa0onal Ins0tu0on –  Paris Conservatoire, Moscow Conservatory, Juilliard School, St Petersburg Conservatory, Bmus, Prague Conservatory, Leipzig Conservatory, Vienna Hochschule für Musik, ...

Knowledge Graph: En0ty Relevance

37

•  Biography subject –  Haydn, Claude Debussy, Arnold Schoenberg, Robert Stevenson, Paul Hindemith, Giovanni Pierluigi da Palestrina, Gustav Mahler, Maurice Ravel, Jean-‐Philippe Rameau

–  Mozart?, Bach?, Wagner? •  Genre

–  chamber music, cappella, jazz, folk music, avant garde, baroque music, electronic music, musical theatre, plainchant

Knowledge Graph: En0ty Relevance

38

•  PageRank algorithm and Maximal Common Subgraph •  Arnold Schoenberg: Anton Webern, Paul Hindemith, Gustav

Mahler, Alban Berg, Claude Debussy •  Guido Adler: Heinrich Jalowetz, Eusebius Mandyczewski,

Robert Fuchs, Karl Weigl, Anton Wranitzky •  Manuel de Falla: Ricardo Viñes, Juan Vicente Lecuna, Enrique

Granados, Miguel Llobet Soles, Luigi Russolo •  Miles Davis: Dizzy Gillespie, Herbie Hancock, Paul Chambers,

Tony Williams, Cannonball Adderley

Knowledge Graph: En0ty Similarity

39

Rela0ons Graph: Informa0on Visualiza0on

40

Rela0ons Graph: Informa0on Visualiza0on

41

•  Crea0on and cura0on of a library from Web content •  FlaBase: Flamenco Knowledge Base

Mixing Different Data Sources

42

•  Data Acquisi0on –  APIs –  Web crawling –  SPARQL endpoints

•  En0ty Resolu0on –  String similarity between labels –  Graph similarity (context informa0on)

Mixing Different Data Sources

43

FlaBase: Flamenco Knowledge Base

44

•  Data gathered –  1,174 Ar0sts (text biography) –  76 Palos (flamenco genres) –  2,913 Albums –  14,078 Tracks –  771 Andalusian loca0ons

•  Knowledge Extracted –  Place of birth –  Date of birth –  En0ty men0ons in text

FlaBase: Flamenco Knowledge Base

45

•  Number of ar0sts by year of birth

FlaBase: Data Analy0cs

46

FlaBase: Data Analy0cs

47

•  Flamenco expert evalua0on

FlaBase: Ar0st Relevance

48

Precision Values

•  Music Digital Libraries can benefit from seman0c approaches •  Music Digital Libraries are s0ll in an early stage of

development compared to the Web (Linked Open Data, Google Knowledge Graph)

•  Knowledge acquisi0on from Digital Libraries can help musicologists not only to search content, but also to discover new knowledge

Conclusions

49

•  This work was partly funded by the COFLA2 research project (Proyectos de Excelencia de la Junta de Andalucía, FEDER P12-‐TIC-‐1362).

Aknowledgments

50

•  Oramas S., Sordo M., Espinosa-‐Anke L., Serra X. (2015). A Seman,c-‐based approach for Ar,st Similarity. Interna0onal Society for Music Informa0on Retrieval Conference ISMIR 2015. In Press.

•  Oramas S., Gomez F., Gomez E., Mora J. (2015). FlaBase: Towards the crea,on of a Flamenco Music Knowledge Base. Interna0onal Society for Music Informa0on Retrieval Conference ISMIR 2015. In Press.

•  Sordo, M., Oramas S., & Espinosa-‐Anke L. (2015). Extrac,ng Rela,ons from Unstructured Text Sources for Music Recommenda,on. Interna0onal Conference on Applica0ons of Natural Language to Informa0on Systems NLDB 2015.

•  Oramas S., Sordo M., Espinosa-‐Anke L. (2015). A Rule-‐based Approach to Extrac,ng Rela,ons from Music Tidbits. 2nd Workshop on Knowledge Extrac0on from Text at WWW 2015.

•  Oramas S., Sordo M., Serra X. (2014). Automa,c Crea,on of Knowledge Graphs from Digital Musical Document Libraries. Conference in Interdisciplinary Musicology CIM 2014.

Bibliography

51

Knolwedge Acquisi0on from Music Digital Libraries

Sergio Oramas, Mohamed Sordo

[email protected]

@sergiooramas

Thanks!

Data & Analytics

Knowledge Acquisition from Music Digital Libraries