A Case Study in Pragmatism: exploring the practical failure modes of Linked Data as applied to classical music catalogues

For the last decade efforts have been made in building the Linked Data enabled Semantic Web. Some of the earliest efforts in publishing Linked Data included music-specific corpora, modeled using e.g. the Music Ontology. Given that, one might assume creating a Linked Data client in the music domain would be a reasonably straight-forward process. The reality of the situation is not so simple. In this paper, we present a case study in building a real-world web-service that makes use of third-party Linked Data resources: libmus. We describe the variety of difficulties in reliably linking to several external sources. These range from poor data coverage, to serialization consistency, to sporadic availability. We outline the ways in which we have coped with these problems within the libmus system. These include: local caching of third-party data to enable search, screen-scraping to expand data availability, and automatically generated regular expressions to emulate fuzzy search. Finally, we will consider how the Linked Data ecosystem can be improved to mitigate these problems in other systems and what this means for the future of the Linked Data Web.

Transcript

1. A Case Study in Pragmatism: exploring the practical failure modes of Linked Data as applied to classical music catalogues Ben Fields !/alsothings Sam Phippen !/samphippen Brad Cohen
2. Introduction
4. Where is the friction? What is the distance between specification and realisation?
5. Background
6. Background: Goals and design of libmus
7. Mapping links between agents and works
8. Mapping these concepts across disparate repositories
12. Background: Linked Data ecosystem reliability
13. We care about two kinds of reliability
14. Reliable modelling: resolvable and accessible
15. Basically
16. Use obvious self-documenting class and attribute names
17. Reliable repositories: availability :
18. Uptime: time service is reachable
19. Downtime: time service is unreachable
20. Availability: percentage of uptime in given length of time
21. Example!
22. 10 minutes of downtime in 1 week
23. 10800 - 10 10800 x 100 = 99.9%
24. SPARQL Endpoint Current Availabiltiy 0 25 50 75 100 up down SPARQL availability index live at http://sparqles.ai.wu.ac.at/availability this data was fetched on 26 April 2015
26. Linking Data: Theory vs. Reality
27. Theory vs Reality: IMSLP
31. No search endpoint
32. So we downloaded metadata for: 16,716 composers 88,464 works
33. SELECT imslp_id, permalink FROM imslp_composers WHERE imslp_id LIKE <composer_string>
34. search: done ✅
35. search: done ✅ (As long as there aren’t too many items)
36. Theory vs Reality: VIAF
37. No SPARQL endpoint
38. Let’s have some questions
39. Let’s have some questions
40. Let’s have some questions
41. Let’s have some questions
42. Let’s have some questions
43. Let’s have some questions
44. Let’s have some questions
45. Let’s have some questions
46. Theory vs Reality: LinkedBrainz
47. Really good triples!
48. Functional Sparql endpoint
49. Let’s have some questions
50. Let’s have some questions
51. Works notions!
52. Let’s have some questions
53. Let’s have some questions
54. Let’s have some questions
55. Sonata BEETHO QUAR O 18 N 3 VARIOU ALTE TSCH ORGE HANDEL SUIT 5 AIR + DAN
56. How to align pathological data?
57. Let’s have some questions
58. HANDEL SUIT 5 AIR + DAN
59. (HANDEL (SUIT)? (5)? (AIR)? (\\+)? (DAN)?|HANDEL SUIT (5)? (AIR)? (\\+)? (DAN)?|HANDEL 5 (AIR)? (\\+)? (DAN)?|HANDEL AIR (\\+)? (DAN)?|HANDEL \\+ (DAN)?|HANDEL DAN | SUIT (5)? (AIR)? (\\+)? (DAN)?| 5 (AIR)? (\\+)? (DAN)?| AIR (\\+)? (DAN)?| \\+ (DAN)?| DAN |SUIT 5 (AIR)? (\\+)? (DAN)?|SUIT AIR (\\+)? (DAN)?|SUIT \\+ (DAN)?|SUIT DAN |5 AIR (\\+)? (DAN)?|5 \\+ (DAN)?|5 DAN |AIR \\+ (DAN)?|AIR DAN |\\+ DAN )
60. 393 character regex 35 optional groups 20 choices
61. Performance nightmare
62. Let’s have some questions
64. Let’s have some questions
65. We then filter by Levenshtein and threshold
66. Conclusions
67. Doing Linked Data is harder than it should be
68. When publishing data consider:
69. the obviousness of your model,
70. search and access,
71. availability and reliability of endpoints.
72. Future Work
73. What course of action will help nurture the Linked Data ecosystem?
74. Make likely paths to specific data items as obvious as possible
75. SPARQL endpoint reliability
76. Publish your data!
77. Let’s have some questions Ben Fields !/alsothings Sam Phippen !/samphippen Brad Cohen