1. Judith Klavans 1, Jen Golbeck 1, Susan Chun 2, Rob Stein 3, Ed Bachta 3, Irene Eleta 1, Raul Guerra 1, Rebecca LaPlante 1 University of Maryland 1 Independent Museum Consultant 2 Indianapolis Museum of Art 3 http://umiacs.umd.edu/research/t3/ Art Images Online: Leveraging Social Tagging and Language for Browsing

2. High Level Goals Images in Museums and Libraries Words…words… words – Traditional cataloging – Handbook and other descriptive text – Social tagging 2

3. Record from American Institute for College Teaching Minimal metadata for image, no descriptive terms. 3

4. Nefertiti Gardner (v. 11, pl. 3-33) The famous painted limestone bust of Akhenaton’s queen, Nefertiti (fig. 3-33), exhibits a similar expression of entranced musing and an almost mannered sensitivity and delicacy of curving contour. The piece was found in the workshop of the queen’s official sculptor, Thutmose, and is a deliberately unfinished model very likely by the master’s own hand. The left eye socket still lacks the inlaid eyeball, making the portrait a kind of before-and-after demonstration piece. With this elegant bust, Thutmose may have been alluding to a heavy flower on its slender stalk by exaggerating the weight of the crowned head and the length of the almost serpentine neck… 4

5. Excerpt of descriptive text from Gardner (v. 11, pl. 3- 33), suggested CLiMB terms highlighted in yellow The famous painted limestone bust of Akhenaton’s queen, Nefertiti (fig. 3-33), exhibits a similar expression of entranced musing and an almost mannered sensitivity and delicacy of curving contour. The piece was found in the workshop of the queen’s official sculptor, Thutmose, and is a deliberately unfinished model very likely by the master’s own hand. The left eye socket still lacks the inlaid eyeball, making the portrait a kind of before-and-after demonstration piece. With this elegant bust, Thutmose may have been alluding to a heavy flower on its slender stalk by exaggerating the weight of the crowned head and the length of the almost serpentine neck… 5

6. User Tags Woman Hat One blind eye Beautiful features Long neck Beautiful woman Blue Yellow Blue hat Graceful eyes Ceramic statue of an elegant woman Half an ear I’d like this in my living room 6

7. Many kinds of Words….. Terms informed by art historical criteria: – deliberately unfinished model Ability to find related images – elongated neck – bust Potential for using thesaural resources – painted limestone bust 7

8. CLiMB: Computational Linguistics for Metadata Building Columbia University University of Maryland – UMIACS and Computational Linguistics and Information Processing (CLIP) Lab

9. 9

10. STEVE.MUSEUM 18 museum partners Over 90,000 tags Nearly 1800 images tagged But tags are “unruly” and “chaotic” 10

11. T3: Tags, Terms, and Trust What computational linguistic techniques can be used to bring all these words to use? What is the nature of the tags? What tools can be helpful to the museums and library user communities? How does tagging in different languages compare? 11

12. Judith Klavans Jen Golbeck Irene Eleta Raul Guerra Rebecca LaPlante Museum Partners Rob Stein Ed Bachta Susan Chun … and 18 museums T3 Research Group 12

13. Funding Mellon Foundation – CLiMB-1 (Columbia Univ) – CLiMB-2 (Univ of Maryland, UMIACS-CLIP Lab) IMLS – – Steve.museum (Indianapolis Museum of Art) – T3 – IMA and University of Maryland National Science Foundation 13

14. Major Contribution – Creating Order over Chaos Developed techniques for processing social tags – What tags are related to other tags? – How are tags related (or not)? – What is the impact of cultural and language differences in tagging? Completed user analysis of tagging behavior Explored the value of tags compared with descriptive text 14

15. Highlight Two Specific Areas The value and use of – Computational linguistic analysis for tags – Multilingual social tags 15

16. Computational Processing Pipeline 16 Woman Hat One blind eye Beautiful features Long neck Beautiful woman Blue Yellow Blue hat Graceful eyes Ceramic statue of an elegant woman Half an ear Graceful - [Adjective] {graceful} Eyes – [Noun-Plural] {eye} One - [Number] Blind – [Adjective] {blind} Eye –[Noun-Singular] {eye} Word Count Woman – 3 Eye - 2

17. What is a tag ? 17 Important to treat tags related to the same topic as related, e.g. Line, line, lines, lining How many “tags”? Four, Three, or one? Stemmer leav Leaves Puppies puppi babi Babies Lemmatizer Leaves leave Puppies Babies puppy baby

18. What is a tag ? Comparing Stemmers and Lemmatizers Stemmers: reduce words to roots. – Advantage: fast with big data – Disadvantage: stems are not necessarily words Lemmatizers: reduce words to ‘dictionary’ form. – Advantage: lemmas are more useful – Disadvantage: slower because they rely on external resources like Wordnet 18

19. How do we do? Checking Lemmatizer performance against Lemma Gold Standard of 850 user-tags: Default configuration of the pipeline – 64% accuracy on all tags – 68% accuracy on all the correctly spelled tags Fine grained configuration of the pipeline – 76.47% accuracy on all tags – 81.35% accuracy on all the correctly spelled tags 19

20. Part of Speech Labeling “graceful eyes” - [[Adjective] [Noun]] “blue” – [Adjective] “face” – [Noun], [Verb] 20

21. WORD LEVEL 21

22. Harder than annotating text because of lack of context. Stanford Tagger has accuracy of 97.28% correct on WSJ 19-21 (90.46% correct on unknown words) 22 Part-of-Speech Labeling

23. The famous painted limestone bust of Akhenaton’s queen, Nefertiti (fig. 3-33), exhibits a similar expression of entranced musing and an almost mannered sensitivity and delicacy of curving contour. The piece was found in the workshop of the queen’s official sculptor, Thutmose, and is a deliberately unfinished model very likely by the master’s own hand. The left eye socket still lacks the inlaid eyeball, making the portrait a kind of before-and-after demonstration piece. With this elegant bust, Thutmose may have been alluding to a heavy flower on its slender stalk by exaggerating the weight of the crowned head and the length of the almost serpentine neck… 23

24. Handbook Text Chunking 24 As a result of the way users tag images, – FEW images have LOTS of tags – LOTS of images have FEW tags Can we add tag-like phrases from text – Two types of phrases Names of things (Named Entities) Nouns and all the words that go with it ( Noun Phrases )

25. Results Precision for Named Entities is 46.39% for Full Match and 69.24% for partial matches. – Columbia University in the City of New York – Columbia University Precision for Noun Phrases is 79.95% for Full Match and 93.03% for Partial matches. – Famous painted limestone bust – Limestone bust 25

26. Conclusions Stemmers/Lematizers are the first step for removing variation from the tags. Part of Speech Tagging of social tags useful for getting a better understanding of the tags Handbook Text Analysis helps to add tags for poorly annotated images. The pipeline enables to add information to the user tags that is useful for applications using them. 26

27. Multilingual Social Tagging of Art Images Cultural Bridges and Diversity La Orana Maria, by Gauguin (Metropolitan Museum of Art) Polynesian women baby jungle Polinesia Virgen roja Jesús

28. Questions How similar are the tags provided by different language communities when tagging art images? Does it depend on the type of painting? Do tags reflect cultural differences in the interpretations of the paintings? 28

29. Contributions of this Study Expands T 3 to include Spanish Identifies tags that could be used for multilingual search Identifies tags that could be used for cross-cultural discovery and understanding Proposes the separation of tagging environments by language 29

30. 773 English tags 566 Spanish tags 33 paintings Spanish participantsAmerican participants 5 questions 10 images IN NUMBERS… Genesis First Version By Lorser Feitelson (San Francisco MOMA) Data Collection 30

31. Tag tokenization and normalization Creation of lexical correspondences Processing and Comparing Tags 31 The Cotton Pickers by Winslow Homer (Los Angeles Contemporary Museum of Art)

32. Semantic Analysis 32 LaPlante, R., Klavans, J., Golbeck, J (under review). Subject Matter Categorization of Tags Applied to Digital Images from Art Museums.

33. Results Many exact translations The Cotton Pickers by Winslow Homer (LACMA) 33

34. Translation Pairs by Type of Painting 34

35. Differences The Cotton Pickers by Winslow Homer (LACMA) 35

36. Translation pairs > multilingual search –“general person or thing” in realistic paintings –“visual elements” in abstract paintings Different perspectives > design for discovery –“emotions and abstract ideas” 36

37. Publications Computational techniques to examine tags and phrases of importance for browsing on art image collections. Klavans, Judith, Guerra, Raul, LaPlante, Rebecca, Stein, Rob & Bachta, Ed (2011). Beyond Flickr: Not all image tagging is created equal. In AAAI 2011 Workshop: Language-Action Tools for Cognitive Artificial Agents, San Francisco, CA. Klavans, Judith, Stein, Rob, Chun, Susan, & Guerra, Raul (2011). Computational linguistics in museums: Applications for cultural datasets. Museums and the Web 2011, Philadelphia, PA. 37

38. Publications Comparing social tagging patterns in two languages to inform design for multilingual access to art images. Eleta, Irene and Jennifer Golbeck (2012, to appear) A Study of Multilingual Social Tagging of Art Images: Cultural Bridges and Diversity. 2012 ACM Conference on Computer Supported Cooperative Work (CSCW 2012), Seattle, Washington. 38

39. Thank you! Publications: http://umiacs.umd.edu/research/t3/ Tools available through Steve in Action: http://www.steve.museum/ 39