Workset Creation for Scholarly Analysis Project presentation at CNI 2013
- 1. HathiTrust Research Center: Improving Scholarly Inquiry Timothy W. Cole (t-cole3@illinois.edu) Harriett Green (green19@illinois.edu) With slides and other contributions from Stephen Downie, Beth Plale, Colleen Fallaw, Megan Senseney, Katrina Fenlon, et al. CNI Fall 2013 Membership Meeting Washington, D.C. 9 December 2013
- 2. Outline The HathiTrust Digital Library (HT) The HathiTrust Research Center (HTRC) The Workset Creation for Scholarly Analysis (WCSA) Project User needs & requirements Characterization of bibliographic metadata for corpus More about the WCSA RFP & prototyping projects Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 2 CNI 2013 Fall Membership Meeting 9 December 2013
- 3. The HathiTrust Digital Library (hathitrust.org) A digital preservation repository coupled with a highly functional access platform An international partnership of 80+ research libraries & consortia Provides long-term preservation of and access to volumes of member library collections that have been digitized by Google, the Internet Archive, Microsoft & member institutions Currently supports ingest of digitized book and journal content, and similar book-like materials Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 3 CNI 2013 Fall Membership Meeting 9 December 2013
- 4. HT DL by the numbers (as of Nov 2013) 10,973,063 total volumes 6,067,835 distinct bibliographic items: 5,778,450 book (monographic) titles 289,385 serial titles 3,803,630,600 pages 487 terabytes 3,512,404 volumes (~32% of total) digitized from public domain originals
- 5. More than just US Libraries Hebrew Hindi Indonesian Polish 1% 1% Dutch 1% 1% 1% Portuguese 1% Latin Arabic 2% Other 9% 97% of bibliographic records specify resource language Only 7% specify more than 1 1% Italian 3% Japanese 3% Chinese 4% English 49% Russian 4% Spanish 4% French 7% German 10%
- 6. HT DL Searching & Data Availability Web User Interface (http://www.hathitrust.org/home): Full text keyword (includes indexed metadata) Bibliographic metadata keyword Advanced (field-specific) bibliographic metadata searching Bibliographic metadata (http://www.hathitrust.org/data): OAI-PMH & custom bib API – http://www.hathitrust.org/bib_api HathiFiles (tab delimited metadata) – http://www.hathitrust.org/hathifiles Full-text (http://www.hathitrust.org/datasets): ~300,000 digitized volumes in the public domain – contact for bulk download or use API for volume-by-volume access ~ 3,500,000 volumes digitized by Google from public domain – available by arrangement, typically using rsync. Must agree to conditions of use. Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 6 CNI 2013 Fall Membership Meeting 9 December 2013
- 7. How many pages per volume? For volumes digitized from public domain sources Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 8. The HathiTrust Research Center (1) HTRC is a collaboration between HT, Indiana University and the University of Illinois at Urbana-Champaign Goal is to provide computational access to researchers: initially to all content digitized from public domain eventually to the entire HT DL corpus Currently hosts complete copy of HT metadata copy of OCR of all HT volumes digitized from public domain copy of OCR of all public domain volumes in HT Supported by the Sloan Foundation, the Mellon Foundation, IU, & UIUC Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 8 CNI 2013 Fall Membership Meeting 9 December 2013
- 9. The HathiTrust Research Center (2) HTRC end-user access (so far) HTRC Portal https://htrc2.pti.indiana.edu/HTRC-UI-Portal2/ Must login; pull-down login menu (upper right) to sign up (free) HTRC Workset Builder https://htrc2.pti.indiana.edu/blacklight Must login to this interface also; same credentials used. HTRC Sandbox (contact us) http://sandbox.htrc.illinois.edu:8080 Clone of Portal, but accessing 250,000 digital public domain volumes Supports use data api – http://wiki.htrc.illinois.edu/display/COM As well as HTRC Solr Proxy api – http://chinkapin.pti.indiana.edu:9994 More: http://www.hathitrust.org/htrc/faq, http://wiki.htrc.illinois.edu/display/OUT Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 9 CNI 2013 Fall Membership Meeting 9 December 2013
- 10. HTRC Portal (as it is now) Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 10 CNI 2013 Fall Membership Meeting 9 December 2013
- 11. HTRC Workset Builder (as it is now) Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 11 CNI 2013 Fall Membership Meeting 9 December 2013
- 12. Create a small workset / collection Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 12 CNI 2013 Fall Membership Meeting 9 December 2013
- 13. Submit a small workset / collection for analysis Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 13 CNI 2013 Fall Membership Meeting 9 December 2013
- 14. Your completed and pending analytical jobs Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 14 CNI 2013 Fall Membership Meeting 9 December 2013
- 15. Results
- 16. Workset Creation for Scholarly Analysis Premise The ability to slice through a massive corpus constructed from many different library collections, and out of that to construct the precise workset required for a particular scholarly investigation, is an example of the “game changing” potential of the HathiTrust...
- 17. Motivation & Models Collections, corpora, worksets, ...: Scholars & librarians aggregate items in a variety of contexts: Archival Curatorial Experimental Referential Thematic These worksets facilitate, sometimes enable certain kinds of scholarly inquiry Carl Spitzweg. 1850 The Bookworm (Der Bücherwurm) Analogy: HathiTrust worksets for analysis are as the contents of a scholar’s carrel in a library With apologies to Martin Mueller, et al. 2010. Towards a digital carrel: A report about corpus query tools
- 18. Anecdotal feedback from UnCamp 2013 My workset should contain… Volumes pertaining to Japan / in Japanese All volumes relevant to the study of Francis Bacon Music scores or notation extracted from HT volumes Images of Victorian England extracted from HT vols. Volumes in HT similar to TCP-ECCO novels 19th c. English-language novels by female authors Representative sample (by pub date & genre) of French language items in HT
- 19. What is a Workset (in context of HTRC)? A workset is an aggregation brought together for the purpose of analysis, i.e., to facilitate inquiry. Worksets are conceptual and need to be expressible in a variety of ways A workset encapsulates the specific materials that share specific attributes or satisfy some set of criteria. May be large, e.g., tens of thousands of items. Can be constructed by machine as well as human agents. Attributes and criteria not always bibliographic Items aggregated may be more granular than a volume
- 20. Why Worksets? The result of a first-level, rough filter Better scale for intensive analytics Provides essential scope for certain analytics Some tools (are trained to) work best on a narrow, homogeneous work-set Eliminate noise that would otherwise arise by asking questions across whole of HT
- 21. Scope
- 22. Workset Creation for Scholarly Analysis Prototyping Project Collection analysis, data modeling and prototype tools & services to facilitate workset creation Principal investigators: J. Stephen Downie, Tim Cole, Beth Plale Funded by a grant from the Andrew W. Mellon Foundation 1 July 2013 - 30 June 2015 Will feature 4 $40K sub-awards for prototyping/demonstration projects illustrating how worksets from HT DL can be created and used and can be useful for scholarly analysis Methods & tools for metadata enrichment, including with links Analytical services over full text useful for defining worksets
- 23. Key research questions for WCSA project Can we formalize the notion of collections and worksets in the HTRC context? What are the attributes that define and describe a workset in the context of HTRC? How can we balance rigor with extensibility & flexibility? What roles do data, metadata, annotations, tags, feature sets, and so on, play in the conception, creation, use and reuse of collections and worksets? Can we demonstrate the utility & practicality of worksets for HTRC?
- 24. WCSA Timeline • July 2013: Project Start • Q1: User needs assessments / focus groups • Q2: HT Corpus characterization Request For Prototype Proposals (RFP) • Q3: RFP Finalist Workshop (Chicago) Prototype experiment funding awarded • Q4-6: Prototype experiments done Metadata workflow & workset modeling • Q7-8: Planning for prototype to production Report out • June 2015: Project ends
- 25. USER NEEDS & REQUIREMENTS Harriett Green, English and Digital Humanities Librarian Preliminary results An early deliverable of WCSA Project
- 26. Who Are Our Researchers? Humanities scholars? Computer programmers and technologists? Digital humanities research teams? Previous research in scholarly use of digital resources (Duff and Cherry 2000; Brockman et al. 2001; Warwick et al., 2008; Sukovic, 2008 and 2011; RIN 2011) Identify use cases for HTRC and large-scale, digitized text corpora
- 27. GOOGLE DIGITAL HUMANITIES AWARDS RECIPIENT INTERVIEWS REPORT Report prepared for the HTRC in 2011 by UIUC researchers at GSLIS’s Center for Informatics Research in Science and Scholarship (CIRSS) Interviewed researchers who were awarded Google Digital Humanities Research Awards on research needs Findings for scholarly requirements included improved metadata, accurate OCR, data curation Report available to download at http://www.hathitrust.org/htrc
- 28. Feedback from UnCamp 2013 My work-set should contain… Volumes pertaining to Japan / in Japanese Music scores or notation extracted from HT volumes Volumes in HT similar to TCP-ECCO novels General Needs: 19th c. English-language novels by female authors User-friendly interfaces Documentation on the portal Avenues for community input in HTRC portal development
- 29. Scholarly Requirements We are interested in understanding how scholars and researchers that use digital book and serials collections decide which texts (or parts of texts) to include in collections used for analysis. This includes: How researchers identify, select and obtain access to texts to include in their analysis Understanding the specific fields/disciplines that work with these sources along with the types of research questions and analysis applied. Desired units of analysis (works, manifestations, pages, n-grams OCR, images, etc.) Transformation and preprocessing steps; Understanding sources and criteria used for identifying texts Specific methods of selection Methods of analysis Challenges to working with these digital collections (e.g., OCR quality, duplication)
- 30. Focus Groups and Interviews Conducted at DH 2013, JCDL 2013, and HTRC Uncamp conferences in summer and fall 2013 Goal: To understand practices of humanities researchers using digital collections, especially in the context of large-scale text corpora Survey instrument queried users about their experiential practices of organizing datasets
- 31. Participant Demographics Positions: Domains: Junior and senior faculty at liberal arts colleges and universities Computer programmers Librarians Data scientists Academic technologists Graduate students English literature, classics, linguistics, library and information science, history Institutions: Academic institutions in Great Britain, Singapore, Germany, France, and United States
- 32. Study Design 1. General types of data, materials, or collections 1. Purposes of collections 2. Selection or inclusion/exclusion criteria 3. Sources, acquisition, and access 4. Pre-processing and analysis 5. Post-analysis 6. Challenges
- 33. Analysis Methodology: Qualitative content analysis of user responses • A ―directed‖ approach based on inductive reasoning to condense raw data (transcriptions of audiorecordings of interviews and focus groups) into categories and themes Goal: To identify common themes and patterns in users’ responses
- 34. Coding (still ongoing) Coding manual consisting of category names, rules for assigning codes, and examples: Challenges — access rights Challenges — OCR quality Collections — comprehensiveness Objects — data Sources — Google Books Sources — Selection Criteria — Language etc.
- 35. Selected examples for categories Category: Challenges— Access Rights User: ―I check to see if a volume has substantial copyrighted text included in it already as quotes or extracts‖ Category: Objects — Temporal User: ―Classic materials‖ User: ―single-authored books of poetry between 1840 and 1900‖ Etc.
- 36. Early Findings • • • Roles of collections Need to implement granular, actionable units of analysis Importance of expert-enriched, shareable metadata
- 37. Figure 1. Selected focus group and interview excerpts on collection- and workset-building.
- 38. Figure 2. Selected focus group and interview excerpts on divisibility and objects of analysis.
- 39. Figure 3. Selected focus group and interview excerpts on metadata enrichment and sharing.
- 40. Use Case 1: Gender Scholar wants to compare works by gender, based on the Library of Congress headings This information is in the metadata, but hard to text mine Questions: How can I track gender of authored texts across time? What correlations are there between gender of the author and sentiment analysis of the text? How people and characters of different genders are treated in books over time?
- 41. Use Case 2: Serials A scholar wants to find a series of an author’s works that were originally serialized across several issues or volumes of a periodical. Serials vs. volumes as manifestations of works Map the pages for content Might be able to investigate questions as: What was the original instantiation of the work in serialized form? How can I text mine for sentiment and themes across the serialized texts?
- 42. Use Case 3: Images Scholar wants to find texts of Victorian travel narratives and the images depicted in them. Investigate questions such as: What are patterns/themes of images depicted in Victorian England travel narratives? What is the frequency of images in travel narratives?
- 43. Use Case 4: Dialogue in Texts Scholar wants to identify conversational dialogue between characters in novels. • Requires OCR that detects boundaries: can we detect quote marks and signal words for dialogue? • Create a training set of curated texts (i.e., TCP texts) matched with HTRC texts, apply detection algorithm • Enable questions such as: How are characters connected across the narrative—who interacts most frequently? What would sentiment analysis or topic modeling reveal about the dialogue in comparative novels of the genre?
- 44. User Needs for Worksets Comments from interviews/focus groups: ―How do I gather works similar to those I currently have in hand? Can I define different kinds of similarity?‖ ―How do I merge a HathiTrust collection of works and metadata with my set of works and tags and my colleague’s annotations?‖
- 45. How useful is existing metadata for creating worksets? HT metadata is bibliographic Built from MARC records provided by members & OCLC Good, consistent quality for author / title / pub info Subject less extensive, less consistent Genre more hit and miss Author gender not present in MARC bib records is present in some MARC authority records MARC records provided for serials are about the serial not about the contents of the serial No visibility over internal elements (e.g., images, embeded language / genre, dialog, ...) of digitized volumes Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 46. Total Records 6.1 million Having at least 1 genre 85% Having no genre Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 5.2 million 0.9 million 15% CNI 2013 Fall Membership Meeting 9 December 2013
- 47. Breakdown of other category, incl. fiction Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 48. 2.6 million (43%) of bib records include LC Class no. Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 49. Not all genres equally described Volumes identified as fiction at least one subject no subject subectGeographic subjectTemporal subjectTopic subjectName Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 270837 70706 200131 26% 74% 25491 12536 61788 13412 9% 5% 23% 5% CNI 2013 Fall Membership Meeting 9 December 2013
- 50. Top genres by country of publication bib records country specified and genre specified Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 6067835 5361473 4776973 CNI 2013 Fall Membership Meeting 9 December 2013
- 51. Top subjects by country of publication bib records 6067835 country specified and LC class no. specified Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 5361473 2354094 CNI 2013 Fall Membership Meeting 9 December 2013
- 52. Fiction as proportion of publications by decade Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 53. Opportunities Computed attributes Author age at time of publication FRBR relationships Add attributes not included in bibliographic records Author gender Author nationality Improve completeness & accuracy of bib records Describe internal components of volumes Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC CNI 2013 Fall Membership Meeting 9 December 2013
- 54. More about RFP 4 awards to teams of scholars, librarians & developers $40,000 each Period of performance 16 April 2014 – 15 Jan 2015 UIUC will supply a testbed of ~250,000 representative volumes; additional volumes (digitized from public domain) available UIUC will collaborate, provide access to HTRC cluster, ... Deliverables: final report; open source software Schedule: Letters of Intent Due (preferred): 16 December 2013 Final Proposals Due: 13 January 2014 Shortlist Meeting Invitations Issued: 20 January 2014 Shortlist Meeting: 20 February 2014 Award Notification: No later than 15 March 2014 Timothy W. Cole (t-cole3@uiuc.edu) University of Illinois at UC 55 CNI 2013 Fall Membership Meeting 9 December 2013
- 55. Questions? Timothy Cole Mathematics and Digital Services Librarian, UIUC t-cole3@illinois.edu Harriett Green English and Digital Humanities Librarian, UIUC green19@illinois.edu Twitter: @greenharr
- 56. Discussion Questions Key questions to look for in the data Alternative approaches and methodologies Knowing what we know about user needs to date, what are the implications for formalize the notion of workset How does this translate across domains? (e.g., Worksetlike objects in science and elsewhere...) What are the re-usability and re-producibility implications for such highly individualized and complex digital objects
Editor's Notes
- #5: HathiTrust is a partnership of major research institutions and libraries working to ensure that the cultural record is preserved and accessible long into the future. There are more than 80 partners in HathiTrust, and membership is open to institutions worldwide.
- #30: Ray soun de tra
- #32: Subfields: statistical text analysis; 19th c. American and British history, women’s history, digital humanities, digital collections, HCI, visualization, literary studies, digital philology, digital classicsAges: Ranged from 26 to 62 but the average was about 48, median was 50Besides US, nationalities included French, German, Indian, British, and ItalianExperience in the area ranged from 2 years to 30