OBJECTIVES

• Use full-text availability of journal Éire-Ireland for years 1994-2017, recent releases of public-use Library of Congress MARC XML files, and the publicly-available HathiTrust Digital Library Hathi Files to supplement our understanding of trends in historiography/scholarship in Irish Studies


• Who is being cited? Who is shaping the field (broadly defined)?


• Measures of heterogeneity in citation in particular (are we citing the same people again and again? Is that bad?)


• Are libraries prone to over-collecting certain scholars?

Non-OBJECTIVES

• Focus on scholarly investigation of Ireland -as-subject


• Not a look at what Irish authors were cited or collected (c.f.  Brian Lavoie and Lorcan Dempsey,  "An Exploration of the Irish Presence in the Published Record,"  OCLC, 2018)--at least not for library holdings


• However, we can get a sense of heterogeneity of literary subjects and authors via citations in Éire-Ireland

Corpus Overview

• Published by Irish American Cultural Institute (founded 1962), first issue 1966
• St. Paul, MN, to Morristown, NJ, in 1995
• Special themed issues (c. early 2000s on), with guest editors starting 36 1&2 (~28 guest editors, 2000-17)

Éire-Ireland

Corpus Overview

• 23 years of volumes, 1994-2017 (vols. 29-52)
• Yields 11,011 pages of main essay texts:
  • 696 Essays/Poems/Translations
  • 7,405 pages with footnotes or works cited
  • 1.9 million words
  • 454 unique authors
• Online publishers:
  • Project Muse, JHU (1994 - present)
  • EBSCOhost and Gale Cengage (via print subscription publisher; 1998 - present)

Back Archive

Corpus Overview

Corpus Prep Workflow

• Extract embedded text and HTML using Xpdf (open-source PDF suite) pdftotext utility  and pdftohtml utilities.

• Assemble metadata for each issue from filename, and for each essay from file contents

Text Treatment

Corpus Prep Workflow

• Scrape page text from body of plain-text files (due to inferior quality of HTML main text); pull footnotes from HTML

• Marry footnotes to page text, join essay and issue metadata to pages

Text Treatment

Corpus Prep Workflow

• Initial preparation workflow had struggled with detecting start of components within footnotes -- easy to identify start of note, but not subsequent citations in same note.

• Labeling of citation components beyond first two-to-three tokens. In initial 2017 report, only these tokens, where identified as personal names, could be used

Updates, 2017-2019

Corpus Prep Workflow

• Implementation of conditional random fields (CRF) in 2018: create training data in which components of footnotes were labeled (~20 footnotes, or ~700 components)

• "Score" unlabeled footnote components based on token features, e.g.:
  • Is it a number? Does it have four digits and start with viable century?
  • Is it a delimiter (, ; : .)
  • Is it capitalized?
  • Does it have quotations marks around it?
  • What type of tokens  surround it?

Updates, 2017-2019

Corpus Prep Workflow

• Use probabilistic modeling (specifically, Python CRFSuite to assign each microcomponent a label based on its similarity to components of labeled training data.
• Results: some improvement:
  • Good in separating journal articles from books
  • Enables us to grab subsequent citations in single note number
• But continued problems:
  • Separate and identify constituents in multi-author works
  • Ignore non-source components, e.g. direct-source quotations being confused for journal article titles (need to implement with further training data)

Updates, 2017-2019

Recap: 2017 Findings

Newspapers & Periodicals

"First-position," full-citation footnotes, name-like tokens

n = 21,850 footnotes

Recap: 2017 Findings

Primary Source Names

Scholars: By Number of Full-Citation, Head-of-Note

Recap: 2017 Findings

Scholars: By Average Placement of First Footnote

Recap: 2017 Findings

Newspapers & Periodicals

CRF-LabelED Data

Citation components labeled "authors"

n = 21,850 footnotes; ~9,000 "authors"

Subject-Authors

CRF-LabelED Data

Scholars

CRF-LabelED Data

Library of Congress

• MARC XML "Open Access" Distribution, "Book Files" (see http://loc.gov/cds/products/marcDist.php)
• Number of records: 10 million +
• LoC: This is an undercount
• Extract built on any title field (MARC 245) with "Irish" or "Ireland", n = 18,576
• Considered only scholar-authors here

Library of Congress

Book Holdings w/ title "Irish" or "Ireland", Scholar-Authors

HathiTrust Digital LIbrary

• Dominated by R1 (especially U Michigan, University of California System) holdings
• Disproportionate representation of pre-1924 holdings (libraries initially afraid to digitize copyright holdings)
• Number of records: 16 million +
• No authors(!) in public use Hathi files
• Extract built on any title field (HT takes from MARC 245) with "Irish" or "Ireland", n = 35,937

Book Holdings w/ title "Irish" or "Ireland", by Title

HathiTrust Digital LIbrary

Takeaways

• Scholar-authors are not as diverse, citation-wise as library holdings enable.
• Who gets cited is dominated by a shorter list of male historians and literary critics
• ...but, the proportion of footnotes covered by each of these authors is still small (~1.25%)
• Generalist histories predominate, and key edited collections (e.g. Field Day Anthology)
• Over-reliance on Dublin-based national newspapers
• Predominance of author-subjects of usual suspects (Yeats, Swift, Joyce) but also some that overlap with historians' interests (Mitchel)