At the library technology institutes referenced above, despite dissenting views that authority control would be prohibitively difficult and expensive, the attendees believed that such files, if properly controlled, would give structure to the bibliographic universe and the universe of knowledge.³ One well-known definition of authority control is “the process of maintaining consistency in the verbal form used to represent an access point in a catalog and the further process of showing the relationships among names, works, and subjects.”⁴ The practical (if anecdotal) experience of librarians did lead to research into the high cost of authority files. The proliferation and popularity of local authority files have increased the breadth of authority control over the names of both individuals and corporate bodies. A special issue of Cataloging and Classification Quarterly followed the 2003 international conference “Authority Control: Definitions and International Experiences” held in Florence, Italy. ⁵ Projects reported included creating local authority files for historical corporate bodies in the Bibliothèque Nationale de France, formulating corporate and personal names associated with the worldwide Roman Catholic Church, experiments in interoperability of disparate Italian authority files and disparate Chinese, Japanese, and Korean catalog standards, and developing supportive theoretical arguments in favor of the practice of authority control.

Several studies have tested automatic processes to create authority files, with mixed results. Snyman and Rosenberg first addressed the need to develop new technological and automatic methods to control the cost of maintaining authority control.⁶ Veve reported on a project at the University of Tennessee Library and concluded that, despite various efforts to automate authority work, levels of human intervention were still required, though perhaps some automation could hold down costs.⁷ Patton and colleagues attempted to explore automated processes to assist catalogers in name authority control by automatically calculating the probability of matches between metadata strings and LC authority files.⁸ Their matching algorithm was successful 58 percent of the time. Rodriguez, Bollen, and Van de Sompel explored a more general solution to propagate metadata from environments rich in metadata onto resources whose metadata are sparse.⁹ Specifically dealing with corporate name authorities, Blake and Samples reported on a project at North Carolina State University to normalize organization names within the university libraries’ electronic resources management (ERM) module, seeking the benefits of greater data integrity in their management of vendors and acquisitions.¹⁰

Progress also has been made in the internationalization and aggregation of name authority control. The Name Authority Cooperative (NACO) was founded in 1976 as a library consortium that, under the leadership of the LC, maintains an extensive name authority database; Bynum offered a prospectus of the history and operation of the organization.”¹¹ The presence of corporate names within the NACO authority file has materially aided the construction of the OCLC Publisher Name Authority File (PNAF). More recently, OCLC Research has led and hosted the construction of the Virtual International Authority File (VIAF). After a prototype was launched that virtually combined the national name authority files of the LC, the Deutsche Nationalbibliothek, and the Bibliothèque Nationale de France, the VIAF has grown into a virtual collaboration of eighteen national-level cataloging organizations.

French, Powell, and Shulman, at the University of Virginia, attempted to use mathematical techniques of clustering theory to aggregate different authority files.¹² They worked with Astrophysics Data System records, using a subset of the database comprising 85,000 refereed articles from seven different journals. They admitted at the outset that it was impossible for them to completely automate the process of clustering corporate names “by lexical techniques alone.”¹³ Instead, they used an iterative variety of programmatic techniques for string clustering and matching, and approximate word matching, followed by expert review of the results. Their complete technique achieved cost savings of approximately half of the human effort in constructing an authority file. Similar clustering techniques also were featured in the early stages of the PNAF (see below).¹⁴

Authority control over the names of publishers in current cataloging practice continues to present difficult issues. The MARC field 260, subfield b ($b), contains the name of the publisher in “the shortest form of the name that it can take to be understood internationally,” with the added complexity that local practices may stipulate how much text to transcribe exactly from the title page or other chief source of information.¹⁵ In addition, changes in the rules of cataloging practice compound the difficulty of automatically identifying matches in the data strings found in this part of the record (see table 1). A study by Jin found a discrepancy rate of 25 percent between corporate names found on official company websites and the corporate names in the LC authority file.¹⁶ However, collection development in libraries often depends on the specialized nature of a publisher's output.¹⁷ The Association for Library Collections and Technical Services (ALCTS) has reported on the need for better authority control for library acquisitions.¹⁸

The work by OCLC Research to normalize publisher names has involved data mining and programmatic clustering of bibliographic records, supplemented by manual review of the results. Data mining appeared first as a tool for business intelligence, only later to be adopted by libraries; the success of Google and Amazon has taught the library field that greater value exists within bibliographic data as well. Libraries have made huge investments in creating and maintaining rich, structured information describing the resources in their collections. These data embody considerable value by supporting basic local access and inventory control. They also represent potential value in terms of knowing more about the characteristics of library collections. OCLC's Office of Research has invested significant effort in the area of data mining.¹⁹

Specifically, research projects have demonstrated the value of the WorldCat database as an “aggregate collection” of bibliographic data.²⁰ It thus has utility as a global-scale dataset of potential value that can “not only provide librarians data for decision-making for collection and service development, but also provide users with enhanced discovery and access methods.”²¹ The WorldCat database is an increasingly global and increasingly comprehensive source of bibliographic data and remains strongest in its data on books. As of February 2011, WorldCat contained more than 217 million records, with more than 1.68 billion distinct library holdings of those resources; 57.5 percent are non-English catalog records, illustrating the increasingly global reach of the “aggregate collection.”²² Its member libraries are located in more than one hundred countries, and the data go beyond those countries to include works from countries that are collected in other OCLC member libraries.

The first OCLC research into publisher name data was performed on an earlier snapshot of the WorldCat database from July 2005. Researchers mined bibliographic records that had a value of “English” in the language fixed field, resulting in 35,434,911 records (61 percent of the database). The next criterion was to explore the presence of valid ISBNs in the 020 field, because ISBN prefixes provided a more consistent way of identifying publishers. Researchers determined that in 2006 more than 22 percent of WorldCat records contained an ISBN, and more than 99 percent of those ISBNs contained a valid check digit.²³ (The check digit in an ISBN is the final digit of the number; it is not assigned by the publisher, but rather is computed according to rules set out by the International ISBN Authority, and serves as a validation for ISBN data points.)

Having demonstrated the prevalence of ISBN data within the bibliographic records, researchers made a first experimental attempt at programmatic clustering of records with a single ISBN prefix to gather variant forms of publisher names. The Free Dictionary defines data clustering to be “the science of extracting useful information from large data sets or databases.”²⁴ By partitioning the data into different subsets (i.e., clusters), the data in each subset ideally shares some common trait. Because most ISBN prefixes are uniquely assigned to a single publishing entity for the assignment of full ISBN numbers, the ISBN prefix seemed to be a good common trait for clustering bibliographic records. Relatively few exceptions occurred in the case of ISBN prefixes assigned to vanity presses and to some publishing communities outside the English-speaking West. However, for the initial research, the researchers examined only English-language cataloging records, and the ISBN prefix was a powerful hook into the overall bibliographic data.

ISBN prefix 019, which belongs to Oxford University Press, was used as a first test group for clustering. In the July 2005 WorldCat snapshot, prefix 019 was the most frequently occurring prefix in WorldCat; it was the prefix for one or more ISBNs within 84,276 records (0.15 percent of WorldCat). The researchers extracted the contents of subfield b of MARC field 260 from these records and deemed these data the publisher name. This process resulted in 91,528 unique strings of text. The publisher names were then normalized according to NACO normalization rules to account for differences in capitalization and punctuation, resulting in 1,550 unique normalized publisher names. The normalized publisher names were clustered using the Levenshtein Distance value.²⁵ This value measures the similarity between two strings by counting the number of deletions, insertions, or substitutions of characters needed to transform one string to the other. The publisher strings were then clustered by this distance metric.

Researchers then attempted automatic resolution of the data across the WorldCat database into a set of variant names for each publishing organization. After refinements to the clustering algorithm to better account for noise phrases and punctuation, the ISBN prefixes most frequently appearing in WorldCat as of January 2006 were automatically clustered (see table 2). In the case of the 019 ISBN set, the program yielded an 85 percent agreement. However, an application of the same program to the next 4 largest groups of ISBN prefixes achieved less than 5 percent success in identifying matches. In other words, this step of the project achieved a workable definition of a number of distinct entities, with their nested inter-relationships, directly from WorldCat data. However, this only was possible with a high level of human intervention.

ISBN prefixes were retained as a data mining technique with different algorithms during the construction of the OCLC PNAF database, as follows. The research team concentrated on a group of high-occurrence publishers. Accordingly, a program was developed to extract sets of ISBN publisher prefixes that represent the highest-occurring ISBN prefixes within the set of database records as sorted by country of publication. (The country is defined by its current political boundaries as coded in the MARC fixed field Place of Publication.) Researchers constructed a list of the most prominent publishers, seeding the PNAF database with high-incidence publishers from a dozen countries around the world, the top 10 research university presses, and any publisher involved in a merger or acquisition during the time of research (under the working assumption that the footprint in the global bibliographic world of any publisher purchasing another would be increasing).²⁶ A large part of WorldCat was thus clustered according to large subsets of ISBN prefixes.

In the case of each publishing entity identified, an authoritative Preferred Form of the name was first assigned. If the publisher already existed in the NACO National Authority File (NAF) as a corporate name (44 percent of the publishers in the PNAF were included), that authoritative form also was selected for the PNAF Preferred Form. All variant strings mined from the 260 $b bibliographic data were then compared to the Preferred Form; comparisons were made according to a tri-dist fuzzy matching program and given further manual review afterward. Tri-dist compares strings on the basis of three-letter sequences called trigrams. When two strings are compared, the strings are typically normalized in some way, for example to eliminate differences in capitalization. Then the two strings to be compared are broken up into overlapping trigrams; for example, using the underscore character to represent a space, the string “Al Smith” generates eight trigrams: “al,” “al_,” “l_s,” “_sm,” “smi,” “mit,” “ith,” “th_.” The trigrams from each string are then compared and a score that estimates the probability of a match is computed on the basis of the proportion of trigrams in common. The fuzzy matches were subjected to human review to assure data quality.

The team then worked outwards from these initial publishing entities, researching all known hierarchical structures related to them, current and past. Relationships were recorded between publishing entities for imprints, acquisitions, and subsidiary divisions, and were collected from a variety of sources. New strings for imprints and related publishing entities were harvested from the MARC 260 subfield b (Name of publisher, distributor, etc.) data and a variety of published business intelligence sources and published company materials were consulted. Each was then cited in the PNAF for each instance of source for a data point.

The number of publishers, imprints, and other publishing entities with records in the PNAF that included data on their hierarchical relationships and other data totaled 1,854 (see below, under “Results”). To further increase the bibliographic data related to these clusters, the table of all variant strings mapped to each entity was then compared once more to the complete bibliographic database. The verified publisher name strings were compared to all 260 $b contents to capture records that do not have an ISBN but still may be associated with the publisher via the 260 $b field.

This process yielded final data clusters totaling some 16.3 million records, which in turn represent 550 million holdings, slightly more than 33 percent of worldwide library holdings as reflected in WorldCat.²⁷ This richness is a direct benefit from the decision to begin with high-occurrence entities.

The automatic methods of data mining and clustering enabled researchers to build an experimental database of publisher data. Records were resolved into clusters via ISBN prefixes and via previously identified publisher name description strings; this process identified the issues associated with the bibliographic data relating to publisher descriptions. Both the automatic parsing of name clusters and the more complex second procedures, which led successfully to the construction of the OCLC Publisher NAF, validated the approach of using ISBN prefix as an initial data element for mining and clustering bibliographic records by publisher. However, in both cases the amount of manual review required hampered research efforts to fully automate the process at a global scale. Researchers had intended to develop a completely automatic process to map publisher name authority information into bibliographic records, but they found the task at this point too costly in terms of human intervention. This finding is in line with many earlier projects reported in the literature on the difficulty of fully automating the practice of matching strings to construct authority records.

The planning of the PNAF database as developed in this project included the decision to concentrate on high-incidence publishers. This decision did yield a very robust dataset to support the PNAF as it stands. The data-mined table of more than 60,000 variant forms of the 260 $b data allowed more than 16 million bibliographic records worldwide, representing more than 550 million global holdings, to be mapped to the 1,854 publishers in the PNAF database. The large number of strings identified with the cataloging data supports at least one of the theoretical arguments commonly made in favor of authority control: it reduces the amount of data clutter, both the labor-intensive clutter of catalogers entering unregulated strings and the user- and system-unfriendly presence of clutter in the resulting bibliographic records. The data from 260 fields resolved within the PNAF had provided a barrier to access that the resolved form could solve. In addition, the complexity of the hierarchical relationships surrounding many publishers in the current world of mergers and acquisitions makes the organizational chart data an extremely valuable component of the PNAF. Any similar projects or further development of the PNAF can only help librarians better assess their collections by publisher.

The construction of publisher profiles verified the method by comparing clusters of records assigned to different publishers. The profiles of the subjects, authors, and languages of a publisher's works in the global bibliographic universe as reflected in WorldCat demonstrated in great detail the differences between the clusters of bibliographic records parsed via the PNAF variants. The differences observed tended to fall along predictable lines, given the specific publishers involved. Such profiles of each publisher's footprint in the bibliographic universe as reflected in WorldCat, of course, cannot statistically prove the completeness of these data clusters. However, they offer more detailed and nuanced profiles of the publishers’ history than are available anywhere else in the publishing or bibliographic world. At the most granular levels of subject analysis, the profiles offer a detailed picture of a publisher's character, and, pointedly, each of the four publishers’ characters were quite distinct.

These differences also tend to validate the unique intelligence present within the PNAF data. Librarians, publishers, and users can view a portrait of the publisher's output in terms of the authors most associated with the publisher, the languages published, and most importantly, the subjects in which a publisher offers the most expert concentrations. As expected in the research goals, the profiles can be valuable to ERM systems because the name authority controls can help inform collection analysis and development and approval plans. The profiles can be useful to publishers as they consider their competitive position in the library marketplace and also can aid users and public service librarians in discovering publisher outputs. As stated by Blake and Samples, “OCLC's publisher name authority server nonetheless demonstrations [sic] a need for organization name authorities and may provide context for librarians whose methods and research have already prompted similar projects.”³⁵

The value of the publisher profiles led to the construction of the WorldCat Publisher Pages. All of the publishers in the experimental PNAF database are represented by a single webpage containing their data from the PNAF (including the Preferred Form of the name, cities of operation, and most importantly, the hierarchical organization chart) and their publication profile in global WorldCat, authors, languages, and subjects.

For future work, researchers have been considering recommendations on ways to code the authorized form of publisher names directly into MARC records, if a completely automatic process for resolving the names could be developed. The most obvious place would be MARC field 710 Corporate Name Added Entry, with the publisher name perhaps occupying a new $6, with NACO in $2 where the Preferred Form of the Name also may be linked to the NACO Authority file (44 percent of the current PNAF). The place of publication, if it could be similarly standardized in the future, could occupy a 752 Added Entry Hierarchical Place Name, with $2 for the FAST terminology currently embedded in PNAF place names.³⁶ The database currently uses its own unique identifiers, but researchers have been in discussion with those developing the International Standard Name Identifier (ISNI) system regarding incorporation of the PNAF publisher names. However, the reliance on human intervention to update and maintain the database is a detriment to inclusion in other systems and services.

This research on publisher names both confirmed them as a difficult issue in Anglo-American cataloging and set a potential example for providing authority control over them. The researchers set out to construct a database containing authoritative strings for publisher names and a variety of data relating to their publication output, and they accomplished this goal. Each automatic method helped generate clusters of items based on an assigned publisher, first via ISBN prefix and then via further matches of 260 $b data, leading to a robust database of high-incidence publishers. Though the process could not be fully automated on a global scale, some 1,854 high-impact publishing entities were profiled by their publishing output, with detailed differences emerging between the profiles. The profiles as a research output are freely available on the web via the WorldCat Publisher Pages.

The data captured for each publisher provide a model service for advanced collection analysis and provide additional value for user access to library resources. Tens of thousands of variant strings were resolved to the small number of publishers in the database, potentially reducing cataloging time by providing automatic suggestion of Preferred Forms for publisher names to catalogers. Further applications of this authority control procedure in the Publisher Description could code the Preferred Form of the publisher name directly into the MARC records, even if a fuller, more informative string were entered in the publisher description area. Such an application of authority control, even for the limited number of (high-impact) publishers in the PNAF, would offer benefits to both publishers and collection development librarians by increasing the power of collection analysis tools to parse a collection by publishing agency. Such an application of authority control also would benefit users and academic public service librarians by allowing better access to searches by publisher name.