3.1. Methodology Used in Research Trends Analysis

The research methods used to analyze trends in research disciplines are distinguished largely by bibliometrics, content analysis, and social network analysis.

3.2. Research Trends in Knowledge Organization

Studies to investigate research trends regarding knowledge organization have not been carried out to any great extent. Pattuelli (2010) analyzed 34 courses related to knowledge organization in LIS schools in the United States, and outlined the topics and readings taught in the courses. Cho (2013) examined the knowledge organization literature in Japan and the Republic of Korea by network analysis. In addition, Hunter (2003) studied metadata research trends by survey, and found XML semantic web, metadata harvesting, web services, and so on as the main research areas. Parlmer, Zavalina, and Mustafoff (2007) performed an analysis of Institute of Museum and Library Services digital collections projects conducted in 2003 and 2008. They surveyed the project managers and gained an understanding of the metadata audience, metadata application, decision factors, and problems for metadata schema development.

3.3. Topic Modeling Utilization in Research Trends Analysis

Topic modeling has not been much employed in research trend analysis. Most studies have utilized the literature as a dataset to investigate the applicability of topic modeling algorithms.

Griffiths and Steyvers (2004) analyzed abstracts from Proceedings of the National Academy of Sciences of the United States of America (PNAS) published from 1991 to 2001. They employed LDA and a Markov chain Monte Carlo algorithm to infer about topic modeling. Topics that continue to decrease and increase in the dataset were presented along with topic related terms. Mimno and McCallum (2008) investigated 300,000 articles related to artificial intelligence. They argued that topic modeling can be usefully applied to discover main authors, topics, and predict lead authors for topics. Hall et al. (2008) analyzed 12,500 papers from a journal of the Association for Computational Linguistics Anthology. He employed the LDA technique and discovered hot topics that have been emphasized in anthology, cold topics, and the decline of the leading conference in the field. Daud (2012) carried out topic modeling analysis based on DataBase systems and Logic Programming dataset as well as temporal analysis by year for the main topic, and determined the key authors, key topics, and changes in key topics in the literature.

Previous studies have contained a number of flaws. Studies to understand research trends have been conducted in various areas but there is little specific literature on the research trends of metadata, and studies have relied on limited methodology such as surveys and descriptive content analysis. To unlock key areas, changes in key research areas, and interlinking among research areas, more intensive studies of metadata research need to be done. Looking at the research on topic modeling performed to date, topic modeling has employed literature as data set. Through this, it can be seen that topic modeling can be used to analyze research tendencies.

4.1. Dataset

In order to investigate topics addressed related to metadata in LIS literature, we constructed our dataset by searching three databases commonly used in the LIS field—LISA, Library and Information Science Source (LISS), and Library, Information Science & Technology Abstracts (LISTA). The three databases have comparable search features, allowing us to construct the same or equitable queries. The searches were done in February 2015. From each database, we retrieved all the records for peer-reviewed scholarly papers written in English that have the term ‘metadata’ in the title, keyword, or subject fields. The initial dataset had a total of 5,473 records including 1,059 records from LISA, 2,065 records from LISTA, and 2,349 records from LISS. From the initial dataset, we removed those records that did not contain an abstract, either having an empty abstract field or having text other than an actual abstract (e.g., copyright notes, accession number, etc.) in the abstract field. Since the three databases overlap in their coverage of LIS literature, there were numerous duplicate records in the initial set. We detected duplicates by comparing both title and abstract and removing them. We also removed five papers published before 1995 and 19 papers published in January/February, 2015. The resulting dataset included 2,713 records published in 370 journals between 1995 and 2014. Fig. 1 shows the number of papers in the final dataset by year.

As can be seen in the table, the research literature on metadata grew steeply in the first ten years, from only a handful of papers per year in the mid-1990s to almost 200 papers in 2005. The number of yearly publications has stabilized around 200 since then. The growth can also be illustrated by the number of journals represented in our dataset. Until the mid-1990s the papers on metadata appeared in only a few journals, but drastic increases occur in the last few years of the 1990s and early 2000s. For instance, between the year 2000 and 2001, the number of journals in our dataset rises from 24 to 41, coinciding with a notable increase of papers in 2001.

4.2. Topic Modeling

4.2.1. Preprocessing

The combined title and abstract from each record in our dataset constituted a single document for topic modeling. The documents were preprocessed prior to the topic modeling. After removing punctuation, numbers, and special characters, we deleted words belonging to a stop-word list or with a length less than 3 letters. Since most, if not all, documents contain the word ‘metadata,’ we also decided to remove the term as it has little value in differentiating topics. For the remaining words (tokens), we performed stemming. All data processing and analysis were done in the R computing environment (R Core Team, 2014). The tm package (Feinerer & Hornik, 2014) was used to preprocess the documents.

Fig. 1.

The final number of papers by years removing duplicates.

4.3. Research Trends and Topic Networks

In addition to finding out metadata-related topics addressed in research literature, we are interested in how these topic areas have evolved over time. As in Griffiths and Steyvers (2004), we conducted a post hoc analysis on the topic modeling outcomes to track changes in topic distributions over time. More specifically, the proportion of topics assigned to documents were aggregated by the publication year of the documents, and each topic’s share in yearly publications was plotted to uncover any trends.

In order to look at the relationships between topics and the dynamics of the relationships over time, we employed network analyses of topics, using the topic distribution over documents. In topic modeling, each document is represented as a mixture of topics. In some cases, a document has a single prevalent topic with other topics having marginal proportions. In some other cases, a document comprises multiple topics each with a non-trivial proportion. We suppose that, if two topics are frequently addressed together in substantial proportions in the same documents, it indicates possible relationships between those topics. That is, relationships between topics are established based on the documents in which the given topics appear together.

5.1. Prominent Topics

In this section, we review topics 1 to 8 in detail, as they turned out to be the most prominent topics in metadata literature. Collectively, these topics account for 57.6 % of entire words (tokens) in the corpus, with each topic assuming more than 5%.

Topic 1 is about digital collection or digital library projects, as can be seen in the words falling in this topic with high probability. Included in the top third probable or relevant words are a set of words referring to the organizations in charge of such projects, including library, archive, cultural, heritage, and institution, as well as a group of words for information objects, such as object, material, and image. In addition, the words related to some key functions of metadata, such as describe, access, or preservation, also appeared high in the lists. The words associated with this topic occupy more than 10% of the corpus, making it the most prominent topic in the metadata literature in our dataset.

As expected, many of the representative papers of this topic report and share the experiences of building a digital collection or setting up a digital library, often discussing metadata related issues or challenges encountered in the process (e.g., Boyd & King, 2006; Woodley, 2002). There are also papers addressing specific aspects of a project, such as the implications of technical choices for digitization or the need for documenting metadata decisions (e.g., Lalitha, 2009; Symonds & May, 2009).

Topic 2 concerns the development or management of information systems/services. The main keywords for this topic include nouns such as system, service, software, and application as well as verbs like develop, manage, implement, and integrate. Also ranked high on the list of probable/relevant words are design, architecture, model, and framework. Combinations of these terms give a fairly good idea as to the topic area.

The papers in this topic discuss metadata based approaches or solutions for specific problems at hand in relation to information systems or services: for instance, access control (Yagüe, Maña, & Lopez, 2005), content management (Yeh, Chen, Sie, & Liu, 2014), or a federation of distributed resources (Aktas, Fox, & Pierce, 2010). Issues concerning the design and implementation of such solutions, or the proposed models or architectures, are commonly found in those papers.

Topic 3 deals with role of metadata or metadata librarians. The most probable words for this topic are library, resource, librarian, service, develop, technology, and digital, but this topic can be better understood when the relevant words distinctive to this topic, such as role, skill, future, and profession are considered.

The papers in this topic appear to divide broadly into two groups, while commonly noting the challenges that the proliferation of electronic resources have brought to library services. One group tackles the problem of organizing electronic resources and discusses the increasingly important role of metadata (e.g., Medeiros, 2003; Emery, 2007). The other group centers on the discussion of professional roles of librarians in the digital era, reflecting on the influence of technologies on library and information services. Many state the need for reconfiguring technical services or cataloging practices to better meet current and future challenges, and call for attention to the changing roles and competencies of librarians (e.g., Schottlaender, 2003; Han & Hswe, 2011).

Topic 4 covers evaluation of metadata quality. The top ten words with the highest probabilities of being associated with this topic are study, paper, quality, analysis, research, result, data, find, survey, and evaluate. The rank order of relevant words differs slightly, with quality, analysis, survey, and evaluate placed higher.

The most representative papers of this topic report the results of studies on metadata quality. Some analyze a number of metadata records and identify patterns of problems or errors therein. Often certain criteria such as accuracy, consistency, and completeness are used to evaluate the quality of metadata (e.g., Chuttur, 2012). Some further suggest and/or test mechanisms for quality assurance (e.g., Park & Tosaka, 2010; Chuttur, 2014). Many papers present empirical studies where a variety of methods including experiments, focus group interviews, and surveys are employed. Yet, some discuss quality issues based on a review and an analysis of research and practice in the field (e.g., Park, 2009).

Topic 5 is focused on semantic web and ontology. Top keywords for this topic are semantic, web, ontology, model, knowledge, base, paper, relation, concept, structure, domain, and so on. As can be seen in the above list of keywords, where all except for one rather general term paper represent a coherent theme, there is little to no ambiguity about this topic area.

The papers in this topic review various semantic web technologies (e.g., Kanellopoulos & Kotsiantis, 2007), or discuss ontology modeling or implementation, including a conversion of existing controlled vocabulary or metadata (e.g., Qin & Paling, 2001).

Topic 6 is about metadata standard development. The first glance at the list of probable words in this topic does not give a clear idea as to its topic content, as it includes rather generic terms that appear in multiple topics, such as project, develop, resource, and access. Only the term standard is relatively unique to this topic. However, when we examine further down the list of the terms relevant to this topic, along with its representative papers, it becomes evident that this topic centers on metadata standards, especially the development of various national or international standards. The names of standards organizations, such as NISO and ISO, are often mentioned in the papers and related terms like initiative, committee, (working) group, or programme appear high on the list of relevant terms.

Some papers in this group discuss the importance of developing and adopting metadata standards (Lagace, Breeding, Romano Reynolds, & Han, 2013), some introduce a newly developed standard or provide updates on a standard under development (e.g., Feick, Henderson, & England, 2011), and some provide a review of an existing standard, often with a discussion of emerging issues (e.g., Mullen, 2001).

Topic 7 mainly concerns record management. The top keywords in this topic include manage, record, paper, and system. Additional words particularly relevant to this topic are recordkeeping, context, and scheme.

A variety of questions regarding metadata for record management or record keeping have been addressed, including the role, purpose, or capacities of metadata in the context of electronic record management, the specifications for records metadata, the methods for acquiring or capturing record keeping metadata, the need for standards and tools, and so on (e.g., Evans & Rouche, 2004; Evans, 2007; Cumming, 2007).

Topic 8 is related to cataloging or bibliographic description of resources, as clearly shown in the lists of probable and relevant topic words including catalog(ue), record, library, bibliographic, MARC, RDA, access, control, description, resource, and so on.

Although sharing the central theme, cataloging, research problems addressed in the papers range broadly from the pertinence of specific aspects of cataloging rules and/or principles in modern catalogs (e.g., Conners, 2008), to the applicability of cataloging tools to organization of Internet resources (e.g., Ferris, 2002), and to future directions for library catalog or cataloging rules (e.g., Wakimoto, 2009). Many of the recent papers report testing and implementation of the new cataloging standard, Resource Description and Access (RDA) (e.g., Danskin, 2014).

5.2. Research Trends

The topics of prominence we looked at in the previous section are based on the proportion of words associated with a given topic within the entire corpus. That is, the result shows the distribution of the twenty topics across all the papers in the corpus that are published in the span of 20 years. In order to see the dynamics of these topics, we now turn our attention to the changes in the topic proportions in the corpus over time.

Since topics are assigned to each document with their respective proportions, when we aggregate the proportions of a given topic in the documents published in a year, we can obtain the share of the topic in that year’s literature. This provides quantitative measures of rise and/or fall of topics in popularity over time, as well as their relative prevalence.

For the overall trend analysis, we plotted topic shares from 2000 to 2014. The publications from 1995 to 1999 were not included in the plots, since the number of publications during the period was too small for this analysis. The plots for all twenty topics were first drawn, and among them, we identified topics with an upward trend, a downward trend, and those reaching a peak at different points in time. In the following, we present those topics.

5.2.1. Emerging Topics

Topic 4 (evaluation of metadata quality), Topic 10 (document extraction methods), Topic 12 (Linked Data) have increased in volume since 2000 as shown in Fig. 2.

Topic 4 (evaluation of metadata quality) shows a steep increase from 2000 to 2010, with its topic share being more than tripled. This trend demonstrates that growing attention has been given to quality issues as metadata has become a pillar of information services.

Fig. 2.

Emerging topics showing upward trends.

Topic 10 (document extraction methods) is related to automatic extraction of metadata from documents. Noting both the need for and the expenses of creating metadata, the papers in this topic explore various strategies and methods for automatically generating metadata using different parts of documents per se. This topic shows a good deal of fluctuation, yet the overall topic share has increased considerably over time.

The papers associated with Topic 12 (Linked Data) show a rapid increase since 2006, the year when Tim Berners-Lee coined the term and outlined the key principles for publishing and connecting structured data on the web. Also included in this topic are papers on data mining, or on curation or retrieval of special data including scientific data and geographic data. This explains why the topic was present before 2006 and has a spike in 2002. The vast majority of the representative papers (24 out of the top 30 papers) in this topic, however, were published after 2006, many touching on the application or adoption of Linked Data concepts, technologies, and practices to the creation, transformation, and use of Libraries, Archives, Museums (LAM) metadata.

5.2.2. Declining Topics

Interestingly, topics showing a downward trend (Fig. 3) are related to metadata standards: topic 6 (standard development), topic 15 (XML and encoding standards), and topic 16 (Dublin Core). It appears that these topics attractedthe most interest at the early stage of metadata research, before or around the beginning of the 2000s, but began to fade at least from the research front.

As described in the previous section, topic 6 has to do with the development of metadata standards, and many papers in this topic introduced a then new standard or reported updates on the development or implementation of a standard.

Topic 15 is about XML and encoding standards. There is a surge of interest in this topic in the metadata community from 2000 to 2002, demonstrated by a special section on XML in Library Hi Tech (volume 19 in 2001). Many of the papers in this period provide an introduction to the suite of XML specifications and technologies, along with a discussion of their implications for metadata sharing. In addition, the release of the XML-compliant EAD version 1 in 1998 seems to kindle the interests of archivists in this topic, resulting in a series of papers on XML encoding of archival resources in following years.

Topic 16 about Dublin Core has declined steadily. As one of the first metadata standards that received international recognition and enjoyed wide adoption, Dublin Core took a central place in the early phase of metadata research, but has given its share to other emerging topics as the research horizon expands. In fact, this topic reached its peak in 1997 (not shown in the plot), where its topic share amounted to 16.8% of the entire corpus. The topic share continued to decrease to 8% in 2000, 3.8% in 2006, and finally to 2% in 2014.

Fig. 3.

Declining topics showing downward trends.

Fig. 4.

Topics with peaks and valleys showing dynamic changes.

5.3. Topic Networks

Having identified the main themes of metadata discourse and having looked at the changes in prominence of such topics over time, we now are interested in how these topics are interrelated and how the strengths of the connections between topics have changed during the fifteen year period.

As described in the method section, in order to look into the relationships between topics, we adopted a network analytic method. Using topic proportions in each paper, we derived connections between topics based on the papers addressing two or more topics together. In order to determine those papers spanning multiple topics, we set the threshold at 30%. That is, if a paper’s topic composition consists of two or more topics each in a proportion of 30% or more, the paper constitutes a potential link between the topics. Between 2000 and 2014, about 32% of the papers (838 out of 2,617) fell under this category.

Fig. 5 shows the topic network considering all papers published between 2000 and 2014. The size of a node is proportional to its degree, and the thickness of a link between two nodes denotes the strength of their connections determined by the number of papers addressing the two together. Each link in this network has a minimum value of 10, which means ten or more papers addressed the topics on each side of the link together. The isolated nodes, topics with no connections, are removed from the figure.

Fig. 5.

Topic network considering all papers published between 2000 and 2014.

Given the condition of ten or more shared documents to establish connections, 16 out of 20 topics have at least one connection with other topic(s). The node with the highest degree, the most well-connected topic, is topic 3, having links to eight other topics. This is not surprising since the role of metadata or metadata librarians is a topic that can be discussed in various contexts. Most of the prominent topics assuming a large proportion in the corpus (topics 1 through 8) have a relatively high degree ranging from three to eight, except for topic 7 (record management). Topics 1 through 4 clearly constitute the tightly-knit core of the network, to which smaller topics are connected with varying strengths. It is notable that the more technically oriented topics, topics 5, 9, and 10, form a clique somewhat separated from the core.

In order to look at how the relationships between topics have evolved over time, as well as to get more insights into their composition, we divide the 15-year period into three sub-periods (period 1: 2000–2004, period 2: 2005–2009, period 3: 2010–2014) and draw a topic network for each sub-period. Considering the relative sizes of the document sets in these periods, we adjusted the number of shared documents to create links between topics—a minimum of four shared documents was used for period 1, and a minimum of five for period 2 and period 3. Fig. 6 presents the three sub-period networks.

As can be seen in Fig. 6, the networks show considerable differences not only in the volume of connections but also in their structure. Overall, it is clear that increasingly more connections among topics have arisen as time passes, but at the same time there are notable changes in topics assuming central positions.

In the first five year period (2000–2004), the most noticeable difference compared to subsequent periods is the prominence of topic 6 (metadata standard development) and topic 15 (XML and encoding standards) both in terms of their degree and the strengths of connections that they have with other topics. Topic 16 (Dublin Core) is present only in period 1, being connected to topic 6 and topic 15. Note that these topics all showed a downtrend from around 2000, as described in the previous section. The fact that they remained salient in the network suggests that, while declining in volume, these topics relating to base standards still had an important place in discussion of other topics. These topics, however, become peripheral in subsequent periods.

In period 2, topics 1 (digital library projects) and topic 2 (development or management of information systems/services) moved to the center of the network. Topic 1’s central position, with its links to a variety of topics, indicates that various metadata research topics were often introduced and discussed in a context of digital projects. Topic 2, on the other hand, showed a tendency of having connections with more tech-oriented topics, including topic 5 (semantic web and ontology) and topic 12 (Linked Data). It is also notable that topic 4 (evaluation of metadata quality) started to appear on the network in this period.

In period 3, topic 3 (role of metadata or metadata librarians) emerged as a center of the network, with a degree of nine. It indicates that a discussion of the function of metadata or the professional role of metadata librarians takes place in a variety of topics in more recent literature. In addition, topic 12 (Linked Data) and topic 14 (publishing and access) show a considerable growth in links, reflecting new trends in research.

Fig. 6.

Evolution of topic networks. (a) Period 1 (2000–2004), (b) period 2 (2005–2009), and (c) period 3 (2010–2014).

The composition of networks and the changes therein portray how topics develop in relation to other topics. For instance, looking at topic 13 (metadata harvesting), the topic was first connected to topic 6 in period 1, topics 1 and 2 in period 2, and finally topics 3 and 4 in period 3. That is, papers introducing the ideas and mechanisms of metadata harvesting, including the OAI protocol standard itself, first formed a link between topic 13 and topic 6. Then papers describing a project creating regional or national repositories using OAI-PHM and papers concentrating on design and implementation of services building upon metadata repositories appeared, connecting this topic to topic 1 and topic 2, respectively. Finally, papers focusing on the evaluation of consistency of metadata elements and values across harvested records constitute the link between topic 13 and topic 4 in period 3. In addition, discussions on academic libraries’ role in managing and promoting OAI compliant institutional repositories spanned both topic 13 and topic 3.