Mapping the evolution of topics published by Education for Information: Interdisciplinary Journal of Information Studies

2.1Corpus splitting

Our corpus covers more than 36 years, from 1983 when this journal was created to the first quarter of 2019 when the data collection was done. Thus 2019 data is incomplete as only the first issue of the journal had been published and entered into the database at the time of corpus collection. In such diachronic studies, it is customary to split the corpus into manageable time spans in order to obtain intelligible maps and better follow the trends. We tried three approaches to corpus splitting:

Bearing in mind this skewness, we believe that mapping the titles by EICs can still tell us something about the topics prevalent during the mandate of each EIC during and how the journal’s center of gravity has evolved over time.

A more traditional way of splitting a corpus spanning a long period of time is to partition it into an equal interval of years. Previous studies tended to split the corpus by periods of five or six years, as this time span was judged sufficient to perceive trends (Chen et al., 2010). In this study, we tested two intervals of 6 years and 5 years.

The distribution of titles in the two intervals is given in Table 2 above. P1…Pn correspond to either five or six year intervals of the corpus. For instance, P1 corresponds to the years 1983–1987 for the five-year interval and to the period 1983–1988 for the six-year interval. The titles published in the first quarter of 2019 were not included in the five year interval as the division of the corpus lifespan (36 years) by 5 left us with a period in which there was only one year (2019) and which was incomplete at that. This explains the difference in total number of titles in each interval.

As shown in Table 2, there is not a significant difference in splitting the corpus by 5 or 6 years. In the following, we will analyse the evolution of titles by EICs and by 6 year interval.

2.2Text analysis process

As the focus of our analysis is the titles published by the journal, we used the Vosviewer package (Van Eck & Waltman, 2010; 2020). VosViewer constructs bibliometric maps taking as input several types of bibliographic units such as authors, journals, keywords or text data from titles and abstracts of documents. In the latter case, significant text units have to be extracted first. For this, VosViewer uses the sentence detection algorithm provided by the Apache OpenNLP library to detect sentences and perform part-of-speech tagging (POS). It then identifies noun phrases (NPs) using surface morphological rules. VosViewer seeks to extract compound NPs, i.e., the longest possible NPs found in a sentence. This ensures that more meaningful text units are preferred over uniterms (one word terms) which tend to be more vague. The software then performs some normalisation (stemming) on the NPs in order to regroup identical terms, i.e., convert plural forms to singular, lower case all letters, remove accents and alphanumeric characters. This has the effect of increasing the occurrence of some NPs (Van Eck & Waltman, 2010; pp. 34-35).

After NP extraction, the analyst has to decide which terms should be kept for the later stages of analysis. To this end, the software allows the analyst to set an occurrence threshold below which terms will not be considered for input into the clustering scheme. Habitually, terms that occur only once (hapax) are excluded. Depending on the corpus characteristics, a higher threshold can be set empirically.

In VosViewer, term occurrences can be counted in two ways: full or binary counting. In the first case, the total occurrences of a term in all the documents is calculated (full counting). In the second case, only the number of documents in which a term occurs is recorded irrespective of how many times the term appeared in each document (binary counting). Notice that in the case of our corpus of titles, the impact of this distinction is minimised given that a term is likely to occur only once in a title, although we cannot rule out the probability that some common terms may occur twice in a title but that probability is low.

We opted for the binary counting since our objective is to identify trends in the publications across time. Hence it was more relevant to highlight topical terms that appeared in many documents which can signal recurring topics. We also set a minimum of 2 occurrences for a term to be considered for further stages of analysis.

The next step is term filtering. To do this, VosViewer calculates a relevance score for each term which is similar to the well known tf.idf score used to select index terms in information retrieval (Salton et al., 1975; Sparck et al., 1984). The rationale is to distinguish topical or domain specific terms from generic terms. The former appear frequently but in a small subset of documents while the latter tend to appear in almost all the documents without informing us about the specific topic of each document and without being domain specific. For instance, words like “analysis”, “system”, “conclusion”, “new method” which are just part of the general vocabulary of academic writing will have a low relevance score while domain specific terms such as “Library” and “Information Science” will receive a higher relevance score and are likely to appear on the maps. By setting a threshold on the relevance score, generic terms and stop words can thus be filtered out of the further steps of clustering. Usually, in clustering endeavours, roughly 40–80% of the terms are filtered out using a combination of (co-)occurrence threshold and term weighting.

Given the small size of our corpus, we decided not to apply the relevance score calculated by VosViewer in order to keep all the terms that met the minimum occurrence of 2 and avoid a massive elimination of terms from the analysis.

2.3Clustering process

The terms selected from the titles with an occurrence of ≥ 2 formed the input to the clustering stage. To build the clusters, VosViewer constructs a co-occurrence matrix and offers four options for computing the strength of the association between two items:

While the software recommends not to use the ‘no normalisation’ option, the visualisation of the maps using any of the other three options did not show significant differences regarding the contents (see below), so we opted for fractionalization (option 3) that normalises the forces of the connections between the nodes.

VosViewer builds non-overlapping clusters, also known as “hard clustering” in the literature, and some terms may not be included in a cluster. Clusters are labeled using cluster numbers and are chosen automatically by the program based on the number of occurrences of the term and the relevance score (Van Eck & Waltman, 2020).

With the processes described in Sections 2.1–2.3 above, we obtained the following characteristics for the three methods of corpus splitting that we tried.

When looking at the corpus characteristics by intervals of 5 or 6 years (Tables 3 and 5), the first observable trend is the sharp decline in the productivity of the journal after the first 18 years (1983–2000). The journal published 232 titles in the first five-year period (1983–87) or 280 papers in the first six years (1983–88). By 1998–02 (Table 5) and 2001–2006 (Table 5), these numbers had dropped to 89 (a 62% drop) and 81 (a 72% drop) respectively for the same intervals.

Also, applying a threshold of 2 for minimum occurrence has the effect of reducing the number of terms considered for each period by approximately 80% which is the standard observation in data analysis studies: about 80% of the initial items are filtered out in some way or the other using occurrence or co-occurrence thresholds and do not appear on the final visualisations produced. Only about 20% of the data or less end up on the maps. For instance in Table 3 (corpus analysed by EICs), we went from a total

of 1658 extracted terms for EIC 1 to 341 terms which met the 2 occurrence threshold and furthermore co-occurred with other terms in that period covered by EIC. We will return to the impact of this feature reduction in the discussion section.

2.4Building the maps

VosViewer offers different types of maps for viewing links in a co-occurrence matrix: network, overlay and density (Van Eck and Waltman 2020). The network view represents items by their label and by a circle the size of which reflects the weight of the item. As the display tries to avoid overlapping labels, the labels of some items may not be displayed. The colour of an item is determined by the cluster to which it belongs. Links are materialised as lines between items. The overlay visualization is similar to the network visualization save that items are colored differently. The density view has two variants: item density and cluster density. The item density is similar to the network and the overlay view except for the background colour and the absence of links. This view is more akin to heatmaps. The cluster density view is available only if items have been assigned to clusters. It is similar to the item density view except that the background colours enable an easier identification of items belonging to a cluster. In essence, VosViewer offers either an intuitive view of relations between items (label view) or the density view/cluster density view if the focus is on the structure of the corpus (i.e., on the importance of an item).

VosViewer offers further parameters to display the maps of which we mention attraction and repulsion since they influence the way in which items are located in a map by the VOS layout technique. We selected the values of attraction 6 and a repulsion value of 1 for a best visualization of the terms and their connections. For more details on these options, see Van Eck and Waltman (2020: 21).

3.1Evolution of topics by Editor-in-Chief (EIC)

We recall that the journal has had three EICs since its creation and that we split the corpus along these lines (see Tables 1 and 3 above for details of the corpus parameters per EIC). The first EIC’s mandate covered the period from 1983–2013 (30 years), EIC2 covered the period 2014–2017 (3 years) while EIC 3 only had 1 year at the time of corpus collection (2018–2019).

Figures 1–3 respectively display the prominent terms from the titles published by the three EICs.

EIC 1 published 799 titles yielding 1658 terms out of which only 341 (20%) met the occurrence threshold of 2 and were connected and thus went on to the mapping stage. In the map, terms belonging to the same cluster bear the same colour. Hence in the above map, terms such as “education, conflict and report” are part of the largest cluster with 52 terms. The terms such as “india, botswana, implication, lecturer, past” belong to the second biggest cluster with 39 terms.

Figure 1.

Map of terms in titles published during EIC1 (1983–2013).

The titles published during the first 30 years of EIC 1’s mandate showed a focus on traditional core topics in LIS as evidenced by the size of the terms “education”, “library” and the recurrence of terms evoking library education and institutions such as “library education, bibliography, library schools, reference service, aacr, information source, documentation, information science degree programme, online information, lis education, fid education, research, practice, comparative librarianship”. We also observe the presence of terms such as “latin america, caribbean, botswana, india, denmark, london” which seem to point to the geographic locations concerned by some of the publications.

The maps obtained for EICs 2 and 3 are understandably much sparser as their periods of mandate were much shorter. The occurrence threshold of 2 resulted in the elimination of 90% of the terms contained in the 51 titles published by EIC 2, hence only 18 out of the 177 terms extracted from titles were used for mapping.

Figure 2.

Map of terms in titles published during EIC2’s mandate (2014–2017).

In the period covered by EIC 2, the focus of publications seems to remain on the core issues of education in LIS as evidenced by the terms “education, library, librarian, analysis, service” which belong to the same cluster and the terms “academic librarian, practice, case study, development” in a second cluster. Other terms seem focused on issues of “quality, assessment, information literacy” and student and user experience. The presence of the term “philosophy” is explained by a special issue of the journal on “philosophy of information” published in 2017 (issue 33, number 1), unfortunately, not extracted as a multiword by VoSviewer. Table 7 below shows the top 10 terms by order of frequency in this period, all of which appeared on the map.

Figure 3.

Map of terms in titles published during EIC3 (2018–2019).

Like for EIC2, 95% of the terms extracted from 35 titles published by EIC 3 were eliminated by the occurrence threshold of 2, thus only 6 out of 119 terms were left for the clustering process. While this map is not relevant because of the sparse data, it is noticeable to see the terms such as “big data” appear on the map. This could signal a shift of focus to more technologically current topics.

In the next section, we will analyse the maps obtained by splitting the corpus of titles by interval of six years which should yield a less skewed distribution of terms and thus a more balanced representation of the evolution of topics published by this journal.

3.2Evolution of topics by six years interval

Table 9 below shows the top 10 terms in each period ranked by decreasing order of frequency for each period.

The prevalence of terms such as “library, librarian, librarianship, education, library education, lis education/information science education, university, student, experience, practice” reflect the core themes of the journal. Another group of terms on “information science, information, communication technology” appear to reflect the newer strand of publications on more emergent topics which will eventually displace traditional library science in many institutions all over the world.

Below, we will analyse the six maps were produced for each period, each showing the prominent terms that co-occurred more than twice in titles published during that period.

3.2.1Period 1: 1983–88

Figure 4 below displays the prominent terms in clusters. The topics in this first period are solidly centred on the core themes of the journal, i.e. library training as evidenced by the prominence of terms in the cluster “librarianship, librarian, training, information service, special librarian, educational pluralism, challenge, objective”.

Another group of clusters appear to reflect publications on library services to users as evidenced by the presence of the terms like “library, information workforce, user education, history, research guide change” while the blue cluster reflects publications on online search which was emerging at the time as evidenced by the terms “information management, principle, future, online information retrieval, online searching, archive”. The terms “curriculum, handbook, school, origin, comparative librarianship, reading” focus on the actual education materials. The term “denmark” co-occurred with “new technology, teaching, november” in a cluster, thus suggesting publications on new teaching methods deployed in that country at the time. This map of this first period retained most of the topics that appeared on the map by EIC 1 (see Fig. 1 above).

Figure 4.

Map of prominent topics in 1983–1988.

3.2.2Period 2: 1989–94

Figure 5 below displays the prominent terms appearing in clusters which show continued focus on the core topics of the journal such as “information education, information technology, information service, librarian/library, science and technology”. All these topics appeared in EIC 1’s map (Fig. 1 above).

Figure 5.

Map of prominent topics in 1989–1994.

Figure 6.

Map of prominent topics in 1995–2000.

3.2.4Period 4: 2001–06

The first noticeable difference with maps in the previous 3 periods is the sharp drop in the number of published titles for the same interval of years. The journal went from 280 titles in period 1 to 81 titles in period 4 (thus a 72% fall). Hence it is unsurprising that from this date, the maps became more and more sparse. The topics observed for this period continue some of the trends already observed in the preceding periods (see Fig. 7 hereafter). The terms “library, education, information, information literacy, information professional” are still prevalent. Terms reflecting the geographic locations concerned by the publications also appeared prominently: “africa” is part of the “information” cluster, while “botswana” is in the “information professional, information literacy” cluster, “latin america, caribbean” are correctly placed in the same cluster. Both terms appeared in an article with the title “Management education in Latin America and the Caribbean”.

Figure 7.

Map of prominent topics in 2001–2006.

Figure 8.

Map of prominent topics in 2007–2012.

3.2.6Period 6: 2013–2019

This last period covers titles published in the last year of EIC 1’s mandate (2013), those published during EIC 2’s mandate (2014–2017) and the first year of EIC 3’s mandate (2018–2019). Hence, the map obtained continued to show a continuity in the trends observed above.

Figure 9 hereafter shows the traditional topics on Library education and services with terms such as “education, impact, experience, service, “literature review” and “librarian, library, china” “civic engagement, course project, student, hashemite university”.

The presence of the terms “information science, information, knowledge, big data, data” is explained by the publication of a special issue following a conference in Brazil on Big Data and autonomous action in 2018. The presence of the term “online health information” is explained by a series of special issues on health information services published in 2019. This last period is marked by a broadening of focus and a shift from traditional library education to information studies, to big data, digital humanities, social media and online education issues in the information fields.

Figure 9.

Map of prominent topics in 2013–19.