A cross-language co-word network comparison of Buddhist literature in Digital Library and Museum of Buddhist study

Introduction
This paper reports a co-words domain analysis of Buddhism literature collected by DLMBS (Digital Library and Museum of Buddhist studies) at National Taiwan University. Established in 1995, the DLMBS is one of the most comprehensive online repository of Buddhist research materials. It currently contains over 400 thousand records of books, research papers, theses and dissertations in 45 languages and digitized Buddhist scriptures. A controlled vocabulary, which is in five languages, including Chinese, English, Japanese, German, and French, was used to help users search DLMBS’s bibliographic database. By using cooccurrence data of author assigned keywords in the bibliographic records, this study attempts to generate co-word networks in three different languages, Chinese, English, and Japanese, to compare regional focuses on Buddhist studies.

Co-words analysis has been shown to be effective in mapping the intellectual structure of disciplines (He, 1999; Leydesdroff, 1989). While it has been widely used in the domains of sciences and technologies (e.g. Buitelaar, Bordea, & Coughlan, 2014; Ding, Chowdhury, & Foo, 2001; Bhattacharya & Basu,1998; Looze, & Lemarie, 1997; Peters & van Raan, 1993a, 1993b; Courtial, 1994; Tijssen, 1992; Callon, Courtial, & Laville, 1991; Rip & Courtial, 1984), to the best our knowledge, it has so far not been applied to humanities. Part of the advantage of using co-word in sciences and technologies is the highly codified subject languages, therefore a higher degree of consistency between concepts and terms in these fields. We believe that a cross-language co-word analysis of Buddhist studies literature would be a worthwhile endeavor for a couple of reasons: Firstly, it has been pointed out that there has been a wide variety of methods, perspectives and subject matters within the international communities of Buddhist studies. The heterogeneity of its scholarships can be partly traced to their geographic roots (Cabezón, 1995). It is therefore interesting to empirically study whether and how the intellectual structures reflected in the published literatures in these language communities differ from one another. A comparison of the intellectual structures can shed light on knowledge interests shared and distinct in these three language communities. From the methodological plain, co-word analysis also provides a viable alternative to citation-based network analysis in humanities where the citation structure is known to be much sparser than in sciences and technologies.

Procedures and analysis
Three separate co-word networks were generated in three different languages where nodes denote the keywords and edges the strength of their cooccurrence. Unlike in most of the previous co-words analysis, where keywords were extracted from titles and abstracts, author assigned keywords were used here as it is believed that they are more representative to the content of the articles and tend to have higher degree of consistency than keywords in free-text. For monographs and other types of publications, the subject-headings assigned by human indexers were used as keywords. Edge weights were normalized by both the inclusion and the Jaccard index (Courtial,1986; Callon, Law, & Rip,1986).

Thus three word similarity matrixes were generated so social network analytical methods such as cohesion, centrality and community-detection (Blondel et.al., 2008) could be performed with a view to exploring the social and cognitive structure of Buddhist studies manifested in its published literature in respective languages. Specifically, the study seeks to answer the following interrelated research questions: firstly, are there recognizable branches or specialties in Buddhist studies? If so, what might these areas of research be? Cross-languages comparisons were also made to examine the similarities and differences of the intellectual structure in different language communities.

Results
Descriptive data

Table 1 gives the numbers of items pertinent to various publication types analyzed in three languages in DLMB.

Language

Publication Type '

Chinese

English

Japanese

Journal Article

45,025

18,703

33,311

Book

12,644

13,604

8,869

Thesis and Dissertation

3,757

1,894

49

Research Paper

3,372

200

663

Proceeding Article

2,409

248

84

Journal Article; Book Review

158

2,201

377

Sound Recording

265

516

16

Serial

427

240

96

Reference Book

393

139

39

Audiovisual

45

304

16

Book Review

49

225

22

Internet Resource

59

207

0

Collected Papers

65

15

83

Others

15

33

0

E-Book

2

21

0

Book; Internet Resource

0

7

0

Book; Sound Recording

1

0

0

Internet Resource; Book Review

2

0

0

Internet Resource; Journal Article

0

2

0

Book Review; Internet Resource

0

1

0

Total

68,688

38,560

43,625

Table 1. Types of publications analyzed

Due to the enormous size of the networks, some sorts of filtering are required to make the groupings intelligible; node degrees was used as the filter as many of the little connected nodes tend to generate noises that impairs meaningful interpretation. To determine the proper threshold of node degree, one needs to consider three criteria: the quality of the clustering, the interpretability of the individual clusters, and the preservation of information. A high threshold would filter out large amount of nodes hence the greater loss of information and low modularity values. On the other hand, a low threshold would result in difficulty in interpreting individual clusters as they tend to lump together heterogeneous topics. Thus a trade-off needs to be made. We approached this matter by performing modularity analysis at different threshold levels so the values of their modularity, the resulting number of communities, as well as the size of the networks could be recorded. The following heuristics were used to select the proper thresholds: to preserve about 10

percent of the total nodes, to limit the number of communities from 10 to 20, and to preserve a high degree of modularity, which is commonly used as the

Figure(1-3). Node degree thresholds and resulting network attributes in three languages

Table 2 reports the thresholds and their corresponding attributes of the resulting networks.

# of nodes

# of edge

Threshold

# of

Communities

% of nodes

%of

edges

Chinese

58,808

787,682

20

17

10.66

54.61

English

34,093

325,161

30

11

9.66

42.74

Japanese

75,728

859,469

40

20

10.09

33.04

Table 2. Descriptive statistics of the three networks

After filtering out lesser connected nodes, modularity maximizing community detection method was then performed to identify the subdomains in each language network. A two-stage approach was adopted here. As some of the communities resulting from the first-round of clustering can still be very broad and heterogeneous, a second modularity analysis was performed on these relative “super” clusters (i.e. clusters with more than 400 nodes), the joint results produce a two-levels hierarchical structure. Three experts in Buddhist studies were then interviewed to help us interpret the clusters (See Figure 4 and 5).

and history of TfiilSafi'Bifffl S«fltlana of Tibetan Buddhisn

Figure 4. Visualization of intellectual structure in Chinse Buddhist studies.

Figure 5. Visualization of intellectual structure in Japanese Buddhist studies.

In this study we utilized co-words networks to visually represent the domain of Buddhist studies. A heuristic was proposed to help select the proper threshold in order to filter less significant keywords. A two-stage clustering approach was adopted, which arguably provides a finer representation of the intellectual structure of the domain. Further analysis will be done, with the help of domain experts, to compare the differences in the intellectual structure reflects in three language communities.

Bibliography
Bhattacharya, S. and Basu, P. (1998), “Mapping a research area at the micro level using co-word analysis”, Scientometrics, 43(3), pp. 359-372.

Blondel, V., Guillaume, J., Lambiotte, R. and Lefebvre, E. (2008), “Fast unfolding of communities in large networks”, Journal of Statistical Mechanics: Theory and Experiment, 2008(10), pp. 10008.

Buitelaar, P., Bordea, G., & Coughlan, B. (2014), “Hot Topics and Schisms in NLP: Community and Trend Analysis with Saffron on ACL and LREC Proceedings”. In 9th Edition of Language Resources and Evaluation Conference (LREC2014).

Cabezón, J. I. (1995), “Buddhist Studies as a Discipline and the Role of Theory”, Journal of the International Association of Buddhist Studies, 18(2), pp. 231-268.

Callon, M., Courtial, J. and Laville, F. (1991), “Coword analysis as a tool for describing the network of interactions between basic and technological research: The case of polymer chemistry”, Scientometrics, 22(1), pp. 155-205.

Callon, M., Law, J., & Rip, A. (1986), “Qualitative scientometrics”, in Mapping the dynamics of science and technology, Palgrave Macmillan UK, pp. 103123.

Courtial, J. (1994), “A coword analysis of scientometrics”. Scientometrics, 31(3), pp. 251-260.

Courtial, J. P. (1986), “Technical issues and developments in methodology”, in Mapping the Dynamics of Science and Technology, Palgrave Macmillan UK, pp. 189-210.

Ding, Y., Chowdhury, G. G., & Foo, S. (2001), “Biblio-metric cartography of information retrieval research by using co-word analysis”. Information processing & management, 37(6), pp. 817-842.

He, Q. (1999), “Knowledge discovery through co-word analysis”, Library trends, 48(1), pp. 133-133.

Leydesdroff, L. (1989), “Words and co-words as indicators of intellectual organization”, Research policy, 18(4), pp. 209-223.

Peters, H. P. F., & van Raan, A. F. (1993), “Co-word-based science maps of chemical engineering. Part I: Representations by direct multidimensional scaling”, Research Policy, 22(1), pp. 23-45.

Tijssen, R. J. (1992), “A quantitative assessment of interdisciplinary structures in science and technology: co-classification analysis of energy research”, Research policy, 21(1), pp. 27-44.