Liu, Y., Rafols, I., & Rousseau, R. (2012). A framework for knowledge integration and diffusion. Journal of Documentation, 68(1), 31-44.

Liu, Y., Rafols, I., & Rousseau, R. (2012). A framework for knowledge integration and diffusion. Journal of Documentation, 68(1), 31-44.

Scientometrics

雖然許多研究區分"interdisciplinary"、"multidisciplinary"、"transdisciplinary" 以及 "crossdisciplinary" (Organisation for Economic Co-operation and Development, 2005)，但在實務上可以發現它們之間具有連貫性，很難加以區別(Barry et al., 2008, pp. 27-8)。因此，本研究以interdisciplinary research(跨學科研究)加以統稱。根據美國國家科學院(National Academy of Science of the USA) (2004, p. 2)將跨學科研究(interdisciplinary research)定義為：為了了解或解決解答某些需要超越單一學科或研究實務領域的問題，團體或個人整合來自兩個或以上學科與特殊知識體系的資訊、資料、技術、工具、觀點、概念以及理論的一種方式。在這個定義中，關鍵的概念是知識整合(knowledge integration)。而所謂的學科(discipline)係指科學組織的結果。學科結構可以應用叢集演算法(cluster algorithm)等統計工具來產生，例如Rosvall and Bergstrom (2008) 和 Leydesdorff and Rafols (2009)等研究；也有像是國際十進位分類法(Universal Decimal Classification, UDC)的哲學結果：或是例如Moya-Anego´n et al., 2004 和 Leydesdorff and Rafols (2009)等有關跨學科性的實務研究上經常利用期刊引文報告(Journal Citation Reports, JCR)的期刊主題分類(Subject Categories)。

Rafols and Meyer (2010)提出一個研究跨學科性的分析架構，他們強調跨學科性的關鍵過程是知識整合(knowledge integration)，兩個重要的觀察面向是多樣性(diversity)和凝聚性(coherence)。多樣性所指的是使用類別的廣度，也就是強調整合的知識有多麼不同，Rafols and Meyer (2010)提出使用參考文獻的參考文獻(references-of-references)的JCR分類來測量。Stirling (2007)認為多樣性是描述一個系統的元素如何被分配到的類別的特性，其概念包含三個層次：一為牽涉的類別數量(Variety)，二是以Simpson指標、Shannon熵、Gini指標或是變異量係數(the coefficient of variation)等各種古典多樣性(Classical diversity)來測量均勻度(evenness)，三是同時涵蓋多樣性的variety、balance和disparity等三個面向的最佳測量，能將類別間的距離與差異考慮進來的Rao-Stirling測量

此處的dij是類別 i 和 j之間的差異度， pi 和 pj 分別是所有項目在類別 i 和 j 上的比例。 α 和 β 則是調整重要性的參數，通常設為1。

凝聚性則是在這個研究上元素彼此間透過主題或類別間相互關連的程度，強調研究中不同體系間如何連貫銜接並形成有意義的群體，是由元素所構成網路的特性，Rafols and Meyer (2010)使用在參考文獻網路中的書目耦合(bibliographic coupling)強度來測量，也就是以參考文獻為節點，參考文獻彼此間的書目耦合值的大小決定節點間連結線的存在與否，並且以平均路徑長度(mean path length)測量凝聚性。

The authors propose that in order to characterise knowledge integration and diffusion of a given issue (the source, for example articles on a topic or by an organisation, etc.), one has to choose a set of elements from the source (the intermediary set, for example references, keywords, etc.). This set can then be classified into categories (cats), thus making it possible to investigate its diversity. The set can also be characterised according to the coherence of a network associated to it.

In a recent article, Rafols and Meyer (2010) presented an analytic framework for the study of interdisciplinarity. The two main factors of this framework are diversity and coherence. These authors stress that the key process characterizing interdisciplinarity is knowledge integration (National Academy of Sciences, 2004; Porter et al. 2006).

According to the National Academy of Sciences (2004, p. 2) of the USA, interdisciplinary research is:

[...] a mode of research by teams or individuals that integrates information, data, techniques, tools, perspectives, concepts and/or theories from two or more disciplines or bodies of specialised knowledge to advance fundamental understanding or to solve problems whose solutions are beyond the scope of a single discipline or area of research practice.

In this definition the key concept is knowledge integration.

Although some researchers make a distinction between the terms “interdisciplinary”, “multidisciplinary”, “transdisciplinary“ and “crossdisciplinary“ research (Organisation for Economic Co-operation and Development, 2005) in empirical studies one finds a continuum that makes it difficult to distinguish among these modes (Barry et al., 2008, pp. 27-8). Hence, we just use the term “interdisciplinary” as a general term, comprising all the latter, as was done in Rafols and Meyer (2010).

When studying interdisciplinarity, the notion of “a discipline” comes logically first. A discipline is the result of the organisation of science (Turner, 2000). The disciplinary structure can be captured using statistical tools, for example by applying a cluster algorithm, as in Rosvall and Bergstrom (2008) and Leydesdorff and Rafols (2009), a philosophical result (as the categories used in the Universal Decimal Classification (UDC)) or a practice-based categorisation, supported by statistics, such as the Subject Categories of the Journal Citation Reports (JCR) (Moya-Anego´n et al., 2004; Leydesdorff and Rafols, 2009).

Diversity refers to the breadth in categories used (Stirling, 2007); coherence to the extent that different elements in the research (categories or topics) are interrelated. The notion of diversity puts the emphasis on how different the incorporated knowledge is, while the notion of coherence emphasizes how different bodies of research are consistently articulated and form a meaningful constellation (Rafols and Meyer, 2010).

In this sense, an increase in diversity reflects the divergence of knowledge integration and diffusion, whereas an increase in coherence reflects their convergence.

In Rafols and Meyer (2010), diversity is measured using the JCR categories of the references-of-references and coherence using the strength of bibliographic coupling in the network of references.

In order to capture diversity and coherence we will consider a framework that consists of three entities:
(1) the source or object of enquiry (often an article or set of articles) used as a representation of an author or group of authors;
(2) an intermediary set (IM) derived from the source; and
(3) a target set, defining the notion we want to study.
These three sets are connected by two mappings: one from the source set to the intermediary set, and one from the intermediary set to the target set.

[...] knowledge integration can be captured as a property of an article (Porter et al., 2007). This leads to the question: how does one describe a relation between an article and the set of all cats?
(1) words used in the article;
(2) words used in the articles in the reference list;
(3) the byline;
(4) the byline of the articles in the reference list;
(5) the reference list;
(6) the reference lists of the articles in the reference list; and
(7) the union of items 5 and 6.

If the cats are disciplines delineated by a set of keywords and IM consists of words then each word is either mapped to itself (if it happens to be a keyword) or to that keyword (or keywords) that are closest to it in meaning (to be determined by a specific algorithm).

If the cats are disciplines and IM consists of journals, then each journal is mapped to the discipline associated with this journal (the case of journals covered by the WoS and the corresponding JCR subject categories is an obvious example).

Contrary to the case of knowledge integration, knowledge diffusion with respect to one article is largely determined by outsiders[1]. ... Instead, we have to determine an intermediary set taking into account the properties of the “audience” or users of the article(s).
This may be (and again no exhaustiveness is claimed):
(1) the set of articles citing the article under consideration, denoted CIT;
(2) the union of CIT and all articles citing an article in CIT (hence including several citation generations, as studied, for example, in Hu et al., 2011);
(3) the set of journals citing the article;
(4) all books citing the article;
(5) the set of persons downloading this article;
(6) the departments where downloading has taken place; and
(7) all web pages linked to the article (if it exists in electronic form).

If the cats are countries or regions and IM consists of citing articles (the case of knowledge diffusion), then each citing article is mapped to those countries appearing in the byline.
For the same set of cats and an IM consisting of books that appear in the reference list, each book is mapped to the country of the publisher.
If IM consists of web pages, each web page is mapped to its country domain name (either removing other domain names; or considering.com,.org, etc. as “regions”).

If the cats are journals and IM consists of citing articles, then each citing article is mapped to the journal in which the citing article has been published.

Diversity is the property of how the elements of a system are apportioned into categories (Stirling, 2007). As one aspect of knowledge integration, diversity is now determined on the image of the cats-mapping. There are three levels on which one may work.
(1) Variety: the number of cats involved, or (maybe better) the relative number of cats (with respect to the total number of cats) involved.
(2) Classical diversity (as the opposite of evenness). This quantity can be measured using a classical evenness measure such as the Simpson index, the Shannon entropy measure, the Gini index or the coefficient of variation (Nijssen et al., 1998).
(3) As explained by Rafols and Meyer (2010) the best approach is to take the three aspects of diversity – i.e. variety, balance and disparity – into account. If a distance or dissimilarity measure exists in cats (and this is assumed in our framework) this suggests using the Rao-Stirling measure, or one of the generalisations that can be derived from Stirling’s (2007) framework [2].

Recall that the Rao-Stirling measure is defined as:

Here, dij denotes the dissimilarity between cat i and cat j, and pi and pj denote the proportions of the total number of items under study in cat i and cat j, respectively. Finally, α and β are parameters that adjust the importance given to small distances (α) and weights (β).

Coherence is the property describing how the elements of a system are related to each other. Hence, coherence is a property of networks.

It is independent from the concept of diversity: diversity reflects the distribution of elements in the IM set into categories; coherence reflects how these elements are related to each other (as measured through cats).

Different network measures may be used to capture coherence, such as the mean path length or the mean dissimilarity between elements (or linkage strength).

Rafols and Meyer (2010), the network nodes consist of the set of references of the original article. These references are the nodes of the network, and the relation studied, determining the existence of links, is bibliographic coupling.

In this way each source can be represented on a two-dimensional graph (coherence versus diversity) as in the Rafols and Meyer (2010) study, where they represented mean-linkage strength versus Stirling measure. In the case where the study of coherence reveals a fragmented structure, the diversity of each of the clusters can then be analysed for each of them.

However, for a group of articles, such as those published by one author, one may calculate an interdisciplinarity value per article and study how this measure changes over time (when new articles are added to the group of articles). Hence the time aspect of the group of articles under study is carried over to a time aspect of the corresponding interdisciplinarity, or more general, knowledge integration measure.

Jensen, P., & Lutkouskaya, K. (2014). The many dimensions of laboratories’ interdisciplinarity. Scientometrics, 98(1), 619-631.

Jensen, P., & Lutkouskaya, K. (2014). The many dimensions of laboratories’ interdisciplinarity. Scientometrics, 98(1), 619-631.

Scientometrics

本研究提出六種指標來測量研究機構的跨學科性。最廣義的來說，跨學科性可以視為不同學科某種程度的整合 (Weingart and Stehr 2000; Porter and Rafols 2009; Marcovich and Shinn 2011; Wagner et al. 2011; Rafols et al. 2012)，為了將這個想法轉換為量化的指標，本研究認為需要考慮三個問題：
1. 如何定義一個學科
2. 在什麼層次達到整合
3. 學科連結需要到達什麼程度

在學科的定義上，本研究提出三種方式：一、因為是分析CNRS的實驗室，自然可採用CNRS的學科組織(disciplinary organization)，包括10個研究所(institutes)以及進一步細分成的40個組(sections)；二、如同其他先前的研究，使用WoS(Web of Science)的224種期刊主題分類(Journal Subject Categories, JSCs)；三、將文件根據共同的參考文獻，以叢集演算法(clustering algorithms)由下往上地(bottom-up)歸類成認知叢集(cognitive clusters)。在整合的層次，則探討實驗室與論文兩個層級。

以實驗室的跨領域程度來說，較簡單的方式可以定義為：

此處的pi是實驗室的論文在期刊主題分類JSC i上的比例。

除了上述的定義之外，本研究還使用的Stirling’s (2007)方法來表現多樣性的三個不同面向：不同類別的數量(variety)、在各類別上的分布均勻程度(balance)、以及表現類別間的差異(disparity) (Porter and Rafols 2009)：

此處的是主題分類JSC i 和 JSC j的相似性，並且此一相似性以cosine測量主題分類間的引用情形得到。(Porter and Rafols 2009).

為了進一步了解實驗室的跨領域多樣性是否在單一論文的認知層次達成，如同上述的情形，計算單一論文的跨領域多樣性時，可以利用下面的方式：

此處的pai是此論文引用的參考文獻在期刊主題分類JSC i上的比例。進一步用實驗室發表的論文考慮實驗室的跨領域程度時，可以將所有論文的跨領域多樣性加以平均，如

此處的 #pap 是實驗室發表的論文數量。

此外，另兩種指標分別是主流引用外的主題分類比例以及不同機構的人員合作占論文全體比率，分別如下所示：

最後一種指標，先以書目耦合(bibliographic coupling) (Kessler 1963)產生論文之間的關連，計算方式如下：

此處的where #common_refsij 是論文 i 和 j 共同引用的參考文獻數量， #refsi 和 #refsj 分別是論文 i 和 j 包含的參考文獻數量。接下來以書目耦合關連建立論文網路，希望在網路上引用文獻相似的論文會聚集形成叢集。因此，接下來Blondel et al. (2008)的演算法，劃分網路成論文的叢集。整個方法可參見Grauwin and Jensen (2011)，結果共劃分成250個叢集。然後以下面的方式計算實驗室在認知叢集上的多樣性

此處的 p_i 和 p_j 分別是實驗室的論文屬於叢集 i 和 j 的比例。

以六種指標計算每一個實驗室的跨學科多樣性後，接下來以主成分分析(Principal Component Analysis, PCA)進行分析，四個主要的成分分別是

1) 實驗室在各種多樣性指標的綜合表現

2) 實驗室連結的學科的認知距離(cognitive distance)

3) 實驗室在實驗室層級或論文層級具有跨學科性

4) 論文發表的期刊具有跨學科的主題分類或是與其他不同機構的實驗室合作。

Interdisciplinarity is as trendy as it is difficult to define. Instead of trying to capture a multidimensional object with a single indicator, we propose six indicators, combining three different operationalizations of a discipline, two levels (article or laboratory) of integration of these disciplines and two measures of interdisciplinary diversity.

Interdisciplinarity means, at the most generic level, some degree of integration of different disciplines (Weingart and Stehr 2000; Porter and Rafols 2009; Marcovich and Shinn 2011; Wagner et al. 2011; Rafols et al. 2012).

To transform this idea into quantitative indicators, we need to answer three questions:

1. How to define a discipline?

2. At what level the integration is achieved?

3. What is the degree of disciplinary linkage achieved?

There are several ways to define a discipline from a scientometrics’ point of view. Since we are dealing with CNRS labs, the most natural would seem to use the disciplinary organization of CNRS in 10 ‘‘institutes’’ and 40 subdisciplinary ‘‘sections’’. A convenient alternative is to use the 224 Journal Subject Categories (JSCs) used by Web of Science (WoS). Finally, instead of using institutionally predefined divisions of science, one could use a more bottom-up definition of ‘‘cognitive clusters’’. To obtain these clusters, we use the roughly 300,000 French articles published between 2007 and 2010 and group them into ‘‘cognitive clusters’’ using clustering algorithms based on shared references.

In this paper, we will use three definitions of ‘‘discipline’’ and two integration levels (laboratory and article) to calculate six partial interdisciplinary indicators.

We adopt Stirling’s (2007) approach to capture the different facets of diversity : ‘variety’, ‘balance’ and ‘disparity’.

‘Variety’ characterizes the number of different categories, ‘balance’ characterizes the evenness of the distribution over these categories and ‘disparity’ characterizes the difference among the categories, usually based on some distance.

A simple indicator of the spread of the disciplines where a laboratory publishes is given by:

where pi is the proportion of articles of the laboratory in JSCi.

As we would like to include the idea of ‘‘distance’’ between disciplines, we calculate the diversity indicator (Stirling 2007; Porter and Rafols 2009) which combines both the spread of the disciplines through the pi and the distance between them.

where sij is the cosine measure of similarity between JSCs i and j. Practically, sij is measured through the citations from publications in JSCsi to publications in JSC j (Porter and Rafols 2009).

To further characterize a lab’s interdisciplinarity, it is useful to introduce an indicator of the interdisciplinarity of single articles, to test whether interdisciplinarity is achieved at this cognitive level.

Specifically, the interdisciplinary diversity of a single article is calculated as:

where pai is the proportion of articles’ references in JSCi.

To quantify the interdisciplinarity of the papers published by a lab, we aggregate the articles’ diversity indicator art_div_corr at the laboratory level by averaging over all the articles published by that laboratory:

where #pap is the number of articles of the lab for which at least one reference was identified.

Then, we choose a threshold to define the most common JSCs for each institute. ... We therefore choose a threshold value of 90 %. ... Then, for each laboratory, we count the percentage of articles outside this 90 % list and normalize by the expected value, i.e. the average value 0.1.

whereare the frequencies of the JSCs that do not belong to the Institute’s JSC main list.

Interdisciplinary collaborations can also be detected by copublications between scientists belonging to different CNRS Institutes. We compute a fifth indicator by calculating the proportion of a lab’s publications that involve authors from other Institutes

where the sum counts the number of articles of the lab involving at least two institutes and
#articles is the total number of articles published by the laboratory.

To build these ‘‘cognitive disciplines’’, we use bibliographic coupling (BC) (Kessler 1963) between the 300,000 papers published by French laboratories in the period 2007–2010 and compiled by the WoS.

where #common_refsij is the number of common references for articles i and j, and #refsi,
#refsj are the numbers of references of articles i and j, respectively.

In comparison to a co-citation link (which is the usual measure of articles’ similarity), BC offers two advantages: it allows to map recent papers (which have not yet been cited) and it deals with all published papers (whether cited or not).

This reinforcement facilitates the partition of the network into meaningful groups of cohesive articles, or clusters. A widely used criterion to measure the quality of a partition is the modularity function (Fortunato and Barthe´lemy 2007), which is roughly is the number of edges ‘inside clusters’ (as opposed to ‘between clusters’), minus the expected number of such edges if the partition were randomly produced. We compute the graph partition using the efficient heuristic algorithm presented in (Blondel et al. 2008). The whole method is described in (Grauwin and Jensen 2011).

Applying this algorithm yields in a partition of French papers into roughly 250 clusters containing more than 100 papers each.

where p_i is and p_j are the proportions of the labs’ papers belonging to clusters i and j respectively.

On average, articles refer to papers from almost 10 different disciplines (9.8 JSC). .... However, when considering those JSC that are used in more than 10 % of the reference list, this average drops to 2.7. This means that, on average, an article spreads its references on 3 main JSCs and 7 additional which benefit from roughly a single reference.

An average laboratory publishes in journals belonging to 34 different JSCs ...

PCA1: combined interdisciplinarity The main axis represents a combination of the various interdisciplinarity indicators.

PCA2: short or long cognitive distance This axis distinguishes those labs that connect distant or nearby disciplines.

PCA3: article or laboratory interdisciplinarity This axis distinguishes labs that achieve interdisciplinarity either at the laboratory or article level.

PCA4: diversity of publications’ JSCs or diversity of collaborations This axis distinguishes labs that publish in journals belonging to different JSCs (high lab_jsc_bal) from labs that co-publish with labs from different CNRS Institutes (high lab_inst_cop_bal).

We have computed the six indicators for the 680 laboratories which have published more than 50 papers over 2007–2010. To allow comparisons and statistical analysis, since the absolute values have no intrinsic meaning, we have scaled all the values to achieve an average value of 0 and a variance of 1. We then carried out a principal component analysis of the (680 9 6) matrix using the free software R (www.r-project.org/). More precisely, we used prcomp from the ‘stats’ package, without any axes rotation.

First, let us note that using the first four PCA axes gives an overall view about the interdisciplinarity practices of each lab. This view has been compared to expert knowledge, namely scientists working in those labs or scientific advisors from CNRS. This comparison, carried out for about 20 different labs from all the disciplines, suggests that these indicators characterize interdisciplinarity
in a meaningful way.

A major drawback of our method is that we cannot distinguish real interdisciplinary collaborations, giving rise to new concepts or to a coherent new scientific field, from simple pluridisciplinary practices that merely juxtapose different disciplines, as when historians use characterizing tools from physics. It seems difficult to learn much about the cognitive dimensions of interdisciplinarity from an automatic analysis of metadata of the papers.

Porter, A. L., & Rafols, I. (2009). Is science becoming more interdisciplinary? Measuring and mapping six research fields over time. Scientometrics, 81(3), 719-745.

Porter, A. L., & Rafols, I. (2009). Is science becoming more interdisciplinary? Measuring and mapping six research fields over time. Scientometrics, 81(3), 719-745.

Scientometrics

本研究將跨學科研究(interdisciplinary  research)操作化的定義為：由團隊或個人從兩個或以上的知識體系(bodies of knowledge)或研究實務整合它們的觀點/概念/理論、工具/技術以及資訊/資料的一種研究模式，也就是這類研究其知識來源具有多樣性，然後分析六個研究領域在1975年和2005年的跨學科程度變化。跨學科指標的計算以引用期刊在WoS (Web of Science)上的主題分類(Subject Categories, SCs)為基礎，並且配合科學映射圖(science maps)表現科學產出在主題分類上的分散情形(dispersion)。整個分析的流程包含五個步驟：

一、將跨學科性的測量操作化。

二、建構主題分類間的相似性矩陣，做為計算整合性指標之用。

三、對相似性矩陣進行因素分析(factor analysis)，將主題分類分群成為巨型學科(macro-disciplines)以便進行視覺化。

四、產生科學映射圖。

五、選取六個主題分類，做為目前的基準與未來探索。

針對操作化跨學科性的測量有幾點必須說明：首先根據Stirling的看法，探索跨學科性時，需要針對引用的學科數量、引用在學科間的分布情形、類別的相似性等面向進行研究[RAFOLS & MEYER, FORTHCOMING]。其次，本研究認為知識整合是一種認知範疇(an epistemic category)，因此跨學科性指標應該建立在研究結果的內容，而不是團隊的成員，部門組織或合作上。最後，跨學科性的測量通常以引用文獻的期刊所屬的主題分類為基礎，但書目計量學的研究社群已經提出主題分類有一些問題，例如期刊叢集的研究指出僅有約50%的叢集結果和主題分類相近[BOYACK & AL., 2005; (BOYACK, personal communication, 14 September 2008)]，根據引用網路得到的分類結果和主題分類之間也沒有很好的符合[LEYDESDORFF, 2006, P. 611]。但這些結果僅對科學映射圖產生有限度的影響，並且在測量整合性上，主題分類目前還是最被廣泛使用的分類資源。

本研究用來測量整合性指標[RAFOLS & MEYER, FORTHCOMING]的公式，由Rao-Stirling提出的多樣性測量方式 [STIRLING, 2007]，如下：

此處pi是給定的論文上引用的參考文獻來自主題分類 i 的比例，sij是主題分類 i 和 j 的相似程度，利用cosine測量。由於許多研究 例如[GRUPP, 1990; HAMILTON & AL., 2005, or ADAMS & AL., 2007]都以Shannon或Herfindhal提出的方式測量整合性，Shannon的多樣性測量方式如

Herfindhal的多樣性測量方式如

但這兩種方法都未考慮類別間的不同；反之Rao-Stirling的多樣性則同時考慮類別數量多寡、類別上的分布平衡和類別間的相似性等三個方面。因此本研究比較此一整合性指標與Shannon和Herfindhal多樣性。

本研究以主題分類被引用的次數為資料，對每一對主題分類進行cosine測量這兩個主題分類間的相似性。當兩個主題分類被大部分的論文共同引用時，它們之間便會有很高的相似度；反之，兩個主題分類共同被引用的情形很稀少時，cosine的值接近於零。完成相似性矩陣的建立後，以主成分分析(Principal Components Analysis, PCA)進行因素分析，以最大變異量轉換(Varimax rotation)產生20個因素，將每一主題分類以其具有最高負荷的因素進行歸類。每一個因素對應一個巨型學科，某些無法歸類的主類分類另外歸於一個巨型學科，結果共有21個巨型學科。

然後以主題分類在21個因素上的負荷值為特徵，再以cosine測量主題分類之間的相似性。以Pajek將主題分類之間的相似性映射成網路圖，過濾相似性在0.6以下的連結線，做為科學映射圖。科學映射圖上呈現每一個主題分類、相對的重要性、以及彼此間的關連程度，目的在於在巨型學科間找出特定研究的主體，發現相互關連在時間上的變化以及主要的跨學科關連，更重要的是發現做為知識來源的期刊是來自於密切關連的學科或是跨越完全不同的領域。

本研究選取生物科技與應用微生物學(Biotechnology & Applied Microbiology)、電子電機工程(Engineering, Electrical & Electronic)、數學(Mathematic)、醫學(Medicine – Research & Experimental)、神經科學(Neurosciences)、物理(Physics – Atomic, Molecular & Chemical)等六個主題分類。

研究結果發現：30年間論文的平均作者數、平均參考文獻數和引用的學科數量都有很大幅度的增加，但是從跨學科指標的增加並不大。造成上述現象，可能是由於雖然引用的主題分類數量有明顯的增加，但每篇論文平均引用的參考文獻數量增加地更快，使得在不同主題上的引用比例的實際改變變得不如預期中的重要；另外，許多主題分類的引用較傾向於鄰近的主題分類，但是鄰近區域的主題分類有較高的相似值，對於多樣性的貢獻較低；最後是某些較跨學科研究的領域其測量的整合性已經到達飽和了。從科學映射圖的結果也指出論文引用的分布仍然主要集中於某些鄰近的學科領域。此外，本研究也發現Rao-Stirling的多樣性測量與Herfindhal和Shannon的測量都有很高的相關性，分別為0.91(標準差0.07)及0.88(標準差0.07) 。

Here we investigate how  the degree of interdisciplinarity has changed between 1975 and 2005 over six research domains. ... The results attest to notable changes in research practices over this 30 year period, namely major increases in number of cited disciplines and references per article (both show about 50% growth), and co-authors per article (about 75% growth). However, the new index of 

interdisciplinarity only shows a modest increase (mostly around 5% growth). Science maps hint 

that this is because the distribution of citations of an article remains mainly within neighboring 

disciplinary areas.

We measure how integrative particular research articles are  based on the association of the journals they cite to corresponding Subject Categories  (“SCs”) of the Web of Science (“WoS”)

And, we present a practical way to map  scientific outputs, again based on dispersion across SCs.

This report operationally defined interdisciplinary  research as: 

x a mode of research by teams or individuals that integrates 

x perspectives/concepts/theories and/or 

x tools/techniques and/or 

x information/data 

x from two or more bodies of knowledge or research practice. 

Our approach here is to investigate changes of degree of interdisciplinarity over time  using various established indicators (e.g. number of disciplines cited, percentage of  citations within-field), together with a new indicator developed the NAKFI evaluation  team [PORTER & AL., 2007]: 

Integration – reflecting the diversity of knowledge sources, as shown by the breadth  of references cited by a paper. 

Following Stirling’s heuristic, we have previously argued that in order to explore interdisciplinarity, one needs to investigate multiple aspects, namely: the number of disciplines cited (variety), the distribution of citations among disciplines (balance), and, crucially, how similar or dissimilar these categories are (disparity) [RAFOLS & MEYER, FORTHCOMING]. 

The computation and visualization of the interdisciplinarity measure has taken five  steps, presented consecutively in this section: 

1. Operationalization of an interdisciplinary measure (the Integration index or disciplinary diversity)
2. Construction of a similarity matrix among Subject Categories that is used to compute the Integration index
3. Grouping via factor analysis of the SCs into macro-disciplines using the similarity matrix as a base to facilitate visualization
4. Generating science maps
5. Selection of a bibliometric sample of 6 SCs, to serve as benchmarks here and in future explorations. 

In other words, since knowledge integration is an epistemic category, indicators of interdisciplinarity should be based on the content of the research outcomes rather than on team membership, departmental affiliations, or collaborations (see illustrations in case studies in RAFOLS & MEYER, 2007). 

The bibliometric community has noted that the SCs have some problems. In journal clustering exercises, only about 50% of clusters were found to be closely aligned with SCs [BOYACK & AL., 2005; (BOYACK, personal communication, 14 September 2008)]. Poor matching between SCs and classifications derived from citation networks has also been reported [LEYDESDORFF,
2006, P. 611], but surprisingly the mismatch only has limited effect on the corresponding science maps [RAFOLS & LEYDESDORFF, UNDER REVIEW].

Nonetheless, the SCs offer the most widely available categorization resource that we could ascertain for the purpose of providing an accessible measure of Integration.

As derived in RAFOLS & MEYER [forthcoming], the formula for the Integration index can be expressed as:

where pi is the proportion of references citing the SC i in a given paper. The summation is taken over the cells of the SC x SC matrix. sij is the cosine measure of similarity between SCs i and j (the cosine measure may be understood as a variation of correlation). Here this matrix sij is based on a US national co-citation sample of 30,261 papers from Web of Science as explained below in detail. 

This Integration measure (aka, Rao-Stirling’s diversity) can be compared with Shannon diversity: 

or with Herfindhal’s diversity (the complement of Herfindahl’s concentration):

The power of the Integration index is that it characterizes interdisciplinarity in terms of the diversity of knowledge sources of papers, using a general formulation of diversity [STIRLING, 2007] rather than an ad hoc indicator.

A number of researchers have used these traditional measures of diversity, such as Shannon or Herfindhal, to measure interdisciplinarity [E.G. GRUPP, 1990; HAMILTON & AL., 2005, or ADAMS & AL., 2007]. These measures do not take into account how different the categories are, whereas our Integration measure reveals increased diversity only when added categories are significantly different.

In particular, a broad national sample of articles from WoS is used to create the sij matrix that underlies the metrics used for computing Integration. First we describe the sample used as a basis for the similarity matrix; second, the construction of the matrix.

We combine six separate weeks of all papers in WoS, with one or more authors having a USA address, sampled during 2005–2007, to obtain 30261 articles. This provides a broadly based, yet manageable base sample. We processed the “Cited References” of these abstract records to identify the “Cited SCs.”

Our sample of 30261 WoS articles contains 1,020,528 cited references (an average of 33.7 per article). Of those, our thesauri link 768,440 to a particular Subject Category. Another 28,000 have been checked and assigned to “not being in an SC.”

For our purposes in addressing cited SCs, the list includes a few more than the current set, for a total of 244 SCs. The sample contains 1,114,930 instances of cited SCs.

The 30261 articles, by 244 SCs, described allow for construction of a co-citation similarity matrix, sij, using Salton cosine [SALTON & MCGILL, 1983; AHLGREN & AL., 2003].

The values of sij are high (i.e. closer to one) when SCs i and j are co-cited by a high proportion of articles that cite one or the other. The cosine value approaches zero when two SCs are rarely cited together.

For various purposes and in particular for visualization, it helps to consolidate the narrow research areas of the ISI SCs into larger categories, which we call “macro-disciplines.”

We base our grouping of SCs on a type of factor analysis – Principal Components Analysis (PCA) – following a similar methodology to that developed by LEYDESDORFF & RAFOLS [2008] to cluster SCs into macro-disciplines.

Within VantagePoint, we constructed the matrix of cosine similarities for the 244 cited SCs by 244 cited SCs described in the previous section. ... We explored various factor analysis solutions, eventually adopting a 20-factor solution (Varimax rotation). ... The 21 macro-disciplines reflect this factor solution.

So, to a considerable degree, named sub-disciplines do not fully coalesce within a single macro-discipline. This warns that the evolving research enterprise does not neatly conform to the traditional scholarly disciplines.

These maps present the SCs, their relative importance in size, and how related they are to each other over all science. The main aim of these science maps is to locate particular bodies of research among the macro-disciplines. ... That can help identify changes in degree of interrelationship over time, and key cross-“disciplinary” relationships that might benefit from nurturing. It should also be informative to see whether knowledge sources of a set of publications are coming from research domains that are closely related (little interdisciplinarity) or that span very disparate domains (high interdisciplinarity). 

We then construct a new Salton cosine similarity matrix among SCs using the loadings of each SC on the 21 factors (as discussed in the previous subsection). This matrix is then uploaded into the network analysis software Pajek [BATAGELJ & MVAR, 2008]. In Pajek, the minimum similarity threshold was arbitrarily set to 0.6 (this choice was found to provide a good readability-to-accuracy trade-off) and the SCs were distributed in a 2-D plane according to their similarities, to obtain a base science map.

Since research collaboration is often (and sometimes mistakenly) associated with interdisciplinarity, we examine measures of co-authorship. ... However, within research domain, the number of authors per paper has escalated remarkably, with about 75% average growth. This increase ranges from 48% in Math and 54% in Physics-AMC to 90% in Neurosciences. 

Before turning to Integration scores, we consider the number of distinct SCs that one article cites. ... Table 2 and Figure 4 show a sturdy increase in the breadth of citing in all six of these research domains (about a 50% growth on average). 

Integration scores are tabulated in Table 2 and shown in Figure 5. We see that over time, there is a modest increase in Integration scores and that math researchers are notably less integrative in their citing patterns. However, math has the highest relative growth (39%) whereas other SCs’ growth ranges from 3% to 14% (5% on average). t-tests between the 1975 and 2005 samples show these differences to be highly significant (<.005 for EE, assuming either equal or unequal variances; all others even more highly significant).

Pearson’s correlation between Integration and Herfindhal takes a mean value of 0.91 (standard deviation = 0.07) and between Integration and Shannon, a mean value of 0.88 (standard deviation = 0.07). These high correlations confirm that Integration is very closely associated with traditional diversity indicators – as could be expected by construction.

The main finding is that Integration scores increase over time, but significantly less so than other indicators, such as percentage of single-authored papers, mean authors per paper, and mean number of disciplines per paper.

First, although the number of cited SCs increases significantly, since the average number of references in a paper also shows a quicker increase (see central columns in Table 2), the actual change in the proportions of citation to different SCs is not as important as could be expected.

Second, as we will show in Figures 7 through 10, the citation patterns of a given SC tend to be with SCs in its vicinity. Since these neighboring SCs have high similarity values with the one investigated, their contribution to Integration (to diversity) is smaller than in other indicators. This means that the Integration score “deflates” the diversity recorded by Shannon or Herfindahl because most of the cited SCs are not very different from the SC doing the citing.

This is much easier to convey using science maps that directly show the three aspects of disciplinary diversity, namely:
1. the variety of “disciplines” (i.e., discrete research areas, the SCs, shown by the number of nodes in the map)
2. the balance, or distribution, of disciplines (relative size of nodes)
3. the disparity, or degree of difference, between the disciplines (distance between the nodes)

These maps were created followed the techniques developed in LEYDESDORFF & RAFOLS [2008], in the context of the current interest in science mapping [MOYA-ANEGON & AL., 2004; BOYACK & AL., 2005; MOYA-ANEGON & AL., 2007]. ... In the figures presented in this article, we only label groups of SCs on the basis of macro-disciplines found by factor analysis, as explained in the methodology. 

However, the perspective provided by the Integration score and the science maps suggests that the practice of interdisciplinarity in citations occurs mainly between neighboring SCs and has undergone a much more modest increase (on average only 5%, excluding math).

This is mainly for two reasons: first, although the number of cited SCs has increased, the growth of citations means that the increase in the proportion of citations to new SCs is small; second, the newly cited SCs tend to be in the vicinity of the previous ones – hence they don’t add as much interdisciplinarity as they would if they were very disparate/distant disciplines. Moreover, for already very interdisciplinary SCs, such as Neuroscience, the indicator may have a certain “saturation” effect. 

Bache, K., Newman, D., & Smyth, P. (2013, August). Text-based measures of document diversity. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 23-31). ACM.

Bache, K., Newman, D., & Smyth, P. (2013, August). Text-based measures of document diversity. In Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 23-31). ACM.

Scientometrics

本研究提出一個利用文本內容測量多樣性的架構，其原理是先以文件語料庫訓練主題模型，然後根據主題在文件上的共現情形，計算主題之間的距離，最後再測量文件的多樣性。這個架構的優點是只需要以文件語料庫的文本資料做為輸入，完全是資料驅動(data-driven)的方法；產生人類可讀的(human-readable)結果；並且能夠廣泛地擴大到作者、學系和期刊等應用。
計算某一個群體的多樣性(diversity)程度，目前已經廣被各種學科關注，例如生態學(ecology) [9]、遺傳學(genetics) [12]、語言學 (linguistics) [8]和社會學(sociology) [5]。在評估群體的多樣性時，通常假定群體內的個體可分為T個類別，每個類別的比例為pi：
Stirling [22]提出多樣性應包含三個面向：類別數目(variety)、比例之間的平衡(balance)以及類別間的差異(disparity)。
1. 群體內包含的類別數目：是相對簡單的多樣性測量方式，也就是π裡非零比例的數量。

2. 比例之間是否相對平衡，常用的測量方式包括Shannon的熵(entropy)，或是變異量。

3. 群體上呈現的類別彼此間的差異。
Stirling [22]並說明這三個面向都是多樣性量化的必要條件，但不是充分條件，而由Rao [18]提出的公式可以將三個面向整合起來，

此處pi和pj分別類別i和j在群體中的比例，δ (i, j) 是類別i和j的距離，Δ是 TXT的距離矩陣，t表示矩陣轉置運算。

Rafols and Porter [14]利用引用文獻的ISI 期刊主題分類(journal subject categorizations)分析6個特定的主題分類在1975到2005年間的跨學科情形。在他們的研究中，是參考文獻發表期刊的主題類別(subject category) i，δ (i, j)定義為1減去主題類別i和j的引用計數向量(citation count vector)的cosine距離。結果發現雖然引用文獻數量與共同作者數有明顯的增加，跨學科性的程度以緩慢的速度增加。這種方法仰賴事先定義的分類而有限制 (Rafols and Porter [15])，因為主題分類可能隨時間改變，而無法反應當時的學科界線。當分析的資料沒有適合的分類方式時，此時也會受到限制。此外，引用資料能否準確反應科學文獻的內容也頗受爭議。

In this context we can define Rao's diversity measure for each document d as

本研究提出一個文本為基礎的研究架構，以文件的內容來測量它的多樣性程度。此一方法利用Latent Dirichlet Allocation (LDA)的主題模型(topic model)方法從文件的語料庫(corpus)推論出T個主題，最後產生出一個DXT的矩陣，D是文件的數量，矩陣上的元素ndj表示文件d上的詞語指定給主題j的比例。利用此一矩陣以Cosine距離計算每一對主題間的距離矩陣，然後結合這些距離和文件上的主題分布，利用Rao [18]提出的公式估算文件的多樣性。

此一方法是完全資料驅動(data driven)，產生容易解讀的結果，並且能夠將其推廣應用於作者、學術部門以及期刊的多樣性測量。

In this paper we present a text-based framework for quantifying how diverse a document is in terms of its content.

The proposed approach learns a topic model over a corpus of documents, and computes a distance matrix between pairs of topics using measures such as topic co-occurrence. These pairwise distance measures are then combined with the distribution of topics within a document to estimate each document's diversity relative to the rest of the corpus.

The method provides several advantages over existing methods. It is fully data-driven, requiring only the text from a corpus of documents as input, it produces human-readable explanations, and it can be generalized to score diversity of other entities such as authors, academic departments, or journals.

The quantification of diversity has been widely studied in areas such as ecology [9], genetics [12], linguistics [8], and sociology [5].

The typical context is where one wishes to measure the diversity of a population, where a population consists of a set of individual elements that have been categorized into T types (such as species), with proportions 

A relatively simple measure of diversity is variety, how many different species are present in a population, or the number of non-zero proportions in π.

One can alternatively measure diversity as a function of the relative balance among the proportions (also referred to as `evenness' in ecology [13] or `concentration' in economics [4]), using measures such as Shannon entropy or variance-based quantities such as 

From a more general perspective, Stirling [22] proposed that there are three distinct aspects to diversity: variety, balance, and disparity.

Disparity is the extent to which the categories that are present are different from each other, based for example on distance within a known taxonomy [21].

Stirling argued that each of these three properties is a necessary (but non-sufficient) component in any quantitative characterization of diversity, arriving at a relatively simple mathematical formulation for diversity, a formulation originally proposed in earlier work by Rao [18]:

where pi, pj are the proportions of category i and j in the population, δ (i, j) is the distance between categories i and j, Δ is a TXT matrix of such distances, and t is the transpose of the TX1 vector of proportions .

The contribution of this present paper is to investigate diversity in the context of text documents, using Rao's measure a starting point.

In particular, we will use words as elements, topics as word categories, and documents as collections (or "populations") of words. Specifically, we address the following task: given a corpus of documents, assign a diversity score to each document, where this diversity score can be used to rank documents from most to least diverse.

Indeed, diversity as defined via co-citation counts is the most widely-used approach to quantify interdisciplinarity in practice, based on the notion that disciplines that are co-cited more often by the same article are "closer" than disciplines that are less frequently co-cited.

Journal subject categories are typically used to capture the notion of a discipline, typically using the manually-defined 244 ISI subject categories from Thomson Reuters, with articles being assigned to a subject category associated with the journal the article is published in (e.g., [15, 14, 17, 23]).

Rafols and Porter [14] used journal subject categorizations of citations to analyze how interdisciplinarity has changed between 1975 and 2005 for six specific subject-categories. They concluded that although the number of citations and co-authors per paper was increasing significantly over time, the degree of interdisciplinarity was increasing at a much slower rate, as reflected by citation patterns between subject categories. As a component in their analysis, Rafols and Porter used Rao's diversity index based on a count matrix of D documents by T categories derived from citations: pi was the proportion of citations made by an article to other articles that were published in journals belonging to subject category i, and δ (i, j) was defined as 1 minus the cosine distance between citation count vectors (across documents) of subject categories i and j.

Our work differs from this earlier work and related threads in scientometrics in two specific ways. First, in our approach the categories and distances, δ (i, j), are learned directly from the text content, rather than being based on manually predefined schema such as the ISI subject categories. ...  The second major difference in our approach is our use of word counts rather than citation counts (which are the basis of most prior work in scientometrics on quantifying interdisciplinarity). 

There are obvious limitations to relying on pre-defined taxonomies, as pointed out by Rafols and Porter [15]. Subject categories can change over time and no longer necessarily reflect current disciplinary boundaries.

In addition, in some contexts such as analysis of proposals and grants, there may be very limited or no categorizations available. For analysis of narrow domains (say the field of data mining and machine learning) existing categorization schemes may be too coarse-grained to be useful. In this context, a corpus-driven approach to learning the categories, such as the topic-based method we describe here, is a useful alternative, and in some cases may be the only option.

We expect that using text content will complement citation-based approaches, as both words and citations carry useful signal. There has long been debate over whether citations accurately reflect the content of a scientific article [2, 1]-- arguably the words in an article may provide a more accurate reflection of the author's intentions than the citations the author uses.

Another field which is related to our current work is that of outlier detection. If we consider documents as being represented by T-dimensional vectors of counts, then one approach to quantifying diversity is to look for documents that are outliers in this T-dimensional space, using a multivariate outlier detection algorithm. ... Equivalently, since the pi are the components of a probability vector in a T - 1 dimensional simplex, we can think of high diversity documents as points that lie in the interior of the simplex (in at least 2 of the dimensions) rather than at the edge.

We use the Latent Dirichlet Allocation (LDA) topic model with collapsed Gibbs sampling to learn T topics for the D documents in the corpus [7]. A single iteration of the collapsed Gibbs sampler consists of iterating through the word tokens in the corpus, sequentially sampling topic assignments for each word token in each document while keeping all other topic-word assignments fixed. Using the topic-word assignments from the final iteration of the Gibbs sampler , we create a DX T document-topic count matrix with entries ndj corresponding to the number of word tokens in document d that are assigned to topic j.

In this context we can define Rao's diversity measure for each document d as

where P(j|d) is the proportion of word tokens in document d that are assigned to topic j and δ (i, j) is a measure of the distance between topic i and topic j. Note that δ (i, j) is constant across all documents, and P(i|d) and P(j|d) vary from document to document.

We presented an approach for quantifying the diversity of individual documents in a corpus based on their text content. Empirical results illustrated the effectiveness of the method on multiple large corpora.

This text-based approach for assigning diversity scores has several potential advantages over previous alternatives, such as methods that define diversity based on citations categorized into predefined journal subject categories. The text-based approach is more data-driven, performing the equivalent of learning journal categories by learning topics from text, and can be run on any collection of text documents, even without a prior categorization scheme.

In addition, it produces human-readable explanations and can be easily generalized to score the diversity of other entities such as authors, departments, or journals (e.g., by aggregating counts across such entities).

Van den Besselaar, P., & Leydesdorff, L. (1996). Mapping change in scientific specialties: A scientometric reconstruction of the development of artificial intelligence. Journal of the American Society for Information Science, 47(6), 415-436.

Van den Besselaar, P., & Leydesdorff, L. (1996). Mapping change in scientific specialties: A scientometric reconstruction of the development of artificial intelligence. Journal of the American Society for Information Science, 47(6), 415-436.

Scientometrics

本研究

科學專業可視為對相同研究問題、使用相同方法並且參考相同的科學文獻來進行工作的人所組成的網路。在同一個專業裡的研究者彼此間的傳播要比和其他社群的人來得多，並且參考彼此的研究工作也比專業外的人的研究工作明顯來得多。因此科學專業與學科可以視為是傳播網路，在此一網路內的傳播比和其他專業研究者的傳播強度大。過去的研究已顯示期刊間的引用可以做為科學的學科組織的操作性指標，屬於同一學科內的期刊的引用關係比和其他期刊更強。在利用期刊之間的引用關係的研究上，可以從關連性分析(relational analysis)，確認彼此間有強連結的節點所形成的派系(cliques)，另一種方式則是根據在網路上有相同位置的節點，來發現網路的結構。

本研究採用的方式是利用因素分析(factor analysis)技術對使用期刊間的引用，進行結構式分析，根據期刊集合的逐漸穩定與一致情形，分析人工智慧(Artificial Intelligence, AI)領域在1982到1992年之間的發展。此一分析從選擇Artificial Intelligence做為核心期刊開始，根據Artificial Intelligence引用其他期刊與被其他期刊引用的情形建立關連環境(relational environment)，進行因素分析，關注傳播網路的結構改變。將Artificial Intelligence有最高因素負荷(the highest factor-loading)的因素做為這個專業的代表，如果選擇的Artificial Intelligence愈是在專業的中心，所得到的分析描述將會愈可信，其它的因素則做為相關的專業。

研究結果發現1982到1988年間，期刊集合不僅相當異質的期刊，同時每段時期的變化也很大，幾乎沒有重複，可見得在80年代初期AI並不能算是一個成熟的專業，然而從1988年起，三、四種期刊穩定地成為AI的核心，此外從期刊標題(journal titles)明顯的相關，可以知道整個期刊集合的凝聚力相當高。


We use aggregated journal-journal citations for describing Artificial Intelligence as sets of journals; factor analytic techniques are used to analyze the development of AI in terms of (an emerging) stability and coherency of the journal-sets during the period 1982-1992.

A scientific specialty can be defined as a network of people working on the same set of research questions, using the same methods and referring to the same scientific literature (see, e.g., Kuhn 1962; Price 1965; cf. Chubin, 1983).

Researchers within a specialty communicate more with one another than with researchers in other communities, and they are expected to refer to one another’s work significantly more frequently than to the work of outsiders.

Scientific specialties and disciplines can thus be considered as communication-networks (e.g., Griffith & Mullins, 1972; Shrum and Mullins, 1988). The communication within such a network is much more intensive than the communication with researchers in other specialties (e.g., McCain, 1990).

In a number of studies, it has been shown that journal-journal citations can be used as an operational indicator for the disciplinary organization of the sciences (e.g., Doreian & Fararo, 1985; Borgman & Rice, 1992; Tijssen, 1992; Cozzens & LeydesdorE, 1993). One may expect strong citation relations within and between journals belonging to a discipline, and (much) weaker relations with other journals.

In general. one can analyze communication networks in two different ways: First, the links between the nodes in the network can be analyzed in terms of graphs. In such a relational analysis, the nodes with strong (mutual) links are considered as so called “cliques.” Second, the latent structure of the network can be studied in terms of so called “eigen-vectors” (cf. Studer & Chubin, 1980). This structural analysis reveals which nodes have similar positions in the network.

In this study, we use citation relations among journals for the delination of the relevant domains, but we use the positional or structural approach for the analysis of development patterns.

We use factor-analysis for detecting the structural characteristics of each communication network. Journals that belong to the same specialty are expected to have a similar position within this communication system, that is, resemble each other in terms of their citation behavior. Thus, we obtain an operational definition of scientific specialties, which can be investigated in terms of coherency (in each year), stability and change (over time).

If the journal-set remains relatively stable over time, this suggests that the specialty is more mature and established.

Finally, a major scientific development like a change of paradigms, may be indicated by a radical rearrangement of the relevant journal-set in relation to other subsets. In other words, the question whether a field of research is becoming a coherent specialty or not can be reformulated as the question whether the field under study can be represented by a stable set of journals.

Our methods provide us with criteria to distinguish between gradual change in terms of journals emerging within existing structural units, and structural change in terms of change at the level of groups of journals (cf. Leydesdorff, Cozzens, & Van den Besselaar, 1994).

The analysis begins with a relational analysis of a core journal of the field under study. The entrance journal determines the domain for the analysis. Therefore, the results rely heavily on this choice: The more central this journal is in the journal-set representing the specialty, the more reliable the resulting delineations of the groupings are expected to be (Leydesdorff & Cozzens, 1993).

The matrices which represent the relational communication network of Artificial Intelligence, are factor-analyzed, both in the “citing” and in the “cited” dimension.

The “citing” dimension refers to the aggregated citing behavior of authors who publish in these journals in the current year. We interpret this as the active reproduction of the structure of the specialty. The authors involved have an opportunity to contribute to a change of the position of the journal in which they publish, by relating to other literatures.

The potential change in position is beyond the control of an individual author, and thus we are looking at structural change and not at behavioral intentions (Leydesdorff, 1993, 1995 ) .

In the “cited” dimension, on the other hand, the history of the journal is reflected, and therefore, the “cited” dimension reflects the codification into the archive of the scientific literature.

Thus, the latter is expected to develop more slowly than the “citing” dimension, and citations may have different meanings in different contexts (Small, 1978; cf. Leydesdorff & Amsterdamska, 1990).

The factor on which the entrance journal (i.e., Artificial Intelligence) has the highest factor-loading, is considered as representing the specialty which we try to delineate (i.e., “artificial intelligence”). The other factors that result from the analysis can partly be interpreted as specialties that are relevant for, and related to, the focal specialty.

Between 1982 and 1988, the journal-sets in Table 3 are, apart from the entrance journal, even completely disjointed. ... These sets do not only change, they also consist of heterogeneous journals. Using this criterion, AI cannot be considered as a mature specialty in the first part of the eighties.

However, starting in 1988, a set of three or four journals seems to become a stable core of the AI-cluster. ... Additionally, the journal-set becomes more coherent, in the sense that the journal titles are substantively related.

Our analysis suggests that the subject matter, as well as the methods of AI are gradually becoming more focused and accepted in a specific community of researchers. AI is increasingly covering a delineatable field, and the early speculative claims of AI now seem to be translated into more “normal problem solving” activities. This may support the view that AI is entering a “paradigmatic phase.”

Table 4 exhibits the factor designations in the relational citation environment of Artificial Intelligence in each year. Each factor is interpreted as representing a scientific specialty. Which specialty, is determined by the nature of the subjects of the journals that have their highest loading on this factor.

Among the different specialties the “pattern analysis” factor is most prominently present in all years. It is also rather stable in its composition.

Additionally, one can see in most years a “theoretical computer science” factor, and an “applied computer science” factor, consisting of journals on information systems and software engineering. 

The factor on man-machine studies is visible in all years except 1992. In this year, the International Journal of Man-Machine Studies loads highest on the AI-factor.

“Cognitive psychology” and “cognitive science” are in the environment of AI in most of the years, but in 1982 and 1988 the cognitive science journals and psychology journals are part of the AI-factor itself. 

The predominance of “computer science” clusters in the environment of AI contradicts the view of AI as an (interdisciplinary) set of specialties, all studying intelligence and computational models of intelligence. This result corresponds to the conclusion in an article of Khawarn ( 1991) who, using a different methodology, demonstrated that AI is not an interdisciplinary field, but mainly related to (computer) science. 

Although in every year we see some of these specialties, in the more recent years they are more numerous. This observation suggests that AI-techniques are more and more used in other specialties. The opinions of Schank and Winograd (in Bobrow & Hayes, 1985) that AI might be mainly a methodology, are supported by this result. 

In this article, we defined scientific specialties as communication networks that can be delineated through the analysis of aggregated journal-journal citations. We used a method for mapping the development of scientific specialties and of their respective relations with other specialties over time.

The journal cluster of AI is unstable in a changing environment. But there seems to have been a change in the structure of this communication network between 1986 and 1988.
Before this date, AI has not been an “interreading community of scientists”; AI was presumably in a preparadigmatic phase at the time.
However, in the second part of the decade, research in AI seems to be more focused, and with hindsight this indicates that AI is entering a new phase. The structure of the field becomes more pronounced and stable: The AI-cluster becomes clearly delineatable, and this delineation is stabilized after 1988.

Furthermore, the factor analytical approach (as contrasted with clustering techniques) allowed us to consider journals to be part of more than one journal-set, and therefore it enabled us to study these specialties during phases of emergence and structural change, and within the context of already more established specialties.

In other words, our results do not support Boden’s view that AI has remained a relatively unstructured field of philosophical questions. First, AI has since then become more focused, and both expert systems and the more theoretical aspects of AI are increasingly represented in this core. Expert systems research has become a structural element of AI-research. Robotics has also become more closely related to AI during the later part of the 1980s. Thus, our results contradict the prediction that relatively successful subfields of AI are systematically institutionalized as separate specialties.

On the other hand, Sloman’s view of AI as cognitive science is also not supported by our results. Journals with “cognitive science” in the title appear in the AI-cluster from time to time. These journals, however, have mainly bridging functions with specialties in the environment of AI. More often, psychology and cognitive science play a role as separate specialties in the citation environment of AI.

The view that AI itself is a collection of specialties, all studying “intelligence,” is also not supported by our data. Many of the specialties that study intelligence, like philosophy, linguistics, neural sciences. cultural sciences, are hardly visible in the communication network surrounding AI. Only occasionally. we see isolated philosophical or linguistic journals having citation relations with AI-journals.

The specialties that are visible in the environment of AI in all or most years are pattern analysis, computer science, and cognitive psychology. As noted, the presence of robotics is becoming more pronounced. The view of Schank and Winograd, i.e., that AI is mainly a methodology. therefore seems to be supported by our indicators. Notably in more recent years, the number of specialties that use AI-techniques increases significantly: Biology, medicine, operations research, statistics, geology, chemistry. etc.