A survey of Chinese interpreting studies: who influences who …and why?

Research questions

Expanding on the major questions outlined at the beginning of this paper, three specific research questions were developed to ascertain how CIS scholars interact with each other and how influence is defined in the community. The following section outlines the authors’ rationale for investigating each research question.

1. How do CIS authors interact with one another? Do they form defined communities? If so, what are the features of those communities?

The advancement of a science relies heavily on its participants’ communicating and collaborating with one another: scholars build on each other’s research, and work together to address common issues or to replicate colleagues’ experiments under different conditions to investigate whether their conclusions can be extrapolated to a larger population. In the context of citation analysis, identifying community structures can help us understand the predominant research themes in a given field and how certain subject matters grow or decline in popularity over time.

2. In terms of citations, who are the most influential scholars in the CIS literature?

Nederhof (2006) observed significant differences in citation behavior between natural and social sciences: members of the former communities (physics, chemistry, etc.) tended to influence each other across geographical boundaries, whereas those of the latter (sociology, linguistics, etc.) generally had very limited influence beyond the countries in which they lived. One might think Interpreting Studies would be an exception to this rule: because of its focus on the interactions between languages and cultures, its authors might reasonably be expected to exert influence across the boundaries of language, geography and culture. Identifying the most influential scholars in CIS ought to reveal whether any Western researchers have an impact on the Chinese field and if so why; it should also prove useful for identifying differences between the most influential Western and Chinese authors’ backgrounds, and for exploring the dominant schools of thought in CIS.

3. Are there any research topics that influential scholars tend to write about? What themes and keywords correlate with author influence?

All researchers would like to see their papers frequently cited by others, and hope that their colleagues might be inspired to pursue their work and address any questions that may remain unanswered. However, the reality is that few articles published are highly influential: a far higher number are rarely read or cited by others. Studying what makes an article influential is a useful exercise from three perspectives: firstly, every scholar would like to make a mark within his research community, and so would do well to know what makes for a successful paper; secondly, because the bodies which allocate grants and other forms of funding always want to be assured that their investments are money well spent (generally speaking, scholars need to publish when they receive a grant); and lastly, studying these predictors of influence can help to identify the hottest topics in the field. A handful of researchers have already explored the issue of what makes one article influential and another not. Buela-Casal et al. (2009) examined the relative influence of theoretical and empirical papers in three Spanish psychology journals, concluding that the former type received twice as many citations as the latter. Haslam et al. (2008) studied the citation data for 308 articles in social-personality psychology, and found the following factors particularly strong for predicting an article’s impact: (1) the reputation of the first-listed author; (2) the presence of a senior colleague’s name among a new author’s collaborators; and (3) a journal’s ranking. The present study is the first time that the predictors of influence in CIS have been analyzed.

Data collection and organization

Given the paucity of coverage of CIS citation data in existing academic databases, for the present study a near-comprehensive database of 59,303 citations was built from scratch—they represent citations from the 1,289 Chinese MA theses, 32 doctoral dissertations and 2,909 research papers available to the authors. These three bodies of literature, chosen because they best represent the overall state of CIS, were accessed through multiple channels: field trips to university libraries, interlibrary loans, book purchases, and academic databases such as CNKI, Wanfang and the National Digital Library of Theses and Dissertations in Taiwan—thus ensuring wide coverage of academic works in both mainland China and Taiwan. Once collected, the references were manually entered into Excel Spreadsheets, using the idea behind Structured Query Language (SQL) to managing data. The method employs multiple interactive and cross-referenced tables; in the present case there are three such: Documents, Authors, and Citations. For these tables to interact with one another it is important to have unique ‘keys,’ elements that allow each row to be identified: a unique, consistent ‘author key,’ for example, enables us to know whether a particular person in the Authors table is the same as one found in a row of the Documents table.

The key concept behind the present analysis is that of the citation network: the documents are nodes in this network, with arrows interconnecting them when authors cite their predecessors (see Fig. 1). The nodes are labeled with various attributes such as ‘author,’ ‘publication year,’ and ‘keywords.’ The number of nodes associated with an author is the number of documents he has produced. Authors share a node when they have co-authored a document. Arrows are in the direction of the citation, so the cited work is at the receiving end of the arrow. When arrows are used to indicate the presence of a citation between documents, there can be at most one arrow between each pair of nodes; when they are used to indicate a type of citation, there can be multiple arrows, as one document may cite another multiple times. A total of ten Excel Spreadsheets were compiled: three documents table for CIS doctoral dissertations, MA theses and academic papers; six citation tables—CIS doctoral dissertations (English and Chinese citations), MA theses (English and Chinese citations), and academic papers (English and Chinese citations), and one author table. This data format makes it easy to be exported into Tulip for data visualization (see section ‘Author interaction’) and ‘R’ for correlational analysis (see section ‘Research topics and academic influence’). The raw data alongside the source codes has been deposited into GitHub, and can be accessed from this link: https://github.com/danielxu85/CIS.

Description of topic labeling method

Rather than rely solely on the keywords provided by the papers’ authors, the content of each and every publication in the data-set was carefully analyzed with the aim of generating keywords that best captured the topics they covered. The keywords typically chosen by authors can often be too general or specific and mask hidden trends. The keywords tagged for the present study were further grouped into six broad themes: Training, Professional, Language, Socio-cultural, Cognitive and Miscellaneous. This classification system was adapted from the coding scheme used in Gile’s (2000) study—it covers all the issues addressed in CIS and minimizes overlap between categories. It should be noted, though, that each paper under examination may be tagged with multiple themes depending on its content. Consolidating keywords into themed categories can help identify major trends in CIS which might otherwise have been drowned out by the sheer number of keywords generated.

Author interaction

Using near-exhaustive citation data, the present authors wished to determine whether the communities of CIS are best classified according to the hypotheses put forward by earlier scholars (Moser-Mercer, 1994; Gile, 2005; Gile, 2013b). In the first two of these studies, it was noted that there existed two dominant and opposing camps—those who approached research from a liberal arts standpoint and those who leaned towards natural/empirical science—and that there was little communication between the two. In his later study, Gile refined his original classification by suggesting that Translation & Interpreting Studies can be divided into three broad groupings—Human Sciences Culture, Canonical Scientific Culture, and Practice, Reflection, Generalization.

To address this aspect of CIS, visual representations of its citation networks were created. A number of open source software applications are available for analyzing and visualizing networks: some, such as SciMAT and VantagePoint, are well suited for analysis, while others (e.g. Pajek and Gephi) are better for visualization. Tulip, a program designed for analyzing and visualizing relational data, provides a good balance because it incorporates a number of layout and clustering algorithms, in addition to network metrics. More importantly, Tulip was selected over other applications for its capacity to visualize large complex networks—the complete citation network used in the present study contains over 12,000 nodes and in excess of 50,000 edges.

Tulip enabled us to import the entire CIS citation network data in the form of a list of all edges (connections between nodes) in the data-set. Once this was done it gave the option to apply a number of layout algorithms. In the present instance, once the entire CIS citation network had been imported, we selected a force-directed algorithm—Fast Multipole Multilevel Method (FM³)—to lay out the graphs presented in the images seen in Figs. 3 and 4 (Hachul & Jünger, 2005). This algorithm groups nodes that are multiply connected to each other closer together in the layout while distancing ones that are not directly connected. This approach makes the produced images ideally suited for visually detecting community structures in networks.

A panoramic graph of the research interactions between various CIS authors was created using the 2012 citation data. In this graph, vertices represent authors and the edges represent the number of citations between them. To generate a full network image with Tulip software, a placement algorithm was used to lay out the nodes. The edges were rendered invisible to ensure that the nodes could be seen clearly. The node color was set to blue using the property option available in Tulip.

To study the relationships between the most influential nodes in the network, we firstly filtered out the less important ones, retaining only the top 150, then calculated the PageRank score of each that remained. The PageRank algorithm for citation networks measures the importance of an author by gauging the quality and quantity of other authors that cite him. The author whose PageRank is being calculated (‘the target’) is said to be at a distance of 0 from himself, and each author who cites him at a distance of 1. The importance of these Distance 1 authors is in turn calculated by the quality and quantity of authors citing them, who are at a distance of 2 from the target. The underlying idea is that influential authors will be much cited by other influential authors, while non-influential authors will either not be cited or only be cited by other non-influential authors. The process continues to a pre-determined parameter known as the damping factor. In line with the majority of citation studies, which use an average chain length of 10 to 15, for the present study a damping factor of 0.15 was used.

PageRanks were calculated using the Algorithm-Measure-Graph-PageRank option in Tulip. The top-ranking 150 nodes were selected, and an induced sub-graph⁶ was created which showed only these 150. Again, the FM³ algorithm was used to lay out the nodes, and Edge Bundling was used to merge edges which were close to one another to make the layout more readable. Tulip automatically colors nodes as soon as any metric has been calculated. In the present instance, the nodes were colored using a gradient scale from blue (high) to orange (low) values. Since it was not possible to visualize the names of all the authors, we chose to display only nodes with a high value of any metric, which in this case was PageRank.

In addition to visually identifying the communities within CIS, a quantitative analysis was conducted to verify whether the network had two or more clusters. Community detection pinpoints the most natural groupings of individuals present in a network (Schaeffer, 2007). There are a number of ways to evaluate the quality of such groupings, one of the most widely used in recent times being Q Modularity, which was introduced by Newman in 2004.

Modularity is defined as the percentage of all connections that fall within a community, minus the expected percentage of connections in the community. The expected percentage is based on the assumption that connections are distributed completely at random, with no regard for community structure. The modularity cut algorithm determines how to partition a graph into communities with high modularity scores, hence the communities it identifies all contain a higher number of intra-community connections than might be expected to occur purely at random. Modularity values lie between 0 and 1; higher values are desirable and represent better clusterings. Typical values lie in the range 0.3 to 0.7 (Newman & Girvan, 2004), values between 0.0 and 0.2 suggesting that the graph is entirely random with no known community structures (Rényi & Erdős, 1959).

Seven commonly used network clustering algorithms were used to determine the number of clusters and optimize the process of clustering CIS citation data; Spin-glass stood out as the best at defining CIS communities. Modularity values, which were obtained for different numbers of clusters using the Spin-glass algorithm, indicate the optimum number of clusters for CIS.

Finally, to analyze the common features of the discrete communities of CIS, the image of clusters was generated. The Spin-glass algorithm was applied with seven clusters. Subgraphs of each cluster were generated with a unique color for each different cluster. FM³ was used to layout the nodes of each cluster individually by executing the layout algorithm on each of the subgraphs. Again, the edges were rendered invisible to make the graphs more readable, and nodes were labeled with author names. PageRank was employed to identify important nodes and labels were displayed only for the most important from each cluster.

It should be emphasized here that modularity cut does not involve the use of meta-information about authors to determine how they are divided into communities. Previous researchers in SNA have not attempted to use meta-data about each citation to generalize the features of each community, because a variety of factors can drive authors into a certain community: a connection can be established between two authors because they have co-authored an article together, or because they have cited one another’s research, a process known as co-author citation (Newman, 2001). In addition, meta-information, such as the content of each cited paper and background information regarding each cited author, cannot be obtained from commercial databases, which means manual labor is required to obtain and screen thousands of papers. Newman & Girvan (2004) applied the modularity cut algorithm to identify communities based on co-authorship data: authors were added to the network as vertices, edges between them indicating their co-authorship of one or more papers in the data-set. Newman and Girvan were primarily interested in investigating whether researchers from the same community were acquainted with one another. Takeda & Kajikawa (2010) analyzed citation data in the fields of energy and material science by tracking modularity scores obtained from each clustering iteration, but they stopped short at summarizing the features that define each community: clustering only indicates that the nodes in each group are similar, but the similarity is dependent upon whether the nodes are connected, not why they are connected.

Most influential scholars

All the Western and Chinese authors appearing in the citation data were ranked according to their degree centrality (DC)⁷ and weighted degree centrality (WDC)⁸ measures, and to their PageRank Algorithm (PRA) scores. DC and WDC are first-order centrality measures, while PRA is a higher-order measure. The three highlight different aspects of an author’s influence within a network: for example, out-degree and weighted out-degree measures indicate how well authors disseminate information about the work of other scholars, while PRA assigns them scores based on how well connected they are with influential colleagues.

It was expected that some authors, despite publishing prolifically, would be found to have little research impact within the CIS community, while others, despite publishing very little, may nonetheless be widely cited. There were also grounds to expect that the ways in which influential authors are cited may be subject to variation: they might be highly influential thanks to a wide range of much-cited work, or their influence might depend on a small number of seminal works. In addition, the present author ventured to predict that, despite the homogenous background of CIS researchers (Zhang, 2008), authors from other disciplines may have topped the influence polls.

Research topics and academic influence

There are various methods for measuring an author’s influence, such as DC, WDC and PRA (see Question 2 above), but for this analysis only in-degree centrality, out-degree centrality, PRA and EigenVector centrality were used. WDC was excluded because it and DC are both first-order centrality measures: they essentially measure the same aspect of an author’s influence. The present study’s database contains 2,909 research articles, 1,289 MA theses and 32 doctoral dissertations; from these were extracted 978 unique keywords to describe their contents. All the keywords for each author in the database were tallied up, and a keyword profile created for each, representing the relative frequency with which he or she used a keyword, normalized to 1, i.e., the number of times that keyword was used divided by the total number of keywords he or she used—hence a keyword with a value of 0.3 represents 30% of all keywords used by that author. Normalization was added in to prevent there being undue emphasis on the connection between keywords and centrality measures when identifying prolific authors.

Such a large number of keywords has two limitations. Firstly, many of them are synonyms or hypernyms, leading to conceptual overlap. Second, in any regression the larger the number of explanatory variables, the more data they require in order to maintain statistical power—the ability to detect significant relationships between explanatory variables and the response (see for example Hsieh, Bloch & Larsen, 1998). In the present case the explanatory variables, namely the relative frequencies of the almost 1,000 keywords, were too numerous in relation to the number of documents for the ‘bulk’ statistics to yield good results. To avoid this sizable stumbling-block, the keywords were classified into six themed groups: Cognitive, Language, Professional, Socio-cultural, Training and Miscellaneous. The categories were designed to be non-overlapping so as to allow for the drawing-out of meaningful trends which would otherwise be undetectable amid the crowd of keywords present in the documents. A supplementary analysis of the 978 keywords, giving a more fine-grained picture of the data, is reported at the end of this section.

The authors’ network measures (in-degree centrality, out-degree centrality, PRA and EigenVector centrality) were first matched to their theme profiles, which came from separate databases: journal articles, MA theses, and PhD dissertations. Of the 2,277 journal article authors, 1,023 were matched in the network database; of the 1,289 MA theses authors, 1,092 were matched; and finally, of the 32 PhD authors, 29 were matched. In total, roughly 60% of authors in the theme profiles database were matched with network measures. Following the same methodology, a subsequent mapping of keywords to themes was also performed.

The reason that 40% of the authors could not be matched in the network influence database was that no references were available in their works. The majority of these authors did not include any bibliography in their papers, and the works of a very small proportion—mainly authors of theses—are embargoed by their affiliated institutions. This is an inherent property of the data, and it is an important component to document when describing the customs of different academic cultures. Because of China’s unique intellectual traditions in the early stages of CIS’ development as an academic discipline, the overwhelming majority of papers published had no bibliographic references. In addition, many early papers were case studies of the authors’ own experiences, and their documentary nature precluded the need for many citations. However, these early studies were included in the data-set for two reasons. Firstly, these articles were produced during CIS’s initial stage, as per Schneider’s (2009) model of the development of scientific disciplines, and to exclude them would be to miss out on a significant portion of the early literature. Secondly, many of them received citations from later studies, which indicated that they served as the foundation for the development of CIS and brought academic value to the field.

To investigate how well themes act as predictors of the network influences of CIS authors, the most simple approach was linear regression; this is a good starting point for becoming acquainted with the data and is typically used as a first step in statistical analysis for examining the null hypothesis that the explanatory variables have no relationship at all to the response variable. For the preliminary analysis, separate linear regression models were fit with theme profiles as the predictors, using DC, WDC, PRA and EigenVector as a response variable in each model. Since each theme profile was normalized, authors who published numerous papers and those who published only a few had similar-looking theme profiles, therefore the number of papers published by an author was added as a predictor. An F-test was used to determine if any of the regression models could explain variation in author influence in statistically significant terms. The F-test results showed that the linear regressions did not explain the variations in author influence very well. High levels of disparity between authors’ influences was the suspected reason for this. To confirm this suspicion, the disparity was calculated by means of the Gini Coefficient, which measures the concentration of mass in a cumulative distribution, and is borrowed from the field of economics.

A linear model is a simple approach but it makes strong assumptions about the relationship between the response variable and the explanatory variables, therefore it came as no surprise that the linear regression failed to explain the data well. The next approach adopted was one which, though less ambitious in terms of explanatory power, is far less dependent on assumptions. This alternative approach involved dividing network measures into three groups (‘bins’): high-, middle-, and low-ranking. However, rather than assuming that these three bins were equally probable and so spacing their cutoffs regularly, a data-based approach was employed to determine where they should fall. A total of 20 cutoff points were considered, corresponding to the percentiles of network measures, from 0 to 95 in steps of 5. Creating the three groups called for finding specific lower and upper cutoffs for the middle-ranking group. An example of where the cutoffs between the three bins might be placed would be at the 5th percentile between the low- and middle-ranking groups, and at the 85th between middle- and high-ranking (see Fig. 2).

The effect of an author’s theme profile on the probability of his belonging to each of the three groups was estimated by means of a multinomial regression. For each of the different groups, a statistical analysis known as a deviance test was performed to assess whether its divisions were good for explaining authors’ influence based on their theme profiles and number of publications. This procedure allows a numerical quantity known as the p-value to be calculated. For each model, the smaller the p-value, the more likely the model is a good fit.

Given that there were 190 different combinations of cutoff points to consider for each response variable, we could not simply report as significant those cutoff points that had a p-value of less than 0.05. If we did, by the definition of a p-value, we could expect about 10 cutoff points to be significant even if there was no relationship between authors’ meme profiles and influence for any cutoff point. Therefore, a statistical procedure was used to find cutoff points that gave considerable evidence of a good model fit over and above the fact that we were choosing from 190 different models. This procedure can be measured by a quantity called False Discovery Rate (FDR), which is defined as the expected proportion of false discoveries, or cutoff points that are not significantly related to authors’ theme profiles, from all cutoff points detected as significant. In statistics, the FDR can be controlled by using the Benjamini–Hochberg procedure. The smaller a group of models’ FDR, the greater their chance of representing true underlying effects rather than random variation in the data. The procedure assigns to each model a score called the q-value. One way of interpreting this value is that to build a group of models with a certain given maximum FDR, only those models with a q-value below that FDR should be included. Consequently a model is considered good enough for inclusion in a group if it has a small q-value.

Several hypotheses were formulated before the statistical analysis was conducted. One such was that scholars typically perform literature searches by submitting keywords to search engines that rank results from the most recent to the oldest. Under this hypothesis, it was expected that authors writing on commonly studied subjects would, because of sheer weight of numbers, have difficulty becoming highly influential. Conversely, authors writing about rarely-studied subject matters would be far more likely to receive attention from colleagues tackling the same subjects, translating into numerous incoming citations for them. Another reasonable hypothesis was that authors might use other methods of performing literature searches, such as finding citations in existing papers or sorting results based on relevance rather than how recently the items were published. The analytical methodology described in this paragraphs is an important first step towards testing the veracity or otherwise of these hypothetical scenarios.

To examine whether CIS authors’ full keyword profiles were significantly correlated with any of the network measures, one additional analysis was performed—regularized regression. The assumption made at this point was that the majority of keywords were not highly correlated with influence, while a small minority were. In statistics this is called a sparsity assumption (Hurley & Rickard, 2009). Had a simple linear regression of measure of influence been run on keyword profiles, we would have expected to obtain a large number of very small regression coefficients (one for each unique keyword), some medium-sized, and maybe a few large ones. Adding too many non-significant terms into a standard regression would have obscured the signal from significant terms, hence the need to use regularized regression for removing non-significant keywords. Regularized regression addressed this issue by zeroing out many of the insignificant coefficients. More specifically, a regularization technique called Lasso was run for multinomial regression (Tibshirani, 1996) with 10-fold cross-validation to approximate the optimal set of non-significant keywords and set their coefficients to 0. The remaining keywords were considered to be significantly correlated with the network measure. Similar outcomes were hypothesized from the keyword profile analysis as from the earlier theme analysis. Some of the frequently used keywords were expected to be correlated with the low influence group, whereas some rarely used ones were expected to be correlated with high influence. The reason for this predicted outcome was the same as the one for themes described in the previous paragraph: authors whose papers have unique keywords are more likely to be read and cited by fellow researchers than those with common keywords.

Author interaction

The nodes representing authors were situated in Fig. 3 using FM³. Contrary to the expectation that the field of CIS is composed of polarized camps which barely communicate with one another, Fig. 3 suggests rather that its scholars cannot be easily divided into clearly separable communities. In addition, the degree distribution of the entire CIS graph follows a scale-free behavior, which implies that several nodes with high In-Degree and Out-Degree scores perform the function of holding the graph together: these nodes are to be found at the center of Fig. 3, those with lower In-Degree and Out-Degree scores being pushed towards the periphery. To corroborate this finding, the citation patterns of the top 150 CIS authors were also visualized (see Fig. 4). The ranking of these authors was determined using the PageRank algorithm on the entire network. If there were to be well-defined communities, they would have been clearly visible, but in reality Fig. 4 also illustrated a hairball effect, which means that even the top-ranking authors closely cite each other’s works.

This finding suggests that CIS researchers do not form opposing camps marked by a distinct intellectual preference for liberal arts or empirical sciences, and that certain influential scholars are well-cited across the board. This is in sharp contrast to the situation in WIS, where for personal and professional reasons authors may choose not to include certain items of relevant research in their works: scholars from ESIT, France’s most famous IS institute, almost exclusively cite one another’s work and avoid research by scholars from disciplines outside interpreting; the latter reciprocate by ignoring the interpreting specialists’ work, adjudging it to be ‘unscientific’ (Gile, 1999).

To investigate CIS community structures in greater depth, a number of leading network-clustering algorithms were examined; the purpose of these was to optimize the clustering process and determine the optimum number of clusters. Table 1 shows the modularity values that resulted from testing each algorithm using CIS citation data to cluster the network. Spin-glass yielded the highest modularity value with seven clusters, so was adopted for this study (Fortunato, 2010). Another reason for using this algorithm is that it allows the user to input the number of clusters required by specifying the number of spins in the system. Other clustering algorithms do not generally allow the number of clusters to be used as input as they determine the number of clusters by optimizing some objective function or by optimizing some dynamic process.

The Spin-glass clustering algorithm is based on spin models, which are popular in statistical mechanics (Reichardt & Bornholdt, 2004). The underlying idea is based on the principle that nodes connected to each other should belong to the same cluster and ones not directly connected should belong to different clusters. If Potts spin variables are assigned to the vertices of a network, and if the interactions are between neighboring spins, structural clusters can be found from spin alignment of the system. Spins of nodes within clusters are similar and different across clusters with the purpose of maximizing Potts energy.

Figure 5 displays the seven clusters obtained when the Spin-glass clustering algorithm was applied to the CIS citation data network. Nodes in each of these clusters were assigned their positions using the aforementioned FM³ algorithm. Their closely-packed appearance represents the dense structure of these communities as they exchange frequent citations with other members of the same community. The names of the influential authors in each cluster are displayed to make it easy to identify the principal communities the network can be subdivided into. This was achieved by applying the PageRank algorithm to the entire network and ranking the importance of each node based on its PageRank value.

Daniel Gile had the highest PRA scores in the green cluster, the largest community in the data-set with 2,366 scholars. Authors such as Franz Pöchhacker, Jean Herbert, Roderick Jones, Robin Setton, Yang Chengshu, David Carroll, George Miller, John Andreson also belonged to this cluster. Most of these top influencers’ works are concerned with cognitive psychology or psycholinguistics, which would appear to indicate that those disciplines have a dominant influence on CIS research. Most notably, George Miller’s The Magic Number Seven, Plus or Minus Two (Miller, 1956), and John Anderson’s Cognitive Psychology and Its Implications (Anderson, 1980), received 103 and 106 incoming citations, respectively, suggesting that the interdisciplinary approach to interpreting research is one favored by many a scholar. It should be noted here that this community features outliers, but further research would be required to ascertain why, alongside scholars whose particular focus is cognition, CIS authors cite writers such as Herbert and Jones, neither of whom specializes in cognitive science, though both have produced popular textbooks on interpreting techniques.

Only one of the top ten scholars in this cluster was Chinese—this was Yang Chengshu, who has served as director of the Graduate School of Translation and Interpretation at Fu Jen Catholic University in Taiwan since 2006. She received a total of 195 incoming citations from CIS scholars, her two most often-cited works being Research into Interpreter Training (Yang, 2005) (In-Degree = 240) and Types of and Rules for Syntactic Linearity in SI (Wan & Yang, 2005) (In-Degree = 2). Her high-ranking in this cluster—6th out of 2,366 scholars—is worthy of remark: given that Taiwan has a somewhat different socio-cultural background from that of mainland China, and that her specialty is Japanese–Chinese interpreting, her great popularity among CIS scholars was unexpected. Several factors may explain this: though Yang has worked exclusively at academic institutions in Taiwan, she received her PhD in Interpreting Studies from BFSU under the supervision of Wang Kefei in 2006. She also has close ties with scholars from the PRC, evidenced by the fact that the second of her most frequently cited works mentioned above was co-authored with Wan Hongyu, a scholar from Shanghai. In addition, the popularity of her other most-cited work indicates that the general principles of interpreter training know no linguistic boundaries.

Most influential scholars

The following 30 people were identified as the most influential Western scholars in CIS based on their PRA scores (see Table 2).

Gile, Lederer, and Seleskovitch head the table, indicating that their research is highly influential among their Chinese colleagues. When it comes to degree centrality (DC) measures the ranking has shifted slightly: Gile, cited by 1,077 Chinese scholars, continues on top, while Seleskovitch, cited by 756, now ranks second. She is trailed by Lederer, who was cited by 674 researchers. Their ranking by weighted degree centrality (WDC) remained consistent with their DC ranking: Gile scored 2,116, Seleskovitch 1,207 and Lederer 1,020. The first had a total of 75 papers and books receive incoming citations. The most widely cited of his works was Basic Concepts and Models for Interpreter and Translator Training, whose main contribution to the field is theoretical, though it does have a number of empirical components (See Table 3). Seleskovitch and Lederer formulated the Interpretive Theory of Translation, which stresses the importance of disregarding the linguistic structure of the original when interpreting into a foreign language. Despite both having published empirical research, almost all Chinese scholars exclusively cite the translations of their non-empirical works such as Interpreting for International Conferences: Problems of Language and Communication (Seleskovitch, 1978a), Interpréter Pour Traduire, and Pédagogie Raisonné de l’Interprétation—these were among the earliest Western interpreting theories introduced into China. Their influence on CIS research is perhaps attributable to the wide availability of the translated versions of their works. Upon further analysis, of the top 30 Western researchers, their works alone were cited in translation; all the others were cited in the original language—for example, though two of Gile’s French papers received incoming citations, they were both cited in French. The citation data revealed that 27 of Seleskovitch’s works and 23 of Lederer’s received citations. Tables 4 and 5 list the top five most cited works of these two authors.

Being cited by a large number of scholars does not necessarily translate into a particular author’s having a high PageRank score or being perceived as highly influential by his confreres. For instance, David Gerver was cited by 160 Chinese scholars and received 257 citations, ranking 20th in terms of DC and 16th in WDC, but his ranking dropped to 40th when calculated by the PRA. Gile’s (2000) paper indicates that Gerver’s work is also reasonably popular in the West. There is no doubt that Gerver was an influential pioneer in introducing methodologies from experimental psychology into interpreting research in the 1970s, but his work was heavily criticized by working interpreters for a lack of ecological validity, his manner of selecting research participants, and the methods he employed for evaluating their performances (Gile, 1994). This may explain his low PRA score, which indicates that the most influential Chinese scholars seldom cited his work—most citations of him were made by less influential researchers such as graduate students. The difficulty of accessing his publications in China may be another factor that contributes to his low score. A similar situation was observed for the rankings of Setton and Moser-Mercer: The former received the seventh most citations, as shown by his WDC score, but ranked only 13th in terms of influence. Having received his PhD from the Chinese University of Hong Kong in 1997 and subsequently taught in Taiwan and Shanghai, he remained active in CIS up to the time of his departure from the Shanghai International Studies University in 2011. While he published a number of texts on interpreting over the years, only his PhD dissertation, which was later published as a monograph, received much attention from the CIS community; the other papers he wrote received fewer than 15 incoming citations each (See Table 6). Moser-Mercer was cited by 221 Chinese colleagues, receiving a total of 381 citations, giving her a ranking of 12th in DC and 6th in WDC, but her position fell to 21st in the PRA. Examination of her works revealed that 39 of them received citations. The most cited was Bridging the Gap: Empirical Research in Simultaneous Interpretation—the first collected volume on empirical research into interpreting, on which she worked as a co-editor (see Table 7).

Working interpreters not engaged in research are still highly influential among Chinese academics. Roderick Jones is a case in point: only his monograph Conference Interpreting Explained (Jones, 1998), a manual that provides tips and techniques for honing interpreting skills, received citations. However, his PRA ranking was 6th, immediately after Franz Pöchhacker. The professional’s intuitive understanding of working interpreters is sometimes deeper and more comprehensive than that gained by collecting and analyzing empirical data, which might explain why he is highly cited by influential Chinese scholars. In addition, many may find his writing easy to relate to and his suggestions simple to follow in interpreting practice, contributing to his wide popularity in the CIS community.

A few translation scholars, such as Eugene Nida, Peter Newmark and Mona Baker, were also reasonably influential among Chinese academics. The first is noted for his Dynamic Equivalence Theory in translation, and works of his such as Language, Culture, and Translating (Nida, 1993), Dynamic Equivalence in Translating (Nida, 1995), and Language and Culture: Context in Translating (Nida, 2001) were widely cited by Chinese colleagues. This contributed to his PRA ranking of 4th among all Western scholars, behind Seleskovitch. He ranked 5th in DC, having been cited by 467 Chinese scholars, and received the same ranking for WDC with 730 citations. Though Pöchhacker outranked him in both DC and WDC, Nida’s PRA score seems to suggest that he has a higher research impact than Pöchhacker in the CIS community. As mentioned earlier, the reason for the discrepancy between their DC/WDC and PRA scores is that Nida received more citations from influential CIS scholars than Pöchhacker, whereas the latter is more popular among low-ranking CIS scholars.

In addition, Western scholars in linguistics, sociology, cognitive science and psychology played an appreciable role in CIS research. For example, Lyle Bachman’s research on language testing was often cited in work on the assessment of interpreting competence. Dan Sperber, a sociologist and cognitive scientist, developed the Relevance Theory in collaboration with Deidre Wilson, a psychologist by training. This theory has been used by numerous Chinese scholars to shed light on the processes of listening comprehension, note-taking and language production in interpreting.

The data reveal the following to be the top 30 most influential Chinese scholars (see Table 8).

In comparison with their Western colleagues, the composition of the top-ranking Chinese scholars in PRA was quite homogenous. Of the top 30, no fewer than 25 (all but Mei Deming, Liu Miqing, Wu Bing, Hu Gengshen and Zhuang Enping) are interpreting scholars.

Cai Xiaohong ranked 5th in DC and 6th in WDC, but had the second highest PRA score of all, indicating that she had a very large research impact among her fellows. In 2000, she defended her doctoral dissertation, in which she studied the development of competence in consecutive interpreting by conducting an experiment with 12 participants of different skill levels. Since then, she has published a number of monographs and papers, of which the top five most frequently cited are shown in Table 9. The present research revealed that as was the case for Moser-Mercer, Cai’s most cited work was a collective volume which she co-edited. All the other top five works were research papers she has published over the years.

Bao Gang ranked 1st in DC, having been cited by 543 people and receiving 592 citations, the majority of them on his monograph An Overview of Interpreting Theories, but his ranking dropped to 4th in PRA. A possible explanation for this is that sadly he passed away in 1999 at the age of 42: had he lived, he would surely have contributed a great deal more to CIS. In Table 10, which lists his top five most cited works, it is interesting to observe that the citation distribution for his works is lopsided: the overwhelming majority of scholars cite his aforementioned work, an introduction to the leading Western theories prevalent in the 1990s. A possible explanation for this highly skewed distribution is, once again, his untimely death.

Liu Heping ranked 2nd in both DC and WDC, receiving a total of 843 citations from 513 scholars, but her ranking dropped to 8th in PRA. Her most cited works are shown in Table 11. Liu’s incoming citations seem much more balanced than those of the previously mentioned CIS scholars. Though her monograph Interpreting Techniques: Scientific Thinking and Reasoning received the majority of incoming citations, her papers, which occupied the rest of the top five spots, were also cited by numerous CIS scholars. Liu obtained her PhD in Translation Studies from the University of Paris III, and is an active member of the CIS community, regularly appearing as a keynote speaker at various conferences. A former student of Seleskovitch’s, she continues to advocate the Interpretive Theory of Translation in her research.

As is the case for their Western colleagues, the opinions of Chinese practicing interpreters are also highly valued by CIS scholars. Zhang Weiwei, who ranked 9th in PRA, once served as a staff interpreter at the United Nations’ duty station in Geneva: he was cited almost exclusively for his handbook English-Chinese Simultaneous Interpreting (Zhang, 1999), which highlights the use of segmentation and syntactic linearity with numerous practical examples. In addition, his high rankings in DC (9th) and WDC (12th) suggest that he was equally popular with scholars both influential and less so.

Zhang Wei’s rankings in PRA, DC and WDC merited some investigation. He topped the list in WDC, his ranking having been significantly boosted by his weighted out-degree centrality (WODC)—he made a total of 854 citations of other scholar’s works in all his papers and in his doctoral dissertation, far exceeding the second highest in WODC, Ren Wen, whose total was 557. However, in DC Zhang slipped to 6th position, and in PRA to 44th. It should also be noted that he was the most productive CIS author, with 28 papers recorded in the data-set. The discrepancy between his influence score and research productivity might be explained by his being a relative newcomer to CIS, having received his PhD in 2007—it is common that the older a work is, the larger its readership and the more it will have been cited. Another plausible reason for the discrepancy is that his doctoral dissertation, in which he investigated the relationship between simultaneous interpreting and working memory, was highly technical and used a combination of experimental, observational and questionnaire-based research methods; the unfamiliar techniques he used may have deterred other CIS authors from adopting his work. Furthermore, the research carried out for Question 3 of the present study revealed that authors writing about cognitive issues were slightly more likely to end up in the bottom PageRank influence group than those writing on other topics.

There are also a number of scholars who, though trained and specializing in other disciplines, occasionally venture into the field of interpreting research, namely Mei Deming, Hu Gengshen, Liu Miqing, Wu Bing and Zhuang Enping. Mei topped the Chinese scholars’ PRA rankings. He is among the most prolific MA thesis advisors in China, and was involved in launching the Shanghai Interpreting Exams. A PhD graduate in linguistics and rhetoric from the University of Indiana, Mei has taught numerous classes on linguistics, public speaking and movie appreciation. In the data-set there were no research papers on interpreting written by him alone, though he co-authored a few with his doctoral students. In addition to all the above, Mei was the editor-in-chief for An Advanced Course in Interpreting (Mei, 1996), which was first published in 1996 (new editions in 2000, 2005 and 2011), and for An Intermediate Course in Interpreting (Mei, 1998), first published in 1998 (new editions in 2003, 2008 and 2010). These are the mandatory test preparation books for the Shanghai Interpreting Exams, and other Chinese authors’ citing of them contributed to his becoming one of the most influential CIS authors in the country. The case of Wu Bing follows the same lines: she is mainly cited for her textbook on Chinese-to-English interpreting. Neither author has published any empirical studies on the subject. Zhuang Enping specializes in cross-cultural communication, but he has also written articles on the principles of interpreting and how the differences between Eastern and Western styles of communication affect it. His high degree of influence among Chinese scholars (17th in PRA) indicates that a good number of researchers draw inspiration from Communication Theory.

Unlike the situation with the top 30 most influential Western scholars, the top 30 Chinese list had only one Chinese scholar whose research was completely unrelated to interpreting. Chen Siping, who ranked 29th in PRA, focused on the application of the Relevance Theory to reading comprehension; her high ranking would seem to indicate that influential Chinese scholars frequently used her work as the theoretical underpinnings for their research.

Research topics and academic influence

Multinomial logistic regression: procedure for stratifying authors into high-, middle- and low-ranking groups

Though linear regression was the first choice—and most straightforward—method of explaining authors’ levels of influence based on their theme profiles and numbers of publications, as we have seen it turned out to be lacking when it came to explaining their influence within networks. An alternative approach was to classify the authors into three influence groups—low, middle and high—using multinomial regression.

The three groups can be defined in different ways depending on which cutoffs are used to separate them. Only two parameters are needed to define the three groups: one cutoff value to separate the low and middle groups and another to separate the middle and the high. N.B.: By knowing these two values we can know both the length and the midpoint value of the middle influence group.

The multinomial models represent the probability of an author’s being in each of the three groups, given an author’s theme profile and number of papers written. The coefficients of each of the multinomial models were determined by fitting each model to the data, where the response variable is now an indicator of each author’s allocation to the three groups corresponding to each model. The coefficients depended on how the three groups were divided, and each pair of cutoffs was associated with a different multinomial model. In Figs. 6 and 7, each square represents a model and is thus associated with a certain definition of the three groups of authors, or equivalent to certain values of the two cutoffs.

Figure 6 shows the results of fitting each of the 209 group divisions for PageRank, and calculating a p-value for the null hypothesis of no significant relation between theme profile and group membership. Since we expect about 21 groups to have p-values less than 0.1 completely by chance, it is not instructive to report all groups with low p-values as indicative of significance. Many of these low p-values will not relate to significant effects and would confuse results.

Instead, the quality of a group of models was evaluated by a so-called False Discovery Rate (FDR) analysis. In the present study, the FDR of a group of models is the expected proportion of models that are not good, i.e., are not statistically significant ones, therefore the lower the FDR the better the group of models. The quality of each model was measured by a quantity called q-value. Simply put, to build a group of models with an FDR lower than a certain threshold value, only models with a q-value lower than that value can be included. The results of this FDR analysis for PageRank are summarized in Fig. 7.

Figure 7 shows that the FDR-based approach makes it easy to identify groups of good models: there are three stripes of squares associated with a small FDR. The furthest (red) stripe to the right corresponds to the models whose group divisions yield the most statistically significant results, i.e., which best explain the connection between an author’s theme profile and his influence.

The red squares correspond to q-values lower than 0.05, so the group of models corresponding to the red squares has an FDR below 5%. In other words, no more than 5% of these models are expected to be non-significant. Similarly, since the orange squares correspond to q-values above 0.05 and below 0.10, all groups of models, whether they correspond only to orange squares or to orange and red ones, have an FDR below 10%. The same reasoning can be repeated for each color of square and its associated q-values. In Fig. 5, the three diagonal stripes of red and orange squares correspond to the best and relatively good models described in this paragraph.

Once groups of relatively good models (each defined by a pair of cutoffs) had been determined, the selection of a ’stand-out’ model, i.e., one with an exceptionally good pair of cutoffs, was still required. A clustering procedure known as k-medoids, by which similar pairs of cutoffs are divided into different groups, was employed at this point (Kaufman & Rousseeuw, 1987). The chosen divisions corresponded to the most central point in the cluster with the smallest FDR (i.e., the cluster with the best statistical significance). The best pair of cutoffs for PageRank are the 20th percentile for low-ranking authors and the 85th for high-ranking ones (see Fig. 7).

Similar analysis revealed the optimal cutoffs for In-Degree to be [0.6, 0.95] and for Out-Degree [0.6, 0.8]. Theme profiles best explain the differences between the bottom 60% of Out-Degrees (less than two incoming citations), the top 5% (more than 33 incoming citations), and the middle group (those between 60 and 95%). Similarly, they best explain the differences between the bottom 60% of In-Degrees (0 outgoing citations), the top 20% of Out-Degrees (more than 24 outgoing citations), and the middle group. Note that for Out-Degree we can choose any cutoff below 64% as that percentage of authors had 0 Out-Degree citations. Hence, any such quantile choice would result in absolute cutoffs of 0 and 24 outgoing citations. For EigenVector, no cutoffs were found that had a chance of less than 20% of being non-null hypotheses—in fact, no significant p-values at all came to light. As a result, the predictors worked better with PageRank than EigenVector for measuring author influences in the CIS network; for this reason, the latter measurement was not adopted for the following analysis. Table 18 shows optimal cutoff quantiles and their corresponding measure values.

Table 18:

Optimal cutoff quantiles and their measure values.

	Low cutoff (percentile)	Low cutoff (Val)	High cutoff (percentile)	High cutoff (Val)
PageRank	20	0.0000659	85	0.0000877
In-degree	60	2	95	33
Out-degree	60	0	80	24

DOI: 10.7717/peerj-cs.14/table-18

The output shown in Tables 19, 20 and 21 below summarizes the results of multinomial regression of theme profiles on the three network measures. The coefficients are symmetric relative odds on the log scale.

Table 19:

PageRank multinomial regression.

	coef	se	p values
Low (<20th percentile)
Intercept	−0.270	0.110	0.014
No. of papers	0.000	0.031	1.000
Cognitive	0.002	0.002	0.508
Language	0.002	0.002	0.432
Miscellaneous	0.000	0.003	1.000
Professional	0.000	0.004	0.978
Socio-cultural	0.000	0.003	1.000
Medium (>=20th, <85th percentile)
Intercept	0.988	0.095	0.000
No. of papers	−0.036	0.028	0.207
Cognitive	−0.004	0.002	0.068
Language	−0.001	0.002	0.763
Miscellaneous	−0.003	0.003	0.276
Professional	−0.002	0.003	0.569
Socio-cultural	−0.006	0.002	0.009
High (>=85th percentile)
Intercept	−0.719	0.129	0.000
No. of papers	0.003	0.035	0.936
Cognitive	0.000	0.003	1.000
Language	0.000	0.003	1.000
Miscellaneous	0.007	0.003	0.029
Professional	0.000	0.005	1.000
Socio-cultural	0.002	0.003	0.512

DOI: 10.7717/peerj-cs.14/table-19

Table 20:

In-degree multinomial regression.

	coef	se	p values
Low (<60th percentile)
Intercept	1.237	0.097	0.000
No. of papers	0.000	0.028	1.000
Cognitive	−0.001	0.002	0.677
Language	0.000	0.002	1.000
Miscellaneous	−0.003	0.003	0.309
Professional	−0.002	0.003	0.473
Socio-cultural	0.000	0.002	1.000
Medium (>=60th percentile, <95th percentile)
Intercept	0.194	0.101	0.054
No. of papers	0.041	0.029	0.153
Cognitive	0.000	0.002	1.000
Language	0.006	0.002	0.011
Miscellaneous	0.000	0.003	1.000
Professional	0.001	0.004	0.799
Socio-cultural	−0.001	0.003	0.613
High (>=95th percentile)
Intercept	−1.431	0.238	0.000
No. of papers	−0.100	0.100	0.318
Cognitive	0.002	0.005	0.739
Language	−0.003	0.005	0.601
Miscellaneous	0.005	0.005	0.308
Professional	0.000	0.007	1.000
Socio-cultural	0.011	0.004	0.006

DOI: 10.7717/peerj-cs.14/table-20

Table 21:

Out-degree multinomial regression.

	coef	se	p values
Low (=0th percentile)
Intercept	1.014	0.096	0.000
No. of papers	0.000	0.029	1.000
Cognitive	0.000	0.002	1.000
Language	−0.004	0.002	0.094
Miscellaneous	−0.004	0.003	0.180
Professional	0.000	0.003	1.000
Socio-cultural	−0.002	0.002	0.372
Medium (>0th percentile, <80th percentile)
Intercept	−0.469	0.130	0.000
No. of papers	−0.029	0.046	0.528
Cognitive	−0.001	0.003	0.837
Language	0.000	0.003	1.000
Miscellaneous	0.001	0.003	0.813
Professional	0.008	0.004	0.036
Socio-cultural	0.001	0.003	0.647
High (>=80th percentile)
Intercept	−0.545	0.117	0.000
No. of papers	0.075	0.031	0.014
Cognitive	0.009	0.002	0.000
Language	0.002	0.003	0.339
Miscellaneous	0.000	0.003	1.000
Professional	−0.001	0.005	0.900
Socio-cultural	0.000	0.003	1.000

DOI: 10.7717/peerj-cs.14/table-21

We present all of our results on the relative odds scale, which can be found from the transformation exp(coef). This makes it easier to compare the relative effect of different theme profile variables in different subgroups. For example, a coefficient of −0.004 for cognitive in the middle group for PageRank means that authors who have 1% less cognitive articles have on average 1 − exp(−0.004) = 0.006 percent smaller odds. It is straightforward to transform odds into raw percentages by taking into account the intercept coefficient and an author’s theme profile. Consider an author who has 100% cognitive articles: They would have exp (0.988 − 100∗.004) = 1.8 to 1 odds, or about a 1.8/(1 + 1.8) = 0.62 chance of being in the middle group.

For PageRank, it was found that trading off 1% of Cognitive for 1% of any other paper gave us exp (−0.004) = 0.996 lower odds, or roughly a 0.4% lower likelihood of being in the middle group of PageRanks. The same trade-off for Socio-cultural gave us exp (−0.006) = 0.994 lower odds, or roughly a 0.6% lower likelihood. The only positive significant relationship between theme profile and placement in the high PageRank group came from trading for Miscellaneous papers. Every percentage point increase in Miscellaneous papers gave roughly 1.007 greater odds of being in the top 15% of PageRanks. This is attributable to the fact that some highly influential papers were classified as Miscellaneous. Earlier research (Xu, 2014; Xu, 2015) indicated that the Miscellaneous theme accounts for only 9% of the total themes mentioned in MA theses, 0% in doctoral dissertations and 5.1% in research papers, meaning that non-mainstream topics have been explored by few CIS researchers, whereas numerous papers on Cognitive, Language, Professional and Socio-cultural issues have competed for attention. It appears to be much easier for those authors with papers on Miscellaneous subjects to get noticed and be cited by other researchers.

Another way of describing a 0 coefficient for a theme profile variable for a certain group is as follows: when one percentage point of this variable is exchanged for one of the Training theme, there is no change in the probability of an author’s belonging to the group under consideration. For example, when an author trades 1% of Cognitive or Language for 1% of Training, the probability of his belonging to the high-ranking PageRank group is not affected at all, but that probability is reduced by 0.007 when he trades 1% of Miscellaneous for 1% Training theme.

The next item to be scrutinized was the regression on In-Degree. The In-Degree multinomial results showed that trading off any theme profile for Language papers gave 1.006 higher odds of being in the middle In-Degree group (greater than 2 but fewer than 33 incoming citations). It was also observed that if an author traded off any theme profile for the Socio-cultural themes, they would have 1.011 higher odds to be in the the high In-Degree group (having more than 33 incoming citations). This was the most significant effect of any theme on any group, suggesting that authors on Socio-cultural issues are easily identified and cited by their confreres. As was the case for the Miscellaneous theme, Socio-cultural issues received little—though slightly greater—attention from CIS researchers and authors: 10.7% for research papers, 11.1% for MA theses and 12.4% for doctoral dissertations. Socio-cultural issues play an important part in interpreting, which can all too easily be affected by factors such as contexts and hidden power relations between various actors in the dialogue (Pym, Shlesinger & Jettmarová, 2006).

Finally, the same procedure was applied for Out-Degree measure. Here it was observed that if authors wrote more Language papers, they were exp (−0.004) = 0.996, or had 0.004 lower odds to end up generating the least number of citations (low rank group). It should be noted, however, that since the p-value in this case was 0.09, this finding is marginal at best. In addition, an author had a 1.008 higher probability of having between 1 and 24 outgoing citations (middle rank group) by writing more Professional papers, 1.009 higher odds of at least 24 citations (high rank group) by writing more Cognitive papers, and 1.077 higher odds of at least 24 outgoing citations by writing more papers in general. These results were found generally to tally with those of the linear regression described in section ‘Linear regression: a first approach to modeling the relationship between network influence and memes.’