A Survey of Chinese Interpreting Studies: Who influences who ... and why?

This paper describes how scholars in Chinese Interpreting Studies (CIS) interact with each other and form discrete circles of influence. It also discusses what it means to be an influential scholar in the community and the relationship between an author’s choice of research topic and his academic influence. The study examines an all-but-exhaustive collection of 59,303 citations from 1,289 MA theses, 32 doctoral dissertations and 2,909 research papers, combining traditional citation analysis with the newer Social Network Analysis to paint a panorama of CIS. It concludes that the community cannot be broadly divided into Liberal Arts and Empirical Science camps; rather, it comprises several distinct communities with various defining features. The analysis also reveals that the top Western influencers have an array of academic backgrounds and research interests across many different disciplines, whereas their Chinese counterparts are predominantly focused on Interpreting Studies. Last but not least, there is found to be a positive correlation between choosing non-mainstream research topics and having a high level of academic influence in the community.


Introduction
The earliest documentary record of interpreting dates back as far as 3000 BCE-the Ancient Egyptians had a hieroglyph for it (Delisle & Woodsworth, 1995)-but it can be assumed reasonably safely that the first interpreters started work as soon as cavemen realized they could not be sure to make themselves understood by neighboring tribes using gestures and signs alone. Given its extremely long history, it is somewhat surprising that it only became an independent field of academic enquiry in the 1990s, when scholars began consciously to use the term Interpreting Studies, to distinguish it from the original 'parent' Translation Studies. Despite the PeerJ PrePrints | https://dx.doi.org/10.7287/peerj.preprints.941v2 | CC-BY 4.0 Open Access | rec: 24 Jun 2015, publ: 24 Jun 2015 mystique which still surrounds the profession to a certain degree 1 , since the Second World War interpreters have been increasingly in demand to bridge communicative divides wherever they might arise-war crimes tribunals, peace-keeping operations, high-level international trade negotiations, low-level sightseeing trips … the list is endless.
Chinese interpreters came to prominence on the international stage when the People's Republic of China (PRC) regained its seat at the United Nations (UN) in 1971. As a result of China's return the UN was instrumental in establishing a dedicated training program to meet the demand for conference interpreting services from various of its offices all over the world (Wang, 2006).
The first research article on Chinese Interpreting Studies (CIS) archived by CNKI 2 was published in the late 50s (Tang & Zhou, 1958), and since then the discipline's growth has been explosive: a total of over 3,600 scholars have to date produced nearly 3,000 journal articles and conference proceedings, 1,300 MA theses and over 30 dissertations on the subject. Given its rapid evolution and ever-heightening academic status it is of crucial importance to study the structure of this scientific community. The purpose of the present scientometric survey is to marry the traditional technique of citation analysis with the newer one of Social Network Analysis (SNA) to obtain a fuller picture of the ways in which CIS scholars communicate with each other both formally and informally.

Major questions
To gain an understanding of how scholars in CIS communicate with one another to generate learning and advance the field, both citation analysis, which describes formal networks of influence (Baumgartner & Pieters, 2003), and social network analysis, which identifies informal communities (Otte & Rousseau, 2002), have been used in this study.
Using an all-but-exhaustive collection of citation data, we ask how authors interact, how we can characterize who is influential, and what being influential means. These questions have long 1 Until the release in 2005 of the movie The Interpreter, starring Nicole Kidman, many an outsider was no doubt unsure of the difference between written translation and spoken interpreting. 2 The China National Knowledge Infrastructure is by far the nation's most comprehensive academic database, archiving conference proceedings, journal articles, MA theses and doctoral dissertations dating back to the early 1900s.
6 researchers were less inclined to interdisciplinarity.

The Present Study
There is clearly a growing interest in scientometrics among CIS researchers. Aside from Gao, most authors to date have applied its more basic methods and principles, mainly article counts, in their research. This is the logical place to start, but there is ample room to employ the more complex approaches scientometrics offers. Doing so will shed light on the true impact made by individual scholars, and will provide other more finely tuned information about the evolution of specific fields of research in a relatively objective way. As (Lowry et al., 2007) point out, mere numbers do not permit a nuanced analysis of influences within a given discipline; for example, an author may have published numerous articles but have little influence among his peers. In addition, a small sample population may cause significant biases and affect the outcome of any analysis conducted. The present research study is intended to contribute to CIS by carrying out, for perhaps the first time, a thorough scientometric survey of the literature, including journal articles, theses and dissertations. Its aim is to provide scholars with a comprehensive and objective overview of the interactions between scholars in the field, and of which academics are the most influential and how their choice of particular subjects of enquiry relates to their impact on research as a whole.

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 each other and how influence is defined in the community. The following section outlines the author's rationale for investigating each research question. 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.

How do CIS authors interact with one another
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.

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 PeerJ PrePrints | https://dx.doi.org/10.7287/peerj.preprints.941v2 | CC-BY 4.0 Open Access | rec: 24 Jun 2015, publ: 24 Jun 2015 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  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.

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 study (2000)-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, 2005Gile, & 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  (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 noninfluential 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-graph 6 was created which showed only these 150. Again, the FM^3 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. 6 An induced sub-graph is one that highlights only a certain number of nodes and all the edges connecting them, but omits all the other nodes and non-connecting edges.
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, 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. FM^3 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 and 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 and 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 7 Degree centrality calculates the number of edges connected to a particular node in the network. It has two subcategories: an author's in-degree centrality represents the number of other authors citing him, while his out-degree is the number of others cited by him. 8 Weighted degree centrality is the total number of citations an author makes and receives. Weighted in-degree centrality is the total number of citations of his work made by others, while weighted out-degree centrality is the total number of citations he makes in all his publications.
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 9 . 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 three reasons. Firstly, these articles were produced during CIS's initial stage, as per Schneider's model of the development of scientific disciplines (2009), 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 Figure   2).  The effect of an author's theme profile on the probability of his or her 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 qvalue. 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 nonsignificant 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 Figure 3 using FM3. Contrary to the expectation that the field of CIS is composed of polarized camps which barely communicate with one another, Figure 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 visualized (see Figure 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 Figure 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 a 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 networkclustering 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

Most influential scholars
The following 30 people were identified as the most influential Western scholars in CIS based on their PRA scores:     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 paper (2000) 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    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:     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 6.3.1. Summary of significant findings
As stated in the methodology section, inferences drawn from linear regressions were somewhat unsatisfactory: F-tests proved only the linear model for Out-Degree to be statistically significant.
As discussed then, we consequently turned to multinomial regressions for three influence groups-low, middle and high-for each of the network measures: PageRank, In-Degree, Out-Degree, and EigenVector Centrality. We delimited the group divisions by determining the cutoff points that led to the most statistically significant regression models.
For each measure and each influence group we report below the themes that had statistically significant coefficients.

Influence group Theme group Change in likelihood of belonging to an influence group (%) p-value
Medium (>= 20th, < 85th percentile) Cognitive -0.4 0.068 Socio-Cultural -0.6 0.009 High (>= 85th percentile) Miscellaneous 0.7 0.029  The change in likelihood column shows that authors who had 1% more publications in the listed theme were x% more or less likely to be in that influence group for that influence measure.
Taking PageRank (see Table 12) as an example: when authors wrote 1% more publications that fell into the Cognitive theme category, they were 0.4% less likely to be part of the Mediumranking influence group. In other words, an author with no Cognitive keywords in 100% of his publications would be 40% more likely to have medium PageRank than an author all of whose keywords were Cognitive. The Lasso multinomial regressions for keywords (see section 6.3.5) also supported the earlier multinomial regression results for theme category analysis. The consistency of these analyses is supported by the fact that the majority of the keywords, and the theme categories that these keywords belong to, share the same correlation sign (either positive or negative) as the influence measure group of the authors who wrote them. To illustrate with an example, the keyword "Theory" was positively correlated with an author's placement in the High-ranking group of PageRank, and the theme category of this keyword (Miscellaneous) was also found to be positively correlated with an author's likelihood of belonging to that same High-ranking group. It should also be acknowledged that the Lasso regression analysis failed to detect a couple of findings from the theme category analysis. For example, the positive relationship between Social-cultural keywords and the High In-Degree influence group, and the negative relationship between Language keywords and the Low Out-Degree group, were not picked up by the Lasso model. It is likely that those keywords were not significant on their own, but collectively they contributed to boosting an author into a certain influence group.
In sum, from the findings above it was found that the most influential authors are those who write about Social-cultural and non-mainstream topics. In particular, authors whose papers cover Theory are more likely to be placed in the High influence group than those who do not. In addition, the analysis revealed that those writing about Cognitive issues, especially Sensory Memory, are more likely to include a significant number of bibliographic references.

Linear regression: A first approach to modeling the relationship between network influence and memes
For the preliminary analysis four standard linear regressions of each of the metrics were performed using the theme profiles as predictors. The fact that each theme profile was normalized might result in prolific and non-prolific authors having very similar profiles.
Therefore number of papers published was added as an extra predictor. This was done to separate the effects of an author's theme profile from his or her overall frequency of publishing.  The only regression model which turned out to be significant was Out-Degree. In general terms this means that the topics an author writes about are a good predictor of the number of outgoing citations in his papers. For example, when people write about cognitive issues, the coefficient of this theme (see Table 16 below) suggests that they tend to have higher Out-Degree scores, meaning that they cite more papers. In sum, the coefficients in  The coefficients of themes in Table 16 represent the changes that would occur in Out-Degree if that particular theme proportion were increased by 1%. For example, a 1% shift in an author's theme profile to Cognitive would result in a 0.08381 increase in Out-Degree. A similar 1% shift to Language would result in a 0.06429 increase. It should be noted, however, that the coefficient for the number of papers represents the amount of Out-Degree shifted as a result of increasing the number of papers by 1. More specifically, the coefficient for the number of papers indicates that having one more paper published might corresponds to a 0.59 increase in predicted Out-Degree for that particular author.
It should be noted that a coefficient for Training does not appear in the regression summary.
This is due to the profile normalization discussed earlier. If an author's theme profile is known for any five terms, their score for the sixth theme can be precisely inferred by 1 -sum (scores for 5 themes). Hence when an author's theme profile is parameterized to five dimensions, as is necessary for computational reasons, it contains the same information as a full 6-theme profile.
Each theme profile group was treated as a numeric variable so that the effect of changing the relative proportion of that theme would be clearly visible. Any positive coefficient in the regression means that trading off Training keywords for keywords in that particular theme profile results in an increased Out-Degree.
Trading off 1% of Cognitive theme for 1% of Language theme lowers Out-Degree by 0.02. If this shift in themes were to be repeated 50 times, the model predicts that the paper would have its Out-Degree score reduced by 1 (that is, one fewer outgoing citation).
The analysis thus reveals that using more Cognitive or Language themes in papers is a predictor of having higher Out-Degree counts, i.e. authors writing on those subjects generate more references. The data also suggest that having larger numbers of papers published is correlated, although only marginally, with higher Out-Degree, an interesting finding in that one would expect writing more papers to lead to an inevitable increase in the total number of outgoing citations. There are two possible explanations for this marginal effect. Firstly, the fact that 30% of the CIS papers in the data-set do not have any references suggests that a large proportion of authors do not necessarily have a high number of outgoing citations, even if they have written multiple publications. Secondly, it is possible that a few highly prolific authors generate only small numbers of outgoing citations, which may have an effect on the correlation between the number of publications and the Out-Degree measure.
It should be noted that the multiple and adjusted R-squared values in Table 16 are rather lowless than 1% of the variance in the data is explained by the model, even though the p-values for several predictors are very small. This means that the results of standard linear regression are not very promising: while the trend detected is significant, the model still cannot explain the data very well. Huge disparities between authors' network influence measures were a likely cause for this inadequacy. A large discrepancy or inequality in a response variable makes a linear regression inadequate because estimated effects are highly influenced by a few extreme values, and hence effects which pertain to the rest of the population are subsumed. In the following section we examine this inadequacy further, and describe mathematically the amount of network measure inequality through an analysis of Gini coefficients.

Gini Coefficient: Measuring disparities in each network measure across authors
The Gini Coefficient has typically been used to calculate income inequality in populations, by converting the cumulative distribution of wages into a single number. In such cases a Gini value of 0 corresponds to complete income equality, i.e. every individual is earning exactly the same  evaluating the impact of an author's theme profile on his or her rankings.

Multinomial logistic regression: Procedure for stratifying authors into high-, middleand 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 Figures 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 qvalue. 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 Figure 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 Figure 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 Figure 7).   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 greaterattention 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

Regularized multinomial regression for predicting influence by keywords
A regularization technique called Lasso was run for multinomial regression (Tibshirani, 1996) with 10-fold cross-validation to approximate the optimal set of keywords which were truly significant. The Lasso works by applying a penalty to the absolute value of coefficients, providing a principled way to set the coefficients of non-significant keywords to 0. Any remaining keywords were considered to be significantly related to the network measure. For the dependent variables the same optimal cutoffs were used as those found by the FDR analysis examined in Table 18. By this technique, keywords that were not relevant enough for the present analysis could be discarded.
Only the positive or negative character of each significant keyword's association with influence is given in  The PageRank model (see Table 22) summarizes the most significant keywords and the role they play in deciding which theme group a particular author is most likely to belong to. Both Attention and Nominalization were correlated with authors belonging to the low influence group, suggesting that those who write about these two topics tend to end up with low influence as measured by PageRank. This finding coincides with the earlier theme profile analysis, where both Cognitive (the theme for Attention) and Language (the theme for Nominalization) issues were positively correlated with the low influence group. At the other end of the spectrum, Theory was positively correlated with authors in the high influence group, suggesting that scholars writing in that vein were very likely to receive high PageRank scores, a finding in line with the earlier one that Miscellaneous themes are positively correlated with the high influence group-Theory belongs to the Miscellaneous category.
The second model indicates that 20 keywords were significantly associated with the In-Degree measure. The largest effect was that Language-related words were more likely to be found in the Middle group, which coincides with the theme profile analysis. However, the Lasso regression was not able to detect that Socio-cultural themes predicted placement in the High In-Degree group. A possible explanation for this is that many Socio-cultural words were not significant on their own but collectively contributed to boosting an author into the high influence group.

Conclusion
The present study demonstrates that CIS authors do indeed form discrete clusters among themselves, but the norms that usually govern the clustering of Liberal Arts and Natural/ Empirical Sciences scholars (Moser-Mercer, 1994;Gile, 2005) cannot be used for classifying these groupings. Close examination of the citation data revealed that the majority of members of each cluster displayed one or more of the following defining characteristics: (1) their areas of research were similar; (2) they were influenced by the same theory; (3) they authored or cited 'classic' textbooks that contain the established fundamentals of the subject. All of which indicates that the CIS community is a diverse one, with scholars forming into groups based on their shared characteristics. Despite this diversity, a small number of individuals stood out as the most influential. While the top 30 Western authors exerting the most influence in CIS had a wide range of distinct areas of interest and expertise, all but one of the Chinese top 30 specialized in research into interpreting. It is also worth noting that a substantial proportion of the overall total of researchers had several of their works cited, while a minority was influential thanks to only one or two publications.
This paper also contributes to better understanding how research topics are associated with a CIS author's influence. It concluded that authors writing about non-mainstream topics (i.e. Miscellaneous themes) were more likely to be found in the high-influence PageRank group than those tackling 'bread-and-butter' subjects, and those writing on Socio-cultural themes were more likely to be placed in the high-influence In-Degree group. The study also identified several keywords significantly correlated with an author's network measures: Theory (high-influence PageRank group); Attention and Nominalization (low-influence PageRank); Language-related keywords (middle-ranking In-Degree); and Sensory Memory and Trade Association (highranking Out-Degree). The findings for keywords were broadly in line with those for themes, suggesting that authors who wish to make their mark in the academic community would do well to embrace certain topics while avoiding others.
When Interpreting Studies was in its infancy in the 1960s only a handful of isolated authors, scattered throughout the world, were conducting research (Gile, 2013a); today, despite its still relative youth in comparison with 'old timers' such as linguistics, philology, etc., it is well on the way to becoming a mature discipline in its own right (Moser-Mercer, 2011), and China's contribution to its rapid development has undoubtedly been considerable. The aim of this scientometric survey has been to provide a panorama of the evolution of Chinese Interpreting Studies while demonstrating the merits of blending traditional citation analysis with Social Network Analysis to produce such a survey. It is hoped that its findings might help authors better appreciate the trade-offs they need to make when choosing research topics and the potential academic influence that may result from their choices, as well as, more importantly, offering policy-makers new insights and food for thought as they chart the future course of CIS research.