Assessing Author Research Focus Using Vector Space Modeling Kun Lu School of Information Studies U. of Wisconsin-Milwaukee P.O. Box 413, Milwaukee, WI USA 53201 e-mail: kunlu@uwm.edu ABSTRACT A method for comparison of research focus by a group of authors using vector space modeling is presented. The body of published work by an author or group of authors in a given field may be represented as a vocabulary based on language use. Using vector space modeling, publications may be represented as vectors, resulting in a topic space. The works of a given author may be mapped onto the space, resulting in an author space. The density of a given author space provides an indication of the coherence of the author’s body of work, where a high density author space is indicative of a more focused research agenda and a low density author space indicates more varied areas of investigation or language use. This concept was applied to a set of publications appearing in eight high impact factor journals in information science from which the 100 most prolific authors were identified. Their author space characteristics were determined and compared to identify those with the most focused research. The findings demonstrate significant correlations in authors’ numbers of publications, vocabulary usage, and resulting author space density, with some exceptions. Keywords Informetrics, Author space, Vector space modeling INTRODUCTION The purpose of this research is to demonstrate the application of word analysis to publications in a vector space environment to identify the level of focus of authors’ research. To date, studies have relied on word analysis for mapping of disciplinary structure (e.g., Braam, Moed, van Raan, 1991a, 1991b; Janssens, Leta, Glänzel, & De Moor, 2006), but to the best of our knowledge not for author analysis using vector space modeling and author space This is the space reserved for copyright notices. ASIST 2011, October 9-13, 2011, New Orleans, LA, USA. Copyright notice continues right here. Dietmar Wolfram School of Information Studies U. of Wisconsin-Milwaukee P.O. Box 413, Milwaukee, WI USA 53201 e-mail: dwolfram@uwm.edu density outcomes. In the vector space model, publications related to a certain topic that are included in an indexed data set constitute a topic space. The focus of an author’s research output can be represented by similarities among his/her works. If an author’s publications are similar to each other, his/her research agenda is assumed to be more focused and vice versa. A mathematical representation of this measure can be thought of as a density, where a highly dense space represents higher homogeneity in the objects and thus more focus in an author’s research output. A low density space indicates greater variation. The space density is defined as the average similarities between each object in the space and the centroid, or average document vector. The mathematical representation of the density is: Density ( S ) Sim( D, C ) DS n (1) where Sim() is the similarity function, cosine is used as the similarity function in the current study, S is the author space, C is the centroid of that space, D is the document in the space and n is the number of objects in the space. We note that vocabulary conventions vary from field to field. Some fields have more limited vocabularies than others. However, we believe that the consistent use of a similar vocabulary in a similar way provides evidence of a focused research agenda. Note that the density calculation takes into account the number of publications, so outcomes are not affected by the number of publications produced by an author. More information about densities in an indexed space can be found in Wolfram & Zhang (2008). The author space density provides a measure by which different authors may be compared for the extent of their research focus. We hypothesize that highly prolific authors will have a lower author space density because they are more likely to contribute to a broader range of topics. Authors with a higher density space are believed to have a more focused or narrower research agenda. THE VECTOR SPACE MODEL FOR INFORMATION RETRIEVAL Author Pubs. (Rank) Density (Rank) VocSize (Rank) THELWALL, M 97 (1) 0.318 (91) 1482 (1) EGGHE, L 74 (2) 0.287 (98) 937 (6) GLANZEL, W 61 (3) 0.310 (92) 965 (5) ROUSSEAU, R 61 (3) 0.244 (100) 1008 (4) LEYDESDORFF, L 60 (5) 0.328 (87) 1092 (2) JACSO, P 56 (6) 0.291 (94) 799 (11) OPPENHEIM, C 48 (7) 0.252 (99) 1020 (3) SPINK, A 48 (7) 0.378 (72) 907 (7) BAR-ILAN, J 40 (9) 0.322 (90) 875 (9) NICHOLAS, D 34 (10) 0.384 (64) 889 (8) BORNMANN, L 31 (11) 0.377 (73) 724 (16) DANIEL, HD 30 (12) 0.386 (62) 682 (21) METHOD HUNTINGTON, P 30 (12) 0.392 (58) 832 (10) Journals with the highest impact factor in the category of “information science & library science” appearing in the Journal Citation Report 2009 Social Sciences Edition were identified. Journals associated with allied subject areas such as Management Information Systems and Medical Informatics, were excluded. Eight journals were selected for inclusion in the study (Journal of Informetrics, Annual Review of Information Science and Technology, Journal of the American Society for Information Science and Technology (Not including JASIS), Scientometrics, Information Processing & Management, Journal of Information Science, Online Information Review, and Journal of Documentation). Bibliographic records for documents published in these journals between 2000 and 2010 were downloaded. Records downloaded were further limited to three document types: articles, proceedings papers and reviews. The other document types were less likely to represent research contributions by the authors. In total, 5227 records were downloaded from Thomson Reuters Web of Science (WoS). The raw WoS records were processed and only three fields were kept, namely, the article title (i.e. “TI” field), the Keywords Plus (i.e. “ID” field) and the abstract (i.e. “AB” field). The records then were indexed with a widely used Lemur information retrieval toolkit. Stop words were removed and Porter stemming was applied. JANSEN, BJ 27 (14) 0.438 (25) 690 (19) SCHUBERT, A 27 (14) 0.290 (96) 522 (41) JARVELIN, K 26 (16) 0.339 (85) 790 (12) CHEN, HC 24 (17) 0.291 (95) 760 (13) FORD, N 24 (17) 0.384 (65) 738 (15) WILSON, CS 24 (17) 0.308 (93) 721 (17) ZHANG, J 24 (17) 0.322 (89) 580 (28) BURRELL, QL 23 (21) 0.391 (59) 498 (47) CRONIN, B 23 (21) 0.288 (97) 510 (43) VAUGHAN, L 22 (23) 0.440 (24) 602 (25) MEYER, M 21 (24) 0.481 (11) 670 (21) MOED, HF 21 (24) 0.373 (76) 661 (23) The vector space model is one of the most influential models in information retrieval (Salton and McGill, 1983). In this model, each document is represented as a vector and the elements of the vector consist of words appearing in the collection. The document vectors in a collection constitute a document term matrix. The value of each element represents the term significance in the document. By virtue of the vector space model, documents are transformed into vectors. Traditional measures like angle (e.g. based on a cosine measure) and distance (e.g. Euclidean distance) can be used to measure the similarity between documents. In the vector space, a number of documents constitute a document space. In the current study, each author will be viewed as a document space consisting of the articles he/she has written. This space is named the author space. From the 5227 records downloaded, we were able to identify 6282 different author names using string matching. Because it is impractical to list all of the authors in our collection, we focused on the 100 most prolific authors according to WoS “analyze results” function. We selected the most prolific authors because the more an author writes the better the algorithm used “understands” his/her interests, and thus the more accurate our measure will be. The 5227 records served as the basis of the topic space, TF*IDF term weighting was employed to assign term significance in the space. Terms that were single characters or only consisted of digits (e.g. “2001”) were filtered out. It Table 1. Author Space Density Outcomes – 25 Most Prolific Authors is believed that these terms add noise into the space rather than meaning. The author space for each of the 100 most prolific authors was generated, which consisted of all the articles an author wrote represented in the topic space. RESULTS A summary table for the 25 most prolific authors identified appears in Table 1. Pubs represents the number of publications included in the author space. VocSize represents the number of indexed content-bearing terms or vocabulary size for an author. Density is defined as in equation (1). Ranks are included in parentheses. There is a significant negative correlation (Kendall’s tau = -0.466, α<0.01) between productivity and author space density, where more prolific authors are more likely to have a less dense author space (Figure 1). There is also a significant positive correlation between vocabulary size and number of publications (Kendall's tau = 0.628, α<0.01) and a significant negative correlation between vocabulary size and author space density (Kendall's tau = -0.655, α<0.01). Among the 100 most prolific authors, the authors with the 10 lowest density values all appear in the top 25 list for productivity, lending evidence that the authors have contributed to a broader range of areas. This agrees with our hypothesis; however, there are exceptions among prolific authors. Of note, three prolific authors who appear in the top 25 (Jansen; Vaughan; Meyer), also have high author density values that appear in the top 25, suggesting a more focused research agenda, or a more consistent use of vocabulary in their research output. In addition, several authors (Schubert, Burrell, Cronin) demonstrate that a large number of publications and a relatively low author space density need not translate to an equally large indexed vocabulary. Density 0.6 judgment for a given author’s contributions. It is a reflection of the language use in a given author’s research, which can serve as a means to identify the breadth of topics he or she has undertaken. Furthermore, the density value should not be considered in absolute terms, but instead as a comparative measure. Density outcomes will depend on the documents and authors included in the topic space, which influence the characteristics of the author space. In the case of the current study, the data may not reflect the full oeuvre (body of work) of the identified authors because only eight journals were used. The developed approach also has applications to explore how authors’ breadth of language use in their research changes over time. This aspect may be studied in a future investigation. 0.5 REFERENCES 0.4 Braam, R.R., Moe, H.F., & van Raan, A.F.J. (1991a). Mapping of science by combined co-citation and word analysis. 1. Structural aspects. Journal of the American Society for Information Science, 42(4), 233-251. 0.3 0.2 0.1 0 0 20 40 60 80 100 # of Publications Figure 1. Scatter plot for author productivity and author space density DISCUSSION & CONCLUSION This study has demonstrated how one may use vector space modeling, classically applied in information retrieval, to study scholarly communication to compare the breadth of prolific author contributions to a topic area. The density of a given author space should not be considered a value Braam, R.R., Moe, H.F., & van Raan, A.F.J. (1991b). Mapping of science by combined co-citation and word analysis. 1. Dynamical aspects. Journal of the American Society for Information Science, 42(4), 252-266. Janssens, F. Leta, J. Glänzel, W., & De Moor, B. (2006). Towards mapping library and information science. Information Processing & Management, 42, 1614–1642. Salton, G., & McGill, M. J. (1983). Introduction to modern information retrieval. New York: McGraw-Hill. Wolfram, D., & Zhang, J. (2008). The influence of indexing practices and term weighting algorithms on document spaces. Journal of the American Society for Information Science and Technology, 59(1), 3-11.