Gubanov, Alexander Valerevich MsC, Tomsk State University, Tomsk, Russia derzhiarbuz@yandex.ru
Abstract One of the common objects describing a scientific discourse in some field is a word co-occurrence network. However, the network by itself is hard to interpret. Scholars usually need a clear and readable map of the discourse. To project an object with network topology (such that only some of pairwise similarities/dissimilarities defined) into a plane one can purposely define a pair of numerical characteristics (for example, a popular density/centrality coordinates mapping) or do it automatically using multidimensional scaling method or one of its analogs (i.e. VOS). In contrast with methods of dimension reduction in euclidean space (PCA, factor analysis), which allow scholars to choose and interpret coordinate axes on the resulting map, multidimensional scaling methods do not allow one to manage a “point of view” to effectively interpret the data. In this talk we consider a support dichotomies (reference points) method which allows scholars to choose the factors mostly influencing distances between terms. Chosen factors mostly affect map distances, obtained by multidimensional scaling or VOS. The basic idea of the method is to “stretch” all pairwise dissimilarities between two preliminary chosen poles (terms). Such an approach allows us to bypass the restriction of a graph topology where no directions can be defined (in contrast with euclidean space). Operating only by pairwise distances it is possible to exaggerate dissimilarities determined by closeness of terms to opposite poles. As an example we use the discourse of charity. Applying the method to stretch differences along “charity-philanthropy” dichotomy we obtain a map with axes interpretable by an expert in the field.