Abstract: A tool derived for the technological watch Tetralogie (http://atlas.irit.fr) is applied to a dataset of three consecutive years of article abstracts published in the European Journal Astronomy and Astrophysics. This tool is based on a two step approach: first, a pretreatment step that extracts elementary items from the raw information (keywords, authors, year of publication, etc.); second, a mining step that analyses the extracted information using statistical methods. It is shown that this approach allows one to qualify and visualize some majors trends characterizing the current astronomical literature: feature of multi-author collaborative work, impact of observational projects, thematic maps of publishing authors.
Electronic publication has become an essential aspect of the distribution of astronomical results (see Heck, 1997). The users who want to exploit this information need efficient information retrieval systems in order to retrieve relevant raw information or to extract hidden information and synthesize thematic trends. The tools providing the former functionalities are based on query and document matching (Salton et al. 1983, Frakes et al. 1992; Eichhorn et al. 1997). The latter functionalities (called data mining functionalities) result from data analysis, data evolution analysis and data correlation principles and allow one to discover a priori unknown knowledge or information (Shapiro et al. 1996). We focus on these latter functionalities.
A knowledge discovery process can be broken down into two steps: first, the data or information selection and pre-treatment; second, the mining of these pieces of information in order to extract hidden information. The main objectives of the mining are to achieve classification (i.e. finding a partition of the data, using a rule deduction from the data characteristics), association (one tries to find data correlations) and sequences (the objective is to identify and to find the temporal relationships between the data). The information resulting from an analysis have then to be presented to the user in the most synthetic and expressive way, including graphical representation.
TETRALOGIE is an information mining tool that has been developed at the Institut de Recherche en Informatique de Toulouse (IRIT). It is used for science and technology monitoring (Dousset et al. 1995; Chrisment et al. 1997) from document collections. In this paper it is not possible to present in detail all the system functionalities. We will rather focus on some key features and on an example of the results that can be obtained from astronomical records.
The information mining process is widely based on statistical methods and more precisely on data analysis methods. First, the raw information have to be selected and pre-treated in order to extract the relevant elements of information and to store them in an appropriate form.
Different kinds of crossing tables can be performed according to the kind of information one wants to discover, for example:
The different mining functions used are described in Chrisment et al. (1997). They include: Principal component analysis (PCA), Correspondence Factorial Analysis (CFA), Hierarchical Ascendant Classification, Classification by Partition, Procustean Analysis.
In that abstract sample, it is possible to extract about 1600 different authors, and 200 can be selected as the most prolific. Topics of the documents can be extracted either from the title, from the keywords or from the abstract field. Titles have been considered as too short to be really interesting for the study. In addition, the use of keywords was considered too restrictive, as they belong to a controlled set. Indeed, we preferred to automatically extract the different topics from the words or series of words contained in the abstracts.
The first crossing was done using all the authors. A first view is obtained by sorting the author correlations in order to find the strong connexities. One can see in Figure 1 strongly related groups (about 15 groups appear on the diagonal). They are almost all weakly linked via at least one common author (see the several points above and below the diagonal line, linking the blocks). The isolated groups appear on the bottom right corner. This strong connexity is typical of a scientific domain including large international projects and strong cooperative links.
Figure 1 : Overview of the sorted connexity table, displaying on the diagonal about 15 groups of collaborating authors (see text).
One can go further and study in depth one of these collaborative groups: Figure 2 shows, for instance, some features of the collaborative work around the Hipparcos project in the years 1994-96 (i.e. before the publication of the final catalogues) as can be viewed by grouping together authors having papers co-authored with M. Perryman (Hipparcos project scientist) and L. Lindegren (leader of one of the scientific consortia). The system allows to extract 25 main authors (with more than two publications, and at least one with one of the selected central authors) and to cross them with all possible co-authors.
The figure illustrates these correlations: the center is crowded with the `core' co-authors, and on the periphery appear those who have specific contributions, with their own external co-authors.
Figure 2: View of the `Hipparcos' collaborative environment in 1994-96, resulting from a CFA of the (main author / author) crossing.
In Figure 3, we illustrate, for the authors of the Hipparcos collaboration extracted in the previous section, the main keywords of the `peripheral' authors (those who bring specific outside collaborations).
Figure 3 : Crossing main authors of the `Hipparcos' collaboration with topical key words.
In this paper, we have tried to show the usefulness of the TETRALOGIE system for discovering trends in the astronomical literature. We focused on several functionalities of this tool that allow one to find some hidden information such as the teams (through the multi-author collaborative work) or the topical maps of publishing authors. This tool graphically displays the discovered relationships that may exist among the extracted informations.
Schulman et al., 1997, using classical statistical approaches, have extracted significant features from an analysis of subsequent years of astronomy literature. In a forthcoming study, we will show how the Tetralogie system can also be used to discover thematic evolutions in the literature over several years.
The Web version (http://cdsweb.u-strasbg.fr/publi/tetra-1.htx) contains additional figures for illustration.