This invited paper summarizes a recent synthesis [8] on the topic of scienti c domain delineation. See also [5].
For decades, Bibliometrics and Information Retrieval developed an increasingly close relationship. Information retrieval (namely its application to science and technology) rst emphasized data organization, indexing and query systems.
Bibliometrics, driven by exploration of science networks (Price) or evaluation of research systems, was boosted by Gar eld's \citation index", a co-word which symbolizes the match between IR and bibliometrics. Kessler's work was a beacon of their opening to mapping techniques.
Bibliometric investigation, whether for cognitive or for evaluative purposes, typically targets particular \domains" de ned at some scale. The meso-level scope covers sub-disciplines, elds, large research areas; the frontier with the micro-level of research fronts or topics is quite fuzzy. Field delineation is usually understood in terms of publication sets. Delineation requirements may be quite basic or more demanding, depending on the study's ambitions and the type of domain. High-level requirements are associated to large projects on \dif
cult" elds such as emerging, transversal or complex areas. Operationalization of delineation typically exploits three models, on their own or in combination.
All of them rely on quantitative methods in statistics-probability, data analysis (clustering and factor techniques) and graph-network theory.
Model A covers institutional classi cations and nomenclatures that stem from the cooperation between scientists, S&T policy experts, librarians and information scientists. Examples: aggregate classi cation from a few national and international bodies; national or international patent o ce classi cations; classi cations from producers of disciplinary or general databases | with various degrees of disaggregation. A distinction should be made between institutional, national and international schemes of classi cation. Schemes that are linked to national or institutional evaluation, especially, are politically loaded and even \international" standards may be biased towards the points of view of stakeholders. In this respect, the way in which disciplines are grouped or broken down bears witness to social and national stakes at a given period. Given a bibliometric project, prede ned categories sometimes happen to meet the requirements, although this is not generally the case on transversal or emerging areas.
Model B, in contrast with this sedimented knowledge, relies on ad-hoc IR investigations in a particular query system. In some cases, this appears as a re
nement brought to existing groupings. For example, whereas some databases o er categories based on journals lists, variants may be established using some expertise | the same goes for patent categories. It is quite common to delineate relatively simple domains in this way, keeping in mind the limitations of Bradfordian ranked lists. A better compromise requires lexical query (or citation) facilities. A domain of science or technology is seldom captured by global terms so that it becomes necessary to combine restricted queries, using the various adaptive techniques of up-to-date IR, in order to reach reasonable precision and recall.
Model C is based on bibliometric mapping and clustering, which makes \invisible colleges" visible through a variety of bibliometric networks (actors, texts, citations, etc.). A domain is then expected to emerge as a delimited area on a map or graph at the proper scale. In practice, the mental expectations of users seldom match a single deus ex machina mapping exercise, which inevitably conveys implicit points of view and technical artefacts. Variable adaptive combinations of top-down phases (cutting the domain out of an extended map of science) and/or bottom-up ones (aggregation of themes in low-level maps) will help. Mapping techniques exploit developments in clustering and community detection (graph unfolding: Louvain, Infomap, SLMA. . . ; spectral methods: LSA, pLSA, LDA. . . ) as well as classical clustering and factor analysis.
The modes of supervision (by commissioners, bibliometricians, scientists from the domain. . . ) are crucial to validate the ndings at various stages. Model A mostly uses \frozen" IR knowledge, which does not spare discussion on its pertinence in the particular case. Model B queries require experts' mental representation of the eld, if only to avoid missing signi cant subareas. Model C has to deal with peculiarities of the mapping methods. Border areas, particularly, need supervision.
Multi-networks. Scienti c IR as well as bibliometrics bring into play the various networks that are explicitly or implicitly created by published S&T: actors, papers, texts, citations, classi cation categories, sometimes funding information, etc. Citation certainly remains the iconic network of bibliometrics, while IR tradition mostly exploited nomenclature-based indexes and lexical terms. The Internet era shu ed the cards, with the rise of webometrics. Quasi-citations (URL linkages) migrated from journal bibliometrics to Google; PageRank then irrigated webometrics and bibliometrics. The cognitive theory of scienti c inforCopyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
BIR 2020, 14 April 2020, Lisbon, Portugal.
130
mation [
Thus, many producers of scientometric indicators rely on the default SCI1 or Scopus classes, or ad-hoc lists of journals, for a rough delimitation. Pragmatic mixes improve the process: lexical or nomenclature index queries; supervised lists of prominent authors, or most cited ones, in multistage and adaptive protocols.
For example, a high-precision detection of a core literature may be completed by various recall-oriented extensions (query-expansion, network proximity, etc.) with the possible help of science maps. Protocols involving multi-network design are fruitful. Two networks are generally held as the most precise for thematic analysis: lexical terms (words/phrases from controlled or natural language) and paper-level citations; the author networks are also valuable. Citations and words exhibit analogies and also di erences: granularity level, statistical properties, uni cation issues especially in natural language, dynamic capability | more direct for citation. Citation biases are severe in an evaluation context but somewhat less in a mapping context. Lexical analysis is prone to natural language traps, homonymy, synonymy, metaphors, etc. There is huge literature on the properties of the two universes.
Sequential word-citation protocols give a few examples of how themes
delimitation and mapping take advantage of those properties. Among those mentioned
in the aforementioned synthesis, the sequence that starts with careful core
building, using supervised lexical queries, and that goes on with a quasi-automatic
citation-based extension, proved e cient ([
Alternatively, one may rst compare techniques, by exploring the
citationway and the lexical way in parallel, and then compare and possibly combine
the results. We explore this at a cluster delineation level, namely the detection
of areas within a given domain (beforehand delimited by hybrid techniques:
nanoscience, genomics. . . ). We established that word and citation techniques
shape fairly convergent breakdowns but without coincidence, a phenomenon well
characterized in cross-maps [
The \full hybrid" approaches, where citation and text tokens are mixed, di er
from the mildly hybrid ones. Even in naive form, hybrid bibliographic coupling
outperformed simple b.c. in Boyack and Klavans studies [
Most classical techniques rely on citation keys and, in the lexical realm, on word n-grams or more sophisticated linguistic analysis. Instead, featureless representation of information such as character n-grams or bit sequences, proved e cient in our experience. Related n-gram or compression distances are used.
1 SCI, WoS, WoK product from ISI, Thomson, Thomson-Reuters, now Clarivate.
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
BIR 2020, 14 April 2020, Lisbon, Portugal.
131 The cost is the black-box e ect, losing any direct semantic interpretation. This can be operated eld by eld (e.g. actors/ all text/ references) or for the whole document. In the latter case, the representation is savagely hybrid.
Delineation ultimately aims at de ning a domain in terms of documents,
but this objective may be met through indirect paths (rather than inter-articles
coupling) and probabilistic model. Citation-in-context studies, following Small's
investigations ([
As previously stated, the frontier with micro-level small topic delineation is fuzzy. Meso-scale bibliometric studies, which involve a breakdown into many topics, cannot generally a ord detailed supervision at this micro-level, where themes are detected by some automatic data analysis. However, using crossmaps or hybrid designs, may be helpful to look for robust \strong forms" or at least cores in multinetworks, resisting to change of settings and vantage points.
Supervision is nevertheless necessary at the key stages, and costly. For both
validity, acceptability and involvement of actors, the careful dimensioning and
conduct of supervision is a key factor in large studies.
An extensive bibliography is found in [
BIR 2020, 14 April 2020, Lisbon, Portugal.
132