Bibliographic coupling edges using the titles of the references (DescriptiveBC)

The DescriptiveBC measure of (Liu 2017) . In the standard bibliographic coupling, two references are either the same or completely different. Here two different references can be partly similar if their titles share words, so that two documents can be coupled even if they cite few identical references.

Usage

coupling_descriptive(
  dt,
  source,
  ref,
  title,
  targets = NULL,
  min_weight = 0,
  output_in_character = TRUE,
  stopwords = NULL,
  max_documents = 5000
)

Arguments

dt: A data frame or a data.table of direct citations, one row per link between a citing document and a cited reference. It is not modified.
source: Character scalar. The column of the citing documents: they are the nodes of the coupling network.
ref: Character scalar. The column of the cited references.
title: Character scalar. The column with the title of the cited reference (one title per reference).
targets: Optional vector of identifiers of documents (values of source). If given, only the pairs involving at least one target are computed.
min_weight: Number. Only pairs with a weight strictly above it are returned (0 by default: all the pairs with some similarity). There is no weight_threshold here, since two documents need not share any reference.
output_in_character: Logical. If TRUE (default), the identifiers are also returned as character columns from and to, to build a tidygraph graph more easily.
stopwords: Character vector of words to ignore in the titles (NULL: none).
max_documents: Number. The maximum number of documents allowed without targets.

Value

A data.table with one row per pair of documents with a weight above min_weight: from and to (character if output_in_character = TRUE), the weight, and the original identifiers in Source and Target (Source < Target).

Details

The similarity of two references $r_1$ and $r_2$ is 1 if they are the same reference, and otherwise the Jaccard index of the sets of terms of their titles, $\mathrm{Sim}(r_1, r_2) = |T(r_1) \cap T(r_2)| / |T(r_1) \cup T(r_2)|$. The weight between documents $A$ and $B$, with reference sets $R_A$ and $R_B$, is

$$w_{AB} = \frac{\sum_{r \in R_A} \max_{r' \in R_B} \mathrm{Sim}(r, r') + \sum_{r' \in R_B} \max_{r \in R_A} \mathrm{Sim}(r, r')}{|R_A| + |R_B|}$$

(Eq. 5 of the paper). It lies between 0 and 1. If no two references are similar except identical ones, it is $2 |R_A \cap R_B| / (|R_A| + |R_B|)$, the Dice coefficient of the bibliographies.

Terms are the words of the titles, in lower case, without the stopwords that you give (for instance stopwords::stopwords("en")): without them, words such as "the" or "of" create similarities between unrelated titles. The paper does not specify other pre-processing (no stemming here). A reference without title is only similar to itself. If a reference has several titles in the table, the first is used, with a warning.

Computing cost. Unlike the other coupling measures, there is no natural set of candidate pairs: two documents sharing no reference can be similar through the words of the titles. All the pairs of documents are therefore potentially concerned, and the cost grows with the square of the number of documents. The function (i) computes the similarity of the pairs of references sharing a term with sparse matrices, and (ii) for each target document, finds the best matches of its references. With a very large corpus, give the targets (the documents for which you want the neighbours, as in the paper, where each article is compared with candidates) or restrict the corpus by time windows. Removing stop words reduces the number of similar references a lot. By default, more than max_documents documents without targets is an error.

Why use it in research

Two papers on the same question often cite different but related works (the same idea in a working paper and in a journal article, two papers on the same model). Classical coupling misses them, and the sparsity of citations makes it dominated by a few shared references (Liu 2017) . Using the words of the titles of the references recovers part of this relatedness. It also lets you say why two documents are related: see descriptive_terms(). The price is a dependence on the quality of the titles (language, abbreviations, missing titles) and a much higher computing cost, since almost any two documents can have some similarity (see below). Liu evaluates the measure on biomedical articles, against expert judgements of relatedness: check on your own field that title words are informative (in the social sciences titles are often metaphorical).

Intuition

Imagine that you compare two reading lists, and that you are allowed to match each book of the first list with its closest book in the second: the same book is a perfect match (1), a book whose title shares half of its words is a half match, unrelated titles do not match (0). The measure is the average quality of the best matches, in both directions.

References

Liu R (2017). “A New Bibliographic Coupling Measure with Descriptive Capability.” Scientometrics, 110(2), 915–935. doi:10.1007/s11192-016-2196-7 .

Examples

library(biblionetwork)
# a and b cite different references, with partly shared title words
toy <- data.frame(
  doc = c("a", "a", "b", "b"),
  ref = c("r1", "r2", "r1", "r3"),
  title = c("inflation expectations", "oil supply shocks",
            "inflation expectations", "supply shocks prices")
)
# 1 for r1; r2 and r3 share 2 of 4 terms: w = (1 + 0.5 + 1 + 0.5) / 4
coupling_descriptive(toy, source = "doc", ref = "ref", title = "title")
#>      from     to weight Source Target
#>    <char> <char>  <num> <char> <char>
#> 1:      a      b   0.75      a      b