Reading bibliometric data into bibnets

1. Introduction and the standard schema

Each network builder in bibnets (author_network(), keyword_network(), reference_network(), document_network(), source_network(), country_network(), institution_network(), conetwork()) requires a data frame with a fixed set of columns and a small number of list-columns holding multi-valued fields. The readers convert source-specific exports (Scopus CSV, Web of Science plaintext, OpenAlex flat or nested, Dimensions, Lens.org, BibTeX, RIS, Crossref) into this common representation.

The standard schema returned by every reader is:

Source-specific extras (e.g. index_keywords, keywords_plus, affiliations, countries, language) follow the standard columns. The contract that downstream functions such as build_bipartite() and author_network() rely on is that each multi-valued field is stored as a list-column whose elements are character vectors.

This vignette covers all nine readers, the read_biblio() entry point, the generic-CSV path, the split_field() helper, and the manual construction of bibnets-compatible data frames.

2. read_biblio()

read_biblio() accepts a single file, a vector of file paths, or a directory. When format = "auto" (the default) it detects the format from the contents of the file:

data <- read_biblio("export.csv")          # auto-detect format
data <- read_biblio("scopus_dir/")         # entire directory, rbind'd
data <- read_biblio(c("a.csv", "b.csv"))   # multiple files, rbind'd
data <- read_biblio("file.csv", format = "scopus")   # force a format

When given a directory, read_biblio() collects every .csv, .txt, .bib, .ris, .xls, and .xlsx file in it, reads each one, and combines the results with rbind(). For more than one file a summary message is emitted:

Read 3 files: 1247 rows total

Format detection is performed on the first non-empty line of the file:

• BibTeX: line 1 begins with @
• RIS: line 1 begins with TY -
• Web of Science plaintext: line 1 begins with FN or PT
• CSV-based: detection inspects the header row. When the first line matches the Dimensions preamble ("About the data: ..."), line 2 is used instead. Header tokens determine the format: eid for Scopus, lens id for Lens.org, publication id or dimensions url for Dimensions, authorships.author.display_name for the OpenAlex flat CSV.

If detection fails, read_biblio() raises an error that lists the supported formats and indicates how to pass format explicitly or use format = "generic" with actors.

Two readers are not dispatched by read_biblio():

• read_openalex() accepts an in-memory tibble from openalexR::oa_fetch(), not a file path.
• read_crossref() accepts the data element of rcrossref::cr_works().

Both take R objects rather than files and are called directly.

3. Worked example — OpenAlex flat CSV

The package includes a 30-row OpenAlex flat CSV at inst/extdata/openalex_works.csv, corresponding to the export produced by downloading “Works” results from the OpenAlex web interface. Multi-valued fields use | as the delimiter.

f <- system.file("extdata", "openalex_works.csv", package = "bibnets")
oa <- read_openalex_csv(f)
str(oa, max.level = 1)
#> 'data.frame':    30 obs. of  13 variables:
#>  $ id            : chr  "W2769342982" "W2264893711" "W2612059685" "W3118164373" ...
#>  $ title         : chr  "Open University Learning Analytics dataset" "Educational Data Mining and Learning Analytics in Programming" "Predicting Student Performance using Advanced Learning Analytics" "Predicting Student Performance Using Data Mining and Learning Analytics Techniques: A Systematic Literature Review" ...
#>  $ year          : int  2017 2015 2017 2020 2022 2016 2020 2024 2016 2020 ...
#>  $ journal       : chr  "Scientific Data" "" "" "Applied Sciences" ...
#>  $ doi           : chr  "10.1038/sdata.2017.171" "10.1145/2858796.2858798" "10.1145/3041021.3054164" "10.3390/app11010237" ...
#>  $ cited_by_count: int  432 312 235 417 247 163 122 133 131 177 ...
#>  $ abstract      : chr  NA NA NA NA ...
#>  $ type          : chr  "article" "article" "article" "article" ...
#>  $ authors       :List of 30
#>  $ references    :List of 30
#>  $ keywords      :List of 30
#>  $ affiliations  :List of 30
#>  $ countries     :List of 30
head(oa[, c("id", "title", "year", "journal", "type")], 5)
#>            id
#> 1 W2769342982
#> 2 W2264893711
#> 3 W2612059685
#> 4 W3118164373
#> 5 W4300484403
#>                                                                                                                title
#> 1                                                                         Open University Learning Analytics dataset
#> 2                                                      Educational Data Mining and Learning Analytics in Programming
#> 3                                                   Predicting Student Performance using Advanced Learning Analytics
#> 4 Predicting Student Performance Using Data Mining and Learning Analytics Techniques: A Systematic Literature Review
#> 5                           Artificial Intelligence and Learning Analytics in Teacher Education: A Systematic Review
#>   year            journal    type
#> 1 2017    Scientific Data article
#> 2 2015                    article
#> 3 2017                    article
#> 4 2020   Applied Sciences article
#> 5 2022 Education Sciences  review

The list-columns:

oa$authors[[1]]
#> [1] "Jakub Kužílek"  "Martin Hlosta"  "Zdeněk Zdráhal"
oa$affiliations[[1]]
#> [1] "The Open University"                 
#> [2] "Czech Technical University in Prague"
#> [3] "The Open University"                 
#> [4] "The Open University"                 
#> [5] "Czech Technical University in Prague"
oa$countries[[1]]
#> [1] "CZ" "GB" "GB" "CZ" "GB"

Two limitations of the flat export should be noted:

all(vapply(oa$references, function(x) length(x) == 0 || all(is.na(x)), logical(1)))
#> [1] TRUE
all(is.na(oa$abstract))
#> [1] TRUE

references is empty and abstract is NA because these fields are not included in the OpenAlex web download. Cited references and abstracts from OpenAlex are obtained via openalexR::oa_fetch() and processed with read_openalex() (described in section 6).

The remaining fields support several network constructions that do not require references — co-authorship, country, institution, keyword, source, and document networks:

co <- country_network(oa, counting = "fractional")
head(co, 5)
#> # bibnets network: country_collaboration | 8 nodes · 5 edges | counting: fractional 
#>    from  to  weight  count
#> 1  GB    NO     1.5      3
#> 2  CA    US   1.167      2
#> 3  AU    CN       1      1
#> 4  AU    EC       1      1
#> 5  CZ    GB       1      1

4. Scopus

sc <- read_scopus("scopus.csv")

read_scopus() ingests the standard Scopus CSV export (File -> Export -> CSV from the Scopus search UI). Mappings from Scopus columns to the bibnets schema:

Scopus stores each cited reference as one semicolon-delimited string in a single cell. read_scopus() splits on ; and applies standardize_refs() to each entry: uppercasing, whitespace normalisation, and removal of a trailing DOI where present. References differing only in case or trailing DOI then resolve to the same node in co-citation and reference networks.

5. Web of Science

WoS exports come in two shapes:

wos1 <- read_wos("savedrecs.txt")                       # plaintext (default)
wos2 <- read_wos("savedrecs.tsv", format = "tab")       # tab-delimited

The plaintext format is a tagged record syntax. Each record begins with a PT (publication type) tag and ends with ER (end record). Within the record, every field is introduced by a 2-letter tag at the start of a line, with continuation lines indented:

read_wos() walks the file, splitting on ER boundaries, and emits one row per record. The tab-delimited variant carries the same fields in a flat CSV-like grid. Either way the output schema is identical.

6. OpenAlex — two paths

OpenAlex ships data through two routes that bibnets supports separately.

library(openalexR)
raw  <- oa_fetch(entity = "works", search = "learning analytics", per_page = 200)
data <- read_openalex(raw)

7. Dimensions

dm <- read_dimensions("dimensions_export.csv")

The Dimensions CSV begins with a metadata row of the form

"About the data: This export was generated on YYYY-MM-DD ..."

before the column header. read_dimensions() detects this preamble and skips it. If the line has been removed (for example, by manual editing of the file), the reader continues to function because it identifies the column row by the Dimensions header tokens Publication ID and Dimensions URL.

Extras returned: affiliations and countries as list-columns, analogous to the OpenAlex schema.

8. Lens.org

ln <- read_lens("lens_export.csv")

Key Lens columns and how they map:

9. BibTeX & RIS

bt <- read_bibtex("library.bib")
ri <- read_ris("savedrecs.ris")

read_bibtex() parses @type{key, field = {value}, ...} blocks. read_ris() parses tagged TY - ... ER - blocks; the structure is equivalent to WoS plaintext, but with a different tag dictionary.

Standard BibTeX and RIS do not contain cited-reference data, so the references column in the resulting data frame is empty on every row. These formats are sufficient for co-authorship and keyword co-occurrence networks. For co-citation, coupling, or direct citation networks, the appropriate sources are Scopus, Web of Science, OpenAlex (via oa_fetch()), Dimensions, Lens, or Crossref.

10. Crossref via rcrossref

library(rcrossref)
raw  <- cr_works(query = "graph neural networks", limit = 100)
data <- read_crossref(raw$data)

read_crossref() accepts the data element of the cr_works() result (a data frame, not the wrapping list). The function handles the two field-naming variants Crossref returns (container.title vs container-title; is.referenced.by.count vs is-referenced-by-count) and maps both to the standard schema.

11. Generic CSV — read_biblio(format = "generic", ...)

For CSV files that do not match any of the recognised signatures (in-house exports, custom dumps, public datasets), the generic path provides explicit column-name mapping. The identifier column is named via id; columns to be treated as list-columns are named via actors. sep is the delimiter applied inside those cells.

Hypothetical call:

data <- read_biblio(
  "my_data.csv",
  format  = "generic",
  id      = "doc_id",
  actors  = c("Authors", "Keywords"),
  sep     = ";"
)

Demonstrated on the bundled OpenAlex CSV (which uses | as the delimiter):

f <- system.file("extdata", "openalex_works.csv", package = "bibnets")
generic <- read_biblio(
  f,
  format = "generic",
  id     = "id",
  actors = c("authorships.author.display_name", "primary_topic.display_name"),
  sep    = "|"
)
names(generic)[1:6]
#> [1] "id"                              "apc_list.value"                 
#> [3] "apc_paid.value"                  "authorships.author.display_name"
#> [5] "authorships.author.id"           "authorships.author.orcid"
generic$authorships.author.display_name[[1]]
#> [1] "Jakub Kužílek"  "Martin Hlosta"  "Zdeněk Zdráhal"

The named id column is copied to a top-level id. Each column listed in actors is split on sep and stored as a list-column. Other columns are retained unchanged. The resulting frame is therefore not in the standard schema; it is a wider source-specific table. Network constructors can either be pointed at the relevant columns directly, or the frame can be post-processed into the standard schema.

12. Building data manually

When data does not come from any of the supported sources, a bibnets-compatible data frame can be constructed directly. The requirement is: standard scalar columns are character or integer; multi-valued fields are list-columns whose elements are character vectors.

df <- data.frame(
  id    = c("p1", "p2", "p3"),
  title = c("Paper A", "Paper B", "Paper C"),
  year  = c(2020L, 2021L, 2022L),
  stringsAsFactors = FALSE
)
df$authors <- list(
  c("ALICE", "BOB"),
  c("BOB", "CAROL"),
  c("ALICE", "CAROL", "DAVE")
)
df$references <- list(
  c("R1", "R2"),
  c("R1", "R3"),
  c("R2", "R3", "R4")
)
df$keywords <- list(
  c("graph", "network"),
  c("network", "embedding"),
  c("graph", "embedding", "neural")
)

author_network(df, "collaboration")
#> # bibnets network: author_collaboration | 4 nodes · 5 edges | counting: full 
#>    from   to     weight  count
#> 1  ALICE  BOB         1      1
#> 2  ALICE  CAROL       1      1
#> 3  BOB    CAROL       1      1
#> 4  ALICE  DAVE        1      1
#> 5  CAROL  DAVE        1      1
keyword_network(df)
#> # bibnets network: keyword_co_occurrence | 4 nodes · 5 edges | counting: full 
#>    from       to       weight  count
#> 1  EMBEDDING  GRAPH         1      1
#> 2  EMBEDDING  NETWORK       1      1
#> 3  GRAPH      NETWORK       1      1
#> 4  EMBEDDING  NEURAL        1      1
#> 5  GRAPH      NEURAL        1      1
reference_network(df)
#> # bibnets network: reference_co_citation | 4 nodes · 5 edges | counting: full 
#>    from  to  weight  count
#> 1  R1    R2       1      1
#> 2  R1    R3       1      1
#> 3  R2    R3       1      1
#> 4  R2    R4       1      1
#> 5  R3    R4       1      1

build_bipartite() applies toupper(trimws(...)) to every entity label before constructing the sparse matrix, so "graph", "Graph", and "GRAPH" are mapped to the same node "GRAPH". Tests or comparisons that reference node names should use uppercase strings.

13. The split_field() helper

split_field() converts a character column with semicolon-delimited (or otherwise delimited) values into a list-column without going through read_biblio(format = "generic"):

split_field(c("Alice; Bob; Carol", "Dave; Eve"))
#> [[1]]
#> [1] "Alice" "Bob"   "Carol"
#> 
#> [[2]]
#> [1] "Dave" "Eve"
split_field(c("a|b|c", "d|e"), sep = "|")
#> [[1]]
#> [1] "a" "b" "c"
#> 
#> [[2]]
#> [1] "d" "e"

This is the same operation that read_scopus() and the other readers apply internally to multi-valued columns; it is exported for use in custom pipelines.

14. Combining data from multiple sources

Different readers expose different extras: WoS provides keywords_plus, Scopus provides index_keywords, OpenAlex provides countries. To combine sources, restrict each frame to the standard columns and bind:

common <- c("id", "title", "year", "journal", "doi", "cited_by_count",
            "abstract", "type", "authors", "references", "keywords")

data(biblio_data)
b1 <- biblio_data
b2 <- biblio_data
b2$id <- paste0(b2$id, "_dup")

cols <- intersect(common, names(b1))
combined <- rbind(b1[, cols], b2[, cols])
nrow(combined)
#> [1] 20

Two practical notes:

1. When document IDs overlap across sources (which occurs when Scopus and WoS both index the same article), prefixing the IDs as shown prevents duplicate documents from inflating co-occurrence counts.
2. Source-specific extras (e.g. WoS keywords_plus) should be retained on the per-source frame and merged selectively rather than coerced into the combined frame.

15. Inspecting and sanity-checking

After reading, basic checks on the list-column sizes and the scalar columns help detect silent corruption. Empty list-columns and out-of-range years are common indicators that an export is incomplete.

data(scopus_quantum_cloud)
sc <- scopus_quantum_cloud

range(lengths(sc$authors))
#> [1]  0 40
range(lengths(sc$references))
#> [1]   0 245
range(lengths(sc$keywords))
#> [1]  0 20

head(sort(table(sc$journal), decreasing = TRUE), 5)
#> 
#> IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 
#>                                                                            24 
#>                   IEEE Transactions on Circuits and Systems I: Regular Papers 
#>                                                                            20 
#>                                                                   IEEE Access 
#>                                                                            18 
#>              IEEE Transactions on Very Large Scale Integration (VLSI) Systems 
#>                                                                            14 
#>                                               IEEE Internet of Things Journal 
#>                                                                            12
range(sc$year, na.rm = TRUE)
#> [1] 2020 2025
table(sc$type)
#> 
#>           Article              Book      Book chapter  Conference paper 
#>               279                 1                15               191 
#> Conference review            Review 
#>                 3                10

Indicators to check:

• A lengths() of 0 on every row of references for a Scopus or WoS file indicates that the export did not include the references column. Re-export from the source with the references field selected.
• A year of 0 or NA indicates an empty source field.
• A single dominant document type (e.g. only "article") is expected for filtered searches; broader mixes are expected for thematic searches.

16. Troubleshooting

Further reading

• The companion vignette, vignette("bibnets"), covers network construction on the in-package datasets.
• All network builders (author_network(), keyword_network(), reference_network(), document_network(), source_network(), country_network(), institution_network(), conetwork()) accept the same set of arguments (type, counting, similarity, threshold, top_n, format), so switching between network types on data already in the standard schema requires only a function-name change.