Skip to content

Experiments to determine a new BibTeX parser formula for Citation.js -- to be applied to other formats as well

License

Notifications You must be signed in to change notification settings

citation-js/bibtex-parser-experiments

Repository files navigation

BibTeX Parser Experiments

Experiments to determine the new BibTeX parser formula. The result may be applied to other formats as well in the future.

Parsing stages of BibTeX

Above is some pseudo-grammar describing the stages of parsing. The reason I distinguish between parsing the file and parsing values is best demonstrated in the "entry value with mid-command concatenation" test case:

Input:

@book{a,
  title = "foo \\copy" # "right{} bar"
}

Output:

[{
  type: 'book',
  id: 'a',
  properties: {
    title: 'foo © bar'
  }
}]

Participants

Note: citationjs-idea (Citation.js Idea #1) and citationjs-nearley are skipped in some tables, as they are only in here for historic reasons.

Citation.js (old)

At the time of starting these experiments, the TokenStack class was utilized, together with a simple RegExp that tokenizes commands.

Citation.js Ideas

The idea was to explore tokenization without introducing a formal grammar, as formal grammars introduce extra build steps, runtime dependencies and large swaths of generated code. However, used as I was to the syntax of PEG.js and nearley.js, I made some unnecessarily complicated features like consumeAnyRule(), and some weird loops in the rules. This was partly due to bad tokenization.

Idea #2, currently active in Citation.js has new tokens, a simpler Grammar class and simplified rules. It also has more features, including more commands and diacritics, including more ways to write them.

Citation.js with nearley

In parallel to reworking the idea, I used the tokenizer in a nearley.js grammar, which failed miserably. This is probably the result of bad grammar-writing on my part, and not a reflection of the capabilities of nearley.js. However, an additional downside of this route is that it introduces an extra build step (nearleyc) and a runtime dependency — nearley itself.

astrocite

The astrocite-bibtex package by @dsifford uses PEG.js. It is capable of returning an AST.

fiduswriter

Fiduswriter's biblatex-csl-converter seems to perform very poorly on the larger file. However, it does return lossless values, although I am not a fan of the (lack of) difference between arrays representing single and multiple values:

[
  { type: 'text', text: 'foo' }
]
// vs
[
  { literal: [{ type: 'text', text: 'foo' }] },
  { literal: [{ type: 'text', text: 'bar' }] }
]

This causes testing such as

literal in value[0]
// or more properly
value.every(part => literal in part)

Zotero

Zotero Translators are relatively hard to use stand-alone, as they depend on a Zotero framework in the global scope. It immediately converts to Zotero API JSON while parsing the syntax. Not shown in the performance table is that Zotero requires initialization, not counting the time it takes to import files, and that this takes relatively long.

Better BibTeX for Zotero (BBT)

Using @retorquere/bibtex-parser, this performs very well. It is capable of returning an AST. I have not had a chance to test out all the parser features for literal/text/name values yet.

JabRef

JabRef is reference management software with Bib(La)TeX as the internal representation, so one can assume their support for parsing it to be pretty good. However, I have not found a way to export their internal representation in the level of detail required for passing the syntax tests. Similarly, as their program only partly supports a CLI (no stdio) and is written in Java, a performance comparison would not be very fair. For now, it is possible to test syntax features by uncommenting the jabref entry in test/feature.js and running npm run features -- parser jabref.

API Features

citationjs-old citationjs astrocite fiduswriter zotero bbt
Sync/Async sync sync sync both async both
AST output
Lossless schema¹
Lossless values
Error recovery

¹ specifically the schema used to represent data entries and not value syntax (commands, formatting), and disregarding AST

Syntax Features

Empty cells indicate a choice to follow either natbib or biblatex for certain behavior, this becomes clear from context. If both cells are empty, this is may be an error, but that should be indicated by a different test fixture. Auto-generated by npm run features -- fixtures, see also the fixture file.

citationjs-old citationjs astrocite fiduswriter zotero bbt
entry with lowercase type
entry with mixed-case type
entry with uppercase type
entry with parentheses
entry with spacing
entry with trailing comma
string key with colon
entry key with colon
entry value with annotation
entry label with number
entry label with colon
entry label with double quotes
entry value of quoted string
entry value of braced string
entry value of number ✘¹ ✘¹ ✘¹
entry value with mid-and concatenation ✘² ✘² ✘² ✘²
entry value with mid-command concatenation ✘² ✘² ✘² ✘² ✘²
entry value with sentence-casing (real title) ✘¹ ✘¹ ✘¹ ✘¹
entry value with sentence-casing (artificial title) ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
entry value with sentence-casing (markup) ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
entry value with sentence-casing (env markup) ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
entry value with markup ✘¹ ✘¹ ✘¹ ✘¹
entry value with envs ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
entry value with env overrides ✘¹ ✘¹ ✘¹ ✘¹
entry value with literal names
entry value with truncated names ✘¹ ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
entry value with extended names (biblatex)
entry value with verbatim fields
entry value with uri fields
entry value with pre-encoded uri fields
entry value with diacritics
entry value with escapes
entry value with sub/superscript
entry value with multi-argument commands
entry value with verbatim-argument commands
entry value with unbracketed-argument commands ✘¹ ✘¹ ✘¹ ✘¹ ✘¹
TODO
string with lowercase type
string with mixed-case type
string with uppercase type
string with parentheses
string value with string
string value with concatenated string
preamble with quoted string
preamble with string
preamble with concatenated string
comment before entry
comment around entry (natbib)
comment around entry (biblatex)

¹ undefined representation, actual support may vary
² very unlikely to matter

Performance

Data from npm test, as run on Travis CI.

Init Time (single entry) Time (3345 entries)
citationjs-old 0.795ms ± 2.2% 0.842ms ± 1.6% 1.97e+3ms ± 5.5%
citationjs 4.42ms ± 2.3% 0.455ms ± 1.5% 1.20e+3ms ± 11.7%
astrocite 1.73ms ± 2.2% 0.665ms ± 2.3% 2.17e+3ms ± 7.0%
fiduswriter 17.7ms ± 12.5% 7.52ms ± 24.7% 1.16e+5ms ± 4.0%
zotero 4.16ms ± 12.1% 2.13ms ± 1.3% 1.00e+4ms ± 0.8%
bbt 21.7ms ± 23.4% 1.55ms ± 6.5% 2.27e+4ms ± 5.4%

About

Experiments to determine a new BibTeX parser formula for Citation.js -- to be applied to other formats as well

Topics

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published