Balisage Paper: JATS Blue Lite

Introduction

The initial spark for this study was a discussion on the JATS mailing list [JATS-list 2020], when someone opposed the “watering down” of Blue, that is, vocabulary items that have only been available in Green are moving to Blue and there are multiple allowed ways to tag a given piece of content. Typically the requests to amend Blue go in the other direction, that is, to add items. But also the JATS dogma of backwards compatibility tends to favor additions as opposed to deletions of vocabulary items. “We have bent over backwards to be backward compatible for years!” [Lapeyre 2019]

Restricting choices for tagging content can be done by restricting the vocabulary or by limiting choice and optionality, depending on context. Apart from documenting allowed use and restrictions, the latter can be achieved more robustly by adapting the DTD itself or by superimposing Schematron rules [Schwarzman 2017], as JATS4R [JATS4R] does, for example.

Although subsetting a DTD may also impose context-dependent restrictions such as which elements may be present in which order, which attributes may be present and which values they may assume, this work views customizations only as a restriction of the vocabulary, that is, the elements and attributes that are permitted by the subset anywhere in the document. Although this is a very coarse and limited approach to customizations, a reduced vocabulary can be beneficial in several ways; in particular it reduces the effort necessary to develop and maintain renderers and editors for the given vocabulary. The JATS authoring subset (“Pumpkin”) has a significantly reduced vocabulary, but it is not designed for publishing since it omits most of the metadata vocabulary. Quoting [Lapeyre 2019]:

Such an attempt has already been made for the Texture editor [Texture 2017]. The proposed customization removed 228 elements and 147 attributes from Green, among them question/answers, index terms, and MathML. As will be discussed later, the authors decided to treat the entirety of MathML as a single vocabulary item. Doing so will reduce the number of elements omitted by Texture to 36 and the number of attributes to 30. This amounts to a vocabulary reduction of about 14%, far away from the stated 40% reduction target.

The authors decided to analyze the actual usage frequencies of the JATS vocabulary items. They sent an email to the mailing list [JATS-list 2021,Reinhardt 2021a], calling on publishers to either supply representative full-text XML articles or to run an analysis XSLT on their articles that produces lists of distinct element and attribute names used in the publications.

For a given collection of articles, the names of the elements and attributes that are used in the collection constitute a “de-facto customization.” This way, the authors hoped to be able to identify less popular vocabulary items and consider them for omission in a future “community consensus customization,” a synthetic customization that is both compact and well suited as a schema for creating, editing, and validating content, or as a starting point to derive the other de-facto customizations from without the nee to add many items.

How does one measure well-suitedness though? Not using available vocabulary items will not hurt, therefore customization A will be better suited as a starting point to derive customization B from if A is a superset of B. So the best synthetic customization will be the union of all element and attribute names of all collections?

If all collected articles used in fact only the same 60% of the JATS vocabulary, the stated goal would haven been reached trivially. It turns out that the collected articles use approx. 82% of the JATS vocabulary though.

In order to identify a smaller de-facto customization – for example the vocabulary used by a specific journal, a specific publisher or a specific subject area – that may serve as starting point to derive many other customizations from with as little vocabulary item additions as possible, the authors established a metric for the aptness of a given customization to serve as a starting point.

When considering vocabulary items on the basis of their popularity, one needs to be aware of the fact that items introduced in later JATS versions might not be widespread yet in published articles, or that other aspects such as promoting accessibility suggest that items be added to the consensus customization despite infrequent use in currently available articles. This is why the authors will suggest a minimal customization that comprises significantly more vocabulary items than the “naive minimal customization” that was determined as adequate for tagging many of the articles actually analyzed.

Sources

The authors have received or downloaded JATS articles from the following sources:

The articles have been tagged according to diverse versions of JATS Blue or Green, sometimes with proprietary extensions.

When there is n/a in a table cell this means that the authors didn’t have access to full-text articles; therefore they were unable to count the articles or to classify them.

The subject classification was only applied handwavingly to individual journals or sources. The authors hoped that such a rough classification will offer insights with respect to varying tagging practices in the disciplines, but the analysis didn’t reveal large discrepancies, except that STEM (science, technology, engineering, mathematics and/or medicine) uses more vocabulary items than HUM (humanities and social sciences) or ECON (economics), but this is primarily due to the much larger sample size of STEM articles, and due to the wholesale categorization as STEM of all articles from PMC, which might not be vindicated.

The PubMed Central bulk packages have been downloaded from their FTP site [PMC].

The Science Open input consists of open-access articles that have been downloaded randomly from scienceopen.com.

In the PMC archives, the authors ignored journals with 40 or fewer articles, which only marginally decreased the number of articles processed.

As schemas, the authors considered the JATS 1.3d2 Blue, Pumpkin, and Green customizations, and also the Green customization that provides OASIS tables. This DTD should be the ultimate superset, unless a publisher uses proprietary extensions. The authors did also include the aforementioned Texture customization.

In addition, a synthetic schema “minimal” is included. It represents the proposed minimal Blue subset. It was only conceived after analyzing the collections and schemas according to the metrics and methods described in the next section, and after identifying an existing de-facto customization as a starting point.

A derivative of this minimal customization is the “naive minimal” customization that was prepared by omitting strategically important but infrequently used items from “minimal.”

Metrics and Methodology

“Supersetticity” and Aptness Metrics

A conceivable metric for the aptness of a given customization is to measure whether its vocabulary is sufficient to mark up 90% of all collected articles. Given that there are more than 1,000 de-facto customizations (when considering the vocabulary of each PMC journal as a de-facto customization on its own), the computational effort necessary to compare each of these 1,000 lists (one per journal) of several 100 items with approximately 630,000 lists (one for each article) of several 100 items seemed prohibitive.

Therefore the authors looked for a metric that was a cheaper to apply while still making the adequacy of a customizing measurable.

The difference is that each of the more than 1,000 customizations will be compared not to each article, but to all other more than 1,000 customizations. If an item needs to be added to arrive at the other customizing, it needs to be penalized more than if an item may be removed.

After four attempts of defining such a metric, the authors arrived at the following quantities and computation rules:

s	The “supersetticity” defines the degree to which one customization is a superset of another. The maximum value here is 2.0, which means a customization is a complete superset of another. The supersetticity of a customization j with respect to a customization i is calculated as follows: sij = (1+rji) / (1+rij) with rji = aji / dji, rij = aij, aij: additions to j towards obtaining i, aji: additions to i towards obtaining j, dij = dji = aji + aij = edit distance between i and j. Example 1 Green (j) contains 764 items, Blue (i) contains 755 items. aij is 0, aji is 9. dij and aji are both 9. rij is 0, rji is 9/9 = 1. sij = (1 + 1) / (1 + 0) = 2. A supersetticity sij of 2 means that j is a superset of i. Example 2 Swapping Green (now i) and Blue (now j) of the previous example. aij is 9, aji is 0. dij and aji are both 9. rij is 9, rji is 0. sij = (1 + 0) / (1 + 9) = 0.1. The supersetticity of Blue with respect to Green is 0.1, which means that Blue is not a superset of Green (otherwise it would be equal to 2). Note rij was originally intended as the relative number of items that needed to be added to j in order to move it closer to i. The asymmetry between rij and rji is due to the consideration that if something needs to be added to j in order to get to  i, it should drag the supersetticity of j wrt i quickly down to zero. On the other hand, an item that needs to be added to Blue on the way to Green doesn’t affect Green’s supersetticity wrt Blue since an increase by one in aji are roughly compensated by the same increase of dji. That is, if Blue lacks 10 items of Green and Green lacks none of Blue, rji = 10 / 10 = 1. If we add a missing item to Blue, rji = 9 / 9 = 1. Example 3 Let i be a de-facto customizing (list of elements and attributes that occur) as constituted by an issue of De Gruyter’s Journal New Crystal Structures. Let j be a de-facto customizing as constituted by an issue of De Gruyter’s Journal Nanophotonics. i contains 122 items (elements and attributes with distinct names that occur at least once), j contains 176 items. There are 5 items in i (New Crystal Structures) that are missing in j (Nanophotonics): The elements <funding-statement> and <pub-id> and the attributes @orientation, @publication-format, and @valign. There are 59 items in j (Nanophotonics) that are missing in i (New Crystal Structures), for example <alternatives>, <mml:math>, @fig-type, and @mathvariant. aij is 5, aji is 59. dij and dji are both 64. rij is 5, rji is 59 / 64 = 0.92. sij = (1 + 0.92) / (1 + 5) = 0.32. The supersetticity of j (Nanophotonics) wrt i ( New Crystal Structures) is 0.32. Swapping both customizations, let j be New Crystal Structures and i be Nanophotonics: aij is 59, aji is 5. dij and dji are both 64. rij is 59, rji is 5 / 64 = 0.078. sij = (1 + 0.078) / (1 + 59) = 0.018. The supersetticity of j (New Crystal Structures) wrt i (Nanophotonics) is 0.018. This exemplifies: Even with the moderate amount of 5 additions necessary, the supersetticity of Nanophotonics drops quickly from the optimal value of 2 (full superset) to 0.32. The need to add almost 50% of the initial item count to New Crystal Structures in order to be able to arrive at Nanophotonics lets the supersetticity of New Crystal Structures wrt Nanophotonics drop to 0.018.
q	defines the aptness of a customization as a starting point for another customization. Being a superset to another should favor a customization. But of two customizations with the same supersetticity wrt a third customization, the one that has the least edit distance should be favored even more. Therefore we define qij := 100 * sij / dij as the aptness of customization j to serve as the starting point for deriving customization  i. Example 1 (as above) qij = 100 * 2 / 9 = 22.2, that is, Green’s aptness to serve as a starting point for Blue is 22.2. Example 2 (as above) qij = 100 * 0.1 / 9 = 1.11, that is, Blue’s aptness to serve as a starting point for Green is 1.11. Example 3 (as above) Nanophotonics’ aptness to serve as a starting point for New Crystal Structures is 100 * 0.32 / 64 = 0.5, while New Crystal Structures’ aptness to serve as a starting point for Nanophotonics is 100 * 0.018 / 64 = 0.03. Figure 1 shows that qij will drop more quickly if items need to be added to j than if items need to be removed from j (= added to i in order to obtain j). It needs to be stressed that the q metric does not claim absolute truth (“the higher, the better starting point in all circumstances” is not necessarily true; it also depends on other factors such as: It is easier to add the whole of MathML than to add MathML partially). The q function was modeled after the requirement that both a small editing distance and “subsetting over supersetting” should be honored. In some figures and tables q5 or q5 can be seen. The 5 stems from the fact that the metric described here was the authors’ fifth attempt at finding appropriate supersetticity and aptness metrics.
p	defines the percentage of aptness of a customization j to serve as a starting point relative to the best starting point’s aptness, which is arbitrarily set to 100. Let qmax,i be the maximum aptness of all other customizations k to serve as the starting point for i. Then the relative aptness of customization j wrt i is pij = 100 * qij / qmax,i Example 1 With i Blue and j Green, qmax,i is 22.2 (that is, Green is the best starting point for Blue, see the green square in the second table row of the table sample-conf.details.xhtml), and pij = 100 * 22.2 / 22.2 = 100. Example 2 With i = Green and j = Blue, qmax,i is 1.11 (third row in said table). In this case, pij = 100 * 1.11 / 1.11 = 100, that is, Blue is the best starting point for deriving Green. The other p values in the same row are lower than 100. Example 3 With i = New Crystal Structures and j = Nanophotonics, we need to identify the table row that corresponds to New Crystal Structures and then spot the cell with the green background. It is in the column with the table head “DeGruyter”. That is, the de-facto customizing that corresponds to the collection of all ten De Gruyter journal issues that we examined is the best starting point for New Crystal Structures. Compared to its p score of 100, the customization Nanophotonics only receives a p score of 25.28 to serve as a starting point for New Crystal Structures. With i = Nanophotonics and j = New Crystal Structures, DeGruyter is again the best starting point of all. Compared to its p score of 100, the customization New Crystal Structures only achieves a p score of 0.66 to serve as a starting point for Nanophotonics.
Average p	This value is shown for each customization in the last column of the table at sample-conf.xhtml. The individual p values for a given customization to serve as a starting point for all the other customizations can be read column-wise in the other, detailed table. These p values are averaged. The customization with the highest score is deemed to be most suitable to derive the other customizations from.

Figure 1: Dependency of q_ij on the number of items to be added to  j or to be removed from j in order to arrive at i

The outcomes of these computation rules have been compared to what one would intuitively think is a good customization starting point, and the authors think they are appropriate in helping identify few candidate de-facto customizations that may serve as a basis for the synthetic minimal customization.

For evaluating the true aptness of a given or putative customization, other factors need to be considered, too, such as compactness (small number of items), support for recent additions to JATS, or subjective factors, such as the conviction that <array> may never substituted with a caption-less <table-wrap> or that <sans-serif> must never perish despite infrequent use and the availability of styled-content/@style for literal CSS.

There is no objective truth in this metric; it is just a means to reward customizations that may act as a starting point for other customizations without the need to add many items.

Minimal Subset Criteria and Choices

A candidate customization with a relatively low item count and a relatively high average aptitude p will be selected as a starting point for the minimal customization.

For synthesizing the proposed minimal subset, these choices have been made:

“Strategic” vocabulary items	Although their frequency is low, vocabulary of certain important areas has been retained: Recently added items (that are not considered aquafication symptoms – index terms and questions/answers may be regarded as such), as shown in Table II; Accessibility items (<alt-text>, <long-desc>, <textual-form>); Open access items (<open-access>); Data citation items (<data-title>, <version>); Machine readability items (@iso-8601-date);
Potentially neglected subject area: Computer science	Although they occur surprisingly unfrequently in the data set, <code> with several attributes such as @language and <preformat> have been retained.
Frequency	Otherwise, let the element and attribute frequencies (Figure 2 to Figure 10) guide the decision whether an given Blue vocabulary item is retained in the synthesized minimal customization. Apart from the frequency itself, a secondary criterion may be whether the item’s frequency is significant in more than one of the source collections considered. (Note that only collections for which full-text XML articles have been obtained could be analyzed by item frequency.)

Table II

Vocabulary items added to Blue in recent JATS versions

Newly introduced elements or attributes in…
JATS version 1.2		JATS version 1.3d1		JATS version 1.3d2
article-version article-version-alternatives contributed-resource-group event event-desc index-term index-term-range-end inline-media pub-date-not-available pub-history resource-group resource-id resource-name resource-wrap see see-also support-description support-group support-source	@article-version-type @degree-contribution @event-type @indent-level @index-type @resource-id-type @resource-type @style-detail @support-type @term-status @term-type @use-typ @vocab @vocab-identifier @vocab-term @vocab-term-identifier @xsi:noNamespaceSchemaLocation	answer answer-set block-alternatives explanation option question question-preamble question-wrap question-wrap-group	@audience @award-id-type @correct @hreflang @pointer-to-explained @pointer-to-question @question-response-type	award-desc award-name extended-by issue-subtitle issue-title-group processing-meta restricted-by	@base-tagset @custom-type @math-representation @mathml-version @table-model @tagset-family

Figure 2: Element frequencies, part 1

Data set (tab-separated values)

Figure 3: Element frequencies, part 2

Figure 4: Element frequencies, part 3

Figure 5: Element frequencies, part 4

Figure 6: Element frequencies, part 5

Figure 7: Element frequencies, part 6

Figure 8: Attribute frequencies, part 1

Data set (tab-separated values)

Figure 9: Attribute frequencies, part 2

Figure 10: Attribute frequencies, part 3

Results

The analysis results presented in this section already contain the proposed minimal subset and its derivative, the naive minimal subset.

Figure 11: Individual PMC journals, consider vocabulary items not in Blue

Figure 12: Individual PMC journals, ignore vocabulary items not in Blue

Figure 13: PMC as a single de-facto customization, consider vocabulary items not in Blue

Figure 14: PMC as a single de-facto customization, ignore vocabulary items not in Blue

The authors selected a compact yet high-scoring de-facto customizing of PMC Transfusion (items with less than 1% frequency ignored) as a starting point. Items that met the criteria presented in section “Minimal Subset Criteria and Choices” have been included. This led to a significantly larger and lower-scoring “minimal” customization. This customization still contains fewer items that the union of all items found in the collections. This is achieved by dropping more unpopular items than adopting strategic items. The item count (324) is still significantly lower than Blue’s (453), and the average aptness is not much worse than that of other collection-based or schema customizations. The reduction goal of 60% couldn’t be reached though. And it is questionable whether 90% of the analyzed articles are covered by this customization. This minimal customization is a superset of 38% of the de-facto customizations in the “Individual PMC journals, ignore vocabulary items not in Blue” scenario. Unless the use of the vocabulary is distributed very unevenly among a collection’s articles, this means that the goal of 90% article coverage is also missed by a high margin.

The situation looks different if the strategic additions are not considered. This is called the “naive minimal” customization and it comprises only 58% of Blue items. It is a superset to only 27% of the de-facto customizations, therefore the 90% goal is also missed significantly.

While the minimal subset has been prepared as a proper DTD customization, the naive minimal customization has been prepared manually by using the HTML vocabulary lists and commenting out strategically important yet empirically unpopular items again. This boosts the average aptness of minimal, in particular in the per-journal PMC case, to heights that only the all-used-vocabulary union customizations can reach, yet with about 25% less vocabulary.

Conclusion

Empirical JATS usage statistics have been analyzed. There is no clear-cut set of universally unpopular vocabulary items though. It is therefore arbitrary where to cut off the long tail. A thing that the authors consider not arbitrary at all is the question whether newly-added or otherwise strategically important items may be left out in a consensus customization; they may not. Only few items have been considered as maybe a sign of aquafication: index terms and questions/answers. These have also been left out in a previous effort by the Texture team.

More restrictions with respect to canonical usage of the tags within the vocabulary that the proposed minimal customizing provides may be done by fine-tuning the schema or by adding Schematron constraints, including existing ones from JATS4R.

The authors are not sure whether the proposed minimal customizing, together with these Schematron rules, may develop into a valid alternative to mainstream Blue. This probably needs to be field-tested by production staff at publishing houses.

A longer-form version of this paper, that is, Nina Reinhardt’s master’s thesis  Reinhardt 2021 will be available for download from HTWK Leipzig in August, 2021.