2.   VSM  (Visual Syntax Method)

(This text corresponds to the VSM-paper preprint’s Supplementary, plus informal notes and summaries).

i. Goal

With VSM, we aim to develop a method that enables people to consider any unit of information, and reformulate it into a semantically precise, computable form that clarifies its full meaning and context.1 Please read the Introduction page if you still need to understand why this is important.
Or at least read the Blue Boxes with our key concepts: ‘Piece of Information‘, ‘Context details‘, ‘Computable Information‘.

See also the introduction’s Venn diagram.

We aim for VSM to be applicable to the widest range of use-cases, so it works for diverse, heterogeneous, and detail-rich information. At the same time, we aim for VSM to be easily usable by many people with diverse skills and expertise, which requires that it keeps the inherent complexity of real-world information manageable. Like this, VSM enables humans to manually create meaningful computable information, with high ease of use, and high expressive power.

VSM includes selected strengths from existing methods to represent knowledge: natural language, controlled languages, ontologies and RDF triples.1 • Controlled language: Wiki:Controlled_natural_language, Wiki:Formal_language;
• Natural language: e.g. English, Dutch;
• RDF: Wiki:RDF, W3C, tutorial;
• Ontology: Wiki:Ontology_(information_science), Wiki:Open_Biomedical_Ontologies. But in order to meet the requirements above, VSM also intentionally differs in several ways, and defines a fresh set of foundational principles.

One can compare VSM with how a high-level programming language enables people to make an algorithm explicit by encoding it in a way that both humans and computers can manage.2 High-level programming languages have for-loops, if-then-else blocks, functions, etc.
In contrast, low-level, bare-metal assembly languages offer little to no such helpful, intuitive abstractions. Similarly, VSM is a language/tool to make an idea, thought, or piece of information explicit, by encoding it with intuitive, elemental components, in a way that both humans and computers can readily process and manage.

ii. Design

VSM is the combination of a new semantic model (or procedure, or method) and the design for a supporting user interface that together enable people to produce and read computable information.

As a Semantic Model, VSM is like a language. It has a simple, small set of rules for constructing information in the form of a so-called VSM-sentence, which is a semi-linear statement form3 This statement form is inspired by controlled languages. But unlike them, VSM avoids intricate rules on word order or fixed keywords. VSM uses a simpler, generalized way to clarify structure. that is flexible and intuitive for humans to work with. Similar to how a natural language consists of words and a grammar, VSM has two building blocks: VSM-terms and VSM-connectors. In addition, the three VSM Principles describe how terms and connectors can be meaningfully combined into virtually limitless combinations, in order to unambiguously represent essentially any type of complex information.

As a User Interface (UI), VSM defines the design for a software input-component that can be embedded in other software, for instance a web page. It enables people to enter (i.e. make explicit to the computer) a possibly complex piece of information, one per input-component. This component is called a VSM-box.

iii. Current status

We have built a prototype user-interface for curation4 Knowledge Curation = the manual (or semi-automatic) conversion of information from unstructured text etc. into a precise, computer-understandable (i.e. computer-actionable) form. based on VSM, for the purpose of demonstrating VSM‘s utility. It is embedded in the page with interactive examples, and we have been using it for several years at our lab for a specific biocuration project, to test and experience its use in real life. This exercise has given us many insights that are now included in this paper, in which we focus on VSM‘s full design and principles.

Recently, we have also built a production-ready (well-designed, well-documented, and easily reusable) VSM-box web-component, available in the ‘vsm’ organization5 At github.com/vsm : several modules that culminate in the vsm-box web-component (user-interface component). on GitHub. This core user-interface component and its supporting modules result from a large software-design and programming effort.

We hope that this work may be the start of a potentially large ecosystem of supporting and derived modules: of tools that facilitate specific or general use of VSM as an interface for communicating diverse and context-rich knowledge with computers. We therefore welcome the community to contribute to VSM on GitHub, or to create depending projects.6 E.g.:
• UniBioDicts : modules that provide bioscience terminology to the vsm-box component (doi);
• causalBuilder : web app that uses a vsm-box to support curation of molecular causal interactions (doi).

iv. Caveat

The design of VSM incorporates several ideas from existing technologies, but it also includes significant differences. As such, we have repeatedly experienced that experts with backgrounds in related areas are prone to make initial assumptions that cause misunderstanding of key concepts of VSM. In what follows, we hope to preempt such assumptions through extensive clarifications, yet we would like to advise readers familiar with related KR technologies to be especially mindful of the intentional differences.7 Still, a mapping between VSM and existing technologies is one of our active research interests.
As one connection, see a proposed conversion to RDF: vsm-to-rdf.

v. Approach of this text

VSM is a practical way of constructing meaning, inspired by human thinking; i.e. it combines a focus on human usability, with a particular view on semantics. For most of VSM‘s intended audience, the prime interest will likely be its usability aspect. Therefore, we have written this VSM paper mostly from that perspective. Each of the semantic principles that make VSM work, are introduced starting from the perspective of usability, inspired by how humans/curators are used to think. Further elaboration on VSM‘s semantics from a mathematical or formal-logic perspective, technical implementations of its user interface, etc., will be described in specialized publications.

While VSM was initially designed to solve a problem in the biosciences, we believe that VSM is much more generally applicable. Therefore, in what follows, we often use and start from domain-neutral examples.

For a quick overview: just read the ‘Short Story’ texts (yellow boxes); these introduce the ideas in informal language. For a deeper explanation and detailed justifications of design choices: read the ‘Full Story’ texts.

1. VSM-terms

VSM-terms are like the words of a language. But they are clearer (not ambiguous).

The Full Story  (formal story)

We will use a VSM-box as the setting in which we introduce both the semantic model and the user-interaction design of VSM.

A VSM-box, the user-interface of VSM, supports the entering of VSM-terms via an autocomplete function, by way of a drop-down list that lets users select the specific term that matches the concept they want to define (Fig. 1a). Each shown, entered VSM-term (blue rectangle with text, as a user-interface element) represents a single concept2 Wikipedia:Concept , or a unit of meaning, represented by a term (i.e. term-string). A term can consist of one or more words, for instance “homo sapiens”, “is located in”, or “binds to”. A term is in fact a human-friendly representation of a unique identifier (ID): a number or code that computers work with. Each term and its linked ID ideally come from an authoritative list, whose purpose it is to unambiguously define the meaning of each of its member terms, and to give it a unique ID together with known synonyms and other information. Such a list can be a domain ontology (e.g. Gene Ontology), a controlled vocabulary (CV) (e.g. PSI-MI), a list of entities (e.g. genes, NCBI Gene), or a curator-built list (e.g. via PubDictionaries).3 • http://www.geneontology.org
• The Gene Ontology Resource: 20 years and still GOing strong, Nucleic Acids Res, 2019.

• Molecular Interactions Controlled Vocabulary

• https://www.ncbi.nlm.nih.gov/gene

• http://pubdictionaries.org
• Open Agile Text Mining for Bioinformatics: The PubAnnotation Ecosystem, Bioinformatics, 2019.
• https://github.com/UniBioDicts/vsm-pubdictionaries
A VSM-box UI-component – which is meant to be embedded in other software – needs to be linked to one or more external term-lists or dictionaries like these, to give the user access to available terminology and IDs.

It can often happen that a same term is present in multiple lists and therefore linked to multiple IDs. The autocomplete function therefore helps the user to disambiguate between the available concepts or IDs. For example, for a human, a mouse, and a chicken gene with the same name, the autocomplete selection panel would suggest all three names together with the species of each. Once the intended term+ID couple is selected, only the term remains visible to the user (by default), while the ID is stored in the background, by the software or curation system. The UI may still show full information when mouse-hovering a term, or it could display a clarifying icon, or use other customizations.9 Here we discuss features for the VSM-box UI that are either implemented or that are ideas for customization options or extensions. We refer to the ‘vsm’ organization on GitHub (e.g. vsm-box) to follow up on which features are currently available.
Note: the VSM-box UI embedded in the Short Stories here and on the Examples page, is just the prototype.
The vsm demo page includes the refactored, ‘production ready’ vsm-box user-interface component.

In addition to supporting the above-mentioned terms with multiple meanings, or homonyms, a VSM UI also supports synonyms, i.e.: different terms may be used to represent a same concept. This enhances user-friendliness in several ways.4 Jointly creating digital abstracts: dealing with synonymy and polysemy, BMC Res Notes, 2012. – First, it enables VSM-sentences to correspond better to the text of the publication that is curated, because for example: gene and ontology terms often have many synonyms that are frequently used in the biological literature. When curating, synonyms (linked to the correct ID) make it easier to mentally map a long curated VSM-sentence onto terminology used in the publication, especially for other curators who may double-check the VSM-sentence. If needed, one can still add a UI-option to switch to displaying terms by their main, official name. – Second, some official terms may be excessively long (e.g. in GO5 Gene Ontology (geneontology.org) includes a dictionary (controlled vocabulary) of terms about gene functionality etc. – Example of a long term: “positive regulation of transcription from RNA polymerase II promoter“, many are longer. This length originates from GO being designed as a label-based classification system for biological concepts, resulting in a deep tree structure of ever-more-specific terminology. ), and curators may prefer to use an available abbreviation to keep produced statements readable. – Third, it enables to make VSM-sentences look more like natural language sentences. For example, consider a single relation concept that has two synonyms: the verb-form “is-located-in” and the preposition “in”, so that both represent the same ID. Then the VSM-sentence “A activates B in C” is more natural to read when one can use the preposition “in” instead of the verb form (Fig. 1b vs. 1c). This support for synonymous word-forms and also conjugations (as also shown in many following examples), is a key factor in enabling the creation of VSM-sentences that are easily human-readable.

In summary: VSM‘s support for synonyms and homonyms enables a curator to write easily readable information, with reduced cause for confusion; and once specified, each VSM-term in a statement represents one specific ID.

Figure 1. Entering VSM terms. (a) A curator searches and enters a sequence of terms, each linked to an identifier that represents a particular concept or meaning (two of five searches shown). Each selected term+ID combination results in the creation of a VSM-term user-interface element, representing that term+ID couple. (b) The full sequence of VSM-terms. This example is just as easily readable as a natural language sentence, among others because the user was able to use the preposition “in” for the relation “is located in”, both of which would here represent the same identifier. (c) This sequence represents the same five IDs but illustrates a less natural choice of terms.

Note: in text that follows, we enclose example VSM-sentences and VSM-terms in double quotes, and replace spaces in multi-word terms by dashes, like in this example VSM-sentence: “X is-located-in Y”. When we give examples as natural language phrases, we use ‘single quotes’.

2. VSM-connectors

VSM-connectors correspond to the grammar or syntax of a natural language. But they are simpler, yet just about as powerful. There are three main connector types, and you can achieve plenty with just the first one:

2.1. Indicating structure with the trident

The Full Story

A sequence of entered VSM-terms may initially look like a natural language statement, in which words have been disambiguated. But a VSM-sentence is typically not copied verbatim from a paper (read also this note: 12 In fact, we are able to make it resemble natural language, because we are able/allowed to make VSM-terms appear as conjugated verbs, nouns, prepositions, etc.
That ensures that we can keep complex information (i.e. longer sentences) easily readable for us, humans, too.
Computers do not understand conjugations etc. reliably well, so we need to clarify the structure of a VSM-sentence in another way. ). We have often experienced that this needs to be emphasized, so we declare it as:

📖 VSM Principle 1. Bottom-up construction.
A VSM-sentence is a curator’s interpretation, or concise rephrasing, of some information’s essence6 Examples of concise paraphrasings (applied to free text): en.wikipedia.org/wiki/Concision . Curators need to reflect this understanding of knowledge that they acquired from somewhere (like literature) by entering terms in such a way that a syntax (i.e. structure of how VSM-terms link together to create composed meaning) can be specified with VSM‘s syntax connectors.13 VSM-connectors show the conceptual structure that exists in the information explicitly.
As a result, VSM-terms’ text can show a readable formulation, while VSM-terms’ IDs and the VSM-connectors show the conceptualization structure that exists in the information, intuitively and clearly. These VSM-connectors are added through a user interface and are available in three main types: the trident (incl. three bident subtypes), the list-connector (of any length), and the co-reference.

In what follows, we explain how VSM-connectors are used to syntactically specify the meaning of VSM-sentences. We start with simple statements, and next progress toward statements with complex internal structure that convey rich contextual information.

We start with the trident. A curator can add structure to a statement by identifying triples, or subunits consisting of three VSM-terms. At first glance these triples look similar to RDF-triples7 • RDF in general: Wiki:RDF, W3C, tutorial.
• E.g. triples represented in RDF Turtle. , yet we elaborate on that further below. The UI enables a user to specify triples by linking terms with a trident connector. A trident is added via three consecutive clicks, one above each term according to the role it has in the triple, in the order of: subject, relation, and direct object14 An indirect object must be captured via an extra relation; e.g. ‘Ann gives Eve X’ as ‘Ann gives X to Eve’. . The corresponding, attaching parts of the trident, which we call its legs, are: a simple line, a line decorated with an up-pointing filled triangle, and a down-pointing arrow, respectively. Figure 2a shows a simple example one-triple statement that is specified with a trident, and that expresses that a “John” “eats” a “chicken”.

Figure 2. Specifying triples with tridents. (a-b) The user adds trident VSM-connectors that specify triples (subject-relation-object triplets) of individual terms (as indicated by leg positions and leg foot-widths); as subunits defining the intended syntax. The resulting composite semantics is explained in the main text. (b-d) Meaning changes when the rightmost connector’s subject-leg is attached to a different term.

Next we will show by example how to use multiple tridents, after which we can explain VSM Principles 2 and 3. Figure 2b shows the addition of more terms or context via a second trident, again attaching to individual VSM-terms (n.b.: not to any previous subunit as a whole). This specifies a second triple subunit that expresses: this “eating” happens “with” (or “using”), a “fork”. We represented this second triple’s relation by a preposition-type synonym, for natural-language-like readability. In contrast, Fig. 2c shows an alternative version in which a chicken is using a fork instead. This illustrates that attaching a connector leg to a different term fundamentally changes the meaning. Similarly, Fig. 2d states that John uses a fork, although it does not say he eats chicken with it (he could be eating chicken with his hand, while using the fork for another purpose). Only by connecting the second trident to the “eating” term, one captures the correct meaning: ‘the eating (by John, of chicken), is by use of a fork’.

This example introduces an elementary way of thinking, as a basis that can be scaled up to knowledge of any complexity. This is further elaborated in:

📖 VSM Principle 2. Referable entities.
A connector leg always attaches to one particular term, and never to an entire other triple/subunit. This is even true when a leg attaches to a term that is used as a relation in another triple; and this makes VSM different from the typical approach in RDF.15 In RDF one typically makes a reification construct in order to point to a relation; i.e. one creates an additional entity (a blank node) that represents that relation (or in fact, its entire triple), that one can then point to. Because RDF treats relations as fundamentally different things than entities.
(For an overview of four approaches or proposals for referring to relations in RDF, see ref: “Hernández D, et al. Reifying RDF: What works well with Wikidata?, ISWC, 2015″).
In fact Principle 2, pertaining to relations, is similar to the n-ary relations approach in RDF, if that would be applied consistently to all relations (except for a few primitive relations like has-subject/object/etc., which is what the VSM-connectors actually represent). This difference is intentional, and lies at the basis of making the syntax-definition process flexible, scalable, and intuitive, following the thinking of the curator. It is rooted in VSM‘s specific approach to treat all VSM-terms equally: both so-called entities and relations are first-class citizens. In VSM, any term is only viewed as a relation instead of an entity, within a particular subunit, under a particular trident that connects to it with its relation-leg. Because of that, any term that happens to function as a relation under one trident can function as a subject or object (i.e. an entity) when viewed under another connecting trident.

For instance, in Fig. 1b the verb term “activates” can (and should) equally be thought of as a noun: ‘activation’. In Fig. 2b the verb term “eats” is in meaning identical to its noun-form ‘the eating’. If both lexical alternatives are available as terms in a dictionary for VSM, they should carry the same ID.

In summary: every VSM-term should primarily be thought of as an entity – i.e. a referable thing.

Note: the VSM Principles are not verbatim definitions. They may be rephrased using various wordings, for instance depending on the intended readers’ background or on a chosen format of explaining VSM. This can offer additional perspectives that put these principles into proper context, and that together better convey the core idea.
See also this note: 16 Principle 2 is crucial, so we repeat it once more in other words:
Only when some trident attaches to a VSM-term with its relation-leg, only then is the term made to be seen as a relation. And then only so, under that one trident, i.e. locally in that one triple subunit.

What RDF (or formal logic) typically view as a relation (or predicate), VSM views only locally as a relation (under a particular connector, in one subunit) and views it elsewhere as an entity / as ‘reified’ / as a noun.
(The RDF phrase ‘re-ified’, based on Latin, means ‘thing-ified’ or ‘made into an object’).

This was an essential, necessary step for designing VSM‘s expressivity and simplicity in what will follow:
think of all VSM-terms as nouns / ‘thingified’ concepts / something that one can always refer to or mentally point to again. Like this, one can immediately refer to any VSM-term, always in the same way. (<– really, read it).

2.2. Scaling up to more complex information

The Full Story

In order to scale this method up to statements with many more terms, one must keep in mind a third key principle. It provides a way of thinking that will be instrumental for constructing longer and valid VSM-sentences, and for unambiguously interpreting their meaning.20 In particular, Principle 3 will support the creation of clear VSM-sentences in which an entity is described in detail in multiple situations (see section 2.5).

📖 VSM Principle 3. Embedded context.
When a trident (or any VSM-connector except coreference, see later) is added, each of its connected terms receives extra ‘context meaning‘ from the other connected terms. In other words: a VSM-term does not just represent a particular entity (or thing, or instance); instead it represents a particular entity, in a particular situation (or state, or local context) that is specified by all the VSM-terms connected to it, directly but also indirectly. For example, in Fig. 2a the first trident makes the VSM-term “eats” (which without connections would just represent a particular but not further specified ‘eating’) now represent ‘the eating, of a chicken, by John’; i.e. some contextual details are now embedded into this individual VSM-term’s meaning. This applies to all three VSM-terms, so “chicken” becomes ‘the chicken eaten by John’, and also “John” becomes a more specific concept.

Next, when one connects a second trident to any of these three terms, one in fact refers to that term’s full, context-enriched meaning, as created by its connection under the first trident. So with the second trident in Fig. 2b, one expresses: “the-eating-of-chicken-by-John happens-with fork”. By induction, all five terms now carry an enriched meaning, acquired or composed from the other four. Next, each of these terms is individually referable again, in the same way. Continuing like this, one can keep adding further connections and thus further context details to every single VSM-term, recursively.

For instance, to express ‘John eats a chicken, occasionally using a fork’, one would add two terms and then connect a new trident to “using”, to create a new triple that states: “using (of fork, by eating of chicken by John) has-frequency occasionally”. (See Bidents, section 2.3, for how to intuitively add adverbs). All added detail to the meaning, delivered by connections, form what we call a VSM-term’s local context. So each VSM-term is coupled with its own, local, or embedded context in the VSM sense.

Additional perspective is given in notes21 The notion of local context in VSM differs from context in e.g. Conceptual Graphs (CG) or RDF Named Graphs, which take a formal logic based starting-point for representing meaning.
• Their context can be seen as a separate object that is shared among a group of concepts, in order to manage the scope of logical quantifiers. (CG: Sowa JF, “Conceptual Structures: Information Processing in Mind and Machine”, 1984; RDF: Caroll JJ, “Named Graphs”, 2005).
• In contrast, VSM starts from an association-based viewpoint, which is instrumental for prioritizing different objectives. (A formal definition is a future prospect). and22 While Principle 2 defines the nature or meaning of individual VSM-terms, Principle 3 defines how the meaning of each of them changes by attaching connectors to them.
Note that at this point, Principle 3 may be intuitive or even seem trivial for simple cases like Fig. 2. But it provides a crucial way of thinking for handling cases that need the coreference connector (sections 2.5 and 2.6). , on the notion of context in VSM.

It is essential to realize that a trident in itself carries no meaning. This should be evident from the fact that a connector never connects to another connector, but only connects to individual terms. In fact, VSM-connectors only act as operators on individual VSM-terms’ meanings. They specify or narrow down the meaning of a term. They make some of its specifics (which make the term a particular instance of something) explicit. For example, in Fig. 2a the particular “eating” is not specified to be done in any specific way; that context is not given, so as far as we know it could be done with hands, a stick, fork and knife etc. But in Fig. 2b the meaning is narrowed down: by specifying the triple “eats with fork”, it is made explicit what is used for John’s action of eating. It narrows down “eats”‘s meaning by anchoring down at least one more aspect of how it happens. Meanwhile, other unknowns remain: where John eats, if he eats alone or together with someone, how quickly he eats, etc.23 For more on general concepts vs. specific concepts in a partly-explicit, partly-implicit context:
see VSM Principle 4 on a further page.

---

Figure 3 shows longer VSM-sentences created in this way, representing information curated from scientific publications. For example, in Fig. 3a the second trident from the left declares a triple expressing “the-being-twisted-of-leaf-lamina  pertains-to  Oryza-that-underexpresses-YAB3-by-RNA-interference”.24 For non-biologists: this means that plant-leaves are deformed in a particular way in a rice plant, which had the activity of one of its genes turned down, using some technique for achieving that. Here, the term “in” will have been linked by the curator to the concept “pertaining-to” rather than “is-located-in”. Each term represents an ID from a controlled vocabulary, e.g. “leaf-lamina” is provided by PO (Plant Ontology), and “twisted” by PATO (Phenotype And Trait Ontology).11 • PO: The plant ontology as a tool for comparative plant anatomy and genomic analyses”, Plant Cell Physiol, 2013. Link.
• PATO: www.ebi.ac.uk/ols/ontologies/pato.

We want to accentuate that any additional context, like the biological details captured here, helps to increase the quality and value of the information. Added detail could for instance resolve two seemingly contradictory ‘bare triple’ statements through reconciliatory contextual details.25 For example both ‘bare triple’ statements “cat is alive” and “cat is dead” can be true. But they would be true in their own specific context.
We could make this context explicit by expanding the statements to e.g. “cat is alive in 2020″ and “cat is dead in 2100″. Added detail is also vital when the collected information needs to be filtered down; for instance based on various quality measures determined by various use cases’ requirements.

Figure 3. VSM-sentences based on scientific papers. The VSM-sentences are structured, reformulated versions of information from life science publications: (a-b) from12 Dai 2007: ‘A WUSCHEL-LIKE HOMEOBOX Gene Represses a YABBY Gene Expression Required for Rice Leaf Development’. , (c) from13 Jiang 1999: ‘Multistep regulation of DNA replication by Cdk phosphorylation of HsCdc6′. . Some of their terms come from ontologies and controlled vocabularies (CVs), others are yet to be added to official CVs, especially relation terms.

We point out that all connections to a VSM-term together define its contextualized meaning; i.e. all connections are applied simultaneously: they add meaning in a way that order does not matter.26 For cases where the semantics requires a deviation from this, see section 2.6. Therefore, the vertical stacking order in which connectors are added or displayed is irrelevant for the connection structure (topology) and its resulting semantics. Still, some stacking order may look more intuitive than others, for example based on which subunit or term one thinks of as central in the statement (see Head later). Various visual sort algorithms may be developed for this UI aspect. We already implemented one that automatically rearranges the stacking of connectors as shown in Fig. 3, no matter in what order they were added.

Note that although a VSM-sentence can be mapped to RDF, either directly when two non-relation legs connect to a term, or via RDF-reification when a non-relation leg attaches to a term also used as a relation, such RDF graph renditions quickly become unintelligible to a curator (see later in Fig. 10 on the Discussion page for a full example).

---

A consequence of Principle 3’s definition of VSM‘s semantics is that tridents (together with bidents and list-connectors, see later) must never create loops. A loop exists when one can follow a path of connections from one term that leads back to that term. (A loop is called a cycle in graph theory).

To illustrate: consider for instance a VSM-sentence that expresses “John eating apple calls Eve buying apple”, and let us leave aside for a moment whether it is the same apple or not. The trident that connects “Eve buying apple” must always connect to a second “apple” VSM-term (Fig. 4b). It must not ‘reuse’ the first “apple” that would already be connected in “John eating apple” (Fig. 4a). Because if it did, it would make it impossible to interpret the meaning (following VSM Principle 3): the VSM-sentence would be read as: ‘John, eating an apple, → that is being bought by Eve, who is called by John, eating an apple, → …’. It would have to be read like that, because every trident-connection a VSM-term has, must be followed (i.e. applied) when interpreting its meaning. Furthermore, if ‘apple’ would have additional VSM-terms attached (e.g. specifying its color), then this loop would create several interpretation possibilities, i.e.: from which of the two ‘directions of encounter’ in the loop would the color apply: for the situation where John eats it, or the reverse, the situation where Eve buys it (i.e. ‘John calls Eve, buying apple, …’)? This means that a connection loop creates context ambiguity. So in conclusion: with tridents (and bidents and list-connectors; later), one must only create hierarchical structures.

It may already be clear intuitively that when two distinct ‘apple’ entities (physical objects) are at play, then two “apple” VSM-terms (separate, referable concepts) are needed.27 I.e. two instances of “apple” are needed. Each one would have its own instance-ID, plus a class-ID for the general concept it is an instance of.
This, however, is a more technical aspect of VSM that is covered by the vsm-box implementation (and also a bit on the VSM-Graphs page, later). However, also if a sentence would mention one and the same apple twice, then the above shows that two VSM-terms must be used: one for each context. For that, a coreference connector will be needed (see later, section 2.5).

Figure 4. Tridents must never create connection loops. (a) An incorrectly constructed VSM-sentence. Here, the interpretation of meaning, performed by following connectors that further specify terms’ meanings, would result in an infinite loop and set a basis for context ambiguity, as described in the main text. (b) A correct version that can be properly interpreted. (See Fig. 7a for the case with one apple).

2.3. A variation on the trident: bidents

The Full Story

Some information is unnatural to express with triples. For instance, some verbs have no object, as in “plant flowers”. Although this can be captured with a triple (the typical approach in RDF) like “plant belongs-to class-of-flowering-things”, the task of stretching and reformulating such information to fit a triple structure is a burden that is best not imposed on the curator. VSM is focused on usability for the curator, and therefore the VSM UI supports tridents where one leg can be left away: bidents. The remaining two legs still appoint the same semantics, i.e. they assign the same role to the two terms in their subunit, as they would for a full trident.

In the example above, no object needs to be defined and the object leg is omitted (Fig. 5a). A user can create this first type of bident just like a trident, but with the third mouse click outside the VSM-box component, or by pressing Escape instead of the third mouse click.

Bidents also make VSM‘s knowledge representation structurally consistent. For instance, compare the statements “X stimulates: A activates B”, and “X stimulates activation-of B”, where the latter uses a second type of bident: one that omits the subject leg (Fig. 5b-c). Apart from using the appropriate synonym for the “activation” relation (with same ID), the principal difference is that the activator is unspecified in the latter case. Appropriately, the structure of both statements is similar, thanks to the bident. This consistency would not be present when using a second triple “activation having-object B” with an artificial relation inserted.29 Note for experts:
One could also enter “activation  of  B”, by using an explicit VSM-term “of“, which would have the meta-meaning ‘has object’.

Note #2: please do not misunderstand and think that ‘VSM knows’ if two sentences (two graph structures) would represent a same, rephrased meaning. VSM does not ‘know’. VSM is just a representation form, to be used by people and algorithms. – Just like ‘English does not know’ what you mean. Your brain does, and it uses English to represent information at some level. This bident is also useful for capturing passive constructs where a subject is irrelevant, as could be the case in “X accelerates destruction-of B” or “to-write(=writing-of) text is rewarding”. One starts this bident with two clicks above the relation.

A third type of bident omits the relation leg. Again, this helps usability for the curator, in this case by allowing statements to better reflect a basic structure also used in natural language: the attribute. Attributes include adjectives, adverbs, and numbers. For example, a statement about a ‘white mouse’ would require a “mouse having-color white” when using only the triple structure. However, if the term “white” is already classified as a “color” in its vocabulary or ontology, then making curators explicitly insert the relation “having-color” would be a waste of effort, since it can be inferred automatically. Then, a bident structure can be used, as in Fig. 5d-e. The relation is omitted via two clicks above the object term. This bident in fact indicates an implicit relation, most generally seen as: ‘is specified by’.

However, in order to keep the semantics with relation-omitting bidents clear, consider this:
• In general, a category of terms that can be used as attributes, may need to be associated with its own default, more specific implicit-relation; e.g.: numbers (“… [has-count] 5″), measures (“… [has-size] big”), or posttranslational modifications (“… [has-state] phosphorylated”); see also Fig. 5f-g.
• In other cases, this could be associated with the attribute’s linked concept as well. For example: “<tree‑x> leaf” would mean “leaf [from-plant-of-species] <tree‑x>”, e.g. “oak leaf”. Or “<gene‑x> expression” would mean “expression [of] <gene‑x>”, e.g. “Cas9 expression”.
• However, for attribute constructions as in natural language where the implicit relation is not yet automatically inferable to something more specific, one must use either an explicit triple or a single term. For example, ‘actin filament’ would then need to be represented as “filament composed-of actin” or as “actin-filament”.

The above demonstrates that knowledge representation in VSM more closely follows the thinking of the curator, or the conceptualization structure in the human mind. This is one key aspect that makes VSM more intuitive.

Figure 5. Bidents. The user can add a trident and omit any one of its three legs, i.e. create a bident, for ease of use and structural consistency as explained in the main text. (a) The object-omitting bident (or oobi) accommodates a verb without an object. (b-c) The subject-omitting bident (or sobi) specifies a unit with an unknown or unstated subject, here “activation(-of) B”. The bident enables representing this information in a form structurally similar to the case where a subject “A” would be known and can be attached to a full trident’s subject-leg. The availability of bidents makes this kind of structural consistency straightforward to achieve (which is not necessarily the case in RDF, see later in Fig. 10). (d-e) The relation-omitting bident (or robi) implies an implicit relation ‘is specified by’. It frees the curator from the obligation to explicitly define a more specific relation that could be computationally inferred. (f-g) Various attributes can be connected via a robi bident. Here, the more specific implicit-relations would be inferable to be: “has-precision”, “has-count”, and “has-concentration”.30 This figure also shows that numbers can be represented with VSM-terms. This makes sense because one “5″ can be conceptually different from another “5″; because just like other VSM-terms, a numeric concept can be embedded in a particular context. For example, one could be an “approximately  5″, another a “5  ±  2″, and another could be contextually specified or modified to mean “at-least  5″.
For this, the autocomplete system needs to support the entry of numeric concepts. But there are infinitely many of them possible, so they cannot be stored in any prepared vocabulary. So the autocomplete will need to automatically generate an ID + term-label for any of them as needed, based on some pattern and on what the user is typing.
(This was not yet implemented in the prototype used on this web page, but is now supported via the ‘vsm‘ GitHub-project’s modules: vsm-dictionary (see its spec), which supports vsm-box).
Other numeric concepts like dates or timestamps could similarly be provided to the user. For this, additional dictionary software-modules that generate such concepts should be easy to implement. Bidents are drawn with a tiny, light-colored indicator for the omitted leg type, to remind the user of their overall semantic equivalence with tridents, regarding their remaining two legs.

2.4. Group types that combine items: the list‑connector

Consider a statement like ‘A, B, and C react to D’. In order to build this with triples only, as in RDF, one would first make an “A and B”, then connect their “and” to “and C”, and then connect this second “and” to “react-to D”. However, it may be unknown in which order reactants combine. Then this construct with triples is artificial, as it declares that A and B combine into a distinct group, which then combines with C (i.e. the first “and” represents the existence of a ‘combined A+B’). In addition, the longer the list that needs to be made, the more burdensome all the artificial triples would become to the curator.

With VSM‘s list-connector, one can express that any number of list-item terms all come together as one group, in a way specified by the meaning of a list-relation term. For example, Fig. 6a states that reactants all together react to D, as a group with unspecified internal order. The fact that group order is unspecified would be embedded in the meaning (and ID) of the chosen list-relation “and” (here: “and-(as-unordered-set)”). The curator is free to choose other list-relations as well: for instance a term “and-(as-ordered-list)”; or “sum-of”; or “either-or”, which is used in Fig. 6b to state that ‘either A, B, or C bind to D’.  – This term “either-or” is also an example of a single concept that is formulated with two separated parts in natural language, and that must be captured with a single term in VSM.

The position of a list-relation among its list-items does not matter for the resulting meaning, although some placing can make a sentence easier to read; like placing “and” second-to-last in Fig. 6a. Just like with tridents, only the resulting connection- or graph-structure matters. The only difference here is: the horizontal order in which list-items are placed matters when the list-relation attaches meaning to this order; e.g. it would matter for “and-(as-ordered-list)”.

As a semantic operator, a list-connector makes its list-relation VSM-term represent the specific concept of a group of items combined in the way defined by this list-relation. Similar to how tridents operate, the list-connectors in Fig. 6a-b contextualize their list-relation term into a specific “the-A-B-C-group”, and (for lack of a less artificial way of phrasing) “the-eitherOr-union-of-A-B-C”, respectively. Then once again, another connector can link to this term as it represents the specific, whole group.

Figure 6. List-connectors. (a) A list-connector assigns “and” as a list-relation that groups a number of list-elements. The connected list-relation “and” represents the combination of the list-elements as a whole, and as such can represent the left side of a reaction. (b) Other terms can be used as a list-relation. Here, a list-connector contextualizes the VSM-term “either-or” to make it represent the phrase ‘either A, B, or C’.

The UI in Fig. 6 distinguishes list-connectors by drawing them with a double backbone, a leg ending in a filled square that connects to the list-relation, and undecorated legs that connect to list-items. The UI needs to support switching between trident and list-connector creation modes, for instance by holding the Shift button down while first clicking above the list-relation; and then clicking above all list-items. If order is important for the list-relation, then it is not the order of clicking above terms, but the order of terms as they appear in the VSM-sentence that counts.

2.5. Referring and further specifying: the co‑reference

The coreference connector provides essential functionality that makes VSM work, both in terms of usability and in terms of semantics. It performs two related functions.

Its first, basic function is to let users build semantically correct, easily readable sentences with a term that refers to another term, for instance an “it” as in Fig. 7. We call this “it” the child term. The child term is a placeholder that receives meaning from its parent term: the term it refers to. A child term can have any label (“that”, “them”, etc.), and represents the same thing (i.e. entity) as its parent. For example in Fig. 7a, “it” refers to the same “apple” as its parent term (in contrast with Fig. 4b earlier, where two distinct “apple“s were mentioned).  – Note again that the no-loops rule (that followed from Principle 3) does not apply to coreferences, as will become clearer later.  – In Fig. 7, the UI draws the coreference in a dashed line, with an undecorated leg connecting to the child, and a down-pointing empty triangle as arrow connecting to the parent. The child term, drawn with a dashed blue rectangle, is a distinct VSM-term type that is entered via autocomplete, when users type e.g. ‘it’. Users may add a coreference by first Ctrl+clicking above the child term, and next above the parent term.

The second function of the coreference is that it enables us to think about two distinct concepts. Although the parent and child term refer to one and the same thing (=instance of something general) in the world, they refer to that thing as it occurs in two different situations, perspectives, or contexts.

Consider for example “Bob activates device in evening causes it beeps in morning” (Fig. 7c). The meaning embedded into the parent concept “device” is that it is being activated by someone, but should not be that it is beeping yet. Therefore (and because of no-loops), we should not connect “device” directly to “beeps” with a bident. Instead, we create a coreference-linked child concept “it”, which inherits the context of ‘having been activated in the evening’ from its parent, and in addition receives the extra context of “beeping in morning” from its own connection environment. These two concepts can then be separately enriched with more context-details or be referred to later on (from other VSM-sentences), e.g. “Jane charges it (#1)” and “Bob silences it (#2)”.  – A biological example could involve a biomolecule being chemically tagged in one context, e.g. in some cell cycle phase or location, which enables it to form a bond with another molecule in a further context.  – See also Fig. 7b, or read the extra perspective in this note:31 In other words: the two VSM-terms refer to a same entity, whereby each one describes the entity as how it is in a distinct state/situation/context,
e.g.: before vs. after an event; or a conditional state that causes an effect that eventually gives it a new state; or its known real state vs. a hypothetical state (which may be future but also present) as in “Eve pretends she likes Cat”.

The last example shows that the two VSM-terms do not need to represent the entity at two distinct timepoints; rather, they represent the entity in some explicitly stated context of what is known or declared about it. This can cover other aspects than time. Thus, a VSM-term does not make an implicit semantic commitment to time. .

All this is made possible by Principle 3 that defines what a VSM-term really represents. In this light, we can also rephrase its gist in the following ways:
• a VSM-term represents a particular entity, in a particular situation (state, context, etc.) that is detailed (made explicit) by its connections to other VSM-terms;
• a VSM-term represents a particular entity, with a particular set of available details that are given or known about it, at a particular point in a discourse;
• a VSM-term represents a particular state of knowledge about a particular entity, determined by that term’s own connections (primarily) and by its parent term (if it has one).

Figure 7. The co-reference connector. (a-c) The coreference combines two functions. First, it allows to build VSM-sentences in which a same entity is referred to twice (here: same apple/protein/device) in a way that is natural to read (via a placeholder with label like “it”), and that creates no loop (of tridents, bidents, lists). Second, it enables one to work with two distinct concepts, each at a different stage of specification: in (c) the first “device” is not yet beeping, while the second occurrence of the device, represented by the child-term “it”, not only was ‘activated in evening’ but also is ‘beeping in morning’. A curator could still add further specifics about the device, in both of its local contexts. Similarly in (a), the parent-term “apple” is not yet described as being eaten by Eve; and in (b), the first occurrence of “protein B” is not specified to be degrading.

Now we can clarify why the no-loops rule, which followed from Principle 3, only applies to tridents, bidents, and lists-connectors, but not to coreferences. Unlike the other connectors, a coreference does not enrich the meaning of the parent term, and so it should not be followed when interpreting the parent’s meaning. As a semantic operator, the coreference works only on the child concept; and also here it does not enrich an existing concept. It creates a new copy of the parent’s context-embedded meaning and places that at the position of the child, where it can be further contextualized.

The two distinct concepts (parent and child) are not necessarily about past versus present, but about how much context is present in each part of the discourse. Hereby it is helpful to think of the coreference as a unidirectional weathervane or ‘wind arrow’ that points to where context is ‘flowing from’, as in ‘inherited from’. This notion is crucial to understand the following issue: the case where (trident/bident/list-)connector attachments are non-interchangeable, as explained in the next section.

2.6. Non-interchangeable connections   (advanced, but nice topic)

When connecting multiple tridents or bidents to a VSM-term, we use one assumption: that the order of connecting them does not matter. This is usually true. In natural language one can test this by observing that the meaning of some statement does not change when switching attribute positions; e.g. ‘big white mouse’ has the same meaning as ‘white big mouse’. Likewise in VSM, each new connection simply adds context in an interchangeable, additive way: in Fig. 8a, “mouse” is specified to be both “big” and “white” at the same time.

However, in some cases order does matter, as e.g. in the statement “half-of black dogs”. By connecting “half-of” and “black” to “dogs”, one can construct three different meanings (see the sets in Fig. 8b1-3):

1. ‘some half of the group of dogs, that consists of only black ones’:  here “dogs” is specified to “half-of” and “black” at the same time (while still being half-of in total);
2. ‘some half of: black ones of the dogs’;
3. ‘black ones of: some half of the dogs’.

In case 1, the attributes are applied (i.e. add context meaning) simultaneously, while in cases 2 and 3 they are applied in one of the two possible orders.

What happens when order matters, conceptually, is that we first compose an idea, and then we isolate it and refer to it, without letting the further added context mix in with the original idea. This is in fact exactly how also the coreference connector works: in e.g. case 2 in Fig. 8b2 one can first construct a parent concept “black dogs”, and then create a placeholder child concept, e.g. labeled “them”32 “them” is a somewhat arbitrarily chosen label for the referring (child) term.
It makes the text quite nice to read in Fig. 8b2, though less so in S8b3. Any other label can be used too, like “it” or “these”. , that points with a coreference connector to the parent. This isolates the parent’s meaning and transfers it onto the child, unidirectionally. Finally, one can use this child term “them” for building the rest of the VSM-sentence, e.g. by connecting it to “half-of” and “escape”. The coreference connector points from the further-enriched child to its isolated, meaning-providing parent.

In case 3, Fig. 8b3, the same reasoning is applied: here one first constructs an isolated “half-of dogs”, then one limits that conceptual unit of dogs to those being black, and specifies that they escape. Note that one can still further detail the isolated part, as in e.g.: ‘black ones of: some female half-of dogs’.

Figure 8. Interchangeable vs. non-interchangeable attributes. In (a), attribute order does not matter. But in special cases, the order in which multiple connections are added to a term can affect meaning. Here, the attributes “half-of” and “black” can be applied either (b1) simultaneously, or (b2-3) in a particular order. In the latter cases, one applies the first connection to the term “dogs”, then isolates that concept as parent for a coreferencing child placeholder, that is e.g. labeled ‘them’, and one then further connects this child.

---

Notes (which may go a bit too deep for a first reading) :

• The implicit relation between “dogs” and “black” would be automatically inferable to be “limited-to-all-those-having-color” (and not just “having-color”), because black would be classified as a color, and dogs represents not just one dog, but a group of dogs.33 A note on how to create the meaning of some ‘group of dogs’, starting from a single “dog” concept.
Consider a bident-based construct like “11 dog”, or stated explicitly: “dog <has-amount> 11″. The association with the relation “has-amount” is what converts the meaning of “dog” into a group. That is how that relation would work, i.e. how it would have to be semantically interpreted. Alternatively, this relation could be made explicit, but for this example’s brevity and focus we used an implicit relation. Similarly, “half-of” pertains to the group of dogs, so it does not create the meaning of half of one dog.

• Each “half-of” creates an unspecified subset that is some half of the group. So in case 1, Fig. 8b1, where “black” also specifies that this half subset consists of only black dogs, and where more than half of the dogs are black, the final subset remains partly unspecified. Therefore, we had to choose an arbitrary subset of black dogs in Fig. 8b1. This is consistent with the other two cases, where the ‘half’ (dark blue line) is also not further specified and thus had to be chosen arbitrarily in Fig. 8b2-3.

• In a natural language sentence one can imply multiple pieces of information simultaneously. For example, a phrase like ‘the black half of dogs escape’ would not only express the information of case 1, but also tell that (an other) “half-of dogs  not  has-color  black”,34 Here “has-color” would be attached (as subject) with a relation-less bident to “not” (as object). 35 [Advanced note]: One might think of representing this as “exactly half-of dogs has-color black”, instead, inspired by a similar expression in natural language. However, this natural language is ambiguous.
- The primary meaning of “exactly” is: exact, as in not more nor less than the stated amount; so: “some precise half”; (opposite of “around”, which defines uncertainty about the exactness of the amount).
- A secondary meaning would imply, or intend to incorporate a meaning like: for the other part of the group, the opposite of what is said is true. This is what “exactly”‘s meaning/ID would have to be linked to in that sentence. (And from that, an algorithm could infer that “half‑of  dogs  not  has‑color  black”, which is the additional information we proposed in the main text already).

It is up to a VSM-box’s term-provider whether or not to include such intricate, compacted meanings.

In any case, when dealing with concepts that represent groups, curators will need to learn to pay attention to ambiguities in natural language (or use curation software that could flag such pitfalls). They should at least interpret expressions like “exactly half‑of” literally, i.e. as guided by Principle 3. With the primary meaning, it would quite simply express “half‑of <has‑precision> exact”, i.e. ‘a/some exact half’. which would need to be captured with a second VSM-sentence in addition.

• Note: 36 The challenge of non-interchangeable connections arose when the designer of VSM (S.V.) heard someone talk about their ‘old new me’ vs. ‘new new me’. Also here, the ‘new me’ is first contextually isolated, before being referred to again with the extra adjective.
Fig. 8b inspiration: M.K.‘s black dog Boomer.

3. Additional functionality

While VSM-terms and -connectors provide the core functionality of VSM, some additional features further increase both the usability and expressivity of VSM. We built some of these functions into the vsm-box UI already, while some others are initial ideas that could be further developed.

3.1. Templates

The Full Story

While VSM enables flexible information capture, some curation tasks involve the repetitive entry of similarly structured facts. To streamline a routine entry of facts, one could use a VSM-template. This is a predefined structure of connectors, and a series of VSM-terms, including empty fields that need to be filled out with a term+ID by the user.  – In a typical curation project, template designers would accompany a template with set of curation guidelines. These describe what type of term is expected in empty fields, in addition to how to interpret text in relevant scientific literature (which is what requires the most effort). Note that VSM-sentences (including VSM-templates) can closely resemble natural language, which makes it clear for the curator to know what to fill in where, and essentially explicates parts of the adopted curation guidelines. Like this, VSM makes the effort it takes to interact with the information entry software minimal.

Figure 9 shows several templates. Each empty field can be associated with one or more preferred dictionaries, e.g. genes or cell types. This can help the autocomplete function to either rank these terms easily accessible toward the top of the list, or to limit entry to these terms only. Curators need not be restricted to capture only what a template was designed for: because a template is also just an (uncompleted) VSM-sentence, they can still insert additional terms and connect them to the rest of the structure. For broader curation tasks, a curation platform could provide access to several VSM-templates in a menu, each dedicated to a particular information type that the curator may want to capture from a paper. The software may even enable users to design new templates.

Figure 9. VSM-templates. (a) A VSM-template in which the first two empty fields would be associated with the lookup of protein terms. The third field expects cell types and cell lines, which this template also clarifies in grey placeholder text. (b) An example of that template, filled in. The result is a normal VSM-sentence. (c) A template in which the focused, empty field features a list of easily selectable, one-click terms (pre-associated term+ID couples for that field). The last empty field would limit its term-lookup to a controlled vocabulary of experiment types.

3.2. External coreference

In natural language, one can build a story by referring to any concept that was conveyed in an earlier sentence (e.g. ‘John buys cheese’; ‘One hour after that, a mouse eats it‘). Because each term in a VSM-sentence is a specific concept, one can refer back to any of them; and not only from within a VSM-sentence but from other VSM-sentences as well. Software for managing multiple VSM-sentences could support such coreferencing between sentences, visually or through internal IDs. This could for instance be used for unambiguously defining the steps of biological experiment protocols.

3.3. Head

As explained in Principle 3, each term receives extra context-meaning from all its connected terms. As a consequence, every term actually represents the entire content of the VSM-sentence, although from its own perspective: e.g. in Fig. 2b the fifth term can be read as ‘the fork used for eating of chicken by John’. In natural language though, a sentence conveys information only from one particular perspective; e.g. in Fig. 2b, the sentence implies focus on ‘the eating’. This perspective or intended main meaning is embedded in the term “eats”; and this term is also the one that subsequent sentences would refer to when stating e.g. “that was observed in 2021″. The VSM UI could enable a user to assign this focal term or head in a VSM-sentence via an Alt+click on or above the VSM-term, and could draw dashes above the term and under the connectors. Alternatively, software could automatically detect this head by analyzing the word-forms chosen for VSM-terms (e.g. a verb, vs. preposition or noun).

Next, when one wants to refer to a sentence as a whole, e.g. for adding meta-information, software could automatically or by default make an external coreference toward this intended term.

3.4. General and Data terms

VSM-terms (as shown until now) each represent a specific concept, embedded in a particular context. Ontologies, however, define relations between classes (categories, sets that generalize over particular concepts), which may also be viewed as general concepts that are not bound to a particular context. In addition, RDF supports literal text-strings or data, which do not really represent any concept or ID. In order to support these distinctly different and useful types too, a VSM UI may enable setting a VSM-term’s type to one of: Instance (specific), Referring Instance (specific, referring child), Class (general), or Literal (raw data or text); and show these in a distinct rectangle color or style (blue, dashed-border blue, yellow, and red, resp.). A VSM-sentence can then express e.g. “duck [in-general] is-a-type-of [specific] bird [general] according-to [specific] ontology-version-X [specific]”, or “chicken [general] has-alias [specific] hen [literal-string]”, or “proteinX has-sequence MEEPQSDPSVEPPLS [literal]”.

Note that in case one would use VSM as an interface for generating RDF or OWL models, VSM does not impose the use of either OWL Instances or OWL Classes.14 OWL: Web Ontology Language: www.w3.org/TR/owl-ref. Still, it may make sense to translate the default, specific VSM-terms to OWL Instances, as bioscience findings are typically reported in the context of particular experimental or biological conditions. A further treatise on connecting the meaning of VSM concepts to e.g. formal logic is outside the scope of this page. Also the topic of general vs. specific concepts in VSM will be discussed later.39 In the VSM view, the only difference between (VSM‘s) specific and general terms/concepts is whether a concept is thought of as being attached to any explicit or implicit context, or not, respectively.

The current page has presented VSM as a human-oriented method for precise, meaningful construction of contextualized knowledge. It explains VSM‘s human-thought inspired semantics primarily in an end-user oriented form, and describes it together with an enabling, intuitive, general-purpose UI design.

An overview of VSM is on the Summary page

Many interactive examples are on the Examples page

More about VSM‘s implications and roll-out, is on the Discussion page

or go back to the Introduction