by Janina Jacke
drawing on the minutes of Marie Flüh, Jan Horstmann and Mareike Schumacher, and on the summaries of Jan Christoph Meister and Marco Petris
JAN CHRISTOPH MEISTER introduces the workshop topic and its relevance in the context of the forTEXT project: Most text annotation tools, especially in the field of computational literary studies, leave their users the choice between either one of two possibilities for organizing annotation categories: categories can be organized strictly hierarchically, or not at all (i.e., all of them coordinated at the same hierarchical level). Once a hierarchical ontology is created, it is hard to re-organize it by moving categories to other hierarchical levels. This lack of flexibility is likely to impede the use of ontology-based annotation in literary studies: due to the multiplicity of literary studies theories and methods, not all approaches might conform to the ideas of optimal ontology construction developed in natural sciences. Instead, they might call for additional, alternative ways of organizing categories and for more flexibility in bottom-up ontology development.
As Meister explains, the workshop serves as an initial approach to exploring and discussing the conceptual and technical conditions under which such flexibility is possible.
DINO BUZZETTI starts his take on the workshop topic in his paper Text as a self-regulating adaptive system. Towards a connectionist approach to text analysis from the idea that diacritical ambiguity is a basic feature of natural language texts and, at the same time, of embedded (i.e. inline) markup in natural language texts. Embedded markup is, on the one hand, referential and may, for example, explicate an annotated text on the content level in terms of its denotation or meaning. On the other hand, it grants us a license to interpret that very text in a specific way: it has a self-referential operational function. In other words: embedded markup is equally part of and about the text – a feature it shares with diacritical signs. Embedded markup can thus be explicated in logical terms as a type of ambiguous textual expression that can be processed as an inference licence: it informs us what and how to read.
In a logical perspective, the what and the how in an ambiguous, double-readable expression are mutually dependent: the what-aspect of such an expression is that which presents specific “statements of fact” as data; the how-aspect of the expression is that of a general, abstract inference rule which licenses us to draw a particular if-then conclusion. In terms of the how-aspect, the expression is thus on the one hand a rule that serves as a warrant (i.e. an abstract conditional “if D [data] then C [conclusion]” licence); in terms of its what-aspect, the expression provides the data which allow us to operationalize the rule in a particular instance. In doing so, we “back” the abstract, general rule by instantiating the variables of the abstract conditional using the concrete data, i.e. the referential facts which the sentence asserts.
Because of this logical interdependency between the general rule and the specific data in a self-referential expression, Buzzetti argues, the distinction between the metalinguistic and the object-language understanding of one and the same sentence is in fact reduced to “an operational difference”. Following Barwise and Etchemendy, Buzzetti argues that “the object language/metalanguage distinction is inappropriate”: a higher-order (meta-language) statement, where it is formulated from a non-deterministic point of view and explicates a text’s meaning in probabilistic terms and therefore conceptually acknowledges polysemy, turns into a quasi-first-order de re statement about the world as such. Vice versa, a first-order object-language statement, where it takes on the meta-linguistic de voce nature of a “statement on the meaning of words” (Anselm of Canterbury), becomes the inferential equivalent of a second-order “statement as to how things are”. Second-order object language de re-statements are thus inferentially equivalent to first-order metalinguistic de voce-statements.
Statistical data about a text (as derived either from manual annotations or gathered from automatic word embeddings or vector analysis) can now be assumed to have both a factual referential facet and an operational facet such as a rule or interpretation. Assuming that such statistical data is then modelled as a graph structure that formalizes “logical dependencies between a set of entities (concepts), their attributes, and their relations” which in turn can then be mapped onto a formal ontology, we are dealing with textual data or with concepts, depending on whether we focus on the referential or the operational aspect.
Transferring these perspectives to ontologies, the distinction between conceptual and metaphysical ontologies becomes relevant. Conceptual ontologies can be regarded as a nominalist approach that tries to enforce, in the spirit of Occam’s razor, conceptual austerity or “ontological economy”, as Buzzetti puts it, and “admits only statements that do not add to our ontology” – a ‘dogmatic’ approach that enforces hierarchy in ontologies. The other option is that of the realist (Platonist) and ultimately metaphysical approach which, in a bottom-up approach, on the basis of the statistical data and graphs, infers the existence of and introduces new ontological concepts and postulates new relations: the ‘undogmatic’ approach that defies predetermined hierarchy.
Buzzetti implies here that the conceptual (‘dogmatic’) ontology is based on the ‘reading’ of statistically analyzed and ‘graphed’/modelled annotation data as first-order referential statements on ‘what the world (= text) consists of’. By contrast, if we read them as second-order operational statements on ‘how the words and phrases in the world (= text) may be interpreted’ (i.e. as interpretation rules), we are in the field of non-hierarchical ontologies where “different co-occurrence distributions determine multiple dependence relations for each node and so define a non-hierarchical structure that can be conceived of only in terms of possible ambiguity”. Rather than mapping univocal logical dependencies in a referential graph, we will now have to map ambiguity by way of an “adaptive system of graph differentials”, i.e. operationally.
The discussion of Buzzetti’s paper addresses mainly three topics. First, since Buzzetti’s argument refers to inline markup, it is discussed whether the argument is equally true for standoff markup. This leads directly to the question of whether the difference between inline and standoff markup is merely a pragmatic one (i.e. inline markup could be translated to standoff markup and vice versa) or a conceptual one. Bottom line of this strand of discussion is that the two types of markup are in fact mutually translatable so that Buzzetti’s argument can be applied to either type. The discussion then directly proceeds to the question of which type of markup is more apt in humanities contexts. Here it is suggested that embedded markup might be the better choice to express operational facets (e.g. suggested readings of text passages), which is a central use case in the humanities.
Second, taking up the evaluative perspective on the choice between highly formalized/strongly regulated (‘dogmatic’) and less structured but more flexible (‘undogmatic’) versions of modelling data, the discussion tackles the question of whether conceptual or metaphysical ontologies are the better choice in humanities contexts. While some parties advocate the view that both ontology development and ontology-based annotation of texts in the humanities requires substantial flexibility and must even allow for unresolvable contradictions, a competing view is that strict rules developed in science contexts are in principle transferable to humanities disciplines. In the latter view, competing ontologies addressing the same object domain (e.g. literature) should only be admissible if they are organized/related to each other by a meta-ontology explicating their status.
Third, and related to the second issue, the requirements of ontologies and workflows of ontology creation in humanities contexts are discussed. Here, the necessity to move concepts within an ontology is pointed out as an important factor for bottom-up ontology creation. Based on this observation, the need to document decisions in the process of building an ontology and the resulting development of an ontology over time is discussed (– with the open question of whether there can be a ‘final’ version of an ontology with no further changes). While versioning options – as available in tools like CATMA or Protégé – can in theory cover the requirement of documenting the temporal development, suitable visualization options to explore the development are still a desideratum. An interesting aspect in this strand of discussion is the question of whether ‘outdated’/reworked parts of an ontology should remain part of the final ontology in humanities contexts or whether they should be discarded in favor of the most ‘up-to-date’ version of the ontology.
LUDGER JANSEN approaches the workshop topic by asking: Are non-hierarchical ontologies possible? He starts off with some definitional remarks that help answer the question of whether non-hierarchical ontologies are logically possible. Ontologies (in the use of the term that is relevant here) are models of which general features there are in a domain, and they normally consist of classes and the relations that hold between these classes. What follows from Jansen’s terminological differentiations is that non-hierarchical ontologies are logically possible if we use a narrow concept of hierarchy (namely a strict order relation where subclasses must not be co-extensive with their superclass). If, however, we think hierarchy in terms of partial order relations (i.e. superclass and subclass can be co-extensive), then every ontology that contains at least one class is (at least implicitly) hierarchical.
Jansen then addresses the question of whether and in which respects ontologies are usually hierarchical. Here, he first looks at the ontology editor Protégé to see which relations the tool stipulates as ‘necessary’, finding that two types of hierarchical relations (and only those) are actually hardwired into the tool (subclassOf and instanceOf), while the definition of other relations may be freely defined by the user. Jansen then shows how some actual ontologies (from the field of biology) are characterized through a heavy (and sometimes exclusive) use of one type of hierarchical relation, namely the subsumption (or “is a”) relation. He finds, however, that this practice often entails several disadvantages.
This observation leads Jansen to suggest a set of four “good practice rules” for building ontologies. These rules are supposed to ensure that ontologies are, on the one hand, non-trivial, “interesting” in Jansen’s words, and, on the other hand, easy to develop, to overview and to maintain. The rules also imply an answer to the question: Should ontologies be hierarchical – and, if so, in which way?
The first rule is: One should, if possible, use standardized relations between the classes of an ontology, or, in other words: one should re-use relations that have been defined elsewhere; for example in the Relation Ontology as developed by Smith et al. This ontology includes different kinds of hierarchical (or: order) relations as well as other types of relations, i.e. non-hierarchical ones. According to the second good practice rule proposed by Jansen, if new rules need to be introduced, these rules should be well-defined and -documented – on the one hand with the help of natural language descriptions, and on the other hand in terms of their logical properties like reflexivity, symmetry and transitivity. These first two rules do not have any implications for the question of how and, if so, in which way ontologies should be hierarchical.
The third rule Jansen suggests is to avoid the exclusive use of one specific type of hierarchical relation, namely the subsumption / subclass / “is a” relation. This is a partial order relation (which means that it is compatible with a subclass sharing the same extension with the superclass). As Jansen explains, some ontologies from the field of biology make exclusive use of this relation. Together with the attempt of integrating into one ontology different criteria according to which classes can be subdivided, this subsumption overload can easily lead to polyhierarchical ontologies. In this type of ontology, subclasses can have more than one superclass, which makes this kind of ontology difficult to develop, overview and maintain.
This leads to Jansen’s fourth good practice rule: We should stick to monohierarchical ontologies. Jansen then goes on to explain how it is still possible to generate interesting, non-trivial ontologies: We should disentangle the different criteria that are combined in a polyhierarchical ontology into separate monohierarchical ontologies that make use of different kinds of relations, and we should carefully define the classes and relations of these monohierarchical ontologies. We can then use reasoning programs to make all sorts of inferences that are possible with polyhierarchical ontologies – only less “messy” and easier to develop and maintain. Rules three and four thus do have implications for the answer to the question of whether, and, if so, in which way ontologies should make use of hierarchical relations: They should, and they should combine different kinds of hierarchical relations with other well-defined non-hierarchical relations in order to avoid polyhierarchies and still be interesting.
In the discussion following the paper, the differences between research practices in the humanities and in the sciences are the major focus. Many questions revolve around reasoning functionalities for ontologies which are readily available, for example, in the Protégé ontology editor: a component that can be used to check the ontology for consistency and deduce implicit properties of the classes.
A first observation in this context is that in literary studies, reasoners for ontologies are hardly (or not at all) used. One likely reason for this seems to be that literary scholars do not work with complex ontologies that include different types of well-defined relations between the concepts. In general, neither the consistency of an ontology nor deducible statements seem to be a generally upheld goal in disciplines like literary studies: Scholars with certain theoretical backgrounds may deliberately want to keep inconsistencies in ontologies, and deduced statements will often be considered trivial.
This leads the workshop participants to discuss how research in the humanities could specifically profit from using reasoners. A potentially interesting use case mentioned for the context of annotation is that a reasoner could calculate all categories implicit in an explicit annotation (as well as, e.g., classes compatible with or excluded through the explicit annotation, based on the applied ontology). This could be a helpful tool both for the process of bottom-up ontology development and hermeneutic annotation. Currently, however, tools like CATMA do not support a straightforward definition of relations between classes except a subsumption relation – and even the latter does not have a strict semantics, since properties are not inherited from super- to subclass.
A different perspective on the paper that is addressed in the discussion is that, if we assume that the use of reasoning functionalities is not a common use case in humanities, then the arguments for avoiding polyhierarchies offered in the paper may not be as compelling. A central argument is that polyhierarchical ontologies are more difficult to maintain, since all implications of changes made to the ontology have to be recognized and implemented manually while this is not the case for monohierarchies. However, if consistency and deduction from ontologies are of limited interest, then this task could be neglected. This thought, together with the impression that polyhierachies with multiple inheritance could be a powerful tool (e.g. to systematize different perspectives on an object via meta-ontology), leads the participants to discuss whether multiple inheritance polyhierarchies are feasible in tools like CATMA. While the technical implementation would not be a problem, polyhierarchies could prove problematic interface-wise.
In general, it – again – remains an open question whether (and, if so, which and to what end) academic standards developed in science contexts should ‘dogmatically’ be imported to disciplines in the humanities, especially if these standards may conflict with some literary theories.
CLAUS HUITFELDT and MICHAEL SPERBERG-MCQUEEN tackle the question of whether and, if so, under which circumstances ontologies can usefully be modeled as trees or unrestricted graphs – and whether there might be more appropriate solutions for modelling ontologies.
They start with a set of definitions and assumptions. First of all, an ontology is “any attempt to say what kind of things exist in any domain or universe of discourse, and to give some account of their properties and relations to each other”. To find what exists, they propose Quine’s method according to which “to be is to be the value of a variable”. They observe that ontologies may conflict with each other on the concept as well as on the application level.
They argue further that an ontology can be seen as a classification scheme which is used to group items into classes on the basis of certain characteristic properties. The extension of such a class is defined as all items falling under that class. The intension is the set of characteristic properties of that class. Such a property can be seen as a predicate, i. e. a function which maps one or more objects from the object domain to the truth values true or false.
With these definitions and assumptions, Huitfeldt and Sperberg-McQueen start by observing that modeling ontologies as trees is an old and widespread practice because of the ability to partition a class into various subclasses, thus having an organizing principle for the physical arrangement of objects and assistance in preparing aggregate statistics. The problem, however, is that the commitment to one hierarchy immediately excludes all other possible hierarchies, which can prevent the building of classes that are otherwise desired.
The authors then turn to ontologies as unrestricted graphs. Here, we do not have the monohierarchy as a limiting factor, but the possibilities of inference would be very limited. Another problem would be to present such an unrestricted graph to the user of an annotation system in some useful manner.
This insight leads to an analysis of the requirements for structuring ontologies in annotation systems. What are the benefits of a hierarchical structure for an annotation system? First, it should provide guidance for the user interface and the annotation management; second, it should provide logical inference possibilities about the properties of annotations and the annotated text passages.
By showing how subclassing (as a means of specialization) and superclassing (as a means of generalization) can provide a great amount of inference possibilities, the authors introduce a lattice as a useful structuring principle. The lattice is defined in terms of the subset relation. Each node is a subset of those nodes connected upwards and a superset of those nodes connected downwards. The closest common ancestor of two nodes is called their meet, and the closest common descendant is called their join. The lattice provides similar possibilities as monohierarchical structures when it comes to fulfill the requirement of guidance of the user interface, and it also has similar inference possibilities. Additionally, however, it has greater flexibility when new generalizations need to be introduced. These generalizations can be useful either to ease the annotation process or to enhance the inference mechanism, for example when defining new annotation classes by operations on existing classes.
Since Huitfeldt’s and Sperberg-McQueen’s paper promotes exactly the kind of polyhierarchical that is rejected in Jansen’s paper, the discussion can directly follow on from the previous one. The main topics are the added value that lattice structures provide, the appropriateness of the structure for humanities context and the feasibility of implementing lattice structures for ontologies in annotation tools.
One strand of discussion addresses the question of the advantages of lattice structures over monohierarchies. While a lattice structure will facilitate representation of (and inferencing on) more complex subsumption relations, more complex inferencing will only be supported if other relations between classes, axioms and inference rules are brought in. However, since it is still unclear whether more complex inferencing based on ontologies is required in humanities contexts, a lattice could already provide the added value.
A different question raised concerns the openness of lattice structures, since an example given in Huitfeldt’s and Sperberg-McQueen’s paper seems to be based on the closed-world assumption (i.e. everything that exists in the domain is modelled into the ontology). While a closed-world assumption may be appropriate to the (rather special case of) analysis of the annotations of a given document as discussed in the paper, it is agreed that ontologies should support the open-world assumption, and that the use neither of lattices nor of monohierarchies commits one to the closed-world assumption.
A second major point of discussion is the feasibility of implementing lattice structures for ontologies in annotation tools like CATMA. Since there already exist annotation tools allowing for lattice structures/multiple inheritance, this is obviously possible. In CATMA, however, minimum conditions for implementing lattices would be to be able to move categories/tags around in the hierarchy of a tagset and to relate supertags to each other. Also, as already addressed in the discussion of the previous paper, suitable visualizations should be developed to help build and overview increasingly complex lattice tagsets and use them for annotation. While a lattice structure is mathematically defined, it would also be helpful to view such tagsets as lattices (or diamonds) to promote an understanding of the logical structure.
In summary, it can be said that the workshop was a fruitful starting point to engage in interdisciplinary discussions of ontologies in the humanities. While the workshop focused on philosophical-technical ideas concerning alternative ways of organizing ontologies and implementing these systems into annotation tools, further discussion could benefit from broadening the scope: What is the general role of ontologies in (different) humanities disciplines? Which are the criteria for fruitful ontologies in these disciplines? What are common humanities use cases in which ontologies are developed and put to use? These and other questions could be addressed in an edited volume – a call for papers will be distributed via the relevant mailing lists in the course of 2020.