GUM: The generalized upper model

Abstract

GUM is a linguistically-motivated ontology originally developed to support natural language processing systems by offering a level of representation intermediate between linguistic forms and domain knowledge. Whereas modeling decisions for individual domains may need to be responsive to domain-specific criteria, a linguistically-motivated ontology offers a characterization that generalizes across domains because its design criteria are derived independently both of domain and of application. With respect to this mediating role, the use of GUM resembles (and partially predates) the adoption of upper ontologies as tools for mediating across domains and for supporting domain modeling. This paper briefly introduces the ontology, setting out its origins, design principles and applications. The example cases for this special issue are then described, illustrating particularly some of the principal differences and similarities of GUM to non-linguistically motivated upper ontologies.

Keywords

GUM foundational ontology linguistic ontology ontological analysis formal ontology ontological motivation

1. Introduction: Historical background, motivations

The Generalized Upper Model is the current version of an ontology whose development begun in the context of natural language processing AI projects in the 1980s. As attempts were made to build computational systems capable of more sophisticated, and natural, ‘intelligent’ behavior, the need for detailed knowledge concerning the domains of application became increasingly clear. As a consequence, there were several initiatives both to suggest general principles of organization for domain knowledge reusable across domains and to deliver specific well-organized instances of domain modeling that could similarly be reused. Many early examples of proposed ontologies, including foundational ontologies, can be traced back to that time. In the vast majority of areas where natural language processing was attempted, however, there were no ready made domain ontologies available for use. One set of problems here relates to breadth – i.e., when constructing linguistic components capable of dealing with a substantial proportion of the expressive possibilities of a language, it is also necessary to determine how those resources relate to the content they are to communicate or interpret. Indeed, even when there are domain ontologies available, it is by no means straightforward to relate those ontologies to the kinds of knowledge organization supportive of natural language processing.

Echoes of this mismatch in modeling requirements return in a current point of discussion in foundational ontology design. One of the basic ontological design choices discussed in Masolo et al. (2003a, 7) is that between ‘descriptive’ and ‘revisionary’ ontologies. A descriptive ontology is “based on the assumption that the surface structure of natural language and the so-called commonsense have ontological relevance”; a revisionary ontology considers “linguistic and cognitive issues at the level of secondary sources (if considered at all), and does not hesitate to paraphrase linguistic expressions … when their ontological assumptions are not defensible on scientific grounds”. Prototypical examples of these ontology types are DOLCE and BFO respectively. The assumption that surface structures of natural language may have ontological import is widespread, but far from uncontroversial. Less controversial perhaps is the position that for certain applications of foundational ontologies, particularly those more related to human-level tasks and understandings (the ‘mesoscopic’ level according to Masolo et al., 2003a, 13), attending to linguistic patterning may be beneficial. There is, however, no automatic guarantee that the resulting chimera of foundational ontology and linguistically-responsive representations is necessarily a coherent goal.

The Generalized Upper Model adopts the descriptive approach but with an explicit additional separation imposed between modeling decisions that are linguistically motivated and those which are not. The question as to which modeling decisions are then ontologically relevant in the revisionist sense, i.e., as identifying the necessary organization of ‘the world’, is translated to a mapping task between ontological modules. On the one hand, the ‘module’ of the Generalized Upper Model accepts motivations for its internal organization solely on the basis of linguistic evidence, in a very specific sense to be explained below. On the other hand, it is accepted that foundational ontology modules developed on the basis of other criteria might well differ in their internal organizations. The framework as a whole thus presupposes a heterogeneous collection of linked ontological modules, where each might be subject to distinct sources of motivations.

The theoretical position adopted for the Generalized Upper Model has many consequences for how it is applied and, indeed, for the kinds of tasks that it is intended to serve. These lead it to differ in several substantial respects from some other traditional foundational ontologies. In short, the Generalized Upper Model seeks to reveal the necessary organization of any descriptions of the (human) world – and, in particular, of any descriptions mediated by natural language. The Generalized Upper Model consequently occupies a very specific theoretical location in relation both to natural language expressions and to contextualized interpretations of the meanings of such expressions. Whereas the latter are commonly related to the domain of foundational ontologies, the assumption that it is straightforward, or ‘philosophically transparent’, to move from linguistic expressions to contextualized interpretations is considered questionable. Formulating accounts in terms of the Generalized Upper Model is then seen as a way of helping to distinguish more systematically between linguistic processes and more traditionally ontological concerns. Since many philosophical forms of argument, particularly in analytic philosophy, rely on linguistic examples (or examples formulated to bring out distinctions in language use), this orientation to method is seen to be beneficial and so will be emphasized in the discussion of the particular case studies below.

The level of representation pursued by the Generalized Upper Model is anchored into patterns of linguistic expressions so that strong methodological guidance can be offered when moving from linguistic expressions to representations of the semantic commitments of those expressions. The tight relationship maintained between the level of linguistic expression and the categories and relationships of the Generalized Upper Model has the consequence that many regularly occurring patterns are explicitly characterized as linguistic processes relating linguistic meaning and form rather than constituting ‘revisionary’, or ‘world’-centered, ontological concerns. These processes range from the very general, such as nominalization, by which almost any elements contributing to a Generalized Upper Model specification can be ‘expressed’ as nouns – including events and relationships such as being the actor of an event and other roles – to more specific linguistic processes, such as the well-known ‘universal grinder’ (Pelletier, 1975), that enables count nouns to be used as mass nouns (‘add apple to the recipe’), and Bunt’s (1981) converse ‘universal sorter’, by which mass nouns can be turned into count nouns (‘we have three wines of note’). Moreover, and of particular importance for the discussion below, it will be shown how following the definitions of the categories and relations of the Generalized Upper Model makes it possible to ‘read off’ of a grammatical analysis of a linguistic expression a corresponding semantic characterization in a reliable fashion that ‘unwinds’ shallow linguistic processes such as nominalization. These semantic characterizations then provide a reference point for relating to further foundational ontological accounts without committing to the modeling decisions made in those accounts. Thus, in many respects, a Generalized Upper Model description may act as a bridge or statement of equivalence relating diverse foundational ontological proposals for the modeling of particular phenomena. This again stresses the importance of accepting a general model of ontology design based on heterogeneous accounts.

Positioning the Generalized Upper Model in relation to linguistic expressions and other foundational components in this way supports the practical task of modeling by providing a well-specified procedure for mapping between linguistic forms and semantic descriptions suitable for further ontological consideration. Such descriptions have been used both for automatic language generation (Bateman, 1997) and as a target representation in automatic language semantic parsing (Bateman et al., 2010). A useful analogy for the use of the Generalized Upper Model in a language processing scenario can then be made with respect to the role of word embeddings in machine learning: just as word embeddings, as continuous characterizations of semantic (and other) relationships, are better suited to subsequent numerical processing, a Generalized Upper Model semantic specification, as a well motivated and richly internally-structured entity, is more suited for subsequent symbolic processing than is the linear string of corresponding linguistic elements. Subsequent processing is explicitly intended to include interfacing with other foundational ontological characterizations relying on contextualized interpretations and situational grounding of various kinds (cf. Pomarlan and Bateman, 2020).

The origins of the Generalized Upper Model lie in a specific knowledge structure developed in William C. Mann’s ‘Penman’ text generation project (cf. Mann, 1983,1985a) for interfacing between domain knowledge and linguistic knowledge. This structure was consequently named the Penman ‘Upper Model’ (documented, for example, in: Bateman et al. 1990; Bateman 1990). Classifying concepts from a domain model under appropriate Upper Model semantic classes allowed those concepts to inherit all the general methods for linguistic expression defined for those classes in the Penman system. Automatic natural language generation could then proceed for the knowledge of the domain without further adjustment. The Upper Model itself thus offered a domain- and task-independent organization of semantic categories already similar in function to many more recently developed foundational ontologies, but with the additional property that a tight relationship with forms of expression in natural language was guaranteed.

The Upper Model organization went through several rounds of expansion. In the late 1980s, a cooperation between the Penman text generation project and the Bolt, Beranek and Newman Inc. (BBN) natural language understanding project explored the extent to which the various distinct sources of knowledge concerning language might be combined within a single architecture known as Janus (looking both ways) for language interpretation and production (Weischedel, 1989). This led to a thorough redesign of the Upper Model called the Janus abstraction structure (Mann, 1985b; Mann et al., 1985). An important facet of this redesign was to make far stronger use of motivations drawn from the linguistic theory adopted within the Penman project, that of Hallidayan systemic-functional linguistics (Halliday and Matthiessen, 1999). This orientation remained and was developed further throughout the 1990s, particularly with respect to including or covering distinctions revealed by the grammars of languages other than English (cf. Bateman et al., 1994,1995; del Socorro Bernados Galindo and Aguado de Ceo, 2001). At this point the organization was renamed the Generalized Upper Model (GUM) in order to reflect the broader range of evidence appealed to. Bateman and Lestrade (2014) then took this further with respect to issues arising in linguistic typology. Finally, in the context of the Collaborative Research Center Spatial Cognition of the Universities of Bremen and Freiburg (2003–2014), the Generalized Upper Model was extended significantly to address spatial semantics more broadly, adopting as before the linguistic (and, in particular, the grammatical) structuring of information as the primary source of motivations for semantic distinctions (Bateman et al., 2010). During this work, GUM was also updated to reflect more current knowledge representation practices and tools available for working with description logics and principles for productively enforcing the modularity that had always been assumed in the design, but not previously implemented.

2. Principles and structure of GUM

The approach adopted for developing GUM takes the underlying organization of the grammars of natural languages as a methodological guide for discovering those semantic distinctions and organizations that need to be maintained regardless of specific domains or applications. Support for this methodology lies in the observation that grammatical organization appears to be non-arbitrary; most current linguistic theories in fact assume that a tight relationship holds between the organizational properties of grammatical resources and those of semantics. Frameworks where this is followed range from generative grammar and treatments of ‘alternations’ (Levin, 1993; Jackendoff, 1999; Levin and Rappaport Hovav, 2005), through cognitive approaches (Langacker, 1988; Frawley, 1992), to the social-functional accounts that form the foundation for GUM (Halliday and Matthiessen, 1999). Particularly in this latter tradition, language is seen as playing an active role in any culture’s construction of its ‘social world’ and is by no means an arbitrary labeling of an independently existing ‘reality’. On the contrary, the social reality within which human experience unfolds is taken to be structured through and through by distinctions drawn in linguistic semantics (for further aspects of this discussion, see, e.g.: Bowerman, 1999; Levinson et al., 2002). It is this property that is argued to make straightforward boundaries between characterizations of the ‘world’, on the one hand, and linguistically-inflected constructs, on the other, problematic – particularly for the human mesoscopic level.

GUM therefore presupposes a view of language in which there is a strong non-arbitrary relationship between the grammatical organization of some language and the ‘experiential world’ constructed with and through that language. Language is seen both as providing access to, and as articulating, the world in a form that renders it intelligible. While a strong conventionalist stance of the kind that all facets of our ‘lifeworld’ (Schutz and Luckmann, 1974) are socially constructed is no doubt too strong, the divisions drawn by language can be seen nevertheless to offer a valuable perspective complementing cognitively-motivated distinctions. In many respects, just as any ‘reality’ is necessarily filtered through our perceptual systems, it is also filtered through the language system—at the latest whenever we wish to communicate (cf. ‘thinking for speaking’: Slobin, 1987). In an analogical sense, then, just as we might construct an ontology informed by the distinctions that our cognitive system appears to deliver to us (colors, shapes, objects, actions, etc.), GUM is informed by the distinctions that language delivers to us. Nevertheless, this remains of necessity a characterization of descriptions and does not commit to a simple construction of ‘the world’. This becomes increasingly relevant when moving into more abstract realms, including categorizations of activities, events and relations, which are often clearly social constructions and can be subject to quite divergent philosophical positions concerning how they might be best captured. Regardless of which philosophical positions are pursued, using GUM, grammatical evidence can be accrued for constructing bridging representations.

The use of linguistic evidence within GUM and the GUM development methodology hypothesizes that if a distinction is drawn in the grammar of a language, then that distinction may be usefully considered for its ontological correlates also. This contrasts with many other linguistically-motivated systems of semantics, such as FrameNet, VerbNet, WordNet and systems descended from these (Bateman, 2010c; Huang et al., 2008), in that GUM relies solely on grammatical distinctions rather than lexical distinctions. This distinction between lexical semantics and grammatical semantics is adopted as central because it is precisely the function of lexical items to be specific, to bundle semantic properties together for ready use in ways that might also be relatively idiosyncratic. Organizations of this kind are consequently considered less likely to reveal the kind of generic semantic configurations necessary for broader, foundational statements. In contrast, grammatical organization needs to generalize across both situations and individual types of entities and so promises a more robust indication of semantic import. Jackendoff argues this as follows:

“…every major phrasal constituent in the syntax of a sentence corresponds to a conceptual constituent that belongs to one of the major ontological categories.” (Jackendoff, 1983, 67)

GUM’s genericity is grounded in this organizational role of grammar.

The GUM approach, and the linguistic theory it builds upon, assumes the most significant grammatical domain where semantically-significant alternations are located to be that of the clause. Thus, to uncover reusable semantic classes, GUM first considers clauses as they occur in natural everyday language usage. Reoccurring patterns of phenomena that appear mutually exclusive in linguistic data are taken as ‘reactancies’ of semantic distinctions; the notion of ‘reactance’ refers to the phenomena that entire collections, or ‘syndromes’, of grammatical choices appear to pattern together: that is, choices in one area will co-occur with choices in other areas. Whorf introduced such syndromes in terms of cryptotypes, covert semantic categories that are only evidenced by sets of overt linguistic devices (Whorf, 1956). Grammatical clusters thus identified are consequently hypothesized to correspond with semantic distinctions.

One of the broadest such reactancies is the three-way division of clauses into a grammatical process (typically expressed by some verbal component), a (small) set of obligatory participants in the process (typically expressed as nominal phrases), and a larger set of optional circumstances (typically expressed as prepositional phrases), setting out the manner, setting, location, time, extent and so on for the occurrence of the process. This grammatical pattern, reoccurring across most languages of the world, is taken by virtue of the non-arbitrary relationship assumed between grammar and semantics to be the most significant class for the subsequent organization of GUM. The covert semantic category corresponding to the pattern is labelled in GUM as a Configuration; this is taken as the basic fragment of grammaticized experience embodying ‘one quantum of change’ (Halliday and Matthiessen, 1999, 128). Consequently, when using GUM for analysis, each clause found in a natural language description is related to a corresponding GUM Configuration category, applying a series of grammatical probes that bring out the grammatical reactancies at play.

Correlating directly with the primary three-way grammatical reactance, Configurations organize three distinct types of elements by defining three roles by which those elements may be brought together within the configuration. Thus, each configuration obligatorily has a process role, one or more participant roles and optionally any number of circumstantial roles. In addition, since the configuration is the primary unit of temporal development, several further relations determine the particular ‘temporal profiles’ of their occurrences. Such profiles have their own possible ‘shapes’ as described in terms of Aktionsart and the classes set out by Vendler (1957) and many since. As with many of the GUM categories, the particular terms identified are seen as interfaces to potential axiomatizations that may be developed and maintained separately from the GUM internal organization. Moreover, there is no requirement, or even expectation, that there will be philosophical agreement concerning just which axiomatizations are ‘best’. This is unnecessary from the heterogeneous representation standpoint. Nevertheless, regardless of axiomatization choices, the distinctions set up in the GUM representation constitute constraints, or requirements, that those axiomatizations will need to satisfy if they are also going to be supportive of the tasks for which GUM is applied, such as natural language comprehension or generation.

The top-level categories of GUM are then as shown graphically in Fig. 1. The categories for temporal profiles appear on the left of the graph, the immediate subtypes of configurations run across the lower middle of the graph, and the further Elements that combine to give configurations their internal ‘constituency’ structures are shown on the right. Subtypes of Configurations are distinguished by restrictions on their participant roles and the fillers of those roles; neither circumstantial roles nor temporal profiles are used to form subclasses.

Fig. 1.

The upper taxonomy of categories in GUM.

Another useful graphical representation for the relationships between categories is that given in Fig. 2, drawing on the notation adopted for DOLCE in Masolo et al. (2003b, Fig. 5). This sets out the formal dependencies among the top-level GUM Configuration categories: labeled arrows correspond to GUM relationships between categories, and nested boxes depict GUM’s is-a hierarchies. The figure shows two of the main types of Configuration, i.e., BeingAndHaving and DoingAndHappening, focusing for purposes of illustration on sub-branches involving spatial information. Thus, whereas BeingAndHaving involves more ‘static’ notions of, for example, placement, DoingAndHappening draws necessarily on an actor role and, for AffectingActions also on an actee, both filled by SimpleThings that either initiate the action or are acted upon, respectively. In addition, spatial DoingAndHappenings may draw on directions and routes along which the actions unfold. In each case, we can see that subtypes are accompanied by different configurations of relations into which the subtypes may enter. The distinct types of Elements shown on the right-hand side of the figure characterize the fillers that are possible for the roles applicable to Configurations; these include Process, SimpleThing, SimpleQuality, and Circumstance. Whereas the former three capture information central to the meaning of a clause employing them, Circumstances are (construed as) relatively peripheral entities that are not focused for the purpose of distinguishing configurations from one another; this is the reason that these elements do not play a role in building further subclasses of Configurations. Instances of the Element class consequently correlate with the semantic commitments of clause constituents, regardless of whether those constituents are nominal phrases, adverbial phrases, prepositional phrases, and so on.

Several very general modeling principles already follow from these distinctions. First, all modeling is oriented to a process view of phenomena – this follows from the centrality of the Configuration for structuring analysis. Second, distinct subclasses of Configurations have differing ‘temporal embeddings’: for example, BeingAndHavings are not compatible with change, whereas DoingAndHappeningrequire change. Third, certain categories are necessarily dependent as they are always accessed via independent entities. These and other principles will be illustrated more in the examples discussed below. Not all of the principles applied are found in the logical formalization of GUM, however. Correct use of the ontology is supported by adherence to more informally specified ‘interpretation policies’, that characterize and document the intention of the modeling decisions that have not yet received formalization.

Fig. 2.

Dependencies between the principal GUM Configuration categories distinguished by spatial relations and their defined roles.

All of the categories shown so far exhibit further diversification motivated by grammatical alternations observed in language use. SimpleThing, as the semantic category of entities that play participant roles in Configurations, tend overwhelmingly to appear as nominal groups. The grammatical properties associated with the use of SimpleThings then construct entities that are relatively stable in time and space and which may also have complex internal organization; stability is grammatically constructed by, for example, the restricted options for temporal modification within the nominal phrase. A further dimension of SimpleThing often registered by the grammar and with substantial discussion in the literature is whether something is decomposable or not. Decomposability refers to the degree to which an object is seen as a collection of its parts (Bateman et al., 1995, 50). A DecomposableObject is then an object that is viewed as a collection of subcomponents that may be taken apart; these parts are often given explicit recognition of their own, for example by receiving distinctive labels. NonDecomposableObjects, on the other hand, include Substance, correlating with ‘mass’ terms in the lexicogrammar, and spatiotemporal entities such as SpacePoint and TimePoint. Crucial here, however, is that it is always the underlying grammatical systems that are correlated with semantics – i.e., the possibility of distinguishing grammatically, for example, between conscious/nonconscious and so on, and not individual choices that a language may make, since these can readily be overridden by lexicalization conventions. All such interpretations are nevertheless construals and cannot be considered as being given simply by some assumed ‘facts’ in the world nor by the organization of that world.

The category SimpleQuality similarly characterizes property-like entities that can be ascribed to simple things in property ascription configurations; these are expressed in the grammar of English, as in many languages, as adjectives or adverbs. Qualities are divided further into the classes of ModalQuality and MaterialQuality. The former refers to “qualities of being able to do something, wanting to do something, having to do something, etc.” (Bateman et al., 1995, 53) and are usually manifest in areas of the grammar to do with commitment and intention. In contrast, the latter are observable, measurable properties that pertain to things. Lexicogrammatical exponents of MaterialQuality subsequently include adjectival constructions with heads such as red, soft, straight, etc. Such entities are seen as acting as interfaces into more foundational treatments of qualities, qualia, and so on, but are always necessarily dependent entities in the interpretation policy sense mentioned above.

The Circumstance class encodes information that is peripheral to the configuration but nevertheless plays a large role in GUM’s account of spatial language. As shown in Fig. 2, GUM formalizes space using two primary categories: GeneralizedLocation and SpatialModality. The concept GeneralizedLocation is used to provide a common semantics for any linguistic expression of a ‘place’—including prepositional phrases, adverbials, nominals and other spatially-relevant constituents when used in appropriate grammatical contexts. SpatialModality identifies just how a place is being construed in relation to other entities. These categories can again be seen as interfaces to any particular axiomatization (or simulation) satisfying the overall spatial requirements GUM identifies (Bateman, 2010b; Bateman et al., 2010).

The examples given so far have emphasized that the linguistic semantics according to GUM must be seen as characterizing descriptions of some articulated events, objects, situations, etc. In the technical terms of the linguistic theory used, the semantics construes events, objects, etc., assigning them particular organizational properties. GUM-descriptions therefore make no direct claims about the world, and so also relate to descriptions in the sense of Gangemi and Mika’s (2003) ‘Descriptions and Situations’ ontological extension for DOLCE. Establishing the precise relationships between these approaches would be a valuable investigation: GUM is then best seen as a foundational ontology of the content of a certain kind of description. Adopting an indirect relationship between semantics and the world in this way reflects and supports the observed flexibility of natural language use while still allowing underspecified semantic commitments to be reliably identified. Depending on their contexts of use, terms may take on quite different properties and so an appropriate semantics needs to be made responsive to this property without compromising the coherence of its own internal organization.

The ontological design principle of rigidity (e.g., Welty and Andersen, 2005) consequently applies here as well and motivates the maintenance of a strict ‘two-level semantics’ (cf., e.g., Lang and Maienborn, 2011) whereby semantic specifications only describe entities in the world; they do not give a referential semantics and properties of entities in the world. A classic statement of this was given by Hobbs over two decades ago in relation to the ‘ontological’ status of a spatial entity such as a ‘road’:

“When we are planning a trip, we view it as a line. When we are driving on it, we have to worry about our placement to the right or left, so we think of it as a surface. When we hit a pothole, it becomes a volume for us.” (Hobbs, 1995, 820)

These quite different ‘conceptualizations’ give rise to correspondingly different linguistic descriptions – such as ‘along the road’, ‘on the road’, ‘in the road’, and so on. It would be undesirable from the perspective of a sensible formalization if entities were free to be classified in mutually inconsistent ways. The often diverging identity criteria necessary for the entities picked out by linguistic classifications and by entities ‘in the world’ therefore call for a clearer ontological separation of tasks (Borgo et al., 1996; Guarino and Welty, 2004). Nevertheless, each of the possibilities for a ‘road’, for example, commits (rigidly) to a certain range of semantic properties for the discourse entity so described. Such discourse entities relate only indirectly to ontological properties of entities in the world, typically in an interactive process of mediation (cf. Bateman, 2004a; Bateman et al., 2007); the importance of ontological mediation is also argued from formal and philosophical perspectives by, for example, Arrighi and Ferrario (2005) and (Atencia and Schorlemmer, 2012).

This means that the selection of some linguistic category, such as ‘road’, cannot be taken as a neutral labelling of the world. There may be considerable uncertainty as to where the limits of applicability of a term might lie and to what follows from such an application once made; this problem has received considerable study in the area of geographic entities (Bennett, 2001; Smith and Mark, 2003). As will be illustrated below, this consideration is also a further reason why it is beneficial to maintain an explicit connection to the mechanisms of linguistic expression. ‘Road’ is a lexical item, not an ontological category, but how we construct grammaticized embeddings of this lexical item leads directly via their GUM modeling to constraints-in-use (such as dimensionality, etc.) for which it should be possible to find further ontological evidence. Precisely which grounding will vary depending on the GUM categories ascribed.

In summary, GUM is an ontology-like organization of categories and relations that has been designed to provide grammatically-responsive semantics for any linguistic expressions. Specifications employing GUM generalize away from specific grammatical forms, while still maintaining sufficient contact with those forms to support natural language analysis and generation more straightforwardly. One of the most important roles of GUM methodologically is consequently to help distinguish and isolate cases of linguistic ambiguity, i.e., underspecification and potential divergences in what any linguistic expression, including those used in more philosophical discussions, are committing to. The resulting GUM descriptions may then provide more secure points of attachment for formalization that are not linguistically bound, since these no longer need to concern themselves with linguistic variability.

In the examples analyzed below we shall see how this can be a valuable step in providing ontological descriptions of particular cases. The areas selected in this special issue are areas where there is already considerable agreement in the formal ontology community concerning just what the phenomena at issue are, but this cannot in general be guaranteed to be the case. Indeed, in many areas there is still extensive discussion concerning modeling choices. This is actually illustrated in the last example discussed, characterizing the nature of ‘marriage’: here modeling proposals vary quite dramatically. In contrast, the characterization of the examples from the GUM perspective proceeds independently of any such discussions: it is not necessary to make philosophical commitments since the motivations drawn on are subject to linguistic analysis rather than requiring contextualized interpretations to begin. Nonetheless, the strong alignment assumed between grammatical organization and semantic distinctions means that the GUM analysis will isolate certain entities and relations, with particular interrelationships, that appear to be intrinsic to the descriptions of the phenomena at issue. This is how these phenomena are construed at the human-level, particularly when engaged for linguistic expression.

Distinct proposals for axiomatizations of the phenomena described can then be inter-related via their mappings to the GUM linguistic semantics – for example, contrasting formalizations of time and events may be related by placing them in correspondence with GUM semantic specifications, since particular kinds of temporal relations and temporal profiles of events are made grammatically manifest. A similar strategy can be explored in terms of simulations (cf. Bergen, 2012) as we mention further below. In this sense, GUM might serve as a ‘meta’-ontology relating both diverse proposals for formalizations and natural language correspondences. The notion of relating ontological modules by mapping rather than direct inclusion is particularly relevant here. Such mappings are to be seen as structured relations between theories as illustrated for spatial representations in, for example, Bateman et al. (2007). An approach of this kind might even be considered as establishing a rather different notion of ‘ontology design patterns’ (cf. Gangemi, 2005; Janowicz et al., 2016) that is more aligned with natural language expressions. Since we can ‘talk about’ any domain, i.e., produce language describing that domain, the GUM organization supporting this is automatically guaranteed a high degree of genericity and re-usability, a re-usability that extends to drawing relations both across potential axiomatizations and between different perspectives on phenomena within single axiomatizations. Just as ontology design patterns are intended to be independent of the ontological commitments made by specific ontologies, GUM is similarly to be seen as offering a further level of abstract description that might be applicable to differing formalizations of particular areas of concern.

This relationship between GUM specifications and other areas of potential foundational ontologies helps characterize the GUM placement of the case study areas as follows. In each case, the distinctions only occur in GUM specifications to the extent to which they are grammaticized.

Continuant vs. occurrent: this distinction is generally construed linguistically by means of ‘verbal’ expressions and ‘nominal’ expressions. In GUM this broad grammatical reactance underwrites the basic Configurationvs.Element distinction. Configurations can have temporal attributes, Elements should not.

Independent entity vs. dependent entity: GUM maintains the grammatically motivated distinction between SimpleThing and SimpleQuality; all of the qualities can be considered dependent entities that must be brought together with their bearers in corresponding Configurations. For example, Color is a SenseANDMeasureQuality, which is in turn a subtype of MaterialWorldQuality. Subtypes of BeingAndHaving Configurations, such as PropertyAscription, then combine Elements with corresponding properties, such as Color. The distinction shows itself in grammatical reactancies such as the asymmetric behavior of “this table’s color” vs. “this color’s table”. Corpus studies can be used to evaluate the reliability of such probes more broadly.

Processes vs. events: GUM maintains the distinction between a Configuration, which is where participating elements, circumstantial information, and temporal and aspectual profiles come together, and the class Process, which takes a particular distinguished role within a Configuration. This is motivated by a broad range of grammatical reactancies which show properties that might be constructed in grammatical clauses and quite distinct properties that are limited to the lexical semantics of the verbs of those clauses. Processes mark out the bare types of activities that are lexically distinguished, such as ‘running’, ‘walking’, ‘ambling’, and so on. Configurations then combine these to form units such as ‘the man walked to the station’, which specifies participants, boundary conditions as well as deictic relations to the ‘here and now’ (cf. Galton, 2018).

Properties, qualities, quantities: properties and qualities of diverse kinds are distinguished grammatically in many languages. These different reactancies are used to populate the SimpleQuality branch of the GUM type hierarchy.

Functions, Roles: many functions and roles are treated in a broad range of languages as analogous to kinds of ‘possession’: and so are placed within the GeneralizedPossession branch of the GUM Configuration type hierarchy. These remain quite underspecified with respect to their semantic commitments as language is very flexible in their use; there are also overlaps with construals of parts. Some lexicalizations of the part relationship incorporate aspects of shape (‘slice’), others are more abstract (‘component’, ‘of’). There would be much to develop further here drawing both from empirical linguistic corpus studies and more explicit engagements with the philosophical literature (cf., e.g., Fine, 2010). Roles are also treated in a specific manner that will be illustrated further below.

For many of these distinctions, it would be possible to consider importing into the GUM definitions stronger sets of constraints taken from more foundational characterizations to the extent that these are considered sufficiently stable; prior to formalization candidate constructs may be maintained as ‘interpretation policies’ as noted above. In general, however, the GUM definitions err on the side of caution in setting constraints so as to avoid compromising the flexibility of the language use supported. In addition, there are certainly areas of ontological interest which appear never to be grammaticized. This is itself an interesting phenomenon worthy of more research and may perhaps be indicative of notions that are so unvarying as part of the human mesoscopic layer that no language would ‘consider’ their grammaticization necessary. Conversely, some distinctions of general ontological interest may turn out to be ‘expressible on demand’ rather than warranting being baked into semantic alternations: many of the distinct kinds of collections identified by Wood and Galton (2009) appear to be of this kind although, again, more refined corpus studies may yet reveal grammatical reactancies here as well.

3. The formalization of GUM and application methodology

The original incarnations of the Penman Upper Model received formalizations in several languages and knowledge representation systems, as standardized description logics were still some time away. Within the Penman project, the upper model was primarily represented in the Loom knowledge representation language (MacGregor, 1993), a representational form itself influenced by the Knowledge Interchange Format (KIF) used in several ontology projects and extended with automatic classification, role reification and further capabilities for reasoning. Currently, versions of GUM are maintained in OWL-DL. This representation is relatively straightforward given the similarly straightforward nature of the axiomatization pursued within GUM. The current version of GUM (including the spatial ontology module) lies within $ALCHQ$ (D). This provides a structural backbone based on the class-subclass relationship (the ‘signature’ of the theory) and the usual description logic possibilities for defining roles between classes and constraints on those roles’ fillers; approximations with more restricted DLs have also been explored. GUM’s signature therefore contains categories (unary predicates), also called classes or concepts, and relations (binary predicates); specific kinds of relationships are defined as relations between classes, and are organized themselves within a relation hierarchy. Many concepts that in other ontologies might be treated as two-place relational predicates are consequently treated in GUM indirectly as unary predicates combined with a distinguished set of ‘participant’ roles. As will be explained below, this follows in the spirit of the commonly adopted Neo-Davidsonian linguistic semantics and simplifies the specification of the compositional semantics of expressions considerably.

Since GUM is not intended to operate independently of other ontological descriptions for the areas it covers, it is presumed that in actual situations of use there will be links defined heterogeneously with further ontologies and reasoning components. GUM only contributes those aspects of a description of any situation, event or object that need to be made explicit when considering verbalization. Other aspects, relevant for other concerns, may best be captured in separate foundational ontologies. The design principles of GUM consider such modularizations to be critical for allowing each ontological module to focus on the classes of distinctions relevant for its aims. Merging potentially incompatible criteria for classification is generally to be avoided. In important respects, this approach is inherently committed to multi-perspectival ontologies and to heterogeneous ontologies in particular and avoids prejudgements concerning which perspectives may be considered ‘basic’ or ‘foundational’ as ideological positions. Within this design approach, each ontological component needs to be free to address its own class of requirements and problems; relating across such ontologies – as is now, for example, supported by the Distributed Ontology Language standard (DOL: Mossakowski et al., 2013) – is a distinct task. This is then the general challenge of mediation between ontological organizations that have been motivated by domain and application, on the one hand, and organizations motivated by the needs of expressing that information in natural language, on the other. These distinct purposes do not necessarily lead to organizational structures that align. In short, use of GUM is always to be seen in combination with further ontological (and other) components in heterogeneous environments of representation and reasoning (cf. Bateman and Space, 2013; Bateman et al., 2018).

Traditionally, the form of semantic expressions used when discussing GUM descriptions of particular cases has been the semantic specification of the original Penman system: Kasper’s (1989) Sentence Plan Language (SPL).1

¹
Many examples of SPL specifications are included in the linguistic resources available with the KPML natural language generation system (Bateman, 1997) as illustrations of use.

This notation also served as the template for the ‘abstract meaning representations’ (AMR) introduced by Langkilde and Knight (1998); AMR is a strictly less abstract notation that has been used to annotate the semantics of linguistic expressions in several corpora; it is related to various first-order logic fragments in Bos (2016). In addition, in the natural language analysis components discussed, for example, in Bateman et al. (2010), the representation for GUM specifications used is the Hybrid Logic Dependency Semantics modal logic defined for combinatorial categorial grammars by Baldridge and Kruijff (2002). Both of these are straightforwardly related to description logic formulations; a current version of the Generalized Upper Model expressed in OWL-DL is available at: http://purl.org/net/gum2. Nevertheless, for ease of comparison with the foundational ontology descriptions presented in the other contributions to this special issue, GUM specifications here will be given in a straightforward first-order logic notation.

Specifications in GUM are centered around ‘events’, or what in linguistic semantics are often termed eventualities: this is strongly motivated by grammatical concerns since events are the most natural correlates of grammatical clauses, which are themselves the units most transparently (i.e., iconically) linked to semantics. Many specifications according to GUM are consequently expressed as Neo-Davidsonian conjunctions of events and relations between those events and further entities (cf. Higginbotham, 1985; Parsons, 1990; Maienborn, 2011). Different types of events are distinguished primarily by the kinds of relations that they may enter into and the types of entities that are permitted to fill those relations. The broadest categories of eventualities used in GUM are the Configurations introduced above. This follows and takes further the Neo-Davidsonian move by allowing a far wider range of types of eventualities than the originally envisaged verbs of change or action. As a consequence, even connections that are commonly seen as two-place relations, such as ‘part’ or spatially being ‘on’ something, are treated in GUM as Configurations identified by unary predicates. Thus, for example, the GUM concept Part is a subtype of GeneralizedPossession, which is itself a subtype of the configuration BeingAndHaving. This means that being a part of something is an eventuality that unfolds in time and which has a particular temporal location, thereby covering statements such as ‘Britain was part of the Roman Empire in the first century’ or ‘Britain stopped being part of Europe’; similarly for all spatial relations. This will be seen in the discussion of the examples below.

Use of GUM to provide semantic specifications for any linguistically expressed descriptions generally begins by identifying the discourse referents that any linguistic description introduces and then adding constraints to those referents as motivated by the grammatical forms deployed. The constraints follow directly from the definitions of GUM categories and determine those semantic configurations that the specific linguistic expressions commit to. Since linguistic expressions are commonly incomplete and underspecified, assigning a GUM-description will also usually provide a fully explicit template of information that is technically ‘missing’ from the information actually present. Filling in such missing information is always a primary task in providing fully contextualized interpretations, which are themselves a prerequisite for further ontological analyses. Assigning GUM descriptions renders this often implicit process open to examination. When GUM is used in actual NLP systems, additional semantic or lexical features required for non-propositional aspects of the semantics (e.g., interaction, speech acts, information statuses and similar) are typically provided; these will not be considered further here although we will need to indicate some aspects in the formalizations offered. This will be shown with certain extra-logical notations that will be introduced as we proceed.

In order to illustrate the approach, and before proceeding to its application for the set example cases discussed below, we will consider the construction of a GUM specification for the simple sentence: “On Tuesday Mary gave the boy a book”. Several discourse entities are straightforwardly projected by this clause: the eventuality of Mary’s giving this specific book to the boy at that time is $e 0$ , Mary is $x 1$ , the boy is $x 2$ and the time interval Tuesday is $t 1$ . The first step is then to determine which category of Configuration best matches the grammatical properties of the clause (and viable alternations generated by grammatical probes associated with each configuration type). In the present case, this is a DispositiveMaterialAction, which involves an actor having some effect on an ‘actee’. As can be seen in Fig. 2, this category is a subtype of AffectingAction and DoingAndHappening. The actor and actee roles are filled straightforwardly by the SimpleThings for Mary and the book; the boy stands in a further similar relation of recipient. This gives the top level organization of the GUM representation as follows, illustrating the standard Neo-Davidsonian ‘unpacking’ of n-place relations into n conjunctions of specially distinguished binary relations: $\begin{array}{l} (1) & DispositiveMaterialAction (e 0) \land actor (e 0, x 0) \land actee (e 0, x 1) \land recipient (e 0, x 2) \end{array}$ The temporal information given in the clause is also straightforwardly translated into temporal circumstantial information given by the corresponding GUM relational categories thus: $\begin{array}{l} (2) & temporalLocating (e 0, t 1) \land OneOrTwoDTime (t 1) \end{array}$ The temporalLocating is the usual semantic projection of a grammatical temporal circumstance while the category OneOrTwoDTime is determined by the preposition used in the corresponding prepositional phrase (‘on’). Grammatical and lexical information also allows the semantic types of the arguments of the eventuality to be filled in employing the relevant GUM categories: $\begin{array}{l} (3) & Female (x 0) \land SimpleThing (x 1) \land Male (x 2) \end{array}$ It is worth noting briefly here that these discourse entities are introduced into the representation without quantifier scoping. As well known, natural language expressions commonly do not commit to particular quantifier scopes in their semantics, leaving this to pragmatic and contextual interpretation. The form of semantics used with GUM reflects this lack of commitment to specific quantifier scopes and is consequently in this respect closely related to the minimal recursion semantics now widely used in natural language processing for (shallow) semantic interpretation (cf. Copestake et al., 2005). Contextualization of scopes may need to make reference to arbitrarily detailed background knowledge and so is best abstracted away from a linguistic semantic specification. This is another of the important differences between fully contextualized interpretations and the commitments entered into by linguistic expressions.

Finally, there are several additional facets of the specification enforced by the GUM definition of Configurations and the temporal information given in the clause. The ‘lexical semantic’ identity of the eventuality is given by the required relation processInConfiguration (cf. Fig. 2); languages supply rich and complex hierarchies of such lexical material whose boundaries from one another are often language-specific and potentially idiosyncratic. In the present case, the process is one of ‘giving’ – this is seen as the bare lexical semantics, which commits to a particular collection of participants and, potentially, some pre-state and post-state constraints. This lexical semantics is abstract and corresponds to the general GUM concept of Process, which must be embedded into a Configuration in order to be associated with actual participants, etc. In addition, and as noted above, GUM specifies that Configurations are linked to a ‘temporal profile’ for determination of Aktionsart and similar commitments. In the present case, the ‘giving’ is anchored into a temporally bounded eventuality – i.e., the giving has an end point (and so combines with certain temporal information and not others). Additional temporal anchoring is provided by the tense of the verb: the simple past indicates that there is a temporal precedence relation between the temporal interval of the described eventuality and the ‘time of utterance’, which is necessarily present as a ‘deictic’ component of any developing discourse. More complex tenses give rise to correspondingly more complex sets of relations between temporal intervals (cf. Matthiessen, 1983); we omit the details here as they are not generally required for the example cases below. The temporal information independent of that necessary for tense is then:2

Since the current paper does not focus on lexical semantics, the precise internal contents of Processes, for example, will not be discussed. The GUM-specifications given here abbreviate these aspects by referring directly to names of lexical items, such as, ‘giving’. Several treatments are suggested in the linguistic literature, such as Pustejovsky (1991).

\begin{array}{l} (4) & processInConfiguration (e 0, p 0) \land p 0 = ‘giving’ \land \\ (5) & configurationTemporality (e 0, t p 0) \land TemporalProfile (t p 0) \land TemporallyBounded (t p 0) \end{array}

The GUM specification for the clause is then the straightforward conjunction of lines (1)–(5). This is taken to be the minimal necessary semantic commitment compatible with the linguistic expression; any further information added to this would move towards contextualization and interpretation. In each of the examples discussed below, a similar process will be employed working from grammatical analysis of the natural language characterizations of the situations to be modeled to a GUM specification corresponding to those characterizations. Particular consequences of the modeling will be discussed as relevant for each case.

4. Analysis and formalization in GUM: Four examples

The designated purpose of GUM of providing a level of linguistic semantics also serves a methodological role during the formalization of specific cases. In many respects, producing a formalization is pursued in an analogous fashion to the task of performing natural language analysis conforming to the guidelines that GUM provides. This means that one does not, at first, consider what the examples might ‘mean’ independently of their linguistic organization. On the contrary, their potential meanings are derived from that analysis. This is then an important contribution to avoiding arbitrary modeling decisions: in all cases, it is the grammatical patterns involved that serve to guide description.

The semantic characterizations resulting from this procedure then mediate between (primarily) linguistic representations of states of affairs, actions, plans and so on, on the one hand, and contextualized interpretations of their meanings on the other. Modularization of this kind is an effective way of avoiding, or at least placing reasonable bounds around, the task of general world modeling. It is necessary to consider only the degree of detail demanded to create the semantic specifications. The method that is pursued for addressing each of the examples in this section follows this strategy. In each case, the linguistic units expressing the situations to be considered are translated into semantic specifications making use of the GUM categories, thereby respecting explicit modeling strategies and decisions.

The basic assumption of the framework is then that the organization supporting such modeling is far from arbitrary: instead it reflects cultures’ implicit and historically-driven ‘theories’ of the human world. Descriptions unfold with respect to developing models of the situation just as the language expresses or reproduces those models. The function of a GUM description is to provide precisely the anchor points, or links, that demarcate where ontological descriptions at other levels of abstraction may hook into human-scale descriptions. Maintaining certain aspects of a complete semantic description outside of the GUM description opens up the door to further experimentation with a range of theoretical treatments of those areas: it is not necessary, for example, to commit to particular models of space, or time, or of properties (cf., e.g., Bateman, 2010b). What is motivated grammatically appears to be restricted to descriptions of temporal entities (which may be extended or not) with respect to which configurations may be situated and which may themselves be placed in particular orders that help determine how collections of Configurations may combine. By such means, the work of description is shared across components with different responsibilities. A GUM-description of a state of affairs then constitutes a triangulation with respect to what is imposed by (some) language and by contextual specifications. Both the modularity and functional purpose of GUM therefore play an important guiding role when pursuing formalizations of particular cases, as will now be illustrated.

4.1. Composition/constitution

“There is a four-legged table made of wood. Some time later, a leg of the table is replaced. Even later, the table is demolished so it ceases to exist although the wood is still there after the demolition.”

As noted above, modeling with respect to GUM is always event-based. Concretely, this means that we consider the example as a short mini-narrative where several things happen. We proceed by analyzing the narrative rather than by assuming that we know what the intended issues here are meant to be. This leads us to explicitly confront areas of potential ambiguity and to demarcate where precisely various sources of knowledge must come from. This in turn provides guidance for the modularities that need to be adopted when working towards an ontological description. We do not, therefore, consider the questions of composition and constitution and their behaviors over time that are intended here before doing the narrative analysis because the GUM-based treatment will lead us to these considerations in any case. This is one of the main methodological gains of the entire approach.

First, the specific mini-narrative at issue brings together six distinct Configurations, most expressed as clauses and one expressed twice, once as a clause and once as a nominalization (‘demolition’). For each configuration, we seek the most fitting Configuration subcategory as illustrated above; these subcategories are, as is often the case, straightforward to ascertain because of the broad range of known grammatical reactancies that can be employed as probes (Halliday and Matthiessen, 2013). The configurations are also related by several ‘argumentative’ or ‘rhetorical’ relations (temporal succession, causality, and concession), which are also part of a complete semantic specification but which cannot be discussed here. The main clause of the first sentence of the mini-narrative then receives the GUM specification: $\begin{array}{l} (6) & Existence (e 1) \land existent (e 1, o 1) \land SimpleThing (o 1) \end{array}$ Existence is a subtype of the configuration BeingAndHaving with a single participant role, existent for the entity asserted to exist. The temporal profiles are also straightforward (maximally unconstrained) and so will not be further discussed. The semantics of the SimpleThing playing the existent participant role of the table ( $o 1$ ) is then built as follows. The pre-modifying ‘four-legged’ asserts that the entity at hand is something that may have parts (and is thereby ‘decomposable’) and, in the case of ‘tables’, some of those parts are called ‘legs’ (l). One could consider iterating the parthood configuration four times giving four potential discourse entities for the legs, but we omit this and other possibilities here and simply constrain the cardinality of the set of leg-parts as a shorthand. The ‘made of’ expresses the ascription of a particular material composition (i.e., ‘wood’: w). In this usage, i.e., as the material constituting the table, the table is construed as being dependent on the wood (MaterialPropertyAscription). These considerations then add the following GUM specifications so as to complete the first description of the discourse referent for the table: $\begin{array}{l} (7) & DecomposableObject (o 1) \land SimpleThing (l) \land MaterialClassQuality (w) \land \\ (8) & MaterialPropertyAscription (r 1) \land domain (r 1, o 1) \land attribute (r 1, w) \land \\ (9) & Part (r 2) \land domain (r 2, o 1) \land attribute (r 2, l) \land | l | = 4 \end{array}$

The adoption of Part here for the relation between the table and the legs is motivated by several grammatical reactancies. Part is a particular subclass of the general GUM concept of GeneralizedPossession. GUM represents all notions of possession with this category. As mentioned above, ‘possession’ is an area studied broadly in linguistics for many languages and there are many subtleties beyond what can be discussed here. Languages commonly grammaticize the difference between alienable possession, which is a more contingent relationship, and inalienable possession, where the part identified is necessarily possessed – for example, a leg necessarily presupposes a body of which it is (conceptually) a part. This is manifested grammatically (in English) by the availability of expressions such as ‘table leg’. Different languages make their own cuts on which items may be considered inalienable and which not and so generalizations here concerning intrinsic properties need to be cautious that they are not simply reflecting the language patterns of the researchers involved. It is also well known that languages tend to be multiply ambiguous in their use of phrases referring to ‘parts’ and so it is important to look at the grammatical reactancies as a whole rather than simple occurrences of lexical items (cf. Winston et al., 1987).

There is a significant further aspect of the consideration of a table’s parts that has much broader implications, however. The fact that tables have parts (and so lend themselves to the corresponding use of grammatical constructions presupposing parthood) is itself a contingent social fact bound primarily to the lexical item (or family of lexical items) that help define what is to be considered a table and what not. An adequate description needs then to separate out the commitments made according to the social definition of table and the ontological characterization that holds of whatever social definition that is adopted. This relates to notions of artifacts and design. For current purposes, we will talk of ‘plans’, ‘blueprints’, or ‘semantic schemas’, which articulate sets of descriptions which establish the properties that are to hold of any entity. These blueprints are considered small narrative units in their own right and so their contents are also to be characterized in terms of the categories provided by GUM. Such ‘narratives’ correspond to what might in some approaches be considered ‘ontologies’ of the entities so defined (e.g., ‘tables’); Bateman (2019) presents more from this, essentially semiotic, perspective on ontological specifications and we will return to this particularly in the discussion of the final example case below.

A blueprint, or semantic schema, for tables might then state that tables have certain parts, such as legs. The GUM-semantics provides the means of saying that certain entities have certain parts, but does not pronounce on this further. It is therefore considered part of lexical knowledge concerning ‘tables’ that they stand in a generalized possession relationship with ‘legs’ and, in a different sense, ‘head’, but not, for example, with ‘bones’, ‘arms’, and so on. Blueprints in general also need to make use of the further possible circumstantial modification of Configurations offered by ModalPropertyAscriptions. These are typically expressed as ‘ought’, ‘must’, ‘should’, etc. Definitions of entity semantic schemas would typically be modalized in this way: i.e., “tables usually have four legs”, and so on. The use of semantic schemas as definitions of how socially constructed entities ‘should’ occur is therefore commonplace. Many lexical items then come to stand in this way for more or less complex descriptions that may themselves be expressed in the form of GUM specifications, but whose precise contents are matters of historical contingency rather than ontology.

Finally for this first sentence, since all of the information additional to the existential is expressed without clauses, their corresponding temporal profiles and anchoring on the timeline are inherited from that of the existence configuration $e 1$ . This means that there is no requirement that the discourse entity of the table stays four-legged or remains made of wood throughout its life. Characterizing these property and part ascriptions as Configurations makes it clear that these have all the usual possibilities of Configurations, including temporal profiles and stages—as in: “the table stopped being 4-legged”, “the table that used to be made of wood”, “the table that will have three legs”, etc. The next sentence of the mini-narrative then begins to modify the situation in precisely these directions.

The corresponding eventuality of the following sentence is described as being straightforwardly anchored into the timeline at some point following the time of reference of the first sentence. ‘Replace’ instantiates a further subcategory of Configuration making use of a combination of participants and circumstances involving actor, actee and exclusive (a subclass of accompaniment). This corresponds to the lexicogrammatical frame or construction ‘X replaces Y with Z’. Although in the sentence as given both the actor and the replacement entity are left unspecified, they are nevertheless implicitly present by virtue of the lexical semantics associated with ‘replace’. We indicate this lack of knowledge, or underspecification, in the specification with variables marked with question marks ( $? x 0$ , $? x 1$ ). These are discourse referents which are necessarily present due, principally, to semantic gaps – i.e., elements which are necessarily present in the semantic specification but which do not appear in the surface form. ‘Replacing’ commits to an exchange but the sentence does not say with what (but could have). This yields the GUM-description: $\begin{array}{l} (10) & DispositiveMaterialAction (e 2) \land processInConfiguration (e 2, p 1) \land p 1 = ‘replacing’ \land \\ (11) & actor (e 2, ? x 0) \land actee (e 2, l^{'}) \land exclusive (e 2, ? x 1) \land \\ (12) & PartOf (r 3) \land domain (r 3, ? x 1) \land attribute (r 3, \underline{o 1}) \end{array}$ The notation $l^{'}$ is used here as an ad hoc shorthand for selecting one member of the identified set (i.e., one of the legs); in contrast to the unspecified $? x 1$ , this discourse referent is present in surface form and so receives an explicit non-anonymous label. Linguistic quantification often serves this kind of selection role in discourse, restricting the current topic under discussion to members of previously introduced sets; this will not be discussed further here. In addition, since this is a new sentence, we cannot assume that the discourse entities identified are simply the same as those seen previously. That is, the connection of the discourse entity $o 1$ with the previous sentence of the mini-text is a matter of discourse pragmatics. Each configuration therefore either introduces or re-uses specific discourse entities which serve as anchor points into situated grounding and help clauses bind together into a coherent text. Such discourse coherence is managed formally by mechanisms such as those set out in segmented discourse representation theory by Asher and Lascarides (2003). This aspect of meaning composition will also not be discussed further here, but must be assumed in order to hold the various sentences of the mini-texts describing the example together. The binding of variables to previously introduced discourse entities will be indicated in the GUM specifications by an extra-logical notation of underlining.3

³
Formally, this indicates that the set of expressions in which such underlined entities occur are placed within a larger discourse structure tree such that the antecedent of the underlined discourse referent is accessible. Variables with generally be numbered successively within each separate example to aid the discussion. All such variables are consequently distinct discourse entities.

Finally, the category PartOf is an inverse of Part explicitly defined in GUM to capture the reciprocal relationship involved; it is therefore, again, a subtype of Configuration with the order of the participants simply reversed.4

⁴

This ‘reversal’ is actually managed at the level of the ‘participant’ roles, domain and attribute. The notion of inverses in the GeneralizedPossession branch was originally introduced to simplify the mapping between Upper Model specifications and linguistic expressions during natural language generation. It might now be a candidate for removal.

The overall meaning of the sentence is then gained from a unification of the specific lexical semantics of the verb ‘replace’ and the semantic frame given by the GUM-description. For the lexical semantic component there are several possibilities discussed in the literature but again, grammatical reactancies lead to a minimal commitment that is necessary regardless of the specific semantics taken. In particular, it must be specified that there is some ‘pre-state’ allowing elements to be picked out, and a ‘post-state’ in which one or more of the elements picked out are no longer present. The importance of having semantic access to internal event structures of this kind is shown in grammatical reactancies such as the ‘imperfective paradox’ discussed by Pustejovsky (1991) and others:

John is running. $⟹$ John has run.

John is building a house. $⇏$ John has built a house.

Although both configurations are ongoing, they are viewed ‘from within’ their unfolding and, at that point, the second case has not yet reached its culmination. Semantic classes of verbs can then be characterized as belonging to one of three basic event types: states, processes and transitions according to their pre- and post-state behaviors (Pustejovsky, 1991, 56). Using these semantic descriptions to construct a model of the example situation concerning the table then gives us two intervals corresponding to the two ‘part-of’ Configurations introduced. The first (

r 2

) is where the table is as would be expected, and the second (

r 3

) is where the table has received a new leg (

? x 1

). Note that in terms of the surface linguistic form all we know is that one of the legs of the table has been replaced with some other entity functioning as a leg, but in the post-state we do not know with what precisely. Pragmatically we might assume that the new entity has many of the properties of the old entity: what those properties are can only be drawn from the blueprint for tables however.

In the next sentence, the table is ‘demolished’, which is also a straightforward case of the Configuration subcategory DoingAndHappening. And, here again, it is necessary to draw on the specific lexical semantics of this verb in order to narrow down the state of affairs that holds following the event. Pustejovsky (2000, 468) sets out several alternatives for the preservation or removal of states of affairs within events that would be relevant for further lexical semantic specifications. Nevertheless, specifying that semantics needs at least to capture the fact that when the overall entity that is demolished no longer exists, its parts no longer exist as well. This is an interpretation of the intended meaning of the example sentence. The selection of a specific reading is what enables an ontological specification to be offered. The current example case was selected precisely to offer an opportunity for ontologies to state how they manage the relation beween entities and those entities’ constituting materiality but relies on the example being interpreted in this way. The GUM treatment makes this explicit and so fixes the terms of discussion.

Different verb choices would lead to quite different characterizations. ‘Demolish’ may, for example, be distinguished from the verb ‘dismantle’, which may well leave the parts intact. In both these cases, GUM says nothing about what happens to the material associated with the entity under discussion. This is in general correct – what is left following the demolition depends entirely on the details of the demolition process. The lexical semantics does not involve any change of the material and so no modification of the material quality attributed to the object is entailed: this is the general approach taken to successive discourse construction, descriptions continue to ‘hold’ until they are retracted. But the situation can be quite different with a different class of verbs: for example, the verb ‘burn’ would typically not leave the material unaffected. This then requires attention to the status of the wood as a dependent entity with respect to the table when combined within a MaterialPropertyAscription (from line (8)): here the bearer of the material property of ‘being made of wood’ should not outlast the wood out of which it is made. The appropriate post-state then varies depending on the extensiveness of the operation performed on the object: dismantling retracts the assertion that the object exists but appears to leave the parts intact; demolishing may retract both the existence of the object and its parts; burning retracts existence of the object, parts and material. Ontologically we would want the removal of parts not (necessarily) to remove the object, as a (temporarily) three-legged table might still be a table; the removal of the material, however, should allow neither the parts nor the object to remain. GUM then provides an overall skeleton for what is occurring appropriate for the specific lexical semantics applied. Any proposed further ontological characterization should be mappable to the entities and relations derived from the GUM characterization, including in particular the contrasting pre- and post-state descriptions and the differences between parts and materials: these are clearly distinguished by their differing grammatical reactancies and so need to be separated in any further formalization as well. The corresponding GUM-categories then act as indices into further ontological accounts, offering a point of connection across the various levels of abstraction involved (cf., e.g., Hois and Kutz, 2008; Pomarlan and Bateman, 2020).

4.2. Roles

“Mr. Potter is the teacher of class 2C at Shapism School and resigns at the beginning of the spring break. After the spring break, Mrs. Bumblebee replaces Mr. Potter as the teacher of 2C. Also, student Mary left the class at the beginning of the break and a new student, John, joins in when the break ends.”

This case study example is intended to focus on how roles, the fillers of roles and organizations interact under change. Again, we will see how the GUM analysis leads to these issues directly. As before, the sentences are analyzed one after the other as if an automatic semantic analysis of the narrative was being performed, thereby anchoring the terms of discussion. The first clause introduces several discourse referents as SimpleThings participating in a configuration. Moreover, even more than was the case with the ‘table’ above, the notion of a ‘school’ and what schools consist of, including the institutional roles that they define, is a contingent definition of a social artifact. It is only with the presupposition of such a semantic schema that particular roles such as ‘teacher’ and a ‘pupil’ make sense. Such roles, regardless of which particular roles any particular schema may define, are distinguished by grammatical alternations and so need to be covered by a corresponding GUM category – in this case, this is the category: RoleOrOffice, a subtype of the GUM concepts Abstraction and NonConciousThing (notwithstanding the fact that one can achieve a reference to a person, i.e,. a ConsciousThing by referring to a role they are playing). There are many such linguistic strategies and, again, these should not be allowed to have undue influence on the semantic and ontological modeling.

Concepts classified as RoleOrOffices are generally the labels of generalized participants within the abstract narrative, i.e., the blueprint or semantic schema, that provides for their recognition. Roles such as ‘the president’, ‘passenger’, ‘teacher’ and so on are consequently grammatical nominalizations, typically of the actors of particular kinds of configurations within the relevant semantic scheme: i.e., ‘teachers’ are the actors of configurations of teachings within the context of a description of schools and their functions. This reflects their relational and definitional dependencies (cf., e.g., Masolo et al., 2005). The nominal phrases expressing such roles are subject to their own sets of grammatical reactancies, which, on the one hand, overlap substantially with those of the fillers of the actor relations and, on the other, allow possessive connections with respect to the framework within which the role exists (e.g., ‘the country’s president’, ‘the school’s teachers’, and so on). In short, the RoleOrOffice concept identifies ‘named’ nominalizations of roles from some definitional blueprint, allowing them to maintain rigidity despite their bearers changing. The connection between roles and their bearers is then modeled with the GUM configuration RolePlaying, a particular subtype of the configuration BeingAndHaving subject to all of the usual temporal contingencies of configurations of that type. Nominalizations of the domain participant role of this configuration (i.e., the bearers) then introduce discourse entities that function in a manner quite similar to qua-individuals (Masolo et al., 2005).

The discourse entities introduced directly by the current example following the principles illustrated previously are: Mr. Potter: $o 1$ , the teacher: $o 2$ , class 2C: $o 3$ , and Shapism School: $o 4$ . A corresponding GUM specification for the first part of the sentence, “Mr. Potter is the teacher of class 2C at Shapism School”, is then: $\begin{array}{l} (13) & RolePlaying (e 1) \land domain (e 1, o 1) \land Male (o 1) \land attribute (e 1, o 2) \land RoleOrOffice (o 2) \land \\ GeneralizedPossessionInverse (r 1) \land domain (r 1, o 2) \\ (14) & \land attribute (r 1, o 3) \land SimpleThing (o 3) \land \\ (15) & SpatialLocation (r 2) \land domain (r 2, o 3) \land attribute (r 2, o 4) \land GeneralizedLocation (o 4) \end{array}$ The discourse referent for ‘the teacher’ is classified as a RoleOrOffice following a standard linguistic construction for attributing roles to fillers. The relation between such roles and the organizational units to which they belong is again a kind of generalized possession: that is, the teacher ‘belongs’ to the specific class mentioned. That class is itself connected with a (generalized) ‘location’, i.e., the school. The discourse entity introduced by the participant-nominalization of the RolePlaying, i.e., the occurrence of $o 1$ as the domain of $e 1$ , would be glossed in natural language as something like ‘Potter-as-teacher-of-class-2C-of-Shapism-School’. This may be formalized further in several ways; an approach using lambda-expressions will be discussed for the next case study example (3.a). Discourse entities built on Elements in this way generally inherit all the usual possibilities for counting, quantification, modification, and so on that the grammar of nominal phrases offers.

The second part of the sentence is then a simple Configuration of ‘doing’ (‘resigning’ as the process of $e 2$ ) without any other affected entity, and with the actor of the configuration or NonAffectingDoing identified discoursally as the teacher ( $o 1$ ). $\begin{array}{l} (16) & NonAffectingDoing (e 2) \land actor (e 2, \underline{o 1}) \land \\ (17) & temporalLocating (r 2) \land domain (r 2, e 2) \land range (r 2, t 1) \land TimeInterval (t 1) \end{array}$ The grammatical semantics of NonAffectingDoing in general permits a further ‘actee’ of a particular, non-affected kind. The lexical semantics of ‘resign’ can make use of this possibility as in ‘resigned her job’ or ‘resigned the post’, where the potential fillers for the relation are typically restricted to include only roles and offices. An alternative grammaticized form for such semantics is where the position resigned from is expressed as the attribute of a RolePlaying circumstantial, as in ‘she resigned as president’, or ‘he resigned as teacher of the class’. This is then an alternative for marking the end point of a RolePlaying configuration directly. Consequently, the pre-state for ‘resign’ is where the actor has a role and the post-state is where that actor no longer has that role, i.e., in the present case, that $e 1$ (from line (13)) holds for some interval that is bounded by the time $t 1$ (‘beginning of the spring break’) given for $e 2$ . The existence of the class and the school are unaffected by this, although more complex ‘blueprint’ narratives can certainly construct dependencies between role-bearers and organizations that might conditionalize the existence of collectives on the existence of dependent role-bearers (cf. Wood and Galton, 2009) – these do not appear to be grammaticized, however.5

⁵
Certain uniqueness conditions are grammaticized, such as ‘the teacher of the class’ vs. ‘a student of the class’, but these follow more generally from identifiability for all descriptions, not only role configurations.

When Mrs Bumblebee ( $o 5$ ) comes on the scene after the spring break, we have again a case of ‘replacement’ as introduced above (line (10)), identified here as $e 3$ . As before, the pre-state requires that some element in a larger collection be identifiable and the post-state has that element swapped for another. Here the clause explicitly uses an ‘as’-circumstantial and so the semantics is straightforward: $\begin{array}{l} (18) & DispositiveMaterialAction (e 3) \land actor (e 3, o 5) \land Female (o 5) \land actee (e 3, \underline{o 1}) \land \\ (19) & RolePlaying (r 3) \land domain (r 3, e 3) \land range (r 3, \underline{o 2}) \land \\ (20) & temporalLocating (r 4) \land domain (r 4, e 3) \land range (r 4, \underline{t 1} + 1) \end{array}$ That is, the newcomer ( $o 5$ ) takes on the discoursally identified role of being a specific teacher ( $o 2$ ) at a time that is also discoursally anchored by virtue of the contrasting ‘at the beginning of the spring break’ and ‘after the spring break’. The temporal precedence relationships are straightforward and could be unpacked using the corresponding GUM classes but we omit this here, using ‘ $t 1 + 1$ ’ as a shorthand notation for some time strictly subsequent to $t 1$ . As indicated above, we have grammatical evidence that there is only one teacher for a class here and that is the role that the newcomer now occupies.

The final sentence offers a good further example of the value of adopting a two-level semantics where the linguistic commitments are not immediately cashed out in referential semantics. The verb ‘leave’ is a straightforward spatial Configuration which, in the GUM treatment of space, corresponds to a GeneralizedLocation with the school class as a reference object and a SpatialModality of moving outside of the reference object. As suggested above and argued in Bateman et al. (2007), the contextualization of this abstract spatial construal can play out in many ways depending on the discourse context involved. In the present case, since the topic of discussion is roles and their fillers, this can also be ‘spatialized’. Being a pupil is then ‘in’ the role, not being a pupil is ‘out’ of the role. Note that this metaphorical usage is certainly the one intended here: the issue is not that any of the pupils physically left the (room of the) class or not. The widespread use of such metaphorical constructions is well documented in a variety of studies of interaction (cf., e.g., Müller and Cienki, 2009) and needs to be seen as a general mechanism for constructing discourse coherence. Much language use is reliant on the explicit bringing together of multiperspectival descriptions of this kind: on the one hand, we have a metaphorical playing out of changes of (metaphorical) location; on the other hand, we have a taking up and laying down of roles. In general, insisting on an ontological treatment in terms of space would then be false – more important is to show how descriptions can invoke such re-mappings (or blends: Kutz et al., 2015) ‘at will’.

Thus, in addition to a simple spatial contextualization, unpacking the metaphor to refer to a change in role status is equally (in fact, given the discourse context, more) likely. With the role playing interpretation foregrounded, the termination of the role relationship can be encoded directly in GUM making use of the temporal profiles defined for all the Configurations involved. With the discourse referents Mary ( $o 6$ ), the role of being a pupil of the class ( $o 7$ ), and other variables either bound by discourse pragmatics as indicated by underlining ( $o 3$ : ‘class 2C’, $t 1$ : the time of Mr Potter’s resignation) or new, this gives the following GUM specification for the first half of the final sentence of the example: $\begin{array}{l} (21) & RolePlaying (e 4) \land domain (e 4, o 6) \land Female (o 6) \land attribute (e 4, o 7) \land RoleOrOffice (o 7) \land \\ (22) & GeneralizedPossessionInverse (r 5) \land domain (r 5, o 7) \land range (r 5, \underline{o 3}) \land \\ (23) & configurationStage (r 6) \land domain (r 6, e 4) \land range (r 6, i) \land ConfigurationEnding (i) \land \\ (24) & temporalLocating (r 7) \land domain (r 7, e 4) \land range (r 7, \underline{t 1} + 2) \end{array}$ The second student joining the class is expressed in a parallel manner, with the staging of the configuration specifying that the temporal extent of the role-playing event is ‘beginning’ rather than ending. – that is, when ConfigurationEnding holds, the configuration holds prior to the time interval of the configuration but does not hold following that time interval; conversely for ConfigurationBeginning. These can be seen as instructions for constructing and combining descriptions of the temporal intervals involved. As characterized above, the fact that students might come and go without one ‘replacing’, or taking the place, of another is only determinable by the ‘blueprint’ description of the role structure of classes. In terms of the GUM modeling, there is no uniqueness associated with either roles or the fillers of roles, although again it is possible that more detailed corpus studies might bring refinements for this view. The student consequently introduces a further instance of a student RoleOrOffice.

4.3. Property change

“A flower is red in the summer. As time passes, the color changes. In autumn the flower is brown.”

GUM models attributions of properties as a further subclass of Configurations, similarly to the attribution of the material to the table above. The difference lies in the kind of quality that the attribution picks out. Wood received no mention as a specific kind of material property because descriptions of entities being constructed from wood do not appear to have any consequences for grammaticization; however, color is very different in that it does have grammatical reactancies that show that many grammars of languages distinguish color as a particular category of experience in its own right. For this reason, there are explicit GUM categories concerned with color. We proceed with the analysis of the mini-narrative of case example (3.a) then as before. All the descriptions constructed in the example are simple events or states anchored into the timeline via their specified temporal positions. As usual, there is no requirement or expectation that properties remain constant over time.

The GUM representation of the first color attribution ( $e 1$ ) of the color ( $a 1$ : ‘red’) to the flower ( $o 1$ ) is as follows: $\begin{array}{l} (25) & ColorPropertyAscription (e 1) \land domain (e 1, o 1) \land attribute (e 1, a 1) \land \\ (26) & SimpleThing (o 1) \land Color (a 1) \land \\ (27) & temporalLocating (r 1) \land domain (r 1, e 1) \land range (r 1, t 1) \land TimeInterval (t 1) \end{array}$ The GUM category Color is a subclass of SenseAndMeasureQuality, which is itself a subclass of MaterialWorldQuality, SimpleQuality and Element. By virtue of being a SimpleQuality, color is considered a dependent entity. Finally, the temporal interval $t 1$ is constructed with the temporal circumstantial information ‘in the summer’ and linked as usual with the relevant configuration via the temporalLocating predicate.

In the second sentence, the nominal phrase ‘the color’ picks out a referent by the type of property attributed. As the phrase is interpreted in context, it can be seen as an abbreviated form of ‘the color of the flower’; this is a pragmatic interpretation and so is a defeasible inference, not a logical entailment. Nevertheless, it is crucial to take this extended semantics into account when producing the semantic analysis corresponding to its use. The nominal phrase acts analogously to a nominalization of the color ascription focusing on the attribute role. This means that it is not talking of any one color that the flower may have, but the filler of the attribute role of the configuration of color ascription whatever that might be. Such constructions occur frequently with descriptions generated from linguistic expressions and have received a variety of treatments in the literature; this is also similar to the role-player nominalizations of the previous example. In the combinatorial categorial grammar treatment of Bateman et al. (2010), ‘unsaturated’ semantic descriptions are taken as lambda expressions which are combined compositionally by function application during parsing (cf. Steedman, 2000). In the present case, therefore, ‘the color of the flower’ may be considered as being equivalent to a lambda expression for ‘x’ with the following GUM specification as the body of the expression: $\begin{array}{l} (28) & ColorPropertyAscription (e x) \land domain (e x, \underline{o 1}) \land attribute (e x, x) \end{array}$ This treatment of nominalizations is always available and corresponds semantically to the distinction between talking of a specific filler of a role and fillers of the role in general.

The second sentence then says that this attribute changes. ‘Change’ is straightforwardly classified as a configuration of type NonAffectingHappening, which indicates that some entity nonvolitionally undergoes something. The overall GUM specification is consequently: $\begin{array}{l} (29) & NonAffectingHappening (e 2) \land actor (e 2, a 2) \land \\ (30) & configurationTemporality (r 2) \land domain (r 2, e 2) \land range (r 2, t 2) \land \\ (31) & DurationInTime (t 2) \land TemporallyUnbounded (t 2) \end{array}$ where $a 2$ is restricted according to the description of the nominalized form ‘the color’ in line (28) above and the temporal interval $t 2$ is again as identified in the sentence; the discourse interpretation of this interval is that it succeeds the interval of the first sentence. The change in attribute follows from the lexical semantics of the verb ‘change’, which must state straightforwardly that some entity designated by its ‘actor’ (in this case, the filler of the designated color ascription) is different in the post-state to the pre-state. ColorPropertyAscriptions, as Configurations, always have the opportunity to be anchored to time intervals and this applies to the body of the nominalization as well. The pre- and post-states of ‘change’ provide such time intervals while the lexical semantics insists that whatever fills the attribute role of the color ascription needs to be different at those time points. In addition, the temporal specification employed in this sentence sets up a semantics where change is continual and unbounded; this might then receive a variety of treatments depending on how time is quantified externally to GUM although the GUM specification remains unchanged. What is fixed are the two endpoints of the quantum of change: the value at the beginning of the interval considered (which we could designate anonymously as $? a_{b}$ ) and the value at the end of the interval ( $? a_{f}$ ). Here also we could consider a metaphorical construction in terms of space. This would be constructed directly if the linguistic expression were, for example, “the color changes from red to brown” because this deploys the grammar of spatial descriptions explicitly and so would be represented in GUM as a corresponding path expression. Such treatments of colors as positions in a space are also naturally related to treatments of quality spaces as proposed by Gärdenfors (2000).

In the final sentence, the GUM specification is very similar to that of the first sentence, but with a different explicit color given (‘brown’: $a 3$ ): $\begin{array}{l} (32) & ColorPropertyAscription (e 3) \land domain (e 3, \underline{o 1}) \land attribute (e 3, a 3) \land Color (a 3) \land \\ (33) & temporalLocating (r 3) \land domain (r 3, e 3) \land range (r 3, t 3) \land TimeInterval (t 3) \end{array}$ With the temporal development of the narrative, the eventualities of the contributing temporal intervals are combined in a succession. This anchors the final value of the color attribute that remained open from the former sentence. Taken as a single unfolding change, we can see a traversal of color ‘values’ along a path anchored at the outset at ‘red’ and at the end at ‘brown’. This is quite analogous to the separation followed in DOLCE whereby the meaning of ‘color’ is maintained as time-independent ‘values’ ( $a 1 = ? a_{b}$ , $? a_{f} = a 3$ ) that are then linked in a time-dependent manner ( $e 1$ , $e 3$ ) to objects.

“A man is walking when suddenly he starts walking faster and then breaks into a run.”

This example is also focused on change and turns out, in many respects, to be quite similar to the previous example. The GUM modeling focuses again on the temporal profiles of the Configurations constituting the mini-narrative. In the first sentence, an activity ‘walking’ is embedded into a Configuration that is temporally unbounded. The GUM specification, with

e 1

as an eventuality with process ‘walking’ and

o 1

as the man, is:

\begin{array}{l} (34) & NonAffectingDoing (e 1) \land actor (e 1, o 1) \land Male (o 1) \land \\ (35) & configurationTemporality (r 1) \land domain (r 1, e 1) \land range (r 1, t 1) \land \\ (36) & DurationInTime (t 1) \land TemporallyUnbounded (t 1) \end{array}

As noted above, Configurations along the DoingAndHappening branch of GUM, such as NonAffectingDoing, are necessarily concerned with temporal unfolding. This information, although not explicitly mentioned in the shallow natural language expression, can always be made present according to the GUM definition of possible ‘circumstantial’ relations. When information is not explicitly given, it is a default discourse assumption that values are assigned as ‘normal’ for whatever the activity involved is.

The second sentence makes explicit just what component of the range of possible additional circumstantial relations is relevant for the present case. The adverbial ‘faster’, as a comparative for ‘speed’, introduces a discourse referent related to a SenseAndMeasureQuality. This is combined into the embedding Configuration by means of the designated GUM relation manner. The phrase ‘suddenly’ anchors the temporal interval of this second Configuration into the time line as immediately following the interval of the first Configuration. The ‘start’ orients to the beginning of the configuration’s temporal unfolding but is otherwise irrelevant for the specification.6

⁶
Phase terms (‘start’, ‘finish’, ‘continue’) have distinctive grammaticalizations; some languages, e.g., Chinese, build these into complex components of their aspect systems. It would also be interesting, therefore, to contrast expressions involving ‘start’, ‘breaking into’ and so on in languages that make different temporal commitments; a consideration of, for example, the Chinese aspect system within a similar framework is given in Yang and Bateman (2002).

Taken together, the GUM specification for the second sentence is then:

\begin{array}{l} NonAffectingDoing (e 2) \land actor (e 2, \underline{o 1}) \land manner (e 2, a 2) \\ (37) & \land SenseAndMeasureQuality (a 2) \land \\ (38) & configurationTemporality (r 2) \land domain (r 2, e 2) \land range (r 2, t 2) \land \\ (39) & DurationInTime (t 2) \land TemporallyUnbounded (t 2) \land \\ (40) & configurationStage (r 3) \land domain (r 3, e 2) \land range (r 3, t 3) \land ConfigurationBeginning (t 3) \end{array}

The grammatical form of a comparative entails semantically that a ‘standard’ can be ascertained with respect to which the comparison can be made. This is commonly left implicit and so must be filled in by discourse pragmatics: i.e., ‘faster’ than what? Discourse pragmatics must then draw the evidently intended connection with the speed of the previous event

e 1

as the only speed available in the immediate discourse context. One could, however, just as easily make the connection explicit with phrasing such as “The man walked at a normal walking speed. Suddenly he walked faster.” The discourse connection of these references is relatively straightforward, however, and so will not be discussed further here.

Finally, in the third sentence, ‘breaking into’ expresses an intensified variant of ‘start’ with more or less the same semantic description. What is then different is that the classification of the configuration also changes. Note, however, that the fact that ‘walking faster’ can become ‘running’ is a matter for the lexical semantics of the respective categories. This might be represented in many ways – for example, by capturing the commonality between ‘walking’ and ‘running’ in terms of a mode of locomotion using legs and feet with a (socially defined) boundary depending on speed (or, more conventionalized, whether at least one foot is on the ground at all times, etc.). The precise conditions for selecting ‘walking’ or ‘running’ are then, again, just as with tables and their ‘legs’, or with classes and their teachers, subject to the social definitions given in a corresponding semantic schema for types of locomotion. What the three configurations together describe is a kind of activity (foot-based locomotion) whose rate of unfolding increases over the three time periods constructed. One might express this in various ways, such as, for example: “the man moved faster and faster” or “the man started moving slowly, then moved faster, and then faster still”. The fact that different ‘modes’ of locomotion, or gaits, correspond to different solutions to combined body-part movements in a system (cf., e.g., Granatosky et al., 2018) is considered the responsibility of a strictly separate module, whose transitions may, or may not, correlate with the socio-cultural linguistic labeling.

4.4. Event change

“A man is walking to the station, but before he gets there, he turns around and goes home.”

In this example, there is more explicit use of spatial semantics and, in particular, the combination of spatial and lexical verb semantics to determine whether events are bounded, continuous, completed, and so on. Pustejovsky (1991) gives examples of such phenomena with the sentences:

Mary walked.

Mary walked to the store.

Mary walked for 30 minutes.

These all have different temporal ‘profiles’ and so combine differently with further grammatical constructions. In GUM terms, the first is a straightforward unbounded Configuration (more specifically, a NonAffectingSpatialDoing) without further restrictions beyond the past tense. The second and third are also NonAffectingSpatialDoings but, in contrast, are ‘bounded’, i.e., they have an endpoint beyond which the situation described no longer (necessarily) applies. The two latter eventualities are, however, bounded in different ways: the former by the destination of the motion event, the latter by a temporal extent. The second example is characterized according to Vendler’s (1957) well-known categories as an ‘accomplishment’; the latter example Pustejovky describes as ‘bounded’. The precise classification for a given case varies according to the combination of grammatical phrases involved: for this reason, it is clear that these properties are not lexical properties of verbs but are grammatically constructed (Moens and Steedman, 1988). This helps show why it is beneficial to separate out eventualities, i.e., Configurations, from the processes they build on.

Without further information, therefore, the use of ‘walk’ in the present tense in the first part of the sentence in example (4) would be unbounded and continuous. However, since it is modified further by the spatial location ‘to the station’, the eventuality described becomes an accomplishment. The closing boundary of the eventuality is defined by the actor of the eventuality being in or at the location given. GUM models this as a particular SpatialModality (corresponding to ‘being located at’) relating something that is located and a reference object; the precise treatment is given in Bateman et al. (2010). In the present case, the relevant SpatialModality introduces a motion qualitatively and functionally ‘towards’ and ending at the given reference object. Further semantic information relevant for this may be filled in by a broad range of potential models, such as, for example, any logically specified spatial calculus containing the appropriate predicates (Eschenbach et al., 2000) or an appropriately specified simulation (Bateman et al., 2019). The GUM-semantics for the first part of the sentence is consequently of the form: $\begin{array}{l} (41) & NonAffectingSpatialDoing (e 1) \land actor (e 1, o 1) \land Male (o 1) \land \\ (42) & destination (r 1) \land domain (r 1, e 1) \land range (r 1, o 2) \land GeneralizedLocation (o 2) \land \\ (43) & configurationTemporality (r 2) \land domain (r 2, e 1) \land range (r 2, t 1) \land \\ (44) & DurationInTime (t 1) \land TemporallyBounded (t 1) \end{array}$ The semantics does not commit to the destination ( $o 2$ : ‘the station’) actually being reached; all that is specified is the condition that would need to be fulfilled for the event to have been achieved. In addition, the present progressive tense explicitly places the ‘speaking time’ (cf. Matthiessen, 1996) within the temporal interval $t 1$ of the Configuration of walking ( $e 1$ ): that is, at the time of speaking or uttering this sentence, an eventuality is ongoing that does have a specified end-condition given by the specified destination but that end-condition has not been reached.

The expression ‘before he gets there’ in the second part of the sentence builds a temporal relation between the descriptions of the times associated with the two configurations. More specifically, it may be interpreted as constructing a description of a time within the time interval over which $e 1$ extends prior to its ‘right’ boundary. The expression ‘he gets there’ is discourse pragmatically anchored to reaching the previously specified destination (‘there’: $o 2$ ). The time with ‘before’ straightforwardly picks out some time within $t 1$ and, at that time, a further non-affecting spatial action configuration occurs that changes the orientation of the actor ( $e 2$ : ‘turn’), followed by a motion ( $e 3$ ) of that actor to a new destination ( $o 3$ : ‘home’): $\begin{array}{l} (45) & NonAffectingOrientationChange (e 2) \land actor (e 2, \underline{o 1}) \land \\ (46) & NonAffectingSpatialDoing (e 3) \land actor (e 3, \underline{o 1}) \land \\ (47) & destination (r 3) \land domain (r 3, e 3) \land range (r 3, o 3) \land GeneralizedLocation (o 3) \land \\ (48) & configurationTemporality (r 4) \land domain (r 4, e 3) \land range (r 4, t 2) \land \\ (49) & DurationInTime (t 2) \land TemporallyBounded (t 2) \end{array}$ The ‘before’ phrase as well as the ‘and’ between the final two clauses additionally help determine the temporal succession relationships between the eventualities. These together with the tenses selected by the clauses position the ‘speaking’ time of the final event after the culmination of $e 3$ , resulting in a post-state for the situation where the man should indeed end up at home if the eventuality is successfully completed. The modeling constructed thus clearly shows the difference between traveling towards a destination, reaching that destination, and defining an activity in terms of being oriented towards a destination. The fact that times can be constructed to refer freely to segments within each of the temporal intervals created straightforwardly allows for events to remain incomplete. The use of particular descriptions of the unfolding events is nevertheless permeated throughout with lexical classifications and it is probably precisely these lexical classifications that are employed during planning and explanation of any activities observed or undertaken.

4.5. Concept evolution

“A marriage is a contract that is regulated by civil and social constraints. These constraints can change but the meaning of marriage continues over time.”

On the surface, this example is a case of a label being maintained while the criteria for the application of the label change, although there is also a suggestion that the label has some ‘core’, or ‘basic’ meaning that does not change. The translation of this mini-narrative into corresponding GUM specifications is straightforward but also revealing of a considerable range of ambiguity in what might be intended. On the one hand, this is again considered one of the benefits of adopting the GUM method – i.e., one is forced to consider ambiguities or unclarities in what one is attempting to model. On the other hand, it also enables important parallels to be drawn with certain aspects of what we have seen in the previous examples. In particular, the central role of descriptions is made very clear.

The first sentence is a statement of identity, which is modelled in GUM with the BeingAndHaving configuration Identity. This configuration relates an Element in the domain role to a further Element in the attribute role: the specific intended meaning (interpretation policy) of this configuration is, however, that what is described is a relation between levels of abstraction – that is, this is by no means a mathematical equality. Extensive discussion of this relationship and its role for the construction of scientific or technical discourses is given in the literature (cf. Halliday and Matthiessen, 1999). The second sentence applies the predicate ‘change’, a NonAffectingHappening configuration with pre- and post-conditions as discussed above, to an entity described as ‘the meaning of marriage’. The GUM configuration involved here is Signification, another subtype of BeingAndHaving. Significations, as configurations, have all the temporal possibilities of relational configurations in general, and so can change over time.

The example is multiply ambiguous, however: particularly in its unpacking of just what is being given a meaning, i.e., which entity stands in the domain role of the signification configuration. First, one might assume that what is intended is something of the form ‘the meaning of “marriage”’, which is just an association of a lexical item (or Word in the current GUM definition: a particular subclass of Abstraction) with some description of what is intended with the lexical item. Second, however, the example does not actually place ‘marriage’ in quotation marks and so may call for interpretation in a more sophisticated fashion. In that case, it is not a lexical item that is being described but ‘marriage’ as some set of activities or relations that the society has come to group under the term. This is then parallel to the various activity labelings that have been used in the previous examples, such as ‘walking’, ‘running’, and so on: the distinction between ‘walking’ and ‘running’ relies upon where a particular language draws its boundaries. Identifying such activities is far from straightforward even for physical activities, as attempts to provide automatic labelling of actions from raw data demonstrate. Doing this for ‘marriage’ is not going to be any easier.7

⁷
There are further readings of ‘meaning’: for example, the statement might be read as saying the significance of marriage as an institution does not change; this moves in a different direction again.

It is consequently plausible that a society would attempt to regulate activities with particular social significance in various ways, for example, by establishing contracts. The first sentence is an attempt to fix (by identification) marriage in this way. This leads into another very different area of formalization concerned with narratives of social order and regulation. The example states that these may change in their details, which is again unexceptional since any description may be changed. The scope or range of a lexical label is a social convention which may change over time, possibly quite idiosyncratically. More interesting is the role that such descriptions may be expected by a society to play – i.e., as providing identity criteria for the acceptance or rejection of certain sets of activities and relationships in relation to the designated class (contract). This belongs to the meaning of ‘contracts’, which is yet another narrative. This situation is consequently a further, more explicitly articulated form of social artifact design: an abstract semantic schema is created as a contract and it is this narrative that defines acceptable instances. The contents of the schema is constituted by a set of definitions and so these definitions might also be given a GUM description. This is to move quite explicitly in the direction of embedding GUM categories within explicitly modelled descriptions in the sense of the Descriptions and Situations ontological extension for DOLCE mentioned above (Gangemi and Mika, 2003; Presutti and Gangemi, 2016). A detailed exploration of an explicit embedding of GUM specifications as formal ‘descriptions’ would consequently be very interesting but has not been undertaken so far.

At a somewhat deeper level, however, the idea that there is some ‘meaning of marriage’ that is constant is not tenable despite the example’s assertion: it is simply the case that there are ‘folk’-narratives independently of the formulated contracts that similarly offer ways of explaining what is intended with ‘marriage’ and which distinguish situations where it would be said to apply. There are just as few grounds for believing that these are unchangeable as there are for assuming that social contracts do not change. It would be difficult even to express this with a GUM specification: the grammatical reactancies for ‘meaning’ are similar to those for color and several other properties and temporal dependence is consequently baked in (cf. Bateman, 2004b). The force of the last part of the example is then the bare assertion that ‘marriage’ has some core meaning which never changes. If this is arguable at all, it would be a task of sociology or anthropology rather than of ontology.

The discussion here consequently leads directly to a consideration of the semiotically self-embedding nature of ontology (Presutti and Gangemi, 2016; Bateman, 2019). Ontologies are similar to the social contracts of marriage in that they are descriptions, expressed with certain more or less formal languages, that attempt to fix their objects of description in specific respects. We might then offer a final extended case study example for discussion:

“An ontology is a formalization that is regulated by ontologists. These formalizations can change but the meaning of ontology continues over time.”

Positions probably vary on this: the GUM position is that this would be modeled precisely as the ‘marriage’ case was, with judgements of the possibility of such core meanings following similarly. Moreover, although it might be thought that replacing ‘marriage’ or ‘ontology’ with ‘walking’ moves us to safer (ontological) ground, the GUM position is, again, that such boundaries are by no means as clear as often assumed.

5. Ontology usage and community impact

In the 1980s and 1990s the Penman Upper Model, the predecessor to GUM, was often adopted as an overarching set of categories for a variety of application domains – thereby serving some of the organizational roles of a foundational ontology – even though the design principles upon which GUM was built explicitly anchored the details of its categories and relationships to linguistic concerns. GUM’s aim was to cover all linguistic expressions, thus providing a suitable level of abstraction for engagement with resources exhibiting broad linguistic coverage. In part, this usage reflected a lack of appropriate ontological resources at that time, which is now a less acute problem. It also reflected, however, the fact that the categories provided by GUM were achieving a degree of generality and reusability uncharacteristic of many domain ontologies or models proposed. The reasons for this follow directly from the design principles and the linguistic theory informing GUM’s development process in that the process of interpretation is turned into a process of grammatically-guided semantic analysis.

Practically, GUM was consequently used as the type hierarchy standing behind typed semantic specifications deployed for automatic natural language generation systems (in various languages), as a target for analysis for automatic analysis components, and as a point of comparison across languages for studies of translation and cross-language relationships. This involves, for example, providing semantic type checks to establish the semantic well-formedness of semantic expressions. Building on this basic functionality, uses and applications of GUM and its predecessors to date then include:

creation of the large-scale Sensus interlingual knowledge base for machine translation, which drew on several components including the Penman Upper Model (Hovy and Knight, 1993),

adoption as meta-organization for modeling various domains for natural language generation and dialogue systems (DiMarco et al., 1995; Bateman et al., 2007),

use as a flexible interlingua for multilingual natural language generation in instructional texts (cf., e.g., Rösner and Stede, 1994; Hartley and Paris, 1997; Kruijff et al., 2000),

development of two-level semantics for spatial language (Bateman, 2010a),

experiments in heterogeneous reasoning spanning symbolic representations and simulation for robotics and human-robot interaction (Bateman et al., 2018,2019; Pomarlan and Bateman, 2020).

The introduction of a layer of domain-independent semantics between syntactic analysis/generation and domain knowledge as pursued within GUM has emerged in several distinct approaches. Within a semantic parsing and interpretation context, it has been argued that such a semantic layer improves portability and re-use of components within dialogue systems (Dzikovksa et al., 2007); within a generation context, it has been argued similarly that compositional semantics needs characterizations that capture how language decomposes entities and that this is, again, independent of domain-specific organization (Stone, 2003). Within the spatial domain, we also see language-motivated characterizations proposed by Mavridis and Roy as a kind of ‘parsing’ of “situations into ontological types and relations that reflect human language semantics” (Mavridis and Roy, 2006). Here, just as in the GUM case, the relationship to language is intended to support automatic natural language processing, while the relationship to situations and ontological types is intended to ease their formal interpretation and contextualization. The Generalized Upper Model approach as a whole also shows some similarities with the proposals of Cimiano and Reyle (2006), who argue that linguistic semantics should incorporate aspects of foundational ontologies. This they term foundational semantics, which

“…is concerned with identifying that abstract meaning layer which remains constant across domains and applications. … From a theoretical point of view, foundational semantics aims at identifying the core components of the domain-independent meaning layer as well as to clarify their interplay, thus contributing to the understanding of the principles of semantic construction.”

In distinction to GUM, however, the starting point they adopt for their proposed account is given by established non-linguistic foundational ontologies, such as DOLCE. This is as a consequence only weakly connected to the concrete linguistic demands of the semantics of particular languages. Foundations of this kind should also assist in improving the adequacy of linguistically-motivated semantic organizations, but there remain many open questions about how these levels of description are best to be brought together while still respecting the productive modularity and heterogeneity of methods encouraged by GUM.

Footnotes

Acknowledgements

The Generalized Upper Model is the result of efforts of many people and institutions that have supported the longterm development of the ontology. The ideas reported here would not have been possible without that development, including the work of William Mann, Christian Matthiessen, Michael A.K. Halliday, Robert Kasper, Johanna Moore, Eduard Hovy, Yigal Arens, Renate Henschel, Fabio Rinaldi, Robert Ross, Thora Tenbrink, Mihai Pomarlan and others. Work on the ontology has been carried out at locations including USC/ISI (Los Angeles), the former GMD-Institute IPSI in Darmstadt, and Bremen University as components of research supported by a broad range of funding agencies, including the US NSF, DARPA, AFOSR, the GMD, the European Commission, and the German DFG – particularly in the DFG collaborative research centers for ‘Spatial Cognition’ (SFB/TR8) and ‘Everyday Activity and Science Engineering’ (SFB 1320).

References

Arrighi, C. & Ferrario, R. (2005). The dynamic nature of meaning. In

Magnani and

Dossena (Eds.), Computing, Philosophy and Cognition: Proceedings of the European Computing and Philosophy Conference (ECAP 2004) (pp. 295–312). College Publications.

Asher, N. & Lascarides, A. (2003). Logics of Conversation. Cambridge: Cambridge University Press.

Atencia, M. & Schorlemmer, M. (2012). An interaction-based approach to semantic alignment. Journal of Web Semantics, 12(C), 131–147. doi:10.1016/j.websem.2011.12.001.

Baldridge, J. & Kruijff, G.-J. (2002). Coupling CCG and hybrid logic dependency semantics. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics, Philadelphia, Pennsylvania (pp. 319–326).

Bateman & Space, J.A. (2013). Language and ontology: A response to Davis. Spatial Cognition & Computation, 13(4), 295–314. doi:10.1080/13875868.2013.808491.

Bateman, J.A. (1990). Upper modeling: Organizing knowledge for natural language processing. In Proceedings of the Fifth International Natural Language Generation Workshop, Pittsburgh, PA (pp. 54–60). http://acl.ldc.upenn.edu/W/W90/W90-0108.pdf .

Bateman, J.A. (1997). Enabling technology for multilingual natural language generation: The KPML development environment. Natural Language Engineering, 3(1), 15–55. doi:10.1017/S1351324997001514.

Bateman, J.A. (2004a). Realist linguistics and dynamic ontology. In

Graumann ,

Holz and

Plümacher (Eds.), Towards a Dynamic Theory of Language: A Festschrift in Honour of Wolfgang Wildgen Diversitas Linguarum, Bochum (pp. 79–117). Universitätsverlags N. Brockmann.

Bateman, J.A. (2004b). The place of language within a foundational ontology. In

Achille , Varzi and

Vieu (Eds.), Formal Ontology in Information Systems: Proceedings of the Third International Conference on Formal Ontology in Information Systems (FOIS-2004) (pp. 222–233). Amsterdam: IOS Press.

10.

Bateman, J.A. (2010a). Language and space: A two-level semantic approach based on principles of ontological engineering. International Journal of Speech Technology, 13(1), 29–48. doi:10.1007/s10772-010-9069-x.

11.

Bateman, J.A. (2010b). Ontological diversity: The case from space. In

Galton and

Mizoguchi (Eds.), Formal Ontology in Information Systems. Proceedings of the 6th International Conference (FOIS 2010) Frontiers in Artificial Intelligence and Applications (pp. 5–16). Amsterdam: IOS Press.

12.

Bateman, J.A. (2010c). Ontologies of language and language processing. In

Poli ,

Healy and

Kameas (Eds.), Theory and Applications of Ontology: Computer Applications (pp. 393–410). Dordrecht, Heidelberg, London and New York: Springer. doi:10.1007/978-90-481-8847-5_17.

13.

Bateman, J.A. (2019). Ontology, language, meaning: Semiotic steps beyond the information artifact. In

Borgo ,

Ferrario ,

Masolo and

Vieu (Eds.), Ontology Makes Sense. Essays in Honor of Nicola Guarino. Frontiers in Artificial Intelligence and Applications (Vol. 316, pp. 119–135). IOS Press.

14.

Bateman, J.A., Beetz, M., Beßler, D., Bozcuoğlu, A.K. & Pomarlan, M. (2018). Heterogeneous ontologies and hybrid reasoning for service robotics: The EASE framework. In

Ollero ,

Sanfeliu ,

Montano ,

Lau and

Cardeira (Eds.), ROBOT 2017: Third Iberian Robotics Conference (Advances in Intelligent Systems and Computing 693) (pp. 417–428). Springer. doi:10.1007/978-3-319-70833-1_34.

15.

Bateman, J.A., Borgo, S., Lüttich, K., Masolo, C. & Mossakowski, T. (2007). Ontological modularity and spatial diversity. Spatial Cognition and Computation, 7(1), 97–128. doi:10.1080/13875860701337991.

16.

Bateman, J.A., Henschel, R. & Rinaldi, F. (1995). Generalized Upper Model 2.0: Documentation. Tech. rep. GMD/Institut für Integrierte Publikations- und Informationssysteme Darmstadt, Germany. http://purl.org/net/gum2.

17.

Bateman, J.A., Hois, J., Ross, R.J. & Tenbrink, T. (2010). A linguistic ontology of space for natural language processing. Artificial Intelligence, 174(14), 1027–1071. https://dx-doi-org.web.bisu.edu.cn/10.1016/j.artint.2010.05.008 . doi:10.1016/j.artint.2010.05.008.

18.

Bateman, J.A., Kasper, R.T., Moore, J.D. & Whitney, R.A. (1990). A general organization of Knowledge for natural language processing: The PENMAN Upper Model. Tech. rep. USC/Information, Sciences Institute Marina del Rey, California. http://www.fb10.uni-bremen.de/anglistik/langpro/kpml/um89/um89-penman.pdf.

19.

Bateman, J.A. & Lestrade, S. (2014). The linguistic ontology of space: General methods and the role of comparative linguistic evidence. In

D.R.

Montello ,

K.E.

Grossner and

G.J.

Donald (Eds.), Space in Mind: Concepts for Spatial Learning and Education (pp. 49–71). Cambridge, MA: MIT Press.

20.

Bateman, J.A., Magnini, B. & Fabris, G. (1995). The generalized upper model knowledge base: Organization and use. In

N.J.I.

Mars (Ed.), Towards Very Large Knowledge Bases: Knowledge Building and Knowledge Sharing (pp. 60–72). Amsterdam: IOS Press.

21.

Bateman, J.A., Magnini, B. & Rinaldi, F. (1994). The generalized {Italian, German, English} upper model. In Proceedings of the ECAI94 Workshop: Comparison of Implemented Ontologies, Amsterdam. http://citeseer.ist.psu.edu/bateman94generalized.html .

22.

Bateman, J.A., Pomarlan, M. & Kazhoyan, G. (2019). Embodied contextualization: Towards a multistratal ontological treatment. Applied Ontology, 14(4), 379–413. doi:10.3233/AO-190218.

23.

Bateman, J.A., Tenbrink, T. & Farrar, S. (2007). The role of conceptual and linguistic ontologies in discourse. In Discourse Processes (Vol. 44, pp. 175–213).

24.

Bennett, B. (2001). What is a forest? On the vagueness of certain geographic concepts. Topoi, 20(2), 189–201. doi:10.1023/A:1017965025666.

25.

Bergen, B.K. (2012). Louder than Words: The New Science of How the Mind Makes Meaning. New York: Basic Books.

26.

Borgo, S., Guarino, N. & Masolo, C. (1996). Stratified ontologies: The case of physical objects. In Proceedings of the Workshop on Ontological Engineering at ECAI’96, Budapest, Hungary.

27.

Bos, J. (2016). Expressive power of abstract meaning representations. Computational Linguistics, 42(3), 527–535. doi:10.1162/COLI_a_00257.

28.

Bowerman, M. (1999). Learning how to structure space for language: A crosslinguistic perspective. In

Bloom ,

M.A.

Peterson ,

Nadel and

M.F.

Garrett (Eds.), Language and Space (pp. 385–436). Cambride, MA: MIT Press.

29.

Bunt, H. (1981). The formal semantics of mass terms.

30.

Cimiano, P. & Reyle, U. (2006). Towards foundational semantics – ontological semantics revisited. In

Bennett and

Fellbaum (Eds.), Proceedings of the International Conference on Formal Ontology in Information Systems (FOIS) (pp. 51–62). IOS Press. http://www.aifb.uni-karlsruhe.de/WBS/pci/Publications/fois06.pdf .

31.

Copestake, A., Flickinger, D., Pollard, C. & Sag, I. (2005). Minimal recursion semantics: An introduction. Research on Language and Computation, 3, 281–332. http://lingo.stanford.edu/sag/papers/copestake.pdf . doi:10.1007/s11168-006-6327-9.

32.

del Socorro Bernados Galindo, M. & Aguado de Ceo, G. (2001). Adapting the generalized upper model to Spanish. In

Angelova ,

Bontcheva ,

Mitkov ,

Nicolov and

Nikolov (Eds.), Proceedings of the Euroconference Recent Advances in Natural Language Processing (RANLP-2001) (pp. 103–107). Bulgaria: Tzigov.

33.

DiMarco, C., Hirst, G., Wanner, L. & Wilkinson, J. (1995). HealthDoc: Customizing patient information and health education by medical condition and personal characteristics. In First International Workshop on Artificial Intelligence in Patient Education, Glasgow.

34.

Dzikovksa, M.O., Allen, J.F. & Swift, M.D. (2007). Linking semantic and knowledge representations in a multi-domain dialogue system. Journal of Logic and Computation.

35.

Eschenbach, C., Tschander, L., Habel, C. & Kulik, L. (2000). Lexical specification of paths. In

Freksa ,

Brauer ,

Habel and

K.F.

Wender (Eds.), Spatial Cognition II – an Interdisciplinary Approach to Representing and Processing Spatial Knowledge (pp. 127–144). Berlin, Heidelberg: Springer. http://link.springer-ny.com/link/service/series/0558/tocs/t1849.htm .

36.

Fine, K. (2010). Towards a theory of part. Journal of Philosophy, 107(11), 559–589. doi:10.5840/jphil20101071139.

37.

Frawley, W. (1992). Linguistic Semantics. Hillsdale, New Jersey: Lawrence Erlbaum Associates.

38.

Galton, A. (2018). Processes as patterns of occurrence. In

Stout (Ed.), Process, Action, and Experience (pp. 41–57). Oxford University Press.

39.

Gangemi, A. (2005). Ontology design patterns for semantic web content. In

Gil (Ed.), Proceedings of the International Semantic Web Conference 2005 (ISWC 2005). LNCS (Vol. 3729, pp. 262–276). Berlin/Heidelberg: Springer. http://metokis.salzburgresearch.at/files/papers/gangemi_2005_ontology_design_patterns.pdf .

40.

Gangemi, A. & Mika, P. (2003). Understanding the semantic web through descriptions and situations. In Proceedings of ODBASE 2003 www.loa-cnr.it/Papers/ODBASE-CONTEXT.pdf .

41.

Gärdenfors, P. (2000). Conceptual Spaces: The Geometry of Thought. Cambridge, MA: MIT Press.

42.

Granatosky, M.C., Bryce, C.M., Hanna, J., Fitzsimons, A., Laird, M.F., Stilson, K., Wall, C.E. & Callum, F.R. (2018). Inter-stride variability triggers gait transitions in mammals and birds. Proceedings of the Royal Society B, 285(1893). https://dx-doi-org.web.bisu.edu.cn/10.1098/rspb.2018.1766 .

43.

Guarino, N. & Welty, C. (2004). An overview of OntoClean. In

Staab and

Studer (Eds.), Handbook on Ontologies (pp. 151–171). Heidelberg and Berlin: Springer. doi:10.1007/978-3-540-24750-0_8.

44.

Halliday, M.A.K. & Matthiessen, C.M.I.M. (1999). Construing Experience Through Meaning: A Language-Based Approach to Cognition. London: Cassell.

45.

Halliday, M.A.K. & Matthiessen, C.M.I.M. (2013). Halliday’s Introduction to Functional Grammar (4th ed.). New York: London and Routledge.

46.

Hartley, A. & Paris, C. (1997). Automatic text generation for software development and use. In

Somers (Ed.), Terminology, Translation and LSP: Studies in Language Engineering in Honor of J.C. Sager (pp. 221–242). Amsterdam: Benjamins.

47.

Higginbotham, J. (1985). On semantics. Linguistic Inquiry, 16, 547–593.

48.

Hobbs, J.R. (1995). Sketch of an ontology underlying the way we talk about the world. International Journal of Human-Computer Studies, 43(5/6), 819–830. doi:10.1006/ijhc.1995.1076.

49.

Hois, J. & Kutz, O. (2008). Counterparts in language and space – similarity and E-connection. In

Eschenbach and

Grüninger (Eds.), Proceedings of the International Conference on Formal Ontology in Information Systems (FOIS) (pp. 266–279). Amsterdam: IOS Press.

50.

Hovy, E.H. & Knight, K. (1993). Motivating shared knowledge resources: An example from the Pangloss collaboration. In Proceedings of IJCAI Workshop on Knowledge Sharing and Information Interchange, International Joint Conference on Artificial Intelligence.

51.

Huang, C., Calzolari, N., Gangemi, A., Lenci, A., Oltramari, A. & Prevot, L. (Eds.) (2008). Ontologies and the Lexicon. Cambridge: Cambridge University Press.

52.

Jackendoff, R. (1983). Semantics and Cognition. Cambridge, MA: The M.I.T. Press.

53.

Jackendoff, R. (1999). The architecture of the linguistic-spatial interface. In

Bloom ,

M.A.

Peterson ,

Nadel and

M.F.

Garrett (Eds.), Language and Space (pp. 1–30). Cambride, MA: MIT Press.

54.

Janowicz, K., Gangemi, A., Hitzler, P., Krisnadhi, A. & Presutti, V. (2016). Introduction: Ontology design patterns in a Nutshell. In

Hitzler ,

Gangemi ,

Janowicz ,

Krisnadhi and

Presutti (Eds.), Ontology Engineering with Ontology Design Patterns: Foundations and Applications. Studies on the Semantic Web, xi–xvi (Vol. 25). Amsterdam: IOS Press.

55.

Kasper, R.T. (1989). A flexible interface for linking applications to PENMAN’s sentence generator. In

Hirschman (Ed.), Proceedings of the DARPA Workshop on Speech and Natural Language, San Mateo, CA: Morgan Kaufmann. http://www.cs.mu.oz.au/acl/H/H89/H89-1022.pdf. Available from ACL Anthology H89-1022.

56.

Kruijff, G.-J., Teich, E., Bateman, J.A., Kruijff-Korbayová, I., Skoumalová, H., Sharoff, S., Sokolova, L., Hartley, T., Staykova, K. & Jiří, H. (2000). A multilingual system for text generation in three Slavic languages. In Proceedings of the 18th. International Conference on Computational Linguistics (COLING’2000), Saarbrücken, Germany (pp. 474–480). doi:10.3115/990820.990889.

57.

Kutz, O., Bateman, J.A., Neuhaus, F., Mossakowski, T. & Bhatt, M. (2015). E pluribus unum: Formalisation, use-cases, and computational support for conceptual blending. In

T.R.

Besold ,

Schorlemmer and

Smaill (Eds.), Computational Creativity Research: Towards Creative Machines (Atlantis Thinking Machines 7) (pp. 167–196). Springer. https://link-springer-com.web.bisu.edu.cn/chapter/10.2991/978-94-6239-085-0_9 .

58.

Lang, E. & Maienborn, C. (2011). Two-level semantics: Semantic form and conceptual structure. In

Maienborn ,

von Heusinger and

Portner (Eds.), Semantics. An International Handbook of Natural Language Meaning (HSK 33.1) (Vol. 1, pp. 709–740). Berlin and New York: de Gruyter Mouton.

59.

Langacker, R.W. (1988). A view of linguistic semantics. In

Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 49–90). Amsterdam: John Benjamins. doi:10.1075/cilt.50.04lan.

60.

Langkilde, I. & Knight, K. (1998). Generation that exploits corpus-based statistical knowledge. In Proceedings of the 36th Annual Meeting of the ACL and the 17th International Conference on Computational Linguistics (COLING), Montreal, Quebec (pp. 704–710).

61.

Levin, B. (1993). English Verb Classes and Alternations: A Preliminary Investigation. Chicago and London: University of Chicago Press.

62.

Levin, B. & Rappaport Hovav, M. (2005). Argument Realization. Cambridge: Cambridge University Press.

63.

Levinson, S.C., Kita, S., Haun, D.B.M. & Rasch, B.H. (2002). Returning the tables: Language affects spatial reasoning. Cognition, 84, 155–188. doi:10.1016/S0010-0277(02)00045-8.

64.

MacGregor, R.M. (1993). Representing reified relations in Loom. Journal of Experimental and Theoretical Artificial Intelligence, 5, 179–193. doi:10.1080/09528139308953767.

65.

Maienborn, C. (2011). Event semantics. In

Maienborn ,

von Heusinger and

Portner (Eds.), Semantics. An International Handbook of Natural Language Meaning (HSK 33.1) (Vol. 1, pp. 802–829). Berlin and New York: de Gruyter Mouton.

66.

Mann, W.C. (1983). An overview of the PENMAN text generation system. In Proceedings of the National Conference on Artificial Intelligence (pp. 261–265). AAAI. Also appears as USC/Information Sciences Institute, RR-83-114.

67.

Mann, W.C. (1985a). An introduction to the Nigel text generation grammar. In

J.D.

Benson and

W.S.

Greaves (Eds.), Systemic Perspectives on Discourse: Selected Theoretical Papers from the 9th. International Systemic Workshop (pp. 84–95). Norwood, NJ: Ablex Pub. Corp.

68.

Mann, W.C. (1985b). Janus Abstraction Structure – Draft 1. An informal project technical memo of the Janus project at ISI.

69.

Mann, W.C., Arens, Y., Matthiessen, C.M.I.M., Naberschnig, S. & Sondheimer, N.K. (1985). Janus abstraction structure — draft 2. Tech. rep. USC/Information, Sciences Institute Marina del Rey, California. (Circulated in draft form only.).

70.

Masolo, C., Borgo, S., Gangemi, A., Guarino, N. & Oltramari, A. (2003a). Ontologies library (final). WonderWeb Deliverable D18 ISTC-CNR Padova, http://wonderweb.semanticweb.org/deliverables/documents/D18.pdf.

71.

Masolo, C., Borgo, S., Gangemi, A., Guarino, N., Oltramari, A. & Schneider, L. (2003b). The WonderWeb library of foundational ontologies: Preliminary report. WonderWeb Deliverable D17 (2.1) ISTC-CNR Padova, Italy.

72.

Masolo, C., Guizzardi, G., Vieu, L., Bottazzi, E. & Ferrario, R. (2005). Relational roles and qua-individuals. In

Boella ,

Odell ,

Van Der Torre and

Verhagen (Eds.), Proceedings of AAAI Fall Symposium on Roles, an Interdisciplinary Perspective (pp. 103–112). Menlo Park, CA: AAAI.

73.

Matthiessen, C.M.I.M. (1983). Choosing primary tense in English. Studies in Language, 7(3), 369–430. doi:10.1075/sl.7.3.03mat.

74.

Matthiessen, C.M.I.M. (1996). Tense in English seen through systemic-functional theory. In

Butler ,

Berry ,

Fawcett and

Huang (Eds.), Meaning and Form: Systemic Functional Interpretations, Norwood, NJ: Ablex.

75.

Mavridis, N. & Roy, D. (2006). Grounded situation models for robots: Where words and percepts meet. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 4690–4697).

76.

Moens, M. & Steedman, M.J. (1988). Temporal ontology and temporal reference. Computational Linguistics, 14(2), 15–28.

77.

Mossakowski, T., Kutz, O., Codescu, M. & Lange, C. (2013). The distributed ontology, modelling and specification language – DOL. In Proc. of the 7th International Workshop on Modular Ontologies (WoMO 2013), Corunna, Spain. Co-located with LPNMR 2013.

78.

Müller, C. & Cienki, A. (2009). Words, gestures, and beyond: Forms of multimodal metaphor in the use of spoken language. In

C.J.

Forceville and

Urios-Aparisi (Eds.), Multimodal Metaphor (pp. 297–328). Berlin and New York: de Gruyter Mouton.

79.

Parsons, T. (1990). Events in the Semantics of English: A Study in Subatomic Semantics. Cambridge, MA and London: MIT Press.

80.

Pelletier, F.J. (1975). Non-singular reference: Some preliminaries. Philosophia, 5(4), 451–465. doi:10.1007/BF02379268.

81.

Pomarlan, M. & Bateman, J.A. (2020). Embodied functional relations: A formal account combining abstract logical theory with grounding in simulation. In Formal Ontology in Information Systems (pp. 155–168). Amsterdam: IOS Press.

82.

Presutti, V. & Gangemi, A. (2016). Dolce+D&S ultralite and its main ontology design patterns. In

Hitzler ,

Gangemi ,

Janowicz ,

Krisnadhi and

Presutti (Eds.), Ontology Engineering with Ontology Design Patterns: Foundations and Applications. Studies on the Semantic Web (Vol. 25, pp. 81–103). Amsterdam: IOS Press.

83.

Pustejovsky, J. (1991). The syntax of event structure. Cognition, 41, 47–81. doi:10.1016/0010-0277(91)90032-Y.

84.

Pustejovsky, J. (2000). Events and the semantics of opposition. In

Tenny and

Pustejovsky (Eds.), Events as Grammatical Objects: The Converging Perspectives of Lexical Semantics and Syntax (pp. 445–482). Stanford, CA: CSLI Publications.

85.

Rösner, D. & Stede, M. (1994). TECHDOC: Multilingual generation of online and offline instructional text. In Fourth International Conference on Applied Natural Language Processing (4th. ANLP), Stuttgart (pp. 209–210). doi:10.3115/974358.974413.

86.

Schutz, A. & Luckmann, T. (1974). The Structures of the Life-World. London: Heinemann. Translated by R.M. Zaner and H.T. Engelhardt.

87.

Slobin, D.I. (1987). Thinking for speaking. In

Aske ,

Beery ,

Michaelis and

Filip (Eds.), Proceedings of the 13th Annual Meeting of the Berkeley Linguistics Society Meeting (pp. 435–445).

88.

Smith, B. & Mark, D.M. (2003). Do mountains exist? Towards an ontology of landforms. Environment and Planning B (Planning and Design), 30(3), 411–427. http://ontology.buffalo.edu/smith/articles/Mountains.htm . doi:10.1068/b12821.

89.

Steedman, M.J. (2000). The Syntactic Process. Cambridge, MA: MIT Press.

90.

Stone, M. (2003). Ontology and description in computational semantics. In Proceedings of the 3rd Workshop on Knowledge and Reasoning in Practical Dialogue Systems at the 18th IJCAI. http://www.ida.liu.se/labs/nlplab/ijcai-ws-03/papers/stone.pdf .

91.

Vendler, Z. (1957). Verbs and times. Philosophical Review, 66, 143–160. doi:10.2307/2182371.

92.

Weischedel, R.M. (1989). A hybrid approach to representation in the JANUS natural language processor. In 27th Annual Meeting of the Association for Computational Linguistics, 26–29 June (pp. 193–202). Columbia, Vancouver, British: The Association for Computational Linguistics. doi:10.3115/981623.981647.

93.

Welty, C. & Andersen, W. (2005). Towards OntoClean 2.0: A framework for rigidity. Applied Ontology, 1(1), 107–116.

94.

Whorf, B. (1956). Whorf: Language, Thought, and Reality: Selected Writings. Cambridge, MA: The M.I.T. Press. (Edited by John Carrol).

95.

Winston, M.E., Chaffin, R. & Herrman, D.J. (1987). A taxonomy of part-whole relations. Cognitive Science, 11(4), 417–444. doi:10.1207/s15516709cog1104_2.

96.

Wood, Z. & Galton, A. (2009). A taxonomy of collective phenomena. Applied Ontology, 4(3–4), 267–292. doi:10.3233/AO-2009-0071.

97.

Yang, G. & Bateman, J.A. (2002). The Chinese aspect system and its semantic interpretation. In

S.-C.

Tseng (Ed.), Proceedings of the 19th International Conference on Computational Linguistics (COLING-2002). Association of Computational Linguistics and Chinese Language Processing Academica Sinica (Vol. 2, pp. 1128–1134). Taipei, Taiwan: Association of Computational Linguistics and Chinese Language Processing.