Frameworks for knowledge representation
An ontology is an explicit specification of a conceptualization (Gruber 1993, p. 199).
Moreover, I have always disliked the word ‘ontology’ (Hacking 2002, p. 1).
This chapter, along with several of those that follow, is directed towards the problem of representation and translation across knowledge systems and frameworks, with a particular focus on those used in the emerging world of the semantic web. Knowledge systems are all too frequently characterised in essentialist terms—as though, as the etymology of ‘data’ would suggest, they are merely the housing of neutral empirical givens. In this book we maintain, on the contrary, that systems always carry with them the assumptions of cultures that design and use them—cultures that are, in the very broadest sense, responsible for them. This is the case for knowledge systems in general, as well as the specifics of what has come to be called the ‘semantic web’, and the ontologies, schemas, taxonomies and other representations of knowledge it supports. This chapter begins by setting the scene for modern approaches to knowledge representation, constructing a broad historical frame which both inspired and motived these approaches. It then introduces the semantic web, arguably the most significant of these approaches. It describes both the affordances and challenges of the semantic web, and outlines several key concepts, which will be mobilised in later chapters—semantics, ontologies and commensurability. This chapter also outlines some of the claims and interventions this book intends to make about both the semantic web specifically, and knowledge representation, management and use generally.
Before the semantic web is described more formally, it is useful to try to articulate what it is in broad brush strokes. At its most general, it is an encompassing vision which imagines a network of connected, federated and integrated databases (Berners-Lee, Hendler and Lassila 2001). It is motivated by the desire to simplify the integration of information from the myriad variety of existing data sources and formats on the web. In the language of the semantic web these structured data sets are termed ontologies, picking up on the analogy with philosophical ontology—how a region of the world is explicitly conceptualised in a series of codified commitments (Gruber 1993). Semantic web ontologies use formal languages—the Resource Definition Framework (RDF) and Ontology Web Language (OWL) to express these commitments (Berners-Lee, Hendler and Lassila 2001).
Ontologies are taken here to be only an exemplary species of the broader genus of knowledge systems—a genus which can be extended to include other types of database models, Extensible Markup Language (XML) schemas, expert systems and electronic classification systems generally. So while the focus is often directed towards semantic web ontologies, since they are not yet as commonly used in organisations as other types of systems, casting a broader net aims to extend the generality of the research findings without loss of semantic specificity. As the argument goes on to show, moreover, even the different formal properties of rival system types— semantic web ontologies compared with older database information models, for instance—can involve important assumptions of a philosophical ontological kind as well.
While shared and standardised ontologies may simplify the job of system integrators connecting data services, without explicit acknowledgement of their epistemological assumptions and conditions—how it is that systems claim to know what they know—there will remain significant impediments to the realisation of the semantic web. By adopting the standpoint that knowledge is a culturally constructed and negotiated process, this book aims to find some heuristic guidelines for finding points of similarity and difference in the systems which codify knowledge. But first, it is helpful to understand something of the background against which the desire to codify, organise and construct baroque informatic systems arose to begin with.
Putting things in order
In The Order of Things, Foucault (1970) writes of the ‘great tables of the seventeenth and eighteenth centuries, when in the disciplines of biology, economics and philology the raw phenomena of experience was classified, categorised, organised and labeled’. At the start of the twenty-first century, when the classificatory technologies of the file system, spreadsheet, database and internet search engine have superseded those of the ruler and pencil, these descriptions of ‘great tables’ and their accompanying heroic taxonomic enterprises can seem quaint and anachronistic. The experience of lists, tables, hierarchical trees and networks and other informational structures as organisational aids is now unremarkable, quotidian, a tacit quality of a modern sensibility, reflected in the acquired facility to navigate everything from baroque scientific taxonomies and global standards to organisational directories and personalised databases. Consumers of electronic devices invest heavily in their repositories of music, books, photos and film, marking individual entries with qualifications of genre, commentary, ratings, biographical snippets and a host of other conceptual distinctions and demarcations. Business, governments and other organisations are necessarily technocratic taxonomists on a grand scale, investing in and managing large knowledge bases, processes and infrastructure. Such fervent activity has even inspired the emergence of a dedicated industry and academic discipline—that of knowledge management. Biology, one of the fields of scientific enterprise Foucault himself analyses, features ever-expanding databases of proteins, genomes, diseases and other biological entities, so vast in size that any single human attempt to review the data would fail by orders of magnitude (Arunguren 2005). It is hard therefore to share Foucault’s wonder at the ambition and scope of classical scholarship, without making an equally wondrously empathic leap back into the past. A modern-day reaction might instead regard these old classificatory systems as historical curiosities; at most, as experimental preludes, for better or worse, to the immense contemporary and global industries which serve an insatiable demand for information.
Yet our current age is also heir to the efforts of those classical scholars. Since Leibniz, the development of symbolic systems to represent knowledge has been a recurring motif of philosophy and, later, of other more applied disciplinary studies. From his universal symbolism, to Kant’s categories, to Frege’s descriptions of a formal logic, to the development of logical positivism in the 1920s, to, finally, the recent developments of the relational database, artificial intelligence and the semantic web, it is possible to trace a distinct and particular epistemological tradition. That tradition has sought to develop increasingly refined formal languages to represent statements about the world unambiguously. Rigorous empiricism—recording only observable facts—would, when coupled with an automatic deductive procedure based on a logical formalism, simplify the production of all knowledge to a series of mechanical acts. In Leibniz’s famous dictum, once this point had been reached even philosophers would be able to settle arguments by appealing to machination: ‘Let us calculate!’ (Lenzen 2004).
There have been at least two notable impediments to the realisation of this vision up until the end of the twentieth century. The first is the development of feasible logic systems and technical implementations systems for representing these concepts. This has been the subject of considerable research and application in artificial intelligence, knowledge representation and broader information technology over the last 50 years. Such research, and the practical consequences of it, have produced in turn a series of pivotal technologies for the emergence of what Castells (1996) terms the ‘Network Society’: the relational database—the current paradigmatic means for storing structured organisational information; the spreadsheet—a metaphor which pervades the construction of tabular data in the personal computing era; XML—a near-ubiquitous format for describing and transmitting data on the internet; and semantic web ontologies, the emerging standardised mechanism for representing knowledge on the internet.
The second impediment is development of consensual arrangements of concepts against which facts can be faithfully recorded. As the many successful cases of technical standards ratified by the ISO and other bodies show, there has been considerable success in efforts to develop standards. However, unlike the production of logical systems and implementations, consensus over such arrangements is frequently a brittle social dynamic, reliant on what (Davidson 2006) terms ‘the principle of charity’ adopted between heterogenous cultures and actors, as they seek to exchange meaning with each other.
The development of computational classification systems and standards has experienced at least partial success because they facilitate a distinctly modern taxonomic impulse—an apparently unceasing desire to order, organise, catalogue, coordinate and control. What makes this desire distinctively modern? In response it could be argued, in a deflationary fashion, that the urge to put things in order is inherent in human language—nouns, and names particularly, express implicit taxonomies. However, natural languages are taxed with many functions other than that of articulating some state of affairs in the world—they must also issue imperatives, pose interrogatives, invoke vocatives and generally perform a host of more esoteric speech acts; and even in the case of assertoric utterances, they must also permit statement modulation according to tense, mood, aspect and a range of sociolinguistic inflections. In contrast, artificial formal languages are deliberately designed with both a minimal syntax—how statements can be expressed—and rigorous semantics—how those statements must be interpreted—in order to make electronic taxonomies easily constructible and unambiguously interpretable. These features are not coincidentally related to the rising informational needs of modern institutions, departments, bureaucracies and organisations. Indeed the tendencies of late capitalism suggest a self-reinforcing chain of multiple factors which stimulate this impulse towards order and categorisation: the operational benefits of the ‘network externalities’ brought about by global communication networks; legal directives towards greater transparency and accountability; competitive pressures towards greater efficiencies; and improved control and regulation of people and objects, effected through ever more fine-grained classificatory structures. These factors both motivate and, in turn, are facilitated by the great affordances of information technology in the post-industrial era.
At the same time, the modernist conception of an organisation as a highly regulated machine-like entity has been challenged by new, postmodern metaphors, which imagine the organisation as open, interconnected, self-reflexive, fluid, relational and networked (Ashkenas et al. 1995; Castells 1996). The organisation is tasked with new, contemporary demands: to be visible, transparent, connected and accountable. It is to be audited regularly and stringently; it must be open to public inspection and accountable to numerous stakeholders—not only its direct constituents or shareholders, but a complex network of those with ‘stakes’ in organisational governance and performance. It must also deal more directly with its members, constituents, customers, partners, employees, suppliers, regulatory bodies and press organisations, via a host of increasingly immediate, ubiquitous, connected and ‘on-demand’ technologies. Information is the pivotal part of this equation, the connection between the modernist imperative to control, order and organise, and the postmodern desire to connect what is controlled, both within and between organisational boundaries. Accordingly, the desire to organise large amounts of information has led to interest, funding and prestige to be associated with information technologies, processes and management. These in turn have been seen as central to development of more successful organisations—organisations at any rate capable of greater performativity in a capitalistic environment. The twin development of the modern organisation and information technologies has been mutually reinforcing, to the extent that neither could any longer be imagined without the other. They are both features of a distinct phase of modernity.
Yet, just as these developments show a trend towards ever greater adoption of common, standardised and homogenised technical artefacts—informational and otherwise—they do not preclude an inverse tendency towards greater differentiation, in which various organisational cultures, brought into engagement within a globalised electronic landscape, both recognise and indeed actively produce perspectival differences towards the world they share. Like painters describing a landscape from different angles, these diverse orientations found both the conditions and limitations of the kinds of facts and observations which can be asserted about the world. Accumulating a base of information—a database—enables organisations to retrieve and analyse data rapidly; yet the price of this is a certain rigidity or reification at work in the deployment of concepts used to structure the database. The record of a person in a database, for instance, captures only some facets, properties, attributes and variables about the person—those typically deemed salient for the use of the system. Moreover these properties ‘slice’ the person in predefined ways, based on assumptions held by the culture responsible for the database. As the system is used over time, as more records are added, and other systems are adapted to fit the particular conceptualisation employed by the system, it becomes increasingly difficult to re-engineer or ‘refactor’ it. Consequently the conceptualisation becomes reified—appearing naturalised through the resilience to change of the system it is deployed in. Lost, or at least obscured, is the potential for other kinds of descriptions of entities to emerge. Nothing indicates, with the passing of time, that this is only one possible way among many others of ‘carving nature at its joints’.
Viewed from the standpoints of either relativism or stark realism, this is either tautologically true or oxymoronically false—true if all expressions of facts are regarded as at best a partial and fragmentary glimpse of things as they are; false if some objective measure is accepted for why one concept is used instead of others. The objective here is to avoid any concomitant commitments along these metaphysical lines, but rather to recognise that in practice social convention determines a range of intermediate possibilities. To take one example, which is examined in further detail in one of the case studies: electronic documents are cultural objects, which are described in a variety of ways—as official records in records management systems; as collectible items in bibliographic databases; as consumable objects in distribution systems like Amazon; and as complex textual objects in word-processing applications. All of these systems can be said to adopt a different standpoint—a metaphorisation of a different set of concepts and conceptual relations—of documents. Yet, equally, none of these views captures the whole truth of a particular document for its author (the possible difficulties of writing it), or a reader (the interpretive reading of it), or indeed the various features of a document required for many other purposes. Rather they capture the particular ‘facticities’—to employ a Foucauldian term—needed to exercise socially instrumented practical functions around documents: to retrieve them, catalogue them, edit them, print and bind them, distribute them, sell them, account for them, and so on. However, the conceptualisations engaged to describe documents for various functions are not at the same time discrete and self- contained bundles of properties or, in philosophical jargon, qualia, separate and unrelated to each other. To retain the geometric and spatial metaphor which is used throughout the study, conceptualisations frequently connect at orthogonal conceptual junctions and splices. They may share common concepts and properties—in the same example, books, authors and titles might be common terms across different system conceptualisations—and yet they may stand in different configurations and relations, which more or less line up depending on the context of their translation and use. How to assess this ‘more-or-less-ness’, the degree of commensurability, between the conceptual configurations operationalised by different systems is then a question that information system ‘translators’—system analysts, engineers and programmers—increasingly face in a world where the prolixity of systems and the range of functions performed by them is ever-growing.
Between these opposing trends—towards standardisation, regulation, connectivity and unification on the one hand, and differentiation, customisation and individuation on the other—the promise of knowledge systems for these organisations has only been partially fulfilled. The digitisation of records management, the development of sophisticated data warehousing tools, the agreement on protocols for transmission of data across networks—among other things—has led to vast increases of scale in both the amount of data captured and the level of analysis which can be performed on this data. And yet here, too, in the age of the internet, the quantitative problems—cost and complexity—of communicating meaningful information across organisational boundaries have remained prohibitive, frustrating the aims of these very organisations. The semantic web is a technology platform explicitly designed to overcome the dilemmas of inter-system translation: a set of standards designed to allow translation and migration of data between systems and applications with the minimum of cognitive impedance or dissonance. Conceptual schemes are rendered as ‘ontologies’, collections of concepts, properties and individual data records, which can be developed using the existing technical infrastructure of the World Wide Web. Even here, however, interpretation, translation, coordination and negotiation of meaning cannot be relegated to the domain of purely technical and engineering considerations. While the systems themselves are technological artefacts, assessment of their commensurability leads from a concern over purely technical compatibility to broader questions of social meaning—what background cultural beliefs and practices motivate, justify and orient these systems? Along what dimensions can systems be said to be commensurable? What must be investigated, negotiated and made explicit in order for systems to be commensurable, translatable and interoperable? What elements of meaning might be sacrificed or abandoned in these negotiations? Together these questions compose a frame for exploring the central concern of this study—whether a holistic notion of commensurability, embracing both sociological and technological dimensions, can be usefully applied to the translation and coordination of organising schemes in the digital age.
Introducing the semantic web
A web of meaning
The semantic web ‘provides a common framework that allows data to be shared and reused across application, enterprise, and community boundaries’ (W3C 2009b). It is constructed on the existing scaffolding of the World Wide Web: it makes use of the whole infrastructure of Extensible Markup Language (XML), Uniform Resource Identifiers (URIs) and, to a degree, Hypertext Markup Language (HTML). Two key formal language specifications for making and connecting assertions about the world comprise the foundational building blocks of the semantic web: the Resource Definition Framework (RDF) and Ontology Web Language (OWL). Several derivative standards describe rules, trust and proof conditions for reasoning and sharing the resulting assertional networks. Any system that supports these standards should be able to write and save data, which can in turn be processed and reasoned over by other compliant systems—bringing about, in theory, a level of interoperability not possible previously. Both RDF and OWL have been developed to be compatible with XML, another standard and common language for the encoding of data and documents (W3C 2009a). One way of viewing the relationships between these standards is that XML supplies a standardised syntax, and RDF and OWL supply standardised semantics for data. Other syntaxes are also available for encoding RDF and OWL (Berners-Lee 2009). However, widely available support, in the form of software tools and libraries, make XML a convenient choice for many purposes.
RDF is designed for describing resources—documents, images, audio and video files, as well as real-world objects which ‘can be identified’— on the web (Miller and Manola 2004). Descriptions take the conventional logical form of subject-predicate-object, where the subject and object are generally identified via a web address or, more formally, a uniform resource identifier. RDF does not supply an explicit vocabulary of terms such as ‘author’, ‘publisher’ or ‘creation date’. Instead it operates at a higher level of abstraction, ‘specif[ying] mechanisms that may be used to name and describe properties and the classes of resource they describe’ (Guha and Brickley 2004). In other words, it provides well-defined abstract and formal structures—such as ‘class’, ‘property’, ‘string’, ‘date’ and ‘collection’—for composing such terms (Powers 2003). OWL, in turn, extends RDF to handle descriptions of ontologies—a central concept for this study, which warrants a more extended introduction below. Together RDF and OWL form a basis for the standardisation of structured data on the web, in such a way that human and machine agents can share, query, navigate, manipulate and conduct inferences with it.
The semantic web is typically explained in terms of consumer convenience. In a now famous statement heralding the advent of the semantic web, Berners-Lee, Hendler and Lassila (2001) describe how it makes possible, for example, the aggregation of book catalogue records across multiple websites, or the merging of contact information from one application with calendaring data from another. The same article, written in the promissory and optimistic tones of technology evangelism, outlines how the semantic web will, more broadly, simplify the electronic life of a prototypical user:
The semantic web will bring structure to the meaningful content of Web pages, creating an environment where software agents roaming from page to page can readily carry out sophisticated tasks for users. Such an agent coming to the clinic’s Web page will know not just that the page has keywords such as ‘treatment, medicine, physical, therapy’ (as might be encoded today) but also that Dr. Hartman works at this clinic on Mondays, Wednesdays and Fridays and that the script takes a date range in yyyy-mm-dd format and returns appointment times. And it will ‘know’ all this without needing artificial intelligence on the scale of 2001’s Hal or Star Wars’s C-3PO. Instead these semantics were encoded into the Web page when the clinic’s office manager (who never took Comp Sci 101) massaged it into shape using off-the-shelf software for writing semantic web pages along with resources listed on the Physical Therapy Association’s site (Berners-Lee, Hendler and Lassila 2001).
Arguably, though, the semantic web has greater application to the costly problems of system integration which preoccupy organisational IT departments or enterprises: for example, how to get the accounts system to ‘talk to’ the human resources system, how to integrate two customer databases after a company merger, or how to represent transaction details across different national taxation regimes. These translation scenarios are common areas of complexity and cost in system integration, and stand to benefit from the kinds of interoperability at least promised by the semantic web. A key example of this use has been the widespread adoption of the semantic web, and of ontologies in particular, among the bioinformatic research community. Chapter 10, ‘Describing knowledge domains’, explores this usage in further detail.
It is worth adding a cautionary note: in the decade since the early euphoric pronouncements of the semantic web, its adoption has been heavily fragmented. Research communities, such as those of the life sciences mentioned above, have been quick to embrace it. But the broader enterprise and consumer markets, targetted in the pitch cited above for instance, have stumbled over the apparent complexity and acronymic soup of its many recommendations and proposals. More specific causes have also been raised within the informal channels of the blogosphere (Shirky 2003), some of which are discussed in more detail in the comparison of knowledge systems in Chapter 7, ‘An historical introduction to formal knowledge systems’. Suffice to say, the degree to which the semantic web remains a research project, limited to scientific and academic applications, remains a highly contested issue. The recent more catholic usage of ‘ontology’, evident in the studies presented here, is indicative of a more general desire to explore possibilities of many semantic webs, and many ontologies, inspired but not necessarily constrained to the specific proposals of the Semantic Web—in its proper noun form.
Ontology—computing ‘what is’
Bearing in mind the preceding exhortation, before describing how ontologies are represented in OWL specifically, it is useful to describe the term ‘ontology’ in its more general computer science usage. The term has been appropriated from its philosophical roots to describe knowledge systems. Despite the shift in meaning from its traditional moorings—where it is far from being an unambiguous term—this appropriation is not without basis: an ontology, for knowledge representation, is a series of statements which purport to describe how the world is. The canonical definition for computer science usage of ‘ontology’ comes from Gruber (1993): ‘an ontology is an explicit specification of a conceptualization’. Elsewhere he elaborates: ‘Pragmatically, a common ontology defines the vocabulary with which queries and assertions are exchanged among agents. Ontological commitments are agreements to use the shared vocabulary in a coherent and consistent manner’ (Gruber 1995, p. 909).
‘Ontology’ is therefore, even in its computer science usage, conceived in broad terms. In this study, it is generally treated as an umbrella term for a range of electronic classification systems: from those with minimal explicit semantics through to ontologies developed in OWL with a highly explicit semantics. It therefore includes taxonomies, controlled vocabularies, thesauri, classification systems, catalogues, XML specifications, software designs, database and knowledge-base schemas, logical deduction systems and, finally, knowledge representation formats such as OWL. Unlike programs, ontologies or knowledge bases do not generally contain procedures, though they may include rules which can be processed by programs.
In its computer science usage, then, an ontology is a representation of some knowledge about a domain or field, sometimes also referred to as a knowledge base or database, and encoded using some derivative of a first-order logical formalism. Further, an ontology is typically composed of a series of logical axioms—the basic building blocks of the knowledge base:
classes—names designating sets of objects (for example, ‘cell’)
properties—names an attribute of a class (for example, ‘label’), or a relationship between two classes (for example, ‘is_contained_by’)
individuals—names an individual object (for example, ‘genid7566’— ‘An enzyme complex which in humans and yeast consists of at least five proteins’).
Briefly, these axioms can be related in various ways. Classes can be defined through relations of subsumption (parent–child relations) to form a traditional taxonomic hierarchy; properties can be assigned to classes as part of their definition in terms of ‘necessary and sufficient conditions’; and individuals can be stipulated in terms of their class association (what classes they are members of), and their attributes and relations with other objects.
Over the kinds of entrenched data systems commonly used—relational databases or spreadsheets, for example—ontologies offer several advantages:
Formal mathematical models provide the foundational semantic anchorings of ontologies, which allow unambiguous and—under certain conditions—tractable interpretation of ontological axioms.
Existing internet infrastructure provides some of the technical plumbing of ontologies—for example, object references or identifiers use web addresses (URIs), while the canonical syntactic format for ontologies is XML, a well-supported data markup language.
The underlying formalisms of ontologies also have well-defined means for stating relationships between them. One ontology can ‘import’, and subsequently reuse, definitions and data from another. Logical relationships of subsumption, equivalence and disjointness can be declared between axioms of different ontologies, as much as within a single given ontology.
These features make possible very large-scale reasoning over heterogenous data sets, which can improve structured searches, interoperability between research results and the automatic discovery of new relationships. There are particular trade-offs however—the particular pros and cons of ontologies, databases and other forms of knowledge representation are discussed further in Chapter 7, ‘An historical introduction to formal knowledge systems’.
Ontologies, at any rate, form a cornerstone of the promise of the semantic web—that in a global world of federated data sets, properly organised under shared ontologies and underlying conceptual schemes, users can publish and query information more quickly, more reliably and with the added benefit of computer-aided inferencing. In the many diverse fields of knowledge and informatics, with very large and ever-growing data sets, active and growing amendments to taxonomic structures, and a globally distributed network of researchers, this promise holds a special allure.
The Ontology Web Language—a language for the semantic web
The Ontology Web Language (OWL) was developed—or, more accurately, was derived from several earlier language initiatives—in order to provide a standardised way of representing ontologies on the semantic web. OWL itself comes in three language variants: OWL Lite, OWL DL (Description Logic) and OWL Full (W3C 2009b). These variants exhibit increasing degrees of computational complexity and expressivity. All variants originate in a long tradition of research into knowledge representation and machine learning, culminating in Description Logics in the 1980s and 1990s. RDF was developed for different, more pragmatic purposes of sharing data between systems, and early in the history of the semantic web efforts were made to harmonise RDF with OWL—since RDF permits a wide range of constructs, it is more expressive than OWL Full, but also computationally intractable. Both OWL and RDF can be rendered in a variety of syntaxes, such as XML, N-Triples and Notation3 (Berners-Lee 2009). While all of these syntaxes are humanly readable and writeable, in practice RDF and OWL files are typically generated by ontology editing tools such as Protégé (Gennari et al. 2003).
The purpose of the resulting OWL ontology is to provide a standards-based model of some set of facts about the world. Precisely because it is a standard, other systems can process an OWL ontology and make inferences about it; importantly, not only about a particular ontology itself, but also about how that ontology might connect with others. The semantic web envisages a network of ontologies which function like a large, distributed database. The current web of course might be said to supply this in a more amorphous form—but the whole point of the semantic web is precisely that any facts in an ontology have a certain amount of context made explicit. Sally is a Person; a Person is a sub-class of Animal; a Person may own a number of Pets; and so on. Each of these concepts, properties and individuals are defined in the ontology; therefore systems with no prior knowledge of Sally, Persons or Animals can nonetheless infer the relations between these things automatically.
Put another way: the semantics of data are encoded in ways that are specific or unique to the organisations, people or agents generating them. Ontologies attempt to solve this problem in a general way. They supply the mechanism by which agents can understand the semantics of any given domain—without prior knowledge of this domain. Consequently agents can consume arbitrary data, expressed in an ontology, and then infer various characteristics about the data. Such inferences include whether the data is in fact valid according to the ontology in which it is declared; how the data, as a series of statements, may relate to other statements (by denying, confirming, extending or refining those statements); and what sort of relations exist in the data itself (where such relations may include instantiation, generalisation, specialisation, composition and attribution).
As an example, a minimal ontology which conforms to the definition above could include the following set of concepts, properties and individuals:
The individuals can be assigned properties as follows:
This ontology now expresses seven statements about the world in a semiformal way—formal enough that an algorithm could be devised to process the statements in some way. Compare this with the range of natural language expressions which could express the same statements, both more succinctly and more verbosely: ‘Sally owns two pets, a grey cat called Samba and a blonde dog called Fido’; ‘There exists in the world a person called Sally’; ‘There exists in the world a cat called Samba’, and so on. As a minimal ontology, it also establishes the primordially ontological distinctions between classes, properties and individuals. In practice these distinctions, while on the one hand being crucial for the performance of current state-of-the-art knowledge representation systems, are on the other not always very clear-cut or easy to disambiguate.
These constructs are, then, sufficient to develop a broad definition of an ontology. It is possible to add further constructs, such as the ability to specify subsumption relations between concepts (e.g. ‘Person’, ‘Cat’ and ‘Dog’ are all sub-classes of the more general concept ‘Animal’). The subsumption construct in particular is integral to providing reasoning services over an ontology, which enables inference about the relationship between two objects in the world (e.g. Sally, Fido and Samba are all instances of the concept ‘Animal’). As one of the aims of OWL is to enable these kinds of inferences, it includes the subsumption construct. However under this broader definition of ontology, it is also to include two of the more common formal models for knowledge representation: the relational database and XML Schema. As one of the purposes of the semantic web, and OWL in particular, is to provide a general way to express any data, it is useful to have just such a broad working definition for the time being.
Networked ontologies—towards the web of meaning
A key goal of ontologies is that they are shared, reuseable repositories of data. In the short history of the semantic web a large number of ontologies have been developed for a range of fields and disciplines. Some of these ontologies define generic concepts, so-called ‘upper-level’ or foundational ontologies. These are designed to be applied across many or all domains, and might include concepts such as Process, Object, Time and Space. Others are quite specific to a given domain, such as the life sciences or linguistics. Upper-level ontologies can be incorporated or imported into more specific ontologies, which can be imported by other ontologies again—forming a lattice-like network of interconnected concepts. Ontologies that import other ontologies can also reuse their conceptual definitions, analogous to the world of object- oriented programming, where programming structures are reused in a similar fashion (Booch et al. 2007). This is one way in which concepts and data can be put towards purposes their original authors would not have envisioned. However, this relies on explicit directives from the authors of the importing ontology, who also take responsibility for the explicit conceptual relations and translations they establish between their own and the imported ontology.
In other contexts, two ontologies which have been independently authored often need to be integrated, translated, aligned or merged. Developing points of connection between two ontologies can be a time-consuming and error-prone task, particularly if the ontologies are large—containing many concepts, relations or individual data records. A specific sub-disciplinary area, ontology matching, has been established to find automatic means of associating concepts from multiple ontologies. In addition to the explicit authoring of connections between ontologies described above, ontology matching holds promise for the explicitation of otherwise implicit connections between ontologies. Together these two approaches make it possible to envisage a global knowledge base—one of the declared aims of the semantic web. Chapter 3, ‘The meaning of meaning’, distinguishes several specific ontology matching approaches; in spite of these distinctions, though, the common underlying feature of these algorithms is the production of a set of individual concept matches. This set, referred to as an overall ‘alignment’ of the ontologies (Shvaiko and Euzenat 2005), can in turn be used to generate a translation from concepts in one ontology to concepts in another. Ignored in this translation process is the general degree of fit between the ontologies—how their overall conceptualisations are commensurable.
The question of commensurability—weaving ontologies together
Commensurability as a concept originates in the field of geometry, meaning ‘of common measure’. Wikipedia (2009), for example, defines this mathematical usage as follows: ‘If two quantities can be measured in the same units, they are commensurable.’ Kuhn introduces the term to talk about scientific paradigms:
The hypotenuse of an isosceles right triangle is incommensurable with its side or the circumference of a circle with its radius in the sense that there is no unit of length contained without residue an integral number of times in each member of the pair.
The incommensurability of these quantities does not mean one cannot be derived from the other however. In these two cases, hypotenuse = the root of 2 × side and circumference = 2 × PI × radius express the relations of these quantities. Since in both cases there is a residue, i.e. the equation does not result in an integer, the quantities are incommensurable (Kuhn 1970, p. 189).
Kuhn (1970) makes use of commensurability as a metaphor for how scientific theories, ‘conceived of as sets of sentences, can be translated without residue or loss’. Following Kuhn, the term ‘commensurability’ is used—in place of synonyms like compatibility, congruence or consistency—to connote a deeper level of cultural perspectival alignment between knowledge systems, while allowing for surface-level differences, such as differences in nomenclature. When faced with two matching ontologies, for instance, commensurability suggests there exists some deep conceptual equivalence between them, even if there are no shared terms or concepts. By contrast, incommensurability suggests substantial differences between their underlying cultural conceptions—differences requiring greater effort to effect translation between them. This study presents a similar argument to that of Kuhn’s Structure of Scientific Revolutions: that semantic web ontologies and other formal representations of knowledge are not always commensurable, and that some form of social negotiation is needed to effect translation when this is the case.
Like scientific paradigms, ontologies can be treated as holding a particular orientation towards the slice of the world they describe. Such an orientation bears any number of assumptions which are properly ontological in the philosophical sense—assumptions about how the world is, derived from the cultural backdrop in which the orientation is formulated. Together these assumptions form the epistemic conditions under which ontologies—of the semantic web kind—can be developed. To give a hypothetical example, which is explored further in the work as a case study, two separate ontologies could be developed to describe the world of documents. The first ontology uses the term Author, while the second ontology uses the alternative of Collaborator. Authors are people specifically engaged in the creation of the document—they write the text, capture and insert the images, structure the document and so on. Collaborators have a looser relationship—they may edit, publish, translate, typeset or perform any number of other activities in relation to the document. At this stage—without further knowledge or recourse to context—it is possible to interpret the difference in at least two ways. On the one hand, the difference could be viewed as contingent and accidental—a question of near-synonymic variants. In the second case, a more general term was chosen, which includes the specific term of the first—all Authors are also Collaborators. On the other hand, the difference could also mark a more fundamental ontological difference. Here, in the second ontology, there is no suitable translation for Author. Instead Collaborators simply collaborate to create a document— which could mean writing it, editing it, typesetting it, and so on. Indeed, possibly the concept of authorship is explicitly denied; there are no Collaborators bearing the special distinction of authorship. The first interpretation suggests there is in fact some underlying commensurability between these ontologies, in spite of the different terms chosen. They share the same view of the world, in which documents have both Authors and Collaborators, and Authors are particular kinds of Collaborators. The second interpretation instead suggests that at least in relation to these particular concepts of the two ontologies, the question of translatability is ambiguous. Consequently, commensurability is a less settled question, requiring at the very least further supplementary information.
According to the literal meaning, it could be argued that all knowledge systems, insofar as they employ different conceptual schemes, are trivially incommensurable. In the sense used here, however, commensurability is a question of degrees rather than kind—what matters is the extent of difference and, by extension, the cost, time and effort of translation between those systems. To assess this means going beyond the explicit representations of the systems themselves, inferring something about the implicit background network of assumptions which underpin them—variously termed their ‘world views’, ‘conceptual schemes’, ‘paradigms’, ‘epistemes’, ‘gestalts’ or ‘frames of reference’. This study aims to demonstrate that using the metaphor of commensurability is a helpful way to conceive of both the explicit and tacit differences in the design of knowledge systems; helpful insofar as it provides practitioners with ways of identifying and bridging those differences—or, just as importantly, identifying when such differences are not practically translatable. Here, incommensurability does not imply a slippery slope into relativism or solipsism—a world in which knowledge systems, no less than the cultures that construct and use them, forever remain trapped in their particular hermetic conceptualisations. On the contrary, proper analysis of ontologies can lead to productive insights into the sorts of differences between them, and whether such differences can be readily reconciled in a given situational context.
The question of commensurability is directed towards the same sorts of problems identified by field of ontology matching. Ontology matching approaches seek to develop algorithms to match the terms of two or more ontologies, based on exploitation of terminological similarities. As discussed further in Chapter 3, ‘The meaning of meaning’, concept-by-concept matches generated by these approaches are a necessary but insufficient means of solving certain problems of ‘semantic heterogeneity’ (Shvaiko and Euzenat 2005). The ontology commensurability framework developed here is intended to augment these approaches by considering translation from a semantic holistic perspective—where not only individual conceptual matches but overall schematic commensurability can be assessed.
Towards a framing of semantics
The semantic web makes bold claims about solving problems of system interoperability—a ‘silver bullet’ solution, in effect, for an industry in which software incompatibilities, project failures, patchwork solutions and ‘semantic heterogeneity’ are sources of significant costs (Shvaiko and Euzenat 2008). Moreover it provides a means for weaving together the rich tapestry of existing data on the internet, by providing transparent means for making the structure of that data explicit. A subsidiary discipline has developed, ontology matching, which has sought various algorithmic solutions to the problem of integrating related ontologies. Here, we argue that translation in some contexts needs a holistic regard for the general cultural conceptualisations underpinning ontologies, which can usefully augment concept-by-concept matching algorithms. The Kuhnian term commensurability has been introduced in order to describe the overall degree of fit between two ontologies, assessed across a range of cognitive, social and technical dimensions. The next chapter explores some of these dimensions as they are understood with various academic discursive fields, and how they variously understand and articulate the meaning of meaning.
References
Arunguren, M.E. technical reportOntology Design Patterns for the Formalisation of Biological Ontologies. University of Manchester, 2005.
Ashkenas, R., Ulrich, D., Jick, T., Kerr, S. The Boundaryless Organization: Breaking the Chains of Organizational Structure. San Francisco: Jossey-Bass; 1995.
Berners-Lee, T. Notation3 (N3): A Readable RDF Syntax. http://www.w3.org/TeamSubmission/n3/. 2009. [(accessed 29 October 2009)].
Berners-Lee, T., Hendler, J., Lassila, O. The Semantic Web. Scientific American. 2001; 284:34–43.
Booch, G., Maksimchuk, R., Engle, M., Young, B., Conallen, J., Houston, K. Object-Oriented Analysis and Design with Applications. Indianapolis, IN: Addison-Wesley Professional; 2007.
Castells, M. The Rise of the Network Society. Cambridge, MA: Blackwell; 1996.
Davidson, D. The Essential Davidson. Oxford: Oxford University Press; 2006.
Foucault, M. The Order of Things: An Archaeology of the Human Sciences. New York: Vintage Books; 1970.
Gennari, J.H., Musen, M.A., Fergerson, R.W., Grosso, W.E., Crubézy, M., Eriksson, H., Noy, N.F., Tu, S.W. The Evolution of Protégé: An Environment for Knowledge-Based Systems Development. International Journal of Human-Computer Studies. 2003; 58:89–123.
Gruber, T.R. A Translation Approach to Portable Ontology Specifications. Knowledge Acquisition. 1993; 5:199–220.
Gruber, T.R. Toward Principles for the Design of Ontologies Used for Knowledge Sharing. International Journal of Human Computer Studies. 1995; 43:907–928.
Guha, R.V., Brickley, D. RDF Vocabulary Description Language 1.0: RDF Schema. W3C recommendation, W3C http://www.w3.org/TR/rdf-schema/. 2004. [(accessed 29 October 2009)].
Hacking, I. Historical Ontology. Cambridge, MA: Harvard University Press; 2002.
Kuhn, T.S. The Structure of Scientific Revolutions. Chicago, IL: University of Chicago Press; 1970.
Lenzen, W., Leibniz’s Logic. Handbook of the History of Logic; vol. 3. North-Holland Publishing Company, Amsterdam, 2004. [‘The Rise of Modern Logic from Leibniz to Frege’].
Miller, E., Manola, F. RDF Primer. W3C recommendation, W3C http://www.w3.org/TR/rdf-primer/. 2004. [(accessed 29 October 2009)].
Powers, S. Practical RDF. Sebastopol: O’Reilly Media; 2003.
Shirky, C. The Semantic Web, Syllogism, and Worldview. http://www.shirky.com/writings/semantic_syllogism.html. 2003. [(accessed 25 November 2009)].
Shvaiko, P., Euzenat, J. A Survey of Schema-Based Matching Approaches. Journal on Data Semantics. 2005; 4:146–171.
Shvaiko, P., Euzenat, J., Ten Challenges for Ontology Matching. OTM 2008: Proceedings of the 7th International Conference on Ontologies, Databases, and Applications of Semantics (ODBASE). Springer, Berlin and Heidelberg, 2008:1164–1182.
W3C. Extensible Markup Language (XML). http://www.w3.org/XML/. 2009. [(accessed 29 October 2009)].
W3C. W3C Semantic Web Activity. http://www.w3.org/2001/sw/. 2009. [(accessed 29 October 2009)].
Wikipedia. Commensurability. http://en.wikipedia.org/wiki/Commensurability. 2009. [(accessed 29 October 2009)].