Liam Magee

Ordering the world by relations

After the Second World War, and partially in response to the emergence of a new kind of conflict in which information was to become a central rather than peripheral military asset, the United States embarked on a continuous and unabated course of research into a wide range of computing applications. At one end of the research spectrum was feverish, speculative and sometimes disappointing research into artificial intelligence, conducted by technology-oriented institutes like MIT. At the other end, companies and organisations like IBM, Xerox, Digital Equipment Corporation (DEC) and the RAND Corporation developed their own more commercially oriented but still highly experimental research incubators, which frequently coordinated with their academic counterparts, often hiring bright PhD candidates with a firm eye on commercial applications. Government departments, particularly those associated with the military, often engaged researchers on various diverse computing projects, including cryptography, cybernetics, game theory and large-scale networking (Ceruzzi 2003). It was the work of the Defense Advanced Research Projects Agency (DARPA) in the 1960s which gave rise to the first widespread computing networks, forerunners of the modern internet. Frequently more prosaic areas of research, like networking, user interface development, typography and operating systems, yielded long-term and substantial gains. Attempts to emulate concepts as nebulous and little understood as human intelligence repeatedly hit low-lying hurdles—in the same period Chomsky was demonstrating just how complex one area of cognition, language acquisition and use could be (Chomsky 1965).

The rise of large commercial organisations operating over international territories increased the imperative to develop technologies for managing the expansive quantitative growth of information (Lyotard 1984). In the 1960s, as computers grew in processing power and storage capacity, different methods were developed for managing volumes of structured information. Principal users of these technologies were the banking and insurance industries, for whom the need to provide reliable and systematic information on customers and transactions was paramount. Data storage at this time used a network or hierarchical model, where data records were linked in a parent–child relationship. Navigating from parent records (for example, from a customer) to children (to the customer’s transactions) was relatively simple programmatically. However, the ad hoc aggregation of records based on relations which had not been previously defined in this manner—for instance, a request for all successful transactions within a given period—was time-consuming and computationally expensive. An industry group, the Database Task Group, comprising a number of leading technology companies, but notably excluding the market leader, IBM, proposed a so-called ‘network’ model in the late 1960s. Both this approach and IBM’s subsequent alternative hierarchical model nonetheless continued to suffer the same limitations of performance and feasibility (National Research Council 1999).

In 1969, in response to these limitations, an IBM employee, Edgar Codd, developed a rich algebra and calculus for representing data using what he termed a ‘relational model’ (Codd 1970). The model comprises several key concepts: relations, attributes and tuples—concepts which became known to designers and users of databases and spreadsheets under more familiar monikers of tables, columns and rows. Although this paper essentially proposes the application of set theoretical constructs to data models, it has a practical purpose—it explicitly aims to provide a better model for ‘non-inferential systems’, unlike the early progenitors of research into artificial intelligence. The principal benefits of this model were to provide a sufficiently abstract series of informational constructs which, once standardised, could allow for true data independence, and a rigorous axiomatic system which could ensure data consistency (although as subsequent database administrators would discover, many other factors intrude on the problem of maintaining a consistent data set). IBM itself was slow to follow up on the promise of its own innovation, and other elaborations of it which followed from its laboratories in the early 1970s. The company did develop a prototype system based on the relational model, called System R, but this failed to be adopted by IBM commercially (National Research Council 1999). It did however publish its research on the relational model; and in 1977, after reading some of this research, three entrepreneurs founded what was soon to become the largest database company in the world, and one of the largest in the information technology sector (Oracle 2007). Like Microsoft, Oracle’s growth through the 1980s and 1990s was staggering—ironically, both companies profited from costly miscalculations at IBM. Though IBM was soon to catch up somewhat with its own commercial relational database system, Oracle’s success was largely driven by, and indeed cleverly anticipated, the unrelenting drive of large organisations to manage enormous data sets. This was complemented by the increasing affordances of ever cheaper and more powerful hardware.

One of the principal advantages of the relational model is the provision of a standard and well-defined query language, Structured Query Language (SQL). In spite of fierce competition among database vendors, and a corresponding emphasis on product differentiation, SQL quickly became—and remained—an essential part of all modern database systems (Date 2007). SQL was ratified as a US ANSI standard in 1986, and an international ISO standard in 1987, ensuring a minimal base compliance for manipulating and querying data across rival systems (Wikipedia 2009). Unhappily for users of these systems, vendors like Oracle, IBM and Microsoft continued to extend the subsequent sequence of standards with various ‘proprietary’ extensions. The preponderance of industry support for SQL demonstrates that even in the heavily competitive and nascent database software industry, vendors were prepared to trade off short-term competitive advantage against the longer term positive network externalities of a larger marketplace built around selective feature standardisation and differentiation (Katz and Shapiro 1985). As SQL became an integral component of modern computing degrees and certification processes, it demonstrated that English-like formal languages could achieve widespread adoption, with the significant incentive of a burgeoning job market in the 1990s and early part of the twenty-first century. Moreover it was no longer just the domain of large organisations— small-to medium-sized businesses and even keen individuals increasingly adopted relational databases as the basis for data management, usually with convenient user interfaces overlaid. The permeation of the relational data model into all aspects of computing culture represents the overwhelming and ongoing success of this paradigm for managing structured data (National Research Council 1999). Indeed, in a very direct sense, the relational model represents the culmination and fruition of the modernist dream to order and organise knowledge systematically.

Early threads of the semantic web

The evolution of the semantic web, and its methods for representing knowledge, follow a decidedly different route. Since its earliest developments, when machines first replaced human ‘computers’ (Davis 2001), theorists and philosophers had been pondering the question of whether—and how—artificial intelligence was possible. The Turing Test, developed as early as 1950, suggested a range of criteria for determining how machine intelligence could be tested (Turing 1950). Research in artificial intelligence was to proceed down numerous different lines over the remainder of the century, in pursuit of the often elusive goal of emulating human behaviour and cognition. Frequently denigrated for not realising its lofty ambitions, many artificial intelligence innovations nevertheless filtered down from these comparatively abstract areas of research into everyday practical technologies. One area that received particular attention was the problem of modelling or representing knowledge—an essential step towards building computational equivalents of memory and reasoning processes.

In the 1960s Quillian (1967) pioneered the idea of ‘semantic networks’—graphs of nodal concepts connected by associations, a precursor to neural networks. These were followed by more detailed models, such as semantic frames (Minsky 1974). Semantic frames added the notion of ‘slots’ to concepts, where their attributes could be stored. Abstractly, both attributes and relations could be considered as properties of a concept, distinguished only on the basis of whether a property could take a data value (attribute) or object value (relation). Semantic network and frame approaches were the basis of a number of early expert systems, with Minsky’s proposals in particular galvanising interest in artificial intelligence circles about the possibility of engineering computational approximations to human cognitive processes (Sowa 2000). From another angle, there were various endeavours to instrumentalise theorem proving through use of declarative or logic programming languages (Colmerauer and Roussel 1996). For whatever reason—perhaps because it was easier to separate knowledge bases from the procedures which reasoned over them—logic programming approaches were to remain a niche market. On the other hand, it soon became apparent that the sorts of things which constitute an ‘association’ or ‘relationship’ between concepts need greater semantic specificity than existing semantic network or frame approaches allowed. In 1979 a new system, KL-ONE, emerged with more expressive semantics, where the kinds of relationships between conceptual nodes is explicitly stipulated (Brachman and Schmolze 1985). This was a step closer towards greater levels of interoperability between multiple knowledge systems; however, it was still possible to interpret constructs differently across different systems.

Over the course of the 1980s and 1990s, researchers began developing restricted forms of logic for representing knowledge (Sowa 2000). ‘Terminological’ or ‘description logics’, as they became known, were fragments of first order logic with specific constructs for representing concepts, properties and individuals (Brachman et al. 1991). Significantly, description logics were directly derived from Tarksi’s work on model theory, discussed earlier, providing unambiguous interpretation of the effect of logical operations within conforming systems (Nardi and Brachman 2003). For example, if a concept is stipulated as being subsumed by two other concepts, its extension—the objects denoted by the concept—must be interpreted as the union of the objects denoted by the parent concepts. For systems implementing these semantics, and for users of these systems, this feature ensured consistency in the handling of queries, and remedied many of the derivative problems which emerged in the implementations of earlier models (Nardi and Brachman 2003).

Still, at this stage knowledge systems were invariably small scale— much too small to capture the many background facts assumed to sit behind the kind of commonsense reasoning humans typically undertake. In 1984 Doug Lenat began a project called Cyc, which was intended to contain all of the facts which constitute an ordinary human’s understanding of the world. Development of the Cyc knowledge base is still ongoing, part of the commercial intellectual property of its owner, Cycorp, and represents a substantial undertaking to codify knowledge under the auspice of a single overriding conceptualisation (Lenat 2001).

Proposals for the semantic web built on the work of description logics even more explicitly. Unlike Cyc, the vision of the semantic web involves many authors and conceptualisations, linked together by a common model-theoretic foundation in RDF and OWL, explicit references and shared pointers to web resources. The explicit design goals of the semantic web were to provide a very general mechanism by which knowledge could be represented, shared and reasoned over computationally. The first published version of OWL came with different description logic variants, with different levels of expressivity and tractability—as the logic becomes more expressive, there are fewer guarantees that in the worst case reasoning problems are tractable, or can be resolved in finite time (Levesque and Brachman 1987). Though its precursors Ontology Inference Language (OIL) and DARPA Agent Markup Language (DAML) were first motivated by a combination of academic and military research, OWL itself quickly became sold to the broader web community as a facilitator for a new range of ‘intelligent’ services—process automation, e-commerce, system integration and enhanced consumer applications (Berners-Lee, Hendler and Lassila 2001; Fensel et al. 2001). RDF had similar, if more pragmatic, origins as a language for data markup, with less emphasis on automated reasoning. It was, however, also motivated by the need to model information more flexibly than the highly structured models of preceding generations of data technology—notably the relational model—would allow. The document-centric nature of the World Wide Web suggested that significant amounts of information could not conform to the strictures of a relational view of the world. While being compatible with existing structured information sources was one constraint on the design of RDF, so too was the need to permit modelling of flexible, semi-structured, document-like and inter-connected data. The involvement of the World Wide Web Consortium (W3C) ensured that the semantic web architecture technically and philosophically built on the foundations of the already—by the late 1990s, when the first formal semantic web technical recommendations were drafted—hugely successful precursor of the World Wide Web.

Shifting trends or status quo?

Broadly, then, the semantic web can be seen as the consequence of three dominant broad trends in information technology and management over the later decades of the twentieth century. First, organisations had grown in size and, commensurately, the burgeoning fields of business, management and information studies had encouraged the use of disciplined techniques for obtaining greater predictability over key variables in organisational operations. From the 1970s onwards the explosive quantitative growth of data, the availability of and demand for computing resources for storing and processing it, and the increasing awareness of the opportunistic value of analysing it led in turn to enormous investments in scientific research and development, as well as considerable financial speculation in the data management industry. This was primarily oriented around the relational data model—although other data storage models continued to be prevalent in specific fields, the general applicability of the relational model led to its near ubiquity as a mechanism for storing structured information.

Meanwhile, a second, less perspicuous trend took place in the ongoing research in artificial intelligence and knowledge representation, which in turn made feasible a well-defined and consistent notation for describing facts and permitting sophisticated inferencing operations. In fields with large numbers of concepts and very large amounts of data, such as medical, financial and military applications, the need for expert systems had long been evident.

Finally, the rise of the World Wide Web—the third, and arguably most disruptive of these trends—also brought new applications and therefore greater commercial relevance for deductive reasoning systems. Where the relational model had typically been used in intra-organisational (or even intra-departmental) settings, the advent of a global network, and the academic, economic and political advantages to be gained through its exploitation, made more evident than ever the need to supply unambiguous definitions of data through a highly expressive formal notation. The application of deductive reasoning to information from a myriad of sources gave rise to new problems of trust, proof and authentication, but also provided the tantalising prospect of unprecedented data being accessible to reasoning algorithms. The development of RDF and OWL was driven by the competing demands of providing notations simple enough, on the one hand, to be used and developed by software engineers and web developers untrained in knowledge representation, and expressive enough, on the other, to permit the kind of deductive power envisaged by the pioneers of artificial intelligence, and indeed by their precursors, the foundational logicians.

In spite of these evolutions, it would be premature to conclude that the semantic web is in the process of replacing the relational database. In fact the relational model has proved remarkably resilient in an industry recognised for inevitable if not always planned technological obsolescence. The massive commercial database industries still dwarf the largely academic and entrepreneurial world of the semantic web, and considerable work has been devoted to building bridging technologies between the respective formalisms—to promote, in fact, further use of the semantic web through connections to existing relational repositories of data (Malhotra 2008). Meanwhile, many in the broader web community are also now examining alternatives to the semantic web itself, suggesting a more complex picture marked by overlapping, shifting trends rather than any clear pattern of technology phase-out (Khare and Çelik 2006). At this stage it is more likely that both relational databases and semantic web ontologies will continue to be developed—making the question of their commensurability, discussed in the section below, highly pertinent.

Systems of knowledge: modern and postmodern

These two models—the first, representing the fulfilment of modernist logicism, the second, a postmodern response and would-be successor—have special interest within a broader historical trajectory of formal systems. Relational systems hold, as mentioned above, a dominant position in the market of information systems. By one indicator—worldwide ‘total software revenue’ (incorporating licences, subscriptions, support and maintenance)—the relational database market grossed an estimated US$15.3 billion in revenue in 2006, US$16.8 billion in 2007 and US$18.8 billion in 2008, continuing to show strong growth rates in spite of a global economic downturn (Gartner 2007, 2009). While this figure includes immediately subsidiary revenues accompanying software, it does not include the many development and maintenance tools, services, related or derivative systems which depend on relational database systems—a value likely to be much higher. Moreover, even an eventual meteoric rise of the semantic web does not imply the eclipse of the database industry, since, as suggested above, logical semantic web data structures like ontologies can be physically stored in a relational database; however, as the analysis above shows, ontologies do present a rival conception of knowledge representation at a logical level, and a separate tradition in a historical sense. Although they represent different logical formalisms, with few ontological commitments, nevertheless they can be distinguished on the basis of certain minimal and abstract assumptions. Perhaps the most contentious of these concerns the use of so-called ‘closed’ versus ‘open’ world assumptions, a distinction which has received substantial attention in the literature on logic, databases and knowledge representation (Reiter 1987; Sowa 2000). The following review examines this distinction in greater detail.

Knowledge systems in social context

To round out the discussion of knowledge systems, the following summary also teases out what was an underlying thread in the account above—the relationship between technological innovation and broader social shifts. These shifts exhibit a complex network of causal relationships to the general processes of technological design, development and innovation, and hence to the question of commensurability between rival systems that emerge from these processes. These relationships, tenuously charted in this study, are more explicit in the studies that follow.

In the last quarter of the twentieth century the development of formal knowledge systems has been precipitous. The preceding discussion showed how this ascent was premised on the foundational work in mathematical logic in the late nineteenth and early twentieth centuries. Leibniz’s dream—of a single symbolic language in which thoughts and argument could be conducted without ambiguity—was a constant motif throughout the evolution of this tradition. Symbolic logic, then, represents a pristine formal component of a long-ranging historical epistemological ideal, while an endless accumulation of ‘sense-experience’ supplies the matter. The semantic web represents a modern-day recasting of this ideal, in which precise agreement about meaning forms the underlying substrate for sharing information and deducing inferences. It receives its most emphatic expression from Ayer, who envisioned philosophy and science of ardent empiricism: ‘The view of philosophy which we have adopted may, I think, fairly be described as a form of empiricism. For it is characteristic of an empiricist to eschew metaphysics, on the ground that every factual proposition must refer to sense-experience’ (Ayer 1952, p. 71).

The unfolding of this tradition in the account above describes three key phases—classicism, modernism and postmodernism. These phases show an increasing impulse towards the development of ‘taxinomia’—indexable, searchable and interoperable knowledge systems which span from the globally networked enterprise down to the fragmentary databases of commercial and social interactions managed by individual consumers. By tracing this tradition through a purely intellectual history, it is possible to suggest several causal factors internal to the tradition itself: the production of particular fortuitous mathematical results, or a sense of exhaustion with the preceding metaphysical speculations of Kant and Hegel, for example. It is equally possible, though, to plot lines of concordance between this intellectual history and broader transitions in economic and political history. Is it purely fortuitous that the search for logic formalisms coincided with a reciprocal drive towards standardisation, in a host of technological, communicative and legal fields, that is related to modern capitalism—specifically, of its relentless need and demand for predictability and efficiency? For Foucault, the modern taxonomic impulse originates alongside the great social and political shifts of the Enlightenment:

By the time of the emergence of formal logic in something like its rigorous modern form in the nineteenth century, the world was also undergoing a period of rapid economic expansion, industrialisation, scientific endeavour and technological innovation (Hobsbawm 1975). Already the opportunities of standardisation were being considered in a host of practical contexts—rail gauge standardisation, currency exchange, scientific notation, legal charters and academic disciplinary vocabularies. The counterweight to international and inter-corporate competition was the beneficial network externalities—greater efficiency, information transparency and intelligibility—these standards would bring. Since these first standards emerged, their growth has been rapid— the ISO website alone currently advertises 17,500 separately catalogued standards (ISO 2009).

While standardisation might rightly seem, then, to be an inextricable feature of modernity, coupled with economic globalisation and cultural homogenisation, it can equally be argued that capitalism also harbours countervailing trends towards systemic differentiation. Most notably in the case of the quintessential capitalist organisation, the company, product or service differentiation forms the foundation for market share, profit, and thus for increasing shareholder value. To take one metric of the extent of differentiation, at the level of invention and innovation: the US Patent Office has filed over seven million utility patents alone since 1836 (US Patent and Trademark Office 2009), with an average rate of increase in the number of patent applications between 1836 and 2008 of 23.5 per cent. There were 436 patent applications filed in 1838, and 158,699 applications filed in 2008, an overall increase of 36,399 per cent over 170 years (the raw data has been taken from the US Patent and Trademark Office (2009), while the percentile calculations are my own). Whatever explanation of drift towards standardisation can be drawn from modern capitalism, there is an equivalent burden for explaining a similar level of hyper-activity towards proprietary protection of intellectual capital and assets.

Equivalent, if more tenuous motives for differentiation can be found in other organisational types—political affiliations, methodological distinctions and sublimated competitive instincts exist in government, scientific and academic institutions as much as in corporate ones. The development and coordination of knowledge systems—formalised representations of meaning—has its origins, in one side of modern capitalism, in the impulse to order, organise and predict. The proliferation of multiple systems represents, then, another facet of capitalism—the need for differentiation and competition. Schumpeterian ‘creative destruction’, describing the process by which capitalism continually cannibalises its own monuments with successive waves of technological and procedural innovation, captures something of these apparently contradictory impulses towards both standardisation and differentiation at the level of systems of meaning. However, as this and the following studies show, other, less tangible vectors can also be seen influencing the mutations of these systems.

At this stage, though, it is perhaps sufficient to draw out the coincidental tendencies between the specific phenomenon of the emergence of knowledge systems and the much broader chameleonic shifts of capitalism, without pursuing too strong an attribution to determining causes. The following studies bring out other complicating and more fine-grained features of the contexts in which these systems emerge, and of the factors which influence their respective differentiation.

References