Balisage Paper: Putting it all in context

Large Scale Information Sharing

Large-scale and even global information sharing has become a major challenge to the world's governments, both internally and internationally. Poor responses of emergency services in the wake of the 9/11 and Katrina disasters have made information sharing a priority for the U.S. Government [GAO385, DODIS07], and as the single largest purchaser of information technology services in the world (US$ 65 billion proposed for FY2008 budget [OMBAP08]), it is expected that the U.S. Federal Government will require management of vast quantities of information and interoperability across an extremely diverse set of technology platforms.

We were interested in the applicability of the ISO 13250 Topic Maps standard to this and similar problems, a discussion which threw up a key question: To what extent is Topic Maps applicable in large-scale information sharing environments such as the Internet? Indeed, does it scale? Furthermore, what specific commercial opportunities exist for technology vendors in large-scale information sharing environments? And finally, which technologies support large-scale information sharing?

Scope and Context

Scope allows multiple views on a particular topic map as well as the filtering and adapting of information contained in a topic map based on the needs of the user [PEPP00, PEPP02, PARK03].

Hence the Topic Maps specifications [ISOTM, XTM, TMDM] provide a mechanism in the scope construct for representing the idea of context. Accordingly the purpose of scope is to represent "the context within which a statement is valid. Outside the context represented by the scope, the statement is not known to be valid. Formally, a scope is composed of a set of topics that together define the context. That is, the statement is known to be valid only in contexts where all the subjects in the scope apply" [TMDM].^[2]

Related questions that have been discussed were:

And concerning general modelling questions:

Furthermore the discussion explored the question of how one might represent context using associations rather than the scope construct, as it seemed potentially to be more dynamic and adaptable than what is syntactically allowed by [XTM]. In discussing a topic map's ability to capture context explicitly enough to be useful outside a small community using either approach, the discussion naturally gravitated to the question of what exactly is meant by "context".

What Is Context

Context is one of those concepts that most people feel is essential to understanding something, but which sparks great debate when one attempts to define it. The main issue in the scientific and technical communities revolves around whether context should be regarded as subjective or objective [MENZ99]. Once a decision is made regarding one's view on context, there is the question of how to represent it computationally.

Subjective context is what most people understand context to be. It is experiential in nature and includes the total of meanings (ideas, assumptions, preconceptions and associations) that are related to a thing or event and which serve to influence our attitudes, judgements, perspectives, and general knowledge of something [AKMA00]. While Menzel argues that subjective context always implies an "outer context" which exists outside any given context, recursively, that enables individuals or intelligent agents to adopt a broader perspective to reason about the original context, Bohm's theory supports the idea that context and those perceiving it are part of a unified whole in much the same way as the observer and the observed are related in relativity theory and quantum theory [BOHM80].

The ability to leverage separate views of a given situation is often useful or indeed necessary, but we need to be aware that these separations are not necessarily accurate or complete reflections of reality. They must be seen as simply a useful mechanism in a particular domain or situation to aid in understanding it.

Objective context, interpreted as "outside the skull of any reasoning individual", is much more directly related to the works of Barwise, Perry, Seligman and Devlin on Situation Theory and Information Flow Theory [MENZ99]. The focus of the objective context is the setting in which the objects and available information reside and also encompasses the relationships between them. The focus of Menzel's work [MENZ99] is to define a logic for reasoning within objective contexts such that a given statement is true in the portion of the real world described by a particular situation rather than deal with the complexity of the perceptions and assumptions involved in how the participants view a given interaction. In order to achieve this, it is necessary to create a definition of context which exists outside these subjective perceptions and which will remain invariant for all the participants.

Clearly, both definitions of context are attempting to capture information about the environment in which an interaction takes place in order that some level of shared understanding between the participants be established. However, there is a considerable difference in the types of information that would be included in each definition. Before determining which perspective would be the most useful in terms of sharing information and knowledge with the Topic Maps standard, it is necessary to consider what the goals really are of including contextual information in specific topic maps.

Context, Fragmentation and Communication

By any account, context is an essential part of successful communication. In practice, it determines how much information must be explicitly stated between the participants and how much information is shared between the participants. Akman provides the example of late arrival to a meeting [AKMA00]. In order for the person to understand and effectively participate in the ongoing discussion, they must establish some of the background on what has already happened during the meeting.

Establishing an objective context would attempt to capture the situation of the meeting, independently of the attitudes, beliefs and motivations of the participants because the goal of the objective context is to establish "complex, structured pieces of the world" [MENZ99]. It is essentially the view that an objective outside observer might have of the situation. For example, some of the following information would likely be aspects of the objective context:

In respect to the classic journalistic Five W's^[3], a truly objective observer attempting to capture an objective context would establish each of the "who", "what", "where", "when" and "how", but would be unable to capture any meaningful understanding of the "why" beyond any surface remarks made during the meeting, e.g. "We are here today to finalize the implementation plan for Project X". While this may make any given context more easily suited to use by autonomous agents and deductive reasoning, it may not accurately reflect the complete reality of the situation.

Therefore, the real question is how accurately one wants to capture "reality" in a particular context. Akman cites several examples where the actions and behaviour in a particular situation reflect the dramatically different views of reality held by the participants [AKMA00]. Attempting to understand these interactions from an objective perspective would likely result in confusing and conflicting interpretations, with the person trying to understand the situation asking themselves, "now, why did they do that?". Searching for the answer to that particular question would truly result in insights and new knowledge for the person who asked it.

In contrast, establishing the subjective context of the same meeting would involve some level of understanding the participant's mental models. An individual's mental model include all of their experiences, beliefs and assumptions about the world. Mental models are important to establishing subjective contexts because they both reflect and shape an individual's understanding of reality [SENG90], i.e. their concepts of what is true and what is not. For the above meeting, knowing the motivations, intents and attitudes of the participants may dramatically alter one's understanding of actions taken and information disclosed during the meeting. This information provides the answers to the "why" question seemingly ignored by the objective context.

Menzel's assertion that subjective context correlates to logical theories made up of sets of propositions [MENZ99] exactly echoes Bohm's earlier theory on fragmentation, itself developed as a way to explain some of the inconsistencies he found in relativity theory and quantum theory. According to Bohm, fragmentation is a result of holding to a theory or set of propositions as being a true reflection of reality rather than viewing it as insights into a given domain and "which are neither true nor false but, rather, clear in certain domains, and unclear when extended beyond these domains" [BOHM80]. These insights are shaped by our mental models, and must therefore be subjective. According to Bohm, the true reality can never be grasped explicitly; it is an implicit whole that is exposed through each subjective context. Recognition of context as subjective is a key to greater understanding of a particular domain, because it is the interactions and boundaries of existing contexts that provide new insights of understanding into a particular domain.

It would therefore seem that the subjective interpretation of context cannot be ignored. Topic maps are a communications medium to encapsulate and convey information, understanding and knowledge about a particular domain. They are ultimately created by humans - whether directly authored by people or created automatically by software created by people - and therefore reflect a subjective view of that domain shaped by the implicit factors inherent in any form of communication. This means that they will be vulnerable to many contextual ambiguities "born out of a misplaced belief on the part of the author that the reader has access to the author's collective dimension" [AKMA00].

This vulnerability is potentially problematic because of the merging process defined by the Topic Maps specification: unlike a document or other work created with a particular intention and with assumptions about context shared between author and readership, information in a specific topic map can find itself far outside the context the original author had in mind. Consequently, any implicit or shared context will be lost. Since context is so critical to effective communication, determining how to best capture it is also critical to creating effective Topic Maps.^[4]

It may well be that there are two, arguably conflicting, views on the role of topic maps. Is a specific topic map intended as something to be studied for greater understanding of a given domain, or is it intended to capture a body of information to support ad-hoc queries to answer questions about a domain—limited only to the information at hand? How this question is answered may have significant implications for the type of information and manner used in expressing context within the instance.

In the first case, the goal is to engage users so that they become active participants in the topic map. In this case, the information in the specific topic map would act similarly to the text in Paul Auster's literature and become "'a springboard for the imagination'" [AKMA00]. The intended use for the topic map is to help the user understand something new about the topic map's domain. Knowing as much as possible about the domain-view of those capturing the information and relationships - including the ontology - of the topic map will greatly enhance the process of cross-domain information sharing. In this respect, support for a more subjective interpretation of context would be valuable.

If, in the second case, the goal of the topic map is essentially to be a sophisticated, multi-dimensional database that is able to respond to queries over its contents, then the level of contextual information captured may not need to be as great, as indicated above. Its goal is not to support deep analysis of a domain, but rather to provide, initially at least, more superficial views of the domain by offering answers to operational needs. Further analysis of the responses may prompt a deeper exploration of the information within the domain of the topic map, but it is not likely to prompt deeper reasoning about the domain^[5]. Therefore, it may be more reasonable to assume a more objective interpretation of context for this second scenario.

In concluding this section, it is worth noting that many of these issues are not the exclusive prerogative of the Topic Maps standard.

Context and Ontologies

Following Bohm's view that there can never be an explicit view of reality as it is and that anyone's individual or shared world-view of reality is simply one way or view of the implicit whole, ontologies must essentially define the boundaries of a particular "domain view", but can never represent that domain as the one, true representation. This view - sometimes referred to as the "information territory" - would seem to confirm that any search for a unifying, fixed upper ontology is necessarily futile. From Bohm's viewpoint, it may be possible to define a more unified view of the world which treats previous views as specializations of the new unified view (e.g. Newtonian physics vs. quantum theory), but which will still have essential differences and distinctions from other views of the same domain [BOHM80]. It does not imply that this new insight is final, however.

In this respect, working with Topic Maps benefits from the fact that thestandard does not require content within a particular topic map to conform to a fixed ontology. Any ontology used is merely a guideline. Therefore, it would seem that Topic Maps has potentially the greatest chance of being useful ways of capturing knowledge without falling prey to Bohm's theory of fragmentation. In his words, "to be confused about what is different and what is not, is to be confused about everything" [BOHM80]. The flexibility of Topic Maps would tend to allow someone with an awareness of fragmentation to be able to view a particular domain and explore the relationships not only between the topics and their occurrences, but also between the different, fragmentary views captured in multiple ontologies. Taken as a device to reason about a particular domain, this could be an extremely useful feature of the Topic Maps specification.

In a scenario where an ontology is used to define context, it becomes critically important to identify the boundaries or context in which a particular ontology applies vs. any other ontology. While this may be straightforward enough when using explicit ontologies, it will obviously prove to be more of a challenge for topic maps using implicit ontologies. In the same way that every software system has an architecture, explicit or not, so too does every topic map have an ontology.

Context Representation

While the two mechanisms for representing context defined by the Topic Map specifications have been discussed above, there has also been considerable work done in the domain of Situation Theory in attempting to define a mechanism for representing context computationally [TIN94, AKMA00, MENZ99]. Much of that work includes, directly or indirectly, the concept of an infon, a basic unit of information [DEVL92], as a way to represent context. One of these systems, BABY-SIT, uses the infon directly as a primitive system object to capture the definition of a situation, or, in parameterized form, a set of possible situations [TIN94]. The motivation for BABY-SIT was to create a system for the development and testing of programs based on a situation theory framework within the domains of linguistics and artificial intelligence.

One of the differences between scope in Topic Maps and infons as used by BABY-SIT is that BABY-SIT infons consist of a relation, one or more objects corresponding to a role in the relationship and a polarity. In this way, they define a context (situation) that can be used as a primitive object to determine relevance of other relationships and situations within the system. While a similar notion could possibly be expressed as a Topic Map query, this construct does not seem sufficiently rich with which to represent context within a specific topic map. However, it could potentially be used as a primitive construct within a Topic Map query. It is also not clear what implications the introduction of a new context representational form may have for existing topic maps.

A Problem of Scale & Centralization

In attempting to examine how the Topic Maps standard could be used to enable information sharing, one of the first questions is how best to address the cross-organizational requirements inherent in information sharing scenarios. In contrast to most knowledge management environments which, by definition, are designed to enable an organization to accomplish its own strategic objectives, the information sharing environment extends outside organizational boundaries and interfaces with external partners in a dynamic manner. The phrase "from external partners...whose role and nature may not be known prior to an event" in the DoD definition is extremely important, because it implies that an organization may not know with whom it will be needing to exchange information before it needs it. In this type of environment, it is not possible to operate a static system under centralized control.

Unfortunately, most implementations of the Topic Maps standard use it, or XTM, as a means for information exchange between proprietary systems that are relatively monolithic and centralized in nature. Information exchanged, using particular topic maps, is integrated into the proprietary system, and that information is accessed either via proprietary APIs, languages or tools that are not generally directly interoperable with other topic map implementations. It is worth mentioning the TMAPI effort [TMAPI] in this respect, however since it is an API, it is specific to the implementation language - in this case Java - and while it will provide portability across compliant applications, it does not provide interoperability with other compliant applications (see [TOWN05] for a more technical discussion of the differences between interoperability and portability, specifically relating to reliable messaging).

The problems with this scenario are that in order to allow a user or application to interact with a topic map, that topic map must be managed by the implementation's software, and it can only be used by that implementation's tools. Most Topic Maps tools are designed to be part of a closed system provided by a single source, and assume that management of all topic map information will be ultimately centralized. Not surprisingly, this approach introduces some potential issues relating to scalability and data ownership in an information sharing scenario as, according to Annex F.5 of the XTM specification, every topic and association in any remotely referenced topic map is added to the local topic map, triggering additional merging and duplicate suppression [XTM].

In many ways, Topic Maps implementations share characteristics of hypermedia systems of the 1980's in that they are closed environments, and they ultimately assume ownership of the content they manage. This issue was one of the main reasons cited by Berners-Lee in writing the software that led to the birth of the World-Wide Web [WRIG01]. It would seem that the same issues could prevent Topic Maps from directly addressing the large-scale information sharing problem - even across a small number of organizations, let alone on the scale of the Internet.

In an information sharing scenario, decentralized control over the information is a necessity. Information exchange between two organizations will mean that both organizations will need to have a shared representation of that information at a specific point in time, but that representation must ultimately remain under the control of the originating organization. It cannot be subsumed into a merged topic map as an understanding of the way the world is. It can only be understood by the consumers of that information that it was valid when it was accessed, and it may not express the same information if accessed again. This highly decentralized model is already present in the World-Wide Web, and is "a key aspect of the Web's success" [BERN06].

Clearly, one of the other key aspects of the Web's success is the underlying architecture, Representational State Transfer (REST), which is designed to support operation and deployment on Internet-scale. Specifically, this includes the ideas of anarchic scalability, to address the needs for the architecture to continue operating under unanticipated load and malformed or maliciously constructed data, and independent deployment to facilitate "gradual and fragmented change" as the underlying hardware and software architectures evolve [FIEL00]. Much of these capabilities are achieved through the open standardization of HTTP, HTML and the URI specifications. Without such a stable foundational core, the Web would not be able to support independent evolution, and interoperability between software and hardware platforms we often take for granted today.

Such a stable platform for interoperability does not yet exist for topic maps. However, many lessons from the Web could likely be successfully applied to the existing specifications in order to allow them to meet the needs of large-scale information sharing, including better cross-domain navigation, explicit information asset ownership, managed access control of those information assets, building netwroks and communities of interest. Such an environment could potentially create "Topic Webs" within a global information space using hypermedia constructs that were the aggregate of information sources supplied by many different organizations and leveraging the highly decentralized model so successful in the Web. Thus, navigation within the map would be handled uniformly through the use of hypermedia links and standardized XML vocabularies leveraging the underlying transfer protocols such as HTTP or its potentially forthcoming successor, Waka^[6] and the views of any given map would be controlled by its publisher, based on the user's context of access, including the topic map's authorization and identity assertions.

Revisiting the Context Problem

When extending the Topic Maps model to information sharing scenarios, adequate capture and representation of context becomes even more important. Not only will the overall information space be much more volatile, the potential cultural and sociological backgrounds of those using the topic maps are likely to be much more diverse than many Topic Maps applications today.

In an information sharing scenario, there are at least three different types of context that will be involved in any interaction. Three types of context are present in any interaction with a particular topic map, and it will be even more important to explicitly capture these in information-sharing scenarios. They are:

This could be achieved by ensuring that the specific context, as represented in Topic Maps by a "scope bundle", can itself be reified as a distinct subject and thus be represented as a distinct topic. The Topic Maps Reference Model [TMRM] certainly points in this direction. Such an approach however brings with it potentially enormous scalability issues, given the number and complexity of such scope bundles and the additional danger of possible recursive contextualization.

In each of the above three steps, authoring, publishing and access, there is the potential for misinterpretation of the relationships within the topic map, because each relationship represents a potential loss of some of the implicit context which contributed to the topic map. However, as the audience gets wider a counter phenomenon is likely to kick in and help regain some balance - that of the "wisdom of crowds": or more precisely, the wisdom of groups, as it is more likely to be groups of users with similar - dare we suggest - contexts, who will share their take on the intended context.

Any of the potential usability and manageability problems inherent in encoding context information into a topic map in the local model will be magnified many times in an information space "whose size can be measured in thousands of terabytes, and the numbers of users in the hundreds of millions" [BERN06]. Therefore, there needs to be a mechanism where, while the necessary contextual information will be available, it should only be referenced as needed. Shared contexts between individuals and organizations must therefore be represented explicitly. The real meaning of "changing" a context (adding or removing references to situations, objects, information or any subjective views) needs to be determined with sufficient clarity so that the resulting large-scale information space represents an improvement on the current Web when attempting to find relevant information.

Commercial Opportunities and Potential Architectures

A highly distributed model for Topic Maps would require similar core, slow-evolving protocols and specifications similar to HTTP, URI and HTML. However, there would be few restrictions on the speed of evolution in both the content and the software implementing those specifications.

The main idea is to separate as much as possible the monolithic view currently present in Topic Maps platforms into the essentials, so that they could evolve independently and provide a more dynamic market for more focused products. It is not hard to envisage multiple Topic Maps engine vendors specializing in particular aspects of performance, scalability, availability and management and maybe specialization around certain user communities. Likewise, authoring and navigation environments would be decoupled from a particular Topic Maps engine, allowing for the market to be developed. Such environments could also address more potential niches in the market through more vendors providing solutions. An increase in technology implementations and proven interoperability between them should lead to increased adoption of the technology and in a much larger way than has currently happened with the existing, monolithic approach.

Leveraging the capabilities of hypermedia and potentially even some of the existing XML vocabularies intended to support interaction such as XForms or XHTML, would mean that clients would not be dependent on proprietary HTTP or SOAP-based APIs in order to author and navigate topic maps. There are many opportunities to leverage the "self-descriptive" and extensible nature of existing XML vocabularies to allow hypermedia documents to completely describe the interaction with a Topic Maps implementation.

Additionally, unlike the Web, such a system designed to support large-scale information sharing would need to be designed with a way to represent the legitimacy principles and requirements specified by each content provider. Legitimacy is the context defining what is considered "fair" behaviour within a community or other common environment [WHIT03, TOWN06]. This notion of fairness can also be expressed in many ways, but in a Topic Maps environment, it would be a good candidate for representation using the context mechanisms described above, and requiring that agents provide compliant implementations of those enforcement mechanisms. Many of the presented ideas align very well with the ideas expressed in the discussion paper on a "Personal Data Services Model" [BROW06].