A lot of research has been done in the area of context-aware computing. Even though, the term context seems often not to be well defined. We attribute this problem partly to the fact that research often focuses on syntactical and technical issues of contextuality and does not take a knowledge level perspective on context. When including the knowledge level, some sort of analysis is required on what aspects need to be modelled. In this paper, we propose the use of an Activity Theory (AT) based approach on modelling components, and outline how it can be combined with the AmbieSense context modelling framework we have proposed earlier.
A major short fall of the research in context-aware systems, and in many other disciplines as well, is the lack of a common understanding of what context is, and perhaps more importantly, what it is not. This shortfall is a very natural one, since no common understanding of what context is and how it is used in the real world exists, it is no surprise that it is hard to agree on the artificial world that IT systems, most often, represent.
Most of the research today has been focused on the technical issues associated with
context, and the syntactic relationships between different concepts. Not so much
attention has been given to context from a knowledge level [
This is the main reason for the approach chosen here. It should be feasible to look at how we can use socio-technical theories to design context-aware systems to supply better services to a user, in a flexible and manageable way.
Context-aware IT Systems are usually designed for specific purposes and with
specific tasks in mind where the system has to support human users. It is used by people
with specific needs and qualifications, and it should preferably adapt to changes in these
needs over time [
This paper is organised as follows: first some background work on the use of context in cognition is covered. Secondly, the knowledge model, including context employed in this work is described. Thirdly, Activity Theory is briefly introduced. This is followed by an explanation of how Activity Theory can be utilised to model contextual information. Finally, some pointers for future work are presented. 2
The concept of context is a closely related to reasoning and cognition in humans. Even though, context might be important for reasoning in other animals, it is common knowledge that context is of huge importance in human reasoning.
Beside the more mechanical view on reasoning advocated by neuro-science, psychology and philosophy play important roles in understanding human cognition. It might not be obvious how computer science is related to knowledge about human cognition. However, many sub-fields in computer science are influenced by our knowledge about humans; and other animals.
The field of Artificial Intelligence has the most obvious relations to the study of reasoning in the real world, most prominently psychology and philosophy. Since AI and psychology are very closely related and context is an important aspect of human reasoning, it should come as no surprise that context also plays an important role in the understanding and implementation of artificial intelligence.
AI has historically been closely connected to formal logic. Formal logic is
concerned with explicit representation of knowledge. This leads to the need to codify all
facts that could be of importance. This strict view on objective truth is also known in
certain directions within philosophy, where such a concept as knowledge as an objective
truth exists. This comes as no surprise, since the father of logic Aristotle, believed that
some subset of knowledge had that characteristic (Episteme). This view stands in stark
contrast to the views advocated by people such as Polanyi, who argues that no such
objective truth exists and all knowledge is a some point personal and hidden (tacit) [
Since context is an elusive type of knowledge, where it is hard to quantify what
types of knowledge is useful in a certain situation, and possible why, it is obvious that it
does not fit very well with the strict logical view on how to model the world. According
to Ekbia and Maguitman [
The paper by Ekbia and Maguitman is not a recipe on how to incorporate contextual reasoning into logistic systems, rather an attempt to point out the deficiencies and suggest possible directions AI could take to include context. The work by Ekbia and Maguitman builds on the work by the American philosopher John Dewey.
According to Ekbia and Maguitman, Dewey distinguishes between two main categories of context: spatio and temporal context, together know as background context; and selective interest. The spatio context covers all contemporary parameters. The temporal context consists of both intellectual and existential context. The intellectual context is what we would normally label as background knowledge, such as tradition, mental habits, and science. Existential context is combined with the selective interest related to the notion of situation. A situation is in this work viewed as a confused, obscure, and conflicting thing, where a human reasoner attempts to make sense of this through the use of context. This view, by Dewey, on human context leads to the following suggestion by the pragmatic approach [7, p. 5]: 1. Context, most often, is not explicitly identifiable. 2. There are no sharp boundaries among contexts. 3. The logical aspects of thinking cannot be isolated from material considerations. 4. Behaviour and context are jointly recognisable.
Once these premises have been set, the authors show that the logical approach to
(artificial) reasoning has not dealt with context in any consistent way. The underlying
argument is that AI has been using an absolute separation between mind and nature,
thus leading to the problems associated with the user of context. This view on the
inseparability of mind and nature is also based on Dewey’s work. This view is not unique
for Dewey. In recent years this view has been proposed in robotics as situatedness by
Brooks [
Through the discussion of different logical-based AI methods and systems, the authors argue that AI has not yet parted company with the limitations of logic with regards to context. Furthermore, they stress the point of intelligence being action-oriented; based on the notion of situations described above.
The notion of intelligence being action-oriented, thus making context a tool for
selecting the correct action, is shared by many people within the computer science milieu.
Most notably the work by Strat [
Strat [
Strat defines three main categories of context: physical, being general information about the visual world independent of the conditions under which the image was taken; photogrammetric, which is the information related the acquisition of the image; and computational, being information about the internal state of the processing. The main idea in this work is to use context to guide the selection of the image-processing algorithms to use on particular images. This is very must in line with the ideas proposed by Ekbia and Maguitman, where intelligence is action-oriented, and context can be use to bring order to diffuse situations.
This action-orientated view on reasoning and use of context is also advocated by
O¨ztu¨rk and Aamodt [
The authors build a taxonomy of context categories based on this merger of the two different worlds of information (internal vs. external). Beside this categorisation, the authors impose the action, or task, oriented view on knowledge in general, and contextual knowledge in particular. The goal of an agent focuses the attention, and thereby the knowledge needed to execute tasks associated with the goal. The domain use in this paper is medical diagnostics, where a doctor attempts to diagnose a patient by the hypothesise-and-test strategy. The particular method of diagnostics in this Case-Based Reasoning system, is related to the strategy used by Strat. Though with the minor modification that Strat used contextual information to select the algorithms to use, whereas O¨ztu¨rk and Aamodt have, prior to run-time, defined the main structure of a diagnostic situation, and only uses context to guide the sub-tasks in this process.
Zibetti et al. [
The implementation used to exemplify this approach contains a number of agents ”living” in a two-dimensional world, where they try to make sense of the world by assessing the spatial changes to the environment. Obviously the acquisition of knowledge starting with a tabula rasa is a long and tedious task for any entity. To speed up the process the authors predefined some categories with which the system is instantiated.
All in all, this approach lies in between a complete bottom-up and the more topdown approaches described earlier. 3
The context model used in this work draws on the subjective view proposed above. This
system proposes that the understanding of a given situation is based on a personal view.
Thus, the CBR agent utilised to assess the situation is personal. However, to avoid the
problem of a tabula rasa we have chosen a pragmatical view on how to model context
and introduced a taxonomy that is based on the definition of context given by Dey [
Context is the set of suitable environmental states and settings concerning a user, which are relevant for a situation sensitive application in the process of adapting the services and information offered to the user.
Even though this definition from Dey do not explicitly state that context is viewed
as knowledge, we adhere to the view advocated by Bre´zillon and Pomerol [
We believe that this pragmatic definition of context allows application developers to efficiently rule out information that is not context in their particular application domain (or their context). At design time, developers can ask the question; is this information relevant for adapting our services and information? If the answer is no, the information is discarded as not being context, and excluded from the context model. This flexibility leads to an open context model that only defines the taxonomic structure in the design phase (see Fig. 1).
User Context Task Context
Social Context
Personal Context
Spatioï Temporal Context
We argue a context model where context is not a special type of information. However, this view is not contradictory to a need to structuring our knowledge model with context in mind. Since we are focusing on applications utilising contextual information to improve services provided to users, we have chosen to structure our model around a taxonomy inherited from the context-aware tradition.
The context is divided into five sub-categories (a more thorough discussion can be
found in [
The model depicted in Fig. 1 shows the top-level ontology. To enable the reasoning in the system this top-level structure is integrated with a more general domain ontology, which describes concepts of the domain (e.g., Airport Hall, Gate, Restaurant, Newsstand) as well as more generic concepts (Task, Goal, Action, Physical Object) in a Action
Part of Part of Task Context Social Context
0 1 TriggTer*assk * AchiGevoesal *1 R*ole Physiolog1ical Context
Results in
Isa Isa
Serves Food Serves Drinks
Performs Performs
Isa
Drink Service
multi-relational semantic network. The model enables the system to infer relationships
between concepts by constructing context-dependent paths between them. One
important use of this is to be able to match two case features that are syntactically different,
by explaining why they are similar [
A part of a domain model —in which the context model is integrated— is illustrated in Fig. 2.
This work postulates that a goal or task exists in every situation. It would be futile to identify a situation unless there is some task connected to it – no matter how mundane. This is most obvious when dealing with users, where a situation implies that there is a problem that needs to be solved; such as the possible situation ”hungry user”, which implies the goal of not hungry user, leading to the task provide food, with a subtask locate food.
The problem we face now is to identify the tasks connected to a particular situation, the goals of the user, and the artefacts and information sources used. The different approaches outlined before do not deal with modelling as such. They primarily focus on how context can be represented and utilised. However, knowledge acquisition is an important part of knowledge intensive systems in general and context-aware systems in particular.
Activity Theory has proven itself as a useful tool in modelling and understanding
interaction between humans and their use of artefacts in work place situations [
Isa
Has a Has a Has a Has a
Context
Isa
Tag Context
Isa Isa Has a User
Context Tag
Isa A la Carte
Restaurant Isa Isa Cafe Bar
In this section, we concentrate further the use of Activity Theory (AT) to support the modelling of context. We can use AT to analyse the use of technical artefacts as instruments for achieving a predefined goal in the work process as well as the role of social components, like the division of labour and community rules. This helps us to understand what pieces of knowledge are involved and the social and technological context used when solving a given problem.
First, we give a short summary of aspects of AT that are important for this work.
See [
Activity Theory is a descriptive tool to help understand the unity of consciousness and activity. Its focus lies on individual and collective work practise. One of its strengths is the ability to identify the role of material artefacts in the work process. An activity (Fig. 3) is composed of a subject, an object, and a mediating artefact or tool. A subject is a person or a group engaged in an activity. An object is held by the subject and motivates activity, giving it a specific direction. – Activity: This is the topmost level. An individual activity is for example to check into a hotel, or to travel to another city to participate at a conference. Individual activities can be part of collective activities, e.g. when someone organises a workshop with some co-workers. – Actions: Activities consist of a collections of actions. An action is performed consciously, the hotel check-in, for example, consists of actions like presenting the reservation, confirmation of room types, and handover of keys. – Operations: Actions consist themselves of collections of non-conscious operations. To stay with our hotel example, writing your name on a sheet of paper or taking the keys are operations. That operations happen non-consciously does not mean that they are not accessible.
It is important to note that this hierarchical composition is not fixed over time. If an action fails, the operations comprising the action can get conceptualised, they become conscious operations and might become actions in the next attempt to reach the overall goal. This is referred to as a breakdown situation. In the same manner, actions can become automated when done many times and thus become operations. In this way, we gain the ability to model a change over time.
An expanded model of Activity Theory, Cultural Historical Activity Theory (CHAT),
covers the fact that human work is done in a social and cultural context (compare e.g.
[
In order to be able to model that several subjects can share the same object, we add the community to represent that a subject is embedded in a social context. Now we have relationships between subject and community and between object and community, respectively. These relationships are themselves mediated, with rules regarding to the subject and the division of labour regarding to the object.
This expanded model of AT is the starting point for our use of AT in the modelling of context for intelligent systems.
The next step is to identify which aspects of an Activity Theory based analysis can help us to capture a knowledge level view of contextual knowledge that should be incorporated into an intelligent system. This contextual knowledge should include knowledge about the acting subjects, the objects towards which activities are directed and the community as well as knowledge about the mediating components, like rules or tools.
For example, we want the contextual knowledge to contain both information about the acting subject itself (like the weight or size) and the tools (like a particular software used in a software development process). To this end, we propose a mapping from the basic structure of an activity into the taxonomy of contextual knowledge as depicted in Table 1. We can see that the personal context contains information we would associate with the acting subject itself.
We would like to point out that we do not think that a strict one to one mapping exists or is desirable at all. Our view on contextual knowledge is contextualised itself in the sense that different interpretations exist, and what is to be considered contextual information in one setting is part of the general knowledge model in another one. Likewise, the same piece of knowledge can be part of different categories based on the task at hand.
The same holds for the AT based analysis itself: the same thing can be an object and a mediating artefacts from different perspectives and in different task settings. The mapping suggested here should lead the development process and allow the designer to focus on knowledge-level aspects instead of being lost in the modelling of details without being able to see the relationship between different aspects on a socio-technical system level.
As an example, let us consider a software development setting where a team is programming a piece of software for a client. The members of the team are all subjects in the development process. They form a community together with representatives for the client and other stake-holders. Each member of the team and personal from other divisions of the software company work together in a division of labour. The object at hand is the unfinished prototype, which has to be transformed into something that can be handed out to the client. The task is governed by a set of rules, some explicit like coding standards some implicit like what is often referred to as a working culture. The programmers use a set of mediating artefacts (tools), like methods for analysis and design, programming tools, and documentation.
When we design a context-aware system for the support of this task, we include information about the different team members (subjects) in the personal context. Aspects regarding the special application he is working on (objects) are part of the task context, it will change when the same user engages in a different task (lets say he is looking for a restaurant). The rules are part of the task context since they are closely related to the task at hand – coding standards will not be helpful when trying to find a restaurant. We find the tool aspects (mediating artefacts) in the environmental context since access to the different tools is important for the ability of the user to use them. Knowledge about his co-workers and other stake-holders (community) are modelled in the spatio-temporal context. Finally, his interaction with other team members (division of labour) is found the social context.
Activity Theory is also capable of capturing changing contexts in break-down situations. Lets consider that a tool used in the development process, such as a compiler, stops working. The operation of evoking the compiler now becomes a conscious action for the debugging process. The focus of the programmer shifts away from the client software to the compiler. He will now be involved in a different task where he probably will have to work together with the system administrators for his work-station. In this sense other aspects of the activity, such as the community, change as well. It is clear that the contextual model should reflect these changes. The ability of Activity Theory to identify possible break-down situations makes it possible for the system designer to identify these possible shifts in situation and model the anticipated behaviour of the system. 6
We have outlined how the design of context-aware systems can benefit from an analysis of the underlying socio-technical system. We have introduced a knowledge-level perspective on the modelling task, which makes it possible to identify aspects of knowledge that should be modelled into the system in order to support the user with contextual information. We have furthermore proposed a first mapping from an Activity Theory based analysis to different knowledge components of a context model. The basic aspects of our socio-technical model fits nicely to the taxonomy of context categories we have introduced before, thus making AT a prime candidate for further research.
The use of Activity Theory allows for system designers to develop the general models of activities and situations. General models are necessary to support the initial usage of the system. They are an important prerequisite for the Case-Based Reasoning system to integrate new situations; thereby adapting to the personal and subjective perspective of the individual user.
In Section 3 we have formulated the problem of identifying the tasks connected to a particular situation, the goals of the user, and the artefacts and information sources used. We argue that our Activity Theory based approach is capable of integrating these cognitive aspects into the modelling process.
The integration of an a posteriori method of analysis with design methodologies
is always challenging. One advantage AT has is that it is process oriented, which
corresponds to a view on systems design where the deployed system itself is not static
and where the system is able to incorporate new knowledge over time [
Nevertheless, one of the next steps is to formalise the relationship between different elements of an AT based analysis and the knowledge contained in the different contextual aspects of our model. This more formalised relationship should be put to the test on a context modelling task, using an AT based analysis of a socio-technical system to support the design of a context-aware intelligent system.
We have recently initiate a project where everyday situations in health care are being observed and documented. These observations will be used to test the situation assessment capabilities of our system. We will use a modelling approach based on Cultural Historical Activity Theory. This will allow us to identify the different activities the medical staff is involved with and the artefacts and information sources used.
Part of this work was carried out in the AmbieSense project, which is supported by the EU commission (IST-2001-34244).