Library-style ontologies to support varying model views Hermi J.M. Tabachneck-Schijf and Linda C. van der Gaag Department of Information and Computing Sciences, Utrecht University, P.O. Box 80.089, 3508 TB Utrecht, The Netherlands e-mail: {hermi,linda}@cs.uu.nl Abstract In this paper we argue that to support model views for varying tasks, a suite of Bayesian networks should be built The next development in building Bayesian net- rather than a single network. We further argue that in the works will most likely entail constructing multi- first step of developing such a suite, knowledge elicitation purpose models that can be employed for vary- will necessarily result in task-specific information mostly, ing tasks and by different types of user. We ar- although also some task-neutral knowledge may emerge. gue that the development of an ontology to or- Structuring the elicited knowledge into a library-style on- ganize the knowledge needed for such a multi- tology of task-specific and task-neutral modules then is best purpose model is crucial for the management of suited to empower reuse of knowledge segments and to fa- the model’s content. This ontology should pre- cilitate composition of model views. We reiterate our view serve all elicited knowledge and be accessible to that this ontology should capture all elicited knowledge and both domain experts and knowledge engineers. be accessible to domain experts and engineers alike. Based on the different ways in which people We begin by defining different types of model view in Sec- learn and gain expertise, we further argue that tion 2, and outline the task model view under discussion knowledge elicitation will result in task-specific in the current paper. We argue that a single multi-purpose knowledge mostly, although some task-neutral model would quickly become too large and unyieldy to af- knowledge will emerge as well. To support vary- ford the knowledge engineers and the domain experts an ing model views, this combination of knowledge overview of its contents. We therefore advocate building a is best stored in a library-style ontology of task- suite of models to support multiple task model views, rather specific and task-neutral modules. than a single Bayesian network. In Section 3 we outline our view of ontologies. We rational- ize why an ontology should be constructed of the elicited 1 Introduction knowledge, before actually developing a suite of Bayesian networks. This rationalization is much in line with our While in the early years of the field of Bayesian networks earlier arguments for developing ontologies for single net- attention focused primarily on algorithmic issues, the last works [9]. The ontology provides as a well-structured doc- decade has seen an increasing interest in methods to sup- umentation of all elicited knowledge and includes also any port the construction of such networks. The field also has background information that is not captured explicitly in a become more and more experienced in building decision- network. This background information supports, for exam- support systems that include a Bayesian network. Bayesian ple, viewing the elicited knowledge from different perspec- networks by now have evolved beyond laboratory settings tives, as required for different tasks. The well-structured and are being employed by non-academic users. In turn, documentation then scaffolds the building of different task users of these network-based decision-support systems are model views for a suite of Bayesian networks. starting to see the possibilities that these systems offer, and begin to ask for more. For example, for various of our We are not the first to suggest the use of ontologies. Ontolo- biomedical applications, we have been asked whether we gies are being developed for a variety of purposes, ranging could perhaps adapt the model for teaching purposes. It from providing a portal for the semantic web to document- is therefore likely that the next development in the field of ing elicited knowledge for the development of knowledge- Bayesian networks will entail building multi-purpose mod- based models; see for example [4, 6, 8, 18]. For many els which can be employed for varying tasks and, in all of these purposes, a rigorously formal logic-based or other likelihood, by varying types of user. mathematical ontology language is used to allow for auto- mated processing. For our purpose of supporting the devel- structure of the suite of models to be developed, and then opment of a suite of networks by well-structured documen- representing this knowledge in the mathematical formal- tation, however, the ontology should provide as a medium ism of Bayesian networks. The final step is characterized for communication between the engineers and the experts by designing interfaces to the suite of models, that is, the involved in the suite’s construction. Based upon the obser- different ways the models can be presented to someone in- vation that a rigorously formal language is not easily ac- teracting with it, be this an engineer or an end-user. cessed by non-mathematical experts, we advocate, in Sec- We consider a task model view to be one view of a suite tion 3, the use a less formal language for our ontologies. of models. The task model view is the result of carry- We address the knowledge content of our ontologies in Sec- ing out the elicitation and structuring of task-neutral and tion 4. In order to align the content of our ontologies with task-related domain knowledge and of making selections of elicited knowledge, we consider the processes by which the elicited knowledge to support a single or a few closely humans learn and structure their own knowledge. We ob- related tasks. In the medical field, for example, one task serve that the elicited professional knowledge of practic- model view might support diagnostic reasoning, while an- ing experts is mostly both task- and domain-specific, al- other task model view could support teaching diagnostics, though also some task-neutral information may emerge which requires additional modeling of underlying mecha- during knowledge elicitation. nisms so that deeper ‘why’ and ‘what if’ questions can be posed and answered. Interaction model views, on the other In Section 5, we argue that knowledge is best stored in the hand, comprise the interfaces of a model that are tailored to fashion in which it is obtained from the experts. We fur- task and user. For example, for a diagnostics model view, ther argue that the elicited knowledge is best organized into one interaction model view could be optimized for data en- modules. An organization of knowledge in modules is well try and another might support maintenance of the model by suited for storing task-specific knowledge to support mul- the knowledge engineer. tiple tasks. Organizing the modules in a library-style on- tology further encourages reuse of the knowledge elicited In sum, for different tasks to be carried out by differ- for one task model view for the construction of another ent types of user, a suite of models can require several task model view. We would like to note that in our ear- task model views, each of which can need several inter- lier work we proposed the development of a meta-library action model views. In last year’s workshop, we laid of generic knowledge structures complemented with ex- out some methods to construct effective interaction model ample network derivations [11]. To support the evolve- views [16]. In the current paper, we concentrate on the elic- ment of an ontology for a suite of Bayesian networks, such itation and structuring of knowledge, in order to support the generic knowledge structures can guide and speed up en- development of multiple task model views. tering knowledge into the various modules. The paper ends with a discussion and some perspectives 3 Ontologies for Bayesian networks for further elaboration of the presented ideas to a practica- ble knowledge-engineering approach to developing multi- A suite of Bayesian networks that supports several tasks purpose Bayesian networks. with different task model views, is likely to be of a com- plexity necessitating development over multiple years, in- volving possibly different engineers and experts. Build- 2 Model views of Bayesian networks ing and maintaining models of such complexity is a hard and time-consuming process. The knowledge elicited from We distinguish two types of model view, namely task model domain experts constitutes a rich pool of knowledge, seg- views and interaction model views. To explain the differ- ments of which can play varying roles in the domain under ence between the two types, we distinguish three different study. All this elicited knowledge has to be carefully re- states in the development of a suite of models. The first viewed and structured, and ultimately captured in the for- state consists of a stored pool of knowledge relevant to all malism of Bayesian networks. In this process, a multitude tasks to be carried out. The second state encompasses the of modeling decisions are taken as well as numerous deci- actual suite of models that allows computations to be car- sions to demarcate the scope of the model. Such decisions ried out for the various tasks. The third state comprises tend to forestall an overview and thorough comprehension concrete means that allow users to work with the suite of of the model by anyone who has not been intimately in- models. In view of these three states, we also consider the volved in its construction. We have experienced already for steps that need to be taken to proceed from one to the next single larger networks, that construction and maintenance state. The first step reaches the first state and involves elic- are seriously hampered if the elicited domain knowledge iting and structuring knowledge. The second step neces- and the decisions taken are not made explicit by proper doc- sitates first selecting, from the pool of all elicited knowl- umentation [9]. This problem is bound to grow worse if a edge, the knowledge that determines the content and the suite of networks is to be developed and maintained. Having observed the advantages of developing an ontol- mal representation in addition may provide for (semi-)au- ogy before building a single Bayesian network in our ear- tomated derivation of segments of the Bayesian networks lier work [9], we feel that the construction of a suite of under construction. While rigorously formal languages models will especially benefit from an explicit ontology, often have limited expressiveness, an ontology language which then serves not just as a documentation of all elicited should come with a rich semantics to introduce as little bias knowledge but also as a means of ensuring consistency over as possible in the represented contents. If the language in- the models within the suite and as a medium for communi- troduces biases, for example as a result of not allowing the cation between the experts and engineers involved. representation of specific knowledge constructs, then the ontology may not properly reflect the intricacies of the do- 3.1 The role of ontologies main. Since the ontology is to be used for the construction of a network, the resulting model may then be biased as Most generally applicable knowledge-engineering method- well, maybe even in unforeseen ways. The development of ologies, among which is the well-known CommonKADS an independent knowledge model, recommended by most methodology [13], strongly recommend the development knowledge-engineering methodologies, in fact has its ori- of a conceptual model before actually constructing a model gin in this observation. in the knowledge-representation formalism to be used. In The purpose of knowledge sharing provides a strong argu- line with this recommendation, we recently proposed to de- ment for using a less formal language. The ontology should velop an ontology before constructing a Bayesian network be represented in a language that is understandable for both for a domain at hand [9]. the knowledge engineers and the domain experts involved There exist many views of the concept of ontology in gen- in a network’s construction. We argued before that the eral; see for example [4, 7, 8, 18]. In this paper, we use mathematical language of Bayesian networks, for example, the term ontology to refer to an explicit specification of the is very difficult to grasp by non-mathematical persons [16]. elicited domain knowledge that is to be shared by the ex- In our opinion in fact, many of the formal languages com- perts and the knowledge engineers involved in a network’s monly used for ontologies are unsuitable for checking the construction and maintenance. From this perspective, an accumulated knowledge with non-mathematical experts. If ontology plays two distinct roles. One of these is to make the use of a formal language is uncommon in a domain of all elicited domain knowledge explicit. To this end, the application, then a rigorously formal language is unsuited ontology specifies not just the knowledge that is to be cap- for the purpose of knowledge sharing between the knowl- tured in a network, but also the relevant background knowl- edge engineers and the domain experts in the domain at edge of the domain and the meta-level knowledge of its hand and a less formal language had best be used. regularities and organizational structure. Note that captur- To support developing Bayesian networks in the biomedi- ing the elicited knowledge directly in a Bayesian network cal domain, we use a semi-formal ontology language com- would result in a representation from which not all types posed of well-structured tables, depictions, graphs and hi- of domain knowledge are easily recognizable as a result of erarchy representations combined with text [9], which can the modeling decisions taken. Also, some of the elicited be understood by both the domain experts and the knowl- knowledge may not be captured at all in the network. The edge engineers. As an example, Figure 1(a) shows part of other main role of an ontology is to provide as an explicit an ontology for the medical domain of oesophageal can- medium for communication between experts and engineers cer. The depicted graph captures the relationships between alike for further knowledge acquisition, network validation the result of a gastroscopic examination of the circumfer- and maintenance. ence of a patient’s tumour and the underlying true circum- ference. It describes, for example, that a gastroscopic ex- 3.2 The ontology language and an example amination may not result in an image from which the cir- To support the two roles mentioned above, the representa- cumference can be established, as a result of a patient’s tion language to be used for an ontology should be chosen impaired swallowing capabilities. with care. The issue of selecting an appropriate ontology Upon establishing the stage of a patient’s cancer, not only language has been addressed by many researchers. Some the circumference of the primary tumour is investigated. suggest that domain knowledge should be represented by Other diagnostic tests are performed as well. In addition a language that is highly informal, semi-informal, or semi- to the knowledge pertaining to these tests separately, the formal [18]; others argue that ontologies should be speci- domain’s ontology specifies the high-level regularities of fied in a rigorously formal language and, in fact, should be the knowledge involved. The graph capturing these regu- machine readable [14]. larities for the various diagnostic tests is depicted in Figure An important argument for using a formal ontology lan- 1(b). Note that this graph can be exploited upon extending guage is that it allows a highly structured and unambigu- the network with the results of a new test, as it provides ous representation of the elicited knowledge. Such a for- for guiding the elicitation of the knowledge pertaining to oesophageal tumour inducing gastroscopic image inducing result gastroscopy circumference circumference circumference value value inducing enabling enabling enabling laboratory technician test skills gastroscopy passage impairment enabling physician enabling degree interpretation skills gastroscopy gastroscopy interpretation gastroscopic image status status (a) entity inducing presentation inducing result property value value inducing enabling enabling enabling test technician test skills manifestation enabling interpreter enabling degree interpretation skills test interpretation status status (b) Figure 1: Relations between test results and the underlying true values, (a) for a gastroscopic examination of the circum- ference of an oesophageal tumour, and (b) for a diagnostic test in oncology in general the new test. For further details of the oesophageal cancer that adheres to the syntax of Bayesian networks. To this ontology, we refer to [9]. end, the domain concepts from the depiction are translated into stochastic variables, which may involve for example re-defining multi-valued variables. The relations from the 3.3 Ontology-supported construction of networks depiction are translated into arcs between variables in the initial graphical structure. Note that many of these steps Of course it is a daunting prospect to have to capture all can be performed in an automated way. Figure 2 shows, elicited knowledge in two ways, that is, first in an ontol- as an example, part of the initial graphical structure that is ogy and then in a suite of Bayesian networks. A carefully derived from the graph of Figure 1(a). In the final step, the structured ontology, however, can be used to derive the engineer has to verify that the resulting structure correctly graphical structure of the suite in a semi-automated fash- captures probabilistic independence. Also, the initial struc- ion. First, the knowledge that is to be captured in the suite is ture may need further optimization [10]. selected from the ontology; the remainder of the ontology then serves as background knowledge to the suite. Note that this step involves a reflection on the elicited knowledge 4 Eliciting ontology knowledge which must be performed and documented by the knowl- edge engineer. In the next step, the central concepts and Given the prospective advantages of constructing a domain relations from the selected parts of the ontology are com- ontology before building a suite of Bayesian networks, we bined into a single depiction for each envisioned network. now turn to the question of how to organize the elicited From this depiction, an initial graphical structure is derived knowledge in the ontology so that it most usefully supports different task model views for the suite. Gastro-image- Many researchers recommend that ontologies be con- Circumference circumf Gastro-circumf structed independently of the projected use of the ontology and its contents; see for example [3]. Underlying this rec- ommendation is the argument that any commitment to the Passage Test-skills Interpretation-skills problem-solving method that will be applied to the domain knowledge for example, will influence and thereby bias the Figure 2: The initial segment of the graphical structure contents of the ontology. Such commitments thus hamper the extendibility and reuse of the ontology. However, con- structing an ontology without any commitments to a par- ticular task requires either eliciting task-neutral knowledge from domain experts, or stripping the task-specific aspects from the elicited knowledge. In this section, we address the feasibility of the first option; the second option is briefly addressed in Section 5. We consider eliciting task-neutral information, that is, elic- iting knowledge from experts without them having a par- ticular task in mind. To provide task-neutral information, experts should be able to gather such information from their minds, which implies that the knowledge should be stored in their brains in such a way that task-neutral as- pects are readily separated from task-specific aspects. We now briefly lay out the different ways in which people learn information and argue that these learning processes imply that the knowledge stored in the human brain is largely both domain- and task-specific. We then conclude that, given how knowledge is learned and stored, it would be Figure 3: Knowledge acquired by different processes extremely difficult to elicit task-neutral knowledge from an experienced professional. in particular situations, accreted knowledge is largely both task- and domain-specific. Figure 3 summarizes the four 4.1 Human knowledge acquisition processes processes by which humans acquire knowledge. Humans acquire knowledge in four different ways: trans- mission, acquisition, accretion, and emergence [19]. Usu- 4.2 Example: the acquisition of medical knowledge ally people start gathering professional knowledge from While the four learning processes reviewed above relate books and teachers: the knowledge is explicitly transmit- to general educational practices, they are easily mapped ted to them. Except in vocational training, such trans- onto what happens in the course of gathering professional mitted knowledge is mostly task-neutral. Over the course knowledge. Although the exact percentages may vary a of a lifetime, transmission accounts for some 10% of our little, the different processes will create roughly the same knowledge. Further learning done by conscious choice proportions of the knowledge that our domain experts pos- is termed acquisition learning, which is good for about sess. We illustrate this observation with an example from 20% of our knowledge. Acquired knowledge is gathered medicine [2], and also argue that transmission and acquisi- by our own initiative: by exploring, experimenting, self- tion learning in college does not prepare a student for med- instruction, inquiry and the like. Emergence is the result of ical practice, because of the task-neutral nature of the ma- self-constructing new ideas and meanings that did not ex- terial learned in medical school. ist before, which in current educational practices is said to account for just 1-2% of our knowledge. The basics for medical knowledge are taught by transmis- sion in universities. This type of knowledge is explicitly When people are asked to describe learning processes, they task-neutral and consists of biomedical knowledge, which generally mention explicit processes akin to transmission is mostly causal and definitional in nature and describes and acquisition, and perhaps emergence. Accretion, which the functioning and possible dysfunctioning of the human accounts for about 70% of what we know, however, does body. It is this transmitted knowledge that upon elicitation not commonly come to mind. Accretion is the gradual, un- would result in task-neutral knowledge segments. conscious and implicit process by which we learn for exam- ple language, culture, social behavior, and whatever other Next, students are confronted with patients in internships, knowledge comes on our path. Accretion knowledge is where they have to link the transmitted task-neutral infor- picked up simply by living and interacting with the world. mation to clinical knowledge. In contrast to biomedical Within limits, we process and react to all we see, hear, knowledge, clinical knowledge is task-specific in nature. smell, taste and experience. By processing the information It consists of knowledge of symptoms, classification and and reacting to it, it is stored in the brain without our be- treatment of diseases, all embedded in medical situations. ing conscious of the learning process. People consequently In internships, some transmitted information is still offered, often are not even aware they possess this type of knowl- but students are also acquiring knowledge by trying to fig- edge. Because it is unconsciously experienced and learned ure out diagnoses and treatment plans themselves. Accre- tion then is also at work, continually recording knowledge comes to the fore upon elicitation is task- and domain- from all perception instruments. Examples of accreted specific. Of course an engineer can explicitly ask a do- knowledge are how to read symptoms from patients’ look, main expert to provide task-neutral knowledge. If experi- smell, utterances and behavior, and how to communicate ence from practice is requested, however, the engineer is with colleagues, patients and their next of kin, yet also how asking for extra information processing from the expert: to get around in the hospital and many other aspects of the expert has to relate his or her knowledge in a differ- work. All that is learned is now embedded in the task at ent way than is stored in the brain. This, as argued in hand and in the medical culture and practices. In cognitive- the example above of the medical students’ transition from science terms, the knowledge is situated. book knowledge to diagnostic and treatment knowledge, requires non-trivial effort, which, as it is to be done real- It is taking the step from employing task-neutral knowledge time, will at least considerably slow down the elicitation. in college to having to apply task-specific knowledge in a More potentially damaging, however, asking people to re- hospital setting that makes the transition from the univer- lay knowledge in a way that requires them to reason about sity classrooms to practice so problematic for many medi- their stored knowledge, as is done when asking an expert cal students [2]. Students may have learned which disease for task-neutral information, always increases the risk of causes which symptoms, and maybe even have seen pic- introducing errors [5]. We conclude that, except for infor- tures of such symptoms. However, recognizing the symp- mation that was transmitted in a task-neutral fashion, it will toms when exhibited by a patient is a very different matter. be difficult, time-consuming and error-prone to try to elicit Each patient is unique, and may or may not exhibit all of the task-neutral knowledge from domain experts. symptoms. Patients also may exhibit symptoms differently. Patients may further have more than one disease, which Two examples from our own research will serve as illus- may result in an indistinct mixture of symptoms. Last but trations. As a first example, when we asked veterinarians not least, the reasoning required now goes diagnostically to supply us with average disease symptoms for pigs that from symptoms to disease, not causally from disease to were sick, most of them provided us with symptoms be- symptoms. The difficulty of this re-representation is sup- longing to one particular illness rather than a context-free ported by research in various other contexts, from which average; some gave symptoms associated with a particu- it is also clear that switching information from one repre- lar group of closely related diseases such as infections of sentation to another is very difficult. Switching represen- the respiratory tract. What happened is that the veterinar- tations, in fact, does not occur spontaneously and must be ians called a pig having a particular disease to mind, of explicitly and extensively taught [1, 17]. which they provided the symptoms. The veterinarians pro- viding a few more symptoms ostensibly generalized but ac- Professional learning in medicine does not stop with the tually were doing exactly what their colleagues did: they internship phase. It continues by a mixture of accretion provided the symptoms of diseases encountered within the and acquisition during the entire professional career. All same differential diagnosis. The veterinarians unwittingly knowledge picked up in this phase is in a task-specific for- rendered their knowledge in the same situated way it was mat, because it is learned while carrying out specific tasks. stored, rather than following our instructions. The theory of situated learning describes this phenomenon and argues that learning as it normally occurs is a func- As a second example, we relate a knowledge-elicitation tion of the activity, context and culture in which it occurs session where we asked a group of veterinary experts to [12, 15]. In fact, the theory argues specifically that learning reason out loud about particular pig cases of which the clin- never occurs in a task-, context-, and culture-neutral man- ical symptoms were described in terms of variables and ner.1 In a physician, for example, interaction with patients values. When asked what would happen to their assess- is typically stored as examplars of sick people complete ment of the case when a particular symptom was changed with diagnosis, treatment plan, and outcomes. from present to absent, one of the participants asked, in earnest, how he could possibly change the symptoms of a From the above observations, we conclude that the bulk pig. Clearly, the veterinary expert had called the case to of the professional knowledge of an expert is stored in the mind as a concrete pig for which he had to come to a di- mind in a task-specific format. agnosis. Thinking in this task-related setting, he could not imagine physically changing a pig’s symptoms. 4.3 Eliciting task-specific knowledge Since professional knowledge is largely task-specific, it is reasonable to assume that most of the knowledge that 5 Storing the elicited knowledge 1 According to this theory, the knowledge transmitted in med- ical school is also not task-neutral: the task is passing the exam. Having established that it will be rather unlikely that an For our purpose, however, the issue is that the knowledge is inde- engineer will elicit knowledge from a domain expert that pendent of specific medical tasks. is altogether task-neutral, we now address how the elicited knowledge is best stored in an ontology. More specifically, From the above observations, we conclude that although we compare constructing a single task-neutral ontology storing knowledge in a task-neutral fashion is prefered, that is free of task biases, with constructing multiple task- it is infeasible to do so for the bulk of elicited informa- specific ontologies. We then argue that a library-style on- tion. Some of the elicited knowledge may be available as tology best supports the development of a suite of Bayesian task-neutral information, however, for example if originat- networks for multiple tasks. This library-style ontology is ing from the transmission phase of learning professional composed of various modules that are task-specific as well knowledge. Also, some of the elicited information can as domain-specific, supplemented with modules that are ei- be abstracted to segments of task-neutral knowledge. An ther task-neutral or domain-neutral. example from our veterinary applications pertains to the stress effects of handling a pig. Catching a pig will cause stress to the animal, regardless of the task for which it is 5.1 Single or multiple ontologies being caught. The knowledge elicited in the contexts of the various tasks thus is explicitly reusable and can be stored We begin by comparing capturing all elicited knowledge in in a task-neutral fashion. a single task-neutral ontology or in multiple task-specific ontologies. For the construction of a single ontology, be it composed of task-neutral or task-specific knowledge, plead 5.2 A library of ontology modules that no duplication is needed and that it will be easier to en- sure internal consistency upon maintenance and extension. Alternative to either a single task-neutral ontology or a In spite of these advantages, however, we reject building a collection of multiple task-specific ontologies as discussed single ontology. A single ontology is likely to become quite above, is a library consisting of multiple ontology modules. large in size for a suite of Bayesian networks supporting Some of the library’s modules contain background knowl- multiple task model views. Even if it is well organized and edge that is common to all tasks in the domain under study highly structured, its mere size will cause the knowledge yet independent of a specific task. Other modules contain engineers and the domain experts to quickly lose track of knowledge that is common to one task but holds across do- its contents. Another argument against the construction of mains; the graph from Figure 1(b), in fact, showed a seg- a single ontology is that it may be much more difficult to ment of such knowledge, pertaining to the interpretation of build multiple task model views from a single entity than the results of diagnostic tests in biomedicine. The majority from a collection of task-focused entities. of the modules, however, capture knowledge that is both task- and domain-specific. A segment of knowledge may Having rejected developing a single ontology, we now ad- thus be captured in more than one module, described from dress the format of the ontology’s content. There are quite the varied perspectives of different tasks. A task-specific strong arguments for storing knowledge in a task-neutral ontology aimed at supporting a particular task model view, fashion. Task-neutral knowledge need not be captured mul- then is constructed by combining various modules. tiple times for use for varying tasks, as would be required if the knowledge were captured in a task-specific fashion. We illustrate the concept of a library-style ontology using Also, when new task model views need be developed, it is our earlier example in medicine. A library of modules for likely that these can already be supported using the avail- medical applications would include, for example, anatom- able task-neutral knowledge. If the knowledge would have ical knowledge. Anatomical knowledge is descriptive and been stored in a task-specific fashion, developing a new definitional in nature and summarizes the elements of the task-specific ontology would be required. human body. Anatomical knowledge is common to most medical tasks yet is independent of any specific task. In Although there are strong arguments for storing the elicited the library, it would therefore be included in one or more knowledge in a task-neutral fashion, it generally will be task-neutral ontology modules. Knowledge of which dis- highly infeasible to do so. In Section 4, we argued that the eases typically occur in the differential diagnoses of which bulk of the elicited knowledge will be available in a format other diseases is closely linked to the task of diagnosis, and that is both task- and domain-specific. Constructing a task- would be included in a task-specific ontology module for neutral ontology would thus require stripping the elicited diagnostic tasks. Note that gradations of task specificity knowledge from its task biases and integrating the result- may be supported. Knowledge of the relationships between ing segments of neutral knowledge. The task of stripping diseases and symptoms, for example, is common to both the elicited knowledge from its task-related context is non- diagnosis and prognostication, and could be included in a trivial, however. Our opinion in fact is that it is infeasible single ontology module subserving both tasks. since not just the experts but also the engineers will have particular tasks in mind when surveying the various seg- To construct a concrete task-specific ontology for support- ments of knowledge. The engineers moreover are likely to ing a model view of teaching diagnostics, information from be insufficiently knowledgeable in the domain of applica- the task-neutral modules of anatomical knowledge would tion to recognize the various task biases included. be pulled in as well as information from modules related to Figure 4: A library-style ontology for developing task model views: the library of ontology modules is supplemented with a library of generic knowledge structures and a document of modeling decisions; drawn arcs indicate instantiation of modules, dashed arcs indicate selection the tasks of diagnosis and prognostication. The modules of 6 Concluding observations anatomy and prognostication would then subserve simula- tion purposes and answering in-depth ‘what-if’ questions. In this paper, we argued that multiple task model views for Note that the other, unrelated modules of the library need Bayesian networks are best supported by a library-style on- not be considered upon constructing the task-specific on- tology composed mainly of task-specific knowledge mod- tology. For supporting a model view of diagnosis, on the ules, but also including task-neutral modules. other hand, the knowledge from the task-neutral modules of anatomy would most likely not be included explicitly in the In summary, this paper addressed several issues. We began task-specific ontology, as the model to be developed could by reiterating the need for documenting all elicited knowl- leave this knowledge implicit. Now suppose that an ontol- edge. If this knowledge is not properly documented, con- ogy for the new task model view of predicting the effects of struction and maintenance of large suites of networks in- treatment is to be developed. Any task-neutral knowledge evitably becomes problematic. We recommended building required for the new model view ideally is already present an ontology to provide a well-structured explicit specifi- in the library and can be readily pulled in. Also the on- cation of the elicited knowledge and a medium for com- tology module of prognostication, which is already present munication for the knowledge engineers and the experts in the library, captures some of the knowledge for the new involved in the networks’ development. We argued that task and can be used. In addition, however, a new task- the ontology should not only store the knowledge needed specific module needs to be developed and included in the for the different model views, but also any relevant back- library. The knowledge for this new module, describing the ground knowledge; in addition, a modeling-decisions doc- physiological effects of treatment, is elicited from domain ument should be maintained. Documentation of the infor- experts, focusing on just the task at hand. mation that cannot be read off the suite of networks directly is especially important when the development of the suite Boshuizen, T. de Jong (eds). Learning with Multiple extends over several years of research and the suite ulti- Representations. Amsterdam: Elsevier Science, Ch. mately is handed off to industry. 8, pp. 137 – 152. The paper also attended to the language to be used for [2] H.P.A. Boshuizen, M.W.J. van de Wiel (1998). One our ontologies. The necessity of including all types of person, multiple representations: An analysis of relevant knowledge demands a language that allows for a a simple, realistic multiple representation learning rich semantics and permits semi-automated model build- task. In: M.W. van Someren, P. Reimann, H.P.A. ing. We stressed that the language used should be accessi- Boshuizen, T. de Jong (eds). Learning with Multiple ble for non-mathematical domain experts. Earlier research Representations. Amsterdam: Elsevier Science, Ch. had shown that rigorously formal representations, be they 12, pp. 237 – 263. logic-based or stated in another mathematical language, cannot readily be understood by domain experts who are [3] T. Bylander, B. Chandrasekaran (1988). Generic tasks not trained in such representations. When stated in a semi- for knowledge-based reasoning: the ‘right’ level formal language that is accessible for the experts, the on- of abstraction for knowledge acquisition. In: B.R. tology can provide as a means of communication between Gaines, J.H. Boose (eds). Knowledge Acquisition for the knowledge engineers and the experts, which serves to Knowledge-based Systems, vol. 1. Academic Press, minimize the risk of omitting important information and of London, pp. 65 – 77. including erroneous information. [4] P.C.G. da Costa, K.B. Laskey, K.J. Laskey (2005). Next, we pled for aligning the content of the ontology with PR-OWL: A Bayesian ontology language for the Se- how practicing experts learn and store knowledge in their mantic Web, International Semantic Web Conference, minds. Some knowledge, we argued, is stored in a task- Workshop Uncertainty Reasoning for the Semantic neutral fashion, and should also be stored in this way in Web, Galway, pp. 23-33. the ontology. However, we contended that most knowledge of domain experts is inherently related to specific tasks [5] K.A. Ericsson, H.A. Simon (1993). Protocol Analy- and is stored in that way in their brains. Constructing a sis: Verbal Reports as Data. MIT Press, Cambridge, task-neutral ontology would thus require stripping the task- MA. specific professional knowledge from its task biases. This, [6] T.R. Gruber (1993). A translation approach to however, is highly demanding, either on the part of the ex- portable ontologies. Knowledge Acquisition, vol. 5, pert or on the part of the knowledge engineer, and error- pp. 199 – 220. prone. We therefore proposed storing task-specific knowl- edge in a task-specific fashion. [7] Th.R. Gruber (1995). Towards principles for the de- Lastly, we proposed to develop a library-style ontology, sign of ontologies used for knowledge sharing. In- composed of the aforementioned task-neutral and task- ternational Journal of Human-Computer Studies, vol. specific knowledge modules which subsequently are com- 43, pp. 907 – 928. bined into task-specific ontologies to support concrete task [8] G. van Heijst, A.Th. Schreiber, B.J. Wielinga (1997). model views for a suite of Bayesian networks. We il- Using explicit ontologies in KBS development. Inter- lustrated the ease of development of multiple views and national Journal of Human-Computer Studies vol. 46, demonstrated that reuse of information is encouraged by pp. 183 – 292. organizing the domain knowledge in modules. [9] E.M. Helsper, L.C. van der Gaag (2002a). A case In the near future, we intend to further develop our con- study in ontologies for probabilistic networks. In: M. cept of ontology library by using it in the development of Bramer, F. Coenen, A. Preece (eds). Research and a suite of Bayesian networks in the field of veterinary sci- Development in Intelligent Systems XVIII. Springer- ence. By doing so, we hope to initiate a publicly available Verlag, London, pp. 229 – 242. collection of ontology modules and inspire the uncertainty community to contribute. [10] E.M. Helsper, L.C. van der Gaag (2002b). Building Bayesian networks through ontologies. In: F. van Harmelen (editor). Proceedings of the 15th European Conference on Artificial Intelligence. IOS Press, Am- References sterdam, pp. 680 – 684. [1] H.P.A. Boshuizen, H.J.M. Tabachneck-Schijf (1998). [11] E.M. Helsper, L.C. van der Gaag (2005). Generic Problem solving with multiple representations by knowledge structures for probabilistic-network engi- multiple and single agents; an analysis of the issues neering. Proceedings of the Third Bayesian Modeling involved. In: M.W. van Someren, P. Reimann, H.P.A. Applications Workshop, held in conjunction with the Twenty-first Conference on Uncertainty in Artificial [16] H.J.M. Tabachneck-Schijf, P.L. Geenen (2006). Pre- Intelligence, Edinburgh. venting knowledge transfer errors: probabilistic deci- sion support systems through the users’ eyes. In: L.C. [12] J. Lave (1988). Cognition in Practice: Mind, mathe- van der Gaag, R. Almond (eds). Proceedings of the matics, and culture in everyday life. Cambridge, UK: 4th Bayesian Modelling Workshop: Bayesian Models Cambridge University Press. Meet Cognition, Boston, pp. 74 – 82. [13] G. Schreiber, H. Akkermans, A. Anjewierden, R. de [17] H.J.M. Tabachneck-Schijf and H.A. Simon (1996). Hoog, N. Shadbolt, W. Van de Velde, B. Wielinga Alternative representations of instructional material. (2000). Knowledge Engineering and Management: In: D. Peterson (editor). Alternative Representa- The CommonKADS Methodology, MIT Press, Cam- tions: an Interdisciplinary Theme in Cognitive Sci- bridge, Massachusetts. ence. London, England: Intellect Books, pp. 28 – 46. [14] R. Studer, V.R. Benjamins, D. Fensel (1998). Knowl- [18] M. Uschold, M. Gruninger (1996). Ontologies: prin- edge engineering: principles and methods. Data & ciples, methods and applications. The Knowledge En- Knowledge Engineering, vol. 25, pp. 161–197. gineering Review, vol. 11, pp. 93 – 136. [15] L. Suchman (1988). Plans and Situated Actions: The [19] L.O. Wilson (1997). New View of Learning: Problem of Human/Machine Communication. Cam- Types of Learning. Retrieved Jan 25, 2007, from bridge, UK: Cambridge University Press. www.uwsp.edu/education/lwilson/learning