Enabling Subject Matter Experts (SMEs) to formulate knowledge without the intervention of Knowledge Engineers (KEs) requires providing SMEs with methods and tools that abstract the underlying knowledge representation, allowing SMEs to focus on the modeling activities. However, automatically bridging the gap between SME-authored models and their internal representation is not an easy task, especially in the case of complex knowledge types like processes, where aspects like frame management, data, and control flow need to be addressed. In this paper, we present a process representation formalism and method for automatically grounding SME-authored process models in the form of process diagrams into a particular representation language, supporting process representation and reasoning.
Enabling Subject Matter Experts (SMEs) to model processes by themselves without the intervention of Knowledge Engineers (KEs) is a complex problem that needs to be addressed from a multidimensional perspective in order to: i) provide the required knowledge artifacts to acquire process knowledge and ii) develop usable tools enabling SMEs to exploit such artifacts. To this purpose, our work has focused on producing the following models, methods and tools:
A process metamodel, which provides the terminology necessary to express process entities in scientific domains and the relations between them.
A library of Problem Solving Methods [
While we presented the first three outcomes in [
We consider four main types of process reasoning to be supported by the formalism: i) reasoning about process entities, ii) intermediate results, iii) process stages, and iv) process preconditions. For example, the multiple-choice question below, selected from Advanced Placement (http://apcentral.collegeboard.com) exams in Biology, illustrates the third type of process reasoning.
Which part of the animal cell is required only in the first stage of mitosis and what is the name of such stage?
chromatin and prophase chromatid and prometaphase centromere and anaphase plasma membrane and telophase In our formalism, a process consists of a set of process actions, connected in the form of a directed graph, with pre and post conditions whose evaluation both determines the flow of data between process steps and controls the order in which such actions are executed. We define the pre and post states of an action respectively as the content of the process frame immediately before and after its execution. The pre state contains all the process resources in the knowledge base that serve as inputs to the action, while the post state contains the outcomes of its execution, obtained by operating on the contents of the pre state.
At modeling time, our code generation method automatically
synthesizes sound and complete executable code in the form of
Flogic rules associated to each action in a process model. F-logic i)
provides a single entry point for reasoning, supporting the
different knowledge types involved in particular questions, ii)
enables the use of rule knowledge for reasoning within processes,
and iii) keeps introspective properties for retrieval of
metainformation about processes, like subprocesses and intermediate
process results. Process rules manage the process frame [
This work was evaluated by an independent team in the context of project Halo. Six SMEs formulated knowledge on the selected syllabi for the domains of Chemistry, Biology, and Physics, and tested reasoning with it. The quality of the resulting knowledge bases was determined by test sets created by the SMEs themselves through the testing & debugging tool in the system in order to check that their process models actually behaved as expected. 82% of the process models were correct. In all cases, process models were formulated by SMEs without intervention of KEs and only required initial training and sporadic support in the utilization of the tools.
The process modeling environment was rated by SMEs with an
average of 64.5 out of 100 in the System Usability Scale [
We studied the effects of the application of the optimizations described in section 6 to the F-logic code resulting from the process models formulated by the SMEs. We measured response times (Table 1) of a sample of ten queries uniformly distributed across the four reasoning types described in section 2. These queries were executed against the Biology knowledge base produced by SME3, which contained the largest sample of process knowledge produced by the SMEs in the evaluation, with three different configurations of the F-logic reasoner OntoBroker, combining different uses of well-founded evaluation and secondorder reasoning. C0 is the most generic configuration, with the well-founded evaluation mode enabled and concept and attribute names ground disabled. C1 and C2 correspond to the optimization methods described in the previous section. C1 aims at increasing performance with respect to C0 by enabling concept and attribute names ground while C2 extends C1 by additionally disabling well-founded evaluation.
The results of executing this query set with the three different configurations are shown in Table 1 (values equal to 0 stand for queries with response times lower than 1ms) shows an average performance improvement of 25% for C1 and almost 30% for C2. The main reason is that concept and attribute names are ground as in C1. C2, an extension of C1 that also disables well-founded evaluation mode, adds in this case little performance gain since the code generation mechanism already produced most of the code in well-stratified form, hence reducing the need of wellfound semantics.
This work has been funded by Vulcan Inc. as part of Ontoprise’s DarkMatter project within the Halo project (http://www.projecthalo.com).