Abstract
Many techniques for the advanced analysis of process models build on the annotation of process models with elements from predefined vocabularies such as taxonomies. However, the manual annotation of process models is cumbersome and sometimes even hardly manageable taking the size of taxonomies into account. In this paper, we present the first approach for automatically annotating process models with the concepts of a taxonomy. Our approach builds on the corpus-based method of second-order similarity, different similarity functions, and a Markov Logic formalization. An evaluation with a set of 12 process models consisting of 148 activities and the PCF taxonomy consisting of 1,131 concepts demonstrates that our approach produces satisfying results.
Introduction and Problem Definition
Organizations today maintain extensive libraries of business process models to document operations, ensure compliance, and identify inefficiencies. To perform advanced automated analysis, these models must be "semantically annotated"—linked to standardized industry taxonomies like the APQC Process Classification Framework (PCF).
However, manual annotation is a significant bottleneck. Standard taxonomies often contain thousands of concepts, making manual mapping error-prone and difficult to scale. Furthermore, business language is frequently idiosyncratic; general-purpose dictionaries like WordNet often fail to capture the specific nuances of corporate jargon. This research introduces the first automated approach to bridge this gap, transforming the mapping task into a structured optimization problem.
Methodology
The proposed approach utilizes a multi-stage pipeline designed to move beyond simple keyword matching:
- Semantic Decomposition: Activities are broken down into three components: an action (verb), a business object (noun), and additional fragments.
- Second-Order Similarity: Instead of relying on static dictionaries, the system uses a corpus-based method (via the DISCO tool) to calculate similarity. It analyzes word co-occurrence in massive text collections to determine semantic relationships, which is more effective for specialized business terminology.
- Category Weighting (tf-idf): To provide context, the system calculates the similarity between the entire process model and various taxonomy categories using a modified tf-idf approach. This helps the system identify the broad domain (e.g., HR vs. Finance) before analyzing individual steps.
- Markov Logic Networks (MLN): This core reasoning engine combines local similarity scores with global constraints. For instance, a "soft" constraint assumes that if two activities are consecutive in a process, their corresponding taxonomy concepts should likely be close to each other in the hierarchy (the "no-leaps" heuristic).
Experimental Setup
The approach was evaluated using 12 BPMN (Business Process Model and Notation) models covering 148 activities across themes like HR and Product Development. These were mapped against the Process Classification Framework (PCF) version 5.2, which contains 1,131 distinct concepts organized into a four-level hierarchy.
Results and Discussion
The system's performance was compared against a naive baseline of simple highest-similarity matching:
- Top-Level Accuracy (Level 1): The system achieved an f-measure of 0.76, a significant improvement over the baseline's 0.20. This indicates high reliability in identifying broad business functions.
- Sub-Category Accuracy (Level 2): The system achieved an f-measure of 0.45 (vs. 0.11 baseline).
- Granular Levels (Levels 3 & 4): Performance decreased at the most specific levels (0.14 and 0.06 respectively). However, the authors noted that even when the system was technically "wrong" at these levels, it typically selected concepts that were semantically very similar to the human-verified gold standard.
Key insights from the evaluation highlight that context is essential. By looking at the "neighborhood" of the taxonomy and the overall theme of the process model, the system can resolve ambiguities where a single term might appear in multiple business categories.
Conclusion and Future Work
While the system demonstrates strong performance in identifying high-level business contexts, challenges remain regarding domain-specific jargon (e.g., industry abbreviations) and granular leaf-node selection. Future work will focus on incorporating more complex control flow information—such as gateways and loops—into the MLN framework and testing the approach on other ontologies like the MIT Process Handbook.
Summary updated