Article Preview
Top1. Introduction
Many business process management tasks such as process querying, storing, the propagation of process changes and the harmonization of process model variants depend on a matching process (i.e., alignment) (Goncalo et al., 2015; Cayoglu et al., 2013; Weidlich et al., 2012; La Rosa, M et al., 2013; Jin, T et al., 2013). Consequently, Process model matching is a key activity in these business process management tasks (Fakhra et al., 2017). This means that if the business process model matching process is effective then the management tasks are effective too and vice versa. Process model matching refers to the activity that consist on detecting an alignment between process models (i.e., BPs). Computing an alignment between process models is widely recognised as a difficult task (Weidlich et al., 210).
Generally, the proposed matchers in the literature are composed of two sub matchers, a first line matcher and a second line matcher (Fakhra et al., 2017; Gal, Avigdor & Sagi, Tomer, 2010). The first matcher compute similarity matrices that store similarity scores between all pairs of activities in the two process models, these scores are calculated using various similarity metrics proposed in various domains such as NLP (i.e., Natural Language Processing) domain. The second matcher takes as inputs the matrices produced by the first one and compute the final alignment using a set of techniques and strategies.
Consequently, similarity metrics used by the first line matcher and strategies used by the second line matcher are critical for the effectiveness of the designed matcher.
Despite that many powerful similarity measures proposed in the literature have been used to design matchers, the results of the evaluation of the proposed matchers show that their effectiveness needs to be improved (Goncalo et al., 2015; Cayoglu et al., 2013; Daniel Faria et al., 2017). One limitation come from the fact that the proposed matchers are based on similarity metrics that are effective in the context of long texts, while activity labels to be matched are just a short text like check files or send order. Studies show that in practice, these activity labels are divided into three classes: verb-object labels, action-noun labels, and a rest category (Mendling et al., 2009). With these styles of labelling, activity labels tend to be short and ambiguous (Mendling et al., 2009) which make model matching harder.
To overcome this limitation, the proposed solutions in the literature tend to base the design of the matchers on many similarity measures found in the literature or combine various similarity onto new one (Goncalo et al., 2015, Daniel Faria et al., 2017).
While these approaches are promising solutions to build effective matchers, another solution that can increase effectiveness of matchers is to expand activity labels with descriptions that decrease ambiguity in activity labels and making them larger which will make the similarity metrics used by the matchers more effective.
This paper proposes a method based on WordNet glosses (Miller et al., 1993) to improve the effectiveness of matchers.
The proposed method relies on the conjecture that information (verb, nouns, adjectives, etc.) contained in word glosses composing activity labels can be used to expand activity labels to make them longer. This expansion has the potential to help in detecting semantically equivalent activities and consequently improves the effectiveness of matchers.
The proposed method aligns two process models in three steps. In the first step, all activities of the two BPs are extracted. Second, each activity is enriched using WordNet glosses of words contained in the textual description of the activity. Third, correspondences between activities are detected using the cosine similarity metric. The cosine similarity metric is similarity metric widely used in information retrieval task to compute similarity between documents or between queries and documents (Yates & Neto, 1999).
Two experiments were conducted on well-known datasets to validate the effectiveness of the proposed approach. In the first one, an alignment is computed using the cosine similarity metric only and without a process of expansion. While, in the second experiment the cosine similarity metric is applied to the expanded activities using glosses. The results of the experiments were promising and show that expanding activities using WordNet glosses improves the effectiveness of process model matchers.