Supporting Business Process Modeling with Pattern-oriented Recommender Systems (ProPoneRe)

设想的研究项目ProPoneRe的目标是开发一种业务流程建模方法，通过自动支持分析和建模建议来提高流程模型的质量。我们利用历史过程模型并分析其模型元素标签，以使用自然语言处理（NLP）识别过程模式。模式表示更高抽象级别上的典型处理步骤（例如，“关于应用的决定”），并且可以由多个过程活动组成。为了识别模式，我们首先将流程模型元素标签中使用的动作词分配给抽象类别。一个过程模式由包含不同类别动作词的多个活动组成，在技术上，我们通过使用自然语言处理方法识别动词、名词化、模糊表达和主题角色，并通过分析历史过程模式来消除歧义，从而识别典型的过程活动。在第二步中，我们将识别的活动作为过程模式。在建模期间，将识别的模式与已经建模的模型片段进行比较。基于这些模式，我们开发了一个推荐系统，以不同的方式支持建模者：首先，基于从历史过程模型中提取的知识，我们知道哪些模型元素可能或通常是下一个模型。推荐系统向建模者提出这些建议，这种方式确保建模者不会忘记过程的任何方面。由于建议仅仅是建议，系统还将考虑新的过程（即，对于推荐器系统来说还未知的不可预见的代表）。其次，由于历史过程知识，这样的系统可以识别建模者是否在建模时犯了错误，并相应地通知建模者。为了构建推荐器功能，我们需要能够进行预测的技术（例如，关于接下来应该被建模的附加模型元素或可能的错误）。因此，我们将基于n-gram和概率有限自动机多方法地开发推荐功能，这些自动机将被修改和扩展，以使它们适合于过程模型和过程模式。推荐系统将被实施，测试人员和现有的过程模型从实践中进行评估，并提供给学术界作为一个编程库。

The goal of the envisioned research project ProPoneRe is to develop a business process modeling approach to increase the quality of process models through automatic support for analysis and modeling recommendations. We make use of historic process models and analyze their model element labels to identify process patterns using Natural Language Processing (NLP). The patterns represent typical processing steps on a higher level of abstraction (e.g., “decision about an application”) and may consist of multiple process activities. To identify the patterns, we first assign action words used in the labels of the process model elements to abstract categories. One process pattern consists of multiple activities containing action words of different categories then.Technically, we recognize typical process activities by using NLP methods to identify verbs, nominalizations, vague expressions, and thematic roles, as well as of disambiguation by analyzing historical process models. In a second step, we subsume the identified activities as process patterns. During modeling, the identified patterns are compared to the model fragments, which have already been modeled. Based on the patterns, we develop a recommender system that supports the modeler in different ways: Firstly, based on the knowledge extracted from the historic process models, it is known which model elements one would probably or typically model next. The recommender system proposes these to the modeler and this way assures that the modeler does not forget any aspects of the process. As the recommendations are only suggestions, the system will also account for processes that are new (i.e., that are unforeseen rep. unknown to the recommender system yet). Secondly, due to the historical process knowledge, such a system can recognize if the modeler made mistakes while modeling and inform the modeler accordingly. To build the recommender functionality, we need techniques that can make predictions (e.g., about additional model elements that should be modeled next or probable mistakes) based on already modeled process model fragments and the historical process knowledge. Therefore, we will develop the recommender functionality multi-methodically based on n-grams and probabilistic finite automatons, which will be both modified and extended to make them suitable for process models and process patterns. The recommender system will be implemented, evaluated with test persons and existing process models from practice and made available for the academic community as a programming library.