The decision model in the DMN (Decision Model and Notation) standard is a declarative representation of decision knowledge, which is favored across industry and academia to represent operational decisions. Many current modeling approaches rely on a) a human modeler, which is a costly, time-consuming approach and it struggles to keep up with domain changes, and b) a lot of data logs, to apply automated modeling, which is not feasible for all domains due to unavailability of data. Furthermore, natural language is a standard and convenient way to document decision knowledge in organizations such as rules, policies, and regulations. Despite such vast availability, decision knowledge extraction from the text is relatively new in this domain. This research investigates state-of-the-art NLP techniques, Rule-based approaches, and ML-based approaches in relevant domains. We provide a general framework, Text2DMN, to automatically convert the decision descriptions to the Decision Models. Using this approach, we aim to support decision modelers by reducing the cost and time of the modeling process. This approach also allows improving the quality of models generated, guided by domain expert knowledge as heuristics. We also discuss some of the challenges of this research.
Decision Model and Notation (DMN) is a decision modeling standard designed
by the Object Management Group [
Research plan: to answer the above research questions our research is planned in the following steps: { Perform a literature study to gain good understanding of the techniques and current methods used in the NLP domain as well as technical experience in working with existing tools in information extraction And propose a theoretical framework to solve the identi ed problem. { Know the characteristics of decision texts by analyzing the linguistic patterns syntactically and semantically and by mapping text to their corresponding model components. This step helps to understand the various text patterns and challenges in applying NLP. { E ectively handle these patterns by designing appropriate modeling heuristics inspired from DMN modelling guidelines. These rules will be build into a prototype tool to assess the framework. { Experiment with a prototype tool to test the proposed solution to extract decision model component from text. { Further improve and evaluate the tool against human modellers to determine the performance of the framework and quality of the models generated.
This paper is structured as follows. First, the literature study explains the concepts of DMN, NLP and related information extraction in concept modeling context. Next, the research approach explains which problem the program tries to solve. It also explains what tools were used and nally sheds a light on how the program was developed. In the current results section, the program performance is explained, the results are discussed, compared to previous research. In the challenges section, recommendations for future research are given. The last section summarizes the paper with the conclusion.
In the relavent works of NLP in DMN domain, the automatic extraction of
decision dependencies from paragraphs and generation of DRDs is investigated
[
Prior work in related domains processes and rules can be largely grouped into
two types of solutions a rule-based approach and a machine learning approach.
Rule based: There are many applications of Rule-based NLP for information
extraction in the domains such as UML, business rules, and even process
models [
ML Based: when there is a large amount of annotated data, supervised
machine learning could be used for information extraction. In this method, the
ML model learns the patterns from the data. ML-based extraction system will
predict the patterns encountered. [
Hybrid: some approaches in the literature have opted for the hybrid method
for extracting information from text. Here limitations of both rule-based and
ML-based NLP are addressed. [
Challenges in Decision Model Extraction from Text Based on the study of related research works, decision model extraction from natural language text faces a few issues which can be categorised into challenges that are due to natural language and modelling challenges.
Example: Health risk level The health risk level of a patient should be assessed from the obesity level, waist circumference and the sex of the patient. Furthermore, the degree of obesity should be determined from the BMI value and sex of the patient. Patient's height and weight are considered to calculate his BMI value. If the weight of the patient given in kgs and height of patient given in meters, then the BMI value is weight/(height*height).a a taken from https://www.nhlbi.nih.gov/ les/docs/guidelines/prctgdc.pdf Natural language challenges Decision descriptions in policies and guidelines are quite di erent from the spoken text or informal text. These descriptions come with bullet lists, or logic occurs at various parts of the text with references. However, we can assume that the ambiguity of the decision text is less compared to any social media text. But still, the text with various structures can lead to multiple interpretations. Therefore nding the best preprocessing steps is trivial.
Decision Modelling Challenges A decision model is an interconnected requirement diagram. Extracted concepts must be semantically grouped and pragmatically placed in DRD to avoid loops, duplicates, and unconnected components. The directions of dependencies are essential for assembling the model. Additionally, there are challenges related to model evaluation on how to measure the quality of models. sample DRD generated for the health risk level can be seen in gure 1. 4
The challenges discussed in the previous section point out the complexity of
extracting DMN models from decision descriptions. We earlier proposed a solution
called Text2dec in[
This framework takes a decision text and performs concepts, logic, and dependency identi cation. We used patterns corresponding to sentences of interest are used in the development of the system. The three key phases, text selection, text processing, and model generation, are being described below to give an idea of how the extraction works.
Prepossessing text and selecting text: The rst step of the sentence level analysis phase is the preprocessing of the text. In this step, each sentence is analyzed. Then the execution proceeds with the Spacy tool so that the syntactic and grammatical structure of the sentence is acquired. In other words, POS tagging is performed, and then parse tree-based dependencies are extracted. In this step, noun-based phrases are used for concept identi cation. Coreferences and anaphoras are resolved using neuralcoref API.
Text processing and information extraction: There are two parallel steps in this phase, as shown in gure 1, and the rst step is for Logic extraction. For the logic identi cation, we used conditional markers such as "if", "then", "whenever," or "unless" are used to detect logical rules. The second step of this stage is dependency identi cation, here verb phrases such as "determines", "decided from," "depends on" are used to determine decision dependency tuples, where a tuple is (base concept, verb, derived concept). In this step, the previously extracted concepts are matched to base concepts and derived concepts.
Decision Model Generation: The last phase of the transformation approach can also be divided into two parallel steps, decision table builder and DRD builder, as depicted in gure 1. In the rst step, tables are built from the identi ed rules. In the second step, detected dependencies are transformed into DRD and consequently completing the full DMN model. 5
A general quantitative approach of information extraction (IS) is used to evaluate the Text2DMN framework's e ectiveness. Precision and recall were used as evaluation metrics per each component of the decision model identi able in the text. The decision logic extractor and decision dependency extractor were evaluated separately. Comprehensive test data set of 20 example texts are tailored and labeled to evaluate the components|data set containing di erent decision descriptions and manually created models covering various domains. Note that since this methodology is the rst time in the DMN domain, there is no available baseline data to test. The overall F1-score was determined based on the comparison between the desired information and what was automatically extracted by the program. Preliminary results for dependency extractor is shown in table. 1 and decision logic extractor is shown in table. 2.
With an overall F1-score of 0.769, we say that the developed rule-based technique is helpful to extract decision model components, precisely logic, and dependencies, from simple texts that are not ambiguous and straightforward. But real-world texts are seldom simple.
Information
Precision Recall F1-score dependency tuple dependency direction
Overall
Further Research Plan and Potential Contributions.
The next step in the research plan is to build a tool based on the Text2DMN
framework that supports the extraction of decision model components from
realworld textual descriptions. Since the literature [
Contributions:
1. First time application of natural language processing for complete decision
modeling in terms of approach and methodology. Natural language
processing has been regarded highly in the domains of conceptual modeling
because it helps to handle rapid information changes in the organization's
policies, rules, or guidelines, and it can also facilitate communication of domain
knowledge stored in models through chatbots (similar to [
In this work, we explain the current modeling approaches and problems with these approaches. To tackle these problems, we aim to automate the DMN component extraction from textual descriptions. The main idea is that from the syntactic and grammatical structure of a sentence, the decision model components can be derived (i.e., concepts, dependencies, rules, and constraints). The result would be a decision knowledge model represented as DMN - a popular decision modeling standard. We discuss some related work in the area of model extraction and provide a general framework called Text2DMN. We argue that the evaluation of the model extraction process, the results of automated modeling should be close to the original ones that domain experts manually model. Our preliminary results show that the rule-based approach can perform reasonably well on short example descriptions, but they struggle with large real texts. Therefore we plan to incorporate ML techniques and develop hybrid approaches. Acknowledgements This research is supervised by Prof. Dr. Jan Vanthienen and Dr. Johannes De Smedt.