Automatic process modeling method for complex scenario description

Abstract

The application discloses an automatic process modeling method for complex scene description, first, using a method based on shared representation learning and structured prediction, synchronously realizing two important basic works of event discovery and event relationship extraction. Secondly, through the process model construction method, events and event relationships are integrated to obtain the process model extracted from the complex scene description. Finally, in order to detect the effectiveness of the process model extraction, the application obtains the graph similarity score between the extracted process model and the real process model through a process model similarity calculation method based on a graph neural network. The application can automatically extract a business process model from a description text of a complex scene through technologies such as event discovery, event relationship extraction and process model construction. The application uses a novel process model consistency detection method combined with deep learning, which can quickly and objectively evaluate the accuracy of the result after automatic process modeling.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of process mining, and particularly relates to an automatic process modeling method for complex scene description. BACKGROUND

[0002] Process mining provides a bridge between data mining and model-driven process management, and has been widely studied and applied in recent years. The process discovery task in process mining aims to obtain the corresponding process model from the event log captured by the system, and this technology provides a new means for process discovery, monitoring and improvement in various application fields. With the rapid development of network and ubiquitous computing, enterprises are pushed into the highly digital big data era, and Internet of Things and data analysis have become information technologies for enterprises to gain competitiveness in the digital era, thereby driving business process modeling to greatly improve generation efficiency, reduce operation risk and increase customer satisfaction.

[0003] With the continuous development of modern informatization, business process modeling work has reached a quite complex level. In the face of rapidly changing business environment, when the existing business process cannot meet the new business demand, it is necessary to rebuild the business process model that meets the demand. On the one hand, existing methods require modeling personnel to be very familiar with the knowledge of the relevant field, and have a large amount of relevant knowledge structure and work experience. Modeling personnel often have difficulty in fully understanding the content of the demand change of the new business, thereby causing the model to be out of touch with reality. On the other hand, the business process in the information age is very complex, and the elements of each link are intertwined. Completely relying on the operation of modeling personnel to build a business process model often causes the model to lack systematicness and comprehensiveness, resulting in a model with great uncertainty. Such methods not only consume a large amount of manpower, material resources and time, but also have problems such as errors. In addition, although the traditional modeling auxiliary means based on process mining or process retrieval improves the application rate of intelligent technology to some extent, a large amount of manual work is still required. More importantly, existing modeling technologies are based on system event logs and can only be applied to process text and structured text, which has great limitations.

[0004] To overcome the aforementioned bottlenecks, this design proposes an automated process modeling method for complex scenario descriptions. This method integrates deep learning technology with traditional business process modeling methods, extending the application of process modeling from process text to unstructured text, and automating operations from event discovery and event relationship identification to process model extraction. The specific steps of the automated process modeling method for complex scenario descriptions are as follows: First, using a method based on shared representation learning and structured prediction, the two crucial foundational tasks of event discovery and event relationship extraction are simultaneously achieved. Second, events and event relationships are integrated through a process model construction method to obtain the process model extracted from the complex scenario description. Finally, to test the effectiveness of the process model extraction, this invention uses a graph neural network-based process model similarity calculation method to obtain the graph similarity score between the extracted process model and the real process model. This invention overcomes the bottlenecks of previous business process modeling methods based on process text, enabling automatic modeling of process models from natural language text descriptions of complex scenarios, significantly improving the efficiency of business process modeling. Furthermore, a novel process model consistency detection method combining deep learning is used, which can quickly and objectively evaluate the accuracy of the results after automated process modeling. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing an automated process modeling method for describing complex scenarios. By preparing a dataset and then designing an automated process modeling method, the automatic construction of process models can be effectively achieved, thereby improving the modeling efficiency in the field of business process management.

[0006] The objective of this invention is achieved through the following technical solution: an automatic process modeling method for describing complex scenarios, comprising the following steps:

[0007] 1) Prepare a dataset. Based on the requirements of automated process modeling, we need to prepare a dataset for automated process modeling, which should include four elements: text, events, event relationships, and corresponding process models.

[0008] 2) Based on the dataset prepared in step 1), preprocess the input text so that it can be input into the automated modeling model.

[0009] 3) Train the joint extraction model of events and temporal relationships based on steps 1) and 2).

[0010] 4) Based on the work of event discovery and event relationship identification in step 3), use a joint event and event relationship extraction model with shared representation learning and structured prediction to extract events and event relationships from the text.

[0011] 5) The work of building the process model based on the events and event relationships extracted in step 4) mainly includes the following parts: integrating event relationships, constructing an initial flowchart from events belonging to the same text, mining flowchart seeds from the initial flowchart, and adding key nodes to the flowchart to form a process model. At this time, the automatic construction of the process model is completed.

[0012] 6) Based on the process model automatically constructed in step 5), a training data set for consistency detection of the process model is constructed.

[0013] 7) Based on the consistency detection data set constructed in step 6), a process model consistency detection model based on graph neural network is trained.

[0014] 8) Based on the consistency detection model trained in step 7), the similarity of the process model is compared, and the difference between the automatically constructed process model and the real process model is detected to evaluate the accuracy of the automatic process modeling method.

[0015] In step 1), the specific steps of the data set preparation process are as follows:

[0016] 1.1) Prepare a certain number of texts, and these articles meet the requirements of non-structured text. Here, we use news articles commonly used in the field of information extraction, because news articles contain a large number of real events and interrelated events, which can further extract rich process models.

[0017] 1.2) Label the events in each article, and use the trigger words of the events to represent the events.

[0018] 1.3) Based on the labeling of events, find out the different event relationships (including BEFORE, AFTER, EQUAL, VAGUE, which represent before, after, parallel, and irrelevant) between events in the article, and label the detailed information of the event relationship pairs in the corresponding article.

[0019] 1.4) According to the events and event relationships, combine the semantics of the article to construct the associated process model of the article.

[0020] In step 2), the text preprocessing operation is as follows:

[0021] 2.1) Remove stop words.

[0022] 2.2) Remove sentences that do not contain events.

[0023] 2.3) Divide the article according to the length limit.

[0024] In step 3), the training process of the event and event relationship joint extraction model is divided into two parts:

[0025] 3.1) Firstly, train the event and event relation extraction model by optimizing the scoring function locally and using the cross-entropy loss function.

[0026] 3.2) Based on step 3.1), use structured inference learning to train and re-optimize the network to adjust the global attributes.

[0027] In step 4), event discovery and event relation extraction:

[0028] 4.1) Enhance the representation of events by allowing the event discovery and event relation extraction modules to share the same context embedding and neural network representation learner.

[0029] 4.2) It uses structured inference and learning methods to jointly assign event labels and event relation labels, avoiding the error cascading in traditional pipeline methods.

[0030] In step 5), the process model automatic construction process:

[0031] 5.1) Convert all extracted event relations into BEFORE relation expression form. In the process model, the directed edge represents the occurrence order between events, i.e. the event relation that can be directly used for process model construction. For example, if there is an edge from event A to event B in the process model, it means that event A occurs before event B, and the execution of event B must go through event A. Therefore, we need to convert the AFTER, EQUAL, VAGUE four types into BEFORE relation.

[0032] 5.2) According to the converted new event relations, connect all event pairs with event relations in each article according to the time order of event occurrence as the initial flowchart.

[0033] 5.3) Based on the initial flowchart, divide the initial flowchart corresponding to each article into one or more flowchart seeds by depth-first traversal, where each flowchart seed meets the condition of weakly connected graph, i.e. ensuring that there is no relationship between multiple graphs.

[0034] 5.4) Add necessary start nodes, end nodes, and gateway nodes in the flowchart seed, where the gateway node is used to distinguish between selection structure and parallel structure, and at this time the automatic construction of the final process model is completed.

[0035] In step 6), construct the training data set for process model consistency detection:

[0036] 6.1) Randomly combine the real correct process model to form a pair of process models.

[0037] 6.2) Calculate the graph edit distance of process model pairs by Hausdorff Distance and linear programming method.

[0038] 6.3) Use function to convert the normalized graph edit distance into a score in the range of (0, 1] as the similarity score of the real correct process model pair, the similarity score is calculated as follows:

[0039] Where, represents the process model, is the normalized GED, which is calculated as follows:

[0040]

[0041] Where, represents the number of nodes of the process model , and represents the graph edit distance calculation function of the two process models.

[0042] At this time, the training data for process model consistency detection has been constructed, each data contains two process models and their corresponding similarity scores.

[0043] In step 7, the training process of process model consistency detection, input two process models and the corresponding similarity score each time, and use the mean square error loss function to optimize the whole model, which is calculated as follows:

[0044]

[0045] Where, D is the set of process model pairs in the training data, represents the number of process model pairs, is the correct similarity score of the process model pair, represents the predicted similarity score, the goal of training is to minimize the loss L to achieve high accuracy.

[0046] In step 8, the prediction process of process model consistency detection:

[0047] 8.1) Get the node embedding representation of the process model by GCN (Graph Convolutional Networks).

[0048] 8.2) Use Attention mechanism to consider the importance of each node in the whole process model, which emphasizes important event nodes with a specific similarity measure, and maps each process model to the final graph-level embedding, which contains the global information of the process model.

[0049] 8.3) Calculate the similarity vector of two process model graphs' embeddings and the relevance features of nodes using Neural Tensor Network and Pairwise Node Comparison, respectively, that is, compare the similarity of two process models from the global coarse-grained and local fine-grained two angles.

[0050] 8.4) Concatenate the two vector features output in the last step and pass them through a fully connected layer to get the final process model similarity score.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The automatic process modeling method for complex scene description of the present application breaks through the three major bottlenecks in the traditional business process modeling work, that is, only based on process text, high degree of artificial dependence, and high requirement for business knowledge and work experience of modeling personnel, and can automatically model a process model from the natural language text description of a complex scene.

[0053] In addition, the present application uses a novel process model consistency detection method combined with deep learning technology, which can calculate the similarity between the automatically modeled process model and the real process model, and then facilitate us to objectively evaluate the accuracy of the result after automatic process modeling and the usability and effectiveness of the proposed automatic process modeling method. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 For technical roadmap;

[0055] Figure 2 For process model example;

[0056] Figure 3 For the first three steps involved in the automatic construction of process diagram;

[0057] Figure 4 For the final process model; DETAILED DESCRIPTION

[0058] The method proposed by the present application will be further described in detail below in combination with the drawings and specific examples.

[0059] As Figure 1 shown, the automatic process modeling method for complex scene description comprises the following steps:

[0060] 1) Data set preparation module: According to the requirements of automatic process modeling work, we need to prepare a data set for automatic process modeling, which should contain four elements of text, event, event relationship and corresponding process model, and its related data is as followsFigure 2 as shown.

[0061] 2) Based on the relevant corpus information sorted in step 1), pre-processing operations are performed on the input text so that it can be input into the automatic modeling model, including removing stop words, irrelevant sentences that do not contain events, and dividing the text into specific equal-length sentences.

[0062] 3) Based on the pre-processed text in step 2), the training process of the event and event relationship joint extraction model is performed, as shown in the event discovery and event relationship extraction module in Figure 1 .

[0063] 3.1) First, use the event and event relationship scoring function to perform separate optimization of the event discovery and event relationship extraction models through the scoring function, thereby training the event and event relationship extraction models;

[0064] 3.2) Based on step 3.1), use structured inference learning for joint training, and re-optimize the network to adjust the global attributes.

[0065] 4) Based on the pre-processed text in step 2) and the trained joint extraction model in step 3), the prediction process is as follows:

[0066] 4.1) Input the event mention into the event discovery and event relationship extraction module, and the event mention is first passed through BERT to obtain the corresponding text embedding representation.

[0067] 4.2) Pass the text embedding representation into the BiLSTM layer, which can learn the temporal relationship in the text in both directions, and obtain the embedding vector of each word. Put the word embedding vectors of the two words into the event scoring function and the relationship scoring function, respectively. At this time, the model enhances the representation of the event by allowing the event discovery and event relationship extraction modules to share the same context embedding and neural network representation learner.

[0068] 4.3) Through the two scoring functions, the probability that the two words are event trigger words and the candidate event relationship between the two words can be obtained.

[0069] 4.4) Combine an end-to-end trainable neural structure support vector machine model and ILP constraints for structured inference learning, while ensuring the global structure of the event and its event relationship. By jointly assigning event labels and event relationship labels, it avoids the error cascading in traditional pipeline methods. The model structure diagram is shown in the event discovery and event relationship extraction module in Figure 1 .

[0070] The final prediction results of the event and event relationship joint extraction model are shown in the following table:

[0071] Table 1 Prediction results of event and event relationship joint extraction model

[0072]

[0073] From Table 1, it can be seen that the event prediction accuracy of the model reaches 0.8136, achieving good results. The extraction effect of event relationship is relatively weak, and the accuracy is 0.5238.

[0074] 5) Automatic construction of process model, such as Figure 1 Process model construction module in

[0075] 5.1) To integrate the event relationship, all the extracted event relationships are converted into BEFORE relationship expression form. Among them, AFTER, EQUAL, VAGUE three types need to be converted into BEFORE relationship, and the conversion of AFTER relationship is to exchange the position of the two events. The conversion of EQUAL relationship mainly unifies the relationship between the events in the relationship, ensuring that the events in the EQUAL relationship can be replaced with each other in the flowchart. VAGUE relationship indicates that there is no relationship between the two events, so it is discarded directly. After converting all other relationship types into BEFORE relationship, we get a new event relationship set. The final converted new event relationship is shown in Figure 3 .

[0076] 5.2) According to the new event relationship set, all event pairs with event relationship in each article are connected according to the time sequence of event occurrence as the initial flowchart. The initial flowchart constructed is shown in Figure 3 .

[0077] 5.3) Based on the initial flowchart, the initial flowchart corresponding to each article is divided into one or more flowchart seeds by depth-first traversal method. As can be seen from Figure 3 , the initial flowchart includes 3 flowchart seeds, which are separated to obtain three independent graphs.

[0078] 5.4) Add necessary start node, end node and gateway node three key nodes in the flowchart seed, and the final flow model is shown in Figure 4 .

[0079] 6) The training process of the consistency detection model, the true flow models in the data set are composed into flow model pairs, and the graph similarity score between the two is taken as the training data, which is put into model training.

[0080] 7) Process model consistency detection process, the input of the consistency detection model is the predicted flow model automatically constructed in step 5) and the true correct flow model, as shown in Figure 1The process model consistency detection module is shown, and its prediction process is as follows:

[0081] 7.1) Two graphs will get the node embedding representation of the process model through the GCN network. In addition, the Attention attention mechanism is used to consider the importance of each node in the whole process model, and each process model is mapped to the final graph-level embedding.

[0082] 7.2) The similarity vector of the graph-level embedding of the two process models and the correlation features of the nodes are calculated using the Neural Tensor Network module and the Pairwise Node Comparison module, respectively.

[0083] 7.3) The two vector features output in step 7.2) are spliced, and the final similarity score is obtained through a fully connected layer. At this time, the effectiveness of the automatic process modeling method can be measured according to the output similarity score.

[0084] For example, the similarity score calculated between the process model in Figure 4 and the real process model will be 100%. The following table shows the final results of the experiment:

[0085] Table 2: Prediction results of the automatic process model extraction method

[0086]

[0087] In Table 2, represents the average number of nodes contained in the real process model, represents the average number of nodes possessed by the automatically extracted process model. represents the average number of edges contained in the real process model, represents the average number of edges possessed by the automatically extracted process model. Score represents the similarity score of the process model pair, and the larger the value, the better the effect. Among them, the number of nodes contained in the extracted process model is similar, and the average number of edges is 4. The average similarity score between the real process model and the automatically extracted process model can reach 0.518.

[0088] For those skilled in the art, the technical solutions described in the foregoing examples can be modified, or some technical features can be replaced with equivalents. Any modification, equivalent replacement, etc. within the spirit and principles of the invention shall be included in the protection scope of the invention.

Claims

1. A method for automatic process modeling oriented to complex scenario description, characterized in that, The method comprises the following steps: 1) preparing a data set, specifically comprising: 1.1) articles, meeting the requirements of unstructured text; 1.2) marking the events existing in each article, and using the trigger words of the events to represent the events; 1.3) finding various relationships between events in the article; 1.4) constructing a flow model of the article according to the events and event relationships, combined with the semantics of the article; 2) performing a preprocessing operation on the input article based on the data set prepared in step 1) to obtain a preprocessed article; 3) training an event and event relationship joint extraction model based on the data set of step 1) and the preprocessed article of step 2); 4) performing event discovery and event relationship identification on the predicted text by the event and event relationship joint extraction model trained in step 3), and extracting the events and event relationships of the predicted text; 5) constructing a flow model based on the events and event relationships extracted in step 4), specifically comprising: 5.1) converting the extracted events and event relationships into a BEFORE relationship expression form to form new event relationships; 5.2) constructing an initial flowchart according to the new event relationships; 5.3) dividing the initial flowchart corresponding to each predicted text into one or more flowchart seeds through a depth-first traversal method based on the initial flowchart; 5.4) adding key nodes of the flow model to the flowchart seeds to form a flow model; 6) constructing a training data set for flow model consistency detection based on the flow model constructed in step 5), specifically comprising: 6.1) randomly combining the flow models to form a flow model pair; 6.2) Compute the graph edit distance between two process models by Hausdorff Distance and linear programming method , get the normalized graph edit distance by graph edit distance ; is standardized which is calculated as follows: ； wherein, the number of nodes of the flow model is represented. 6.3) Usage The function converts the normalized graph edit distance into a score in the range (0, 1] as a similarity score of the process model to the real correct, similarity score is calculated as follows: ; 7) training a consistency detection model based on the training data set constructed in step 6); 8) comparing the similarity of the flow models based on the consistency detection model trained in step 7), and detecting the difference between the automatically constructed flow model and the real flow model; Based on the consistency detection model trained in step 7), the similarity of the flow models is compared, specifically comprising: 8.1) obtaining the node embedding representation of the flow model through the GCN network; 8.2) considering the importance of each node in the flow model using the Attention attention mechanism, and mapping each flow model to the final graph-level embedding; 8.3) using Neural Tensor Network and Pairwise Node Comparison respectively to calculate the similarity vector of the graph-level embedding and the correlation feature of the nodes of the two flow models; 8.4) concatenating the similarity vector of the graph-level embedding and the correlation feature of the nodes, and passing through a fully connected layer to obtain the final flow model similarity score.

2. The method of claim 1, wherein the method further comprises: In step 2), the input article is preprocessed, specifically comprising: 2.1) removing stop words; 2.2) removing sentences that do not contain events; 2.3) dividing the article according to the limited length.

3. The method of claim 1, wherein the method further comprises: In step 3), the event and event relationship joint extraction model is trained based on the data set of step 1) and the preprocessed article of step 2), specifically comprising: 3.1) input the data set and the preprocessed article, firstly locally pass through the scoring function and use the cross-entropy loss function to optimize, train to obtain the event and event relation extraction model; 3.2) use structured inference learning to train the event and event relation extraction model obtained in step 3.1), re-optimize the network to adjust the global attribute.

4. The method of claim 1, wherein the method further comprises: In step 7, the consistency detection model is trained, specifically including: The training target of the flow model consistency detection is specifically the loss function: ； where D is a set of process model pairs in the training data, denotes the number of process model pairs, is the similarity score of a process model pair being correct, denotes the predicted similarity score of two process models, and the goal of the training is to minimize the loss L to achieve a higher accuracy.

Images