CN117290188A - Event log obtaining method based on task mining technology - Google Patents

Abstract

The application discloses an event log obtaining method based on a task mining technology, which relates to the field of task mining and comprises the following steps: when the implementation process mining project is started, the business operation of the enterprise universe is reserved by using a task mining technology; step 2: setting the format and attribute consistency of the event log data imported by the task mining and saving event log data and the process mining tool; step 3: according to the data format required by importing and cleaning the data by the process mining tool, storing event log data reserved by the task mining technology; step 4: the saved event log data is imported into a process mining tool for format test; step 5: after the test is passed, the imported reserved data can be used, the event log data required by the process mining is normalized and combed, the input data required by the process mining tool is generated, the data can be acquired in a cross-platform and cross-system mode, and the problems that the process mining tool is difficult to comb and clean the data, the effective data is less and the like are fundamentally solved.

Description

Event log obtaining method based on task mining technology

Technical Field

The invention relates to the field of information technology and artificial intelligence, and is suitable for all projects and enterprises using process mining technology, in particular to projects with complicated early event data and inconsistent process mining cleaning conditions.

Background

The process mining is a technology based on the collection and monitoring of event log information generated in work, the restoration of real and detailed business processes, and the analysis and optimization of the restored work processes, and the business processes can be all-round ground-based by using the technology, so that the operation efficiency is greatly improved.

Task mining techniques use user interaction data (also referred to as desktop data) to record tasks in a process, during which the data involved in the process can be saved as desired. This type of data includes keystrokes, mouse clicks, and data inputs that occur during the completion of a given operation, and may be recorded across systems, terminals, accounts, etc. Task mining may use Optical Character Recognition (OCR), natural Language Processing (NLP), and machine learning algorithms to interpret and record these data, and then use event log data that is fully compliant with the flow mining data import format as the underlying event log for the flow mining. Thus, analysts and other stakeholders can identify operating conditions, overall process operating conditions, etc.

The event log information is a leading condition for exerting energy efficiency of the whole process, and belongs to the important importance of the whole process, but the reserved data of an ERP system, a financial processing system and the like used in the early stage of an enterprise are mixed, the data format is disordered, so that the availability of the event log is low, the data quality is low, and the whole efficiency of the process mining technology cannot be fully exerted. The invention solves the problems of standard and efficient and reasonable data format of the event log, namely a method for generating the event log based on a task mining technology.

Disclosure of Invention

In order to solve the above problems in the prior art, the present invention has an object of: event log data required by process mining is normalized and is generated, input event log data required by an enterprise usable and efficient process mining tool can be acquired in a cross-platform and cross-system mode, and the problems that the process mining tool is difficult to comb and clean data, effective data are few and the like are fundamentally solved.

In order to achieve the above purpose, the main technical scheme of the invention is as follows: the event log obtaining method based on the task mining technology comprises the following steps of:

step 1: when the implementation process mining project is started, the business operation of the enterprise universe is reserved by using a task mining technology;

step 2: setting the format and attribute consistency of the event log data imported by the task mining and saving event log data and the process mining tool;

step 3: according to the data format required by importing and cleaning the data by the process mining tool, storing event log data reserved by the task mining technology;

step 4: the saved event log data is imported into a process mining tool for format test;

step 5: after the test is passed, the imported and retained data can be used in a large scale.

Further, the step 1 specifically includes: when an enterprise starts to implement a process mining project, related information of business operations is captured and extracted by using a task mining technology, and recording and retaining of the business operations are ensured, wherein the related information comprises user interaction, system logs and event records.

Further, the task mining technique includes:

step 1.1: preprocessing initial text data through reverse translation, translating the initial text to be subjected to task mining into foreign language, and translating the foreign language back into Chinese, so that the balance of the data is maintained;

step 1.2: the method comprises the steps of marking nouns on a preprocessed text, extracting nouns in the preprocessed text, finding all synonym sets of the nouns on DANet, determining meanings of the nouns in a context by means of word sense disambiguation of glosbert, obtaining correct synonym sets of the nouns, and finally replacing the nouns with synonyms to create a new input text;

step 1.3: encoding the input text using a DeBERTa model; generating a knowledge graph corresponding to the input text by using a knowledge graph generating system, copying the generated knowledge graph to enable the generated knowledge graph to correspond to the input text after data enhancement, and inputting the knowledge graph into a knowledge graph encoder;

step 1.4: the input text is encoded to obtain an input text embedding, the corresponding knowledge graph is encoded to obtain a knowledge graph embedding, and then the input text embedding and the knowledge graph embedding are input into a fusion module;

step 1.5: and sending the combined result at the fusion module into a linear layer, and obtaining the task information of the text through softmax operation.

Further, the fusion module in the step 1.5 includes a self-attention module, a linear layer, a multi-head cross-attention module and a merging layer which are sequentially connected, and text contents are input from the self-attention module and output from the merging layer.

Further, the processing procedure of the fusion module in the step 1.5 is as follows: firstly, an input text is embedded and a knowledge graph is embedded through a self-attention module, then the input text and the knowledge graph are input into a linear layer to obtain the input of a multi-head cross attention layer, the input text is embedded to serve as the query input of the multi-head cross attention layer, the knowledge graph is embedded to serve as two supporting inputs of the multi-head cross attention layer, and then the output of the multi-head cross attention layer and the input text are embedded and combined, so that the original information embedded by the input text can be kept from being lost.

Further, the attribute in the step 2 includes: field naming, data type, and data structure.

Furthermore, the cleaning data in the step 3 is performed by using a large language model LLMs and an iterative processing method.

Further, the method specifically comprises the following steps:

step 3.1: inputting data to be cleaned and a specified format into large language model LLMs, and then cleaning the data according to the provided format by the large language model LLMs to obtain data I;

step 3.2: iterative cleaning is carried out: inputting the obtained data I and a specified format into the large language model LLMs again to obtain data II; then inputting the data II and the specified format into the large language model LLMs again to obtain data III;

step 3.3: and performing cross entropy loss on the first data, the second data, the third data and the original data, and controlling the validity of the data obtained by each iteration.

Compared with the prior art, the beneficial effects of this application are:

[1] to solve the problem of data imbalance in task mining, we use reverse translation to enrich the data in fewer classes.

[2] In terms of task mining data expansion, we use DANet and glosbert to create additional data to train our task mining network by replacing noun synonyms in the original text.

[3] In order to make up for the information which is easy to be lost in the task mining, the knowledge graph is utilized to jointly perform the task mining, so that the task mining performance is greatly improved.

[4] In the data cleaning process, the data is cleaned by using an iterative processing mode, so that the effectiveness of the data is ensured, and the cleaned data can be more suitable for the following flow.

Drawings

Fig. 1 is a schematic diagram of a data flow direction in an event log of a mining tool in the prior art.

Fig. 2 is a flowchart of a log obtaining method based on a task mining technology in the present invention.

FIG. 3 is a schematic overall flow chart of the task mining technique of the present invention.

FIG. 4 is a schematic diagram of a fusion module structure and a workflow according to the present invention.

FIG. 5 is a flow chart of the cross entropy loss step in the present invention.

Detailed Description

The technical scheme of the invention is further explained below with reference to the specification, the drawings and the specific embodiments.

Fig. 1 is a schematic diagram of a data flow direction in an event log of a mining tool in the prior art. The ERP system, the financial processing system and the like used in the early stage of enterprises have mixed reserved data and chaotic data formats, so that the availability of event logs is low, the data quality is low, and the full efficacy of the process mining technology cannot be fully exerted.

Fig. 2 is a flowchart of a log obtaining method based on the task mining technology in the present invention. Event log data required by process mining is normalized and is generated, input event log data required by an enterprise usable and efficient process mining tool can be acquired in a cross-platform and cross-system mode, and the problems that the process mining tool is difficult to comb and clean data, effective data are few and the like are fundamentally solved.

The event log obtaining method based on the task mining technology comprises the following steps:

step 1: and when the implementation process mining project is started, the business operation of the enterprise domain is reserved by using a task mining technology. The method comprises the following specific steps:

the enterprise should ensure the recording and retention of business operations when beginning to implement the process mining project. Task mining techniques may be used to capture and extract relevant information for business operations, such as user interactions, system logs, event records, and the like. The data is retained for subsequent process mining and analysis.

Task mining is an emerging technology. Task mining techniques focus on tasks, i.e., smaller components of a process or sub-process that contain multiple steps, typically performed manually by an employee at their workstation. It enables businesses to better understand how they perform tasks by tracking user activities and collecting user interaction information. Enterprises can use the collected information to review how they manage operations, identify the most common errors in executing jobs, and identify tasks that can be automated.

The specific process of the task mining method used in this embodiment is as follows:

reverse translation techniques are employed to solve the problem of data imbalance. A data enhancement network dant is then employed to enhance the text data. Meanwhile, in order to solve the problems of inaccurate mining and the like caused by directly mining tasks, a knowledge graph is introduced to assist in mining the tasks. The present embodiment uses the latest DeBERTA as the text encoder and KGE module as the knowledge-graph encoder. The overall flow of the task mining method is shown in fig. 3.

In the text of the task mining process, many problems exist, such as uneven keyword distribution, ambiguous mining information, and the like. To solve these problems, preprocessing of text data is required. According to the method, the problem of uneven data is solved through reverse translation, text to be subjected to task mining is translated into English or other languages, then translated back into Chinese, and the balance of the data can be maintained after pretreatment.

More data tends to give better results to the model because the model training data is more data. The embodiment firstly marks nouns on the text to be mined, and then extracts nouns therein. We then find all of their synonym sets on dant, and use the word sense disambiguation of glosbert to determine the meaning of these nouns in the context, resulting in their correct synonym set. Finally, these nouns are replaced with synonyms to create new input text.

To encode the input, the present embodiment uses DeBERTa. It exhibits good performance in various natural language processing tasks. In order to better mine task information, a knowledge graph is preferably introduced to supplement some key information to assist task mining, a knowledge graph generating system is utilized to generate a knowledge graph corresponding to the input text, and then the generated knowledge graph is subjected to copying operation so that the generated knowledge graph can correspond to the input text after data enhancement. And then inputting the knowledge graph into a knowledge graph encoder (KGE module). Wherein the knowledge-graph encoder is composed of the latest large language model.

The input text is encoded and then embedded into the input text. And obtaining the embedded knowledge graph after encoding the corresponding knowledge graph. And then embedding the input text and the knowledge graph into an input fusion module.

The fusion module is shown in fig. 4. Including self-attention modules and multi-headed cross-attention modules. The input text embedding and knowledge graph embedding are first performed by a self-attention module, and then are input into a linear layer to obtain the input of a multi-head cross attention layer. The input text is embedded as a query input for the multi-headed cross-attention layer, and the knowledge graph is embedded as two support inputs for the multi-headed cross-attention layer. The output and input text embeddings of the multi-headed cross-attention layer are then combined, which can preserve the original embedded information of the input text from being lost.

And finally, the combined result passes through a linear layer, and finally, task information of the text is obtained through softmax operation.

Step 2: setting the format and the attribute consistency of the event log data which are saved by the task mining and the event log data imported by the process mining tool.

In order to ensure that the event log data saved by task mining and the process mining tool can be successfully imported and processed, consistency of data formats and attributes needs to be ensured. This includes matching in terms of field naming, data type, data structure, etc.

Step 3: and (3) importing and cleaning the data according to a data format required by the process mining tool, and storing event log data reserved by the task mining technology.

And importing and cleaning the event log data reserved by the task mining technology according to the requirements of the process mining tool. The method comprises the steps of data format conversion, data cleaning, preprocessing and the like, so that the data is suitable for the requirements of subsequent process mining tools.

The method utilizes a large language model LLMs and an iterative processing mode to clean data, and the specific implementation process is as follows:

firstly, inputting data to be cleaned and a specified format into a large language model LLMs, and then cleaning the data according to the provided format by the LLMs to obtain data 1. The format of the data in data 1 has preliminarily satisfied the requirements of the subsequent flow. But in order to obtain more satisfactory data we need to perform iterative cleaning.

Then, the obtained data 1 and the prescribed format are input into the large language model LLMs again to obtain data 2. Then, the data 2 and the prescribed format are input into the large language model LLMs again to obtain data 3, and the iteration is performed. In our method, three iterations are performed.

And finally obtaining the data after cleaning after three iterations, wherein the data basically accords with a specified format and can be used for subsequent operation.

In order to control the validity of the data obtained for each iteration, that is to say the data after each iteration does not change its meaning. We have cross entropy loss of data 1, data 2 and the original data, and the specific flow is shown in fig. 5.

Step 4: the saved event log data is imported into a process mining tool for format test;

and importing the cleaned and converted event log data into a process mining tool, and performing format test. This is to ensure that the data is properly imported into the tool and that the tool is able to properly parse and process the data.

Step 5: after the test is passed, the imported and retained data can be used in a large scale.

Through the scheme, related software and keywords in the workflow can be set by designing the task mining tool, when related operations are triggered, the task mining log generating tool is started automatically, all working operations of an enterprise are recorded by the task mining technology, and formatted event logs are generated and stored. And then, importing the formatted event log generated by the task mining technology into a subsequent flow mining system and tool to perform subsequent operations such as data inspection and the like. The scheme is mainly oriented to the data processing link of the process mining technology, low-quality time log data which is difficult to process and format is rewritten by using the task mining technology.

While the invention has been described with reference to preferred embodiments, it is not intended to be limiting. Those skilled in the art will appreciate that various modifications and adaptations can be made without departing from the spirit and scope of the present invention. Accordingly, the scope of the invention is defined by the appended claims.

Claims (8)

1. The event log obtaining method based on the task mining technology is characterized by comprising the following steps of:

step 1: when the implementation process mining project is started, the business operation of the enterprise universe is reserved by using a task mining technology;

step 2: setting the format and attribute consistency of the event log data imported by the task mining and saving event log data and the process mining tool;

step 3: according to the data format required by importing and cleaning the data by the process mining tool, storing event log data reserved by the task mining technology;

step 4: the saved event log data is imported into a process mining tool for format test;

step 5: the imported persisted data may be used after the test passes.

2. The method for obtaining the event log based on the task mining technology according to claim 1, wherein the step 1 specifically comprises: when an enterprise starts to implement a process mining project, related information of business operations is captured and extracted by using a task mining technology, and recording and retaining of the business operations are ensured, wherein the related information comprises user interaction, system logs and event records.

3. The method for obtaining an event log based on a task mining technique according to claim 2, wherein the task mining technique comprises:

step 1.1: preprocessing initial text data through reverse translation, translating the initial text to be subjected to task mining into foreign language, and translating the foreign language back into Chinese, so that the balance of the data is maintained;

step 1.2: the method comprises the steps of marking nouns on a preprocessed text, extracting nouns in the preprocessed text, finding all synonym sets of the nouns on DANet, determining meanings of the nouns in a context by means of word sense disambiguation of glosbert, obtaining correct synonym sets of the nouns, and finally replacing the nouns with synonyms to create a new input text;

step 1.3: encoding the input text using a DeBERTa model; generating a knowledge graph corresponding to the input text by using a knowledge graph generating system, copying the generated knowledge graph to enable the generated knowledge graph to correspond to the input text after data enhancement, and inputting the knowledge graph into a knowledge graph encoder;

step 1.4: the input text is encoded to obtain an input text embedding, the corresponding knowledge graph is encoded to obtain a knowledge graph embedding, and then the input text embedding and the knowledge graph embedding are input into a fusion module;

step 1.5: and sending the combined result at the fusion module into a linear layer, and obtaining the task information of the text through softmax operation.

4. The method according to claim 3, wherein the fusion module in step 1.5 includes a self-attention module, a linear layer, a multi-head cross-attention module, and a merging layer connected in sequence, and text content is input from the self-attention module and output from the merging layer.

5. The method for acquiring the event log based on the task mining technology according to claim 4, wherein the processing procedure of the fusion module in step 1.5 is as follows: firstly, an input text is embedded and a knowledge graph is embedded through a self-attention module, then the input text and the knowledge graph are input into a linear layer to obtain the input of a multi-head cross attention layer, the input text is embedded to serve as the query input of the multi-head cross attention layer, the knowledge graph is embedded to serve as two supporting inputs of the multi-head cross attention layer, and then the output of the multi-head cross attention layer and the input text are embedded and combined, so that the original information embedded by the input text can be kept from being lost.

6. The method for obtaining an event log based on task mining according to claim 1, wherein the attribute in step 2 includes: field naming, data type, and data structure.

7. The method for acquiring the event log based on the task mining technology according to claim 1, wherein: the cleaning data in the step 3 is performed by using a large language model LLMs and an iterative processing mode.

8. The method for acquiring the event log based on the task mining technology as set forth in claim 7, wherein the method specifically comprises the following steps:

step 3.1: inputting data to be cleaned and a specified format into large language model LLMs, and then cleaning the data according to the provided format by the large language model LLMs to obtain data I;

step 3.2: iterative cleaning is carried out: inputting the obtained data I and a specified format into the large language model LLMs again to obtain data II; then inputting the data II and the specified format into the large language model LLMs again to obtain data III;

step 3.3: and performing cross entropy loss on the first data, the second data, the third data and the original data, and controlling the validity of the data obtained by each iteration.