Lightweight process form intelligent prediction method and system

Abstract

The present invention belongs to the technical field of process form design, and specifically relates to a lightweight process form intelligent prediction method and system. The method includes: S1, obtaining process entity information and component entity information with unified structure from various business systems, and forming a data set; S2, generating a model dictionary, and performing enhanced training on the model dictionary according to the data set and through a machine learning method to obtain a data dictionary; the data dictionary has a data matching function; S3, according to the needs of different enterprises, importing the data dictionary into the system to provide a lightweight data model service; the data model service includes a prediction component list and an output process form. The present invention has the characteristics of enabling users to conveniently build forms and provide intelligent predictions for the building process, and the training process does not rely on a large amount of business data, and can quickly adapt to the changing business needs of enterprises.

Description

Technical Field

[0001] The present invention belongs to the technical field of process form design, and specifically relates to a lightweight process form intelligent prediction method and system. Background Art

[0002] At present, there are four main ways to build forms in common process engines:

[0003] 1. Create a new form and build the form component by component using the form building tool on the front-end page (usually by dragging and dropping);

[0004] 2. Copy the existing form, and then modify the copied form using the form building tool on the page;

[0005] 3. Automatically create a form by identifying files (word, pictures), and then modify it through the form building tool on the page;

[0006] 4. Based on the AI ​​big model to assist form modeling, multiple hypothetical answers are generated for users to choose through intelligent question-and-answer methods to guide users to build forms;

[0007] Except for the first method, which is built completely from scratch, the other methods are based on existing forms, documents or data to generate forms, and then complete the form building with subsequent manual adjustments. The fourth method provides the function of intelligent recommendation of AI big models in subsequent adjustments, and is also committed to giving users an intelligent experience in the process of form generation and subsequent adjustments, providing a form prediction service covering the entire cycle of form building.

[0008] However, it is known that the training of large AI models requires a large amount of complete business data, but most companies' business data cannot provide complete business data for large model training and learning due to the presence of sensitive information; in addition, different business systems of enterprises also have large differences in the expression of business logic at the data level, which will affect the validity of the input data and thus affect the accuracy of the training results; as the company develops, there will be changes in organization and business, and the composition of the forms will also be updated in large quantities. The AI ​​large model will need new fine-tuning or retraining accordingly, and the update cost will be relatively high.

[0009] In summary, the integration of AI technology in enterprise-level form modeling scenarios only applies the general semantic understanding capabilities of large models. However, due to the lack of uniformity and accuracy in the logical representation of various types of business data, training and fine-tuning are very difficult, and it is difficult to meet enterprise application scenarios.

[0010] Based on the above problems, it is very important to design a lightweight process form intelligent prediction method and system that enables users to easily build forms and provide intelligent predictions for the building process. The training process does not rely on a large amount of business data and can quickly adapt to the ever-changing business needs of enterprises. Summary of the invention

[0011] The present invention aims to overcome the problems in the prior art that the existing form building method has the problem that the training process needs to rely on a large amount of business data, the update cost is high and it is difficult to meet the needs of enterprise application scenarios. It provides a method and system for intelligent prediction of lightweight process forms that enables users to build forms conveniently and provide intelligent prediction for the building process. The training process does not rely on a large amount of business data, and can quickly adapt to the ever-changing business needs of enterprises.

[0012] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0013] The lightweight process form intelligent prediction method includes the following steps:

[0014] S1, obtain the process entity information and component entity information with unified structure from each business system and form a data set;

[0015] S2, generating a model dictionary, performing enhanced training on the model dictionary according to the data set and using a machine learning method to obtain a data dictionary; the data dictionary has a data matching function;

[0016] S3, according to the needs of different enterprises, a data dictionary is imported into the system to provide a lightweight data model service; the data model service includes a prediction component list and an output process form.

[0017] Preferably, in step S1, the process entity information includes a process ID and a process name; the component entity information includes a component ID, a component name, a component type and a sequence number.

[0018] Preferably, step S2 comprises the following steps:

[0019] S21, obtain entity information from the data set, and segment the process name and component name to generate semantic tokens; at the same time, maintain the relationship between the process template and the component list, allow the process information and all components to be obtained through the process ID, allow all associated components to be obtained through the token of the process name, allow all process templates with the same token to be obtained through the token of the process name, and allow the component list with the same token to be obtained through the component token; wherein the token refers to multiple single words or phrases generated after the text is decomposed according to semantics;

[0020] S22, permuting and combining the word-grams generated by the process name, and using the TF-IDF algorithm to count the weights of the process word-gram combinations corresponding to the associated component word-grams, screening out the key component word-grams corresponding to the process word-gram combinations, and storing them in the model dictionary;

[0021] S23, taking the process template list corresponding to the process word combination as the filtering range, filtering out the components existing in the corresponding process template range from the component list corresponding to the key component word, thereby obtaining the key component list corresponding to the process word combination;

[0022] S24, using the TF-IDF algorithm, performs algorithm statistics on the component words in the key component list to obtain word weights;

[0023] S25, performing similarity calculation on the components in the key component list, retaining components with similarity greater than a set threshold, and obtaining a similar component list; then assigning weight values ​​to the components in the similar component list by using the word unit weights obtained in step S24, for subsequent selection of components to be retained, to achieve similarity deduplication; the similarity calculation adopts cosine similarity calculation;

[0024] S26, merging the process word-gram combination and the key component word-gram, using the FP-Growth algorithm to perform association rule mining on the merged process word-gram combination and the key component word-gram, obtaining a word-gram relationship combination with a high degree of association, and performing the similarity deduplication processing operation of step S25 on the components associated in the word-gram relationship combination;

[0025] S27, generate a data dictionary mapping process word combinations and key component lists; the data dictionary supports obtaining a matching component list through industry information and process name keywords; the data dictionary supports the association relationship between process words and component words in step S26, so that users can input component names and recommend components corresponding to highly associated words.

[0026] Preferably, in step S25, the similarity deduplication specifically refers to taking the maximum value of the word unit weight for each component in the similar component list, and retaining the component with the largest maximum value of the corresponding word unit weight; the maximum value of the word unit weight is obtained by accumulating the weight values ​​of all words contained in a component.

[0027] Preferably, step S27 includes the following steps:

[0028] S271, constructing an initial data set based on the result of step S26, that is, a combination of component entities and all corresponding key component word-grams with similarities greater than a set threshold;

[0029] S272, classify the category names of the component entities again, and merge the classified data into the data set in step S271;

[0030] S273, according to the word-unit relationship combination obtained in step S26, obtain the corresponding classification label set, sort and number the classification labels, and generate a label vector of the word-unit relationship combination; each label vector is an array with a shape of (n_sames, n_class), where n_sames represents the corresponding word unit and n_class represents the number of corresponding labels;

[0031] S274, splitting the data set obtained in step S272 into a training set and a test set according to a preset ratio; training the model dictionary based on the training set and using a multi-label classification method in machine learning to obtain a data dictionary; during the training process, inputting a parameter X word element vector, a parameter Y label vector, selecting a naive Bayes classifier, and selecting a linear support vector machine model for the model dictionary;

[0032] S275, input the test set into the data dictionary to obtain a test result set, and construct a confusion matrix with the label vector of the original test set to check the difference between the predicted value and the label value;

[0033] S276, predicting the word-unit labels corresponding to the components existing in the current system or the components newly created by the user through the data dictionary, and storing the relationship between the component entity and the word-unit in the data dictionary.

[0034] Preferably, in step S3, the outputting process form includes the following process:

[0035] When the user creates a process template, the component list is matched from the data dictionary by the word element of the process name; at the same time, the final component arrangement is calculated based on the sorting of the components themselves and the components in the associated similar component list, and the component type is taken as the final type based on the high-frequency type in the component itself and the similar component list, thereby predicting a complete process form; the calculation of the final component arrangement specifically refers to finding the standard deviation of each component sort number, removing discrete points, and finding the average of the remaining component sort numbers.

[0036] Preferably, in step S3, the predicted component list includes the following process:

[0037] When a user creates a new component, the word element of the split process name is added to the word element of the new component, and a highly correlated word element relationship combination is matched from the data dictionary. Several highly correlated combination items are intercepted for the user to select the required component from the combination items.

[0038] The present invention also provides a lightweight process form intelligent prediction system including:

[0039] The data set construction module is used to obtain process entity information and component entity information with unified structure from various business systems and form a data set;

[0040] A model dictionary generation and enhancement training module is used to generate a model dictionary and perform enhancement training on the model dictionary according to the data set and through a machine learning method to obtain a data dictionary; the data dictionary has a data matching function;

[0041] The dictionary application module is used to import data dictionaries into the system according to the needs of different enterprises, and to provide lightweight data model services; the data model services include a prediction component list and an output process form.

[0042] Compared with the prior art, the present invention has the following advantages: (1) It has the characteristics of accurate data, lightweight model and easy implementation. The accuracy of data is reflected in the extraction of data sets, which accurately captures the entity associations of processes and components, avoids the entry of redundant information, and avoids the reading of sensitive data. The lightweight model means that there is no fine-tuning and training cost of large models. The information accuracy of the model data set itself is used to build the relationship between semantic words and entities (processes, components). Real-time business data is processed through machine learning to expand the business matching ability of the model dictionary. Easy implementation means that the model dictionary can be deployed on most application servers, and does not require computing support from hardware such as GPU to meet the business scenarios of form prediction. The deployment environment and operating costs are relatively lower than those of large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of the lightweight process sheet intelligent prediction method in the present invention. DETAILED DESCRIPTION

[0044] In order to more clearly illustrate the embodiments of the present invention, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings and other implementation methods can be obtained based on these accompanying drawings without creative work.

[0045] like Figure 1 As shown, the present invention provides a lightweight process form intelligent prediction method, comprising the following steps:

[0046] 1. Obtain process entity information and component entity information with unified structure from various business systems and form a data set;

[0047] 2. Generate a model dictionary, perform enhanced training on the model dictionary based on the data set and through a machine learning method to obtain a data dictionary; the data dictionary has a data matching function;

[0048] 3. According to the needs of different enterprises, a data dictionary is imported into the system to provide lightweight data model services; the data model services include a prediction component list and an output process form.

[0049] In step 1, the process entity information includes the process ID and process name, such as payment application, contract payment, and leave adjustment; the component entity information includes the component ID, component name, component type, and sequence number. In addition, the data set also includes industry information, which is used to describe the industry category of the enterprise, such as design and construction industries, and the template effective date, which is used to filter out outdated historical data.

[0050] Furthermore, step 2 includes the following steps:

[0051] 2-1. Obtain entity information from the data set, segment the process name and component name, and generate semantic tokens; at the same time, maintain the relationship between the process template and the component list, allow the process information and all components to be obtained through the process ID, allow all associated components to be obtained through the token of the process name, allow all process templates with the same token to be obtained through the token of the process name, and allow the component list with the same token to be obtained through the component token;

[0052] Among them, word-units refer to multiple words or phrases generated after the text is decomposed according to semantics; entities refer to objects that store complete information; the association relationship table between word-units and entities is as follows:

[0053] Table 1: The relationship between words and entities

[0054] ;

[0055] 2-2. Arrange and combine the tokens generated by the process name, and use the TF-IDF algorithm to count the weights of the process token combinations corresponding to the associated component tokens, filter out the key component tokens corresponding to the process token combinations, and store them in the model dictionary; for example, "department leave application" has three tokens: "department", "leave", and "application", and there are 6 corresponding combinations: "department", "leave", "application", "department leave", "leave application", and "department leave application".

[0056] The model dictionary is a dictionary used for prediction components, which stores entities, tokens, and their associations, and is used to find matching entities through input feature text;

[0057] TF-IDF (Term Frequency-Inverse Document Frequency) is a weighting technique used in information retrieval and text mining to evaluate the importance of a word to a document set or one of the documents in a corpus.

[0058] TF represents the frequency of a word in a document. It is calculated by dividing the number of times a word appears in an entity by the total number of words in the entity. The formula for calculating IDF is:

[0059] IDF = log (total number of entities in the corpus (component names or process names) / number of entities containing the word + 1).

[0060] TF-IDF is the product of TF and IDF, that is, TF-IDF = TFxIDF. It represents the inverse relationship between the importance of a word in a document and the prevalence of the word in the entire corpus.

[0061] 2-3. Take the process template list corresponding to the process word combination as the filtering range, filter out the components existing in the corresponding process template range from the component list corresponding to the key component word, and thus obtain the key component list corresponding to the process word combination;

[0062] 2-4. Use the TF-IDF algorithm to perform algorithm statistics on the component words in the key component list to obtain the word weight;

[0063] 2-5. Calculate the similarity of the components in the key component list, retain the components with similarity greater than the set threshold, and obtain a similar component list; then assign weight values ​​to the components in the similar component list through the word unit weights obtained in step S24, which are used for subsequent selection of components to be retained to achieve similarity deduplication (i.e., take the maximum value of the word unit weight for each component in the similar component list, and retain the component with the largest maximum value of the corresponding word unit weight; if a component has multiple word units, the weights are accumulated to obtain the maximum value of the word unit weight); the similarity calculation adopts cosine similarity calculation;

[0064] Among them, cosine similarity is an indicator to measure the similarity between two vectors. It expresses their similarity in the vector space by calculating the cosine value of the angle between two vectors. The present invention vectorizes the entity name (using the out-of-the-box method in text2vec), and calculates the cosine similarity between the vector and itself. The larger the cosine angle value, the higher the word-unit similarity. The names with high similarity are combined into a list, which is a similar name list.

[0065] 2-6. Merge the process word combination and the key component word, use the FP-Growth algorithm to mine association rules on the merged process word combination and the key component word, obtain word-unit relationship combinations with high association, and perform similarity deduplication processing in step 2-5 on the components associated with the word-unit relationship combination;

[0066] The basic idea of ​​the FP-Growth algorithm is to build an FP-Tree (frequent item set tree) to represent the frequent item sets in the data set and mine frequent item sets from the FP-Tree. FP-Tree is an efficient data structure that can mine frequent item sets without generating candidate frequent item sets. Here, transactions refer to the combination of process tokens and component tokens, and frequent items refer to the combination of single or multiple tokens.

[0067] 2-7. Generate a data dictionary mapping process word combinations and key component lists; the data dictionary supports obtaining a matching component list through industry information and process name keywords; the data dictionary supports the association relationship between process words and component words in step 2-5, so that users can input component names and recommend components corresponding to highly associated words.

[0068] For steps 2-7, the design converts the prediction task of component keyword meta-data under a specific business into a multi-label classification task for component meta-data. Because the process meta-data is associated with the key component meta-data, these key component meta-data are the classification items of the component names in the dataset. Through the development set training, the entity names are vectorized and classified into keyword meta-data, and the entity names in the test set are also matched and verified accordingly. Specifically, steps 2-7 include the following steps:

[0069] 2-7-1. Construct the initial data set based on the result of step 2-6, that is, the combination of component entities and all corresponding key component word-grams whose similarity is greater than the set threshold; for example, the start time of leave is classified into "start time", "start", and "time". Here, if "leave" is also a key word-gram, the word-gram combination filters out the full name and retains the more common word-gram combination.

[0070] 2-7-2. Classify the category names of component entities again, for example, "start time" is classified into "time" and "start", and merge the classified data into the data set in step 2-7-1;

[0071] 2-7-3.According to the word-gram relationship combination obtained in step 2-6, obtain the corresponding classification label set, sort and number the classification labels, and generate the label vector of the word-gram relationship combination; each label vector is an array with a shape of (n_sames, n_class), where n_sames represents the corresponding word-gram and n_class represents the number of corresponding labels; for example, if the word-gram combination hits the label, it is recorded as 1, otherwise it is 0, and the array "[[0,1,1],[0,0,0]]" is obtained, which means that the 0th word-gram hits the 1st and 2nd labels, and the 1st word-gram does not hit any label.

[0072] 2-7-4. Split the data set obtained in step S272 into a training set and a test set according to a preset ratio, such as 85% of the training set and 15% of the test set; train the model dictionary based on the training set and use the multi-label classification method in machine learning to obtain a data dictionary; during the training process, input parameter X word vector, parameter Y label vector, select the naive Bayes classifier, and select the linear support vector machine model (LinearSVC) for the model dictionary;

[0073] Among them, machine learning methods are used to predict keyword meta-words (that is, labels) that match the input entity name, help classify the entity into the model dictionary, build new associations between meta-words and entities, strengthen the dictionary matching range and provide more real-time business matching capabilities.

[0074] 2-7-5. Input the test set into the data dictionary to obtain the test result set, and construct a confusion matrix (sklearn.metrics.confusion_matrix) with the label vector of the original test set to check the difference between the predicted value and the label value;

[0075] 2-7-6, predict the corresponding word meta tags of the components in the current system or the components newly created by the user through the data dictionary, and store the relationship between the component entity and the word meta tag in the data dictionary. When the user enters a new process name, the system obtains the corresponding keyword meta tag based on the process word meta tag, obtains a list of component entities with high matching degree from the dictionary, and can recommend the entity list to the user or generate a prediction form based on the information of the component entity (name, type, sorting).

[0076] As users use the model, more user data is added to the model, and the matching degree between the model and the current user will be enhanced. As data accumulates, the initial data set can be repeatedly trained to optimize the matching accuracy of user usage.

[0077] About model dictionary application: In response to different enterprise needs, business model dictionaries (data dictionaries) are imported into the system. The scope of dictionary data can be determined mainly by effective time and industry description, providing lightweight data model services. The application covers two major scenarios, including prediction component list and output process form

[0078] The output process form includes the following steps:

[0079] When the user creates a process template, the component list is matched from the data dictionary by the word element of the process name (step 2-7-5); at the same time, the final component arrangement is calculated based on the sorting of the components themselves and the components in the associated similar component list, and the component type is taken as the final type based on the high-frequency type in the component itself and the similar component list, thereby predicting a complete process form; the calculation of the final component arrangement specifically refers to finding the standard deviation of each component sort number, removing discrete points, and finding the average of the remaining component sort numbers.

[0080] The forecast component list includes the following processes:

[0081] When a user creates a new component, the word element of the split process name is added to the word element of the new component, and a highly correlated word element relationship combination is matched from the data dictionary (step 2-7-6), and several highly correlated combination items are intercepted for the user to select the required component from the combination items.

[0082] The present invention also provides a lightweight process form intelligent prediction system including:

[0083] The data set construction module is used to obtain process entity information and component entity information with unified structure from various business systems and form a data set;

[0084] A model dictionary generation and enhancement training module is used to generate a model dictionary and perform enhancement training on the model dictionary according to the data set and through a machine learning method to obtain a data dictionary; the data dictionary has a data matching function;

[0085] The dictionary application module is used to import data dictionaries into the system according to the needs of different enterprises, and to provide lightweight data model services; the data model services include a prediction component list and an output process form.

[0086] The following will refer to Figure 1 Describe the exemplary embodiments of the present application in more detail. As mentioned above, AI technology in enterprise-level form modeling scenarios is also based on the understanding of semantics, but it relies on a large amount of data content that expresses the completeness, and the training and fine-tuning are costly. A lightweight intelligent design based on machine learning can quickly respond to the needs of business development of enterprises. And using a simple data structure to obtain element information can avoid the complexity of business data of different systems. In practical applications, the method of the present invention specifically includes the following steps:

[0087] 1. Obtain a unified data set. Data can be extracted from different business systems in a unified format. The information is semantically divided into three levels: business description, process description, and component description. It also contains effective date information to filter out obsolete data. Data can be obtained in a variety of forms, such as database acquisition or Excel template import. The implementation method is explained using the following data as an example: ["personnel related", "employee leave form", "leave start time", "leave end time"].

[0088] 2. Generate word combinations by segmenting the description information. On this basis, obtain the word combinations of key components through statistical methods. This step builds a model dictionary under the "personnel" business description, and there will be word combinations ["employee", "leave", "employee leave", "start", "end", "time", "leave start", "leave end", "start time", "end time"...].

[0089] 3. Maintain the association between word units and entities (components, processes), as shown in Table 1, and generate a model dictionary to save these data associations. This step will have the relationship between component word units and entities, such as the "time" word unit is associated with the two components "leave start time" and "leave end time".

[0090] 4. Calculate the keyword-meta combination corresponding to the process-meta by statistical methods, that is, perform feature extraction of a layer of process component list. This step will generate keyword-meta combinations. For example, the process-meta “leave” will generate keyword-meta [“leave”, “leave start”, “leave end”, “start time”, “end time”].

[0091] 5. Obtain the key component entity list of the process entity through the process component keyword element. So far, the model dictionary has obtained entities, word elements, keyword elements, key entities and their relationships.

[0092] Based on the model word in step 5, the subsequent business processing is divided into three branches. Component prediction, that is, predicting other subsequent component combinations by using the input component name. Form prediction, that is, predicting possible component combinations by using the input process name. Machine learning empowers form prediction, which further enhances the data of the generated model dictionary through machine learning methods, so that the system usage data can also take effect quickly and be included in the matching range of the model dictionary.

[0093] 6. Traverse the process tokens, merge the process tokens and the associated key component tokens, and obtain high-correlation token combinations through high-correlation mining calculations. There may be multiple groups of correlation token combinations, such as ["leave", "leave start", "leave end"], ["time", "start time", "end time"].

[0094] 7. Obtain the component entity list through the word combination in step 6 and perform similarity deduplication. The association between each word combination and the final entity list is saved in the model dictionary. The result is used for component prediction. This is actually deduplication of the component list associated with the recommended word. For example, in the above example, the word "leave end" corresponds to multiple component instances. Here, the component instance with the highest similarity match with the word is selected.

[0095] 8. Traverse the process tokens, read the associated key component tokens (i.e., step 2-2 above), and use the tokens to obtain the key component list for similarity removal. Save the association between the process tokens and the final component list in the model dictionary. This result is used for form prediction. Here, the dictionary will store the process token "leave" and the component list ["leave start time", "leave end time"].

[0096] 9. This step describes the component prediction scenario. When the user creates a component on the page, the system segments the input component name, obtains the component list through word matching in the model dictionary, and returns it to the front end. After the user selects a component, a corresponding component is created on the form. In this way, when the user enters "leave start time", the component entity corresponding to "leave end" will be recommended; if the user enters "start time", the component entity corresponding to "end time" will be recommended. In this example, "leave end time" is returned.

[0097] 10. This step describes the form prediction scenario. When the user creates a template on the page, the system segments the input process name, obtains the component list through word matching in the model dictionary, and returns it to the front end. The front end creates form components in batches based on the component entity information. In this way, when the user enters "leave application", the form component list ["leave start time", "leave end time"] will be predicted, and then the system will create a form according to the component list information.

[0098] 11. This step introduces the generation of machine learning models (application of machine learning methods). Here, it is assumed that there is a large amount of form modeling data. The model can be trained based on the component keyword metadata associated with the process vocabulary as the sample range, so the model result belongs to the process vocabulary; at the same time, considering the small amount of vocabulary data, all component keyword metadata can also be used as the sample range for training. Determine the vocabulary range to obtain the vocabulary list. Through the model dictionary, you can get a list of all component entities, which is the sample for model training. According to the above introduction, get the corresponding machine learning model.

[0099] 12. In order to enhance the matching degree of the model dictionary to specific business scenarios and enable it to have the ability of "frequent use and constant innovation", the present invention uses a machine learning model to process the newly created component data. When the front-end user creates a new component, the component name is input into the machine learning model to predict its associated keyword meta-combination. Then the predicted keyword meta-combination and the newly created component are incorporated into the model dictionary. This will affect two result sets, one is the component prediction result set of step 7, which stores the relationship between the keyword meta-combination and the component entity; the other is the form prediction result set of step 8, which also updates the relationship between the meta-combination and the component entity into the form prediction result set. The following is a detailed description. For example, if the user creates a new component named "Reason for Leave", it can be input into the trained model to predict the keyword meta-combination, such as returning ["Leave", "Reason"] two keyword meta-combinations (multi-label tasks often return multiple labels). Here, the name of the current process template "Leave Application Form" is obtained through the model dictionary, and there are two process meta-combinations "Leave" and "Application". The model dictionary stores the key component words associated with the process word to match the component words "leave" and "reason". For example, if the key component word list of the process word "leave" contains "leave", then the "leave reason" component is added to the component entity list corresponding to this key word; and if the key component word of the process word "application" contains "reason", then the "leave reason" component is added to the entity list corresponding to "reason". This step enriches the model dictionary to support the prediction scenarios of steps 9 and 10, and achieves "frequently used and always new".

[0100] The innovative features of the present invention are as follows:

[0101] 1. The semantics of the process name and the semantics of the form component are associated with each other by using tokens as the basic units.

[0102] 2. Use word combination as a way to enhance the accuracy of semantic matching.

[0103] 3. Form prediction and component prediction are two sets of calculation logic. Form prediction requires a relatively strict match of the process name, and the prediction process has a stricter double check. First, the process name is used to filter out the key component words, and then the specific key component list is filtered out from the entities corresponding to the key component words and process words. One layer is the filtering between words, and the other layer is the filtering between entities (templates, component entities). This strengthens the constraints of the initial word (process name) on the entire output and effectively reduces redundant information output. Component prediction emphasizes the association between word semantics, weakens the matching of entities, and obtains word combinations through association rule mining, so that users can be given a broader recommendation in the semantic dimension.

[0104] 4. Use machine learning to generate lightweight small models to support business scenarios of form prediction. The user's real-time usage data is included in the data dictionary through model prediction, which reduces the amount of data and training time of the model, and also reduces the complexity of training. At the same time, the unified data set structure has universal adaptability to various business data, and also avoids the reading of sensitive business data. The data dictionary supports multiple trainings as the initial data set, further enhancing the matching accuracy of the model to the user's usage data.

[0105] The above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, according to the ideas provided by the present invention, there will be changes in the specific implementation methods, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. Lightweight process form intelligent prediction method, characterized by: The method comprises the following steps: S1, obtain the process entity information and component entity information with unified structure from each business system and form a data set; S2, generating a model dictionary, performing enhanced training on the model dictionary according to the data set and using a machine learning method to obtain a data dictionary; the data dictionary has a data matching function; S3, according to the needs of different enterprises, import the data dictionary into the system to provide lightweight data model services; the data model services include a prediction component list and an output process form; Step S2 includes the following steps: S21, obtain entity information from the data set, and segment the process name and component name to generate semantic tokens; at the same time, maintain the relationship between the process template and the component list, allow the process information and all components to be obtained through the process ID, allow all associated components to be obtained through the token of the process name, allow all process templates with the same token to be obtained through the token of the process name, and allow the component list with the same token to be obtained through the component token; wherein the token refers to multiple single words or phrases generated after the text is decomposed according to semantics; S22, permuting and combining the word-grams generated by the process name, and using the TF-IDF algorithm to count the weights of the process word-gram combinations corresponding to the associated component word-grams, screening out the key component word-grams corresponding to the process word-gram combinations, and storing them in the model dictionary; S23, taking the process template list corresponding to the process word combination as the filtering range, filtering out the components existing in the corresponding process template range from the component list corresponding to the key component word, thereby obtaining the key component list corresponding to the process word combination; S24, using the TF-IDF algorithm, performs algorithm statistics on the component words in the key component list to obtain word weights; S25, performing similarity calculation on the components in the key component list, retaining components with similarity greater than a set threshold, and obtaining a similar component list; then assigning weight values ​​to the components in the similar component list by using the word unit weights obtained in step S24, for subsequent selection of components to be retained, to achieve similarity deduplication; the similarity calculation adopts cosine similarity calculation; S26, merging the process word-gram combination and the key component word-gram, using the FP-Growth algorithm to perform association rule mining on the merged process word-gram combination and the key component word-gram, obtaining a word-gram relationship combination with a high degree of association, and performing the similarity deduplication processing operation of step S25 on the components associated in the word-gram relationship combination; S27, generating a data dictionary mapping process word combinations and key component lists; the data dictionary supports obtaining a matching component list through industry information and process name keywords; the data dictionary supports the association relationship between process words and component words in step S26, enabling users to input component names and recommending components corresponding to highly associated words; Step S27 includes the following steps: S271, constructing an initial data set based on the result of step S26, that is, a combination of component entities and key component words after deduplication; S272, classify the category names of the component entities again, and merge the classified data into the data set in step S271; S273, according to the word-unit relationship combination obtained in step S26, obtain the corresponding classification label set, sort and number the classification labels, and generate a label vector of the word-unit relationship combination; each label vector is an array with a shape of (n_sames, n_class), where n_sames represents the corresponding word unit and n_class represents the number of corresponding labels; S274, splitting the data set obtained in step S272 into a training set and a test set according to a preset ratio; training the model dictionary based on the training set and using a multi-label classification method in machine learning to obtain a data dictionary; during the training process, inputting a parameter X word element vector, a parameter Y label vector, selecting a naive Bayes classifier, and selecting a linear support vector machine model for the model dictionary; S275, input the test set into the data dictionary to obtain a test result set, and construct a confusion matrix with the label vector of the original test set to check the difference between the predicted value and the label value; S276, predicting the word-unit labels corresponding to the components existing in the current system or the components newly created by the user through the data dictionary, and storing the relationship between the component entity and the word-unit in the data dictionary.

2. The lightweight process form intelligent prediction method according to claim 1 is characterized in that: In step S1, the process entity information includes a process ID and a process name; the component entity information includes a component ID, a component name, a component type and a sequence number.

3. The lightweight process form intelligent prediction method according to claim 1 is characterized in that: In step S25, the similarity deduplication specifically refers to taking the maximum value of the word unit weight for each component in the similar component list, and retaining the component with the largest maximum value of the corresponding word unit weight; the maximum value of the word unit weight is obtained by accumulating the weight values ​​of all words contained in a component.

4. The lightweight process form intelligent prediction method according to claim 1 is characterized in that: In step S3, the output process form includes the following process: When the user creates a process template, the component list is matched from the data dictionary by the word element of the process name; at the same time, the final component arrangement is calculated based on the sorting of the components themselves and the components in the associated similar component list, and the component type is taken as the final type based on the high-frequency type in the component itself and the similar component list, thereby predicting a complete process form; the calculation of the final component arrangement specifically refers to finding the standard deviation of each component sort number, removing discrete points, and finding the average of the remaining component sort numbers.

5. The lightweight process form intelligent prediction method according to claim 4 is characterized in that: In step S3, the predicted component list includes the following process: When a user creates a new component, the word element of the split process name is added to the word element of the new component, and a highly correlated word element relationship combination is matched from the data dictionary. Several highly correlated combination items are intercepted for the user to select the required component from the combination items.

6. A lightweight process form intelligent prediction system, used to implement the lightweight process form intelligent prediction method according to any one of claims 1 to 5, characterized in that: The lightweight process form intelligent prediction system includes: The data set construction module is used to obtain process entity information and component entity information with unified structure from various business systems and form a data set; A model dictionary generation and enhancement training module is used to generate a model dictionary and perform enhancement training on the model dictionary according to the data set and through a machine learning method to obtain a data dictionary; the data dictionary has a data matching function; The dictionary application module is used to import data dictionaries into the system according to the needs of different enterprises, and to provide lightweight data model services; the data model services include a prediction component list and an output process form.

Images