Event processing flow recommendation method and system based on deep learning

Abstract

The invention discloses an event processing flow recommendation method and system based on deep learning, and belongs to the technical field of artificial intelligence and information technology application, and the method comprises the steps: multi-modal feature extraction and fusion: constructing a comprehensive case feature portrait by using the structured information and unstructured text description of a case at the same time; constructing a process recommendation model based on deep learning: constructing a neural network classifier, learning a complex mapping relation from the fused case features to the optimal processing process, and outputting probability distribution of the recommendation process; and closed-loop feedback learning: the system records artificial decision feedback and actual processing effects and is used for continuous iterative optimization of the model, so that the model has self-learning and evolution capabilities. According to the method, instant and automatic recommendation of the case processing flow can be realized, the manual decision-making time is greatly shortened, and the overall case processing throughput is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and information technology applications, specifically a method and system for recommending event processing workflows based on deep learning. Background Technology

[0002] In the fields of judiciary and administration, the efficiency and accuracy of case (or administrative) processing are crucial. Traditional case processing workflows largely rely on human experience. Case handlers select one of several pre-set processing procedures based on the basic case type (e.g., "civil loan dispute," "traffic accident claim") and a cursory reading of the case description. Existing automation solutions are mostly based on simple rule engines or decision trees, essentially solidifying manually generated rules. While this improves efficiency, it fails to address fundamental problems such as rigid rules, inability to handle complex text semantics, and difficulty in adaptive optimization. Therefore, there is an urgent need for a technological solution that can intelligently, accurately, and automatically recommend the optimal case processing workflow. Summary of the Invention

[0003] The technical objective of this invention is to address the above-mentioned shortcomings by providing a method and system for recommending event processing procedures based on deep learning. This method and system can achieve real-time and automatic recommendation of case processing procedures, significantly shortening the time for manual decision-making and improving the overall case processing throughput.

[0004] The technical solution adopted by this invention to solve its technical problem is:

[0005] A deep learning-based method for recommending event handling processes, the implementation of which includes:

[0006] Multimodal feature extraction and fusion: Simultaneously utilizing structured information and unstructured text descriptions of cases to construct a comprehensive case feature profile;

[0007] A process recommendation model based on deep learning is built: a neural network classifier is constructed to learn the complex mapping relationship from the fused case features to the optimal processing flow, and outputs the probability distribution of the recommendation flow.

[0008] Closed-loop feedback learning: The system records human decision-making feedback and actual processing effects, and uses this information for continuous iterative optimization of the model, enabling it to learn and evolve on its own.

[0009] This method integrates machine learning and natural language processing technologies to automatically recommend optimal processing procedures for cases, addressing the problems of high subjectivity, low efficiency, and inability to deeply utilize textual information and historical data caused by the reliance on human experience in traditional case processing procedures. By quantitatively analyzing the structured features (type, amount, etc.) and unstructured text features (descriptive content) of cases through machine learning models, it deeply mines their correlation with the optimal process, reduces subjective assumptions, ensures consistency in recommended processes for similar cases, and improves decision-making efficiency.

[0010] Furthermore, the multimodal feature extraction and fusion includes:

[0011] The structured feature vector includes case categories and numerical data. Suppose there are m structured features, which, after One-Hot encoding and standardization, form an m-dimensional structured feature vector S = [s1, s2, ..., sm]. m ];

[0012] The text feature vector includes the case description text; the case description D is encoded using a pre-trained BERT model, and the hidden state labeled [CLS] is taken as the semantic summary vector of the text; let the dimension of the BERT output be d, then the text feature vector T = BERT(D) ∈ R d ;

[0013] Feature fusion: The vectors from the two different sources mentioned above are fused to form a unified case representation vector.

[0014] Furthermore, the case representation vector is obtained by concatenation followed by dimensionality reduction using a fully connected layer:

[0015] X = f(W) f ·[S;T]+b f )

[0016] Where: [S;T] means concatenating vectors S and T in terms of dimension to obtain a (m+d) dimensional vector;

[0017] W f It is the weight matrix, b f It is the bias vector;

[0018] f(·) is an activation function (such as ReLU);

[0019] X is the final fused case feature vector.

[0020] Furthermore, the process recommendation model built based on deep learning aims to learn a mapping function F: X→Y, where Y is a probability distribution representing whether a case belongs to any predefined process P. iThe probability of.

[0021] Furthermore, the deep learning-based process recommendation model employs a deep neural network as a classifier:

[0022] Z1 = f(W1·X + b1)

[0023] Z2 = f(W2·Z1 + b2)

[0024] Y = Softmax(W out ·Z k +b out )

[0025] Where: Z i It is the output of the i-th hidden layer;

[0026] W i ,b i These are the weights and bias parameters of the i-th layer, respectively;

[0027] Y = [y1, y2, ..., y n ] represents the probability distribution of the output, where n is the total number of processes, and y i This indicates that the case is recommended as the i-th process P. i The probability of.

[0028] Furthermore, the process of training the recommendation model is as follows:

[0029] During the offline phase, a process with a true label (i.e., the historical best process P) is used. true The model is trained using data from the dataset; the loss function used is the cross-entropy loss function, which aims to minimize the difference between the model's predicted distribution Y and the true distribution (One-Hot encoding).

[0030] L=-Σ(I(label=P i )·log(y i ))

[0031] in:

[0032] L is the loss value;

[0033] I(·) is an indicator function, which is used when the true label of the sample is P. i The value is 1 when the time is right, and 0 otherwise.

[0034] y i The model predicts the process P. i The probability of;

[0035] The model parameters W,b are continuously updated using the backpropagation algorithm and gradient descent optimizer (such as Adam) to minimize the overall loss L.

[0036] Furthermore, the closed-loop feedback learning includes online recommendation and feedback learning.

[0037] When making online recommendations, the system will combine the fusion features of new cases. new Input the pre-trained model to obtain the probability distribution Y;

[0038] The system recommends the K processes with the highest probability (Top-K) and their confidence scores to the user;

[0039] The user's (case handler's) final choice will be recorded by the system. If the user does not select the Top-1 recommendation, the sample will be stored in the feedback database along with the final selected process label. The model will be incrementally trained or fine-tuned regularly using feedback data so that the model can learn the latest experience of human experts and achieve self-evolution.

[0040] This invention also claims a deep learning-based event processing flow recommendation system, comprising:

[0041] The multimodal feature fusion module is used to simultaneously utilize the structured information and unstructured text descriptions of a case to construct a comprehensive case feature profile;

[0042] The model building module is used to build a neural network classifier, learn the complex mapping relationship from the fused case features to the optimal processing flow, and output the probability distribution of the recommendation process.

[0043] The feedback learning module is used to record human decision feedback and actual processing effects, and is used for continuous iterative optimization of the model, enabling it to have self-learning and evolution capabilities.

[0044] This system can implement the methods described above.

[0045] The present invention also claims a deep learning-based event processing flow recommendation device, comprising: at least one memory and at least one processor;

[0046] The at least one memory is used to store a machine-readable program;

[0047] The at least one processor is used to call the machine-readable program to implement the above method.

[0048] The present invention also claims a computer-readable medium storing computer instructions that, when executed by a processor, enable the implementation of the above-described method.

[0049] Compared with existing technologies, the event processing flow recommendation method and system based on deep learning of the present invention have the following advantages:

[0050] 1. Intelligentization and Automation: The system has automated and intelligentized the case handling process recommendation, significantly reducing reliance on human experience.

[0051] 2. Precise and efficient: By deeply mining the multimodal information of cases, especially the deep semantics of the text, the recommendation results are more accurate and consistent, which greatly improves decision-making efficiency.

[0052] 3. Self-evolution capability: A feedback learning mechanism is introduced, enabling the system to continuously optimize, adapt to changes in business rules and the external environment, and have a long lifespan.

[0053] 4. Interpretability: The model can output recommendation probabilities and combine techniques such as attention mechanisms to display key decision-making criteria in the text, enhancing the interpretability and credibility of the recommendation results. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the event processing flow recommendation method based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0055] This invention provides a deep learning-based event handling process recommendation method. By integrating machine learning and natural language processing technologies, it automatically recommends the optimal handling process for cases, solving the problems of high subjectivity, low efficiency, and inability to deeply utilize textual information and historical data caused by the reliance on human experience in traditional case handling process allocation. The implementation of this method includes:

[0056] Multimodal feature extraction and fusion: Simultaneously utilize the structured information of the case (type, amount, etc.) and unstructured text description (using models such as BERT to extract deep semantic features) to construct a comprehensive case feature profile;

[0057] A process recommendation model based on deep learning is built: a neural network classifier is constructed to learn the complex mapping relationship from the fused case features to the optimal processing flow, and outputs the probability distribution of the recommendation flow.

[0058] Closed-loop feedback learning: The system records human decision-making feedback and actual processing effects, and uses this information for continuous iterative optimization of the model, enabling it to learn and evolve on its own.

[0059] This method uses machine learning models to quantitatively analyze the structured features (type, amount, etc.) and unstructured text features (description content) of cases, deeply explores their correlation with the optimal process, reduces subjective assumptions, ensures the consistency of recommended processes for similar cases, and improves decision-making efficiency.

[0060] Combination Figure 1 As shown, the specific implementation of this method is as follows:

[0061] 1. Multimodal feature extraction and fusion: For each case, its feature vector is formed by fusing two parts:

[0062] The structured feature vector (S) comes from the case's category and numerical data. Suppose there are m structured features, which, after One-Hot encoding and standardization, form an m-dimensional structured feature vector S = [s1, s2, ..., sm]. m ].

[0063] The text feature vector (T) comes from the case description text. A pre-trained BERT model is used to encode the case description D, and the hidden state labeled [CLS] is taken as the semantic summary vector of the text. Let the dimension of the BERT output be d, then the text feature vector T = BERT(D) ∈ R. d .

[0064] Feature fusion: The vectors from the two different sources are fused to form a unified case representation vector X. This embodiment uses concatenation followed by dimensionality reduction through a fully connected layer.

[0065] X = f(W) f ·[S;T]+b f )

[0066] Where: [S;T] means concatenating vectors S and T in terms of dimension to obtain a (m+d) dimensional vector;

[0067] W f It is the weight matrix, b f It is the bias vector;

[0068] f(·) is an activation function (such as ReLU);

[0069] X is the final fused case feature vector.

[0070] 2. Construct a process recommendation model:

[0071] The goal of the model is to learn a mapping function F: X → Y, where Y is a probability distribution representing whether a case belongs to a predefined process P. i The probability of.

[0072] This embodiment uses a deep neural network as the classifier:

[0073] Z1 = f(W1·X + b1)

[0074] Z2 = f(W2·Z1 + b2)

[0075] Y = Softmax(W out ·Z k+b out )

[0076] Where: Z i It is the output of the i-th hidden layer;

[0077] W i ,b i These are the weights and bias parameters of the i-th layer, respectively;

[0078] Y = [y1, y2, ..., y n ] represents the probability distribution of the output, where n is the total number of processes, and y i This indicates that the case is recommended as the i-th process P. i The probability of.

[0079] 3. Model training and loss function:

[0080] During the offline phase, a process with a true label (i.e., the historical best process P) is used. true The model is trained using data from [data source name]. The loss function used is the cross-entropy loss function, which aims to minimize the difference between the model's predicted distribution Y and the true distribution (One-Hot encoding).

[0081] L=-Σ(I(label=P i )·log(y i ))

[0082] in:

[0083] L is the loss value;

[0084] I(·) is an indicator function, which is used when the true label of the sample is P. i The value is 1 when the time is right, and 0 otherwise.

[0085] y i The model predicts the process P. i The probability of;

[0086] The model parameters W,b are continuously updated using the backpropagation algorithm and gradient descent optimizer (such as Adam) to minimize the overall loss L.

[0087] 4. Online recommendation and feedback learning:

[0088] When making online recommendations, the system will combine the fusion features of new cases. new Input the pre-trained model to obtain the probability distribution Y;

[0089] The system recommends the K processes with the highest probability (Top-K) and their confidence scores to the user;

[0090] The user's (case handler's) final choice will be recorded by the system. If the user does not select the Top-1 recommendation, the sample will be stored in the feedback database along with the final selected process label. The model will be incrementally trained or fine-tuned regularly using feedback data so that the model can learn the latest experience of human experts and achieve self-evolution.

[0091] This invention also provides a deep learning-based event processing workflow recommendation system, comprising:

[0092] The multimodal feature fusion module is used to simultaneously utilize the structured information and unstructured text descriptions of a case to construct a comprehensive case feature profile;

[0093] The model building module is used to build a neural network classifier, learn the complex mapping relationship from the fused case features to the optimal processing flow, and output the probability distribution of the recommendation process.

[0094] The feedback learning module is used to record human decision feedback and actual processing effects, and is used for continuous iterative optimization of the model, enabling it to have self-learning and evolution capabilities.

[0095] This system can implement the deep learning-based event processing flow recommendation method described in the above embodiments.

[0096] The multimodal feature fusion module, for each case, fuses its feature vector from two parts:

[0097] The structured feature vector (S) comes from the case's category and numerical data. Suppose there are m structured features, which, after One-Hot encoding and standardization, form an m-dimensional structured feature vector S = [s1, s2, ..., sm]. m ].

[0098] The text feature vector (T) comes from the case description text. A pre-trained BERT model is used to encode the case description D, and the hidden state labeled [CLS] is taken as the semantic summary vector of the text. Let the dimension of the BERT output be d, then the text feature vector T = BERT(D) ∈ R. d .

[0099] Feature fusion: The vectors from the two different sources are fused to form a unified case representation vector X. This embodiment uses concatenation followed by dimensionality reduction through a fully connected layer.

[0100] X = f(W) f ·[S;T]+b f )

[0101] Where: [S;T] means concatenating vectors S and T in terms of dimension to obtain a (m+d) dimensional vector;

[0102] W f It is the weight matrix, b f It is the bias vector;

[0103] f(·) is an activation function (such as ReLU);

[0104] X is the final fused case feature vector.

[0105] The model building module aims to learn a mapping function F: X→Y, where Y is a probability distribution representing whether a case belongs to any predefined process P. i The probability of.

[0106] This embodiment uses a deep neural network as the classifier:

[0107] Z1 = f(W1·X + b1)

[0108] Z2 = f(W2·Z1 + b2)

[0109] Y = Softmax(W out ·Z k +b out )

[0110] Where: Z i It is the output of the i-th hidden layer;

[0111] W i ,b i These are the weights and bias parameters of the i-th layer, respectively;

[0112] Y = [y1, y2, ..., y n ] represents the probability distribution of the output, where n is the total number of processes, and y i This indicates that the case is recommended as the i-th process P. i The probability of.

[0113] Model training and loss function:

[0114] During the offline phase, a process with a true label (i.e., the historical best process P) is used. true The model is trained using data from [data source name]. The loss function used is the cross-entropy loss function, which aims to minimize the difference between the model's predicted distribution Y and the true distribution (One-Hot encoding).

[0115] L=-Σ(I(label=P i )·log(y i ))

[0116] in:

[0117] L is the loss value;

[0118] I(·) is an indicator function, which is used when the true label of the sample is P. i The value is 1 when the time is right, and 0 otherwise.

[0119] y i The model predicts the process P. i The probability of;

[0120] The model parameters W,b are continuously updated using the backpropagation algorithm and gradient descent optimizer (such as Adam) to minimize the overall loss L.

[0121] The feedback learning module,

[0122] When making online recommendations, the system will combine the fusion features of new cases. new Input the pre-trained model to obtain the probability distribution Y;

[0123] The system recommends the K processes with the highest probability (Top-K) and their confidence scores to the user;

[0124] The user's (case handler's) final choice will be recorded by the system. If the user does not select the Top-1 recommendation, the sample will be stored in the feedback database along with the final selected process label. The model will be incrementally trained or fine-tuned regularly using feedback data so that the model can learn the latest experience of human experts and achieve self-evolution.

[0125] This invention also provides a deep learning-based event processing flow recommendation device, comprising: at least one memory and at least one processor;

[0126] The at least one memory is used to store a machine-readable program;

[0127] The at least one processor is used to call the machine-readable program to implement the deep learning-based event processing flow recommendation method described in the above embodiments.

[0128] This invention also provides a computer-readable medium storing computer instructions. When executed by a processor, the computer instructions cause the processor to perform the deep learning-based event processing flow recommendation method described in the above embodiments. Specifically, a system or apparatus equipped with a storage medium storing software program code that implements the functions of any of the above embodiments can be provided, and the computer (or CPU or MPU) of the system or apparatus can read and execute the program code stored in the storage medium.

[0129] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0130] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0131] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0132] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0133] The present invention has been shown and described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above embodiments, those skilled in the art will know that more embodiments of the present invention can be obtained by combining the code review methods in the different embodiments. These embodiments are also within the protection scope of the present invention.

Claims

1. A deep learning-based event processing flow recommendation method, characterized in that, The implementation of this method includes: Multimodal feature extraction and fusion: Simultaneously utilizing structured information and unstructured text descriptions of cases to construct a comprehensive case feature profile; A process recommendation model based on deep learning is built: a neural network classifier is constructed to learn the complex mapping relationship from the fused case features to the optimal processing flow, and outputs the probability distribution of the recommendation flow. Closed-loop feedback learning: The system records human decision-making feedback and actual processing effects, and uses this information for continuous iterative optimization of the model, enabling it to learn and evolve on its own.

2. The event processing flow recommendation method based on deep learning according to claim 1, characterized in that, The multimodal feature extraction and fusion includes: The structured feature vector includes case categories and numerical data. Suppose there are m structured features, which, after One-Hot encoding and standardization, form an m-dimensional structured feature vector S = [s1, s2, ..., sm]. m ]; The text feature vector includes the case description text; the case description D is encoded using a pre-trained BERT model, and the hidden state labeled [CLS] is taken as the semantic summary vector of the text; let the dimension of the BERT output be d, then the text feature vector T = BERT(D) ∈ R d ; Feature fusion: The vectors from the two different sources mentioned above are fused to form a unified case representation vector.

3. The event processing flow recommendation method based on deep learning according to claim 2, characterized in that, The case representation vector is obtained by concatenating the data and then reducing its dimensionality using a fully connected layer. X=f(W f ·[S;T]+b f ) Where: [S;T] means concatenating vectors S and T in terms of dimension to obtain a (m+d) dimensional vector; W f It is the weight matrix, b f It is the bias vector; f(·) is the activation function; X is the final fused case feature vector.

4. The event processing flow recommendation method based on deep learning according to claim 3, characterized in that, The deep learning-based process recommendation model aims to learn a mapping function F: X→Y, where Y is a probability distribution representing whether a case belongs to any predefined process P. i The probability of.

5. The event processing flow recommendation method based on deep learning according to claim 4, characterized in that, The deep learning-based recommendation model uses a deep neural network as a classifier. Z1 = f(W1·X + b1) Z2 = f(W2·Z1 + b2) Y =Softmax(W out ·Z k +b out ) Among them: Z i It is the output of the i-th hidden layer; W i ,b i These are the weights and bias parameters of the i-th layer, respectively; Y = [y1, y2, ..., y n ] represents the probability distribution of the output, where n is the total number of processes, and y i This indicates that the case is recommended as the i-th process P. i The probability of.

6. A deep learning-based event processing flow recommendation method according to claim 1 or 4, characterized in that, The process of training the recommended model is as follows: In the offline phase, the model is trained using data with real labels; the cross-entropy loss function is used to minimize the difference between the model's predicted distribution Y and the true distribution. L=-Σ(I(label=P i )·log(y i )) in: L is the loss value; I(·) is an indicator function, which is used when the true label of the sample is P. i The value is 1 when the time is right, and 0 otherwise. y i The model predicts the process P. i The probability of; The model parameters W,b are continuously updated using the backpropagation algorithm and gradient descent optimizer to minimize the overall loss L.

7. The event processing flow recommendation method based on deep learning according to claim 1, characterized in that, The closed-loop feedback learning includes online recommendation and feedback learning. When making online recommendations, the system will combine the fusion features of new cases. new Input the pre-trained model to obtain the probability distribution; The system will recommend the K processes with the highest probability and their confidence levels to the user; The user's final choice will be recorded by the system. If the user does not select the Top-1 recommendation, the sample will be stored in the feedback database along with the final selected process label. Regularly use feedback data to incrementally train or fine-tune the model, enabling it to learn from the latest experiences of human experts and achieve self-evolution.

8. A deep learning-based event processing flow recommendation system, characterized in that, include: The multimodal feature fusion module is used to simultaneously utilize the structured information and unstructured text descriptions of a case to construct a comprehensive case feature profile; The model building module is used to build a neural network classifier, learn the complex mapping relationship from the fused case features to the optimal processing flow, and output the probability distribution of the recommendation process; The feedback learning module is used to record human decision-making feedback and actual processing effects, and is used for continuous iterative optimization of the model, enabling it to have self-learning and evolution capabilities. This system can implement the method described in any one of claims 1 to 7.

9. A deep learning-based event processing flow recommendation device, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that, The computer-readable medium stores computer instructions that, when executed by a processor, enable the implementation of the method described in any one of claims 1 to 7.

Images