A loan application process anomaly correction method based on conditional adversarial generation network

Abstract

The application discloses a loan application process abnormality correction method based on a conditional adversarial generative network, and the method comprises the following steps: calling historical loan application process data of normal execution, firstly injecting abnormal behaviors in each normal execution track, and saving the original normal execution track; then performing word segmentation on the original normal execution track and the constructed abnormal execution track to carry out word embedding, and constructing an abnormal correction data set; then correcting the abnormal execution track by using the conditional adversarial generative network, constructing a generative model, pre-training a generator, alternately training a discriminator and the generator, and maximizing the correction ability of the generator; finally, using the generator after the adversarial training to correct the abnormal execution track. The application has the advantages that: the application can be used as downstream work of abnormality detection, and can make up for the deficiency that the abnormality detection algorithm can only detect abnormalities but cannot correct abnormalities.

Description

Technical Field

[0001] This invention belongs to the field of business process management technology and relates to a method for correcting anomalies in business processes. Background Technology

[0002] Companies rely on business processes to ensure the smooth operation of their business and guarantee its normal functioning. However, abnormal situations may occur during the execution of business processes, causing interruptions or deviations from the intended course of action. This undoubtedly poses a certain risk to the goals achieved through the business processes. Typically, business process management systems store a large number of process execution logs, containing information such as which events were executed and in what order. By analyzing these logs, companies can identify abnormal execution sequences within a business process and rectify them, thereby optimizing the process.

[0003] Traditional anomaly detection methods model normal behavior and then analyze the differences between the modeled data and normal data to detect anomalies. Because the model is trained primarily on data representing normal behavior, without modeling abnormal behavior, traditional anomaly detection methods struggle to establish a correspondence between abnormal and normal behaviors. This limits their ability to detect anomalies to effective correction. Anomaly correction typically requires a large amount of abnormal process data, leading to a data insufficiency in traditional methods. Therefore, there is an urgent need for a business process anomaly correction method that injects abnormal behavior into normal execution trajectories to avoid this data shortage problem. Furthermore, this method can learn the correspondence between abnormal and normal trajectories, reconstructing the normal execution trajectory corresponding to the abnormal one within the business process. Summary of the Invention

[0004] To more quickly and accurately correct execution trajectories in the loan application process, this invention proposes an anomaly correction method based on a conditional generative adversarial model. This method corrects abnormal execution trajectories in the bank loan application process, overcoming the limitation of anomaly detection which can only detect abnormal behavior but cannot effectively correct it. By training a conditional generative adversarial network to learn the correspondence between abnormal and normal execution trajectories in the bank loan application process, it achieves automatic correction of abnormal execution trajectories, providing a reference for subsequent optimization of the bank loan application process.

[0005] A method for correcting anomalies in a loan application process based on conditional adversarial generative networks includes the following steps:

[0006] (1) Retrieve historical process data that conforms to the bank's loan application process execution standards. First, inject abnormal behavior into each trajectory that conforms to the bank's loan application process execution standards, and save the original execution trajectory.

[0007] (2) Determine whether the construction of the anomaly correction dataset has been completed. If yes, proceed to step (4); otherwise, execute step (3).

[0008] (3) The saved abnormal and normal execution trajectories are segmented and word-embedded to construct an abnormal correction dataset.

[0009] (4) A conditional adversarial generative network is used to correct abnormal execution trajectories. The generator is pre-trained in the model, and the discriminator and generator are trained alternately to maximize the generator’s correction capability.

[0010] (5) Use the generator trained against adversarial behavior to correct existing abnormal execution sequences: If the generator performs poorly, the hyperparameters need to be readjusted or the dataset needs to be rebuilt. Jump back to step (3).

[0011] Furthermore, step (1) consists of the following steps:

[0012] (1.1) Constructing functions for injecting abnormal behaviors. Recent research summarizes five common abnormal behaviors in business processes: event skipping, event insertion, duplicate event, event premature execution, and event delay. Based on this, five corresponding abnormal behavior functions are constructed: (a) an event skipping function, which skips some events in the process, preventing their execution; (b) an event insertion function, which inserts irrelevant events into the process; (c) a duplicate event function, which executes some events repeatedly in the process; (d) an event premature execution function, which advances the execution order of some events in the process; and (e) an event delay function, which postpones the execution order of some events in the process. (These functions for injecting abnormal behaviors are used to simulate common process anomalies in actual bank loan application processes.)

[0013] (1.2) Construct abnormal loan process execution trajectories and retain the corresponding execution trajectories that conform to the process execution specifications. First, copy the process that conforms to the bank loan application process execution specifications n times. Then, using different random number seeds, randomly call the above five abnormal behavior functions. The functions randomly select the location of the abnormal injection to obtain abnormal loan process execution trajectories. Finally, save the abnormal loan process execution trajectories (abnormal execution trajectories) and the corresponding execution trajectories that conform to the process execution specifications (normal execution trajectories).

[0014] Furthermore, step (3) consists of the following steps:

[0015] (3.1) The execution trajectory is segmented into words. All events appearing in the execution trajectory of the loan application process are traversed, and each event is treated as a word unit. At the same time, four special lexical symbols are added as additional words units. The four special symbols include “__PAD”, “__GO”, “__EOS” and “__NUK”, which are respectively the padding symbol, the start symbol of the sequence, the end symbol of the sequence, and the unknown symbol.

[0016] (3.2) Word embedding of all execution trajectories in the loan application process: First, construct the corresponding vocabulary based on the lexical units in step 3.1). The vocabulary contains all event types in the process data and four special lexical symbols. In step 1.2, inject abnormal process event trajectories, i.e., behaviors that do not conform to the process specifications in the loan process. Add the word vector _GO to the beginning, and add _EOS to the end of the corresponding execution trajectory that conforms to the bank's process execution specifications in step 1.2. Set the effective length M, truncate event trajectories with a length exceeding M, and fill event trajectories with a length less than M with __PAD to the length M. Then, convert the event execution trajectory into word indexes according to the vocabulary to complete the word embedding. Finally, divide the dataset with completed word embedding into training set, validation set and test set according to the proportion, as the final dataset for anomaly correction.

[0017] Furthermore, step (4) consists of the following steps:

[0018] (4.1) A Seq2Seq model is used to construct the generative model and pre-train the generator. The Seq2Seq model consists of an encoder and a decoder. In the encoder, the sequence of anomalous event trajectories is converted into a fixed-length vector, and then the decoder converts this vector into a normal event trajectory sequence for output. The input of the encoder is the sequence of anomalous trajectories, and each event in the trajectory is encoded to obtain a word vector v. i Output a hidden state h at each time step i , Here, the function f represents the transformation of the RNN hidden layer. The input word vectors v are processed in chronological order. i Compared to the hidden state h of the previous time step i-1 Then, the hidden states at each time step are transformed into vectors using the encoder's custom function q. : In the decoder, the input at each time step is the encoder's output c and the decoder's output from the previous time step. And the vector of the word predicted in the previous time step. If the first word is being predicted, the input word vector is "_GO", marking the start of decoding. The function g represents the hidden layer transformation of the decoder. The decoding process continues until "_EOS" is decoded, marking the end of the decoding process. At each time step, the decoder predicts the probability distribution of the output tokens, uses a softmax function to obtain the distribution, and measures the generation effect by calculating the cross-entropy loss function. The decoder's goal is to generate a sequence Y that is as close as possible to the target sequence. This stage requires pre-training the generator until it converges.

[0019] (4.2) The discriminator is trained to score the generator's results, guiding the generator to generate sequences. Given an input sequence V, the generator produces a sequence Y, and the discriminator learns to maximize the score. And minimize the score The generator's goal is to generate sequence Y, maximizing the score as much as possible. Specifically, as shown in Formula 1:

[0020] (Formula 1)

[0021] Where G and D represent the generator and discriminator, respectively; E represents the mathematical expectation; and The probability distribution of the input sequence V in the training data and The joint probability distribution of ; The probability distribution of the generator generating sequence Y given input sequence V; the discriminator scores. It is the reward value of all states and actions within the entire generated sequence length M. The average value is calculated as shown in Formula 2:

[0022] (Formula 2)

[0023] Among them, the state action reward value It is a state , is the return value of the candidate word vectors generated at time t, and M is the sequence length.

[0024] (4.3) Train the generator to maximize the score. The generator reads in the input sequence V and then generates the event sequence. Predict the next event The generator's update gradient is shown in Equation 3:

[0025] (Formula 3)

[0026] in, These are weighting coefficients related to time t. The sequence of events generated from time 1 to time t-1. The event predicted at time t; This indicates that the generator operates on the input sequence V and... Under the conditions, generate The probability distribution.

[0027] (4.4) Train the generator and discriminator alternately to maximize the generator’s correction capability. As the generator continues to improve, the discriminator needs to be retrained periodically to keep pace with the generator. If the loss of the generator and discriminator remains stable, proceed to step (5); otherwise, continue training.

[0028] The technical concept of this invention is as follows: In anomaly detection of business processes, anomaly detection models are typically trained using most normal execution sequences. This leads to difficulties in determining the correspondence between abnormal and normal behaviors, resulting in the ability to detect abnormal behaviors but not effectively correct them. Training for anomaly correction usually requires a large amount of abnormal process data, thus leading to a shortage of abnormal behavior data. Based on five common abnormal behavior types identified in recent research, abnormal behaviors are injected into normal loan application process data, creating a large number of loan execution trajectories with anomalies. This yields data combining abnormal execution trajectories with their corresponding normal loan execution trajectories, solving the problem of insufficient abnormal data. A conditional generative adversarial network is then used to train the model against the data, enabling it to learn the correction process from abnormal execution trajectories to normal execution trajectories, thereby achieving the correction of abnormal loan processes.

[0029] 1) Training for anomaly correction requires a large amount of anomaly data. Therefore, injecting artificially simulated abnormal behavior into normal loan execution trajectory data can create a large number of loan execution trajectories with anomalies, solving the problem of insufficient anomaly data. 2) The conditional generative adversarial network (GAN) consists of a discriminator and a generator. Both the generator and discriminator are conditioned on loan execution trajectories with anomalies. The generator's output needs to learn the correspondence between loan execution trajectories with anomalies and their corresponding normal loan execution trajectories in the training set. 3) The discriminator in the GAN aims to distinguish the correction result output by the generator from its original normal loan execution trajectory. The two networks compete against each other and continuously adjust their parameters, making it impossible for the discriminator to determine whether the abnormal loan process correction result output by the generator is correct, ultimately achieving the correction of the abnormal loan execution process.

[0030] The advantages of this invention are: it overcomes the limitation of anomaly detection, which can only detect abnormal behavior but cannot effectively correct it. By simulating abnormal behavior, it solves the problem of insufficient abnormal data during the training process for anomaly correction, thereby improving the accuracy of anomaly correction. Using generative adversarial networks (GANs) enables the model to learn the generation relationship between abnormal execution trajectories and corresponding normal execution trajectories, achieving the correction of abnormal loan processes and providing a valuable reference for process improvement. Attached Figure Description

[0031] Figure 1 This is the overall flowchart of the present invention;

[0032] Figure 2 A flowchart for bank loan operations;

[0033] Figure 3 A diagram illustrating the injection of anomalies into the normal loan application process;

[0034] Figure 4 Examples of five different abnormal loan processes and their corresponding correct loan application processes;

[0035] Figure 5 A schematic diagram of the generator training process and the generator error correction process;

[0036] Figure 6 This is a schematic diagram of the training process of the discriminator. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings.

[0038] Reference Appendix Figure 1

[0039] A method for correcting anomalies in a bank loan process based on conditional adversarial generative networks, the method comprising:

[0040] (1) Retrieve historical process data that conforms to the bank's recognized loan application process execution standards. First, inject abnormal behavior into each trajectory that conforms to the bank's loan application process execution standards, and save the original execution trajectory. Specifically, this includes the following steps:

[0041] (1.1) Constructing functions for injecting abnormal behaviors. Recent research summarizes five common abnormal behaviors in business processes: event skipping, event insertion, duplicate event, event premature execution, and event delay. Based on this, five corresponding abnormal behavior functions are constructed: (a) an event skipping function, which skips some events in the process, preventing their execution; (b) an event insertion function, which inserts irrelevant events into the process; (c) a duplicate event function, which executes some events repeatedly in the process; (d) an event premature execution function, which advances the execution order of some events in the process; and (e) an event delay function, which postpones the execution order of some events in the process. (These functions for injecting abnormal behaviors are used to simulate common process anomalies in actual bank loan application processes.)

[0042] (1.2) Construct abnormal loan process execution trajectories while retaining the corresponding execution trajectories that conform to the process execution specifications. First, copy the process that conforms to the bank loan application process execution specifications n times. Then, using different random number seeds, randomly call the five abnormal behavior functions mentioned above. The functions randomly select the location for abnormal injection to obtain abnormal loan process execution trajectories. Finally, save the abnormal loan process execution trajectories and the original normal loan process execution trajectories. The normal loan process execution trajectory is the trajectory that conforms to the loan process execution specifications.

[0043] (2) Determine whether the construction of the anomaly correction dataset has been completed. If yes, proceed to step (4); otherwise, execute step (3).

[0044] (3) The saved abnormal and normal execution trajectories are segmented and word-embedded to construct an abnormality correction dataset. The specific process is as follows:

[0045] (3.1) The execution trajectory is segmented into words. All events appearing in the execution trajectory of the loan application process are traversed, and each event is treated as a word unit. At the same time, four special lexical symbols are added as additional words units. The four special symbols include “__PAD”, “__GO”, “__EOS” and “__NUK”, which are respectively the padding symbol, the start symbol of the sequence, the end symbol of the sequence, and the unknown symbol.

[0046] (3.2) Word embedding of all execution trajectories in the loan application process: First, a corresponding vocabulary is constructed based on the lexical units in step 3.1. The vocabulary contains all event types in the process data and four special lexical symbols. Abnormal process event trajectories are injected in step 1.2, which are common behaviors in the loan process that do not conform to the process specifications. The word vector _GO is added to the beginning, and _EOS is added to the end of the corresponding execution trajectory that conforms to the bank's process execution specifications in step 1.2. An effective length M is set, and event trajectories with a length exceeding M are truncated. Event trajectories with a length less than M are padded to the length M with __PAD. Then, the event execution trajectories are converted into word indexes according to the vocabulary to complete the word embedding. Finally, the dataset with completed word embedding is divided into training set, validation set and test set according to the proportion, which are used as the final dataset for anomaly correction.

[0047] (4) A conditional adversarial generative network is used to correct abnormal execution trajectories. The generator is pre-trained in the model, and the discriminator and generator are trained alternately to maximize the generator's correction capability. The specific steps are as follows:

[0048] (4.1) A Seq2Seq model is used to construct the generative model and pre-train the generator. The Seq2Seq model consists of an encoder and a decoder. In the encoder, the sequence of anomalous event trajectories is converted into a fixed-length vector, and then the decoder converts this vector into a normal event trajectory sequence for output. The input of the encoder is the sequence of anomalous trajectories, and each event in the trajectory is encoded to obtain a word vector v. i Output a hidden state h at each time step i , Here, the function f represents the transformation of the RNN hidden layer. The input word vectors v are processed in chronological order. i Compared to the hidden state h of the previous time step i-1 Then, the hidden states at each time step are transformed into vectors using the encoder's custom function q. : In the decoder, the input at each time step is the encoder's output c and the decoder's output from the previous time step. And the vector of the word predicted in the previous time step. If the first word is being predicted, the input word vector is "_GO", marking the start of decoding. The function g represents the hidden layer transformation of the decoder. The decoding process continues until "_EOS" is decoded, marking the end of the decoding process. At each time step, the decoder predicts the probability distribution of the output tokens, uses a softmax function to obtain the distribution, and measures the generation effect by calculating the cross-entropy loss function. The decoder's goal is to generate a sequence Y that is as close as possible to the target sequence. This stage requires pre-training the generator until it converges.

[0049] (4.2) The discriminator is trained to score the generator's results, guiding the generator to generate sequences. Given an input sequence V, the generator produces a sequence Y, and the discriminator learns to maximize the score. And minimize the score The generator's goal is to generate sequence Y, maximizing the score as much as possible. Specifically, as shown in Formula 1:

[0050] (Formula 1);

[0051] Where G and D represent the generator and discriminator, respectively; E represents the mathematical expectation; and The probability distribution of the input sequence V in the training data and The joint probability distribution of ; The probability distribution of the generator generating sequence Y given input sequence V; the discriminator scores. It is the reward value of all states and actions within the entire generated sequence length M. The average value is calculated as shown in Formula 2:

[0052] (Formula 2);

[0053] Among them, the state action reward value It is a state , is the return value of the candidate word vectors generated at time t, and M is the sequence length.

[0054] (4.3) Train the generator to maximize the score. The generator reads in the input sequence V and then generates the event sequence. Predict the next event The generator's update gradient is shown in Equation 3:

[0055] (Formula 3);

[0056] in, These are weighting coefficients related to time t. The sequence of events generated from time 1 to time t-1. The event predicted at time t; This indicates that the generator operates on the input sequence V and... Under the conditions, generate The probability distribution.

[0057] (4.4) Train the generator and discriminator alternately to maximize the generator’s correction capability. As the generator continues to improve, the discriminator needs to be retrained periodically to keep pace with the generator. If the loss of the generator and discriminator remains stable, proceed to step (5); otherwise, continue training.

[0058] (5) Use the generator trained against adversarial forces to correct existing abnormal execution sequences. If the generator performs poorly, it is necessary to readjust hyperparameters such as the learning rate or rebuild the dataset, and then return to step (3).

[0059] Appendix Figure 2 This flowchart illustrates the standard execution process of various events in a bank loan operation.

[0060] Appendix Figure 3 This diagram illustrates the injection of anomalies into a normal loan application process. First, the normal bank loan application process is copied n times. Then, by randomly selecting five anomaly type functions—event insertion, event skipping, event advancement, event lag, and event duplication—and randomly choosing the injection point, loan application process data with anomalies is obtained.

[0061] Appendix Figure 3 This means that in the appendix Figure 2 This document describes five example diagrams illustrating loan application processes with different exception types, each with its corresponding correct exception, obtained by injecting anomalies into a normal loan application process. The diagrams include three types of events in the loan application process: the first event type, represented by a rectangle, describes the status of the loan application; the second event type, represented by a parallelogram, describes the status of the loan application quote; and the third event type, represented by a trapezoid, describes the execution status of the loan application project.

[0062] Table 1 lists all event types and their associated events in the loan application process.

[0063] Table 1

[0064] Loan application status Status of loan application quote Project execution status Submit application Price selection Fill in the application information Submit partial application Quotation creation Evaluation application Application rejected Quotation Submission Quotation Response Pre-acceptance Quotation feedback Please add more information when replying to the application. Acceptance Quote Cancellation / Cancel application Reject the offer / Application completed Accept the offer / Approval application / / Registration application / / Activation application / / ;

[0065] There is an additional event category, indicated by darker colors in the diagram, which describes anomalies. These anomalies are further categorized into five types: The first is event insertion anomaly in the loan application process, where an irrelevant event is executed, such as in (a) where an irrelevant event, "Submit Application," is inserted after "Quote Response" and "Assessment Application." The second is event advancement anomaly in the loan application process, where events are executed prematurely, such as in (b) where "Approve Application," which should have been executed after "Accept Quote," is executed after "Quote Response." The third is event skipping anomaly in the loan application process, where some events are skipped, such as in (c) where the "Register Application" event, which should have been executed after "Approve Application," is skipped. The fourth is event lag anomaly in the loan application process, where events are executed late, such as in (d) where the "Acceptance Completed" event, which should have been executed after "Acceptance," is delayed until after "Quote Feedback." The fifth type is the event duplication anomaly in the loan application process, that is, some events in the process are abnormally executed repeatedly, such as the "quote submission" event in (e) which should not be executed repeatedly but is executed repeatedly.

[0066] Figure 4This diagram illustrates the generator training and anomaly correction processes. During training, the generator uses a loan application process with anomalies as a condition. It calculates cross-entropy loss based on the corrected sequence generated and its corresponding normal loan application process. Minimizing this loss allows the generator to learn the correspondence between the loan application processes with anomalies and their corresponding normal loan application processes in the training set as closely as possible. The anomaly correction process diagram shows the generator, after adversarial training, being input with a sequence of loan application processes with anomalies. The output sequence of corrected loan application processes is consistent with its corresponding normal loan application process sequence.

[0067] Figure 5 This diagram illustrates the training process of the discriminator. The discriminator uses a loan application process with anomalies as a condition, the corresponding normal loan application process as positive input for training, and the anomaly-corrected sequence generated by the generator for the abnormal loan application process as negative input for training. The purpose of training the discriminator is to distinguish the corrected result of the abnormal loan application process output by the generator from its corresponding normal loan application process.

[0068] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for correcting anomalies in a loan application process based on conditional adversarial generative networks, characterized in that, Includes the following steps: 1) Retrieve historical process data that conforms to the bank's loan application process execution standards. First, inject abnormal behavior into each trajectory that conforms to the bank's loan application process execution standards to obtain abnormal loan process execution trajectories, and save the original normal execution trajectory, that is, the corresponding execution trajectory that conforms to the process execution standards. Step 1) specifically includes the following steps: 1.1) Constructing functions for injecting abnormal behaviors: Summarize five common abnormal behaviors in business processes, including event skipping exception, event insertion exception, event duplication exception, event early execution exception, and event delay exception; construct five corresponding abnormal behavior functions: a) Event skipping exception function, which skips some events in the process so that they cannot be executed; b) Event insertion exception function, which inserts irrelevant events into the process; c) Event duplication exception function, which executes some events in the process repeatedly; d) Event early execution exception function, which advances the execution order of some events in the process; e) Event delay exception function, which postpones the execution order of some events in the process. 1.2) Construct abnormal loan process execution trajectories and retain the corresponding execution trajectories that conform to the process execution specifications; First, copy the process that conforms to the bank loan application process execution specifications n times, then use different random number seeds to randomly call the above five abnormal behavior functions. The function randomly selects the location of the abnormal injection to obtain the abnormal loan process execution trajectory. Finally, save the abnormal loan process execution trajectory and the corresponding execution trajectory that conforms to the process execution specifications. 2) Determine whether the anomaly correction dataset has been constructed. If yes, proceed to step 4; otherwise, execute step 3. 3) The saved abnormal and normal loan process execution trajectories are segmented and word-embedded to construct an abnormality correction dataset; 4) A conditional adversarial generative network is used to correct abnormal execution trajectories. The generator is pre-trained in the model, and the discriminator and generator are trained alternately to maximize the generator's correction capability. 5) Use the adversarially trained generator to correct existing anomalous execution sequences: If the generator's performance is unsatisfactory, the hyperparameters need to be readjusted or the dataset needs to be rebuilt. Go back to step 3.

2. The method for correcting anomalies in a loan application process based on a conditional adversarial generative network as described in claim 1, characterized in that, Step 3) The specific process is as follows: 3.1) Tokenize the execution trajectory, traverse all events that appear in the execution trajectory of the loan application process, treat each event as a word unit, and add four special lexical symbols as additional word units; the four special symbols include __PAD, __GO, __EOS and __NUK, which are respectively the padding symbol, the start of the sequence symbol, the end of the sequence symbol and the unknown symbol; 3.2) Word embedding for all execution trajectories in the loan application process: First, construct a corresponding vocabulary based on the lexical units from step 3.1). The vocabulary includes all event types in the process data and four special lexical symbols. Add the word vector _GO to the beginning of the execution trajectory of the loan process that was injected with anomalies in step 1.2, and add _EOS to the end of the corresponding execution trajectory that conforms to the bank's process execution specifications in step 1.

2. Set an effective length M, truncate execution trajectories longer than M, and fill the length of execution trajectories shorter than M with __PAD until the length is M. Then, convert the event execution trajectories into word indexes by referring to the vocabulary, thus completing the word embedding. Finally, the dataset with completed word embeddings is divided into training, validation, and test sets according to the proportions, which serve as the final dataset for anomaly correction.

3. The method for correcting anomalies in a loan application process based on a conditional adversarial generative network as described in claim 1, characterized in that, Step 4) The specific steps are as follows: 4.1) A Seq2Seq model is used to construct the generative model and pre-train the generator: The Seq2Seq model consists of an encoder and a decoder. In the encoder, the abnormal execution trajectory sequence is converted into a fixed-length vector, and then the decoder converts this vector into a normal execution trajectory sequence for output. The input of the encoder is the abnormal execution trajectory sequence, and each event in the trajectory is encoded to obtain a word vector v. i Output a hidden state h at each time step i , Where the function f represents the transformation of the RNN hidden layer; the input word vectors v are processed in chronological order. i Compared to the hidden state h of the previous time step i-1 Then, the hidden states at each time step are transformed into vectors using the encoder's custom function q. : In the decoder, the input at each time step is the encoder's output c and the decoder's output from the previous time step. And the vector of the word predicted in the previous time step. If the first word is being predicted, the input word vector is the word vector of _GO, marking the start of decoding. The function g represents the hidden layer transformation of the decoder. The decoding process continues until the _EOS is decoded, marking the end of the decoding process. At each time step, the decoder predicts the probability distribution of the output tokens, uses the softmax function to obtain the distribution, and measures the generation effect by calculating the cross-entropy loss function. The goal of the decoder is to generate a sequence Y that is as close as possible to the target sequence. This stage requires pre-training the generator until it converges; 4.2) Training the discriminator to score the generator's results and guide the generator in generating sequences: Given an input sequence V, generate a sequence Y, and the discriminator learns to maximize the score. And minimize the score The generator's goal is to generate sequence Y, maximizing the score as much as possible. Specifically, as shown in Formula 1: （1）； Where G and D represent the generator and discriminator, respectively; E represents the mathematical expectation; and The probability distribution of the input sequence V in the training data and The joint probability distribution of ; The probability distribution of the generator generating sequence Y given input sequence V; the discriminator scores. It is the reward value of all states and actions within the entire generated sequence length M. The average value is calculated as shown in Formula 2: （2）； Among them, the state action reward value It is a state , The return value of the generated word vectors for the candidate words at time t, where M is the sequence length; 4.3) Training the generator to maximize the score: The generator reads in the input sequence V, and then generates a sequence of events based on the input sequence V. Predict the next event The generator's update gradient is shown in Equation 3: （3）； in, These are weighting coefficients related to time t. The sequence of events generated from time 1 to time t-1. The event predicted at time t; This indicates that the generator operates on the input sequence V and... Under the conditions, generate The probability distribution; 4.4) Alternately train the generator and discriminator to maximize the generator's corrective ability: As the generator continues to update and improve, the discriminator needs to be retrained periodically to keep pace with the generator. If the loss of the generator and discriminator remains stable, proceed to step 5); otherwise, continue training.

Images