Training method and device of call detection model

Abstract

The embodiment of the invention provides a call detection model training method and device, and the method comprises the steps: firstly obtaining a user call sample, inputting the user call sample into an encoder of a to-be-trained model, and carrying out the feature extraction, thereby obtaining the features of the call sample, and inputting the random sample features into a generator of a to-be-trained model to generate a random call sample, then inputting the user call sample, the call sample features, the random call sample and the random sample features into a discriminator of the to-be-trained model to carry out anomaly detection to obtain a detection result, and finally carrying out loss calculation based on the detection result and a result label. And performing parameter adjustment on the to-be-trained model according to the loss obtained by calculation to obtain a call detection model, thereby improving the efficiency and accuracy of call detection.

Description

Technical Field

[0001] This document relates to the field of artificial intelligence, and in particular to a method and device for training a call detection model. Background Art

[0002] With the development of information technology, people have higher and higher requirements for call security and call quality. Any factors that affect call clarity may reduce user experience and even lead to business loss, such as the caller asking the called party to transfer funds to an abnormal account, the caller or the called party mentioning words in the gray list, or the call content containing invalid fields. Therefore, it is crucial to effectively monitor and quickly respond to various abnormal situations that may occur during the call. Traditional call detection methods usually rely on rule matching and expert experience. Although this method is effective in some scenarios, it is powerless in the face of ever-changing fraud methods. In recent years, with the development of machine learning and natural language processing technology, data-driven methods have gradually become mainstream. This type of technology can be used for call detection by learning from a large number of known cases. Therefore, how to further improve the effect of call detection has become a focus of attention for all parties. Summary of the invention

[0003] The purpose of one embodiment of the present specification is to provide a method and device for training a call detection model to solve the problems of low efficiency and low accuracy of call detection.

[0004] To solve the above technical problems, an embodiment of this specification is implemented as follows: In a first aspect, an embodiment of the present specification provides a method for training a call detection model, the method comprising: Get user call samples; Inputting the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and inputting the random sample features into the generator of the model to be trained to generate random call samples; Inputting the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; Loss calculation is performed based on the detection results and result labels, and parameters of the model to be trained are adjusted according to the calculated loss.

[0005] The training method of the call detection model provided in this embodiment first obtains multiple types of user call samples so that the model can perceive multiple types of data during the training process. Secondly, on the one hand, the user call samples are input into the encoder of the model to be trained for feature extraction to obtain call sample features, and the encoder is used to extract rich data features, while data compression is achieved, thereby improving data processing efficiency. On the other hand, random sample features are input into the generator of the model to be trained to generate random call samples, so that real user call samples are simulated by the generator to increase the diversity and richness of the data set, thereby improving the generalization ability of the model. Again, the real user call samples, call sample features, and simulated random call samples and random sample features are simultaneously input into the discriminator of the model to be trained for anomaly detection to obtain detection results, so that the discriminator learns the ability to determine whether the generated samples are real. Finally, loss calculation is performed based on the detection results and result labels, and the parameters of the model to be trained are adjusted according to the calculated losses to obtain a call detection model, so that the model to be trained learns more abundant knowledge, thereby improving the accuracy and effectiveness of the trained call detection model.

[0006] In a second aspect, another embodiment of the present specification provides a call detection method, the method comprising: Get user call data; Inputting the user call data into an encoder of a call detection model for feature extraction to obtain call features; Inputting the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The generated call data is input into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

[0007] The call detection method provided in this embodiment first obtains user call data, and then inputs the user call data into the encoder of the call detection model for feature extraction to obtain call features, so as to learn meaningful representations from the original data. The call features are again input into the generator of the call detection model for data generation to obtain generated call data, so as to create new training samples through the generator. These samples can increase the diversity and richness of the data set, thereby improving the generalization ability of the model. Finally, the generated call data is input into the discriminator of the call detection model for anomaly detection to obtain call detection results, and the discriminator is used to evaluate whether the samples generated by the generator are authentic and reliable, so as to achieve efficient and accurate call detection.

[0008] In a third aspect, an embodiment of the present specification provides a training device for a call detection model, the device comprising: A sample acquisition module, configured to acquire user call samples; A feature extraction module is configured to input the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and input the random sample features into the generator of the model to be trained to generate random call samples; an anomaly detection module, configured to input the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; The parameter adjustment module is configured to perform loss calculation based on the detection result and the result label, and adjust the parameters of the model to be trained according to the calculated loss.

[0009] In a fourth aspect, another embodiment of the present specification provides a call detection device, the device comprising: A data acquisition module is configured to acquire user call data; A feature extraction module is configured to input the user call data into an encoder of a call detection model to extract features and obtain call features; A data generation module is configured to input the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The anomaly detection module is configured to input the generated call data into the discriminator of the call detection model for anomaly detection to obtain a call detection result.

[0010] In a fifth aspect, an embodiment of the present specification provides a training device for a call detection model, comprising: a memory, a processor, and computer executable instructions stored on the memory and executable on the processor, wherein the computer executable instructions, when executed by the processor, implement the steps of the call detection model training method as described in the first aspect above.

[0011] In the sixth aspect, another embodiment of the present specification provides a call detection device, comprising: a memory, a processor, and computer executable instructions stored in the memory and executable on the processor, wherein the computer executable instructions, when executed by the processor, implement the steps of the call detection method as described in the first aspect above.

[0012] In a seventh aspect, an embodiment of the present specification provides a computer-readable storage medium, which is used to store computer-executable instructions. When the computer-executable instructions are executed by a processor, the steps of the training method of the call detection model as described in the first aspect above are implemented.

[0013] In an eighth aspect, another embodiment of the present specification provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the call detection method as described in the first aspect above are implemented.

[0014] In a ninth aspect, an embodiment of the present specification provides a computer program product, wherein the computer program product includes a training program for a call detection model, and the training program for the call detection model is executed by a processor to implement the steps of the training method for the call detection model as described in the first aspect above.

[0015] In the tenth aspect, another embodiment of the present specification provides a computer program product, wherein the computer program product includes a call detection program, and the call detection program is executed by a processor to implement the steps of the call detection method as described in the first aspect above. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in one or more embodiments of the present specification, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 A processing flow chart of a method for training a call detection model provided in one embodiment of this specification; Figure 2 A schematic diagram of the structure of a call detection model provided by one embodiment of this specification; Figure 3 A processing flow chart of a training method for a call detection model applied to a training scenario of a call detection model provided by one embodiment of this specification; Figure 4 A call detection method processing flow chart provided in one embodiment of this specification; Figure 5 A processing flow chart of a call detection method applied to a call detection scenario provided by an embodiment of this specification; Figure 6 A schematic diagram of a training device for a call detection model provided in one embodiment of this specification; Figure 7 A schematic diagram of a call detection device provided in one embodiment of this specification; Figure 8 A schematic diagram of the structure of a training device for a call detection model provided in one embodiment of this specification; Fig. 9 A schematic diagram of the structure of a call detection device provided in one embodiment of this specification. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0019] One or more embodiments of a ride processing method provided in this specification are as follows: Reference Figure 1 , which shows a processing flow chart of a ride processing method provided in this embodiment, specifically including the following steps S102 to S108.

[0020] Step S102, obtaining user call samples.

[0021] In this embodiment, the user call sample may be call content such as audio and / or text. Specifically, the user call sample may be an audio sample collected during the user call, or a text sample obtained by performing speech recognition on the collected audio sample. In addition, the user call sample may also be a recording sample obtained by sampling the recording after the user call is completed, or a recording text obtained by performing speech recognition on the recording sample.

[0022] In actual applications, abnormal call content may be generated during a user call, such as the caller asking the called party to transfer funds to an abnormal account, or the caller or the called party mentioning words in the gray list, or the call content containing invalid fields, etc. In view of this, it is necessary to detect the user's call process to avoid abnormal calls that reduce user experience and cause user losses. Therefore, the call process can be detected with the help of a model to improve call detection efficiency and avoid user losses. In this embodiment, a call detection model that can be used for call detection is obtained by training the model to be trained, and in the training process, by introducing positive label call samples and negative label call samples, the model can perceive abnormal data and normal data, effectively guiding the model to learn accurate anomaly detection.

[0023] Optionally, the user call samples include labeled call samples and unlabeled call samples; the labeled call samples include positively labeled call samples and negatively labeled call samples. Specifically, the positively labeled call samples may be normal call samples, such as call content that has been verified to be normal, the negatively labeled call samples may be abnormal call samples, such as call content that has been verified to be abnormal, and the unlabeled call samples may be call content that has not been verified.

[0024] It should be noted that the call detection service provided by the call detection model involved in this application complies with the requirements of laws and regulations and respects social morality and public order and good customs. The information and data used by the call detection model involved in this application are all authorized by the user or fully authorized by all parties.

[0025] Step S104: input the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and input the random sample features into the generator of the model to be trained to generate random call samples.

[0026] After obtaining the user call samples as mentioned above, in this step, on the one hand, the user call samples are input into the encoder of the model to be trained for feature extraction to obtain call sample features, so as to utilize the encoder to extract rich data features, while realizing data compression and improving data processing efficiency; on the other hand, the random sample features are input into the generator of the model to be trained to generate random call samples, thereby simulating real user call samples through the generator to increase the diversity and richness of the data set.

[0027] The model to be trained in this embodiment can be a model composed of a bidirectional generative network (BiGANs, BidirectionalGenerative Adversarial Networks), and the model to be trained includes the following components: an encoder, a generator, and a discriminator, wherein the encoder is used for feature extraction, that is, mapping the input sample to a low-dimensional latent variable space to obtain a latent variable code to achieve data compression, the generator is used to generate random call samples, that is, decoding and reconstructing the latent variable code to obtain a random sample, and the discriminator is used for anomaly detection, that is, identifying whether the input sample belongs to a normal or abnormal sample, and is used to guide the generator to learn the probability distribution of normal data. In addition, the model to be trained may also include a label flipping module, which is used to flip the sample labels of the positive label call samples and the sample labels of the negative label call samples based on the flipping probability, that is, flipping the sample label of at least one positive label call sample to a negative label, or flipping the sample label of at least one negative label call sample to a positive label, so as to avoid overfitting. Here, the structure of the call detection model obtained after the training of the model to be trained can be referred to. Figure 2 shown.

[0028] In the specific feature extraction process, the positive label call samples, negative label call samples and unlabeled call samples can be respectively input into the encoder of the model to be trained for feature extraction to obtain the first sample feature, the second sample feature and the third sample feature. The positive label call samples, negative label call samples and unlabeled call samples can also be spliced ​​and then input into the encoder of the model to be trained for feature extraction to obtain the spliced ​​sample features. In addition, the two methods can be combined for processing at the same time, which is not limited in this embodiment.

[0029] For example, define the encoder as E, the generator as G, the discriminator as D, and the positive label call sample , negative label call samples And unlabeled call samples After inputting encoder E for feature extraction, the sample features are obtained: , .

[0030] In addition, by introducing random sample features extracted from the random sample feature space, and generating random call samples according to the random sample features, the sample diversity is enriched and the model training effect is enhanced. In an optional implementation provided in this embodiment, the random sample features are obtained in the following manner: Construct random sample feature space; Sampling is performed from the random sample feature space to obtain the random sample feature.

[0031] Specifically, after the random sample feature space is constructed, the random sample feature may be extracted from the random sample feature space in a random manner.

[0032] Furthermore, in an optional implementation manner provided by this embodiment, the random sample feature space is constructed in the following manner: Sampling a preset sample feature space to obtain a preset sample feature; Inputting the preset sample features into the generator to generate a preset call sample, and inputting the preset call sample into the encoder to extract intermediate sample features; Inputting the preset sample features and the intermediate sample features into the discriminator to perform anomaly detection to obtain anomaly detection results; The random sample feature space is constructed based on the anomaly detection result, the preset sample feature and the intermediate sample feature.

[0033] Optionally, the preset sample feature space may be a prior distribution such as a Gaussian distribution or a normal distribution. For example, a Gaussian distribution: .

[0034] Specifically, in the process of anomaly detection performed by the discriminator here, the similarity between the preset sample features and the intermediate sample features can be detected. If the similarity is lower than or equal to the similarity threshold, a random sample feature space is constructed based on the preset sample features. If the similarity is higher than the similarity threshold, a random sample feature space is constructed based on the intermediate sample features.

[0035] In a specific implementation, after extracting random sample features from the random sample feature space, the random sample features are input into the generator of the model to be trained to generate random call samples. For example, the random sample feature z is input into the generator G to generate random call samples. .

[0036] In this embodiment, on the one hand, by inputting negative label call samples into the model to be trained, the discriminator can perceive the existence of abnormal data, and can suppress the high-density area caused by the corresponding abnormality, so that it falls back into the low-density area. At the same time, by inputting positive label call samples into the model to be trained, the discriminator can perceive the existence of normal data, and while suppressing abnormal data, the diversity of normal data is not destroyed, making it more suitable for the task of anomaly detection of multiple types of normal data, and enabling the generator to learn a more accurate distribution of normal data.

[0037] Step S106: input the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result.

[0038] After the user call samples are input into the encoder of the model to be trained for feature extraction to obtain call sample features, and the random sample features are input into the generator of the model to be trained to generate random call samples, in this step, the user call samples, call sample features, random call samples and random sample features are simultaneously input into the discriminator of the model to be trained for anomaly detection to obtain detection results, so that the discriminator learns the ability to determine whether the generated samples are real.

[0039] During the specific execution process, in order to improve the training efficiency of the model, the samples and features can be spliced ​​and then input into the discriminator. In an optional implementation provided in this embodiment, feature splicing is performed in the following manner: Performing feature splicing on the user call sample and the call sample feature to obtain a first splicing feature, and performing feature splicing on the random call sample and the random sample feature to obtain a second splicing feature; The first splicing feature and the second splicing feature are input into the discriminator for abnormality detection to obtain a first detection result and a second detection result.

[0040] Specifically, the positive label call sample and the first sample feature can be spliced ​​to obtain the first sub-splice feature, the negative label call sample and the second sample feature can be spliced ​​to obtain the second sub-splice feature, the unlabeled call sample and the third sample feature can be spliced ​​to obtain the third sub-splice feature, the first sub-splice feature, the second sub-splice feature, and the third sub-splice feature can be used as the first splicing feature, and the random call sample and the random sample feature can be spliced ​​to obtain the second splicing feature, the first sub-splice feature, the second sub-splice feature, the third sub-splice feature, and the second splicing feature can be respectively input into the discriminator for anomaly detection to obtain the first sub-detection result, the second sub-detection result, the third sub-detection result, and the second detection result.

[0041] Using the above example, we can add positive label call samples , negative label call samples Unlabeled call sample After inputting encoder E for feature extraction, the sample features are obtained: , And generate random call samples after inputting random sample feature z into the generator , concatenate them and obtain as well as And input the discriminator to obtain the discrimination result: , .

[0042] Step S108, performing loss calculation based on the detection result and the result label, and adjusting parameters of the model to be trained according to the calculated loss.

[0043] After obtaining the above-mentioned test results, in this step, a loss calculation is performed based on each test result and a result label corresponding to each test result, and parameters of the model to be trained are adjusted according to the calculated loss.

[0044] In the specific implementation process, during the model training process, in order to improve the training efficiency of the significant feature model and reduce the workload of model training, in an optional implementation provided by this embodiment, the loss is calculated based on the detection results and result labels in the following manner: Determine a first result label and a second result label; The first loss is calculated based on the detection result and the first result label, and the second loss is calculated based on the detection result and the second result label.

[0045] Furthermore, in an optional implementation manner provided by this embodiment, adjusting parameters of the model to be trained according to the calculated loss includes: Update based on the second loss and the parameters of the discriminator; Parameters of the generator and parameters of the encoder are updated based on the first loss.

[0046] Optionally, the first result label may be a result label of a discriminator, the first loss may be a loss of a discriminator, the second result label may be a result label of an encoder and a generator, and the second loss may be a loss of an encoder and a generator.

[0047] For example, the loss of the discriminator is calculated using the following formula:

[0048] For another example, the following formula is used to calculate the loss of the encoder and generator:

[0049] in, , They are positive label call samples , negative label call samples , unlabeled call samples , random call samples The discriminant score of , , Corresponding positive label call samples , negative label call samples , unlabeled call samples , random call samples The joint probability distribution of. c. Represents the selected parameters, and the model parameters include the parameters of the discriminator D, the parameters of the encoder E, and the parameters of the generator G.

[0050] Referring to the above model training method, the above training process is repeated to train the model to be trained, and the encoder, generator, and discriminator are adjusted until the loss function converges. After the loss function converges, the training is completed to obtain the call detection model. Specifically, after calculating the loss of the objective function, the model parameters are updated by gradient descent, and the parameters of the discriminator, generator, and encoder are updated alternately by gradient back propagation.

[0051] For example, when the model is trained to convergence and reaches Nash equilibrium, the optimal discriminator is:

[0052] The corresponding distribution of the optimal generator is: .

[0053] In addition, during the model training process, this embodiment is also provided with a label flipping module. In order to prevent the discriminator from overfitting on the labeled call samples, the label flipping module is added to make the discriminator not overconfident and thus remember the entire labeled call samples. An optional implementation provided by this embodiment also includes: based on the label flipping module, the sample labels of the positively labeled call samples and the sample labels of the negatively labeled call samples are flipped.

[0054] Specifically, the label flipping module adjusts the difference in the discrimination scores output by the discriminator for the positive label call sample and the negative label call sample. If it exceeds the set threshold, overfitting is detected and the flipping probability is increased until the difference in the discrimination scores of the two is less than the threshold, then the flipping probability is reduced.

[0055] In summary, the training method of the call detection model provided in this embodiment first obtains multiple types of user call samples so that the model can perceive multiple types of data during the training process. Secondly, on the one hand, the user call samples are input into the encoder of the model to be trained for feature extraction to obtain call sample features, and the encoder is used to extract rich data features, while realizing data compression and improving data processing efficiency. On the other hand, the random sample features are input into the generator of the model to be trained to generate random call samples, so that the real user call samples are simulated by the generator, the diversity and richness of the data set are increased, and thus the generalization ability of the model is improved. Again, the real user call samples, call sample features, and simulated random call samples and random sample features are simultaneously input into the discriminator of the model to be trained for anomaly detection to obtain the detection results, so that the discriminator learns the ability to judge whether the generated samples are real. Finally, the loss is calculated based on the detection results and the result labels, and the parameters of the model to be trained are adjusted according to the calculated loss to obtain the call detection model, so that the model to be trained learns more abundant knowledge, and the accuracy and effectiveness of the trained call detection model are improved. Furthermore, by introducing labeled call samples, that is, positively labeled call samples and negatively labeled call samples to participate in training, the discriminator can perceive the existence of abnormal data and normal data, and can better guide the generator to generate samples, which is more suitable for scenarios with multiple types of samples; Furthermore, by introducing a label flipping module, the corresponding outputs of the discriminator for positive label call samples and negative label call samples are calculated to perceive whether the discriminator is in the overfitting stage, and gradually adjust the probability of label reversal of the auxiliary data set to alleviate the overfitting phenomenon of the discriminator.

[0056] The following combination Figure 3, taking the application of the training method of the call detection model provided in this embodiment in the training scenario of the call detection model as an example, the training method of the call detection model provided in this embodiment is further described. Figure 3 The training method of the call detection model applied to the training scenario of the call detection model specifically includes steps S302 to S314.

[0057] Step S302, obtaining user call samples.

[0058] Optionally, the user call samples include labeled call samples and unlabeled call samples; the labeled call samples include positively labeled call samples and negatively labeled call samples.

[0059] Step S304: input the user call sample into the encoder of the model to be trained to extract features to obtain call sample features.

[0060] Step S306, inputting the random sample features into the generator of the model to be trained to generate random call samples.

[0061] Step S308: concatenate the user call sample and the call sample feature to obtain a first concatenated feature.

[0062] Step S310, concatenating the random call sample and the random sample feature to obtain a second concatenated feature.

[0063] Step S312: input the first splicing feature and the second splicing feature into a discriminator for abnormality detection to obtain a first detection result and a second detection result.

[0064] Step S314, performing loss calculation based on the first detection result, the first result label, the second detection result, and the second result label, adjusting parameters of the model to be trained according to the calculated first loss and second loss, and obtaining a call detection model.

[0065] It should be noted that any one of steps S302 to S314 or any combination of multiple steps can be replaced by the corresponding technical means provided in the above steps S102 to S108 according to the needs of implementation and deployment, and will not be described one by one here.

[0066] It should be supplemented that, when reading the training method of the call detection model provided in the above embodiment, reference can be made to the relevant content of the call detection method provided in the following embodiment. Correspondingly, when reading the call detection method provided in the following embodiment, reference can be made to the relevant content of the training method of the call detection model provided in the above embodiment.

[0067] This specification provides a call detection method embodiment: Reference Figure 4, which shows a processing flow chart of a call detection method provided in this embodiment. The call detection method provided in this embodiment specifically includes the following steps S402 to S408.

[0068] Step S402, obtaining user call data.

[0069] In this embodiment, the user call sample may be call content such as audio and / or text. Specifically, the user call sample may be an audio sample collected during the user call, or a text sample obtained by performing speech recognition on the collected audio sample. In addition, the user call sample may also be a recording sample obtained by sampling the recording after the user call is completed, or a recording text obtained by performing speech recognition on the recording sample.

[0070] In actual applications, abnormal call content may be generated during a user call, such as the caller asking the called party to transfer funds to an abnormal account, or the caller or the called party mentioning words in the gray list, or the call content containing invalid fields, etc. In view of this, it is necessary to detect the user's call process to avoid abnormal calls that reduce user experience and cause user losses. Therefore, the call process can be detected with the help of a model to improve call detection efficiency and avoid user losses. In this embodiment, a call detection model that can be used for call detection is obtained by training the model to be trained, and in the training process, by introducing positive label call samples and negative label call samples, the model can perceive abnormal data and normal data, effectively guiding the model to learn accurate anomaly detection.

[0071] Optionally, the user call samples include labeled call samples and unlabeled call samples; the labeled call samples include positively labeled call samples and negatively labeled call samples. Specifically, the positively labeled call samples may be normal call samples, such as call content that has been verified to be normal, the negatively labeled call samples may be abnormal call samples, such as call content that has been verified to be abnormal, and the unlabeled call samples may be call content that has not been verified.

[0072] It should be noted that the call detection service provided by the call detection model involved in this application complies with the requirements of laws and regulations and respects social morality and public order and good customs. The information and data used by the call detection model involved in this application are all authorized by the user or fully authorized by all parties.

[0073] Step S404: input the user call data into the encoder of the call detection model for feature extraction to obtain call features.

[0074] Step S406: input the call features into the generator of the call detection model to generate data, and obtain generated call data.

[0075] Step S408: input the generated call data into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

[0076] Optionally, in an optional implementation manner provided by this embodiment, the discriminator performs anomaly detection in the following manner: Calculating the similarity between the generated call data and the user call data; If the similarity is greater than a preset error threshold, an abnormal detection result is output; if the similarity is less than the preset error threshold, a normal detection result is output.

[0077] Specifically, the similarity between the generated call data and the user call data can be calculated in the following manner: ;in, express The generated call data is obtained through encoder E and generator G in sequence.

[0078] The following combination Figure 5 , taking the application of the call detection method provided in this embodiment in the call detection scenario as an example, the problem solving method provided in this embodiment is further explained. Figure 5 ,The problem handling method applied to the call detection scenario specifically includes the following steps.

[0079] Step S502, obtaining user call data.

[0080] Step S504: input the user call data into the encoder of the call detection model for feature extraction to obtain call features.

[0081] Step S506, input the call features into the generator of the call detection model to generate data, and obtain generated call data.

[0082] Step S508: Input the user call data and the generated call data into the discriminator of the call detection model to calculate the similarity between the generated call data and the user call data.

[0083] Step S510: if the similarity is greater than a preset error threshold, the abnormal detection result is output.

[0084] Step S512: if the similarity is less than the preset error threshold, a normal detection result is output.

[0085] It should be noted that any one of steps S502 to S512 or any combination of multiple steps can be replaced by the corresponding technical means provided in the above steps S402 to S408 according to the needs of implementation and deployment, and will not be described one by one here.

[0086] Figure 6A schematic diagram of a call detection device provided by an embodiment of the present invention is shown in FIG. Figure 6 As shown, the device comprises: The sample acquisition module 602 is configured to acquire user call samples; The feature extraction module 604 is configured to input the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and input the random sample features into the generator of the model to be trained to generate random call samples; The anomaly detection module 606 is configured to input the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; The parameter adjustment module 608 is configured to perform loss calculation based on the detection result and the result label, and adjust the parameters of the model to be trained according to the calculated loss.

[0087] The call detection device provided in this embodiment first runs the anomaly detection module 602 to obtain user call samples, then runs the feature extraction module 604 to input the user call samples into the encoder of the model to be trained for feature extraction to obtain call sample features, and inputs the random sample features into the generator of the model to be trained to generate random call samples, and then runs the anomaly detection module 606 again to input the user call samples, the call sample features, the random call samples and the random sample features into the discriminator of the model to be trained for anomaly detection to obtain detection results, and finally runs the parameter adjustment module 608 to calculate the loss based on the detection results and result labels, and adjusts the parameters of the model to be trained according to the calculated losses to obtain a call detection model, thereby improving the efficiency and accuracy of call detection.

[0088] The training device for the call detection model provided in one embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0089] Figure 7 A schematic diagram of another training device for a call detection model provided by an embodiment of the present invention is shown in FIG. Figure 7 As shown, the device comprises: The data acquisition module 702 is configured to acquire user call data; The feature extraction module 704 is configured to input the user call data into the encoder of the call detection model to extract features and obtain call features; The data generating module 706 is configured to input the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The anomaly detection module 708 is configured to input the generated call data into the discriminator of the call detection model to perform anomaly detection and obtain a call detection result.

[0090] The training device of the call detection model provided in this embodiment first runs the data acquisition module 702 to obtain user call data, then runs the feature extraction module 704 to input the user call data into the encoder of the call detection model for feature extraction to obtain call features, then runs the data generation module 706 to input the call features into the generator of the call detection model for data generation to obtain generated call data, and finally runs the anomaly detection module 708 to input the generated call data into the discriminator of the call detection model for anomaly detection to obtain call detection results, thereby improving the efficiency and accuracy of call detection.

[0091] The call detection device provided in an embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0092] Furthermore, an embodiment of the present specification also provides a training device for a call detection model. Figure 8 A structural diagram of a training device for a call detection model provided in an embodiment of this specification, such as Figure 8 As shown, the device includes: a memory 801, a processor 802, a bus 803 and a communication interface 804. The memory 801, the processor 802 and the communication interface 804 communicate through the bus 803, and the communication interface 804 may include an input and output interface, which includes but is not limited to a keyboard, a mouse, a display, a microphone, a loudspeaker, etc.

[0093] Figure 8 In the embodiment, the memory 801 stores computer executable instructions that can be run on the processor 802, and when the computer executable instructions are executed by the processor 802, the following process is implemented: Get user call samples; Inputting the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and inputting the random sample features into the generator of the model to be trained to generate random call samples; Inputting the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; Loss calculation is performed based on the detection results and result labels, and parameters of the model to be trained are adjusted according to the calculated loss.

[0094] The training device for the call detection model provided in this embodiment, through the cooperation of the memory 801, the processor 802, the bus 803 and the communication interface 804, first obtains the user call sample, then inputs the user call sample into the encoder of the model to be trained for feature extraction to obtain the call sample feature, and inputs the random sample feature into the generator of the model to be trained to generate a random call sample, and then inputs the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained for anomaly detection to obtain the detection result, and finally calculates the loss based on the detection result and the result label, adjusts the parameters of the model to be trained according to the calculated loss, and obtains the call detection model, thereby improving the efficiency and accuracy of call detection.

[0095] The call detection device provided in an embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0096] Furthermore, an embodiment of the present specification also provides a call detection device, Fig. 9 A structural diagram of a call detection device provided in an embodiment of this specification is shown as follows: Fig. 9 As shown, the device includes: a memory 901, a processor 902, a bus 903 and a communication interface 904. The memory 901, the processor 902 and the communication interface 904 communicate through the bus 903, and the communication interface 904 may include an input and output interface, which includes but is not limited to a keyboard, a mouse, a display, a microphone, a loudspeaker, etc.

[0097] Fig. 9 In the embodiment, the memory 901 stores computer executable instructions that can be run on the processor 902. When the computer executable instructions are executed by the processor 902, the following process is implemented: Get user call data; Inputting the user call data into an encoder of a call detection model for feature extraction to obtain call features; Inputting the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The generated call data is input into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

[0098] The call detection device provided in this embodiment, through the cooperation of the memory 901, the processor 902, the bus 903 and the communication interface 904, first obtains the user call data, then inputs the user call data into the encoder of the call detection model for feature extraction to obtain call features, then inputs the call features into the generator of the call detection model for data generation to obtain generated call data, and finally inputs the generated call data into the discriminator of the call detection model for anomaly detection to obtain call detection results, thereby improving the efficiency and accuracy of call detection.

[0099] The call detection device provided in an embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0100] Furthermore, an embodiment of the present specification also provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store computer-executable instructions, and when the computer-executable instructions are executed by a processor, the following process is implemented: Get user call samples; Inputting the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and inputting the random sample features into the generator of the model to be trained to generate random call samples; Inputting the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; Loss calculation is performed based on the detection results and result labels, and parameters of the model to be trained are adjusted according to the calculated loss.

[0101] The computer-readable storage medium provided in this embodiment first obtains user call samples, then inputs the user call samples into the encoder of the model to be trained for feature extraction to obtain call sample features, and inputs the random sample features into the generator of the model to be trained to generate random call samples, and then inputs the user call samples, call sample features, random call samples and random sample features into the discriminator of the model to be trained for anomaly detection to obtain detection results, and finally calculates losses based on the detection results and result labels, adjusts the parameters of the model to be trained according to the calculated losses, and obtains a call detection model, thereby improving the efficiency and accuracy of call detection.

[0102] The computer-readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0103] The computer-readable storage medium provided in an embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0104] Furthermore, another embodiment of the present specification provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store computer-executable instructions, and when the computer-executable instructions are executed by a processor, the following process is implemented: Get user call data; Inputting the user call data into an encoder of a call detection model for feature extraction to obtain call features; Inputting the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The generated call data is input into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

[0105] The computer-readable storage medium provided in this embodiment first obtains user call data, then inputs the user call data into the encoder of the call detection model for feature extraction to obtain call features, then inputs the call features into the generator of the call detection model for data generation to obtain generated call data, and finally inputs the generated call data into the discriminator of the call detection model for anomaly detection to obtain call detection results, thereby improving the efficiency and accuracy of call detection.

[0106] The computer-readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0107] The computer-readable storage medium provided in an embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0108] Furthermore, an embodiment of the present specification also provides a computer program product, which implements the following process when executed by a processor: Get user call samples; Inputting the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and inputting the random sample features into the generator of the model to be trained to generate random call samples; Inputting the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; Loss calculation is performed based on the detection results and result labels, and parameters of the model to be trained are adjusted according to the calculated loss.

[0109] The computer program product provided in this embodiment first obtains user call samples, then inputs the user call samples into the encoder of the model to be trained for feature extraction to obtain call sample features, and inputs the random sample features into the generator of the model to be trained to generate random call samples, and then inputs the user call samples, call sample features, random call samples and random sample features into the discriminator of the model to be trained for anomaly detection to obtain detection results, and finally calculates losses based on the detection results and result labels, adjusts the parameters of the model to be trained according to the calculated losses, and obtains a call detection model, thereby improving the efficiency and accuracy of call detection.

[0110] The computer program product provided in one embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0111] Furthermore, another embodiment of the present specification also provides a computer program product, which implements the following process when executed by a processor: Get user call data; Inputting the user call data into an encoder of a call detection model for feature extraction to obtain call features; Inputting the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The generated call data is input into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

[0112] The computer program product provided in this embodiment first obtains user call data, then inputs the user call data into the encoder of the call detection model for feature extraction to obtain call features, then inputs the call features into the generator of the call detection model for data generation to obtain generated call data, and finally inputs the generated call data into the discriminator of the call detection model for anomaly detection to obtain call detection results, thereby improving the efficiency and accuracy of call detection.

[0113] The computer program product provided in one embodiment of this specification can implement each process in the aforementioned method embodiment and achieve the same functions and effects, which will not be repeated here.

[0114] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0116] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0117] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0118] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0119] Memory may include non-permanent storage in a computer-readable storage medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable storage medium.

[0120] Computer-readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable storage media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0121] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0122] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0123] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for training a call detection model, characterized in that: The method comprises: Get user call samples; Inputting the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and inputting the random sample features into the generator of the model to be trained to generate random call samples; Inputting the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; Loss calculation is performed based on the detection results and result labels, and parameters of the model to be trained are adjusted according to the calculated loss.

2. The method for training a call detection model according to claim 1, characterized in that: The user call samples include labeled call samples and unlabeled call samples; the labeled call samples include positively labeled call samples and negatively labeled call samples.

3. The method for training a call detection model according to claim 1, characterized in that: The step of inputting the user call sample, the call sample feature, the random call sample, and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result includes: splicing the user call sample and the call sample feature to obtain a first splicing feature, and splicing the random call sample and the random sample feature to obtain a second splicing feature; The first splicing feature and the second splicing feature are input into the discriminator for abnormality detection to obtain a first detection result and a second detection result.

4. The method for training a call detection model according to claim 1, characterized in that: The loss calculation based on the detection result and the result label includes: Determine a first result label and a second result label; The first loss is calculated based on the detection result and the first result label, and the second loss is calculated based on the detection result and the second result label.

5. The method for training a call detection model according to claim 4, characterized in that: The step of adjusting parameters of the model to be trained according to the calculated loss includes: Updating the parameters of the discriminator based on the second loss; Parameters of the generator and parameters of the encoder are updated based on the first loss.

6. The method for training a call detection model according to claim 2, characterized in that: Also includes: Based on the label flipping module, the sample labels of the positive label call samples and the sample labels of the negative label call samples are flipped.

7. The method for training a call detection model according to claim 1, characterized in that: The random sample characteristics are obtained in the following way: Construct random sample feature space; Sampling is performed from the random sample feature space to obtain the random sample feature.

8. The method for training a call detection model according to claim 7, characterized in that: The constructing of the random sample feature space comprises: Sampling a preset sample feature space to obtain a preset sample feature; Inputting the preset sample features into the generator to generate a preset call sample, and inputting the preset call sample into the encoder to extract intermediate sample features; Inputting the preset sample features and the intermediate sample features into the discriminator to perform anomaly detection to obtain anomaly detection results; The random sample feature space is constructed based on the anomaly detection result, the preset sample feature and the intermediate sample feature.

9. A call detection method, characterized in that: The method comprises: Get user call data; Inputting the user call data into an encoder of a call detection model for feature extraction to obtain call features; Inputting the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The generated call data is input into the discriminator of the call detection model for abnormality detection to obtain a call detection result.

10. The call detection method according to claim 9, wherein the discriminator performs abnormality detection in the following manner: Calculating the similarity between the generated call data and the user call data; If the similarity is greater than a preset error threshold, outputting a detection abnormality result; If the similarity is less than a preset error threshold, a normal detection result is output.

11. A training device for a call detection model, characterized in that: The device comprises: A sample acquisition module, configured to acquire user call samples; A feature extraction module is configured to input the user call sample into the encoder of the model to be trained to extract features to obtain call sample features, and input the random sample features into the generator of the model to be trained to generate random call samples; an anomaly detection module, configured to input the user call sample, the call sample feature, the random call sample and the random sample feature into the discriminator of the model to be trained to perform anomaly detection and obtain a detection result; The parameter adjustment module is configured to perform loss calculation based on the detection result and the result label, and adjust the parameters of the model to be trained according to the calculated loss.

12. A call detection device, characterized in that: The device comprises: A data acquisition module is configured to acquire user call data; A feature extraction module is configured to input the user call data into an encoder of a call detection model to extract features and obtain call features; A data generation module is configured to input the call features into the generator of the call detection model to generate data, thereby obtaining generated call data; The anomaly detection module is configured to input the generated call data into the discriminator of the call detection model for anomaly detection to obtain a call detection result.

13. A training device for a call detection model, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer executable instructions, and when the computer executable instructions are executed on the processor, the steps of the method described in any one of claims 1 to 8 can be implemented.

14. A call detection device, characterized in that: It comprises a memory and a processor, wherein the memory stores computer executable instructions, and when the computer executable instructions are executed on the processor, the steps of the method described in any one of claims 9 to 10 can be implemented.

15. A computer-readable storage medium having computer-executable instructions stored therein, characterized in that: When the computer executable instructions are executed by a processor, the steps of the method described in any one of claims 1 to 8 can be implemented.

16. A computer-readable storage medium having computer-executable instructions stored therein, characterized in that: When the computer executable instructions are executed by a processor, the steps of the method described in any one of claims 9 to 10 can be implemented.

17. A computer program product, characterized in that The computer program product includes a training program for a call detection model, and when the training program for the call detection model is executed by a processor, the steps of the method described in any one of claims 1 to 8 can be implemented.

18. A computer program product, characterized in that The computer program product comprises a call detection program, and when the call detection program is executed by a processor, the steps of the method as described in any one of claims 9 to 10 can be implemented.

Images