Feature extraction model training, user intent prediction method, apparatus, and electronic device

Abstract

The application provides a feature extraction model training and user intention prediction method and device, and electronic equipment, and relates to the field of artificial intelligence. The method comprises the following steps: obtaining a first sample data set comprising at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information; using the first sample data set and a recovery module, training a first preset model to obtain a trained first feature extraction model; the first feature extraction model is used for language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence; the target language feature is used for determining a target user intention; in the i-th round of training, the recovery module restores the sample language feature output by the first preset model to obtain a text restoration feature sequence corresponding to the sample language feature; the i+1-th round of training is performed according to the text restoration feature sequence obtained in the i-th round of training; and the application improves the accuracy of user intention prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a feature extraction model training method and a user intention prediction method and device and electronic equipment. BACKGROUND

[0002] Predicting user intention can improve the quality of service to users and thus improve user experience. Currently, existing intelligent user intention prediction methods mainly include user intention prediction based on a K-means clustering algorithm and user intention prediction based on a support vector machine method. However, the existing user intention prediction methods all have the problem of poor accuracy. SUMMARY

[0003] The present application provides a feature extraction model training method and a user intention prediction method and device and electronic equipment to solve the problem of poor accuracy of the existing user intention prediction methods.

[0004] In a first aspect, the present application provides a feature extraction model training method, which comprises:

[0005] obtaining a first sample data set; the first sample data set comprises at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information;

[0006] using the first sample data set and a restoration module, performing N rounds of training on a first preset model to obtain a trained first feature extraction model; the trained first feature extraction model is used to perform language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence; the target text feature sequence is obtained based on user language description information; and the target language feature is used to determine a target user intention of the user;

[0007] in the i-th round of training, the sample language feature output by the first preset model based on the input sample text feature sequence is restored by the restoration module to obtain a text restoration feature sequence corresponding to the sample language feature;

[0008] in the (i+1)-th round of training, the text restoration feature sequence obtained in the i-th round of training and the first sample data set are used to perform the (i+1)-th round of training on the first preset model; N is an integer greater than or equal to 1, and i is an integer greater than or equal to 1 and less than N.

[0009] In a second aspect, the present application provides a user intention prediction method, which comprises:

[0010] obtaining user language description information;

[0011] obtaining a target text feature sequence according to the user language description information;

[0012] performing language feature extraction on the target text feature sequence by using the trained first feature extraction model to obtain a target language feature corresponding to the target text feature sequence; the first feature extraction model is obtained by training based on the feature extraction model training method in any one of the first aspect;

[0013] determining a target user intent of the user according to the target language feature.

[0014] In a third aspect, the present application provides a feature extraction model training device, which comprises:

[0015] The obtaining module is configured to obtain a first sample data set; the first sample data set comprises at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information;

[0016] The training module is configured to perform N rounds of training on a first preset model by using the first sample data set and the restoration module to obtain a trained first feature extraction model; the trained first feature extraction model is configured to perform language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence; the target text feature sequence is obtained based on user language description information; and the target language feature is configured to determine a target user intent of the user;

[0017] In the i-th round of training, the restoration module is configured to restore sample language features output by the first preset model based on an input sample text feature sequence to obtain a text restoration feature sequence corresponding to the sample language features;

[0018] In the i+1-th round of training, the first preset model is trained by using the text restoration feature sequence obtained in the i-th round of training and the first sample data set; N is an integer greater than or equal to 1, and i is an integer greater than or equal to 1 and less than N.

[0019] In a fourth aspect, the present application provides a user intent prediction device, which comprises:

[0020] The obtaining module is configured to obtain user language description information;

[0021] The processing module is configured to obtain a target text feature sequence according to the user language description information;

[0022] The extraction module is configured to perform language feature extraction on the target text feature sequence by using a trained first feature extraction model to obtain a target language feature corresponding to the target text feature sequence.

[0023] The determination module is configured to determine the target user intent of the user according to the target language feature.

[0024] In a fifth aspect, the present application provides an electronic device, which comprises a memory and a processor.

[0025] The memory stores a computer program.

[0026] The processor is configured to execute the method of any one of the first aspect and the second aspect by using the computer program.

[0027] In a sixth aspect, the present application provides a computer readable storage medium, which stores computer execution instructions, and when the computer execution instructions are executed by a processor, the method of any one of the first aspect and the second aspect is implemented.

[0028] In a seventh aspect, the present application provides a computer program product, which comprises a computer program, and when the computer program is executed by a processor, the method of any one of the first aspect and the second aspect is implemented.

[0029] The feature extraction model training, user intent prediction method, apparatus, and electronic device provided in this application train a first preset model using a sample text feature sequence obtained based on sample language description information and a restoration module, resulting in a trained first feature extraction model. During the i-th round of training of the first preset model, the restoration module restores the sample language features output by the first preset model, obtaining a text restoration feature sequence corresponding to the sample language features. Then, during the (i+1)-th round of training, the first preset model is trained again using the text restoration feature sequence obtained in the i-th round and the aforementioned first sample dataset. The process of obtaining the sample language features corresponding to the sample text feature sequence through the first preset model can also be called encoding. Therefore, this method achieves encoding-restoration training of the first preset model, ensuring that while extracting rich language features, the sample text feature sequence obtained based on the sample language description information remains undistorted, thus improving the accuracy of training the first preset model based on the sample text feature sequence to obtain the first feature extraction model. By improving the accuracy of the first feature extraction model, the accuracy of language feature extraction of the target text feature sequence based on the first feature extraction model can be improved, thereby improving the accuracy of the user's target intent determined based on the target language features. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0031] Figure 1 A flowchart illustrating a feature extraction model training method provided in this application;

[0032] Figure 2 A flowchart illustrating another feature extraction model training method provided in this application;

[0033] Figure 3 A flowchart illustrating a user intent prediction model training method provided in this application;

[0034] Figure 4 A flowchart illustrating a user intent prediction method provided in this application;

[0035] Figure 5 A flowchart illustrating another user intent prediction method provided in this application;

[0036] Figure 6 A schematic diagram of the structure of a feature extraction model training device provided in this application;

[0037] Figure 7A structural schematic diagram of a user intention prediction device provided in the present application is shown in the following figure.

[0038] Figure 8 A structural schematic diagram of an electronic device provided in the present application is shown in the following figure.

[0039] The specific embodiments of the present application have been shown in the above figures, and will be described in more detail hereinafter. These figures and the written description are not intended to limit the scope of the present application concept in any way, but to illustrate the present application concept to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0040] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to designate the same elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not representative of all embodiments consistent with the present application. Rather, they are merely examples of devices and methods consistent with some aspects of the present application, as detailed in the appended claims.

[0041] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0042] It should be noted that the feature extraction model training, user intention prediction method, device and electronic equipment provided by the present application can be used in the field of financial technology, and can also be used in any field other than financial technology. The application field of the feature extraction model training, user intention prediction method, device and electronic equipment provided by the present application is not limited.

[0043] Predicting the user intention can improve the service quality for the user, and thus improve the user experience. At present, the existing user intention prediction method mainly includes manual prediction. However, the efficiency of manual prediction is often low, and when the number of users is large, the method of manually predicting the user intention cannot meet the demand. Therefore, some existing technologies propose automatic user intention prediction (also known as intelligent prediction).

[0044] At present, existing intelligent prediction methods mainly include user intention prediction based on a K-means clustering algorithm (also referred to as a K-means clustering algorithm) and user intention prediction based on a support vector machine method. When the K-means clustering algorithm is used to predict user intention, only user language description and some simple statistical methods are relied on for prediction. However, user language description is sometimes ambiguous, information transmission may have noise, and the K-means clustering algorithm is extremely susceptible to outliers, so the user intention prediction result determined based on the K-means clustering algorithm is unstable and has poor accuracy. When the support vector machine method is used to predict user intention, the support vector machine method has great difficulty in training a large-scale data set, and the manually selected parameters and kernel functions have a great influence on the user prediction result determined based on the support vector machine method. Therefore, the accuracy of user intention prediction by the support vector machine method is also poor.

[0045] That is, the existing user intention prediction method has the problem of poor accuracy.

[0046] In view of the above problems of the existing user intention prediction method, the present application proposes a method of extracting language features in user language description by using a first feature extraction model with high accuracy, and determining user intention based on the language features. By using the above-mentioned first feature extraction model with high accuracy, language features with high extraction accuracy can be extracted from user language description, and the influence of manually selected parameters and kernel functions on the prediction result is avoided, thereby improving the accuracy of user intention prediction based on user language description.

[0047] The feature extraction model training technical solution of the present application and how the technical solution of the present application solves the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings. Optionally, the execution subject of the feature extraction model training method can be any electronic device such as a terminal or a server, and the present application does not limit this.

[0048] Figure 1 A flowchart of a feature extraction model training method provided by the present application is shown in FIG. 1. As shown in FIG. 1, the method can include the following steps: Figure 1

[0049] S101, obtain a first sample data set.

[0050] ​The first sample data set can include at least one sample text feature sequence. The sample text feature sequence can be obtained based on sample language description information. The sample language description information can be in a text format. The sample text feature sequence can be obtained by encoding the sample language description information in the text format by using any existing natural language processing method, such as a word embedding algorithm. For example, the sample language description information can be a text corresponding to a voice recording of a user's communication with a customer service personnel.

[0051] Optionally, the electronic device can receive the first sample data set input by the user through an application programming interface (API) or a graphical user interface (GUI).

[0052] Alternatively, the electronic device can first obtain at least one piece of sample language description information. Then, for any piece of sample language description information, the electronic device can perform a word embedding operation on the sample language description information, and then add position encoding to the language feature sequence obtained by the word embedding operation to enhance the position information of each element in the sequence, to obtain a sample text feature sequence corresponding to the sample language description information.

[0053] S102, using the first sample data set and the recovery module, performing N rounds of training on the first preset model to obtain a trained first feature extraction model.

[0054] The N can be an integer greater than or equal to 1. Optionally, the first preset model can be any existing neural network model, which is not limited in the present application. For example, the first preset model can be an encoder in a transformer model.

[0055] In the i-th round of the training process (i is an integer greater than or equal to 1 and less than N), the electronic device can restore the sample language feature output by the first preset model based on the input sample text feature sequence through the restoration module to obtain a text restoration feature sequence corresponding to the sample language feature. That is, in the i-th round of the training process, the electronic device can input the sample text feature sequence into the first preset model. The first preset model can extract features from the sample text feature sequence according to the sample text feature sequence, and output a sample language feature. Then, the electronic device can use the restoration module to restore the sample language feature to obtain a text restoration feature sequence corresponding to the sample language feature. Optionally, the form of the sample language feature can be a feature vector or a feature matrix, which is not limited in the present application.

[0056] Optionally, the restoration module can include at least one fully connected neural network. That is, in this implementation, the electronic device can restore the sample language feature through at least one fully connected neural network (or fully connected network) to obtain a text restoration feature sequence corresponding to the sample language feature. Alternatively, the restoration module can also be any other existing method capable of restoring the sample language feature to obtain a text restoration feature sequence corresponding to the sample language feature, which will not be described here.

[0057] In the i+1-th round of the training process, the electronic device can perform i+1-th round training on the first preset model according to the text restoration feature sequence obtained in the i-th round of training and the first sample data set.

[0058] The trained first feature extraction model can be used to extract language features from a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence. The target text feature sequence is obtained based on user language description information. The target language feature can be used to determine the target user intent of the user. Optionally, how to determine the target user intent of the user based on the target language feature can refer to the user intent prediction method described later.

[0059] In this embodiment, the first preset model is trained based on the sample text feature sequence obtained based on the sample language description information and the restoration module to obtain the trained first feature extraction model. In the i-th training process of the first preset model, the restoration module can restore the sample language feature output by the first preset model to obtain a text restoration feature sequence corresponding to the sample language feature. Then, in the (i+1)-th training process, the first preset model can be trained based on the text restoration feature sequence obtained in the i-th training process and the first sample data set. The process of obtaining the sample language feature corresponding to the sample text feature sequence by the first preset model can also be referred to as encoding. Therefore, by the above method, the first preset model is trained in an encoding-restoration manner, which ensures that the sample text feature sequence obtained based on the sample language description information is not distorted while rich language features are extracted, thereby improving the accuracy of training the first preset model based on the sample text feature sequence to obtain the first feature extraction model. By improving the accuracy of the first feature extraction model, the accuracy of language feature extraction based on the first feature extraction model for the target text feature sequence can be improved, and thus the accuracy of determining the target user intent of the user based on the target language feature can be improved.

[0060] The following describes how the electronic device trains the first preset model based on the text restoration feature sequence obtained in the i-th training process and the first sample data set in the (i+1)-th training process:

[0061] As a possible implementation, the electronic device can input the sample text feature sequence used in the i-th training, the cross-entropy of the sample language feature output by the first preset model in the i-th training process, and the text restoration feature sequence obtained in the i-th training to a preset loss function. Then, the electronic device can train the first preset model in the (i+1)-th training process based on the loss function.

[0062] Optionally, the electronic device can obtain the cross-entropy of the sample language feature by a preset cross-entropy calculation method after outputting the sample language feature by the first preset model. It should be understood that the type of the loss function is not limited in the present application. In addition, the specific implementation of training the first preset model in the (i+1)-th training process based on the loss function can refer to any existing neural network training method, which will not be described herein.

[0063] By the above method, the sample text feature sequence used in the i th round of training, the cross entropy of the sample language feature output by the first preset model in the i th round of training, and the text restoration feature sequence obtained in the i th round of training are input into the preset loss function, realizing the construction of the loss function based on the output of the first preset model (the cross entropy of the sample language feature) and the sample text feature sequence and the text restoration feature sequence. By the above method, the sample language feature and the sample text feature sequence are as different as possible, that is, the gap between the sample language features output by the first preset model is as large as possible, thereby improving the richness of the language features. By generating as many sample language features as possible, the selectability of subsequent network training based on the sample language features is higher, and in the reverse propagation adjustment stage of the training process, the possibility of covering better sample language features is higher, thereby further improving the accuracy and efficiency of training the first preset model.

[0064] Further, as a possible implementation manner, the electronic device can further train the second feature extraction model based on the historical operation record of the user, so that the second feature extraction model can be used to extract the historical behavior feature of the user based on the historical operation record of the user. Then, the user's intention can be predicted based on the historical behavior feature and the aforementioned language feature, thereby further improving the accuracy of predicting the user's intention.

[0065] For example, the electronic device can first obtain a second sample data set, and then use the second sample data set to train the second preset model for M (M is an integer greater than or equal to 1) rounds to obtain the trained second feature extraction model.

[0066] The second sample data set can include at least one sample historical operation feature sequence. The sample historical operation feature sequence can be obtained based on the historical operation record. The format of the historical operation record of the user can be, for example, a text format. The sample historical operation feature sequence can be obtained by encoding the historical operation record in the text format by using any existing natural language processing method (for example, a word embedding algorithm). For example, the historical operation record can include a record of at least one operation of the user on a target application (Application, APP) within a historical preset time period. For example, the operation can include at least one operation of clicking a transfer control, querying an account balance, modifying an account password, and clicking a setting control.

[0067] For example, it is assumed that each operation of the user on the target APP corresponds to a different identifier, as shown in Table 1 below:

[0068] Table 1

[0069] Operation Identification Operation 1 Identification 1 Operation 2 Identification 2 Operation 3 Identification 3

[0070] For example, assuming that the user sequentially performs operation 2, operation 1, and operation 3 on the target APP within a preset time period, the historical operation record can include identifier 2, identifier 1, and identifier 3.

[0071] Optionally, the electronic device can receive the second sample data set input by the user through, for example, an API, or a GUI, and the like. Alternatively, the electronic device can first obtain at least one historical operation record. Then, for any historical operation record, the electronic device can add position encoding to the historical operation feature sequence obtained through word embedding of the historical operation record, so as to enhance the position information of each element in the sequence, and obtain a sample historical operation feature sequence corresponding to the historical operation record.

[0072] The trained second feature extraction model can be used to extract historical behavior features from a target historical operation feature sequence corresponding to the user, to obtain a target historical behavior feature corresponding to the user. The target historical operation feature sequence is obtained based on the historical operation record of the user. The target historical behavior feature is used to determine the target user intent of the user. That is, in this implementation manner, the target historical behavior feature and the target language feature can be used to determine the target user intent of the user in a subsequent stage.

[0073] In this implementation manner, as one possible implementation manner, the electronic device can further train a third preset model to obtain a trained user intent prediction model, so that the user intent can be predicted based on the user intent prediction model in a subsequent stage.

[0074] In some embodiments, the electronic device can input at least one sample text feature sequence into the trained first feature extraction model to obtain at least one sample language feature. The electronic device can input at least one sample historical operation feature sequence into the trained second feature extraction model to obtain at least one sample historical behavior feature. It should be understood that the order in which the electronic device obtains the sample language feature and the sample historical behavior feature is not limited in the present application. Optionally, the electronic device can also obtain the sample language feature and the sample historical behavior feature at the same time.

[0075] Optionally, the sample text feature sequence can be a sample text feature sequence in the first sample data set, or a sample text feature sequence other than the first sample data set, which is not limited in the present application. Optionally, the sample historical operation feature sequence can be a sample historical operation feature sequence in the second sample data set, or a sample historical operation feature sequence other than the second sample data set, which is also not limited in the present application.

[0076] Then, for any sample language feature of any user, the electronic device can splice the sample language feature of the user with a sample historical behavior feature of the user to obtain a spliced feature.

[0077] For example, the electronic device may, for example, splice the sample language feature of the user with the sample historical behavior feature of the user by a channel stacking splicing method to obtain a spliced feature. For example, assuming that the dimension of the sample language feature is 2x3 and the dimension of the sample historical behavior feature is also 2x3, the dimension of the spliced feature obtained by splicing the sample language feature and the sample historical behavior feature by the above channel stacking splicing method is 2x2x3. Alternatively, if the dimension of the sample language feature and the dimension of the sample historical behavior feature are different, the sample language feature or the sample historical behavior feature can be first zero-padded to make the dimension of the sample language feature and the dimension of the sample historical behavior feature the same, and then the sample language feature and the sample historical behavior feature after the dimension is made the same are spliced by the channel stacking splicing method to obtain the dimension of the spliced feature.

[0078] Alternatively, the electronic device may, for example, directly splice the feature matrix corresponding to the sample language feature with the feature matrix corresponding to the sample historical behavior feature. For example, still assuming that the dimension of the sample language feature is 2x3 and the dimension of the sample historical behavior feature is also 2x3, the dimension of the spliced feature obtained by directly splicing the matrices is 4x3.

[0079] After obtaining at least one spliced feature, the electronic device may, according to the at least one spliced feature, train the third preset model to obtain a trained user intent prediction model.

[0080] Optionally, the second preset model and the third preset model may each be any existing neural network model, and the present application does not limit this. For example, the second preset model may also be an encoder in a transformer model. The third preset model may, for example, include a decoder in a transformer model.

[0081] The user intent prediction model may be used to determine a target user intent of a user according to a spliced target feature corresponding to the user. The spliced target feature is obtained by splicing a target language feature of the user with a target historical behavior feature.

[0082] In some embodiments, for any one spliced feature, the electronic device can take the spliced feature as an attribute (i.e., key, k) and an attribute value (i.e., Value, V) in a key-value pair. That is, in the key-value pair, k and v can both be the spliced feature.

[0083] Then, the electronic device can train a third preset model according to the key-value pair and a query vector to obtain a trained user intent prediction model. The query vector can include identifiers of a plurality of preset user intents. The plurality of preset user intents can include the target user intent.

[0084] For example, the electronic device can first obtain a plurality of preset user intents in a text format. Then, the electronic device can perform input embedding (i.e., the aforementioned word embedding) on the plurality of preset user intents in the text format, and then add position encoding to strengthen the position information of each preset user intent in the query vector. It should be understood that the present application does not limit the specific implementation of the electronic device training the third preset model according to the key-value pair and the query vector to obtain the trained user intent prediction model. Optionally, the electronic device can refer to any existing training method of a user intent prediction model, which will not be described herein.

[0085] Taking the above restoration module (also referred to as a restorer) including a three-layer fully connected network as an example, Figure 2 Another feature extraction model training method provided by the present application is shown in the flowchart. Figure 3 A user intent prediction model training method provided by the present application is shown in the flowchart.

[0086] As shown in Figure 2 and As shown in Figure 3 The electronic device can collect user information, including user language description (sample language description information) Xw and recent operation record (i.e., historical operation record) Xb. The electronic device can filter the user's recent operation record Xb and language description Xw according to the user identifier (e.g., user ID) on the server side, and accordingly develop accurate prediction of user intent.

[0087] Then, the electronic device can use a language feature encoder (trained first feature extraction model) to extract a user language feature vector. As shown in Figure 2 The structure of the first preset model can be as shown in the encoder. Wherein Qx can represent that the encoder can include Q modules of "multi-head attention + residual sum and normalization". As shown in Figure 2As shown, before the sample language description information is input into the encoder, input embedding (i.e., word embedding) and position encoding can be performed to enhance the position information of each element in the sequence. Then, the electronic device uses a multi-head self-attention module to model the self-correlation between the elements in the sample text feature sequence, performs non-uniform feature weighting, and uses residual connection and layer normalization processing on the weighted sample text feature sequence; finally, after multi-layer feature abstraction and extraction, it is encoded into sample language features.

[0088] Then, in order to ensure that the user language information is not distorted while extracting rich sample language features, an encoding-decoding strategy is used for training. As shown, Figure 2 The input is the sample text feature sequence and the output is the text restoration feature sequence, and the input and output feature dimensions remain the same after encoding-decoding operation. After training, the parameters of the encoder are assigned initial values according to the training results. The encoding-decoding strategy is to keep the prediction results consistent with the input data, thereby reducing the loss of user language features in the feature extraction process, i.e., to ensure that the extracted language features are not distorted and the original corpus information of the user is restored. At the same time, by adding the cross-entropy of the language feature encoder output result in the loss function in the encoding-decoding stage, the features output by the encoder are as different as possible from the input, thereby extracting rich language features of the user.

[0089] The training process of the above-mentioned second feature extraction model can refer to the above-mentioned process, which will not be described again. In addition, it should be understood that the user behavior features can also be their generalized features, which can not only include the recent operation records of the user on the mobile phone, but also include some other information of the user, such as user age, location, account flow, recent consumption and transfer records, etc. By integrating these information, the self-attention mechanism is used to mine the correlation between the information, thereby accurately describing the behavior features of the user, and providing a basis for accurately predicting the user's intention in the next step.

[0090] As shown, Figure 3 When training the user intention prediction model, the collected recent operation records of the user can be input into Figure 3 The behavior feature encoder (the second feature extraction model trained) as shown. Before that, first, the position encoding can be added according to the different recent operation records of the user, to maintain the position uniqueness of different information. Then, based on the multi-head self-attention module, the self-correlation coefficients between the information are extracted, so as to weight different information, pay attention to useful information, and ignore irrelevant information. Then, the behavior features are mapped to a high-dimensional abstract space through a feedforward network, and finally the normalized mapped historical behavior features are used for the final user intention prediction task.

[0091] The sample language features and a sample historical behavior feature of the user are spliced (as shown in Figure 3The concatenated features are input into a multi-head attention module as a key vector K and a value vector V, and multi-head attention is performed on the key vector K and the value vector V and the query vector. The weighted feature vectors are processed using residual connection and normalization operation, and then a feedforward network with residual connection is used for nonlinear mapping and normalization operation. Finally, the abstract feature information extracted is mapped to the final user intent prediction space through a fully connected network, and a Softmax operation is performed to process the confidence of various prediction results, thereby training the third preset model.

[0092] Before training the third preset model, the pre-trained user intent prediction initial encoding can be subjected to an intent prediction initial embedding operation (i.e., word embedding), and then position encoding is added to strengthen the position information of the sequence. A residual connection mask multi-head attention module is used to perform non-uniform weighting on the feature sequence, and the normalized operation is used to input the query vector into the multi-head attention module of the decoder.

[0093] The concatenated features are input into a multi-head attention module as a key vector K and a value vector V, and multi-head attention is performed on the key vector K and the value vector V and the query vector. The weighted feature vectors are processed using residual connection and normalization operation, and then a feedforward network with residual connection is used for nonlinear mapping and normalization operation. Finally, the abstract feature information extracted is mapped to the final user intent prediction space through a fully connected network, and a Softmax operation is performed to process the confidence of various prediction results, thereby training the third preset model.

[0094] In the stage of using the trained user intent prediction model, 20 user intent prediction results can be obtained through the Softmax output, and the three user intent prediction results with the highest possibility are output, thereby obtaining the accurate prediction result of the user intent.

[0095] The user intent prediction technology and how the technology solves the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in detail in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings. Optionally, the execution subject of the user intent prediction method can be any electronic device such as a terminal or a server, and the present application does not limit this. In addition, it should be understood that the execution subject of the user intent prediction method and the execution subject of the feature extraction model training method can be the same electronic device or different electronic devices, and the present application does not limit this.

[0096] Figure 4 A flowchart of a user intent prediction method provided by the present application is shown in FIG. 1. Figure 4 As shown in FIG. 1, the method can include the following steps:

[0097] S201, obtaining user language description information.

[0098] Optionally, the electronic device can receive the user language description information input by the user through an API, a GUI, or the like. Alternatively, the user language description information can also be pre-stored in the electronic device. In this implementation, the electronic device can read the user language description information from the data stored in itself.

[0099] S202. Obtain a target text feature sequence according to the user language description information.

[0100] For example, the electronic device can perform word embedding on the user language description information. Then, the electronic device can add position encoding to the language feature sequence obtained through word embedding to enhance the position information of each element in the sequence, and obtain a target text feature sequence corresponding to the user language description information.

[0101] Alternatively, the electronic device can also obtain a target text feature sequence according to the user language description information by referring to any existing text feature sequence acquisition method, which will not be described herein.

[0102] S203. Perform language feature extraction on the target text feature sequence through the trained first feature extraction model to obtain a target language feature corresponding to the target text feature sequence.

[0103] The first feature extraction model is trained based on the feature extraction model training method described in any of the preceding embodiments. The electronic device can input the target text feature sequence into the trained first feature extraction model, perform language feature extraction on the target text feature sequence through the trained first feature extraction model, and obtain a target language feature corresponding to the target text feature sequence.

[0104] S204. Determine a target user intent of the user according to the target language feature.

[0105] For example, the user intent can be querying a credit card bill, modifying a bank card password, or querying a recent service point.

[0106] Optionally, the electronic device can determine multiple target user intents or one target user intent of the user according to the target language feature.

[0107] As a possible implementation, the electronic device can determine the target user intent of the user based only on the target language feature. Alternatively, the electronic device can also determine the target user intent of the user based on the target language feature and historical behavior operation records of the user.

[0108] In this embodiment, the target text feature sequence can be obtained through the user language description information, and the target language feature can be obtained by performing language feature extraction on the target text feature sequence through the trained first feature extraction model. Then, the target user intent of the user can be determined based on the target language feature. Through the above method, the target language feature is obtained based on the first feature extraction model with high accuracy, so that the accuracy of the target language feature is improved, and the accuracy of the target user intent determined based on the target language feature is further improved.

[0109] As a possible implementation manner, after determining the target user intent of the user, the electronic device can also output the target user intent of the user. For example, the electronic device can output the k user intents with the maximum probability as the target user intent of the user according to the target language feature, as an example. For example, the electronic device can output the target user intent to the target APP, so that the user can perform subsequent operations according to the target user intent, thereby improving the user experience. Alternatively, the electronic device can also output the target user intent to the service personnel, so as to provide a basis for the service personnel to serve the user, thereby further improving the user experience. Through the above method, the intelligent customer service or the artificial customer service can be provided with the function of user intent prediction, thereby solving the problem of low accuracy of user intent prediction caused by less reference information or various user language descriptions.

[0110] The following describes how the electronic device determines the target user intent of the user according to the target language feature:

[0111] As a possible implementation manner, before determining the target user intent of the user according to the target language feature, the electronic device can also obtain the historical operation record of the user, and obtain the target historical operation feature sequence corresponding to the user according to the historical operation record. For example, the historical operation record and the manner of obtaining the target historical operation feature sequence corresponding to the user according to the historical operation record can refer to the foregoing embodiments, and will not be described herein.

[0112] Then, the electronic device can perform historical behavior feature extraction on the target historical operation feature sequence through the foregoing trained second feature extraction model, to obtain the target historical behavior feature corresponding to the user. The second feature extraction model can refer to the method described in the foregoing embodiments, and will not be described herein. Optionally, the electronic device can input the target historical operation feature sequence into the trained second feature extraction model, to perform historical behavior feature extraction on the target historical operation feature sequence through the second feature extraction model, to obtain the target historical behavior feature corresponding to the user.

[0113] Then, electronic devices can determine the user's target intent based on the target language characteristics and the target's historical behavioral characteristics.

[0114] In some embodiments, the electronic device can concatenate target language features and target historical behavior features to obtain a concatenated target feature. Optionally, the specific implementation of the electronic device concatenating target language features and target historical behavior features to obtain a concatenated target feature can refer to the specific implementation of "concatenating the user's sample language feature with the user's sample historical behavior feature to obtain a concatenated feature" described in the foregoing embodiments, which will not be repeated here.

[0115] Then, the electronic device can input the concatenated target feature into a trained user intent prediction model to obtain the user's target user intent. As described in the previous embodiment, this user intent prediction model is obtained by training a third preset model based on at least one concatenated feature. The concatenated feature is obtained by concatenating any sample language feature of any user with one sample historical behavior feature of that user. The sample language feature is obtained by inputting a sequence of sample text features into a trained first feature extraction model. The sample historical behavior feature is obtained by inputting a sequence of sample historical operation features into a trained second feature extraction model.

[0116] For example, let's call the trained first feature extraction model the language feature encoder, the trained second feature extraction model the behavior feature encoder, and the user intent prediction model the decoder. Figure 5 A flowchart illustrating another user intent prediction method provided in this application. Figure 5 As shown, the electronic device can input the target text feature sequence into the language feature encoder to obtain the target language features; and input the target historical operation feature sequence into the behavior feature encoder to obtain the target historical behavior features. Then, the target language features and the target historical behavior features are concatenated and input into the decoder to obtain the target user intent P.

[0117] In this embodiment, a first feature extraction model is used to extract user language description features, and a second feature extraction model is used to extract user historical behavior features. Then, a user intent prediction model is used to abstract and map these two types of feature information to obtain the final target user intent prediction result. By jointly extracting and abstracting features from user language descriptions and historical operation records, accurate prediction of user intent is achieved, improving the accuracy of user intent prediction. Using the above method, employing the high-performance first and second feature extraction models to extract user language description and behavioral features, compared to existing statistical methods, this application demonstrates stronger feature extraction capabilities and generalization performance in addressing issues such as rich user language descriptions and noise in information exchange.

[0118] Taking the first feature extraction model as an example of a transformer model encoder, by adding an additional encoder structure to the original transformer model and cascading two encoder structures, the user's language description features and recent operation record features are comprehensively considered, enriching the sources of information that can be relied upon in the user intent prediction problem and improving the accuracy of user intent prediction. Furthermore, the user intent prediction method based on the transformer model relies on massive amounts of data for its training process. As user data grows, the model weights can be updated and iterated, thereby improving the prediction accuracy and robustness of the method.

[0119] Figure 6 This is a schematic diagram of the structure of a feature extraction model training device provided in this application. Figure 6 As shown, the device may include: an acquisition module 31 and a training module 32. Wherein,

[0120] The acquisition module 31 is used to acquire a first sample dataset. The first sample dataset includes at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information.

[0121] The training module 32 is configured to use the first sample data set and the restoration module to perform N rounds of training on the first preset model to obtain a trained first feature extraction model; in the i-th round of training, the restoration module is used to restore the sample language feature output by the first preset model based on the input sample text feature sequence to obtain a text restoration feature sequence corresponding to the sample language feature; in the (i+1)-th round of training, the first preset model is trained based on the text restoration feature sequence obtained in the i-th round of training and the first sample data set. The N is an integer greater than or equal to 1, and the i is an integer greater than or equal to 1 and less than the N. The trained first feature extraction model is configured to perform language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence. The target text feature sequence is obtained based on user language description information. The target language feature is used to determine a target user intent of the user.

[0122] Optionally, the training module 32 is specifically configured to input the sample text feature sequence used in the i-th round of training, the cross-entropy of the sample language feature output by the first preset model in the i-th round of training, and the text restoration feature sequence obtained in the i-th round of training into a preset loss function, and perform the (i+1)-th round of training on the first preset model based on the loss function.

[0123] Optionally, the restoration module includes at least one fully connected neural network.

[0124] Optionally, the acquisition module 31 is further configured to acquire a second sample data set. The second sample data set includes at least one sample historical operation feature sequence. The sample historical operation feature sequence is obtained based on historical operation records. Optionally, the training module 32 is further configured to use the second sample data set to perform M rounds of training on a second preset model to obtain a trained second feature extraction model. The trained second feature extraction model is configured to perform historical behavior feature extraction on a target historical operation feature sequence corresponding to a user to obtain a target historical behavior feature of the user. The target historical operation feature sequence is obtained based on historical operation records of the user. The target historical behavior feature is used to determine a target user intent of the user. The M is an integer greater than or equal to 1.

[0125] Optionally, the obtaining module 31 is further configured to input at least one of the sample text feature sequence into the trained first feature extraction model to obtain at least one sample language feature, input at least one of the sample historical operation feature sequence into the trained second feature extraction model to obtain at least one sample historical behavior feature, and splice, for any sample language feature of any user, the sample language feature of the user and a sample historical behavior feature of the user to obtain a spliced feature. Optionally, the training module 32 is further configured to train a third preset model according to at least one of the spliced features to obtain a trained user intention prediction model. The user intention prediction model is configured to determine a target user intention of a user according to a target spliced feature corresponding to the user, and the target spliced feature is obtained by splicing a target language feature of the user and a target historical behavior feature.

[0126] Optionally, the obtaining module 31 is specifically configured to splice, by a channel stacking splicing method, the sample language feature of the user and the sample historical behavior feature of the user to obtain a spliced feature.

[0127] Optionally, the training module 32 is specifically configured to, for any spliced feature, take the spliced feature as an attribute and an attribute value in a key-value pair, and train the third preset model according to the key-value pair and a query vector to obtain a trained user intention prediction model. The query vector includes identifiers of a plurality of preset user intentions, and the plurality of preset user intentions include the target user intention.

[0128] The feature extraction model training apparatus provided in the application is used to execute the foregoing feature extraction model training method embodiments, and has similar implementation principles and technical effects, which will not be described herein again.

[0129] Figure 7 A structural schematic diagram of a user intention prediction apparatus provided in the application is shown in FIG. 4. Figure 7 As shown in FIG. 4, the apparatus can include an obtaining module 41, a processing module 42, an extracting module 43, and a determining module 44.

[0130] The obtaining module 41 is configured to obtain user language description information.

[0131] The processing module 42 is configured to obtain a target text feature sequence according to the user language description information.

[0132] The extraction module 43 is configured to perform language feature extraction on the target text feature sequence by using the trained first feature extraction model, to obtain a target language feature corresponding to the target text feature sequence. The first feature extraction model is trained based on the feature extraction model training method in any one of the preceding embodiments.

[0133] The determination module 44 is configured to determine the target user intent of the user according to the target language feature.

[0134] Optionally, the processing module 42 is further configured to obtain a historical operation record of the user before determining the target user intent of the user according to the target language feature, obtain a target historical operation feature sequence corresponding to the user according to the historical operation record, and perform historical behavior feature extraction on the target historical operation feature sequence by using a trained second feature extraction model, to obtain a target historical behavior feature corresponding to the user. Optionally, the determination module 44 is specifically configured to determine the target user intent of the user according to the target language feature and the target historical behavior feature. The second feature extraction model is trained based on a second sample data set. The second sample data set includes at least one sample historical operation feature sequence, and the sample historical operation feature sequence is obtained based on a historical operation record.

[0135] Optionally, the determination module 44 is specifically configured to splice the target language feature and the target historical behavior feature to obtain a spliced target feature, and input the spliced target feature into a trained user intent prediction model to obtain the target user intent of the user. The user intent prediction model is obtained by training a third preset model according to at least one spliced feature. The spliced feature is obtained by splicing a sample language feature of any user and a sample historical behavior feature of the user. The sample language feature is obtained by inputting a sample text feature sequence into the trained first feature extraction model. The sample historical behavior feature is obtained by inputting a sample historical operation feature sequence into the trained second feature extraction model.

[0136] The user intent prediction apparatus provided in the present application is used to execute the user intent prediction method embodiments, and has similar implementation principles and technical effects, which will not be described herein again.

[0137] Figure 8 An electronic device structure schematic diagram is provided in the present application. As shown in the figure, the electronic device 500 can include at least one processor 501 and a memory 502. Figure 8

[0138] ​The memory 502 is configured to store a computer program. Specifically, the program can include program code including computer operation instructions.

[0139] The memory 502 can include a high-speed RAM memory, and can further include a non-volatile memory such as at least one disk memory.

[0140] The processor 501 is configured to execute the computer operation instructions stored in the memory 502, so as to implement the feature extraction model training or the user intention prediction method described in the foregoing method embodiments. The processor 501 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0141] Optionally, the electronic device 500 can further include a communication interface 503. In a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are independently implemented, the communication interface 503, the memory 502 and the processor 501 can be connected with each other through a bus and complete communication therebetween. The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc., but does not mean that there is only one bus or one type of bus.

[0142] Optionally, in a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are integrated on a chip, the communication interface 503, the memory 502 and the processor 501 can complete communication through an internal interface.

[0143] The present application further provides a computer readable storage medium, which can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various media that can store program codes, and specifically, the computer readable storage medium stores program instructions, and the program instructions are used for the method in the foregoing embodiments.

[0144] The application also provides a program product including execution instructions stored in a readable storage medium. At least one processor of an electronic device can read the execution instructions from the readable storage medium, and the at least one processor executes the execution instructions to enable the electronic device to implement the feature extraction model training or the user intention prediction method provided by the various embodiments described above.

[0145] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A feature extraction model training method characterized by, The method comprises: obtaining a first sample data set; the first sample data set comprises at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information; using the first sample data set and a recovery module, performing N rounds of training on a first preset model to obtain a trained first feature extraction model; the trained first feature extraction model is used for performing language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence; the target text feature sequence is obtained based on user language description information; and the target language feature is used to determine a target user intent of the user; in the i-th round of training, the recovery module is used to restore sample language features output by the first preset model based on an input sample text feature sequence to obtain a text restoration feature sequence corresponding to the sample language features; in the (i+1)-th round of training, the first preset model is trained based on the text restoration feature sequence obtained in the i-th round of training and the first sample data set; N is an integer greater than or equal to 1, and i is an integer greater than or equal to 1 and less than N; The method further comprises: obtaining a second sample data set; the second sample data set comprises at least one sample historical operation feature sequence; the sample historical operation feature sequence is obtained based on historical operation records; using the second sample data set, performing M rounds of training on a second preset model to obtain a trained second feature extraction model; the trained second feature extraction model is used for performing historical behavior feature extraction on a target historical operation feature sequence corresponding to a user to obtain a target historical behavior feature corresponding to the user; the target historical operation feature sequence is obtained based on historical operation records of the user; and the target historical behavior feature is used to determine a target user intent of the user; and M is an integer greater than or equal to 1; The method further comprises: inputting at least one sample text feature sequence into the trained first feature extraction model to obtain at least one sample language feature; inputting at least one sample historical operation feature sequence into the trained second feature extraction model to obtain at least one sample historical behavior feature; for any sample language feature of any user, splicing the sample language feature of the user with a sample historical behavior feature of the user to obtain a spliced feature; training a third preset model based on at least one spliced feature to obtain a trained user intent prediction model; the user intent prediction model is used to determine a target user intent of a user based on a spliced target feature corresponding to the user; and the spliced target feature is obtained by splicing a target language feature of the user with a target historical behavior feature.

2. The method of claim 1, wherein, The method further comprises: The sample text feature sequence used in the i-th round of training, the cross-entropy of the sample language features output by the first preset model during the i-th round of training, and the text reconstruction feature sequence obtained in the i-th round of training are input into the preset loss function. The first preset model is trained for the (i+1)th round according to the loss function.

3. The method of claim 1, wherein, The restoration module includes at least one fully connected neural network layer.

4. The method of claim 1, wherein, The step of concatenating the user's sample language feature with the user's sample historical behavior feature to obtain a concatenated feature includes: By using the channel stacking and splicing method, the language feature of this user sample is spliced ​​with the historical behavior feature of this user sample to obtain a spliced ​​feature.

5. The method of claim 1, wherein, The step of training the third preset model based on at least one of the concatenated features to obtain a trained user intent prediction model includes: For any of the concatenated features, the concatenated feature is used as an attribute in a key-value pair, and the attribute value is used as an attribute value. Based on the key-value pairs and the query vector, the third preset model is trained to obtain a trained user intent prediction model; wherein, the query vector includes: identifiers of multiple preset user intents, and the multiple preset user intents include the target user intent. 6.A user intention prediction method, characterized by, The method includes: Obtain user language description information; Based on the user language description information, the target text feature sequence is obtained; Using a trained first feature extraction model, language features are extracted from the target text feature sequence to obtain the target language features corresponding to the target text feature sequence; the first feature extraction model is trained based on the feature extraction model training method described in any one of claims 1-5; the target user intent of the user is determined based on the target language features; The method further includes: Obtain the user's historical operation records; Based on the historical operation records, the target historical operation feature sequence corresponding to the user is obtained; By using the trained second feature extraction model, historical behavior features are extracted from the target historical operation feature sequence to obtain the target historical behavior features corresponding to the user; the second feature extraction model is trained based on a second sample dataset, which includes at least one sample historical operation feature sequence; the sample historical operation feature sequence is obtained based on historical operation records. Based on the target language features and the target historical behavior features, the target user intent of the user is determined; Determining the user's target user intent based on the target language features and the target historical behavior features includes: The target language features and the target historical behavior features are concatenated to obtain a concatenated target feature; input the spliced target feature to the trained user intention prediction model to obtain a target user intention of the user; the user intention prediction model is obtained by training a third preset model according to at least one spliced feature; the spliced feature is obtained by splicing a sample language feature of any user and a sample historical behavior feature of the user for any sample language feature of the user; the sample language feature is obtained by inputting a sample text feature sequence into the trained first feature extraction model; and the sample historical behavior feature is obtained by inputting a sample historical operation feature sequence into the trained second feature extraction model.

7. A feature extraction model training apparatus characterized by comprising: The device comprises: an acquisition module configured to acquire a first sample data set; the first sample data set comprises at least one sample text feature sequence; the sample text feature sequence is obtained based on sample language description information; a training module configured to train a first preset model for N rounds using the first sample data set and a restoration module to obtain a trained first feature extraction model; the trained first feature extraction model is configured to perform language feature extraction on a target text feature sequence to obtain a target language feature corresponding to the target text feature sequence; the target text feature sequence is obtained based on user language description information; and the target language feature is used to determine a target user intention of the user; in the i th round of training, the first preset model outputs a sample language feature based on an input sample text feature sequence, and the restoration module restores the sample language feature to obtain a text restoration feature sequence corresponding to the sample language feature; in the i+1 th round of training, the first preset model is trained for the i+1 th round according to the text restoration feature sequence obtained in the i th round of training and the first sample data set; N is an integer greater than or equal to 1, and i is an integer greater than or equal to 1 and less than N; the acquisition module is further configured to acquire a second sample data set; the second sample data set comprises at least one sample historical operation feature sequence; the sample historical operation feature sequence is obtained based on historical operation records; the training module is further configured to train a second preset model for M rounds using the second sample data set to obtain a trained second feature extraction model; the trained second feature extraction model is configured to perform historical behavior feature extraction on a target historical operation feature sequence corresponding to a user to obtain a target historical behavior feature corresponding to the user; the target historical operation feature sequence is obtained based on historical operation records of the user; and the target historical behavior feature is used to determine a target user intention of the user; and M is an integer greater than or equal to 1. The acquisition module is further configured to input at least one of the sample text feature sequences into the trained first feature extraction model to obtain at least one sample language feature, input at least one of the sample historical operation feature sequences into the trained second feature extraction model to obtain at least one sample historical behavior feature, and splice, for any sample language feature of any user, the sample language feature of the user with a sample historical behavior feature of the user to obtain a spliced feature. The training module is further configured to train a third preset model according to at least one of the spliced features to obtain a trained user intention prediction model, and the user intention prediction model is configured to determine a target user intention of a user according to a target spliced feature corresponding to the user, and the target spliced feature is obtained by splicing a target language feature of the user with a target historical behavior feature.

8. A user intention prediction apparatus characterized by comprising: The apparatus comprises: An acquisition module configured to acquire user language description information; A processing module configured to obtain a target text feature sequence according to the user language description information; An extraction module configured to perform language feature extraction on the target text feature sequence by using a trained first feature extraction model to obtain a target language feature corresponding to the target text feature sequence, and the first feature extraction model is trained based on the feature extraction model training method in any one of claims 1 to 5; A determination module configured to determine a target user intention of the user according to the target language feature; The processing module is further configured to acquire historical operation records of the user, obtain a target historical operation feature sequence corresponding to the user according to the historical operation records, and perform historical behavior feature extraction on the target historical operation feature sequence by using a trained second feature extraction model to obtain a target historical behavior feature corresponding to the user, and the second feature extraction model is trained based on a second sample data set, and the second sample data set comprises at least one sample historical operation feature sequence, and the sample historical operation feature sequence is obtained based on historical operation records; The determination module is configured to determine the target user intention of the user according to the target language feature and the target historical behavior feature. The determining module is specifically configured to splice the target language feature and the target historical behavior feature to obtain a spliced target feature; and input the spliced target feature into a trained user intent prediction model to obtain the target user intent of the user; the user intent prediction model is obtained by training a third preset model according to at least one spliced feature; the spliced feature is obtained by splicing any sample language feature of any user and a sample historical behavior feature of the user; the sample language feature is obtained by inputting a sample text feature sequence into the trained first feature extraction model; and the sample historical behavior feature is obtained by inputting a sample historical operation feature sequence into the trained second feature extraction model.

9. An electronic device, comprising: The electronic device comprises a memory and a processor; The memory stores a computer program; The processor is configured to execute the method of any one of claims 1-6 through the computer program.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer execution instructions, and when the computer execution instructions are executed by the processor, the method of any one of claims 1-6 is implemented.

11. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-6.

Images