Appeal processing method and apparatus, and storage medium

Abstract

The application provides a claim processing method and device and a storage medium. The method comprises the following steps: obtaining a plurality of groups of consultation item data of a current customer by predicting the current customer's historical behavior related data; obtaining a plurality of groups of pre-training data by using the plurality of groups of consultation item data and a preset user historical data set and a claim historical data set; and recursively deriving target claim reply information of the current customer by sequentially performing weak contrast learning on the plurality of groups of pre-training data. Since the current conversation information of the customer is ignored, the target claim reply information is determined by performing more fine-grained claim analysis in multiple levels including a scene type, an item type and a question type, so that the soothing information matched with the customer can be accurately provided.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of Internet, and in particular, to a complaint processing method and device and storage medium. BACKGROUND

[0002] In the e-commerce field, how to appease customer emotions is an important content to improve customer experience and increase customer repurchase rate. The method of appeasing customer emotions in the related art mostly relies on the experience value of customer service. Experienced customer service can quickly find out the problems of customers and solve customer problems in a better way. Some customers are reasonable, while some customers are more demanding. The existing methods that can identify customer emotions and appease them mainly include the following:

[0003] (1) Appeasing customers based on the training or experience of customer service.

[0004] (2) Appeasing customers based on sentiment analysis of customers.

[0005] However, the differences in customer personality and regional language expression make the same or similar customer speech represent different meanings, and the model processes them without distinction, which not only introduces a lot of noise to the ability of the model itself, but also misleads the customer service. Therefore, the existing method is prone to introduce a lot of noise when analyzing and appeasing different users and different items, resulting in inaccurate analysis and providing appeasing information with large errors. SUMMARY

[0006] The complaint processing method, device and storage medium provided by the embodiments of the present application can accurately provide appeasing information matched with customers.

[0007] The technical solution of the present application is implemented as follows:

[0008] The complaint processing method provided by the embodiments of the present application comprises:

[0009] By extracting the historical behavior related data corresponding to the current customer, a plurality of sets of consultation item data of the current customer are predicted. Different sets of consultation item data include different scene types, matter types and problem types of information predicted corresponding to the current customer.

[0010] A plurality of pre-training data sets are extracted from the plurality of sets of consultation item data and the preset user historical data set and the appeal historical data set; each pre-training data set includes consultation item information identical to the scene class, the matter class, and the question class information, and appeal reply information corresponding to the consultation item information; the user historical data set includes a plurality of first information sets; each first information set includes first scene class, first matter class, and first question class information; the appeal historical data set includes a plurality of second information sets, and each second information set includes second scene class, second matter class, second question class information, and corresponding appeal reply information; weakly supervised contrast learning is sequentially performed on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer.

[0011] In the above solution, before the weakly supervised contrast learning is sequentially performed on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer, the method further includes:

[0012] The plurality of pre-training data sets are sequentially subjected to vectorization processing to obtain a scene class vector corresponding to the scene class in each pre-training data set, a matter class vector corresponding to the matter class, and a question class vector corresponding to the question class information.

[0013] The scene class vector, the matter class vector, and the question class vector are combined to obtain a first matrix corresponding to each pre-training data set.

[0014] In the above solution, the weakly supervised contrast learning is sequentially performed on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer, including:

[0015] The first matrix is combined with a weakly supervised contrast learning model and pre-obtained scene encoding, matter encoding, and question encoding to obtain a scene class first vector, a matter class second vector, and a question class third vector corresponding to each pre-training data set.

[0016] The scene class first vector, the matter class second vector, and the question class third vector are combined with parameter information of the weakly supervised contrast learning model to calculate a minimum confidence of each pre-training data set, and further obtain the minimum confidence of the plurality of pre-training data sets.

[0017] Based on the minimum confidence, the target appeal reply information for the current customer is determined from the plurality of pre-training data sets.

[0018] In the above solution, the first matrix is combined with a weakly supervised contrast learning model and pre-obtained scene encoding, matter encoding, and question encoding to obtain a scene class first vector, a matter class second vector, and a question class third vector corresponding to each pre-training data set, including:

[0019] multiplying the first matrix with a square of the scene encoding, a preset first intermediate value, the scene class vector and a predetermined parameter of the weakly supervised contrast learning model, to obtain the scene class first vector; the preset first intermediate value is a probability value equal to a product of the scene class vector and the calculation matrix of the weakly supervised contrast learning model and the scene class vector;

[0020] multiplying the first matrix with a square of the transaction encoding, a preset second intermediate value, the scene class first vector and the predetermined parameter of the weakly supervised contrast learning model, to obtain the transaction class second vector; the preset second intermediate value is a probability value equal to a product of the scene class first vector and the calculation matrix of the weakly supervised contrast learning model and the scene class first vector;

[0021] multiplying the first matrix with a square of the transaction encoding, a preset third intermediate value, the transaction class second vector and the predetermined parameter of the weakly supervised contrast learning model, to obtain the problem class third vector; the preset third intermediate value is a probability value equal to a product of the transaction class second vector and the calculation matrix of the weakly supervised contrast learning model and the transaction class second vector.

[0022] In the above scheme, the combination of the scene class first vector, the transaction class second vector and the problem class third vector, and the parameter information of the weakly supervised contrast learning model, the minimum confidence of each group of pre-training data is calculated, and then the minimum confidence of the multiple groups of pre-training data is obtained, comprising:

[0023] multiplying the scene class first vector corresponding to each group of pre-training data with the scene encoding, the predetermined parameter and the first matrix to obtain a first intermediate value, and multiplying the scene class difference value converted from the scene class first vector with the scene class first vector, the scene encoding and the predetermined parameter to obtain a second intermediate value;

[0024] combining the first intermediate value, the second intermediate value and the problem class third vector, a scene class confidence is calculated;

[0025] multiplying the transaction class second vector with the transaction encoding, the predetermined parameter and the first matrix to obtain a third intermediate value, and multiplying the transaction class difference value converted from the transaction class second vector with the transaction class second vector, the transaction encoding and the predetermined parameter to obtain a fourth intermediate value;

[0026] combining the third intermediate value, the fourth intermediate value and the problem class third vector, a transaction class confidence is calculated;

[0027] multiplying the problem-type third vector with the problem code, the predetermined parameter and the first matrix to obtain a fifth intermediate value, and multiplying the problem-type third vector converted problem-type difference value with the problem-type third vector, the problem code and the predetermined parameter to obtain a sixth intermediate value;

[0028] combining the fifth intermediate value, the sixth intermediate value and the problem-type third vector to obtain a problem-type confidence;

[0029] determining the minimum confidence among the scene-type confidence, the matter-type confidence and the problem-type confidence as the minimum confidence.

[0030] In the above scheme, the combining the first intermediate value, the second intermediate value and the problem-type third vector to obtain a scene-type confidence comprises:

[0031] calculating a first sum of the first intermediate value and the second intermediate value;

[0032] taking a first reciprocal of a second sum of the problem-type third vector and a normalized vector of the problem-type third vector;

[0033] multiplying the first reciprocal with the first sum to obtain the scene-type confidence.

[0034] In the above scheme, the combining the third intermediate value, the fourth intermediate value and the problem-type third vector to obtain a matter-type confidence comprises:

[0035] calculating a third sum of the third intermediate value and the fourth intermediate value;

[0036] taking a first reciprocal of a second sum of the problem-type third vector and a normalized vector of the problem-type third vector;

[0037] multiplying the first reciprocal with the third sum to obtain the matter-type confidence.

[0038] In the above scheme, the combining the fifth intermediate value, the sixth intermediate value and the problem-type third vector to obtain a problem-type confidence comprises:

[0039] calculating a fourth sum of the fifth intermediate value and the sixth intermediate value;

[0040] taking a first reciprocal of a second sum of the problem-type third vector and a normalized vector of the problem-type third vector;

[0041] multiplying the first reciprocal with the fourth sum to obtain the problem-type confidence.

[0042] In the scheme, the determination of the target appeal reply information for the current customer based on the minimum confidence and the plurality of groups of pre-training data comprises:

[0043] Among the plurality of minimum confidences corresponding to the plurality of groups of pre-training data, a target confidence greater than a confidence threshold is determined.

[0044] The target appeal reply information is extracted from the pre-training data corresponding to the target confidence.

[0045] In the scheme, before the plurality of groups of consultation item data of the current customer are predicted by the extracted historical behavior related data corresponding to the current customer, the method further comprises:

[0046] The identification information corresponding to the current customer is extracted.

[0047] Correspondingly, the plurality of groups of consultation item data of the current customer are predicted by the extracted historical behavior related data corresponding to the current customer, comprising:

[0048] The plurality of click behavior information and the corresponding form data of the current customer in a first historical time period are extracted in a local database through the identification information; the historical behavior related data comprises the plurality of click behavior information and the corresponding form data;

[0049] The plurality of click behavior information and the corresponding form data are input into a prediction model to obtain the plurality of groups of consultation item data corresponding to the plurality of click behavior information.

[0050] In the scheme, the plurality of groups of pre-training data are extracted by the plurality of groups of consultation item data and a preset user historical data set and an appeal historical data set, comprising:

[0051] In the user historical data set, the first target scene class, the first target matter class and the first target question class information same as the scene class, the matter class and the question class information in each group of consultation item data are extracted;

[0052] In the appeal historical data set, the second target scene class, the second target matter class and the second target question class information same as the scene class, the matter class and the question class information in each group of consultation item data are extracted, and the appeal reply information included in the second target information group to which the second target scene class, the second target matter class and the second target question class information belong;

[0053] The third scene class, the third matter class and the third question class information are obtained by merging and unifying the each group of consultation item data and the first target scene class, the first target matter class, the first target question class information, the second target scene class, the second target matter class and the second target question class information; the consultation item information includes the third scene class, the third matter class and the third question class information

[0054] The third scene class, the third matter class and the third question class information are obtained by merging and unifying the each group of consultation item data and the first target scene class, the first target matter class, the first target question class information, the second target scene class, the second target matter class and the second target question class information; the consultation item information includes the third scene class, the third matter class and the third question class information

[0055] In the above scheme, before the multiple groups of consultation item data of the current customer are predicted by the extracted historical behavior related data corresponding to the current customer, the method further comprises:

[0056] Obtaining current appeal session information of the current customer;

[0057] Performing sentiment analysis on the appeal session information to obtain an emotion value;

[0058] If it is detected that the emotion value exceeds a negative emotion threshold value, the historical behavior related data of the current customer is extracted.

[0059] In the above scheme, after the target appeal reply information of the current customer is recursively derived by sequentially performing weakly supervised contrast learning on the multiple groups of pre-training data, the method further comprises:

[0060] The target appeal reply information is pushed to a current client corresponding to the current customer.

[0061] The embodiment of the application also provides a kind of appeal processing device, comprising:

[0062] A prediction unit is configured to predict multiple groups of consultation item data of a current customer by extracting historical behavior related data corresponding to the current customer; different groups of consultation item data include different scene class, matter class and question class information predicted for the current customer;

[0063] The processing unit is configured to extract a plurality of pre-training data sets from the plurality of sets of consultation item data and a preset user historical data set and a preset appeal historical data set, wherein each pre-training data set comprises consultation item information identical to the scene class, the matter class and the question class information, and appeal reply information corresponding to the consultation item information; the user historical data set comprises a plurality of first information sets, each first information set comprising first scene class information, first matter class information and first question class information; and the appeal historical data set comprises a plurality of second information sets, each second information set comprising second scene class information, second matter class information, second question class information and corresponding appeal reply information.

[0064] The processing unit is configured to sequentially perform weakly supervised contrast learning on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer.

[0065] The embodiment of the present application also provides a complaint processing device, comprising a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the steps in the above method when executing the program.

[0066] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the above method.

[0067] In the embodiment of the present application, a plurality of sets of consultation item data of the current customer are predicted by extracting historical behavior related data corresponding to the current customer; different sets of consultation item data comprise different scene class information, matter class information and question class information corresponding to the current customer predicted by the current customer; a plurality of pre-training data sets are extracted from the plurality of sets of consultation item data and a preset user historical data set and a preset appeal historical data set, wherein each pre-training data set comprises consultation item information identical to the scene class, the matter class and the question class information, and appeal reply information corresponding to the consultation item information; the user historical data set comprises a plurality of first information sets, each first information set comprising first scene class information, first matter class information and first question class information; the appeal historical data set comprises a plurality of second information sets, each second information set comprising second scene class information, second matter class information, second question class information and corresponding appeal reply information; and weakly supervised contrast learning is sequentially performed on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer. Since the current conversation information of the customer is ignored in the present solution, the target appeal reply information is determined through more fine-grained appeal analysis of multiple levels of scene class, matter class and question class, so that the soothing information matched with the customer can be accurately provided. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0069] Figure 2 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0070] Figure 3 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0071] Figure 4 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0072] Figure 5 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0073] Figure 6 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0074] Figure 7 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0075] Figure 8 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0076] Figure 9 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0077] Figure 10 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0078] Figure 11 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0079] Figure 12 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0080] Figure 13 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0081] Figure 14 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0082] Figure 15 An optional flowchart of the solicitation processing method provided by the embodiment of the present application is shown in FIG. 1;

[0083] Figure 16A hardware entity schematic diagram of the appeal processing device provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0084] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further described in detail below in combination with the drawings and embodiments, and the described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0085] In the following description, "some embodiments" are described, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0086] If the similar description of "first / second" appears in the invention document, the following description is added, in the following description, the terms "first\second\third" only distinguish similar objects, and do not represent the specific order of the objects, and it can be understood that "first\second\third" can be interchanged in the specific order or sequence as allowed, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0087] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0088] Figure 1 An optional flowchart of the appeal processing method provided by the embodiment of the present application is shown, which will be described in combination with the steps shown in the figure. Figure 1 The steps shown in the figure are described.

[0089] S101, a plurality of groups of consulting item data of the current customer are predicted by extracting the historical behavior related data corresponding to the current customer; different groups of consulting item data include: different scene types, matter types and problem types of information predicted corresponding to the current customer.

[0090] In the embodiment of the present application, the server predicts a plurality of groups of consulting item data of the current customer by extracting the historical behavior related data corresponding to the current customer. Different groups of consulting item data include: different scene types, matter types and problem types of information predicted corresponding to the current customer.

[0091] In the embodiment of the present application, the server extracts the historical behavior related data of the current customer in the first historical time period according to the identification information of the current customer in the current session in the local database. The server inputs the historical behavior related data into the prediction model to obtain a plurality of sets of consultation item data. The prediction model can be a general framework multi-perspective double-feedback prediction model.

[0092] In the embodiment of the present application, each set of consultation item data includes different types of scene class, matter class and problem class information predicted for the current customer. The scene class can include after-sales service class scene, promotion scene and hot selling scene, etc. The matter class can include return matter, replacement matter and replenishment matter. The problem class can include product error, product order allocation error, product size error and product expiration, etc. Each consultation item data includes matter class and scene class information corresponding to a problem class information. For example, a certain consultation item data can include three types of information of after-sales scene, return matter and product error.

[0093] The first historical time period can be a time period of half an hour or 1 hour before the current time. The embodiment of the present application does not limit the specific duration of the first historical time period.

[0094] In the embodiment of the present application, the historical behavior related data can be the behavior information clicked by the current customer in the historical time period on the client and the form data information corresponding to each behavior information. For example, the current customer clicks the price protection button in the historical time period, and the button parameter carries the order number, price and other parameters at the time of price protection and is stored in the local database of the server.

[0095] S102, a plurality of sets of pre-training data are extracted by the plurality of sets of consultation item data and the preset user historical data set and the appeal historical data set.

[0096] In the embodiment of the present application, the server extracts a plurality of sets of pre-training data from the plurality of sets of consultation item data and the preset user historical data set and the appeal historical data set. Each set of pre-training data includes the same consultation item information as the scene class, matter class and problem class information, and the appeal reply information corresponding to the consultation item information. The user historical data set includes a plurality of first information groups. Each first information group includes first scene class, first matter class and first problem class information. The appeal historical data set includes a plurality of second information groups. Each second information group includes second scene class, second matter class, second problem class information and corresponding appeal reply information.

[0097] In the embodiment of the application, the server extracts the same scene class, matter class and question class information from the user historical data set and the appeal historical data set respectively through the scene class, matter class and question class information in each set of consultation item data, and then forms pre-training data corresponding to each set of consultation item data by unifying the same scene class, matter class and question class information. Then, the server extracts different class scene class, matter class and question class information combination corresponding appeal reply information from the appeal historical data set and adds it to the corresponding pre-training data. Then, a plurality of pre-training data sets are formed.

[0098] In the embodiment of the application, the server extracts the same scene class, matter class and question class information from the user historical data set and the appeal historical data set respectively through the scene class, matter class and question class information in each set of consultation item data, and then forms pre-training data corresponding to each set of consultation item data by unifying the same scene class, matter class and question class information. Then, the server extracts different class scene class, matter class and question class information combination corresponding appeal reply information from the appeal historical data set and adds it to the corresponding pre-training data. Then, a plurality of pre-training data sets are formed.

[0099] In the embodiment of the application, the appeal reply information can be solution method information corresponding to the scene class, matter class and question class information.

[0100] S103, sequentially performing weakly supervised contrast learning through the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer.

[0101] In the embodiment of the application, the server recursively performs weakly supervised contrast learning through the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer.

[0102] In the embodiment of the application, the server recursively inputs the plurality of pre-training data sets into the weakly supervised contrast learning model to calculate the confidence of the weakly supervised contrast learning model corresponding to each pre-training data set. The server determines a target confidence greater than a confidence threshold value from a plurality of confidences corresponding to the plurality of pre-training data sets. The server extracts the appeal reply information from the pre-training data corresponding to the target confidence as the target appeal reply information. The server sends the target appeal reply information to the current client corresponding to the current customer.

[0103] In the embodiment of the present application, when there are multiple confidence values greater than the confidence threshold, the server determines the maximum confidence value as the target confidence value.

[0104] For example, in combination with Figure 2 An optional flowchart of the claim processing method provided by the embodiment of the present application will be described in combination with the steps.

[0105] S201, user consultation behavior.

[0106]

[0107] S203, consultation item.

[0108] S204, consultation target.

[0109] In this item, the behavior data of the two behaviors of behavior and form are corresponded. A multi-view two-way feedback model is used for multi-level prediction, and the consultation target including scene, matter and question is obtained.

[0110] S205, user history library.

[0111] S206, claim history library.

[0112] This step is mainly to learn the consultation target output by the model in the previous step according to the existing historical behavior. The two standard information libraries are mainly derived from the contents of the well-handled cases in the history and gradually form.

[0113] S207, weakly supervised contrast learning.

[0114] This step is to learn the consultation content, consultation item and history library and output target claim reply information.

[0115] S208, monitoring component.

[0116] S209, auxiliary component.

[0117] S210, feedback component.

[0118] The monitoring component is used to monitor whether the threshold value exceeds the set negative emotion, and after exceeding, the auxiliary component is used to retrieve the target claim reply information, and the feedback component is used to feed back to the current client. After auxiliary solution, the target claim reply information or the corrected target claim reply information is stored in the user history library or the claim history library.

[0119] In this embodiment of the invention, multiple sets of consultation item data for the current customer are predicted by extracting relevant historical behavior data corresponding to the current customer. Different sets of consultation item data include different scenario, matter, and problem information predicted for the current customer. Multiple sets of pre-training data are extracted using multiple sets of consultation item data and preset user historical data sets and request historical data sets. Each set of pre-training data includes consultation item information identical to the scenario, matter, and problem information, and request response information corresponding to the consultation item information. The user historical data set includes multiple first information groups; each first information group includes first scenario, first matter, and first problem information. The request historical data set includes multiple second information groups; each second information group includes second scenario, second matter, and second problem information, and corresponding request response information. Weakly supervised comparative learning is performed sequentially using multiple sets of pre-training data to derive the target request response information for the current customer, and the target request response information is pushed to the current client. Because this solution ignores the customer's current conversation information, it determines the target request response information through a more granular request analysis that includes multiple levels such as scenario, matter, and problem, so it can accurately provide reassurance information that matches the customer.

[0120] In some embodiments, see Figure 3 , Figure 3 This is a schematic diagram of an optional flowchart of the prescription information processing method provided in an embodiment of the present invention. Figure 1 The shown S103 can be implemented through S104 to S108, which will be explained in conjunction with each step.

[0121] S104. The multiple sets of pre-training data are vectorized sequentially to obtain scene class vectors, item class vectors, and question class vectors corresponding to scene class, item class, and question class information in each set of pre-training data.

[0122] In this embodiment of the invention, the server sequentially vectorizes multiple sets of pre-trained data to obtain scene class vectors for scene classes, item class vectors for item classes, and question class vectors for question information in each set of pre-trained data.

[0123] In the embodiment of the present application, the server inputs a plurality of sets of pre-training data into a vector conversion model (for example, a word2vec model, which includes a Continuous Bag-of-Word Model (CBOW Model) and a skip-gram model) to convert the scene class, matter class and question class information in each set of pre-training data into a vector, thereby obtaining a scene class vector, a matter class vector and a question class vector. In the embodiment of the present application, the server can also obtain the scene class vector, the matter class vector and the question class vector by processing the pre-training data through other vector conversion models.

[0124] In the embodiment of the present application, the server obtains a first matrix corresponding to each set of pre-training data by combining the scene class vector, the matter class vector and the question class vector.

[0125] In the embodiment of the present application, the server obtains a first matrix corresponding to each set of pre-training data by combining the scene class vector, the matter class vector and the question class vector.

[0126] In the embodiment of the present application, the server takes the scene class vector corresponding to each set of pre-training data as the first row of the first matrix, takes the matter class vector as the second row of the first matrix, and takes the question class vector as the third row of the first matrix, thereby obtaining the first matrix of each set of pre-training data. The server can obtain a plurality of first matrices corresponding to a plurality of sets of pre-training data by using this method.

[0127] In the embodiment of the present application, the server obtains a scene class first vector, a matter class second vector and a question class third vector corresponding to each set of pre-training data by combining the weakly supervised contrast learning model and the pre-obtained scene encoding, matter encoding and question encoding through the first matrix.

[0128] In the embodiment of the present application, the server obtains a scene class first vector, a matter class second vector and a question class third vector corresponding to each set of pre-training data by combining the weakly supervised contrast learning model and the pre-obtained scene encoding, matter encoding and question encoding through the first matrix.

[0129] In the embodiment of the present application, when predicting a plurality of sets of consultation item data, the server can obtain a scene encoding, a matter encoding and a question encoding corresponding to each set of consultation item data.

[0130] In the embodiment of the present application, the server obtains a scene class first vector of each set of pre-training data by combining the weakly supervised contrast learning model and the first matrix of each set of pre-training data, and then obtains a matter class second vector by combining the scene class first vector and the first matrix, and obtains a question class third vector by combining the matter class second vector and the first matrix.

[0131] S107, in combination with the scene class first vector, the matter class second vector and the question class third vector, and the parameter information of the weakly supervised contrast learning model, calculate the minimum confidence of each group of pre-training data, and then obtain the minimum confidence of the multiple groups of pre-training data.

[0132] In the embodiment of the application, the server combines the scene class first vector, the matter class second vector and the question class third vector, and the parameter information of the weakly supervised contrast learning model, calculates the minimum confidence of each group of pre-training data, and then obtains the minimum confidence of the multiple groups of pre-training data.

[0133] In the embodiment of the application, the server combines the scene class first vector, the matter class second vector and the question class third vector, and the parameter information of the weakly supervised contrast learning model, calculates the minimum confidence of each group of pre-training data, and then obtains the minimum confidence of the multiple groups of pre-training data.

[0134] S108, based on the minimum confidence, determine the target appeal reply information for the current customer from the multiple groups of pre-training data.

[0135] In the embodiment of the application, the server determines the target appeal reply information for the current customer based on the minimum confidence and the multiple groups of pre-training data.

[0136] In the embodiment of the application, the server determines the target confidence greater than the confidence threshold from the multiple minimum confidences corresponding to the multiple groups of pre-training data. The server further extracts the target appeal reply information from the pre-training data corresponding to the target confidence.

[0137] In the embodiment of the application, the server first vectorizes each group of pre-training data to obtain a first matrix corresponding to each group of pre-training data, facilitating the processing of text information by the computer; and the server accurately determines the target appeal reply information according to the minimum confidence corresponding to each group of pre-training data.

[0138] In some embodiments, referring to Figure 4 , Figure 4 An optional flowchart of the prescription information processing method provided by the embodiment of the application is shown in Figure 3 S104 to S108 shown in the figure can be implemented by S211 to S219, which will be described in combination with each step.

[0139] S211, user consultation.

[0140] S212, user history library.

[0141] S213, appeal history library.

[0142] S214, supplementary label.

[0143] S215, scene weakly supervised contrast.

[0144] The scene weakly supervised contrast can be obtained through scene class vector transformation.

[0145] S216, matter weakly supervised contrast.

[0146] The matter weakly supervised contrast can be obtained through matter class vector transformation.

[0147] S217, question weakly supervised contrast.

[0148] The question weakly supervised contrast can be obtained through question class vector transformation.

[0149] S218, threshold output.

[0150] S219, solution.

[0151] First, the user history library and the appeal history library are pre-trained using the model predictive control (MPC) method. The pre-training method is performed using the mark mechanism. Then, if the user consultation is in the user history library, the user consultation content is fine-tuned with the user's historical consultation. Note that the embedding includes Speaker, Segment, and Token embedding methods, which are used to distinguish different users from historical data. Then, the supplementary label is encoded. The supplementary label mainly refers to the encoding of the known order number of the commodity order attribute and some specific form items, which is used to distinguish the differences between different commodities. Then, the scene, matter, and question are respectively subjected to weakly supervised contrast learning.

[0152] In some embodiments, referring to Figure 5 , Figure 5 An optional flowchart of a prescription information processing method provided by an embodiment of the present application is shown in Figure 4 S106 shown can be implemented by S109-S111, which will be described in conjunction with each step.

[0153] S109, multiply the first matrix with the square of the scene code, the preset first intermediate value, the scene class vector and the predetermined parameter of the weak supervision contrast learning model to obtain a scene class first vector.

[0154] In the embodiment of the application, the server multiplies the first matrix with the square of the scene code, the preset first intermediate value, the scene class vector and the predetermined parameter of the weak supervision contrast learning model to obtain a scene class first vector.

[0155] The preset first intermediate value is a probability value equal to the product of the calculation matrix of the weak supervision contrast learning model and the scene class vector.

[0156] The predetermined parameter can be a parameter of the weak supervision contrast learning model.

[0157] In the embodiment of the application, the server can convert the first vector A into a decimal number, and further obtain a corresponding scene weak supervision contrast difference value.

[0158] S110, multiply the first matrix with the square of the transaction code, the preset second intermediate value, the scene class first vector and the predetermined parameter of the weak supervision contrast learning model to obtain a transaction class second vector.

[0159] In the embodiment of the application, the server multiplies the first matrix with the square of the transaction code, the preset second intermediate value, the scene class first vector and the predetermined parameter of the weak supervision contrast learning model to obtain a transaction class second vector.

[0160] In the embodiment of the application, the server converts the transaction class second vector into a decimal number, and further obtains a corresponding transaction weak supervision contrast difference value.

[0161] The preset second intermediate value is a probability value equal to the product of the calculation matrix of the weak supervision contrast learning model and the scene class first vector.

[0162] S111, multiply the first matrix with the square of the transaction code, the preset third intermediate value, the transaction class second vector and the predetermined parameter of the weak supervision contrast learning model to obtain a problem class third vector.

[0163] In the embodiment of the application, the server multiplies the first matrix with the square of the transaction code, the preset third intermediate value, the transaction class second vector and the predetermined parameter of the weak supervision contrast learning model to obtain a problem class third vector.

[0164] The preset third intermediate value is a probability value equal to the product of the calculation matrix of the weak supervision contrast learning model and the transaction class second vector.

[0165] In this embodiment of the invention, the server converts the third vector of the problem class into a decimal, thereby obtaining the corresponding weak supervision comparison difference value of the problem.

[0166] In this embodiment of the invention, the server multiplies the first matrix with the square of the scene code, a preset first intermediate value, the scene class vector, and predetermined parameters of the weakly supervised contrastive learning model to obtain a first scene class vector. Then, it calculates a second item class vector based on the first scene class vector, and a third question class vector based on the second item class vector. Finally, it determines the minimum confidence level based on the third question class vector. Since the server considers the three aspects of scene class, item class, and question class in the process of calculating the minimum confidence level, and determines the minimum confidence level by recursively using the first scene class vector, the second item class vector, and the third question class vector, the minimum confidence level determined by the finer-grained method is more accurate. Therefore, the determined target request response information is also more accurate.

[0167] In some embodiments, see Figure 6 , Figure 6 This is a schematic diagram of an optional flowchart of the prescription information processing method provided in an embodiment of the present invention. Figure 4 The shown S107 can be implemented through S112 to S118, which will be explained in conjunction with each step.

[0168] S112. Multiply the first vector of the scene class corresponding to each group of pre-trained data with the scene code, the predetermined parameters and the first matrix to obtain the first intermediate value. Multiply the scene class difference value transformed from the first vector of the scene class with the first vector of the scene class, the scene code and the predetermined parameters to obtain the second intermediate value.

[0169] In this embodiment of the invention, the server multiplies the scene class first vector corresponding to each set of pre-trained data with the scene code, predetermined parameters and the first matrix to obtain a first intermediate value, and multiplies the scene class difference value transformed from the scene class first vector with the scene class first vector, the scene code and the predetermined parameters to obtain a second intermediate value.

[0170] In this embodiment of the invention, the server can calculate the modulus of the first vector of the scene class, and then obtain the scene class difference value.

[0171] S113. Combine the first intermediate value, the second intermediate value, and the third vector of the problem class to calculate the confidence score of the scenario class.

[0172] In this embodiment of the invention, the server combines the first intermediate value, the second intermediate value, and the third type of problem vector to calculate the scenario class confidence score.

[0173] In the embodiment of the present application, the server calculates the first sum of the first intermediate value and the second intermediate value, calculates the first reciprocal of the second sum of the problem category third vector and the normalized vector of the problem category third vector, multiplies the first reciprocal with the first sum to obtain the scenario category confidence.

[0174] In the embodiment of the present application, the server multiplies the transaction category second vector with the transaction code, the predetermined parameter and the first matrix to obtain the third intermediate value, and multiplies the transaction category difference value converted from the transaction category second vector with the transaction category second vector, the transaction code and the predetermined parameter to obtain the fourth intermediate value.

[0175] In the embodiment of the present application, the server multiplies the transaction category second vector with the transaction code, the predetermined parameter and the first matrix to obtain the third intermediate value, and multiplies the transaction category difference value converted from the transaction category second vector with the transaction category second vector, the transaction code and the predetermined parameter to obtain the fourth intermediate value.

[0176] In the embodiment of the present application, the server can calculate the modulus of the transaction category second vector to obtain the transaction category difference value.

[0177] In the embodiment of the present application, the server combines the third intermediate value, the fourth intermediate value and the problem category third vector to calculate the transaction category confidence.

[0178] In the embodiment of the present application, the server combines the third intermediate value, the fourth intermediate value and the problem category third vector to calculate the transaction category confidence.

[0179] In the embodiment of the present application, the server multiplies the problem category third vector with the problem code, the predetermined parameter and the first matrix to obtain the fifth intermediate value, and multiplies the problem category difference value converted from the problem category third vector with the problem category third vector, the problem code and the predetermined parameter to obtain the sixth intermediate value.

[0180] In the embodiment of the present application, the server multiplies the problem category third vector with the problem code, the predetermined parameter and the first matrix to obtain the fifth intermediate value, and multiplies the problem category difference value converted from the problem category third vector with the problem category third vector, the problem code and the predetermined parameter to obtain the sixth intermediate value.

[0181] In the embodiment of the present application, the server can calculate the modulus of the problem category third vector to obtain the problem category difference value.

[0182] In the embodiment of the present application, the server combines the fifth intermediate value, the sixth intermediate value and the problem category third vector to calculate the problem category confidence.

[0183] In the embodiment of the present application, the server combines the fifth intermediate value, the sixth intermediate value and the problem category third vector to calculate the problem category confidence.

[0184] S118, determine the minimum confidence among the scene type confidence, the matter type confidence and the question type confidence as the minimum confidence.

[0185] In the embodiment of the application, the server determines the minimum confidence among the scene type confidence, the matter type confidence and the question type confidence as the minimum confidence.

[0186] For example, the server can calculate the minimum confidence B corresponding to each set of pre-training data by formula (1).

[0187] (1)

[0188] wherein M is the question type third vector, is a normalized vector of the question type third vector, L is a first matrix corresponding to the preset training data, is the scene type first vector or the matter type second vector or the question type third vector, is the scene code or the matter code or the question code, is a predetermined parameter, is a scene type difference value. Wherein, indicates that the server multiplies the scene type first vector by the first matrix L, the scene code to obtain a first intermediate value corresponding to the scene type, after determining the scene type confidence, the server multiplies the matter type second vector by the first matrix, the matter code to calculate a first intermediate value of the matter type, and then calculates the matter type confidence, until the question type confidence is calculated. Wherein, indicates that the server multiplies the scene type first vector converted scene type difference value by the scene type first vector , the scene code and the predetermined parameter to obtain a second intermediate value, after determining the scene type confidence, the server multiplies the matter type second vector converted scene type difference value by the matter type second vector , the matter code and the predetermined parameter to obtain a second intermediate value, and then calculate the matter type confidence, until the question type confidence is calculated.

[0189] In the embodiment of the application, the server determines the minimum confidence through the scene type first vector, the matter type second vector and the question type third vector in a layer-by-layer recursive manner, and the minimum confidence determined in a more fine-grained manner is more accurate, so the target appeal reply information determined is also more accurate.

[0190] In some embodiments, referring to Figure 7 , Figure 7 An optional flowchart of the prescription information processing method provided by the embodiments of the present application is shown in Figure 6 S113 shown can be implemented through S119 to S121, which will be described in combination with the steps.

[0191] S119, calculating a first sum of the first intermediate value and the second intermediate value.

[0192] In the embodiments of the present application, the server calculates a first sum of the first intermediate value and the second intermediate value.

[0193] S120, calculating a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0194] In the embodiments of the present application, the server calculates a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0195] S121, multiplying the first reciprocal by the first sum to obtain the scenario-type confidence.

[0196] In the embodiments of the present application, the server calculates a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0197] In some embodiments, referring to Figure 8 , Figure 8 An optional flowchart of the prescription information processing method provided by the embodiments of the present application is shown in Figure 6 S115 shown can be implemented through S122 to S124, which will be described in combination with the steps.

[0198] S122, calculating a third sum of the third intermediate value and the fourth intermediate value.

[0199] In the embodiments of the present application, the server calculates a third sum of the third intermediate value and the fourth intermediate value.

[0200] S123, calculating a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0201] In the embodiments of the present application, the server calculates a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0202] S124, multiplying the first reciprocal by the third sum to obtain the matter-type confidence.

[0203] In the embodiments of the present application, the server calculates a first reciprocal of a second sum of the question-type third vector and the normalized vector of the question-type third vector.

[0204] In some embodiments, referring to Figure 9 , Figure 9 An optional flowchart of the prescription information processing method provided by the embodiment of the present application is shown in Figure 6 S117 shown can be implemented through S125 to S127, which will be described in combination with each step.

[0205] S125, calculating the fourth sum of the fifth intermediate value and the sixth intermediate value.

[0206] In the embodiment of the present application, the server calculates the fourth sum of the fifth intermediate value and the sixth intermediate value.

[0207] S126, taking the first inverse of the second sum of the problem type third vector and the normalized vector of the problem type third vector.

[0208] In the embodiment of the present application, the server takes the first inverse of the second sum of the problem type third vector and the normalized vector of the problem type third vector.

[0209] S127, multiplying the first inverse and the fourth sum to obtain the problem type confidence.

[0210] In the embodiment of the present application, the server multiplies the first inverse and the fourth sum to obtain the problem type confidence.

[0211] In some embodiments, referring to Figure 10 , Figure 10 An optional flowchart of the prescription information processing method provided by the embodiment of the present application is shown in Figure 5 S108 shown can be implemented through S128 to S129, which will be described in combination with each step.

[0212] S128, determining a target confidence greater than a confidence threshold from a plurality of minimum confidences corresponding to a plurality of pre-training data sets.

[0213] In the embodiment of the present application, the server determines a target confidence greater than a confidence threshold from a plurality of minimum confidences corresponding to a plurality of pre-training data sets.

[0214] S129, extracting target appeal reply information from pre-training data corresponding to the target confidence.

[0215] In the embodiment of the present application, the server extracts target appeal reply information from pre-training data corresponding to the target confidence.

[0216] In some embodiments, referring to Figure 11 , Figure 11 An optional flowchart of the prescription information processing method provided by the embodiment of the present application is shown in Figure 1S101 shown can be implemented through S130 to S132, which will be described in combination with each step.

[0217] S130, extracting the identification information corresponding to the current customer.

[0218] In the embodiment of the application, the server extracts the identification information corresponding to the current customer.

[0219] S131, extracting the multiple click behavior information and the corresponding form data of the current customer in the first historical time period through the identification information in the local database.

[0220] In the embodiment of the application, the server extracts the multiple click behavior information and the corresponding form data of the current customer in the first historical time period through the identification information in the local database.

[0221] In the embodiment of the application, the local database stores the mapping relationship between the identification information of multiple customers and the click behavior information and the corresponding form data in the corresponding historical time period. Therefore, the server can extract the multiple click behavior information and the corresponding form data of the current customer in the first historical time period according to the identification information of the current customer.

[0222] S132, inputting the multiple click behavior information and the corresponding form data into the prediction model to obtain multiple sets of consultation item data corresponding to the multiple click behavior information.

[0223] In the embodiment of the application, the server inputs the multiple click behavior information and the corresponding form data into the prediction model to obtain multiple sets of consultation item data corresponding to the multiple click behavior information. In the embodiment of the application, the scene code, the item code and the question code corresponding to each set of consultation item data can also be obtained.

[0224] In the embodiment of the application, the server inputs the multiple click behavior information and the corresponding form data into the multi-view bidirectional feedback prediction model. For example, in combination with Figure 12 will be described in combination with steps.

[0225] S220, form data.

[0226] S221, click behavior information.

[0227] S222, encoding.

[0228] S223, feature encoding.

[0229] S224, feature cross.

[0230] S225, residual network.

[0231] S226, forward update module.

[0232] S227, optimal label criterion.

[0233] S228, scenario.

[0234] S229, matter.

[0235] S230, question.

[0236] S231, scenario supplement.

[0237] S232, matter supplement.

[0238] S233, question supplement.

[0239] In the embodiment of the application, the server encodes the form data, the server encodes the click behavior information, and the scenario encoding, the matter encoding and the question encoding are obtained through feature cross processing, residual network, forward update module and optimal label criterion processing, and the scenario supplement information, the matter supplement information and the question supplement information are obtained through the multi-view bidirectional feedback prediction model processing.

[0240] In some embodiments, referring to Figure 13 , Figure 13 An optional flowchart of a prescription information processing method provided by the embodiment of the application is shown in Figure 1 S102 shown in the figure can be implemented by S133 to S135, which will be described in combination with each step.

[0241] S133, in the user historical data set, extract the first target scenario class, the first target matter class and the first target question class information which are the same as the scenario class, the matter class and the question class information in each group of consultation item data.

[0242] In the embodiment of the application, the server extracts the first target scenario class, the first target matter class and the first target question class information which are the same as the scenario class, the matter class and the question class information in each group of consultation item data in the user historical data set.

[0243] In the embodiment of the application, the server traverses the multiple first information groups in the user historical data set, and extracts the first target scenario class, the first target matter class and the first target question class information which are the same as the scenario class, the matter class and the question class information in each group of consultation item data.

[0244] S134, in the appeal historical data set, extract the second target scenario class, the second target matter class and the second target question class information which are the same as the scenario class, the matter class and the question class information in each group of consultation item data, and the appeal reply information included in the second target information group to which the second target scenario class, the second target matter class and the second target question class information belong.

[0245] In the embodiment of the present application, the server extracts the second target scene class, the second target matter class and the second target question class information same as the scene class, the matter class and the question class information in each set of consultation item data from the appeal history data set, and the appeal reply information included in the second target information group to which the second target scene class, the second target matter class and the second target question class information belong.

[0246] In the embodiment of the present application, the server traverses the multiple second information groups of the appeal history data set, extracts the second target scene class, the second target matter class and the second target question class information same as the scene class, the matter class and the question class information in each set of consultation item data, and determines the second target information group to which the second target scene class, the second target matter class and the second target question class information belong. The server extracts the appeal reply information included in the second target information group.

[0247] S135, merging and unifying each set of consultation item data with the first target scene class, the first target matter class, the first target question class information, the second target scene class, the second target matter class and the second target question class information to obtain third scene class, third matter class and third question class information.

[0248] In the embodiment of the present application, the server merges and unifies each set of consultation item data with the first target scene class, the first target matter class, the first target question class information, the second target scene class, the second target matter class and the second target question class information to obtain third scene class, third matter class and third question class information.

[0249] The consultation item information includes the third scene class, the third matter class and the third question class information.

[0250] S136, forming each set of pre-training data using the third scene class, the third matter class and the third question class information, and adding the appeal reply information included in the second target information group into the corresponding set of pre-training data to obtain multiple sets of pre-training data.

[0251] In the embodiment of the present application, the server forms each set of pre-training data corresponding to each set of consultation item data using the third scene class, the third matter class and the third question class information, and adds the appeal reply information included in the second target information group into the corresponding set of pre-training data to obtain multiple sets of pre-training data.

[0252] In some embodiments, referring to Figure 14 , Figure 14 An optional flowchart of the prescription information processing method provided by the embodiment of the present application is shown in Figure 1 S101 before the embodiment of the present application is shown, and S137 to S139 are further implemented, which will be described in combination with each step.

[0253] S137, obtain current appeal conversation information of the current customer.

[0254] In the embodiment of the application, the server obtains the current appeal conversation information of the current customer.

[0255] The current appeal conversation information can be text communication information of the current customer and the customer service personnel, or voice communication information of the current customer and the customer service personnel.

[0256] S138, perform sentiment analysis on the appeal conversation information to obtain an emotion value.

[0257] In the embodiment of the application, the server performs sentiment analysis on the appeal conversation information to obtain an emotion value.

[0258] In the embodiment of the application, the server can input the appeal conversation information into a sentiment analysis model to obtain an emotion value corresponding to the appeal conversation information.

[0259] S139, if it is detected that the emotion value exceeds a negative emotion threshold value, extract historical behavior related data of the current customer.

[0260] In the embodiment of the application, if the server detects that the emotion value exceeds the negative emotion threshold value, the historical behavior related data of the current customer is extracted, that is, the step of S101 is performed, and a plurality of sets of consultation item data of the current customer are predicted by using the extracted historical behavior related data of the current customer.

[0261] In the embodiment of the application, if the server detects that the emotion value does not exceed the negative emotion threshold value, the step of S101 is not performed.

[0262] In the embodiment of the application, after the server recursively obtains the target appeal reply information of the current customer by sequentially performing weakly supervised contrast learning on the plurality of sets of pre-training data, the server pushes the target appeal reply information to the current customer terminal corresponding to the current customer.

[0263] Please refer to Figure 15 The structure of the appeal processing device provided in the embodiment of the application is shown in the figure.

[0264] The embodiment of the application further provides an appeal processing device 800, which comprises a prediction unit 803 and a processing unit 804.

[0265] The prediction unit is configured to predict a plurality of sets of consultation item data of the current customer by using the extracted historical behavior related data of the current customer, wherein the different sets of consultation item data comprise different scene types, matter types and question types predicted for the current customer.

[0266] The processing unit is configured to extract a plurality of pre-training data sets from the plurality of consultation item data sets and a preset user history data set and a preset appeal history data set, wherein each pre-training data set comprises consultation item information identical to the scene class information, the item class information and the question class information, and appeal reply information corresponding to the consultation item information; the user history data set comprises a plurality of first information groups, each first information group comprising first scene class information, first item class information and first question class information; the appeal history data set comprises a plurality of second information groups, each second information group comprising second scene class information, second item class information, second question class information and corresponding appeal reply information; and the processing unit is configured to sequentially perform weakly supervised contrast learning on the plurality of pre-training data sets to recursively obtain the target appeal reply information of the current customer.

[0267] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to sequentially perform vectorization processing on the plurality of pre-training data sets to obtain a scene class vector corresponding to the scene class information, an item class vector corresponding to the item class information and a question class vector corresponding to the question class information in each pre-training data set; and obtain a first matrix corresponding to each pre-training data set by combining the scene class vector, the item class vector and the question class vector.

[0268] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to obtain a scene class first vector, an item class second vector and a question class third vector corresponding to each pre-training data set by combining a weakly supervised contrast learning model and pre-obtained scene encoding, item encoding and question encoding through the first matrix; calculate the minimum confidence of each pre-training data set by combining the scene class first vector, the item class second vector and the question class third vector with parameter information of the weakly supervised contrast learning model, and further obtain the minimum confidence of the plurality of pre-training data sets; and determine the target appeal reply information for the current customer based on the minimum confidence and the plurality of pre-training data sets.

[0269] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is used to multiply the first matrix, the square of the scene encoding, a preset first intermediate value, the scene class vector and a predetermined parameter of the weakly supervised contrast learning model to obtain the scene class first vector; the preset first intermediate value is a probability value equal to the product of the scene class vector and the calculation matrix of the weakly supervised contrast learning model and the scene class vector; multiply the first matrix, the square of the item encoding, a preset second intermediate value, the scene class first vector and the predetermined parameter of the weakly supervised contrast learning model to obtain the item class second vector; the preset second intermediate value is a probability value equal to the product of the scene class first vector and the calculation matrix of the weakly supervised contrast learning model and the scene class first vector; multiply the first matrix, the square of the item encoding, a preset third intermediate value, the item class second vector and the predetermined parameter of the weakly supervised contrast learning model to obtain the problem class third vector; the preset third intermediate value is a probability value equal to the product of the item class second vector and the calculation matrix of the weakly supervised contrast learning model and the item class second vector.

[0270] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is used to multiply the scene class first vector corresponding to each set of pre-training data, the scene encoding, the predetermined parameter and the first matrix to obtain a first intermediate value, multiply the scene class difference value converted from the scene class first vector, the scene class first vector, the scene encoding and the predetermined parameter to obtain a second intermediate value; combine the first intermediate value, the second intermediate value and the problem class third vector to calculate the scene class confidence; multiply the item class second vector, the item encoding, the predetermined parameter and the first matrix to obtain a third intermediate value, multiply the item class difference value converted from the item class second vector, the item class second vector, the item encoding and the predetermined parameter to obtain a fourth intermediate value; combine the third intermediate value, the fourth intermediate value and the problem class third vector to calculate the item class confidence; multiply the problem class third vector, the problem encoding, the predetermined parameter and the first matrix to obtain a fifth intermediate value, multiply the problem class difference value converted from the problem class third vector, the problem class third vector, the problem encoding and the predetermined parameter to obtain a sixth intermediate value; combine the fifth intermediate value, the sixth intermediate value and the problem class third vector to calculate the problem class confidence; determine the minimum confidence among the scene class confidence, the item class confidence and the problem class confidence as the minimum confidence.

[0271] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to calculate a first sum of the first intermediate value and the second intermediate value; calculate a first reciprocal of a second sum of the problem class third vector and the normalized vector of the problem class third vector; multiply the first reciprocal and the first sum to obtain the scene class confidence.

[0272] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to calculate a third sum of the third intermediate value and the fourth intermediate value; calculate a first reciprocal of a second sum of the problem class third vector and the normalized vector of the problem class third vector; multiply the first reciprocal and the third sum to obtain the matter class confidence.

[0273] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to calculate a fourth sum of the fifth intermediate value and the sixth intermediate value; calculate a first reciprocal of a second sum of the problem class third vector and the normalized vector of the problem class third vector; multiply the first reciprocal and the fourth sum to obtain the problem class confidence.

[0274] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to determine a target confidence greater than a confidence threshold value from the plurality of minimum confidences corresponding to the plurality of pre-training data; and extract the target appeal reply information from the pre-training data corresponding to the target confidence.

[0275] In the embodiment of the present application, the prediction unit 803 in the appeal processing device 800 is configured to extract identification information corresponding to the current customer; extract a plurality of click behavior information and corresponding form data of the current customer in a first historical time period from a local database through the identification information; the historical behavior related data includes the plurality of click behavior information and the corresponding form data; input the plurality of click behavior information and the corresponding form data into a prediction model to obtain the plurality of sets of consultation item data corresponding to the plurality of click behavior information.

[0276] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to extract, from the user historical data set, first target scene class information, first target matter class information and first target problem class information identical to the scene class information, the matter class information and the problem class information in each set of consultation item data; extract, from the appeal historical data set, second target scene class information, second target matter class information and second target problem class information identical to the scene class information, the matter class information and the problem class information in each set of consultation item data, and the appeal reply information included in a second target information group to which the second target scene class information, the second target matter class information and the second target problem class information belong; merge and unify each set of consultation item data with the first target scene class information, the first target matter class information, the first target problem class information, the second target scene class information, the second target matter class information and the second target problem class information, to obtain third scene class information, third matter class information and third problem class information; form each set of pre-training data by using the third scene class information, the third matter class information and the third problem class information, and add the appeal reply information included in the second target information group into the corresponding set of pre-training data, to obtain a plurality of sets of pre-training data; and the consultation item information includes the third scene class information, the third matter class information and the third problem class information.

[0277] In the embodiment of the present application, the processing unit 804 in the appeal processing device 800 is configured to obtain current appeal session information of the current customer; perform emotion analysis on the appeal session information to obtain an emotion value; and if it is detected that the emotion value exceeds a negative emotion threshold value, extract the historical behavior related data of the current customer.

[0278] In the embodiment of the present application, the processing unit 804 is configured to push the target appeal reply information to a current client corresponding to the current customer.

[0279] In the embodiment of the present application, the prediction unit 803 is configured to predict a plurality of sets of consultation item data of the current customer by extracting the historical behavior related data corresponding to the current customer; the different sets of consultation item data include: different scene type, matter type and question type information predicted corresponding to the current customer; the processing unit 804 is configured to obtain a plurality of sets of pre-training data by the plurality of sets of consultation item data and a preset user historical data set and a claim historical data set; wherein each set of pre-training data includes: consultation item information same as the scene type, the matter type and the question type information, and claim reply information corresponding to the consultation item information; the processing unit 804 is configured to recursively derive the target claim reply information of the current customer by sequentially performing weakly supervised contrast learning on the plurality of sets of pre-training data. Since the present scheme ignores the current session information of the customer, and directly determines the target claim reply information by more fine-grained claim analysis of multiple levels including scene type, matter type and question type, the soothing information matched with the customer can be accurately provided.

[0280] It should be noted that if the above-mentioned claim processing method is implemented in the form of a software function module and sold or used as an independent product in the embodiment of the present application, it can also be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a claim processing device (which can be a personal computer, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read only memory (Read Only Memory, ROM), a magnetic disk or an optical disk, and various media that can store program codes. Thus, the embodiments of the present application are not limited to any specific hardware and software combination.

[0281] Correspondingly, the embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the above method.

[0282] Correspondingly, the embodiment of the present application provides a claim processing device, which includes a memory 802 and a processor 801, the memory 802 stores a computer program executable on the processor 801, and the processor 801 implements the steps in the above method when executing the program.

[0283] It should be noted here that the description of the above storage medium and device embodiments is similar to the description of the above method embodiments, and has similar beneficial effects to the method embodiments. For technical details not disclosed in the storage medium and device embodiments of the present application, please refer to the description of the method embodiments of the present application for understanding.

[0284] It should be noted that, Figure 16 A hardware entity schematic diagram of the claim processing apparatus provided by the embodiment of the present application is shown in Figure 16 The hardware entity of the claim processing apparatus 800 includes a processor 801 and a memory 802, wherein;

[0285] The processor 801 generally controls the overall operation of the claim processing apparatus 800.

[0286] The memory 802 is configured to store instructions and applications executable by the processor 801, and can also cache data (for example, image data, audio data, voice communication data and video communication data) to be processed by the processor 801 and modules in the claim processing apparatus 800, which can be implemented by FLASH or Random Access Memory (RAM).

[0287] It should be understood that the term "one embodiment" or "an embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present application, the size of the sequence number of each process does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the present application. The sequence number of the above-mentioned embodiment of the present application is only for description, not representing the advantages and disadvantages of the embodiment.

[0288] It should be noted that in this paper, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0289] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented by other manners. The apparatus embodiments described above are merely illustrative, for example, the division of the units is merely a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed components can be indirect coupling or communication connection through some interfaces, apparatuses or units, which can be electrical, mechanical or other forms.

[0290] The units described above as separate components can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place or distributed on a plurality of network units; and part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0291] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware or hardware plus software functional unit.

[0292] Those of ordinary skill in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction-related hardware, and the foregoing program can be stored in a computer-readable storage medium, and the program executes the steps including the above-mentioned method embodiments when executed; and the foregoing storage medium includes mobile storage devices, read-only memories (ROM), magnetic discs or optical discs and various storage media that can store program codes.

[0293] Alternatively, the integrated unit of the present application, if implemented in the form of a software functional module and sold or used as an independent product, can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes mobile storage devices, ROM, magnetic discs or optical discs and various storage media that can store program codes.

[0294] The above merely describes the embodiments of the present application, and the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for handling complaints, characterized in that, include: By extracting relevant historical behavior data corresponding to the current customer, multiple sets of consultation items data for the current customer are predicted; Different groups of consultation item data include: different scenario categories, item categories, and question categories of information corresponding to the current customer prediction; Multiple sets of pre-training data are extracted from the multiple sets of consultation item data and the preset user history data set and request history data set; wherein, each set of pre-training data includes: consultation item information that is the same as the scenario class, the matter class and the question class information, and the request response information corresponding to the consultation item information; the user history data set includes: multiple first information groups; each first information group includes: first scenario class, first matter class and first question class information; the request history data set includes: multiple second information groups, each second information group includes: second scenario class, second matter class, second question class information and the corresponding request response information; The multiple sets of pre-trained data are sequentially input into the weakly supervised contrastive learning model to obtain the confidence level corresponding to each set of pre-trained data. Among the multiple confidence levels corresponding to the multiple sets of pre-trained data, a target confidence level greater than the confidence level threshold is determined. The request response information is then extracted from the pre-trained data corresponding to the target confidence level.

2. The method for handling requests according to claim 1, characterized in that, The method further includes, before extracting the request response information from the pre-training data corresponding to the multiple sets of pre-training data sequentially inputting the multiple sets of pre-training data into a weakly supervised contrastive learning model to obtain the confidence level corresponding to each set of pre-training data, determining a target confidence level greater than a confidence threshold from the multiple confidence levels corresponding to the multiple sets of pre-training data, and the method itself. The multiple sets of pre-training data are sequentially vectorized to obtain scene class vectors for the scene class, item class vectors for the item class, and question class vectors for the question information in each set of pre-training data. By combining the scenario class vector, the event class vector, and the question class vector, a first matrix corresponding to each set of pre-trained data is obtained.

3. The method for handling requests according to claim 2, characterized in that, The process involves sequentially inputting the multiple sets of pre-trained data into a weakly supervised contrastive learning model to obtain the confidence level corresponding to each set of pre-trained data. A target confidence level greater than a confidence threshold is determined from the multiple confidence levels corresponding to the multiple sets of pre-trained data. The request response information is then extracted from the pre-trained data corresponding to the target confidence level, including: Using the first matrix, combined with the weakly supervised contrastive learning model and the pre-obtained scene codes, event codes, and question codes, the first vector of the scene class, the second vector of the event class, and the third vector of the question class corresponding to each set of pre-trained data are obtained. By combining the first vector of the scenario class, the second vector of the event class, and the third vector of the problem class with the parameter information of the weakly supervised contrastive learning model, the minimum confidence of each set of pre-trained data is calculated, thereby obtaining the minimum confidence of the multiple sets of pre-trained data; Based on the minimum confidence level and the multiple sets of pre-trained data, the response information for the target demand of the current customer is determined.

4. The method for handling complaints according to claim 3, characterized in that, The first matrix, combined with a weakly supervised contrastive learning model and pre-obtained scene codes, event codes, and question codes, yields a first vector for the scene class, a second vector for the event class, and a third vector for the question class corresponding to each set of pre-trained data, including: The first matrix is ​​multiplied by the square of the scene code, a preset first intermediate value, the scene class vector, and the predetermined parameters of the weakly supervised contrastive learning model to obtain the scene class first vector; the preset first intermediate value is the probability value that the scene class vector is equal to the product of the calculation matrix of the weakly supervised contrastive learning model and the scene class vector. The first matrix is ​​multiplied by the square of the item code, the preset second intermediate value, the first vector of the scene class, and the predetermined parameters of the weakly supervised contrastive learning model to obtain the second vector of the item class; the preset second intermediate value is the probability value that the first vector of the scene class is equal to the product of the calculation matrix of the weakly supervised contrastive learning model and the first vector of the scene class. The first matrix is ​​multiplied by the square of the item code, the preset third intermediate value, the second vector of the item class, and the predetermined parameters of the weakly supervised contrastive learning model to obtain the third vector of the issue class; the preset third intermediate value is the probability value that the product of the second vector of the item class and the calculation matrix of the weakly supervised contrastive learning model and the second vector of the item class is equal.

5. The method for handling requests according to claim 3, characterized in that, The step of combining the first vector of the scenario class, the second vector of the event class, and the third vector of the problem class with the parameter information of the weakly supervised contrastive learning model to calculate the minimum confidence of each set of pre-trained data, and then obtaining the minimum confidence of the multiple sets of pre-trained data, includes: The scene class first vector corresponding to each group of pre-trained data is multiplied by the scene code, the predetermined parameters and the first matrix to obtain a first intermediate value. The scene class difference value transformed from the scene class first vector is multiplied by the scene class first vector, the scene code and the predetermined parameters to obtain a second intermediate value. The scenario class confidence score is calculated by combining the first intermediate value, the second intermediate value, and the third vector of the problem class. Multiply the second vector of the item class with the item code, the predetermined parameter and the first matrix to obtain a third intermediate value; multiply the item class difference value transformed from the second vector of the item class with the second vector of the item class, the item code and the predetermined parameter to obtain a fourth intermediate value. The confidence level of the event class is calculated by combining the third intermediate value, the fourth intermediate value, and the third vector of the issue class. Multiply the third vector of the problem class by the problem code, the predetermined parameter and the first matrix to obtain a fifth intermediate value; multiply the problem class difference value transformed from the third vector of the problem class by the third vector of the problem class, the problem code and the predetermined parameter to obtain a sixth intermediate value; The confidence level of the problem class is calculated by combining the fifth intermediate value, the sixth intermediate value, and the third vector of the problem class. The minimum confidence level among the scenario-based confidence level, the event-based confidence level, and the problem-based confidence level is determined as the minimum confidence level.

6. The method for handling requests according to claim 5, characterized in that, The scenario class confidence score is calculated by combining the first intermediate value, the second intermediate value, and the third vector of the problem class, including: Calculate the first sum of the first intermediate value and the second intermediate value; For the third vector of the problem class, find the first reciprocal of the second sum of the normalized vector of the third vector of the problem class; Multiply the first reciprocal by the first sum to obtain the confidence score of the scenario class.

7. The method for handling requests according to claim 5, characterized in that, The process of combining the third intermediate value, the fourth intermediate value, and the third vector of the question class to calculate the confidence score of the event class includes: Calculate the third sum of the third intermediate value and the fourth intermediate value; For the third vector of the problem class, find the first reciprocal of the second sum of the normalized vector of the third vector of the problem class; Multiply the first reciprocal by the third sum to obtain the confidence level of the event class.

8. The method for handling complaints according to claim 5, characterized in that, The calculation of the question class confidence score by combining the fifth intermediate value, the sixth intermediate value, and the third vector of the question class includes: Calculate the fourth sum of the fifth intermediate value and the sixth intermediate value; For the third vector of the problem class, find the first reciprocal of the second sum of the normalized vector of the third vector of the problem class; Multiply the first reciprocal by the fourth sum to obtain the confidence level of the problem class.

9. The method for handling complaints according to claim 3, characterized in that, The step of determining the response information for the current customer's target request based on the minimum confidence level and the multiple sets of pre-trained data includes: Among the multiple minimum confidence levels corresponding to the multiple sets of pre-trained data, a target confidence level greater than the confidence threshold is determined; The target request response information is extracted from the pre-training data corresponding to the target confidence level.

10. The method for handling complaints according to claim 1, characterized in that, Before predicting multiple sets of consultation items for the current customer by extracting relevant historical behavior data corresponding to the current customer, the method further includes: Extract the identification information corresponding to the current customer; Accordingly, by extracting relevant historical behavior data corresponding to the current customer, multiple sets of consultation item data for the current customer are predicted, including: Using the identification information, multiple click behavior information and their corresponding form data within the first historical time period of the current customer are extracted from the local database; the historical behavior related data includes: the multiple click behavior information and their corresponding form data; The multiple click behavior information and their corresponding form data are input into the prediction model to obtain the multiple sets of consultation item data corresponding to the multiple click behavior information.

11. The method for handling complaints according to claim 1, characterized in that, The process involves extracting multiple sets of pre-trained data from the multiple sets of consultation data and preset sets of user historical data and request historical data, including: In the user's historical data set, extract the first target scenario class, first target matter class, and first target question class information that are identical to the scenario class, matter class, and question class information in each group of consultation item data; In the historical data set of the requests, extract the second target scenario class, second target matter class, and second target problem class information that are the same as the scenario class, matter class, and problem class information in each group of consultation item data, and the request response information included in the second target information group to which the second target scenario class, second target matter, and second target problem class information belong; The consultation item data of each group is merged and unified with the information of the first target scenario class, the first target matter class, the first target question class, the second target scenario class, the second target matter class, and the second target question class to obtain the information of the third scenario class, the third matter class, and the third question class; the consultation item information includes: the information of the third scenario class, the third matter class, and the third question class; Each set of pre-training data is formed using the information of the third scenario class, the third matter class, and the third question class, and the request response information included in the second target information group is added to the corresponding set of pre-training data to obtain the multiple sets of pre-training data.

12. The method for handling complaints according to claim 1, characterized in that, Before predicting multiple sets of consultation items for the current customer by extracting relevant historical behavior data corresponding to the current customer, the method further includes: Obtain the current customer's current request session information; Sentiment analysis is performed on the aforementioned request conversation information to obtain a sentiment value; If the detected emotion value exceeds the negative emotion threshold, then the historical behavior-related data of the current customer is extracted.

13. The method for handling complaints according to claim 1, characterized in that, The method further includes: sequentially inputting the multiple sets of pre-trained data into a weakly supervised contrastive learning model to obtain the confidence level corresponding to each set of pre-trained data; determining a target confidence level greater than a confidence threshold among the multiple confidence levels corresponding to the multiple sets of pre-trained data; and extracting the request response information from the pre-trained data corresponding to the target confidence level. The response information regarding the target request is pushed to the current client corresponding to the current customer.

14. A complaint processing device, characterized in that, include: The prediction unit is used to predict multiple sets of consultation items for the current customer by extracting relevant historical behavior data corresponding to the current customer. Different groups of consultation item data include: different scenario categories, item categories, and question categories of information corresponding to the current customer prediction; The processing unit is configured to extract multiple sets of pre-trained data from the multiple sets of consultation item data and a preset set of user historical data and a set of request historical data; wherein each set of pre-trained data includes: consultation item information that is the same as the scenario class, the matter class, and the question class information, and request response information corresponding to the consultation item information; the user historical data set includes: multiple first information groups; each first information group includes: first scenario class, first matter class, and first question class information; the request historical data set includes: multiple second information groups, each second information group includes: second scenario class, second matter class, second question class information, and the corresponding request response information; The processing unit is configured to sequentially input the multiple sets of pre-trained data into a weakly supervised contrastive learning model to obtain the confidence level corresponding to each set of pre-trained data, determine a target confidence level greater than a confidence level threshold among the multiple confidence levels corresponding to the multiple sets of pre-trained data, and extract the request response information from the pre-trained data corresponding to the target confidence level.

15. A complaint processing device, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, the processor executing the program to implement the steps of the method according to any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 13.

Images