Customer service answering method and device

Abstract

The present invention discloses a customer service response method and device, and relates to the field of computer technology. A specific implementation of the method includes: obtaining a statement to be answered, and based on a multidimensional intent recognition model, identifying a first multidimensional classification result corresponding to the statement to be answered; obtaining a slot-filling statement corresponding to the statement to be answered, and determining a target multidimensional classification result corresponding to the statement to be answered according to the first multidimensional classification result and the slot-filling statement; based on a mapping relationship between the classification result and the scene, generating a response statement corresponding to the statement to be answered according to the target multidimensional classification result. This implementation can simplify the system architecture, reduce the cost of model training, improve the reusability of the model, and improve the accuracy of the intent recognition results and the accuracy of the response.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a customer service answering method and device. Background Art

[0002] Among the application scenarios of intelligent response systems, the most widely used scenario is intelligent customer service, which establishes a fast and effective technical means based on natural language for communication between enterprises and massive users. In the intelligent customer service response method, the intention recognition model is used to identify the user's intention, and the identified user intention is used as the condition for entering the response scenario. In the response scenario, the slot is filled in according to the scenario requirements. After the user answers the question, the intention recognition model is used to re-identify the scene and complete the slot filling, and finally the response answer is obtained.

[0003] However, the existing intent recognition model is a multi-layer model. Due to the diversification of user needs, the number of layers and models at each layer is increasing, resulting in high system complexity and maintenance difficulty, low model reusability and high training costs. In addition, during the slot filling and rhetorical questioning process, due to the different ways of expressing the user's answers, it is easy to switch scenes, reducing the accuracy of the answer. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a customer service response method and device, which can simplify the system architecture, reduce the cost of model training, improve the reusability of the model, and improve the accuracy of intent recognition results and the accuracy of response.

[0005] To achieve the above objective, according to a first aspect of an embodiment of the present invention, a customer service answering method is provided.

[0006] A customer service response method according to an embodiment of the present invention includes: obtaining a statement to be answered, and identifying a first multidimensional classification result corresponding to the statement to be answered based on a multidimensional intent recognition model; obtaining a slot-filling statement corresponding to the statement to be answered, and determining a target multidimensional classification result corresponding to the statement to be answered based on the first multidimensional classification result and the slot-filling statement; and generating a response statement corresponding to the statement to be answered based on the target multidimensional classification result based on a mapping relationship between the classification result and the scene.

[0007] Optionally, the method also includes: obtaining historical response logs, dimensionally annotating the historical response logs according to preset dimensional attributes, and obtaining multidimensional annotation results; based on a deep learning model, according to the multidimensional annotation results, training a sub-intention recognition model corresponding to the dimensional attributes, and using the sub-intention recognition models corresponding to the dimensional attributes to construct the multidimensional intent recognition model; wherein the input of the multidimensional intent recognition model is a sentence, and the output is a classification result of the sentence on the dimensional attributes.

[0008] Optionally, obtaining the slot-filling statement corresponding to the statement to be answered, and determining the target multidimensional classification result corresponding to the statement to be answered according to the first multidimensional classification result and the slot-filling statement, includes: judging whether scene slot-filling is needed; if so, obtaining the slot-filling statement, using the multidimensional intent recognition model to identify the second multidimensional classification result corresponding to the slot-filling statement, and determining the target multidimensional classification result according to the first multidimensional classification result and the second multidimensional classification result; if not, directly determining the first multidimensional classification result as the target multidimensional classification result.

[0009] Optionally, determining the target multidimensional classification result based on the first multidimensional classification result and the second multidimensional classification result includes: judging whether the first multidimensional classification result is the same as the second multidimensional classification result; if so, directly determining the first multidimensional classification result as the target multidimensional classification result; if not, obtaining the difference classification results and the same classification results in the second multidimensional classification result, calculating the correlation weights between the difference classification results and the same classification results, and determining the target multidimensional classification result according to the correlation weights according to a preset intention inheritance rule.

[0010] Optionally, the calculation of the correlation weight between the difference classification result and the same classification result includes at least one of the following options: according to the historical response log, counting the first number of occurrences of the difference classification result, the second number of occurrences of the same classification result, and the third number of occurrences of the difference classification result and the same classification result at the same time, and then calculating the correlation weight between the difference classification result and the same classification result according to the first number, the second number and the third number; according to the historical response log, calculating the correlation weight table, and then querying the correlation weight between the difference classification result and the same classification result from the correlation weight table.

[0011] Optionally, the target multidimensional classification result is determined according to the relevance weight in accordance with the preset intention inheritance rule, including: judging whether the relevance weight is greater than the preset weight; if so, determining that the second multidimensional classification result is the target multidimensional classification result; if not, de-entity generalizing the slot-filling statement according to the difference classification result, and then identifying it using the multidimensional intention recognition model to obtain a third multidimensional classification result, and determining the target classification result according to the third multidimensional classification result.

[0012] Optionally, based on the mapping relationship between the classification results and the scenes, a response statement corresponding to the statement to be responded to is generated according to the target multi-dimensional classification result, including: based on the mapping relationship between the classification results and the scenes, matching the target scene corresponding to the target multi-dimensional classification result; generating a response statement corresponding to the statement to be responded to in the target scene.

[0013] Optionally, before matching the target scenario corresponding to the target multidimensional classification result, the method also includes: obtaining an online response log and a scenario corresponding to the online response log; classifying the online response log using the multidimensional intent recognition model to obtain a multidimensional classification result corresponding to the online response log; and establishing a mapping relationship between the classification result and the scenario based on the multidimensional classification result corresponding to the online response log and the scenario corresponding to the online response log.

[0014] To achieve the above objective, according to a second aspect of an embodiment of the present invention, a customer service answering device is provided.

[0015] A customer service response device according to an embodiment of the present invention includes: an identification module, which is used to obtain a statement to be answered, and identify a first multidimensional classification result corresponding to the statement to be answered based on a multidimensional intention recognition model; a determination module, which is used to obtain a slot-filling statement corresponding to the statement to be answered, and determine a target multidimensional classification result corresponding to the statement to be answered based on the first multidimensional classification result and the slot-filling statement; and a generation module, which is used to generate a response statement corresponding to the statement to be answered based on the target multidimensional classification result based on a mapping relationship between the classification result and the scene.

[0016] Optionally, the device also includes a training module, which is used to: obtain historical response logs, dimensionally annotate the historical response logs according to preset dimensional attributes, and obtain multidimensional annotation results; based on a deep learning model, according to the multidimensional annotation results, train the sub-intention recognition model corresponding to the dimensional attributes, and use the sub-intention recognition model corresponding to the dimensional attributes to construct the multidimensional intent recognition model; wherein the input of the multidimensional intent recognition model is a sentence, and the output is the classification result of the sentence on the dimensional attributes.

[0017] Optionally, the determination module is also used to: determine whether scene slot filling is needed; if so, obtain the slot filling statement, use the multidimensional intent recognition model to identify the second multidimensional classification result corresponding to the slot filling statement, and determine the target multidimensional classification result based on the first multidimensional classification result and the second multidimensional classification result; if not, directly determine the first multidimensional classification result as the target multidimensional classification result.

[0018] Optionally, the determination module is also used to: determine whether the first multidimensional classification result is the same as the second multidimensional classification result; if so, directly determine the first multidimensional classification result as the target multidimensional classification result; if not, obtain the difference classification results and the same classification results in the second multidimensional classification result, calculate the correlation weight between the difference classification results and the same classification results, and determine the target multidimensional classification result according to the correlation weight according to the preset intention inheritance rule.

[0019] Optionally, the determination module is also used to: count the first number of occurrences of the difference classification result, the second number of occurrences of the same classification result, and the third number of simultaneous occurrences of the difference classification result and the same classification result according to the historical response log, and then calculate the correlation weight between the difference classification result and the same classification result according to the first number, the second number and the third number; calculate the correlation weight table according to the historical response log, and then query the correlation weight between the difference classification result and the same classification result from the correlation weight table.

[0020] Optionally, the determination module is also used to: determine whether the relevance weight is greater than a preset weight; if so, determine that the second multidimensional classification result is the target multidimensional classification result; if not, de-entity generalize the slot-filling statement based on the difference classification result, and then use the multidimensional intent recognition model to identify it to obtain a third multidimensional classification result, and determine the target classification result based on the third multidimensional classification result.

[0021] Optionally, the generation module is further used to: match a target scenario corresponding to the target multi-dimensional classification result based on a mapping relationship between the classification result and the scenario; and generate a response statement corresponding to the statement to be responded to in the target scenario.

[0022] Optionally, the device also includes an establishing module, which is used to: obtain an online response log and a scenario corresponding to the online response log; classify the online response log using the multidimensional intention recognition model to obtain a multidimensional classification result corresponding to the online response log; and establish a mapping relationship between the classification result and the scenario based on the multidimensional classification result corresponding to the online response log and the scenario corresponding to the online response log.

[0023] To achieve the above objective, according to a third aspect of an embodiment of the present invention, an electronic device is provided.

[0024] An electronic device according to an embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the customer service response method according to the embodiment of the present invention.

[0025] To achieve the above objective, according to a fourth aspect of an embodiment of the present invention, a computer-readable medium is provided.

[0026] A computer-readable medium according to an embodiment of the present invention stores a computer program, and when the program is executed by a processor, the customer service response method according to the embodiment of the present invention is implemented.

[0027] An embodiment of the above invention has the following advantages or beneficial effects: it can use a multidimensional intent recognition model to recognize the intent of the sentence to be responded to, thereby solving the problems of high system complexity and maintenance difficulty, low model reusability, and high training cost caused by the multi-layer intent recognition model in the current technology. It can also use the slot filling statement to correct the first multidimensional classification result to obtain the target multidimensional classification result, thereby improving the accuracy of the intent recognition result. In addition, the mapping relationship between the classification result and the scene is used to generate a response statement in combination with the target multidimensional classification result, thereby solving the scene switching problem caused by the different expressions of user answers in the prior art and improving the response accuracy.

[0028] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0030] Figure 1 is a schematic diagram of the main process of the customer service response method according to an embodiment of the present invention;

[0031] Figure 2 is a schematic diagram of the main process of a method for training a multi-dimensional intent recognition model according to an embodiment of the present invention;

[0032] Figure 3 is a schematic diagram of the main process of a method for establishing a mapping relationship between classification results and scenes according to an embodiment of the present invention;

[0033] Figure 4 is a schematic diagram of a mapping relationship between classification results and scenarios stored in a graph database according to an embodiment of the present invention;

[0034] Figure 5 is a schematic diagram of the main process of a method for determining a target multidimensional classification result according to a first multidimensional classification result and a slot-filling statement according to an embodiment of the present invention;

[0035] Figure 6 is a schematic diagram of the main process of the customer service response method according to an embodiment of the present invention;

[0036] Figure 7is a schematic diagram of main modules of a customer service answering device according to an embodiment of the present invention;

[0037] Figure 8 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0038] Fig. 9 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0040] Intelligent customer service mainly consists of two parts: intent recognition and answer session management. Intent recognition is used to identify the user's intention. Intent recognition can use deep learning technology to classify user questions into models, determine user intentions based on classification results, and obtain intent recognition results; answer session management is used to match answer scenarios based on intent recognition results, and fill in slots and ask questions according to scenario requirements. After the user enters the question, the intent recognition model is used to identify the user's intention again, re-enter the scenario and complete the slot filling, and finally return the user's answer. Among them, slot filling refers to the process of completing information in order to convert user intentions into clear instructions for the user. Taking the interaction between users and task-based robots as an example, user information is supplemented through multiple rounds of interaction to ultimately achieve the purpose of task completion. The process of supplementing user information is slot filling.

[0041] Table 1 is an example of an intelligent response method. The intelligent customer service performs intent recognition on the user's question "I want to modify the order" and obtains the intent recognition result "modify the order". The intelligent customer service then fills in the slot and asks back "Please enter the order number". After obtaining the order number "10649835XXXX", the intelligent customer service performs intent recognition again and obtains the result "modify the order". The intelligent customer service then returns the answer "Your order has been shipped and cannot be modified. You can choose to cancel or reject the order and place another order."

[0042]

[0043] Table 1

[0044] In the existing intelligent customer service response method, the intention recognition model used is a multi-layer model. First, the top-level model is used to identify the large business scenario mapped to the user's intention, and then the sub-model corresponding to the large business scenario is entered. Through the sub-model, the sub-scenario mapped to the user's intention is identified. However, as the response becomes more and more refined, the shortcomings of this recognition method become more and more obvious: (1) As the recognition becomes more refined, the top-level model classification and sub-model classification will increase, and the number of levels will also increase, resulting in increasing system complexity and maintenance difficulty; (2) The sub-model is not highly reusable. When the business expands, a large number of training models are required, resulting in a high model training cost. In addition, during the multi-round interaction process of the user in the scene, the configuration based on the scene slot filling will ask the user back. Due to the different ways of expression of the user, it is easy to switch the scene, resulting in irrelevant answers, reduced answer accuracy, and poor user experience.

[0045] In order to solve the above problems, an embodiment of the present invention proposes a customer service response method, which can simplify the system architecture, reduce the model training cost, improve the model reusability, and improve the fluency of the response session and the response experience while ensuring the accuracy of intent recognition and scenario response. Figure 1 FIG. 1 is a schematic diagram of the main process of the customer service response method according to an embodiment of the present invention. Figure 1 As shown, the main process of the customer service response method may include steps S101 to S103.

[0046] Step S101: Obtain a sentence to be answered, and identify a first multidimensional classification result corresponding to the sentence to be answered based on a multidimensional intent recognition model.

[0047] Among them, the multidimensional intention recognition model is composed of multiple sub-intention recognition models. By using the multidimensional intention recognition model to recognize the intent of the sentence to be answered, the multidimensional classification result corresponding to the sentence to be answered can be obtained, that is, by using multiple sub-intention recognition models to recognize the intent of the sentence to be answered, the classification results of the sentence to be answered on multiple sub-intention recognition models can be obtained. For example, the multidimensional intention recognition model M includes three sub-intention recognition models A, B and C. Using M to recognize the intent of the sentence to be answered means using A to recognize the intent of the sentence to be answered to obtain the classification result a, using B to recognize the intent of the sentence to be answered to obtain the classification result b, and using C to recognize the intent of the sentence to be answered to obtain the classification result c. Therefore, using M to recognize the intent of the sentence to be answered can obtain a multidimensional classification result (a, b, c).

[0048] It should be noted that the statement to be answered is equivalent to the statement sent to the intelligent customer service, such as "The address is wrong, how do I change the order address?" In addition, in order to distinguish the multi-dimensional classification results that appear, in the embodiment of the present invention, "the first multi-dimensional classification result", "the second multi-dimensional classification result" and "the third multi-dimensional classification result" are used for limitation.

[0049] In the embodiment of the present invention, a multi-dimensional intent recognition model is used to replace the multi-layer intent recognition model in the prior art, which can reduce the classification within the model. This is because the multi-layer intent recognition model in the prior art is equivalent to a single-dimensional intent recognition model. The segmented intents in a large business scenario need to be identified separately. For example, the sentences to be answered are "modify order", "modify order address", "modify order phone", "modify invoice", "modify invoice address", and "modify invoice header". There are 6 categories to be identified, and for each new business scenario, the content repetition of corpus combing and model training is very high. However, in the embodiment of the present invention, a multi-dimensional intent recognition model is used, and the action sub-intent recognition model only needs to recognize "modify", the business sub-intent recognition model recognizes "order" and "invoice", and the business entity sub-intent recognition model recognizes "address", "phone", and "heading". Other sub-intent recognition models can be identified as null values. In this way, within the model of each dimension, the classification is greatly reduced, the training difficulty is reduced, the accuracy is improved, and the recognition performance is also improved. Moreover, the classification results on the attributes of each dimension can be reused in different business scenarios, reducing the work of repeated training.

[0050] Step S102: Obtain a slot-filling statement corresponding to the sentence to be answered, and determine a target multidimensional classification result corresponding to the sentence to be answered based on the first multidimensional classification result and the slot-filling statement.

[0051] In an embodiment of the present invention, the answering statement can be analyzed in combination with the context, and a slot-filling question can be sent to the user as required, and then the slot-filling statement input by the user can be obtained. Among them, the slot-filling access is to supplement the user information, and the intelligent customer service sends the question and answer statement to the user. For example, the user sends the answering statement "The address is wrong, how can I modify the order address" to the intelligent customer service, and the intelligent customer service can generate a slot-filling question "What is the order number", and then the user enters the slot-filling statement "The order number is 123456XXXX". In addition, based on the dialogue management technology and natural language generation technology, the answering statement and the context can be analyzed, and then the slot-filling question can be performed. The dialogue management technology can estimate the user's goal in each round of the dialogue, manage the input and dialogue history of each round, output the current dialogue state, and then output the next system behavior and updated dialogue state based on the semantic expression input by the user and the current dialogue state; the natural language generation technology can generate a fluent and readable natural language sentence from the semantic expression output by the dialogue strategy, and feedback it to the user.

[0052] After obtaining the slot-filling sentence input by the user, the first multidimensional classification result can be corrected in combination with the slot-filling sentence to obtain the target multidimensional classification result corresponding to the sentence to be answered, thereby improving the accuracy of the multidimensional classification result.

[0053] Step S103: Based on the mapping relationship between the classification results and the scenarios, a response statement corresponding to the statement to be responded to is generated according to the target multi-dimensional classification result.

[0054] In a reference embodiment of the present invention, based on the mapping relationship between classification results and scenes, generating a response statement corresponding to the statement to be responded to according to the target multi-dimensional classification result may include: matching the target scene corresponding to the target multi-dimensional classification result based on the mapping relationship between classification results and scenes; generating a response statement corresponding to the statement to be responded to under the target scene.

[0055] After obtaining the target multidimensional classification result corresponding to the statement to be answered, the target scenario corresponding to the target multidimensional classification result is matched in the mapping relationship between the classification result and the scenario, and then a response statement is generated according to the context and the target scenario. Among them, the target scenario corresponding to the target multidimensional classification result can be matched by a regular expression matching method. Considering that one or more matching scenarios can be returned, the matching scenarios are then sorted, and the matching scenario ranked first is selected as the target scenario, and then the target scenario is entered for answering. For example, the user's statement to be answered is "The delivery address is incorrectly filled in, and the order address needs to be modified", and the target scenario finally obtained is "Modify order address scenario", and the order number is obtained according to the context. Then, in the scenario of modifying the order address, the status of the order number will be queried, and a response statement "Your order has been shipped and cannot be modified; you can choose to cancel or reject the order and place another order" will be generated.

[0056] In the customer service response technical solution of the embodiment of the present invention, a multidimensional intent recognition model can be used to recognize the intent of the response sentence, thereby solving the problems of high system complexity and maintenance difficulty, low model reusability, and high training cost caused by the multi-layer intent recognition model in the current technology. The first multidimensional classification result can also be corrected by using a slot-filling statement to obtain a target multidimensional classification result, thereby improving the accuracy of the intent recognition result. The mapping relationship between the classification result and the scene is used to generate a response sentence in combination with the target multidimensional classification result, thereby solving the scene switching problem caused by different expressions of user answers in the prior art and improving the response accuracy.

[0057] In the embodiment of the present invention, it is necessary to use a multidimensional intent recognition model to perform intent recognition, and to generate a response statement with the help of the mapping relationship between the classification results and the scene. Therefore, training the multidimensional intent recognition model and establishing the mapping relationship between the classification results and the scene can be regarded as data initialization processing. Next, the model training method and the mapping relationship establishment method are described in detail.

[0058] Figure 2 FIG. 1 is a schematic diagram of the main process of the method for training a multi-dimensional intent recognition model according to an embodiment of the present invention. Figure 2 As shown, the main process of the method for training a multi-dimensional intent recognition model may include:

[0059] Step S201, obtaining historical response logs, dimensionally annotating the historical response logs according to preset dimension attributes, and obtaining multi-dimensional annotation results;

[0060] Step S202: Based on the deep learning model and the multi-dimensional annotation results, a sub-intent recognition model corresponding to the dimension attributes is trained;

[0061] Step S203, constructing a multi-dimensional intent recognition model using the sub-intent recognition models corresponding to the dimensional attributes.

[0062] The multidimensional intent recognition model is composed of multiple sub-intent recognition models, and each sub-intent recognition model is obtained by training according to preset dimension attributes. The preset dimension attributes may include: action attributes, business attributes, business entity attributes, scenario attributes, and question attributes. Through training, an action sub-intent recognition model, a business sub-intent recognition model, a business entity sub-intent recognition model, a scenario sub-intent recognition model, and a question sub-intent recognition model can be obtained. Thus, the multidimensional intent recognition model is composed of an action sub-intent recognition model, a business sub-intent recognition model, a business entity sub-intent recognition model, a scenario sub-intent recognition model, and a question sub-intent recognition model.

[0063] The specific training process may be, in step S201, obtaining the historical response log, and then dimensionally annotating the historical response log according to the preset dimension attributes to obtain the annotation result. The historical response log may be the historical response conversation between the user and the intelligent customer service, such as the historical response conversation in the past two months. Then, the historical response log is multi-dimensionally annotated on the preset dimension attributes, such as annotating the historical response log on action attributes, business attributes, business entity attributes, scenario attributes, and question attributes. Table 2 is an example table of multi-dimensional annotation of the historical response log according to an embodiment of the present invention.

[0064]

[0065] Table 2

[0066] After multi-dimensionally annotating the historical response logs in step S201, the multi-dimensional annotation results are obtained, and then in step S202, based on a deep learning model, such as the BERT model (i.e., a new language model, the full name of which is Bidirectional Encoder Representations from Transformers), combined with the multi-dimensional annotation results, the sub-intent recognition model corresponding to each dimensional attribute is trained. Among them, the number of dimensions of the dimensional attributes can be set according to demand, and the number of sub-intent recognition models is the same as the number of dimensions of the preset dimensional attributes. Therefore, in an embodiment of the present invention, the number of sub-intent recognition models in the multi-dimensional intent recognition model can be set according to actual conditions, thereby improving the practicality of this solution.

[0067] By using a multidimensional intent recognition model to recognize the intent of the sentence to be answered, a multidimensional classification result corresponding to the sentence to be answered can be obtained, that is, by using multiple sub-intention recognition models to recognize the intent of the sentence to be answered, the classification results of the sentence to be answered on multiple dimensional attributes can be obtained. For example, the multidimensional intent recognition model is composed of an action sub-intention recognition model, a business sub-intention recognition model, a business entity sub-intention recognition model, a scenario sub-intention recognition model, and a question sub-intention recognition model. The sentence to be answered is "The address is filled in incorrectly, how do I modify the order address?" The intent recognition of the sentence to be answered is performed, and the classification results on the dimensional attributes are: modification (action attribute), order (business attribute), order address (business entity attribute), incorrectly filled in (scenario attribute), and method (question attribute).

[0068] Figure 3 FIG. 1 is a schematic diagram of the main process of the method for establishing a mapping relationship between classification results and scenes according to an embodiment of the present invention. Figure 3 As shown, the main process of the method for establishing a mapping relationship between classification results and scenes may include:

[0069] Step S301, obtaining an online response log and a scenario corresponding to the online response log;

[0070] Step S302, using a multi-dimensional intent recognition model to classify the online response log, and obtain a multi-dimensional classification result corresponding to the online response log;

[0071] Step S303: Establish a mapping relationship between the classification result and the scenario according to the multi-dimensional classification result corresponding to the online response log and the scenario corresponding to the online response log.

[0072] Among them, the online response log is a historical response log that has been matched to a scenario. For example, the online response log is "I want to modify the order address", and the matched scenario is the "modify order address" scenario. In an embodiment of the present invention, the historical response log can be analyzed to obtain response logs that can successfully match the scenario. These response logs that successfully match the scenario are defined as online response logs, and can be combined with their corresponding scenarios to construct a mapping relationship between classification results and scenarios.

[0073] By using a multidimensional intent recognition model to classify the online response log, a multidimensional classification result corresponding to each online response log can be obtained. For example, the online response log is "I want to modify the order address", and the multidimensional classification results obtained are modification (action attribute), order (business attribute), order address (business entity attribute), other (scenario attribute) and other (question attribute). Then, the scenario corresponding to the online response log is mapped with the multidimensional classification result corresponding to the online response log. For example, the multidimensional classification results are: modification (action attribute), order (business attribute), order address (business entity attribute), other (scenario attribute) and other (question attribute), and the corresponding scenario is "modify order address". In the embodiment of the present invention, the corresponding relationship between the multidimensional classification result and the scenario corresponding to each online response log is obtained, and then the mapping relationship between the classification result and the scenario is established. It should be noted that after obtaining the multidimensional classification result corresponding to each online response log, the online response logs with the same multidimensional classification results can be deduplicated, and supplementary processing can also be performed.

[0074] In addition, in an embodiment of the present invention, a mapping relationship between classification results and scenarios may be stored in the form of a graph database, wherein the graph database is a new type of database that can handle large-scale data and ever-changing requirements. Figure 4 It is a schematic diagram of using a graph database to store the mapping relationship between classification results and scenes according to an embodiment of the present invention. Figure 4 In the figure, the big circle represents the scene node and the small circle represents the model node. It can be seen that one scene node is connected to five model nodes, and each model node has different dimensional attributes. The direct connection between the circles represents the relationship between the nodes. Figure 4 It can be seen that by connecting Modify Order Address (scenario node) with Order (business sub-intention recognition model node), Order Address (business entity sub-intention recognition model node), Modify (action sub-intention recognition model node), Others (scenario sub-intention recognition model node) and Others (question sub-intention recognition model node), the scenario {Modify Order Address} corresponding to the multi-dimensional classification result {Order, Order Address, Modify, Others, Others} is obtained.

[0075] After obtaining the first multidimensional classification result, the slot-filling statement can be used to correct the first multidimensional classification result to obtain the target multidimensional classification result, thereby improving the accuracy of user intention recognition. Therefore, in a reference embodiment of the present invention, obtaining the slot-filling statement corresponding to the sentence to be answered, and determining the target multidimensional classification result corresponding to the sentence to be answered based on the first multidimensional classification result and the slot-filling statement, can include:

[0076] Step S1021, determine whether scene slot filling is required, if yes, execute step S1022, if no, execute step S1023;

[0077] Step S1022, obtaining a slot-filling sentence, using a multidimensional intent recognition model to identify a second multidimensional classification result corresponding to the slot-filling sentence, and determining a target multidimensional classification result according to the first multidimensional classification result and the second multidimensional classification result;

[0078] Step S1023: directly determine the first multi-dimensional classification result as the target multi-dimensional classification result.

[0079] First, based on the dialogue management technology and natural language generation technology, the answering statement and the context are analyzed to determine whether the scene needs to be filled in. That is, by analyzing the answering statement and the context, it is determined whether the user information has been supplemented. If so, it means that there is no need to fill in the scene, and the first multi-dimensional classification result is directly determined as the target classification result. If scene filling is required, a slot-filling question is conducted, that is, the intelligent customer service has a dialogue with the user, and the sentence the user answers to the slot-filling question is the slot-filling sentence. For example, the user sends a sentence to be answered to the intelligent customer service "How to modify the order consignee information", the intelligent customer service can generate a slot-filling question "What is the order number", and then the user enters the slot-filling sentence "The order number is 123456XXXX".

[0080] After obtaining the slot-filling statement, the multidimensional intention recognition model can be used to identify the intention of the slot-filling statement to obtain the second multidimensional classification result corresponding to the slot-filling statement. Since the slot-filling statements entered by the user are expressed in different ways, scene switching may occur. For example, the sentence to be answered entered by the user is "The price of the mobile phone I bought has been reduced", and the corresponding scene can be "Apply for price protection", and a slot-filling question "Please enter the order number" is required. The slot-filling statement entered by the user is "I am talking about this 10649835XXXX", and the corresponding scene can be "Order query". It can be concluded that after the user enters the slot-filling statement, the scene switches, which will reduce the user's intention recognition accuracy, and then reduce the scene response accuracy, affecting the user experience. In an embodiment of the present invention, the target classification result can be obtained by comparing the first multidimensional classification result and the second multidimensional classification result, and then the target classification result can be used to generate a response statement.

[0081] When the first multidimensional classification result is the same as the second multidimensional classification result, the first multidimensional classification result can be directly determined as the target multidimensional classification result; when the first multidimensional classification result is different from the second multidimensional classification result, they can be analyzed to determine the target multidimensional classification result. It should be noted that there may be multiple slot-filling statements. For each slot-filling statement, the multidimensional classification result corresponding to the slot-filling statement is obtained, and then two adjacent ones can be analyzed to obtain the target multidimensional classification result. For example, for the sentence N1 to be processed, the corresponding slot-filling sentences are N2, N3 and N4 respectively. The multidimensional intention recognition model M is used to perform intention recognition on N1 to obtain the multidimensional classification result W1; then M is used to perform intention recognition on N2 to obtain the multidimensional classification result W2. W1 and W2 are analyzed to obtain the target classification result W1; then M is used to perform intention recognition on N3 to obtain the multidimensional classification result W3. W1 and W3 are analyzed to obtain the target classification result W3; finally, M is used to perform intention recognition on N4 to obtain the multidimensional classification result W4. W3 and W4 are analyzed to obtain the final target classification result W4.

[0082] In a reference embodiment of the present invention, a target multidimensional classification result is determined based on a first multidimensional classification result and a second multidimensional classification result, including: judging whether the first multidimensional classification result is the same as the second multidimensional classification result; if so, directly determining the first multidimensional classification result as the target multidimensional classification result; if not, obtaining the difference classification results and the same classification results in the second multidimensional classification result, calculating the correlation weights between the difference classification results and the same classification results, and determining the target multidimensional classification result according to the correlation weights according to a preset intention inheritance rule.

[0083] In the embodiment of the present invention, a multidimensional intent recognition model is used to perform intent recognition on the sentence to be processed or the slot-filling sentence, and a multidimensional classification result can be obtained, that is, the classification result on multiple dimensional attributes. Taking into account that the sentence to be processed and the slot-filling sentence have a contextual relationship, the first multidimensional classification result corresponding to the sentence to be processed and the second multidimensional classification result corresponding to the slot-filling sentence may be different classification results on one sub-intent recognition model or multiple sub-intent recognition models. Therefore, in the embodiment of the present invention, the difference classification results and the same classification results in the second multidimensional classification result can be obtained, the correlation weights between the difference classification results and the same classification results can be calculated, and then the target multidimensional classification result can be determined according to the correlation weights according to the preset intent inheritance rules.

[0084] For example, the first multidimensional classification result {action sub-intent recognition model, business sub-intent recognition model} corresponding to the sentence to be processed "The price of the mobile phone I bought has been reduced" is {application, price protection}, and the second multidimensional classification result {action sub-intent recognition model, business sub-intent recognition model} corresponding to the slot-filling sentence "This is the 10649853XXXX I'm talking about" is {application, order}. It can be found that compared with the first multidimensional classification result, the classification result of the second multidimensional classification result on the business sub-intent recognition model has changed, and the difference classification results {business sub-intent recognition model, order} and the same classification results {action sub-intent recognition model, application} in the second multidimensional classification result are obtained, and then the target multidimensional classification result is determined by calculating the correlation weight.

[0085] Considering that the target multidimensional classification result can be determined by calculating the correlation weights between the difference classification results and the same classification results. Therefore, the present invention can refer to the embodiment to calculate the correlation weights between the difference classification results and the same classification results, which may include: step (a), according to the historical response log, counting the first number of occurrences of the difference classification results, the second number of occurrences of the same classification results, and the third number of occurrences of the difference classification results and the same classification results at the same time; step (b) calculates the correlation weights between the difference classification results and the same classification results based on the first number, the second number, and the third number. Definition: When the classification result a under the A sub-intention recognition model is Model (A, a), the classification result b under the B sub-intention recognition model is Model (B, b), and the correlation weights between Model (A, a) and Model (B, b) are the probability value of the appearance of Model (B, b) when Model (A, a) appears, the specific calculation formula can be:

[0086] R(Model(A,a),Model(B,b))=(Count(Model(B,b)) / Count(Model(A,a)))*lg(Count(Model(A,a))

[0087] Among them, Count(Model(A,a)) represents the number of times classification result a appears under the A sub-intent recognition model based on historical response logs; Count(Model(B,b)) represents the number of times classification result b appears under the B sub-intent recognition model based on historical response logs.

[0088] In an embodiment of the present invention, steps (a) and (b) above can also be used to implement the process. Specifically, after determining the difference classification results and the same classification results, the historical response log is queried to count the first number of occurrences of the difference classification results, the second number of occurrences of the same classification results, and the third number of occurrences of the difference classification results and the same classification results at the same time; then, based on the first number, the second number, and the third number, the correlation weight between the difference classification results and the same classification results is calculated.

[0089] For example, for the different classification results {business sub-intention recognition model, order} and the same classification results {action sub-intention recognition model, application}, by querying the historical response log, the first number of orders appearing under the business sub-intention recognition model, the second number of applications appearing under the action sub-intention recognition model, and the third number of orders appearing under the business sub-intention recognition model and applications appearing under the action sub-intention recognition model are counted. Then, based on the first number, the second number, and the third number, the correlation weights of the different classification results {business sub-intention recognition model, order} and the same classification results {action sub-intention recognition model, application} are calculated.

[0090] In an embodiment of the present invention, a correlation weight table can also be calculated in advance based on historical response logs, storing the correlation weights between the classification results of each dimension in the multidimensional intent model. In this way, after obtaining the different classification results and the same classification results, the correlation weights between the different classification results and the same classification results can be queried by querying the correlation weight table.

[0091] In addition, the preset intention inheritance rule refers to a preset rule for determining whether to inherit the intention recognition result. In the reference embodiment of the present invention, according to the preset intention inheritance rule, the target multidimensional classification result is determined according to the correlation weight, which may include: determining whether the correlation weight is greater than the preset weight; if so, determining the second multidimensional classification result as the target multidimensional classification result; if not, de-entity generalization processing is performed on the slot filling statement according to the difference classification result, and then the multidimensional intention recognition model is used for recognition to obtain a third multidimensional classification result, and the target classification result is determined according to the third multidimensional classification result.

[0092] The correlation weight calculates the correlation between the difference classification result and the same classification result. The difference classification result is the different classification result part in the second multidimensional classification result corresponding to the slot filling statement compared with the first multidimensional classification result corresponding to the statement to be answered. Therefore, if the correlation weight is greater than the preset weight, it means that the correlation between the difference classification result and the same classification result is greater, and the second multidimensional classification result can be determined to be the target classification result.

[0093] If the relevance weight is not greater than the preset weight, the slot-filling statement can be de-entitized and generalized according to the difference classification result. For example, the slot-filling statement "I am talking about this 10649853XXXX" obtains the difference classification result {business sub-intent recognition model, order} and the same classification result {action sub-intent recognition model, application}, then the slot-filling statement is de-entitized and generalized to obtain "I am talking about this (other)", that is, the corresponding parts of the difference classification result are replaced with (other), and then the multidimensional intent recognition model is used to perform intent recognition, and the third multidimensional classification result {action sub-intent recognition model, business sub-intent recognition model} is obtained as {application, price protection}. The comparison shows that it is the same as the first multidimensional classification result, which means that the first multidimensional classification result can be directly inherited, that is, the first multidimensional classification result is determined as the target multidimensional classification result. Of course, if the third multidimensional classification result obtained by identification is the same as the second multidimensional classification result, or the de-entitized part obtained by analyzing the third multidimensional classification result can affect the model classification, then the second multidimensional classification result is determined as the target multidimensional classification result.

[0094] It should be noted that in the embodiment of the present invention, the number of difference classification results can be one or more, and the number of identical classification results can also be one or more. For example, when there are two difference classification results Model (A1, a1) and Model (A2, a2), and two identical classification results Model (B1, b1) and Model (B2, b2), the correlation weight R1 between Model (A1, a1) and Model (B1, b1), the correlation weight R2 between Model (A1, a1) and Model (B2, b2), the correlation weight R3 between Model (A2, a2) and Model (B1, b1), and the correlation weight R4 between Model (A2, a2) and Model (B2, b2) can be calculated. If the product of R1 and R2 is greater than the preset weight Y1, and the product of R3 and R4 is greater than the preset weight Y2, it means that Model (A1, a1) and Model (A2, a2) are components of the target multidimensional classification result. If the product of R1 and R2 is not greater than the preset weight Y1, it is necessary to perform entity generalization processing on the part corresponding to Model (A1, a1) to obtain the third multidimensional classification result, and then obtain the target multidimensional classification result.

[0095] Figure 5 1 is a schematic diagram of the main process of a method for determining a target multidimensional classification result based on a first multidimensional classification result and a slot-filling statement according to an embodiment of the present invention. Figure 5 As shown, the main process of the method for determining the target multidimensional classification result according to the first multidimensional classification result and the slot filling statement may include:

[0096] Step S501, determine whether scene slot filling is required, if yes, execute step S502, if no, execute step S504;

[0097] Step S502, obtaining a slot-filling sentence, and using a multi-dimensional intent recognition model to identify a second multi-dimensional classification result corresponding to the slot-filling sentence;

[0098] Step S503, determining whether the first multi-dimensional classification result is the same as the second multi-dimensional classification result, if so, executing step S504, if not, executing step S505;

[0099] Step S504, directly determining the first multi-dimensional classification result as the target multi-dimensional classification result;

[0100] Step S505, obtaining the difference classification results and the same classification results in the second multi-dimensional classification results, and calculating the correlation weights between the difference classification results and the same classification results;

[0101] Step S506, determining whether the correlation weight is greater than a preset weight, if so, executing step S507, if not, executing step S508;

[0102] Step S507, determining the second multi-dimensional classification result as the target multi-dimensional classification result;

[0103] Step S508, according to the difference classification results, perform entity generalization processing on the slot-filling sentence, and then use the multi-dimensional intent recognition model to recognize it, to obtain a third multi-dimensional classification result;

[0104] Step S509: determine the target classification result according to the third multi-dimensional classification result.

[0105] It should be noted that the calculation of the correlation weights between the difference classification results and the same classification results in step S505 can be implemented according to the above steps (a) and (b) after determining the difference classification results and the same classification results, or it can be implemented after pre-calculating the correlation weight table and then querying the specific value from the correlation weight table.

[0106] In an embodiment of the present invention, a slot-filling statement can be used to correct the first multidimensional classification result to obtain a target multidimensional classification result, thereby improving the accuracy of the intent recognition result, solving the scene switching problem caused by different expressions of user answers in the prior art, and thereby improving the accuracy of the response.

[0107] Figure 6 FIG. 1 is a schematic diagram of the main process of the customer service response method according to an embodiment of the present invention. Figure 6 As shown, the main process of the customer service response method may include:

[0108] Step S601, obtaining a sentence to be answered, and identifying a first multi-dimensional classification result corresponding to the sentence to be answered based on a multi-dimensional intention recognition model;

[0109] Step S602, determine whether scene slot filling is required, if yes, execute step S603, if no, execute step S605;

[0110] Step S603, obtaining a slot-filling sentence, and using a multi-dimensional intention recognition model to identify a second multi-dimensional classification result corresponding to the slot-filling sentence;

[0111] Step S604, determining whether the first multi-dimensional classification result is the same as the second multi-dimensional classification result, if so, executing step S605, if not, executing step S606;

[0112] Step S605, directly determining the first multi-dimensional classification result as the target multi-dimensional classification result;

[0113] Step S606, obtaining the difference classification results and the same classification results in the second multi-dimensional classification results, and calculating the correlation weights between the difference classification results and the same classification results;

[0114] Step S607, determining whether the correlation weight is greater than a preset weight, if so, executing step S608, if not, executing step S609;

[0115] Step S608, determining the second multi-dimensional classification result as the target multi-dimensional classification result;

[0116] Step S609, according to the difference classification results, perform entity generalization processing on the slot filling sentence, and then use the multi-dimensional intent recognition model to recognize it, to obtain a third multi-dimensional classification result;

[0117] Step S610, determining a target classification result according to the third multi-dimensional classification result;

[0118] Step S611, matching the target scene corresponding to the target multi-dimensional classification result based on the mapping relationship between the classification result and the scene;

[0119] Step S612: Generate a response statement corresponding to the statement to be responded to in the target scenario.

[0120] Among them, before executing step S601, it is necessary to pre-train the multidimensional intent recognition model. The specific training method has been explained in detail in steps S201 to S203 above, and will not be repeated here. In addition, the calculation of the correlation weights between the difference classification results and the same classification results in step S606 can be implemented according to steps (a) and (b) above after determining the difference classification results and the same classification results, or it can be pre-calculated to obtain the correlation weight table, and then the specific value is queried from the correlation weight table. In addition, before step S611, a mapping relationship between the classification results and the scenes is established. The specific method has been explained in detail in steps S301 to S303 above, and will not be repeated here.

[0121] In the customer service response technical solution of the embodiment of the present invention, a multidimensional intent recognition model can be used to recognize the intent of the response sentence, thereby solving the problems of high system complexity and maintenance difficulty, low model reusability, and high training cost caused by the multi-layer intent recognition model in the current technology. The first multidimensional classification result can also be corrected by using a slot-filling statement to obtain a target multidimensional classification result, thereby improving the accuracy of the intent recognition result. The mapping relationship between the classification result and the scene is used to generate a response sentence in combination with the target multidimensional classification result, thereby solving the scene switching problem caused by different expressions of user answers in the prior art and improving the response accuracy.

[0122] Figure 7 Schematic diagram of the main modules of the customer service response device according to an embodiment of the present invention. Figure 7 As shown, the main modules of the customer service response device 700 may include: an identification module 701 , a determination module 702 and a generation module 703 .

[0123] Among them, the identification module 701 can be used to obtain the sentence to be answered, and based on the multidimensional intention recognition model, identify the first multidimensional classification result corresponding to the sentence to be answered; the determination module 702 can be used to obtain the slot-filling sentence corresponding to the sentence to be answered, and determine the target multidimensional classification result corresponding to the sentence to be answered according to the first multidimensional classification result and the slot-filling sentence; the generation module 703 can be used to generate the answer sentence corresponding to the sentence to be answered according to the target multidimensional classification result based on the mapping relationship between the classification result and the scene.

[0124] In the embodiment of the present invention, the customer service response device 700 may also include a training module (not shown in the figure). The training module can be used to obtain historical response logs, dimensionally annotate the historical response logs according to preset dimension attributes, and obtain multi-dimensional annotation results; based on the deep learning model, according to the multi-dimensional annotation results, train the sub-intention recognition model corresponding to the dimension attributes, and use the sub-intention recognition model corresponding to the dimension attributes to form a multi-dimensional intent recognition model. The input of the multi-dimensional intent recognition model is a sentence, and the output is the classification result of the sentence on the dimension attributes.

[0125] In an embodiment of the present invention, the determination module 702 can also be used to: determine whether scene slot filling is required; if so, obtain the slot filling statement, use the multidimensional intent recognition model to identify the second multidimensional classification result corresponding to the slot filling statement, and determine the target multidimensional classification result based on the first multidimensional classification result and the second multidimensional classification result; if not, directly determine the first multidimensional classification result as the target multidimensional classification result.

[0126] In an embodiment of the present invention, the determination module 702 can also be used to: determine whether the first multidimensional classification result is the same as the second multidimensional classification result; if so, directly determine the first multidimensional classification result as the target multidimensional classification result; if not, obtain the difference classification results and the same classification results in the second multidimensional classification result, calculate the correlation weights between the difference classification results and the same classification results, and determine the target multidimensional classification result according to the correlation weights according to the preset intention inheritance rules.

[0127] In an embodiment of the present invention, the determination module 702 can also be used to: count the first number of occurrences of the difference classification results, the second number of occurrences of the same classification results, and the third number of occurrences of the difference classification results and the same classification results at the same time based on the historical response log, and then calculate the correlation weight between the difference classification results and the same classification results based on the first number, the second number, and the third number; calculate the correlation weight table based on the historical response log, and then query the correlation weight between the difference classification results and the same classification results from the correlation weight table.

[0128] In an embodiment of the present invention, the determination module 702 can also be used to: determine whether the relevance weight is greater than a preset weight; if so, determine that the second multidimensional classification result is the target multidimensional classification result; if not, perform de-entity generalization processing on the slot-filling statement based on the difference classification result, and then use the multidimensional intent recognition model to identify it to obtain a third multidimensional classification result, and determine the target classification result based on the third multidimensional classification result.

[0129] In the embodiment of the present invention, the generation module 703 can also be used to: match the target scene corresponding to the target multi-dimensional classification result based on the mapping relationship between the classification result and the scene; and generate a response statement corresponding to the statement to be responded to in the target scene.

[0130] In the embodiment of the present invention, the customer service response device may further include: an establishment module (not shown in the figure). The establishment module may be used to: obtain the online response log and the scene corresponding to the online response log; classify the online response log using the multidimensional intention recognition model to obtain the multidimensional classification result corresponding to the online response log; and establish a mapping relationship between the classification result and the scene according to the multidimensional classification result corresponding to the online response log and the scene corresponding to the online response log.

[0131] From the above description, it can be seen that the customer service response device of the embodiment of the present invention can use a multidimensional intent recognition model to recognize the intent of the response sentence, thereby solving the problems of high system complexity and maintenance difficulty, low model reusability, and high training cost caused by the multi-layer intent recognition model in the current technology. The first multidimensional classification result can also be corrected by using a slot-filling statement to obtain a target multidimensional classification result, thereby improving the accuracy of the intent recognition result. The mapping relationship between the classification result and the scene is used to generate a response sentence in combination with the target multidimensional classification result, thereby solving the scene switching problem caused by different expressions of user answers in the prior art and improving the response accuracy.

[0132] Figure 8 An exemplary system architecture 800 is shown to which the customer service response method or the customer service response device according to the embodiment of the present invention can be applied.

[0133] like Figure 8 As shown, system architecture 800 may include terminal devices 801, 802, 803, a network 804, and a server 805. Network 804 is used to provide a medium for communication links between terminal devices 801, 802, 803 and server 805. Network 804 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0134] Users can use terminal devices 801, 802, and 803 to interact with server 805 through network 804 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 801, 802, and 803, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only examples).

[0135] The terminal devices 801 , 802 , and 803 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0136] The server 805 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 801, 802, and 803. The backend management server may analyze and process the received data such as product information query requests, and feed back the processing results (such as target push information, product information - only an example) to the terminal device.

[0137] It should be noted that the customer service response method provided in the embodiment of the present invention is generally executed by the server 805 , and accordingly, the customer service response device is generally arranged in the server 805 .

[0138] It should be understood that Figure 8The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0139] Reference below Fig. 9 , which shows a schematic diagram of the structure of a computer system 900 of a terminal device suitable for implementing an embodiment of the present invention. Fig. 9 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0140] like Fig. 9 As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the system 900 are also stored. The CPU 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0141] The following components are connected to the I / O interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage section 908 as needed.

[0142] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 909, and / or installed from the removable medium 911. When the computer program is executed by the central processing unit (CPU) 901, the above-mentioned functions defined in the system of the present invention are executed.

[0143] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0144] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0145] The modules involved in the embodiments of the present invention may be implemented in software or hardware. The modules described may also be provided in a processor. For example, they may be described as follows: a processor includes an identification module, a determination module, and a generation module. Among them, the names of these modules do not constitute a limitation on the modules themselves in certain circumstances. For example, the identification module may also be described as a "module that obtains a statement to be answered and identifies the first multidimensional classification result corresponding to the statement to be answered based on a multidimensional intent recognition model."

[0146] As another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs. When the above one or more programs are executed by a device, the device includes: obtaining a statement to be answered, and identifying a first multidimensional classification result corresponding to the statement to be answered based on a multidimensional intent recognition model; obtaining a slot-filling statement corresponding to the statement to be answered, and determining a target multidimensional classification result corresponding to the statement to be answered based on the first multidimensional classification result and the slot-filling statement; based on the mapping relationship between the classification result and the scene, generating a response statement corresponding to the statement to be answered according to the target multidimensional classification result.

[0147] According to the technical solution of the embodiment of the present invention, a multidimensional intent recognition model can be used to recognize the intent of the sentence to be responded to, thereby solving the problems of high system complexity and maintenance difficulty, low model reusability, and high training cost caused by the multi-layer intent recognition model in the existing technology. The first multidimensional classification result can also be corrected by using a slot-filling statement to obtain a target multidimensional classification result, thereby improving the accuracy of the intent recognition result. The mapping relationship between the classification result and the scene is used to generate a response statement in combination with the target multidimensional classification result, thereby solving the scene switching problem caused by different expressions of user answers in the prior art and improving the response accuracy.

[0148] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A customer service response method, characterized in that: include: Obtaining a sentence to be answered, and identifying a first multidimensional classification result corresponding to the sentence to be answered based on a multidimensional intent recognition model; Obtaining a slot-filling statement corresponding to the sentence to be answered, and determining a target multidimensional classification result corresponding to the sentence to be answered according to the first multidimensional classification result and the slot-filling statement; wherein, identifying a second multidimensional classification result corresponding to the slot-filling statement, and when the first multidimensional classification result is different from the second multidimensional classification result, obtaining a difference classification result and a same classification result in the second multidimensional classification result, calculating a correlation weight between the difference classification result and the same classification result, and determining the target multidimensional classification result according to the correlation weight; The calculating of the correlation weight between the difference classification result and the same classification result includes at least one of the following options: According to the historical answer log, counting the first number of occurrences of the difference classification result, the second number of occurrences of the same classification result, and the third number of occurrences of the difference classification result and the same classification result at the same time, and then calculating the correlation weight between the difference classification result and the same classification result according to the first number, the second number, and the third number; Calculating a correlation weight table according to the historical response log, and then querying the correlation weights of the difference classification results and the same classification results from the correlation weight table; Based on the mapping relationship between the classification results and the scenarios, a response statement corresponding to the statement to be responded to is generated according to the target multi-dimensional classification result.

2. The method according to claim 1, characterized in that: The method further comprises: Obtaining historical response logs, dimensionally annotating the historical response logs according to preset dimension attributes, and obtaining multi-dimensional annotation results; Based on the deep learning model, according to the multi-dimensional annotation results, the sub-intention recognition model corresponding to the dimensional attribute is trained, and the multi-dimensional intention recognition model is constructed using the sub-intention recognition model corresponding to the dimensional attribute; wherein, The input of the multi-dimensional intent recognition model is a sentence, and the output is the classification result of the sentence on the dimensional attributes.

3. The method according to claim 1, characterized in that The step of obtaining a slot-filling statement corresponding to the sentence to be answered, and determining a target multi-dimensional classification result corresponding to the sentence to be answered according to the first multi-dimensional classification result and the slot-filling statement, includes: Determine whether scene slot filling is needed; If yes, then obtaining the slot-filling sentence, using the multidimensional intent recognition model to identify the second multidimensional classification result corresponding to the slot-filling sentence, and determining the target multidimensional classification result according to the first multidimensional classification result and the second multidimensional classification result; If not, directly determine the first multi-dimensional classification result as the target multi-dimensional classification result.

4. The method according to claim 3, characterized in that The step of determining the target multidimensional classification result according to the first multidimensional classification result and the second multidimensional classification result includes: Determining whether the first multi-dimensional classification result is the same as the second multi-dimensional classification result; If yes, directly determining the first multi-dimensional classification result as the target multi-dimensional classification result; If not, obtain the difference classification results and the same classification results in the second multidimensional classification results, calculate the correlation weights between the difference classification results and the same classification results, and determine the target multidimensional classification results according to the correlation weights according to the preset intention inheritance rules.

5. The method according to claim 4, characterized in that The step of determining the target multi-dimensional classification result according to the relevance weight according to the preset intention inheritance rule includes: Determining whether the correlation weight is greater than a preset weight; If yes, determining the second multi-dimensional classification result as the target multi-dimensional classification result; If not, the slot-filling statement is de-entity-generalized based on the difference classification results, and then recognized using the multidimensional intent recognition model to obtain a third multidimensional classification result, and the target multidimensional classification result is determined based on the third multidimensional classification result.

6. The method according to claim 1, characterized in that The mapping relationship between the classification result and the scene is based on the target multi-dimensional classification result, and generating a response statement corresponding to the statement to be responded to includes: Based on the mapping relationship between the classification results and the scenes, matching the target scenes corresponding to the target multi-dimensional classification results; In the target scenario, a response statement corresponding to the statement to be responded to is generated.

7. The method according to claim 6, characterized in that Before matching the target scene corresponding to the target multi-dimensional classification result, the method further includes: Obtaining an online response log and a scenario corresponding to the online response log; Classifying the online response log using the multidimensional intent recognition model to obtain a multidimensional classification result corresponding to the online response log; According to the multi-dimensional classification results corresponding to the online response logs and the scenarios corresponding to the online response logs, a mapping relationship between the classification results and the scenarios is established.

8. A customer service answering device, characterized in that: include: An identification module, used for obtaining a sentence to be answered, and identifying a first multi-dimensional classification result corresponding to the sentence to be answered based on a multi-dimensional intention recognition model; A determination module, used to obtain a slot-filling statement corresponding to the sentence to be answered, and determine a target multidimensional classification result corresponding to the sentence to be answered according to the first multidimensional classification result and the slot-filling statement; wherein, a second multidimensional classification result corresponding to the slot-filling statement is identified, and when the first multidimensional classification result is different from the second multidimensional classification result, a difference classification result and an identical classification result in the second multidimensional classification result are obtained, a correlation weight between the difference classification result and the identical classification result is calculated, and the target multidimensional classification result is determined according to the correlation weight; The calculating of the correlation weight between the difference classification result and the same classification result includes at least one of the following options: According to the historical answer log, counting the first number of occurrences of the difference classification result, the second number of occurrences of the same classification result, and the third number of occurrences of the difference classification result and the same classification result at the same time, and then calculating the correlation weight between the difference classification result and the same classification result according to the first number, the second number, and the third number; Calculating a correlation weight table according to the historical response log, and then querying the correlation weights of the difference classification results and the same classification results from the correlation weight table; A generation module is used to generate a response statement corresponding to the statement to be responded to based on the mapping relationship between the classification result and the scene and according to the target multi-dimensional classification result.

9. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images