A text processing method, apparatus, device and medium

Abstract

The application relates to the technical field of deep learning, in particular to a text processing method and device, equipment and medium. The application can input the corresponding target dialogue text of a user to be detected in completed dialogue content into a target model corresponding to the target identity type of the user, and obtain a target quality indicator value corresponding to the user output in the target model. On the one hand, the target quality indicator value of the user can be intelligently detected, thereby avoiding the problems of inaccurate quality inspection results due to the subjective consciousness of quality inspectors and high labor costs when the quality inspection is performed by using an artificial quality inspection method. On the other hand, the quality inspection is performed based on the deep learning method, so that the target quality indicator value is more accurate compared with a non-deep learning method. On the other hand, the quality inspection is not dependent on the matching result of the target dialogue text and a keyword or a regular expression, thereby avoiding the problem of high maintenance cost of the keyword or the regular expression in the later period.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a text processing method and device, equipment and a medium. BACKGROUND

[0002] With the increasing maturity of the Internet development, compared with offline advertising, offline distribution of flyers and other ways, reaching customers through online ways can often break through the time and space limitations, and is more conducive to the promotion and development of various businesses. Specifically, reaching customers through online ways mainly includes: calling customers through human agents and calling customers through intelligent voice robots. In order to ensure the quality of reaching, it is necessary to accurately determine the professionalism of human agents when calling customers, the accuracy of the response of the intelligent robot, and the emotional preference of the customer.

[0003] In order to ensure the quality of reaching, traditional manual quality inspection methods can be used to supervise the quality of reaching. Specifically, quality inspectors manually listen to a large number of call recordings or chat records, and use their own subjective consciousness as a standard to determine whether the communication work of human agents is compliant, whether the response voice output by the intelligent robot according to the input voice of the customer is correct, and whether the customer has a complaint intention, etc. This method requires quality inspectors to have a considerable amount of professional knowledge, and clear and unified determination standards need to be developed to reduce the differences in subjective consciousness between different quality inspectors. At the same time, when the number of call recordings to be inspected is large, limited human resources often cannot achieve full coverage of quality inspection, so only a sampling inspection method can be used for quality inspection, and the quality inspection result obtained may not be reliable.

[0004] In addition, keywords or regular expressions that can determine the quality of reaching can be pre-stored in a database, and the subsequent conversation text corresponding to the call record or chat record to be inspected is matched with the keywords or regular expressions stored in the database. According to the matching result, it is determined whether the communication work of human agents is compliant, whether the response voice output by the intelligent robot according to the input voice of the customer is correct, and whether the customer has a complaint intention, etc. However, this method has a high dependence on the completeness of the keywords or regular expressions in the database, and the maintenance cost of the keywords or regular expressions in the database is high. In the case of the increasing complexity of the business scenario, the quality inspection result will also gradually be difficult to meet the industry standards. SUMMARY

[0005] The present application provides a text processing method, device, equipment and medium to solve the problem of low quality of quality inspection of the quality of reaching and inaccurate determination of the quality inspection result in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a text processing method, which comprises:

[0007] If an instruction for processing completed dialogue content is received, target dialogue text input by a user of a target identity type is determined according to the dialogue content and a target identity type of the user to be detected carried in the instruction, wherein the target identity type is a customer, an artificial agent or an intelligent robot;

[0008] A target model corresponding to the target identity type is determined according to a correspondence between the target identity type and a pre-stored identity type and model;

[0009] The target dialogue text is input into the target model to obtain a target quality indicator value corresponding to the user.

[0010] Further, the instruction also carries a target quality indicator to be determined;

[0011] Before the target dialogue text is input into the target model to obtain the target quality indicator value corresponding to the user, the method further comprises:

[0012] A target sub-model corresponding to the target quality indicator is determined according to the target quality indicator and a pre-stored correspondence between a quality indicator and a sub-model in the target model, and the target sub-model is used to update the target model.

[0013] Further, the process of training the sub-model corresponding to each quality indicator comprises:

[0014] For the sub-model corresponding to each quality indicator, any first sample dialogue text in a first training set is obtained, the first sample dialogue text is pre-labeled with a first quality indicator value corresponding to the quality indicator, the first sample dialogue text is input into an original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model, and the original sub-model is trained according to the first predicted quality indicator value and the first quality indicator value.

[0015] Further, the method further comprises:

[0016] For any second sample dialogue text in a second training set, the second sample dialogue text is input into the sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model;

[0017] The second sample dialogue text is input into the sub-model trained based on the first training set to obtain a third predicted quality indicator value output by the sub-model;

[0018] train the sub-model based on the second prediction quality indicator value and the third prediction quality indicator value.

[0019] Further, the process of training the model corresponding to each identity type comprises:

[0020] For each identity type, obtain a third sample dialogue text in a third training set, and the third sample dialogue text is pre-labeled with a first feature matrix corresponding to the third sample dialogue text; wherein each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text; input the third sample dialogue text into the original model to obtain a second feature matrix output by the original model; and train the original model according to the first feature matrix and the second feature matrix.

[0021] Further, the method further comprises:

[0022] For any fourth sample dialogue text in a fourth training set, input the fourth sample dialogue text into the model trained based on the third training set to obtain a third feature matrix output by the model;

[0023] input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model;

[0024] train the model based on the third feature matrix and the fourth feature matrix.

[0025] In a second aspect, the embodiments of the present application further provide a text processing device, the device comprises:

[0026] A determination module is configured to, if an instruction of processing completed dialogue content is received, determine target dialogue text input by a user of a target identity type according to the dialogue content and a target identity type of the user to be detected carried in the instruction; wherein the target identity type is a customer, an artificial agent or an intelligent robot; and determine a target model corresponding to the target identity type according to a pre-stored correspondence relationship between the target identity type and the model.

[0027] An acquisition module is configured to input the target dialogue text into the target model to obtain a target quality indicator value corresponding to the user.

[0028] Further, the determining module is further configured to, if the instruction further carries a target quality indicator to be determined, determine a target sub-model corresponding to the target quality indicator according to the target quality indicator and a pre-stored correspondence between a quality indicator and a sub-model in the target model, and update the target model by using the target sub-model.

[0029] Further, the apparatus further comprises:

[0030] The training module is configured to, for each sub-model corresponding to a quality indicator, obtain any first sample dialogue text in a first training set, the first sample dialogue text being pre-labeled with a first quality indicator value corresponding to the quality indicator; input the first sample dialogue text into an original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model; and train the original sub-model according to the first predicted quality indicator value and the first quality indicator value.

[0031] Further, the training module is further configured to, for any second sample dialogue text in a second training set, input the second sample dialogue text into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model; input the second sample dialogue text into the sub-model trained based on the first training set to obtain a third predicted quality indicator value output by the sub-model; and train the sub-model based on the second predicted quality indicator value and the third predicted quality indicator value.

[0032] Further, the training module is further configured to, for each identity type, obtain any third sample dialogue text in a third training set, and pre-label a first feature matrix corresponding to the third sample dialogue text; wherein each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text; input the third sample dialogue text into an original model to obtain a second feature matrix output by the original model; and train the original model according to the first feature matrix and the second feature matrix.

[0033] Further, the training module is further configured to, for any fourth sample dialogue text in a fourth training set, input the fourth sample dialogue text into a model trained based on the third training set to obtain a third feature matrix output by the model; input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model; and train the model based on the third feature matrix and the fourth feature matrix.

[0034] In a third aspect, the embodiments of the present application further provide an electronic device, which comprises a processor configured to implement the steps of the text processing method according to any one of the above described methods when executing a computer program stored in a memory.

[0035] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium storing a computer program, which is configured to implement the steps of the text processing method according to any one of the above described methods when executed by a processor.

[0036] In a fifth aspect, the embodiments of the present application further provide a computer program product, which comprises computer program code, which, when executed on a computer, causes the computer to perform the steps of the text processing method according to any one of the above described methods.

[0037] In the embodiments of the present application, if an instruction of processing the completed dialogue content is received, the target dialogue text input by the user of the target identity type is determined according to the dialogue content and the target identity type of the user to be detected carried in the instruction, the target identity type is a customer, an artificial clerk or an intelligent robot, the target model corresponding to the target identity type is determined according to the correspondence between the target identity type and the pre-stored identity type and model, the target dialogue text is input into the target model, and the target quality indicator value corresponding to the user is obtained. In the embodiments of the present application, the target dialogue text corresponding to the user to be detected in the completed dialogue content can be input into the target model corresponding to the target identity type of the user to obtain the target quality indicator value corresponding to the user output in the target model, on the one hand, the target quality indicator value of the user can be intelligently detected, thereby avoiding the problem of inaccurate quality inspection results and high labor cost caused by the subjective consciousness of the quality inspector when the quality inspection is performed by using the artificial quality inspection method, on the other hand, the quality inspection is performed based on the deep learning method of the present application, compared with the non-deep learning method in the prior art, the target quality indicator value is more accurate, and on the other hand, the present application does not depend on the matching result of the target dialogue text and the keyword or regular expression when performing the quality inspection, thereby avoiding the problem of high maintenance cost of the keyword or regular expression in the later period. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 A text processing process schematic diagram is provided for some embodiments of the present application.

[0040] Figure 2 A model training process diagram combining supervised learning and semi-supervised learning provided for some embodiments of the present application;

[0041] Figure 3 A process diagram for determining each training sample provided for some embodiments of the present application;

[0042] Figure 4 A text processing device structure diagram provided for some embodiments of the present application;

[0043] Figure 5 A structure diagram of an electronic device provided for some embodiments of the present application. DETAILED DESCRIPTION

[0044] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0045] In order to improve the quality of quality inspection and improve the accuracy of the determined quality inspection result, the present application provides a text processing method, device, equipment and medium.

[0046] In the embodiments of the present application, if an instruction of processing the completed dialogue content is received, the target dialogue text input by the user of the target identity type is determined according to the dialogue content and the target identity type of the user to be detected carried in the instruction, wherein the target identity type is a customer, an artificial agent or an intelligent robot, the target model corresponding to the target identity type is determined according to the corresponding relationship between the target identity type and the pre-stored identity type and model, the target dialogue text is input into the target model, and the target quality indicator value corresponding to the user is obtained.

[0047] The acquisition, storage, use, processing and the like of data in each technical solution of the present application comply with the relevant provisions of national laws and regulations.

[0048] Embodiment 1:

[0049] Figure 1 A text processing process diagram provided for some embodiments of the present application, the process includes the following steps:

[0050] S101: If an instruction of processing completed conversation content is received, target conversation text input by a user of a target identity type is determined according to the conversation content and the target identity type of the user to be detected carried in the instruction, wherein the target identity type is a customer, an artificial clerk or an intelligent robot.

[0051] The text processing method provided in the embodiments of the present application is applied to an electronic device, which can be a smart terminal, a PC or a server, etc., and the text processing method provided in the embodiments of the present application is mainly used in a business handling scenario, a question consultation scenario, etc. Specifically, after obtaining completed conversation content to be processed in a certain scenario, a target quality indicator value corresponding to any one of two users completing the conversation content can be determined, wherein the identity types of the two users in the conversation content are different, for example, the identity types of the two users in the conversation content in a business handling scenario can be a customer and an artificial clerk, etc.

[0052] In the embodiments of the present application, in order to determine the quality indicator of which identity type of user in which conversation content to be detected, the electronic device can first receive an instruction of processing completed conversation content, and the instruction carries a target identity type of a user to be detected. The conversation content can be text conversation content, such as chat records, or voice conversation content, such as call records. The user to be detected is at least one of two users completing the conversation content. The target identity type can be a customer, an artificial clerk or an intelligent robot, etc.

[0053] For ease of description, the embodiments of the present application take the target identity type as a customer, or an artificial clerk, or an intelligent robot as an example for illustration, at this time, the completed conversation content in the embodiments of the present application can be conversation content completed by a user of a customer identity type and a user of an artificial clerk identity type, or conversation content completed by a user of a customer identity type and a user of an intelligent robot identity type.

[0054] In order to detect the quality indicator of the user to be detected, in the embodiments of the present application, target conversation text input by the user to be detected is first determined according to the completed conversation content. For example, if the completed conversation content is conversation content completed by a user of a customer identity type and a user of an artificial clerk identity type, and the identity type of the user to be detected is a customer identity type, the conversation text of the user of the customer identity type, i.e., the target conversation text, is determined according to the conversation content.

[0055] In order to determine the target dialogue text, if the completed dialogue content is voice dialogue content, the voice dialogue of the two identity types of users in the voice dialogue content can be separated by using a dual audio track technology, the voice dialogue corresponding to the to-be-detected user and the voice dialogue corresponding to the other user are obtained, and then the voice dialogue corresponding to the to-be-detected user is converted into target dialogue text. The process of separating the voice dialogue of the two identity types of users in the dialogue content by using the dual audio track technology and the process of converting the voice dialogue into text are prior art and will not be described here.

[0056] If the completed dialogue content is text dialogue content, for each dialogue text contained in the text dialogue content, there is corresponding identification information of the user inputting the dialogue text, and the dialogue text with the identification information corresponding to the to-be-detected user is determined as the target dialogue text.

[0057] S102: According to the correspondence between the target identity type and the pre-stored identity type and model, a target model corresponding to the target identity type is determined.

[0058] In order to accurately determine the target quality indicator value of the user of each target identity type, in the embodiment of the present application, the model corresponding to each identity type can be pre-trained, and the target model can be a bidirectional encoding model with attention mechanism. In order to facilitate the determination of which model to use for prediction each time the quality indicator value is predicted, the correspondence between the identity type and the model is also pre-stored in the electronic device, for example, the first model corresponds to the customer identity type, the second model corresponds to the intelligent robot identity type, the third model corresponds to the artificial agent identity type, and so on.

[0059] In order to determine which model to use to determine the quality indicator value of the to-be-detected user, in the embodiment of the present application, the electronic device determines the target model corresponding to the target identity type of the to-be-detected user in the received instruction and the pre-stored correspondence between the identity type and the model, so as to subsequently predict the target quality indicator value of the to-be-detected user based on the target model.

[0060] S103: The target dialogue text is input into the target model to obtain the target quality indicator value corresponding to the user.

[0061] In order to obtain the target quality indicator value corresponding to the to-be-detected user, in the embodiment of the present application, the target dialogue text is input into the target model to obtain the target quality indicator value corresponding to the user output by the target model, wherein the target quality indicator value can be one or more.

[0062] If the target identity type of the user is a customer, the target quality indicator value can be one or several of the emotional preference value, complaint intention degree, service satisfaction degree, etc. of the customer. If the target identity type of the user is an artificial agent, the target quality indicator value can be one or several of the work efficiency, operation compliance, output efficiency value, service quality value, etc. If the target identity type of the user is an intelligent robot, the target quality indicator value can be one or several of the intention recognition accuracy, interaction correctness, etc.

[0063] Specifically, for each model corresponding to an identity type, the model can predict which quality indicator values corresponding to quality indicators are known.

[0064] In the embodiments of the present application, the target dialogue text of the user to be detected in the completed dialogue content can be input into the target model corresponding to the target identity type of the user to obtain the target quality indicator value corresponding to the user output by the target model. On the one hand, the target quality indicator value of the user can be intelligently detected, avoiding the problem of high labor cost when detecting by artificial quality inspection means and inaccurate quality inspection results due to the subjective consciousness of the quality inspector. On the other hand, the quality inspection is performed based on the deep learning method of the present application, and the target quality indicator value determined is more accurate compared with the non-deep learning method in the prior art. On the other hand, the present application does not rely on the matching result of the target dialogue text and the keyword or regular expression when performing quality inspection, avoiding the problem of high maintenance cost of the keyword or regular expression in the later stage.

[0065] Embodiment 2:

[0066] In order to accurately obtain the target quality indicator value corresponding to the user, on the basis of the above-mentioned embodiments, in the embodiments of the present application, the instructions also carry a target quality indicator to be determined.

[0067] Before the target dialogue text is input into the target model to obtain the target quality indicator value corresponding to the user, the method further comprises:

[0068] According to the target quality indicator and the pre-stored correspondence between the quality indicators and the sub-models in the target model, a target sub-model corresponding to the target quality indicator is determined, and the target sub-model is used to update the target model.

[0069] In the embodiment of the present application, the target model used to determine the target quality indicator value corresponding to the user to be detected can be a multi-label model. Specifically, when training the original multi-label model, the sample text input into the original multi-label model is labeled with multiple labels, i.e., the labeled quality indicator values corresponding to multiple quality indicators, and the original multi-label model can simultaneously output the predicted quality indicator values corresponding to the predicted multiple quality indicators. Subsequently, when identifying based on the target model trained, after each dialogue text to be identified is input into the target model, the target quality indicator values corresponding to the multiple quality indicators of the dialogue text can be predicted. The target model can also be a model composed of multiple single-label sub-models, that is, when training each original single-label sub-model, the sample text input into the original single-label sub-model is labeled with one label, i.e., the labeled quality indicator value corresponding to one quality indicator, and the original single-label sub-model only outputs the predicted quality indicator value corresponding to the predicted quality indicator. Subsequently, when identifying based on the single-label sub-model trained, after each dialogue text to be identified is input into the single-label sub-model, the target quality indicator value corresponding to one quality indicator of the dialogue text can be predicted, and the target quality indicator values corresponding to different quality indicators predicted by each single-label sub-model.

[0070] If the target model is a multi-label model, after the target dialogue text is input into the multi-label model, the multi-label model directly outputs the quality indicator value corresponding to each quality indicator of the user, and subsequently, each output quality indicator value corresponding to each quality indicator can be determined as the target quality indicator value, or part of the output quality indicator values corresponding to each quality indicator can be determined as the target quality indicator value.

[0071] If the target model is a model composed of multiple single-label sub-models, in order to determine which quality indicator value corresponding to the user to be detected is the target quality indicator value, in the embodiment of the present application, the instruction received by the electronic device for processing the completed dialogue content also carries the target quality indicator to be determined. Since each sub-model predicts the target quality indicator value corresponding to one quality indicator, for convenience, the electronic device pre-stores the correspondence between the quality indicators and the sub-models in the target model. After the electronic device determines the target quality indicator carried in the instruction, the target quality indicator and the pre-stored correspondence between the quality indicators and the sub-models in the target model are used to determine the target sub-model corresponding to the target quality indicator, and the target model is updated using the target sub-model, so that subsequently, the target dialogue text is input into the updated target model to obtain the target quality indicator value of the target quality indicator corresponding to the user.

[0072] Embodiment 3:

[0073] To obtain a high-precision sub-model, in the embodiments of the present application, based on the above embodiments, the process of training the sub-model corresponding to each quality indicator includes:

[0074] For the sub-model corresponding to each quality indicator, any first sample dialogue text in the first training set is obtained, and the first sample dialogue text is pre-labeled with the first quality indicator value corresponding to the quality indicator; the first sample dialogue text is input into the original sub-model corresponding to the quality indicator, and the first predicted quality indicator value of the quality indicator output by the original sub-model is obtained; and the original sub-model is trained according to the first predicted quality indicator value and the first quality indicator value.

[0075] In the embodiments of the present application, in order to improve the quality of quality inspection, the sub-model corresponding to each quality indicator can be trained first, and a high-precision sub-model is obtained. In order to obtain a high-precision sub-model, a first training set is pre-configured, and each first sample dialogue text in the first training set is used to train the sub-model. In order to facilitate subsequent training of the sub-model, in addition to each first sample dialogue text, the first training set also contains labels corresponding to each first sample dialogue text, wherein the label is used to label the first quality indicator value corresponding to the first sample dialogue text under the quality indicator.

[0076] For example, if the sub-model corresponding to the customer satisfaction quality indicator is trained, the first sample dialogue text used to train the sub-model corresponding to the customer satisfaction quality indicator needs to be manually labeled with the first quality indicator value corresponding to the customer satisfaction; if the sub-model corresponding to the quality of service of the artificial agent is trained, the first sample dialogue text used to train the sub-model corresponding to the quality of service of the artificial agent needs to be manually labeled with the first quality indicator value corresponding to the quality of service of the artificial agent, and so on.

[0077] In order to train the sub-model, in the embodiments of the present application, after obtaining the first training set, for any first sample dialogue text in the first training set, the first sample dialogue text is input into the original sub-model corresponding to the quality indicator, and the first predicted quality indicator value of the quality indicator output by the original sub-model is obtained. The original sub-model is trained according to the pre-labeled first quality indicator value and the first predicted quality indicator value output by the original sub-model. Specifically, the random gradient descent method can be used to adjust the network parameters in the original sub-model until the performance of the sub-model is adjusted to be optimal.

[0078] Embodiment 4:

[0079] In order to reduce the cost of manual annotation and improve the accuracy of the sub-model, on the basis of the above embodiments, the method further includes the following steps in the embodiments of the present application:

[0080] For any second sample dialogue text in the second training set, the second sample dialogue text is input into the sub-model trained based on the first training set, and a second predicted quality indicator value output by the sub-model is obtained.

[0081] The second sample dialogue text is input into the sub-model trained based on the first training set, and a third predicted quality indicator value output by the sub-model is obtained.

[0082] The sub-model is trained based on the second predicted quality indicator value and the third predicted quality indicator value.

[0083] In order to obtain a high-precision sub-model, a large amount of first sample dialogue text needs to be included in the first training set, and a large amount of first sample dialogue text in the first training set needs to be manually annotated. Although the first sample dialogue text annotated by manual annotation has high reliability, it is difficult to achieve large-scale annotation with high efficiency. Therefore, in order to reduce the cost of manual annotation and obtain a high-precision sub-model, in the embodiments of the present application, when training the sub-model, a part of sample dialogue text (first sample dialogue text) with labeled labels can be used to train the original sub-model first, and a sub-model with low precision is obtained. Then, another part of sample dialogue text (second sample dialogue text) without labeled labels is used to continue training the obtained sub-model with low precision, and finally a high-precision sub-model is obtained. The number of first sample dialogue texts included in the first training set can be set according to requirements.

[0084] In order to facilitate description, the training set containing sample dialogue text with labeled labels is referred to as the first training set, the training set containing sample dialogue text without labeled labels is referred to as the second training set, the sample dialogue text contained in the first training set is referred to as the first sample dialogue text, and the sample dialogue text contained in the second training set is referred to as the second sample dialogue text.

[0085] When training the sub-model trained based on the first training set based on the second sample dialogue text in the second training set, for any second sample dialogue text in the second training set, the second sample dialogue text is input into the sub-model trained based on the first training set, and the sub-model outputs a second predicted quality indicator value. After obtaining the second predicted quality indicator value output by the sub-model, the second predicted quality indicator value is used as the pseudo label of the second sample dialogue text, and the sub-model is further trained.

[0086] Specifically, the second sample dialogue text is input into the sub-model trained based on the first training set, a third predicted quality indicator value output by the sub-model is obtained, and the sub-model is trained based on the second predicted quality indicator value and the third predicted quality indicator value.

[0087] To improve the accuracy of the sub-model, after the pseudo labels of the second sample dialogue texts are determined, part of the pseudo label results can be manually sampled and detected, whether the pseudo label results are correct is determined by professional staff, and whether the accuracy of the pseudo label results is greater than a preset accuracy threshold is determined. If yes, the pseudo labels of the second sample dialogue texts are taken as the annotation labels of the second sample dialogue texts, and the sub-model trained based on the first training set is further trained.

[0088] In addition, in the embodiments of the present application, before the original sub-model is trained based on the first training set, the network parameters of the original sub-model can be randomly initialized, then for any first sample dialogue text in the first training set, the first sample dialogue text is input into the original sub-model corresponding to the quality indicator, a first predicted quality indicator value output by the original sub-model is obtained, and the network parameters in the original sub-model are adjusted according to the pre-labeled first quality indicator value and the first predicted quality indicator value output by the original sub-model. The sub-model preliminarily trained based on the first training set and the second training set is further trained, and the parameters of the sub-model are updated. This process is fine-tuning. The sub-model obtained by algorithm convergence in the fine-tuning link is the final trained sub-model, which has the characteristics of high accuracy and high robustness, and can be used for recognition of downstream systems.

[0089] Embodiment 5:

[0090] To accurately obtain the target quality indicator value corresponding to the user, on the basis of the above embodiments, in the embodiments of the present application, the process of training the model corresponding to each identity type includes:

[0091] For each identity type, any third sample dialogue text in the third training set is obtained, and the first feature matrix corresponding to the third sample dialogue text is pre-labeled. Each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text. The third sample dialogue text is input into the original model, and a second feature matrix output by the original model is obtained. The original model is trained according to the first feature matrix and the second feature matrix.

[0092] In the embodiment of the present application, in order to improve the accuracy of quality inspection, the original model corresponding to each identity type can be trained first to obtain a high-precision model. In order to obtain a high-precision model, a third training set is pre-configured, and the original model corresponding to the quality indicator is trained according to each third sample dialogue text included in the third training set. In order to facilitate subsequent training of the original model corresponding to each identity type, the third training set includes, in addition to each third sample dialogue text, a label corresponding to each third sample dialogue text, wherein the label is a quality indicator value annotated for the third sample dialogue text under each quality indicator. In order to facilitate subsequent model training, a first feature matrix can be determined according to the quality indicator value corresponding to the third sample dialogue under each quality indicator, and the first feature matrix is determined as the label of the third sample dialogue, wherein each element in the first feature matrix is the quality indicator value corresponding to the third sample dialogue under each quality indicator, and the first feature matrix can be a 1xN matrix, or an Nx1 matrix, where N is the total number of annotated quality indicators.

[0093] For each quality indicator, in order to train the original model corresponding to the quality indicator and obtain a high-precision model, in the embodiment of the present application, after obtaining the third training set, for any third sample dialogue text in the third training set, the third sample dialogue text is input into the original model corresponding to the quality indicator, and then the original model outputs a second feature matrix, wherein each element in the second feature matrix is a predicted quality indicator value corresponding to the third sample dialogue under each quality indicator. Finally, the original model is trained according to the first feature matrix and the second feature matrix. Specifically, the random gradient descent method can be used to adjust the network parameters in the original model until the model performance is adjusted to be optimal.

[0094] In order to accelerate the convergence speed of the model, help the model to achieve higher accuracy, and improve the robustness, in the embodiments of the present application, when training the original model corresponding to each quality indicator, a plurality of third sample dialogues can also be input into the original model each time, at this time, the plurality of third sample dialogue texts can be labeled based on the first target feature matrix, wherein each row or column element in the first target feature matrix is the quality indicator value of each third sample dialogue under each quality indicator, and the first target feature matrix is an MxN matrix, and can also be an NxM matrix, N is the total number of each quality indicator to be labeled, and M is the total number of the plurality of third sample dialogues input into the original model each time. The plurality of third sample dialogue texts are input into the original model to obtain a second target feature matrix output by the original model, wherein each row or column element in the second target feature matrix is the predicted quality indicator value of each third sample dialogue under each quality indicator. Finally, the original model is trained according to the first target feature matrix and the second target feature matrix.

[0095] Embodiment 6:

[0096] In order to reduce the cost of manual labeling and improve the accuracy of the model, on the basis of the above embodiments, in the embodiments of the present application, the method further comprises:

[0097] For any fourth sample dialogue text in the fourth training set, the fourth sample dialogue text is input into the model trained based on the third training set to obtain a third feature matrix output by the model;

[0098] The fourth sample dialogue text is input into the model trained based on the third training set to obtain a fourth feature matrix output by the model;

[0099] The model is trained based on the third feature matrix and the fourth feature matrix.

[0100] In order to obtain a high-precision model, a large number of third sample dialogue texts in the third training set often need to be manually annotated. Although the third sample dialogue texts annotated by manual annotation have high reliability, it is difficult to efficiently realize large-scale annotation. Therefore, in order to reduce labor costs and obtain a high-precision model, in the embodiments of the present application, when training the model, a part of the sample dialogue texts (third sample dialogue texts) annotated with labels can be used to train the original model to obtain a model with low precision, and then another part of the sample dialogue texts (fourth sample dialogue texts) without labels is used to continue training the obtained model with low precision, and finally a high-precision model is obtained. The number of third sample dialogue texts contained in the third training set can not be too large, and the number of third sample dialogue texts contained in the third training set can be set according to requirements.

[0101] For the convenience of description, the training set containing sample dialogue texts annotated with labels is referred to as the third training set, the training set containing sample dialogue texts without labels is referred to as the fourth training set, the sample dialogue texts contained in the third training set are referred to as the third sample dialogue texts, and the sample dialogue texts contained in the fourth training set are referred to as the fourth sample dialogue texts.

[0102] When training the model trained based on the third training set based on the third sample dialogue texts in the third training set, for any fourth sample dialogue text in the fourth training set, the fourth sample dialogue text is input into the model trained based on the third training set, and the model outputs a third feature matrix. After obtaining the third feature matrix output by the model, the third feature matrix is used as the pseudo-label of the fourth sample dialogue text, and the model is further trained.

[0103] Specifically, the fourth sample dialogue text is input into the model trained based on the third training set, and the fourth feature matrix output by the model is obtained. The model is trained based on the third feature matrix and the fourth feature matrix.

[0104] In order to improve the precision of the model, after determining the pseudo-labels of each fourth sample dialogue text, part of the pseudo-label results can be manually inspected and detected, and whether the pseudo-label results are correct and whether the correctness of the pseudo-label results is greater than a preset correctness threshold can be determined by professional workers. If yes, the pseudo-labels of each fourth sample dialogue text are used as the annotated labels of each fourth sample dialogue text, and the model trained based on the third training set is further trained.

[0105] In addition, in the embodiment of the present application, before the original model is trained based on the third training set, the network parameters of the original model can be randomly initialized first, then for any third sample dialogue text in the third training set, the third sample dialogue text is input into the original model corresponding to the identity type, a second feature matrix output by the original sub-model is obtained, the network parameters in the original model are adjusted according to the first feature matrix and the labeled first feature matrix, and then the model preliminarily trained based on the third training set and the fourth training set is used to train the model again to update the parameters of the model. This process is the optimization. The model obtained by the algorithm convergence in the optimization link is the finally trained model, which has the characteristics of high accuracy and high robustness and can be used for recognition of a downstream system.

[0106] Figure 2 A model training process combining supervised learning and semi-supervised learning provided by some embodiments of the present application will be described below with reference to Figure 2 .

[0107] For ease of description, the sample set used to train the model based on supervised learning is referred to as a supervised learning sample set, each sample in the supervised learning sample set is referred to as a supervised sample, the sample set used to train the model based on semi-supervised learning is referred to as a semi-supervised learning sample set, and each sample in the semi-supervised learning sample set is referred to as a semi-supervised sample.

[0108] Each supervised sample in the supervised learning sample set is manually labeled first, the model is preliminarily trained, and a preliminarily trained model is obtained. The accuracy of the preliminarily trained model is not particularly high. Then each semi-supervised sample in the semi-supervised learning sample set is input into the preliminarily trained model, the quality indicator value corresponding to each semi-supervised sample is estimated, a first target quality indicator value estimated by the preliminarily trained model is obtained, and for each semi-supervised sample, the first target quality indicator value estimated for the semi-supervised sample is determined as the pseudo label of the semi-supervised sample.

[0109] The pseudo labels corresponding to each semi-supervised sample are manually sampled, and after it is determined that the correctness of the pseudo label result is greater than a preset correctness threshold, each semi-supervised sample is input into the preliminarily trained model, a second target quality indicator value estimated by the preliminarily trained model is obtained, and the preliminarily trained model is trained according to the first target quality indicator value and the second target quality indicator value, and a trained model is obtained.

[0110] Figure 3 A process for determining each training sample provided by some embodiments of the present application will be described below with reference to Figure 3 .

[0111] Each user of the customer identity type can have voice interaction with each user of the human agent identity type, and each user of the customer identity type can also have voice interaction with each user of the intelligent robot identity type, to obtain each voice dialogue content obtained by voice interaction of each two users. For each voice dialogue content, the voice dialogue of the two users of different identity types in the voice dialogue content can be separated by using a double audio track technology, to obtain voice dialogue corresponding to the user of the customer identity type and voice dialogue corresponding to another user in the voice dialogue content, where the another user can be the user of the human agent identity type, or the user of the intelligent robot identity type. Then, all voice dialogues obtained according to each voice dialogue content are converted into dialogue texts, and each dialogue text is determined as a training sample.

[0112] Embodiment 7

[0113] Figure 4 A text processing device structure schematic diagram is provided for some embodiments of the present application, and the device includes:

[0114] The determining module 401 is configured to, if an instruction of processing completed dialogue content is received, determine target dialogue text input by a user of a target identity type according to the dialogue content and a target identity type of the user to be detected carried in the instruction, where the target identity type is a customer, a human agent or an intelligent robot, and determine a target model corresponding to the target identity type according to a correspondence relationship between the target identity type and pre-stored identity types and models.

[0115] The obtaining module 402 is configured to input the target dialogue text into the target model to obtain a target quality indicator value corresponding to the user.

[0116] In a possible implementation, the determining module 401 is further configured to, if a target quality indicator to be determined is further carried in the instruction, determine a target sub-model corresponding to the target quality indicator according to a correspondence relationship between the target quality indicator and sub-models in the target model pre-stored, and update the target model by using the target sub-model.

[0117] In a possible implementation, the device further includes:

[0118] The training module 403 is configured to, for each sub-model corresponding to a quality indicator, obtain any first sample dialogue text in the first training set, the first sample dialogue text being pre-labeled with a first quality indicator value corresponding to the quality indicator; input the first sample dialogue text into the original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model; and train the original sub-model according to the first predicted quality indicator value and the first quality indicator value.

[0119] In a possible implementation, the training module 403 is further configured to, for any second sample dialogue text in the second training set, input the second sample dialogue text into the sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model; input the second sample dialogue text into the sub-model trained based on the first training set to obtain a third predicted quality indicator value output by the sub-model; and train the sub-model based on the second predicted quality indicator value and the third predicted quality indicator value.

[0120] In a possible implementation, the training module 403 is further configured to, for each identity type, obtain any third sample dialogue text in a third training set, and pre-label a first feature matrix corresponding to the third sample dialogue text; wherein each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text; input the third sample dialogue text into the original model to obtain a second feature matrix output by the original model; and train the original model according to the first feature matrix and the second feature matrix.

[0121] In a possible implementation, the training module 403 is further configured to, for any fourth sample dialogue text in a fourth training set, input the fourth sample dialogue text into the model trained based on the third training set to obtain a third feature matrix output by the model;

[0122] input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model; and train the model based on the third feature matrix and the fourth feature matrix.

[0123] Embodiment 8

[0124] On the basis of the above-mentioned embodiments, the embodiments of the present application further provide an electronic device, Figure 5 A structural schematic diagram of an electronic device provided by the embodiments of the present application is shown in FIG. 8. Figure 5As shown, the server comprises: a processor 501, a communication interface 502, a memory 503 and a communication bus 504, wherein the processor 501, the communication interface 502 and the memory 503 complete mutual communication through the communication bus 504;

[0125] The memory 503 stores a computer program, and when the program is executed by the processor 501, the processor 501 executes the following steps:

[0126] If an instruction of processing completed conversation content is received, target conversation text input by a user of a target identity type is determined according to the conversation content and a target identity type of the user to be detected carried in the instruction, wherein the target identity type is a customer, an artificial clerk or an intelligent robot.

[0127] A target model corresponding to the target identity type is determined according to a correspondence relationship between the target identity type and a pre-stored identity type and model.

[0128] The target conversation text is input into the target model to obtain a target quality indicator value corresponding to the user.

[0129] In a possible implementation, the processor 501 is further configured to, if the instruction further carries a target quality indicator to be determined, determine a target sub-model corresponding to the target quality indicator according to a correspondence relationship between the target quality indicator and a pre-stored quality indicator and sub-model in the target model, and update the target model by using the target sub-model.

[0130] In a possible implementation, the processor 501 is further configured to, for each sub-model corresponding to a quality indicator, obtain any first sample conversation text in a first training set, wherein the first sample conversation text is pre-labeled with a first quality indicator value corresponding to the quality indicator, input the first sample conversation text into an original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model, and train the original sub-model according to the first predicted quality indicator value and the first quality indicator value.

[0131] In a possible implementation, the processor 501 is further configured to, for any second sample conversation text in a second training set, input the second sample conversation text into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model, input the second sample conversation text into the sub-model trained based on the first training set to obtain a third predicted quality indicator value output by the sub-model, and train the sub-model based on the second predicted quality indicator value and the third predicted quality indicator value.

[0132] In a possible implementation, the processor 501 is further configured to, for each identity type, obtain any one third sample dialogue text in a third training set, and pre-label a first feature matrix corresponding to the third sample dialogue text; each row or column element in the first feature matrix is a plurality of second quality index values corresponding to the third sample dialogue text; input the third sample dialogue text into the original model to obtain a second feature matrix output by the original model; and train the original model according to the first feature matrix and the second feature matrix.

[0133] In a possible implementation, the processor 501 is further configured to, for any one fourth sample dialogue text in a fourth training set, input the fourth sample dialogue text into the model trained based on the third training set to obtain a third feature matrix output by the model; input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model; and train the model based on the third feature matrix and the fourth feature matrix.

[0134] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, or the like. The communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0135] The communication interface 502 is configured to communicate between the above electronic device and other devices. The memory can include a Random Access Memory (RAM) and can also include a Non-Volatile Memory (NVM), for example, at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0136] The above processor can be a general-purpose processor, including a central processing unit, a network processor (NP), and the like; can also be a Digital Signal Processing (DSP), an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and the like.

[0137] Embodiment 9:

[0138] On the basis of the above-mentioned embodiments, the embodiment of the present application further provides a computer readable storage medium, the computer readable storage medium stores a computer program executable by a processor, when the program runs on the processor, the processor executes the following steps:

[0139] If the instruction of processing the completed dialogue content is received, the target dialogue text input by the user of the target identity type is determined according to the dialogue content and the target identity type of the user to be detected carried in the instruction; wherein the target identity type is a customer, an artificial agent or an intelligent robot;

[0140] According to the correspondence between the target identity type and the pre-stored identity type and model, the target model corresponding to the target identity type is determined;

[0141] The target dialogue text is input into the target model, and the target quality indicator value corresponding to the user is obtained.

[0142] Further, the instruction also carries a target quality indicator to be determined;

[0143] Before the target dialogue text is input into the target model to obtain the target quality indicator value corresponding to the user, the method further comprises:

[0144] According to the target quality indicator and the pre-stored correspondence between the quality indicator and the sub-model in the target model, the target sub-model corresponding to the target quality indicator is determined, and the target model is updated by using the target sub-model.

[0145] Further, the process of training the sub-model corresponding to each quality indicator comprises:

[0146] For the sub-model corresponding to each quality indicator, any first sample dialogue text in the first training set is obtained, the first sample dialogue text is pre-labeled with a first quality indicator value corresponding to the quality indicator; the first sample dialogue text is input into the original sub-model corresponding to the quality indicator, and a first predicted quality indicator value of the quality indicator output by the original sub-model is obtained; according to the first predicted quality indicator value and the first quality indicator value, the original sub-model is trained.

[0147] Further, the method further comprises:

[0148] For any second sample dialogue text in the second training set, the second sample dialogue text is input into the sub-model trained based on the first training set, and a second predicted quality indicator value output by the sub-model is obtained;

[0149] input the second sample dialogue text into the sub-model trained based on the first training set, to obtain a third predicted quality indicator value output by the sub-model;

[0150] train the sub-model based on the second predicted quality indicator value and the third predicted quality indicator value.

[0151] Further, the process of training the model corresponding to each identity type includes:

[0152] For each identity type, any third sample dialogue text in a third training set is obtained, and a first feature matrix corresponding to the third sample dialogue text is pre-labeled; each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text; the third sample dialogue text is input into the original model, to obtain a second feature matrix output by the original model; and the original model is trained according to the first feature matrix and the second feature matrix.

[0153] Further, the method further includes:

[0154] For any fourth sample dialogue text in a fourth training set, the fourth sample dialogue text is input into the model trained based on the third training set, to obtain a third feature matrix output by the model;

[0155] The fourth sample dialogue text is input into the model trained based on the third training set, to obtain a fourth feature matrix output by the model;

[0156] The model is trained based on the third feature matrix and the fourth feature matrix.

[0157] Embodiment 10:

[0158] The embodiments of the present application also provide a computer program product, which, when executed by a computer, implements the text processing method described in any of the method embodiments applied to an electronic device.

[0159] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof, and can be implemented in the form of a computer program product in whole or in part. The computer program product includes one or more computer instructions, which, when loaded and executed on a computer, generates the processes or functions described in the embodiments of the present application in whole or in part.

[0160] Since in the embodiment of the present application, the corresponding target dialogue text of the user to be detected in the completed dialogue content can be input into the target model corresponding to the target identity type of the user, and the target quality indicator value corresponding to the user output by the target model is obtained, on the one hand, the target quality indicator value of the user can be intelligently detected, avoiding the problem of high labor cost when detecting by artificial quality inspection means, and the problem of inaccurate quality inspection results caused by the subjective consciousness of the quality inspector, on the other hand, the quality inspection is performed based on the deep learning method of the present application, and compared with the non-deep learning method in the prior art, the determined target quality indicator value is more accurate, and on the other hand, the present application does not depend on the matching result of the target dialogue text and the keyword or regular expression when performing quality inspection, avoiding the problem of high maintenance cost of the keyword or regular expression in the later stage.

[0161] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for performing the functions specified in one or more flows and / or blocks.

[0163] These computer program instructions can also be stored in a computer-readable storage medium that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable storage medium produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for performing the functions specified in one or more flows and / or blocks.

[0164] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide processes for implementing the functions specified in the flowchart Figure 1 flowchart or multiple flows and / or blocks Figure 1 flowchart or multiple flows and / or blocks

[0165] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A text processing method characterized by, The method comprises: If an instruction of processing completed conversation content is received, target conversation text input by a user of a target identity type is determined according to the conversation content and a target identity type of the user to be detected carried in the instruction, wherein the target identity type is a customer, an artificial agent or an intelligent robot; A target model corresponding to the target identity type is determined according to a correspondence between the target identity type and a pre-stored identity type and model; The target conversation text is input into the target model to obtain a target quality indicator value corresponding to the user; The instruction further carries a target quality indicator to be determined. Before the target conversation text is input into the target model to obtain the target quality indicator value corresponding to the user, the method further comprises: A target sub-model corresponding to the target quality indicator is determined according to the target quality indicator and a pre-stored correspondence between a quality indicator and a sub-model in the target model, and the target sub-model is used to update the target model; each sub-model is used to predict a target quality indicator value corresponding to a quality indicator.

2. The method of claim 1, wherein, The process of training each sub-model corresponding to a quality indicator comprises: For each sub-model corresponding to a quality indicator, any first sample conversation text in a first training set is obtained, the first sample conversation text is pre-labeled with a first quality indicator value corresponding to the quality indicator; the first sample conversation text is input into an original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model; and the original sub-model is trained according to the first predicted quality indicator value and the first quality indicator value.

3. The method of claim 2, wherein, The method further comprises: For any second sample conversation text in a second training set, the second sample conversation text is input into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model; The second sample conversation text is input into the sub-model trained based on the first training set to obtain a third predicted quality indicator value output by the sub-model; The sub-model is trained based on the second predicted quality indicator value and the third predicted quality indicator value.

4. The method of claim 1, wherein, Each model corresponding to an identity type is trained in the following manner: For each identity type, any third sample conversation text in a third training set is obtained, and a first feature matrix corresponding to the third sample conversation text is pre-labeled; each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample conversation text; the third sample conversation text is input into an original model to obtain a second feature matrix output by the original model; and the original model is trained according to the first feature matrix and the second feature matrix.

5. The method of claim 4, wherein, The method further comprises: For any fourth sample conversation text in a fourth training set, the fourth sample conversation text is input into a model trained based on the third training set to obtain a third feature matrix output by the model; input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model; train the model based on the third feature matrix and the fourth feature matrix.

6. A text processing apparatus characterized by comprising: The device comprises: The determining module is configured to, if an instruction for processing completed dialogue content is received, determine target dialogue text input by a user of a target identity type according to the dialogue content and the target identity type carried in the instruction; the target identity type is a customer, an artificial agent or an intelligent robot; and determine a target model corresponding to the target identity type according to a pre-stored correspondence between identity types and models. The obtaining module is configured to input the target dialogue text into the target model to obtain a target quality indicator value corresponding to the user. The determining module is further configured to, if a target quality indicator to be determined is also carried in the instruction, determine a target sub-model corresponding to the target quality indicator according to the target quality indicator and a pre-stored correspondence between quality indicators and sub-models in the target model, and update the target model by using the target sub-model; and each sub-model is configured to predict a target quality indicator value corresponding to one quality indicator.

7. The apparatus of claim 6, wherein, The device further comprises: The training module is configured to, for each sub-model corresponding to a quality indicator, obtain any first sample dialogue text in a first training set, the first sample dialogue text being pre-labeled with a first quality indicator value corresponding to the quality indicator; input the first sample dialogue text into an original sub-model corresponding to the quality indicator to obtain a first predicted quality indicator value of the quality indicator output by the original sub-model; and train the original sub-model according to the first predicted quality indicator value and the first quality indicator value.

8. The apparatus of claim 7, wherein, The training module is further configured to, for any second sample dialogue text in a second training set, input the second sample dialogue text into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model. The training module is further configured to, for any second sample dialogue text in a second training set, input the second sample dialogue text into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model.

9. The apparatus of claim 6, wherein, The training module is further configured to, for any second sample dialogue text in a second training set, input the second sample dialogue text into a sub-model trained based on the first training set to obtain a second predicted quality indicator value output by the sub-model. The device further comprises: The training module is configured to, for each identity type, obtain any third sample dialogue text in a third training set, and pre-label a first feature matrix corresponding to the third sample dialogue text; each row or column element in the first feature matrix is a plurality of second quality indicator values corresponding to the third sample dialogue text; input the third sample dialogue text into an original model to obtain a second feature matrix output by the original model; and train the original model according to the first feature matrix and the second feature matrix.

10. The apparatus of claim 9, wherein, The training module is further configured to, for any fourth sample dialogue text in the fourth training set, input the fourth sample dialogue text into the model trained based on the third training set to obtain a third feature matrix output by the model; input the fourth sample dialogue text into the model trained based on the third training set to obtain a fourth feature matrix output by the model; and train the model based on the third feature matrix and the fourth feature matrix.

11. An electronic device, comprising: The electronic device comprises a processor configured to implement the steps of the text processing method of any one of claims 1-5 when executing a computer program stored in a memory.

12. A computer-readable storage medium, characterized in that, The computer program is stored in the memory and is configured to implement the steps of the text processing method of any one of claims 1-5 when executed by the processor.

13. A computer program product, characterised in that, The computer program product comprises computer program code configured to cause a computer to perform the steps of the text processing method of any one of claims 1-5 when the computer program code is run on the computer.

Images