Conversation processing method and related device

Abstract

The invention provides a dialogue processing method and a related device, and relates to the technical field of artificial intelligence, the method of one embodiment comprises the following steps: performing verbal skill prediction on first text content, and outputting a first verbal skill text; under the condition that the preset type of content exists in the first verbal skill text, identifying a scene category corresponding to the first text content; determining a second verbal skill text from the verbal skill text of the scene category based on a first similarity between the verbal skill text of the scene category and the first verbal skill text; and determining a pushed target verbal skill text based on a second similarity between the first verbal skill text and the second verbal skill text. Based on the scheme of the embodiment, the accuracy of the finally obtained target verbal skill text is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a dialogue processing method and related devices, wherein the related devices include a dialogue processing device, a computer device, a computer-readable storage medium, and a computer program product. Background Art

[0002] With the development of artificial intelligence technology, there will be various scenarios of intelligent dialogue, such as determining the speech text that needs to be returned to the user based on the user's input content. At present, when determining the speech text that needs to be returned to the user, it is usually based on the user's input content and the context of the dialogue with the user, and the speech text to be returned to the user is comprehensively analyzed. However, this method has the problem of inaccurate final speech text. Summary of the invention

[0003] The present application provides a dialogue processing method and related devices that can improve the accuracy of a determined speech text, wherein the related devices include a dialogue processing device, a computer device, a computer-readable storage medium, and a computer program product.

[0004] In a first aspect, the present application provides a conversation processing method, wherein the method comprises:

[0005] Perform speech prediction on the first text content and output the first speech text;

[0006] In the case where there is a preset type of content in the first speech text, identifying the scene category corresponding to the first text content;

[0007] Determine a second speech text from the speech text of the scene category based on a first similarity between the speech text of the scene category and the first speech text;

[0008] Based on the second similarity between the first speech text and the second speech text, a target speech text to be pushed is determined.

[0009] In a second aspect, the present application further provides a dialogue processing device, the device comprising:

[0010] A prediction module, used to predict the speech of the first text content and output the first speech text;

[0011] An identification module, configured to identify a scene category corresponding to the first text content if there is a preset type of content in the first speech text;

[0012] An acquisition module, configured to filter out a second speech text from the speech text of the scenario category based on a first similarity between the speech text of the scenario category and the first speech text;

[0013] A determination module is used to determine a target speech text to be pushed based on a second similarity between the first speech text and the second speech text.

[0014] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method in any of the above embodiments are implemented.

[0015] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method in any of the above embodiments are implemented.

[0016] In a fifth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the method in any of the above embodiments are implemented.

[0017] It can be seen that based on the scheme in the embodiment of the present application, when performing dialogue processing, after performing speech prediction processing on the first text content to obtain the first speech text, if there is a preset type of content in the first speech text, it also identifies the scene category corresponding to the first text content, and determines the second speech text from the speech text of the scene category based on the first similarity between the speech text of the scene category and the first speech text, and determines the target speech text to be pushed based on the second similarity between the first speech text and the second speech text, that is, in the process of determining the target speech text to be pushed, on the basis of performing speech prediction processing, if the first speech text obtained by speech prediction still contains the preset type of content, it further adjusts the speech text corresponding to the scene category of the first speech text in combination with the first similarity and the second similarity, so that refined dialogue processing can be achieved in combination with the scene category, thereby improving the accuracy of the target speech text finally obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIG1 is a schematic diagram of an application scenario of a conversation processing method in an embodiment;

[0019] Figure 2 A flowchart of a method for processing a conversation in one embodiment;

[0020] Figure 3 A schematic diagram of a process for obtaining a first speech text in an embodiment;

[0021] Figure 4 A schematic diagram of the principle of training to obtain a first model in one embodiment;

[0022] Figure 5 A schematic diagram of a process of training a first model in one embodiment;

[0023] Figure 6 A schematic diagram of the principle of training to obtain a second model in one embodiment;

[0024] Figure 7 A schematic diagram of a process of training a second model in one embodiment;

[0025] Figure 8 A schematic diagram of a process for determining a target speech text based on a first speech text and a second speech text in one embodiment;

[0026] Fig. 9 A schematic diagram of a process for determining candidate speech texts in one embodiment;

[0027] Fig.10 A schematic diagram of a process for determining a target speech text based on a first speech text and a second speech text in another embodiment;

[0028] Fig.11 A schematic diagram of a process for determining a target speech text in another embodiment;

[0029] Fig.12 A schematic diagram of a process for determining a target speech text in another embodiment;

[0030] Fig.13 is a flowchart of a conversation processing method in a specific example;

[0031] Fig.14 is a structural block diagram of a data processing device in one embodiment;

[0032] Fig.15 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0034] The following embodiments of the technical solution of the present application will be described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" in the specification and claims of this application and the above-mentioned figure descriptions and any variations thereof are intended to cover non-exclusive inclusions.

[0036] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "multiple" is more than two, unless otherwise clearly and specifically defined.

[0037] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0038] In the description of the embodiments of the present application, the term "multiple" refers to more than two (including two). Similarly, "multiple groups" refers to more than two groups (including two groups), and "multiple pieces" refers to more than two pieces (including two pieces).

[0039] In the field of artificial intelligence technology, when it is necessary to determine the speech text to be returned to the user in intelligent question and answering, the speech text to be returned is usually determined by comprehensively analyzing the user's input content and the context of the conversation with the user.

[0040] At present, when determining the speech text, it is usually processed in combination with the speech model. For example, after pre-training the pre-trained model through a large-scale language model, the pre-trained model is continuously reinforced based on manually annotated data to fine-tune the pre-trained model and finally obtain the speech model that can be used in the intelligent question-answering process. However, the current speech model still has inaccurate output problems, and this inaccurate output is also commonly referred to as hallucination output. Especially in highly domain-specific scenarios, due to the diversity of domains, knowledge in different technical fields and / or applications is gradually becoming more refined. If there is a lack of relevant refined expectations during the model training process, the speech model will not learn the instructions of the field accurately, which will affect the accuracy of the output results of the speech model.

[0041] Moreover, with the widespread application of current technologies, the speed of data change is accelerating. Taking computer financial technology as an example, data even changes by the second. In the process of training the speech model, due to its large structure, this model needs to obtain user data in advance for training, but the real-time performance cannot be guaranteed. Even if more user data is obtained for training, in actual application, due to the change of user data, the results of the model output will be biased and the accuracy will not be high. Therefore, for dialogue systems in specific fields, it is extremely important to achieve refined and controllable information.

[0042] After research, it was found that the reason why some important information in the speech text output by the speech model is inaccurate is that in order to meet the needs of information refinement, all scene information usually needs to be considered. For example, the features of the input speech model need to include all parameter features related to the scene. However, due to the complexity and diversity of information, if the information of all scenes is to be included, the input information will be too long, resulting in a decrease in the control ability of the model information and inaccurate output speech information. Increasing the structure of the model layer can improve the control ability to a certain extent, but the output speed cannot meet the real-time requirements. Therefore, in the process of determining the speech text, after performing speech prediction processing to obtain a preliminary speech text (such as obtaining the first speech text), if necessary, such as when there is a preset type of content in the first speech text, the scene category can be further determined, and adjustments can be made in combination with the scene category to improve the accuracy of the final speech text and meet the real-time requirements.

[0043] Based on this, an embodiment of the present application provides a dialogue processing method to solve one or more of the problems mentioned above.

[0044] The dialogue processing method provided in the embodiment of the present application can be applied in the application environment as shown in FIG1. ​​The first terminal 10 and the second terminal 20 communicate through the network, and the first terminal 10 and the second terminal 20 can communicate through the server 30, wherein the first terminal 10 can be a user terminal, the second terminal 20 can be a terminal used by an agent, and the server 30 can be an ICC (Internet call center) server. The user can use the intelligent answering service deployed on the server 30 by accessing the server 30 through the first terminal 10, or the agent can log in to the server 30 through the second terminal 20 and actively initiate a call to the first terminal 10 to provide the first terminal 10 with an intelligent answering service.

[0045] In the process of the server 30 providing the first terminal 10 with intelligent answering service, the first terminal 10 transmits the user's input content to the server 30, the server 30 determines that the speech text of the first terminal 10 needs to be returned, and transmits the obtained speech text to the first terminal 10. It can be understood that when the input content provided by the first terminal 10 to the server 30 is in the form of voice or the like, the server 30 converts the obtained speech text into the corresponding audio content or the like, and then transmits it to the first terminal 10. Among them, the first terminal 10 and the second terminal 20 can be, but are not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 30 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0046] refer to Figure 2 As shown, the embodiment of the present application provides a data processing method, which is described by taking application in a server 30 as an example, wherein the method includes:

[0047] Step S201: perform speech prediction on the first text content and output the first speech text.

[0048] The first text content can be the text content input by the user, and can be the text content input by the user in the current conversation. The current conversation refers to the most recent conversation among multiple rounds of conversations between the user and the customer service agent or the intelligent robot on the server. The first text content refers to the text content of the most recent user. For example, if user A and intelligent robot B have multiple rounds of conversations, and the conversation contents in chronological order are: A: 1***; B: 2***; A: 3***, B: 4***, A: 5***, then the first text content refers to "5***".

[0049] Among them, the method of obtaining the first text content is not limited. When interacting with the user in text mode, the text content input by the user and transmitted to the server can be directly determined as the first text content. The text content input by the user can be the text content input by the user through a device such as a keyboard, or it can be the text content obtained by converting the collected voice into text after collecting the user's voice through a voice collection device, but it is not limited to this. When interacting with the user in audio mode, it can be the text content obtained by converting the audio information input by the user and transmitting it to the server. It can be understood that in other embodiments, the first text content can also be obtained by other methods.

[0050] There is no limitation on the method of performing speech prediction processing on the first text content. For example, a speech prediction model obtained through training is used to perform speech prediction processing on the first text content and output a first speech text.

[0051] Step S202: When there is a preset type of content in the first speech text, identify the scene category corresponding to the first text content.

[0052] Preset type content refers to the content of preset type information, which can be determined based on the need for scene refinement of specific scenarios. Taking the field of computer financial technology as an example, the preset type content can include numerical content. In a specific example, the preset type content can include content of amount data. There is no limit to the method for determining whether the numerical value is amount data. It can be determined in combination with whether the text content after the numerical value is an amount unit, or in combination with the context before and after the numerical value, but is not limited to this.

[0053] There is no limitation on the method of identifying the scene category corresponding to the first text content, such as extracting keywords from the first text content and identifying the scene category based on the extracted keywords. For example, an intention scene recognition model obtained through training can be used to identify the scene category corresponding to the first text content, but it is not limited to this.

[0054] Step S203: Based on the first similarity between the speech text of the scene category and the first speech text, determine the second speech text from the speech text of the scene category.

[0055] In the embodiment of the present application, a speech text may be pre-set for each scenario category, and there may be multiple speech texts. These speech texts may be manually input, or they may be multiple speech texts with the highest degree of association with the scenario category evaluated for several provided speech texts in actual technical scenario applications, or they may be speech texts corresponding to the scenario category set by other means, and the embodiment of the present application does not limit this.

[0056] The first similarity between the first speech text and each speech text of the identified scene category can be calculated respectively, and based on the calculated first similarity, the second speech text can be determined from the speech text of the scene category. The type of the first similarity is not limited, for example, in some embodiments, the first similarity can be cosine similarity.

[0057] The specific method of determining the second speech text from the speech text of the scene category based on the first similarity is not limited, and can be set according to actual needs. In some embodiments, the second speech text may be the speech text in the scene category, which has the greatest first similarity with the first speech text. Among them, there may be only one second speech text with the greatest first similarity, or more than two, and the specific value can be set based on actual needs. In some embodiments, the second speech text may also be the speech text in the scene category, which has the first similarity with the first speech text within a preset first similarity range.

[0058] Step S204: Determine the target speech text to be pushed based on the second similarity between the first speech text and the second speech text.

[0059] Among them, the second similarity between the first speech text and the second speech text reflects the similarity between the second speech text and the first speech text, and the specific type of the specific second similarity is not limited. In some embodiments, the type of the second similarity can be the same as the type of the first similarity, such as cosine similarity. In some embodiments, the type of the second similarity can also be different from the type of the first similarity mentioned above. Taking the first phase velocity as cosine similarity as an example, the second similarity can be the edit distance. In this way, the target speech text to be pushed can be determined by combining two different types of distances, thereby further reducing the deviation caused by screening by considering only one type of distance, and further improving the accuracy of the target speech text obtained on this basis.

[0060] It can be seen that based on the scheme in the embodiment of the present application, when performing dialogue processing, after performing speech prediction processing on the first text content to obtain the first speech text, if there is a preset type of content in the first speech text, it also identifies the scene category corresponding to the first text content, and determines the second speech text from the speech text of the scene category based on the first similarity between the speech text of the scene category and the first speech text, and determines the target speech text to be pushed based on the second similarity between the first speech text and the second speech text, that is, in the process of determining the target speech text to be pushed, on the basis of performing speech prediction processing, if the first speech text obtained by speech prediction still contains the preset type of content, it further adjusts the speech text corresponding to the scene category of the first speech text in combination with the first similarity and the second similarity, so that refined dialogue processing can be achieved in combination with the scene category, thereby improving the accuracy of the target speech text finally obtained.

[0061] Among them, the speech prediction is performed on the first text content, and the method of outputting the first speech text is not limited. For example, the speech prediction is only performed on the first text content. In an embodiment of the present application, in order to improve the accuracy of the predicted first speech text, the speech prediction can be further combined with other data on the basis of the first text content.

[0062] Accordingly, in some embodiments, Figure 3 As shown, the above step S202 of performing speech prediction on the first text content and outputting the first speech text may include:

[0063] Step S2021: Determine basic user information and conversation data according to the user identifier corresponding to the first text content.

[0064] The first text content obtained is content provided by a user of the first terminal, so the first text content corresponds to the user identifier of the user. The user identifier corresponding to the first text content can be obtained, and the user identifier uniquely identifies a user.

[0065] User basic information refers to information related to the basic attributes of the user corresponding to the user identifier. The basic types of user basic information are not limited, such as user category (VIP user or ordinary user), user registration time, user permissions, age, etc. It should be understood that based on the specific technical scenarios and different user data content, the obtained user basic information may also be different, and may also have more information related to the user's basic attributes. The embodiments of the present application do not impose specific restrictions on this.

[0066] Conversation data refers to the data of the conversation content of users in the process of multiple rounds of conversation, such as the above "A: 1***; B: 2***; A: 3***, B: 4***, A: 5***."

[0067] Step S2022: Perform speech prediction based on the first text content, user basic information and conversation data, and output the first speech text.

[0068] There is no limit to the method of predicting speech based on the first text content, basic user information and conversation data. In some embodiments, the first text content, basic user information and historical conversation data can be fused to obtain data to be processed, and speech prediction can be performed on the data to be processed to output the first speech text. There is no limit to the specific fusion processing method. For example, the first text content, basic user information and historical conversation data can be spliced ​​to obtain data to be processed, but it is not limited to this. In other embodiments, speech prediction can also be performed based on the first text content, basic user information and conversation data in other ways.

[0069] Based on the method of the embodiment of the present application as described above, in the process of performing speech prediction processing on the first text content, the basic information of the user of the first text content and the conversation data are combined to perform speech prediction, thereby improving the accuracy of the first text content obtained by speech prediction, and on this basis further improving the accuracy of the target speech text finally obtained.

[0070] In some embodiments, performing speech prediction based on the first text content, basic user information, and conversation data and outputting the first speech text includes:

[0071] The first text content, user basic information and conversation data are input into the first model, and the first speech text is output.

[0072] Accordingly, when performing speech prediction based on the first text content, user basic information and conversation data, it can be performed through a first model, which can also be called a speech prediction model in some embodiments.

[0073] The first model can be obtained through training, wherein the principle diagram of obtaining the first model through training is as follows: Figure 4 As shown, combined Figure 4 , Figure 5 As shown, in some embodiments, the training method of the first model includes:

[0074] Step S501: Encode the first sample data to obtain a coding vector; the first sample data includes basic user information of the sample user and conversation data of the sample user, and the conversation data of the sample user includes text content of the sample user and text content of the agent.

[0075] In some embodiments, the first sample data for training the first model may include basic user information of the sample user and conversation data of the sample user, and the conversation data of the sample user includes text content of the sample user and text content of the agent. For example, the user A and the intelligent robot B have multiple rounds of conversations, and the conversation contents in chronological order are: A: 1***; B: 2***; A: 3***, B: 4***, A: 5***, then the text content of the sample user includes 1***, 3***, and 5***, and the text content of the agent includes 2***, and 4***.

[0076] There is no limitation on the manner of encoding the first sample data, as long as the first sample data can be encoded into a vector form.

[0077] Step S502: modify the encoding value corresponding to the text content of the sample user in the encoding vector to a preset value to obtain a first sample label.

[0078] The encoding value corresponding to the user text content in the encoding vector refers to the encoding value obtained by performing vector encoding on the text content of the sample user in the conversation data of the sample user.

[0079] The specific content of the preset value is not limited, as long as it can reflect the difference from the normal coding value. For example, in some embodiments, the preset value can be set to -100.

[0080] Step S503: training the first model based on the encoding vector and the first sample label.

[0081] It can be understood that during the training process of the first model, the loss is calculated based on the difference between the predicted output of the first model for the encoding vector and the first sample label, and the first model is optimized based on the calculated loss until the training end condition is reached, such as model convergence. The specific method of training the model can be any existing training method or any training method that may appear in the future, and the embodiments of the present application do not make specific restrictions.

[0082] Based on the method of the embodiment of the present application as described above, the first model can be combined to perform speech prediction processing on the data to be processed, and in the training process of the first model, the first sample label of the first sample data is obtained after modifying the encoding value corresponding to the text content of the sample user in the encoding vector of the first sample data to a preset value, that is, the first sample label of the first sample data only considers the text content of the agent, and only learns the text content provided by the agent, so as to use the first model obtained thereby to perform speech prediction, which can improve the accuracy of the first speech text obtained by the speech prediction, and on this basis can also further improve the accuracy of the target speech text finally obtained.

[0083] There is no limitation on the method of identifying the scene category corresponding to the first text content. For example, keywords in the first text content may be extracted and the scene category may be determined based on the extracted keywords. For another example, the first text content may be subjected to scene recognition by the second model to obtain the scene category corresponding to the first text content. In some embodiments, the second model may also be referred to as a scene recognition model.

[0084] Taking the case where the second model is used to perform scene recognition on the first text content as an example, the step S203 of identifying the scene category corresponding to the first text content includes:

[0085] The second model is used to perform intent recognition on the first text content, and a scene category corresponding to the first text content is output.

[0086] The second model can be obtained through training, wherein the principle diagram of the training process of the second model is as follows: Figure 6 As shown, combined Figure 6 , Figure 7 As shown, the training method of the second model in some embodiments includes:

[0087] Step S701: Obtain second sample data and corresponding scene category labels.

[0088] In some embodiments, the second sample data includes the user's single-round conversation content. The single-round conversation content refers to the conversation content input by the user in a certain round of conversation with the user, such as the above 1***, 3***, 5***, that is, 1***, 3***, 5*** are separate second sample data, a total of 3 sample data.

[0089] There is no limit to the way of setting the scene category label, for example, it can be obtained through manual labeling, but is not limited thereto.

[0090] Step S702: training the second model based on the second sample data and the scene category label.

[0091] The specific method for training the second model can be any existing training method or any training method that may appear in the future, and the embodiments of the present application do not make specific limitations.

[0092] Based on the method of the embodiment of the present application as described above, the scene category of the first text content can be identified in combination with the second model, and in the training process of the second model, the training is based on the second sample data and the scene category label of the second sample data, so that the accuracy of the second speech text determined based on the scene category can be improved. Wherein, in the case where the second sample data includes the content of a single round of conversation, the training is based on the user's single round of conversation content and the corresponding scene category label, so that the second model obtained thereby can accurately identify the scene category corresponding to the single round of conversation, so that the accuracy of the second speech text determined based on the scene category can be improved, and on this basis, the accuracy of the target speech text finally obtained can also be further improved.

[0093] In some embodiments, reference Figure 8 As shown, the above step 204 of determining the target speech text to be pushed based on the second similarity between the first speech text and the second speech text may include:

[0094] Step S801: Based on the second similarity between the first speech text and the second speech text, determine a candidate speech text from the second speech text and the first speech text.

[0095] Based on the second similarity between the first speech text and the second speech text, the first speech text may be used as a candidate speech text, or the second speech text may be used as a candidate speech text, which is specifically determined based on the second similarity between the first speech text and the second speech text.

[0096] Step S802: Determine the target speech text to be pushed to the user based on the candidate speech texts.

[0097] In the process of determining the candidate speech text from the second speech text and the first speech text based on the second similarity between the first speech text and the second speech text, the similarity threshold can be combined for determination.

[0098] Among them, the second speech text is determined from the speech text of the scenario category. There can be only one or N of them, where N is a positive integer greater than 1. The value of N is not limited as long as it is greater than 2, and can be set based on the actual technical scenario needs.

[0099] Accordingly, in some embodiments, if there is only one second speech text, the first speech text can be determined as a candidate speech text if the second similarity between the second speech text and the first speech text is greater than a similarity threshold; and the second speech text can be determined as a candidate speech text if the second similarity between the second speech text and the first speech text is less than or equal to the similarity threshold.

[0100] In some embodiments, the number of the second speech texts determined is N, where N is a positive integer greater than 1. Fig. 9 As shown, the above-mentioned determining the candidate speech text from the second speech text and the first speech text based on the second similarity between the first speech text and the second speech text may include:

[0101] Step S901: Based on the second similarity between the first speech text and the second speech text, determine a target second speech text from N second speech texts.

[0102] The second similarity between the second speech text and the first speech text reflects the similarity between the second speech text and the first speech text determined based on the second phase velocity.

[0103] The target second speech text determined from the N second speech texts may be one of the second speech texts. In some embodiments, the target second speech text may be the second speech text with the greatest similarity to the first speech text. Taking the second similarity as the edit distance as an example, since the smaller the edit distance, the greater the similarity, the target second speech text may be the second speech text with the smallest edit distance to the first speech text, and the number of the determined second speech texts may be 1.

[0104] Step S902: When the second similarity between the target second speech text and the first speech text is greater than the similarity threshold, the first speech text is determined as a candidate speech text.

[0105] Among them, the similarity threshold can be set according to actual technical needs, as long as it can distinguish the similarity between the first speech text and the second speech text.

[0106] Step S903: When the second similarity between the target second speech text and the first speech text is less than or equal to the similarity threshold, the target second speech text is determined as a candidate speech text.

[0107] Based on this embodiment, the second similarity between the determined target second speech text and the first speech text is greater than the preset threshold, indicating that the second similarity between the determined target second speech text and the first speech text is very high, and the first speech text is a speech text that can conform to the historical dialogue text. Taking the first speech text as a candidate speech text can make the text expression content of the target speech text more accurate, further improving the accuracy of the obtained target speech text. The second similarity between the determined target second speech text and the first speech text is less than or equal to the preset threshold, indicating that the second similarity between the determined target second speech text and the first speech text is not high, that is, the text content of the user's current round may have changed relative to the context, and the determined target second speech text is a speech text that is more able to conform to the current round of dialogue. Therefore, taking the determined target second speech text as a candidate speech text can be more consistent with the accuracy of the current round of dialogue, further improving the accuracy of the obtained target speech text.

[0108] Among them, the above-mentioned first similarity and second phase velocity can be different types of similarities, so that the final target speech text can be comprehensively determined by combining different types of similarities to further improve the accuracy of the obtained target speech text. The similarity type of the first similarity and the second similarity is not limited, as long as the two are different. In some embodiments, the first similarity can be cosine similarity, and the second similarity can be edit distance. It is understandable that in other embodiments, the similarity type of the first similarity and the second similarity can also be set in other ways.

[0109] Among them, in the process of determining the target speech text to be pushed to the user based on the candidate speech text, in some application scenarios, the candidate speech text can be directly used as the target speech text to be pushed to the user. In some application scenarios, when the first speech text contains data content with high requirements for data accuracy, that is, when the first speech text includes preset type content, the candidate speech text is further processed to obtain the target speech text pushed to the user.

[0110] Accordingly, in some embodiments, reference Fig.10 As shown, the above step S802 of determining the target speech text to be pushed to the user based on the candidate speech texts may include:

[0111] Step S8020: Perform content replacement processing on the preset type content in the candidate speech text to obtain the target speech text.

[0112] Thus, by performing content replacement processing on the preset type of content in the candidate speech text, the target speech text finally obtained contains the data after the content replacement of the preset type of content, thereby improving the accuracy of the target speech text further obtained.

[0113] Among them, when performing content replacement processing on the preset type content in the candidate speech text, it can be replaced with the target data content corresponding to the scene category based on the user identifier corresponding to the first text content.

[0114] The specific method of content replacement is not limited. In some embodiments, reference Fig.11 As shown, step S8020 of performing content replacement processing on the preset type content in the candidate speech text to obtain the target speech text may include:

[0115] Step S1101: When the candidate speech text is the first speech text, the first speech text is provided to the agent terminal.

[0116] In some embodiments, the first speech text may be directly provided to the agent terminal. In some embodiments, the preset type of content in the first speech text may be subjected to data desensitization processing before being provided to the agent terminal. Data desensitization processing refers to deforming the data to achieve reliable protection of the data. The specific method of data desensitization is not limited, such as replacement, rearrangement, encryption, truncation, masking, etc., but is not limited thereto.

[0117] Step S1102: Obtain the third speech text returned by the agent terminal, where the third speech text is the text obtained after the agent terminal replaces the preset type of content in the first speech text.

[0118] In the case where the first speech text is directly provided to the agent terminal, the third speech text obtained is the text obtained after the agent terminal replaces the preset type content in the first speech text. In the case where the preset type content in the first speech text is data-massified and provided to the agent terminal, the third speech text obtained is the text obtained after the agent terminal replaces the preset type content after data-massified processing in the first speech text.

[0119] There is no limit to the way the agent terminal replaces the preset type content. For example, the agent terminal queries the data content corresponding to the user identifier and replaces the data content in the first speech text with the queried data content, such as the preset type content or the preset type content after data desensitization, to obtain the third speech text.

[0120] Step S1103: Use the third speech text as the target speech text.

[0121] Based on this embodiment, when the candidate speech text is the first speech text, it means that the first speech text that needs to be returned to the user contains content that needs to be refined, or content that needs to improve the accuracy of the preset type content. Therefore, the first speech text is provided to the agent terminal, and the agent terminal provides the specific content of the preset type content, so that the content data of the preset type in the third speech text finally obtained is accurate, and the first speech text is a speech text that can conform to the historical dialogue text, so that the text expression content of the target speech text and the content data of the preset type content in the text are more accurate, further improving the accuracy of the obtained target speech text.

[0122] In some embodiments, reference Fig.12 As shown, in step S8020, replacing the preset type of content in the candidate speech text to obtain the target speech text may include:

[0123] Step S1201: When the candidate speech text is the target second speech text, the target data content corresponding to the scene category is obtained according to the user identifier corresponding to the first text content.

[0124] There is no limitation on the method of acquiring the target data content corresponding to the user identification and the scene category. For example, the target data content corresponding to the user identification and the scene category may be searched from a database.

[0125] Step S1202: Use the target data content to replace the preset type content of the target second speech text to obtain the target speech text.

[0126] Based on this embodiment, when the candidate speech text is the target second speech text, it means that there is content in the first speech text that needs to be returned to the user that needs to be refined, or content that needs to improve the accuracy of the preset type content, and the determined target second speech text is not very similar to the first speech text, that is, the text content of the user's current round may have changed relative to the context. Therefore, by directly obtaining the target data content corresponding to the user identifier and the scene category, and replacing the specific content of the preset type content in the determined target second speech text, the content data of the preset type content in the target speech text finally obtained is accurate, and the determined target speech text is a speech text that can meet the current round of dialogue, thereby making the text expression content of the target speech text and the content data of the preset type in the text more accurate, further improving the accuracy of the obtained target speech text.

[0127] Based on the above-mentioned embodiments, combined with Fig.13 As shown, the following is illustrated with reference to a specific application example. In the following description, it is illustrated by taking the example of obtaining a first speech text through a speech prediction model, identifying a scene category through an intention scene recognition model, and the similarity including a first similarity and a second similarity.

[0128] Before performing specific conversation processing, you can first conduct training to obtain a speech prediction model and an intent scenario recognition model.

[0129] When training to obtain a speech prediction model, the user's basic information and historical conversation data can be obtained, and the user's basic information and historical conversation data can be combined to obtain first sample data, encode the first sample data, obtain a coding vector, modify the coding value corresponding to the user's text content in the coding vector to a preset value, such as -100, obtain the first sample label corresponding to the first sample data, and then train the initial speech prediction model based on the coding vector and the corresponding first sample label to obtain a speech prediction model. The initial speech prediction model can be a large language model obtained by pre-training.

[0130] When training to obtain the intent recognition model, the user's single-round conversation content can be used as the second sample data, and the corresponding scene category label can be obtained. The second sample data and the scene category label can be combined into training data. For example, if the user's single-round conversation content is "process early settlement for me", the corresponding scene classification label is "early settlement". Then, based on the second sample data and the scene category label, the initial intent recognition model is trained to obtain the intent scene recognition model. Among them, the initial intent scene recognition model can be a pre-trained model, and the initial intent scene recognition model and the initial speech prediction model can have the same model architecture.

[0131] After training the speech prediction model and the intention scene recognition model, you can deploy the speech prediction model and the intention scene recognition model to the intelligent dialogue scene to conduct intelligent dialogue. Fig.13 As shown, in specific applications, the user's basic information and conversation content are first obtained, where the conversation content includes historical conversation data and the current first text content (i.e., the user's current question, such as: confirming the early settlement), and the user's basic information is combined with the conversation content to obtain the data to be processed.

[0132] Use the speech prediction model to perform speech prediction processing on the data to be processed, and obtain the first speech text, that is, predict the answer to the question corresponding to the user's current question. For example, the first speech text obtained after the speech prediction processing may be: "Dear, please confirm again that we will process the early settlement for you. The amount is: XXXX. The money will be deducted from your XXXX. Is the balance sufficient? Are you sure to deduct the money?"

[0133] Check whether there is a preset type of content in the first speech text, for example, whether it contains important information such as the amount. If not, directly use the first speech text as the target speech text, output it and push it to the user.

[0134] If there is a preset type of content in the first speech text, the scene category of the first speech text is identified through the intention scene recognition model. For example, the scene category of the first speech text in the above example is: early settlement.

[0135] Then, among the multiple speech texts of the scene category, multiple second speech texts with the greatest first similarity to the first speech text are searched, wherein a database, such as a milvus vector database, can be set up, and multiple speech texts of each scene category are stored in the database.

[0136] After obtaining the speech text of the scene category, the first similarity (such as cosine similarity) between the first speech text and the queried speech text can be calculated, and the first N second speech texts with the largest first similarity to the first speech text can be obtained. Taking the pre-N as 3 as an example, the above-obtained 3 second speech texts can be respectively:

[0137] 'Dear, please confirm again. We will process early settlement for you. The amount is: [clear_money]. The money will be deducted from your [bank]. Is the balance sufficient? Are you sure to deduct the money?'

[0138] 'Confirm again, the amount [clear_money] will be deducted from your [bank] repayment bank card for early settlement. Is the balance sufficient?'

[0139] 'Dear, please confirm again, the amount of [clear_money] for early settlement will be deducted from your [bank] repayment bank card. Is the balance sufficient?'

[0140] After obtaining N second speech texts, the second similarity (e.g., edit distance) between each second speech text and the first speech text is calculated, and the second speech text with the largest second similarity is found, which is referred to as the target second speech text in this embodiment of the application. For example, the target second speech text with the shortest edit distance is:

[0141] 'Dear, please confirm again. We will process early settlement for you. The amount is: [clear_money]. The money will be deducted from your [bank]. Is the balance sufficient? Are you sure you want to deduct the money?'

[0142] In order to further ensure the accuracy of the obtained speech text, it is determined whether the second similarity between the target second speech text and the first speech text is greater than a preset threshold (such as whether the edit distance is less than a preset distance, the smaller the edit distance, the higher the similarity).

[0143] If it is greater than the preset threshold, the preset type content (such as amount) in the first speech text will be desensitized and provided to the agent terminal. The agent of the agent terminal will find the user's actual data and fill it in with the user's actual data. That is, the preset type content after data desensitization in the first speech text will be replaced with the user's actual data, so as to obtain the third speech text and provide it to the user.

[0144] If it is less than the preset threshold, the target data content corresponding to the user identifier and the scene category is obtained, for example, by calling the sub-scenario business interface to obtain the target data content corresponding to the user identifier and the scene category, and the retrieved target data content is used to replace the preset type content in the determined second speech text to obtain the target speech text. For example, the target speech text finally determined is:

[0145] "Dear, please confirm again. We will process the early settlement for you. The amount is: 1111. The money will be deducted from your xxxx. Is the balance sufficient? Are you sure you want to deduct the money?"

[0146] Based on the scheme of the embodiment of the present application as described above, it combines the recognition of scene categories on the basis of speech prediction to differentiate the scene into various sub-scenes, and then determines the speech text based on the determined scene category, reducing the interference of too complex customer information and improving the efficiency of obtaining real-time information. Moreover, for the text of speech prediction containing preset type content (such as the first speech text), the target data content corresponding to the user identifier and the scene category can be obtained in real time to replace the preset type content in the candidate speech text, ensuring the real-time and accuracy of the preset type content (such as the amount). In the case of irregular replies, that is, the scene where the first speech text predicted by the speech has a low similarity with the target second speech text determined based on the scene, the preset type content can be desensitized and handed over to the real person seat for processing, further ensuring the accuracy of the speech text returned to the user.

[0147] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0148] Based on the same inventive concept, the embodiment of the present application also provides a conversation processing device for implementing the conversation processing method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more conversation processing device embodiments provided below can refer to the limitations on the conversation processing method above, and will not be repeated here.

[0149] In one embodiment, Fig.14 As shown, a dialogue processing device is provided, comprising: a prediction module 141, a recognition module 142, an acquisition module 143, and a determination module 144, wherein:

[0150] Prediction module 141, used for predicting the speech of the first text content and outputting the first speech text;

[0151] The identification module 142 is used to identify the scene category corresponding to the first text content if there is a preset type of content in the first speech text;

[0152] An acquisition module 143 is used to filter out a second speech text from the speech text of the scene category based on a first similarity between the speech text of the scene category and the first speech text;

[0153] The determination module 144 is used to determine the target speech text to be pushed based on the second similarity between the first speech text and the second speech text.

[0154] In some embodiments, the prediction module 141 is used to determine the user's basic information and conversation data based on the user identifier corresponding to the first text content; perform speech prediction based on the first text content, the user's basic information and the conversation data, and output the first speech text.

[0155] In some embodiments, the prediction module 141 is used to input the first text content, user basic information and conversation data into the first model, and output the first speech text.

[0156] In some embodiments, the device also includes a first model training module, which is used to encode the first sample data to obtain a coding vector; the first sample data includes basic user information of the sample user and conversation data of the sample user, and the conversation data of the sample user includes text content of the sample user and text content of the agent; the coding value corresponding to the text content of the sample user in the coding vector is modified to a preset value to obtain a first sample label; and the first model is trained based on the coding vector and the first sample label.

[0157] In some embodiments, the recognition module 142 is used to perform intent recognition on the first text content through the second model, and output a scene category corresponding to the first text content.

[0158] In some embodiments, the device further includes a second model training module, which is used to obtain second sample data and a scene category label of the second sample data; and train the second model based on the second sample data and the scene category label.

[0159] In some embodiments, the determination module 144 is used to determine a candidate speech text from the second speech text and the first speech text based on the second similarity between the first speech text and the second speech text; and determine a target speech text to be pushed to the user based on the candidate speech text.

[0160] In some embodiments, the number of second speech texts is N, where N is a positive integer greater than 1; the determination module 144 is used to determine a target second speech text from the N second speech texts based on the second similarity between the first speech text and the second speech text; when the second similarity between the target second speech text and the first speech text is greater than a similarity threshold, the first speech text is determined as a candidate speech text; when the second similarity between the target second speech text and the first speech text is less than or equal to the similarity threshold, the target second speech text is determined as a candidate speech text.

[0161] In some embodiments, the determination module 144 is used to perform content replacement processing on preset types of content in the candidate speech text to obtain the target speech text.

[0162] In some embodiments, the determination module 144 is used to provide the first speech text to the agent terminal when the candidate speech text is the first speech text; obtain the third speech text returned by the agent terminal, the third speech text being the text obtained after the agent terminal replaces the preset type of content in the first speech text; and use the third speech text as the target speech text.

[0163] In some embodiments, the determination module 144 is used to obtain the target data content corresponding to the scene category based on the user identifier corresponding to the first text content when the candidate speech text is the target second speech text; and use the target data content to replace the preset type content of the target second speech text to obtain the target speech text.

[0164] Each module in the above-mentioned dialogue processing device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0165] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.15As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data, such as speech texts of different scenario categories, basic user information, user historical conversation data, etc. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a conversation processing method is implemented.

[0166] Those skilled in the art will understand that Fig.15 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0167] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of the dialog processing method in any of the embodiments described above when executing the computer program.

[0168] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the dialog processing method in any of the embodiments described above are implemented.

[0169] In one embodiment, a computer program product is provided, including a computer program, which implements the steps of the dialog processing method in any of the embodiments described above when executed by a processor.

[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0171] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0172] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for processing a conversation, characterized in that: The method comprises: Perform speech prediction on the first text content and output the first speech text; In the case where there is a preset type of content in the first speech text, identifying the scene category corresponding to the first text content; Determine a second speech text from the speech text of the scene category based on a first similarity between the speech text of the scene category and the first speech text; Based on the second similarity between the first speech text and the second speech text, a target speech text to be pushed is determined.

2. The method according to claim 1, characterized in that The step of performing speech prediction on the first text content and outputting the first speech text includes: Determining basic user information and conversation data according to a user identifier corresponding to the first text content; Based on the first text content, the user basic information and the conversation data, speech prediction is performed to output the first speech text.

3. The method according to claim 2, characterized in that The performing speech prediction based on the first text content, the user basic information and the conversation data, and outputting the first speech text, includes: The first text content, the user basic information and the conversation data are input into a first model, and the first speech text is output.

4. The method according to claim 3, characterized in that The training method of the first model includes: Encode the first sample data to obtain a coding vector; the first sample data includes basic user information of the sample user and conversation data of the sample user, and the conversation data of the sample user includes text content of the sample user and text content of the agent; Modify the encoding value corresponding to the text content of the sample user in the encoding vector to a preset value to obtain a first sample label; The first model is trained based on the encoding vector and the first sample label.

5. The method according to claim 1, characterized in that The identifying the scene category corresponding to the first text content includes: The intent of the first text content is recognized by a second model, and a scene category corresponding to the first text content is output.

6. The method according to any one of claims 1 to 5, characterized in that: The step of determining a target speech text to be pushed based on a second similarity between the first speech text and the second speech text includes: Based on a second similarity between the first speech text and the second speech text, determining a candidate speech text from the second speech text and the first speech text; Based on the candidate speech texts, a target speech text to be pushed to the user is determined.

7. The method according to claim 6, characterized in that The number of the second speech texts is N, where N is a positive integer greater than 1; the determining of candidate speech texts from the second speech texts and the first speech text based on the second similarity between the first speech text and the second speech text includes: Determine a target second speech text from N second speech texts based on a second similarity between the first speech text and the second speech text; In the case where the second similarity between the target second speech text and the first speech text is greater than a similarity threshold, determining the first speech text as a candidate speech text; When the second similarity between the target second speech text and the first speech text is less than or equal to a similarity threshold, the target second speech text is determined as a candidate speech text.

8. The method according to claim 6, characterized in that The step of determining the target speech text to be pushed according to the candidate speech texts includes: The preset type of content in the candidate speech text is replaced to obtain the target speech text.

9. The method according to claim 8, characterized in that The step of performing content replacement processing on the preset type of content in the candidate speech text to obtain the target speech text includes: In the case where the candidate speech text is the first speech text, providing the first speech text to the agent terminal; Acquire a third speech text returned by the agent terminal, wherein the third speech text is a text obtained after the agent terminal replaces the preset type of content in the first speech text; The third speech text is used as the target speech text.

10. The method according to claim 8, characterized in that The step of performing content replacement processing on the preset type of content in the candidate speech text to obtain the target speech text includes: In the case where the candidate speech text is the target second speech text, acquiring target data content corresponding to the scene category according to the user identifier corresponding to the first text content; The target data content is used to replace the preset type content of the target second speech text to obtain the target speech text.

11. A dialogue processing device, characterized in that: The device comprises: A prediction module, used to predict the speech of the first text content and output the first speech text; An identification module, configured to identify a scene category corresponding to the first text content if there is a preset type of content in the first speech text; An acquisition module, configured to filter out a second speech text from the speech text of the scenario category based on a first similarity between the speech text of the scenario category and the first speech text; A determination module is used to determine a target speech text to be pushed based on a second similarity between the first speech text and the second speech text.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 10 are implemented.

13. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images