Information processing method, device, storage medium, and program product

Abstract

The application provides an information processing method, device, storage medium and program product. The method comprises the following steps: obtaining input information of a user in a current session with a merchant; processing the input information through a plurality of intent models to obtain reply information corresponding to each intent model; sorting a plurality of reply information corresponding to the plurality of intent models according to historical sessions of the user with the merchant; and outputting target reply information according to the sorted plurality of reply information. The application is suitable for personalized scenarios, can accurately output answers to user consultation questions, and thus meets user needs and improves user service experience.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an information processing method, device, storage medium and program product. Background Technology

[0002] With the continuous development of artificial intelligence technology, intelligent customer service robots have acquired knowledge base question-and-answer and dialogue capabilities, enabling them to answer user questions relatively smoothly and accurately.

[0003] However, when answering user questions, the aforementioned customer service robots typically respond according to a pre-set question-and-answer engine priority. This cannot guarantee that the set priority applies to different users, leading to chaotic response scenarios (such as answering irrelevant questions) and thus reducing the accuracy of the responses.

[0004] Therefore, the existing technology has limitations in its applicable scenarios, cannot accurately output answers to different users' questions, and thus cannot meet users' personalized needs, resulting in a poor user experience. Summary of the Invention

[0005] The main objective of this application is to provide an information processing method, device, storage medium, and program product that are suitable for personalized scenarios, can accurately output answers to different user inquiries, thereby meeting users' personalized needs and improving their service experience.

[0006] In a first aspect, embodiments of this application provide an information processing method, including:

[0007] The system obtains the user's input information in the current conversation with the merchant, processes the input information through multiple intent models, and obtains the response information corresponding to each intent model.

[0008] Based on the user's and merchant's historical conversations, sort the multiple response messages corresponding to multiple intent models;

[0009] Output the target response information based on the sorted responses.

[0010] Optionally, based on the user's historical conversations with the merchant, the multiple response messages corresponding to multiple intent models are sorted, including:

[0011] The user's historical conversations with the merchant are used to rank intents and determine the priority of each intent.

[0012] The multiple response messages are sorted according to the priority of each intent model.

[0013] Optionally, the method further includes:

[0014] Obtain the historical conversations of at least one user, and extract multi-turn dialogues from the historical dialogue data of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models;

[0015] For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue.

[0016] Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical conversations. The network model is then trained to obtain the intent ranking model.

[0017] Optionally, based on the user's historical conversations with the merchant, the multiple response messages corresponding to multiple intent models are sorted, including:

[0018] Input the user's historical conversations with the merchant and multiple response messages into a predefined response sorting model to determine the priority of the multiple response messages;

[0019] The multiple response messages are sorted according to their priority. If at least two response messages have the same priority, the at least two response messages are sorted according to the merchant's intent model strategy.

[0020] The intent model strategy is used to represent the priority of each pre-set intent model.

[0021] Optionally, the method further includes:

[0022] If no historical conversation between the user and the merchant is found, then the merchant's intent model strategy is obtained;

[0023] According to the intent model strategy, the multiple response messages are sorted to obtain multiple sorted response messages.

[0024] Optionally, based on the sorted responses, output the target response information, including:

[0025] For any response information, determine the response type corresponding to the response information; the response type is used to indicate at least one of the following: whether to modify the response information before output, whether to add recommendation information to the user after the response information, and whether to output the added recommendation information after concatenation;

[0026] Based on the sorting results and the determined response type, output the target response information.

[0027] Optionally, the response type corresponding to the response information is determined, including at least one of the following:

[0028] Based on whether the response information contains preset words, determine whether to modify the response information before outputting it;

[0029] Based on the context information corresponding to the response information, determine whether to add recommendation information after the response information is output;

[0030] If multiple recommended answers are added, the added recommended information will be concatenated and output.

[0031] Optionally, based on the sorted responses, output the target response information, including:

[0032] Based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the time interval in the time limit condition; and / or, based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the number of recommendations in the recommendation flow limit condition;

[0033] The method further includes:

[0034] If the user receives the next input, stop outputting the target response for the current input and start searching for the target response for the next input.

[0035] Optionally, the method further includes:

[0036] Obtain the configuration information input by the merchant, which includes at least one of the following: time limit conditions, recommendation traffic limiting conditions, and intent model strategy.

[0037] Secondly, embodiments of this application also provide an information processing method, including:

[0038] The user's input information is obtained, and the input information is processed through multiple intent models to obtain multiple response information;

[0039] Find the user's historical sessions and sort the multiple response messages according to the found historical sessions;

[0040] Output the target response information based on the sorted responses.

[0041] Thirdly, embodiments of this application provide an electronic device, including:

[0042] At least one processor; and

[0043] A memory that is communicatively connected to the at least one processor;

[0044] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, cause the electronic device to perform the method described in any of the above aspects.

[0045] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method described in any of the above aspects.

[0046] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the methods described in any of the above aspects.

[0047] The information processing method, device, storage medium, and program product provided in this application can obtain user input information in the current session with a merchant, process the input information through multiple intent models respectively, and obtain response information corresponding to each intent model; further, based on the user's historical session with the merchant, sort the multiple response information corresponding to the multiple intent models; and further, output target response information based on the sorted multiple response information. This application can call multiple intent models to process user input information to obtain multiple response information, then use the historical session between the user and the merchant to sort the multiple response information, and then output target response information based on the sorted multiple response information. It is applicable to different personalized scenarios, without the need for mandatory rule sorting and outputting answers only after sorting, avoiding inconsistencies caused by simultaneous triggering of answers, and also avoiding confusion in the output response information due to different users' different perceptions of priority, which would prevent maintenance personnel from having a basis for maintenance. Therefore, this application is applicable to personalized scenarios, can accurately output answers to user inquiries, thereby meeting user needs and improving the user's service experience. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 This is a schematic diagram illustrating an application scenario according to an embodiment of this application;

[0050] Figure 2 This is a schematic diagram illustrating the interface of a terminal device provided in an embodiment of this application;

[0051] Figure 3 A flowchart illustrating an information processing method provided in an embodiment of this application;

[0052] Figure 4 A flowchart illustrating a question-and-answer system processing method is provided as an embodiment of this application.

[0053] Figure 5 A schematic diagram of a merchant configuration interface provided in an embodiment of this application;

[0054] Figure 6 A flowchart illustrating another information processing method provided in an embodiment of this application;

[0055] Figure 7 This is a schematic diagram of the structure of an information processing device provided in an embodiment of this application;

[0056] Figure 8 This is a schematic diagram of another information processing device provided in an embodiment of this application;

[0057] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0058] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0059] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0060] First, let me explain the terms used in this application:

[0061] Intent model: refers to an engine that can provide search skills based on corresponding intents, searching for response information corresponding to user and merchant input information.

[0062] Question-and-answer system: This refers to a dialogue system configured by the merchant, which uses an e-commerce website to provide users with product information and suggestions, helping them decide what products to buy, and simulating a salesperson assisting a customer in completing the purchase process. The output response information can be controlled, such as through business rule restrictions, message appending and recommendation management, and timing control, with the output response being the final answer.

[0063] In e-commerce contexts, "user" usually refers to consumers or buyers.

[0064] Through recent developments, intelligent customer service robots have acquired knowledge base question-and-answer and dialogue capabilities, enabling them to answer user questions relatively smoothly and accurately. Specifically, they can automatically generate and output corresponding response information based on user input. To ensure that the output response information meets user needs as much as possible, multiple response information can be searched or generated based on multiple intent models (i.e., question-and-answer engines) and output to the user.

[0065] In some technologies, responses from multiple intent models can be sorted and output to the user according to a pre-defined priority. However, pre-defined priorities are difficult to adjust for various personalized scenarios, and it is also difficult to ensure that the set priorities are applicable to different users or merchants. This can easily lead to confusing response scenarios (such as answering the wrong question), thereby reducing the accuracy of the response.

[0066] Specifically, outputting responses according to pre-defined priorities is a strong rule-based approach. It heavily relies on the absolute accuracy of business outputs to ensure error-free responses. This makes it difficult to reconcile differing opinions on priorities among users or merchants as the number of intent models in the question-and-answer system increases, leading to maintenance difficulties later on. Secondly, this approach is highly susceptible to misinterpretation by high-priority engines. For example, if task-oriented intent models are initially prioritized over frequently asked questions (FAQ) intent models, and a user inquires whether a garment is suitable for adults, the response might be triggered by a low-relevance intent from the task-oriented dialogue. Consequently, the actual high-confidence FAQ answer might not be triggered and output, leaving maintenance personnel unable to determine if a better answer exists.

[0067] Another approach is to schedule and trigger results in parallel using multiple intent models. However, this approach can easily lead to inconsistent output answers in the scenario, resulting in a chaotic response scenario and thus reducing the accuracy of the response.

[0068] Specifically, multiple intent models are scheduled to trigger results in parallel. Different intent models can correspond to different push methods. This may result in the intelligent customer service conversation page outputting a general answer to the user, but sending a text message containing other answers according to other rules. As a result, the answer seen by the user on the customer service conversation page and the text message received are inconsistent, which leads to inaccurate output response information.

[0069] Therefore, the above methods have limited applicability and cannot accurately output answers to different users' questions, thus failing to meet users' personalized needs and resulting in a poor user experience.

[0070] In view of this, this application provides an information processing method that, in order to be applicable to personalized scenarios and to output more accurate answers to different user inquiries, can call multiple intent models to search for response information based on a user's input information, and use historical dialogue records generated in the communication between the user and the merchant to extract content that the user may be interested in, and sort the searched response information according to the user's habitual way of viewing the content that may be of interest to the user, and then output the final answer.

[0071] Compared to the above-mentioned method of outputting answers according to a pre-set priority, the embodiments of this application can simultaneously trigger multiple intent models (such as engines), call their respective skills in parallel to search for answers, and then sort them based on the response information (such as skill answers) and user behavior data (such as historical conversations). This is applicable to different personalized scenarios and does not require mandatory rules that could lead to confusion in the output answers due to different opinions of different merchants or users on priority, thus failing to provide a basis for maintenance personnel to maintain the system.

[0072] Compared to the parallel scheduling methods described above, this application embodiment can schedule individual skills in parallel. Unlike parallel scheduling of triggered answers, it sorts each answer after it is obtained, and then outputs the answers after sorting them according to the user's historical conversations. This is suitable for different personalized scenarios, and outputting answers after sorting avoids inconsistencies caused by simultaneous triggering of answers, which would lead to chaotic output response information and prevent maintenance personnel from providing a basis for maintenance.

[0073] Therefore, sorting based on users' historical behavior data ensures that the answers users receive are probabilistically consistent with their questions, making it easier to obtain answers that satisfy users' personalized needs and making it suitable for personalized scenarios.

[0074] Figure 1 This is a schematic diagram illustrating an application scenario involved in an embodiment of this application. For example... Figure 1 As shown, the information processing method provided in this application embodiment can be applied to, for example... Figure 1In the application scenario shown, the scenario includes: terminal device 101 and e-commerce platform 102. Specifically, e-commerce platform 102 can obtain the dialogue records between terminal device 101 and merchants on the e-commerce platform. These dialogue records can include the current user's input information on the merchant's customer service chat page and the user's historical chat history on the merchant's customer service chat page. Further, based on the input information obtained by terminal device 101 in the current chat with the merchant, multiple intent models are used to search for answers to the input information, resulting in multiple response information corresponding to multiple intent models. For example, intent model 1 outputs response information 1, intent model 2 outputs response information 2, ..., intent model N outputs response information N. According to the historical chat history, the response information 1, response information 2, ..., response information N output by intent model N are sorted. Then, the sorted response information 1, response information 2, ..., response information N output by intent model N are sorted.

[0075] In this customer service conversation page, the target response information is output. The target response information can be the response information 1 output by intent model 1, response information 2 output by intent model 2, ..., response information N output by intent model N, which are output in sequence. It can also be the first few response information selected according to the order, or one or more of the selected response information to be output are modified before the answer is output. There is no specific limitation here. The corresponding target response information can be output according to different scenarios.

[0076] It should be noted that an intent model can output one or more response messages, without specific limitations. Multiple response messages can be combined into a single card, which can be used to carry multiple response messages. The card can be presented in the form of text, list, link, image, etc., without specific limitations.

[0077] In addition, when outputting the target response information, the output can be based on the time interval preset by the merchant to avoid time sequence disorder. Furthermore, if the output target response information consists of multiple additional recommended answers, the multiple additional recommended answers can be merged into a single card before being output, thus avoiding message overload and affecting the reading of effective answers.

[0078] It should be noted that the above-described steps—processing the user's current conversation with the merchant using multiple intent models to obtain multiple response messages, then sorting these responses using historical conversations between the user and the merchant, and finally outputting the execution steps for the target response based on the sorted responses—are merely illustrative. This process could also be achieved by splitting the data into more platforms and more modules; this application does not specifically limit this approach. Furthermore, the acquired historical behavior data, input information, and configuration information are all authorized by the user or merchant.

[0079] Optionally, the terminal device 101 can obtain the input information in the current session with the merchant and the generated historical sessions through the following methods: Figure 2 This is a schematic diagram illustrating the interface of a terminal device provided in an embodiment of this application, such as... Figure 2 As shown, the terminal device can identify the dialogue history from the customer service chat page of a merchant that the user has opened, forming a historical chat history and recognizing the currently entered input information. For example, if a user clicks the "Customer Service" button on a merchant's homepage via touch, they are redirected to the merchant's customer service chat page. The user enters information in the "input window" or "dialog box" on the customer service chat page, such as asking, "What activities are currently available?" The terminal device then obtains the input information. Furthermore, obtaining the user's historical chat history on the customer service chat page can be achieved through at least two of the following methods:

[0080] Method 1: This involves searching and identifying content prior to input on the customer service conversation page as historical conversation data, including dialogue data between the user and the merchant, as well as product interactions. The dialogue data can include the user's historical input information and responses to historical input information. Product interactions can refer to the user's interaction with the intelligent customer service, such as the intelligent customer service sending different product links to the user, and the user browsing different product pages by clicking on the links.

[0081] Method 2: The terminal device uploads the acquired input information to the cloud and stores it. The e-commerce platform can then access historical sessions from the cloud through authorization.

[0082] Therefore, the information processing method provided in this application embodiment can call multiple intent models to process user input information to obtain multiple response information. Then, it can sort the multiple response information using the historical conversation between the user and the merchant, and output the target response information based on the sorted multiple response information. It is applicable to different personalized scenarios, and can make the output response information more in line with user preferences. It reduces the situation where the output response information is chaotic due to forced rule sorting, or the response information is triggered at the same time, resulting in inconsistent answers output by different push methods. Thus, it can output the answer to the user's inquiry more accurately and improve the user's service experience.

[0083] The technical solutions of this application will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0084] For example, Figure 3 This is a flowchart illustrating an information processing method provided in an embodiment of this application. This embodiment can be applied to any device capable of data processing, such as a server. Figure 3 As shown, the method may include:

[0085] S301. Obtain the user's input information in the current session with the merchant, process the input information through multiple intent models respectively, and obtain the response information corresponding to each intent model.

[0086] In this embodiment, a Natural Language Understanding (NLU) model (a general term for all methods, models, or tasks that support machine understanding of text content) is used to identify the user's intent in the current conversation with the merchant (e.g., on the merchant's current customer service conversation page). The intent model is then determined to correspond to the intent (e.g., triggering the appropriate engine module). This intent model can include, but is not limited to: general / industry-specific intent models, FAQ intent models, marketing and sales guide intent models, and custom scenario intent models. The intent model is used to provide response information based on the corresponding intent.

[0087] Examples include general / industry-specific packages such as greetings; FAQs such as size information; marketing guides such as promotional and coupon information; and custom scenarios such as the differences between this merchant's product 1 and other merchants' product 2.

[0088] After one or more engine modules are triggered, the corresponding skills are scheduled in parallel to respond. First, the input information is processed by entity recognition or semantic analysis based on the merchant's custom configuration to identify relevant keywords. Then, based on the identified entities or keywords, the response information is searched through one or more engine modules. Since the intent triggers the engine, but the engine cannot find response information related to the entity, the response may have two types of results, as shown in Table 1.

[0089] Table 1

[0090] type Examples / Explanations User-visible response types Directly providing answers, linking questions, and proactively recommending products. User-invisible response types Skill exit, forced command, command to call external components

[0091] For example, if an engine cannot find an answer, it can output empty information and filter this empty information before sorting the multiple responses corresponding to multiple intent models, so that the sorting process can be performed on engines (intent models) that have answers.

[0092] Among these methods, obtaining user input information during the current session with the merchant can be combined with... Figure 2 As shown, it will not be elaborated further here.

[0093] S302. Based on the user's and the merchant's historical conversations, sort the multiple response messages corresponding to multiple intent models.

[0094] In this embodiment, the historical conversations can include historical dialogue data. By using this historical dialogue data to sort multiple responses corresponding to multiple intent models, the answers obtained by the user are probabilistically consistent with the consumer's question, making it easier to obtain satisfactory answers. Specifically, the sorting method can involve inputting historical dialogue data into a sorting model and outputting a sorting result.

[0095] Optionally, sorting the multiple response information corresponding to multiple intent models may include:

[0096] Based on user historical behavior data, multiple response messages corresponding to multiple intent models are sorted. This historical behavior data can include at least one of the following: historical conversations and page interaction records. Historical conversations can include historical dialogue data, product interaction records, etc.

[0097] The sorting is based on page interactions. These interactions can refer to user interactions with public pages provided by the merchant's platform, such as refund or coupon pages. These interactions indicate user interest, allowing for the recommendation of responses tailored to their needs. Product interaction records reveal the content viewed by the user at the merchant, including browsing time and frequency. For example, the price and features of viewed products can reflect user preferences, leading to personalized recommendations. Historical conversations can be based on set time intervals, such as conversations older than 24 hours, without specific limitations. Each time period's historical conversations can contain multiple rounds of dialogue. The current conversation refers to the dialogue with the merchant during the current time period. For instance, if the current conversation is in the morning, conversations from this morning constitute one time period. If five questions and five answers were asked this morning, a sixth question asked during the morning, even if not consecutively with the previous five, can still be considered part of the current conversation.

[0098] Specifically, page interactions combined with conversation data can constitute user behavior data. By sorting multiple response information corresponding to multiple intent models using user behavior data, it can be applied to personalized scenarios, and the recommended response information is relatively accurate and meets the needs of personalized users.

[0099] Optionally, if the historical conversations include product interaction records, these records can refer to the user's browsing history at the merchant. That is, it can also be done by obtaining the user's browsing history at the merchant and sorting multiple response information corresponding to multiple intent models; or, it can be done by combining historical conversation data and browsing history to sort multiple response information corresponding to multiple intent models, i.e., sorting multiple response information corresponding to multiple intent models based on historical conversations.

[0100] Specifically, the sorting model can be based on the priority of the intent model, and then sort the corresponding multiple response information based on the priority of the intent model; it can also be based on the priority of the response information, and then output the target response information according to the priority of the response information; or, based on the priority of the response information and the priority of the intent model predefined by the merchant, sort the multiple response information and then output the target response information.

[0101] S303. Based on the sorted multiple response information, output the target response information.

[0102] In this embodiment of the application, multiple sorted response information can be output directly in sequence according to their order, or one or more response information with the highest sort order can be selected for output.

[0103] Optionally, to avoid scenarios such as excessive reply content and card segments, message overload affecting the reading of core content, disordered reply information sequence, excessive product recommendations, and generic and inaccurate reply information, the sorted multiple reply information can be controlled. For example, by building a product Q&A central control capability, problems such as message overload, replacement of invalid generic answer output, message disorder, and excessive addition of product and knowledge recommendations can be avoided.

[0104] Specifically, controls can include business rule restrictions, replacement of useless or sensitive content, and message appending recommendation management. For example, business rule restrictions could include the length of the output content (e.g., the content presented on a card cannot exceed half the size of the customer service conversation page) and the time interval between output responses (e.g., a 10-second interval between every two responses). Replacement of useless or sensitive content could involve determining whether the information to be output is generic and meaningless (e.g., outputting useless information like "Please see the product details page" could be replaced with relevant content from the details page, rather than just a simple "Please see the product details page"). Controls could also include whether to forcibly replace inappropriate content (e.g., replacing sensitive words with other content or filtering them out). Message appending recommendation management could involve managing recommended content (e.g., not recommending in scenarios where a recommendation has already been made in the previous round or where direct transfer to human assistance is possible; allowing recommendations but not exceeding 5 cards, with trigger time limits for output cards; not recommending in after-sales scenarios, and merging recommended content).

[0105] Therefore, this application embodiment can obtain the user's input information in the current session with the merchant, process the input information through multiple intent models respectively, and obtain the response information corresponding to each intent model; further, based on the user's historical session with the merchant, sort the multiple response information corresponding to the multiple intent models; further, based on the sorted multiple response information, output the target response information. This application can call multiple intent models to process the user's input information to obtain multiple response information, then use the historical session between the user and the merchant to sort the multiple response information, and then output the target response information based on the sorted multiple response information. It is applicable to different personalized scenarios, without the need for mandatory rule sorting and outputting the answer after sorting, avoiding the situation where answers are triggered simultaneously and resulting in inconsistent answers, and also avoiding the output of answers being chaotic due to different users' different views on priority, which would not provide a basis for maintenance personnel. Therefore, this application is applicable to personalized scenarios, can accurately output the answer to the user's inquiry, thereby meeting the user's needs and improving the user's service experience.

[0106] Taking the above considerations into account, the information processing method provided in this application can be implemented based on a question-and-answer system, which can be configured on the server side, and is not specifically limited here. For example, Figure 4 A flowchart illustrating a question-and-answer system processing method is provided for an embodiment of this application, such as... Figure 4 As shown, the question-answering system can, based on input information such as a question (Query), utilize NLU intent recognition (including: general packages / industry packages, custom scenarios, positive and negative examples, marketing guides, etc.), then perform entity recognition, and combine merchant-customized configurations. Through context maintenance, it can invoke skills from the intent model. These skills can include multi-round troubleshooting skills, activity zone skills, attribute question-answering skills, etc. The system then sorts or selects skill answers: first, it filters out skills with no answers, then it sorts the candidate answers by quality (which can be combined with the priority of the intent model), and finally generates an answer.

[0107] Optionally, to ensure that the final selected high-confidence answer meets the business requirements of the platform and merchants, a message control module has been added. This module handles business rule restrictions, message addition and recommendation management, and timing control, better balancing the requirements of the platform and merchants. For example, by identifying the scenario, answer, and target audience, it determines whether to recommend similar questions, products, promotional activities, or marketing phrases / selling points, generating answers accordingly. Based on the generated answers, the system outputs the answers through the aforementioned restrictions on recommended messages, business rule restrictions, and message timing control. Therefore, to meet platform requirements, business operations can intervene in answer output under special circumstances, and business rules are scalable. To improve merchant inquiry conversion rates, the intelligent customer service (question-answering system) recommendation message control module can determine whether to add a recommendation message based on the user's contextual intent. It recommends suitable products, coupons, or text content based on the identified scenario, target audience, and answer reply. Merchants can also decide to enable or disable this function. To make the communication between users and intelligent customer service more human-like and to avoid merchants outputting answers all at once in a short period, multiple message intervals can be set to make the intelligent customer service dialogue more human-like.

[0108] In this embodiment, based on the multi-engine parallel scheduling of skill answer sorting, the answer quality is comprehensively sorted by combining the user's historical consultation data with the merchant, so that the answers obtained by the user are more in line with the user's historical habits; at the same time, based on the original answer output, a message control module that can be influenced by the business is set up, which comprehensively considers whether the answer output meets the requirements of the platform and the merchant, so that the answer is more compliant and meets the merchant's needs.

[0109] Optionally, based on the user's historical conversations with the merchant, the multiple response messages corresponding to multiple intent models are sorted, including:

[0110] The user's historical conversations with the merchant are input into a preset intent ranking model to determine the priority of each intent model;

[0111] The multiple response messages are sorted according to the priority of each intent model.

[0112] In this embodiment, the intent ranking model is applied to scenarios where intent models are ranked. If a user has historical behavioral data (including historical conversation data and / or browsing behavior data) with the merchant, then based on this historical behavioral data, the ranking model outputs the priority of the intent models, thereby ranking the corresponding response information to meet personalized needs. Determining the priority of each intent model can involve ranking all intent models by priority, with intent models that do not output an answer having a priority of 0.

[0113] Optionally, sorting multiple response messages corresponding to multiple intent models based on the user's historical conversations with the merchant can include: inputting the user's historical conversations with the merchant and the identifiers of multiple intent models into a preset intent sorting model to determine the priority of each intent model. Determining the priority of each intent model here involves prioritizing the triggered multiple intent models.

[0114] Optionally, the method further includes:

[0115] Obtain the historical sessions of at least one user, and extract multi-turn dialogues from the historical sessions of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models;

[0116] For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue.

[0117] Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical conversations. The network model is then trained to obtain the intent ranking model.

[0118] The intent ranking model can be implemented in at least three ways:

[0119] Method 1: The intent ranking model is a single model, and the historical conversations include historical dialogue data, which is an intent ranking model trained based on historical dialogue data.

[0120] Optionally, the training process for this intention ranking model can be as follows:

[0121] Acquire historical dialogue data of at least one user, and extract multi-turn dialogues from the historical dialogue data of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models;

[0122] For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue.

[0123] Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical dialogue data. The network model is then trained to obtain the intent ranking model.

[0124] The operation information here can be user replies or other information, such as the operation information of the user clicking a link to jump to another page.

[0125] Specifically, historical dialogue data between multiple users and the merchant is obtained, and at least one round of dialogue is extracted from the historical dialogue data of multiple users. Each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models.

[0126] Based on the context information corresponding to each round of dialogue, determine the intent model corresponding to the actual response information used by the user and the order in which the actual response information is used from the multiple response information corresponding to the user input information in each round of dialogue.

[0127] Based on the order of the intent models corresponding to the user's input information in each round of dialogue, the priority of the intent models corresponding to each user's historical conversations with the merchant is determined.

[0128] The network model is trained by using each user's historical conversations with the merchant and the priority of the corresponding intent model as training samples to obtain the intent ranking model.

[0129] In this embodiment of the application, for the user input information in each round of dialogue, the multiple response information output by the intelligent customer service to the user is determined. Based on the following, it is determined which intention model corresponding to the response information actually used by the user. For example, if the user clicks on the card pushed by the intelligent customer service, or if the user answers the question asked by the intelligent customer service, the priority of the intention model is determined based on the response information used by the user: the priority of the used model is higher than that of the unused model. If both are used, the priority of the model processed first is higher than that of the model processed later.

[0130] Based on user click behavior, the order of intent models corresponding to user input information in each round of dialogue is determined. Then, the priority of intent models corresponding to all user input information for that user is comprehensively judged, i.e., the priority of the intent model corresponding to that user for that merchant. Then, the historical conversations of each user and the priority of the corresponding intent models are used as training samples to train a ranking model. The input of the ranking model is the user's historical conversations, and the output is the priority of the intent model corresponding to the user.

[0131] Similar to method 1, an intent ranking model can also be trained based on historical browsing data or page interaction data (i.e., page interaction records). The training samples are replaced with historical browsing data or page interaction data. The specific training process is similar to method 1 and will not be described in detail here.

[0132] Method 2: This intent ranking model includes two sub-models, and the historical conversation includes historical dialogue data and historical browsing data (i.e., product interaction records). That is, the two sub-models are the first sub-intent ranking model and the second sub-intent ranking model. The first sub-intent ranking model is an intent ranking model based on historical conversations, and the second sub-intent ranking model is an intent ranking model based on browsing behavior (i.e., historical browsing data). Finally, the output results of the two sub-intent models are weighted. For example, the output results of the first sub-intent ranking model are all assigned a weight of 0.65, and the output results of the second sub-intent ranking model are all assigned a weight of 0.35. Furthermore, the weights for the priority of the output results can be: the highest priority is assigned 1, the second highest priority is assigned 0.9, and so on down, etc.

[0133] For example, the output of the first sub-intent ranking model is: Intent Model 1 > Intent Model 3 > Intent Model 2; the output of the second sub-intent ranking model is: Intent Model 1 > Intent Model 2 > Intent Model 3. Finally, the priority weights of each intent model are: Intent Model 1 is 1*0.65+1*0.35=1, Intent Model 2 is 0.8*0.65+0.9*0.35=0.835, and Intent Model 3 is 0.9*0.65+0.8*0.35=0.865. Therefore, the output of the intent ranking model is Intent Model 1 > Intent Model 3 > Intent Model 2.

[0134] The training process for the first sub-intent ranking model can be found in the training process for the intent ranking model in Method 1, and will not be repeated here. The training process for the second sub-intent ranking model can be as follows:

[0135] Obtain the browsing history data of at least one user at the merchant, and identify each user's browsing information from the browsing history data of at least one user at the merchant;

[0136] Based on the browsing duration and / or number of views corresponding to each user browsing information, determine the intent model corresponding to each user's historical browsing data at the merchant and the order of these intent models.

[0137] Based on the order of each user's intent models at the merchant, the priority of each user's historical browsing data at the merchant is determined.

[0138] The network model is trained using each user's historical browsing data at the merchant and the corresponding intent model priority as training samples to obtain the intent ranking model. This intent ranking model is the second sub-intent ranking model.

[0139] For example, the longer a user spends browsing a merchant's content and the more times they browse, the more interested they are in the content. Therefore, ranking the intent models corresponding to a user's browsing behavior based on the browsing time and / or number of times the content is viewed can make the answers obtained by the user more consistent with their historical habits.

[0140] Similar to method 2, the two sub-models can also be a sub-model corresponding to historical dialogue data (i.e., the first sub-intent ranking model) and a sub-model corresponding to historical page interaction data (i.e., obtained from page interaction records) (i.e., the third sub-intent ranking model, an intent ranking model trained based on historical page interaction data); or, the two sub-models can also be a sub-model corresponding to historical browsing data (i.e., obtained from product interaction records) (i.e., the second sub-intent ranking model) and a sub-model corresponding to historical page interaction data (i.e., the third sub-intent ranking model). Furthermore, this intent ranking model can also include three sub-models: the first sub-intent ranking model, the second sub-intent ranking model, and the third sub-intent ranking model. Finally, the outputs of the three sub-intent models are weighted.

[0141] The specific training process is similar to that of Method 2, and will not be repeated here.

[0142] Method 3: This intent ranking model is a single model, trained based on at least two of the following: historical session data, historical browsing data, and historical page interaction data. The input data consists of at least two of the following: historical session data, historical browsing data, and historical page interaction data. The following detailed explanation uses historical session data and historical browsing data as input data as an example to illustrate the training of this intent ranking model. (Other input data follow a similar training process, but the input data differs, and will not be elaborated upon here.)

[0143] Optionally, the training process for this intention ranking model can be as follows:

[0144] Obtain at least one user's historical conversations with the merchant and at least one user's historical browsing data at the merchant;

[0145] Extracting multi-turn dialogues from the historical sessions of at least one user, wherein each turn of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models; and identifying each user browsing information from the historical browsing data of at least one user in the merchant's database;

[0146] For any given round of dialogue, the order of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue; and the order of intent models corresponding to each user's historical browsing data at the merchant is determined based on the browsing content, browsing duration, and browsing frequency corresponding to each piece of user browsing information.

[0147] Based on the order of multiple intent models in each round of dialogue for each user and the order of intent models corresponding to each user's historical browsing data at the merchant, the priority of the intent models for each user is determined through weighted processing.

[0148] Training samples are constructed based on the priority of the intent model corresponding to each user and the historical sessions. The network model is then trained to obtain the intent ranking model.

[0149] Specifically, at least one user’s historical conversations with the merchant and at least one user’s historical browsing data at the merchant are obtained;

[0150] Based on the context information corresponding to each round of dialogue, determine the intent model corresponding to the actual response information used by the user and the order of the actual response information used from the multiple response information corresponding to the user input information in each round of dialogue; and determine the intent model corresponding to each user's historical browsing data in the merchant and the order of the corresponding intent models based on the browsing duration and / or number of browsings corresponding to each piece of user browsing information.

[0151] Based on the order of the intent models corresponding to the user's input information in each round of dialogue, the priority of the intent models corresponding to each user's historical conversations with the merchant is determined.

[0152] Based on the order of multiple intent models in each round of dialogue for each user and the order of intent models corresponding to each user's historical browsing data at the merchant, the priority of the intent models for each user is determined through weighted processing.

[0153] Training samples are constructed based on the priority of the intent model corresponding to each user and the historical sessions. The network model is then trained to obtain the intent ranking model.

[0154] The process of determining the intent model corresponding to the user's actual response information and the order of these intent models from the multiple responses output by the user in each round of dialogue, based on the context information corresponding to the user's input information in each round of dialogue; and the process of determining the intent model corresponding to each user's historical browsing data at the merchant and the order of these intent models based on the browsing content, browsing duration, and / or number of browsing visits corresponding to the user's browsing information in each round of dialogue, can be found in Method 1 and Method 2 respectively, and will not be elaborated here.

[0155] In addition, based on the order of the intent models corresponding to the user input information in each round of dialogue and the order of the intent models corresponding to the user's historical browsing data at the merchant, the priority of the intent model corresponding to each user is determined through weighted processing. See the weighted process of the two sub-intent ranking models in Method 2, which will not be repeated here.

[0156] It should be noted that the above three methods are merely illustrative and do not impose specific limitations on the selection of training samples for the intention ranking model.

[0157] Optionally, based on the user's historical conversations with the merchant, the multiple response messages corresponding to multiple intent models are sorted, including:

[0158] Input the user's historical conversations with the merchant and multiple response messages into a predefined response sorting model to determine the priority of the multiple response messages;

[0159] The multiple response messages are sorted according to their priority. If at least two response messages have the same priority, the at least two response messages are sorted according to the merchant's intent model strategy.

[0160] The intent model strategy is used to represent the priority of each pre-set intent model.

[0161] For example, if at least two response messages have the same priority, the response messages corresponding to the multiple intent models are sorted by weighting based on the priorities of the multiple response messages and the priorities of the predefined intent models.

[0162] The input data for the response ranking model can be a concatenation of historical conversations and intent model identifiers corresponding to multiple response messages. This concatenated data is then input into the ranking model to obtain a score, which is used as the priority of the intent model. Alternatively, the input data can be a concatenation of historical conversations, intent model identifiers corresponding to multiple response messages, and conversation data from previous conversations within a preset time interval. This concatenated data is then input into the ranking model to obtain a score, which is used as the priority of the intent model. No restrictions are placed on the input data for the response ranking model.

[0163] In this embodiment, the response ranking model is applied to scenarios where response information is ranked. The input to the response ranking model is historical conversations, and the output is the priority of the corresponding response information.

[0164] It should be noted that this response ranking model is also applicable to the three methods mentioned above. The training samples consist of user historical behavior data (historical conversations and / or historical browsing data) and the corresponding priority of response information. The determination of the priority of response information is similar to that of the intent model described above, and will not be elaborated upon here.

[0165] Specifically, if a user has historical behavior data with the merchant, then based on this historical behavior data, the priority of the response information is output through the response ranking model. If at least two responses have the same priority, they can be weighted in combination with the priority of the merchant's predefined intent model, and then the responses can be ranked to meet personalized needs.

[0166] Therefore, by comprehensively judging users' historical behavior data with merchants, appropriate answers can be given, and it can be determined whether to add recommended products or messages for the current scenario. This achieves a balance between improving the user consultation experience and satisfying the merchant's inquiry conversion, avoiding user reading experience problems such as message overload, time sequence errors, and mechanical feeling, and improving satisfaction.

[0167] Optionally, the method further includes:

[0168] If no historical conversation between the user and the merchant is found, then the merchant's intent model strategy is obtained;

[0169] According to the intent model strategy, the multiple response messages are sorted to obtain multiple sorted response messages.

[0170] In this embodiment, the query method can be to find dialogue records based on user identifiers or to find dialogue records of the corresponding user based on all records. The predefined skill priorities are not arbitrarily configured, but are set based on the unique attributes of the target object, group objects, and other characteristics.

[0171] Specifically, if a user has no historical conversation history with a merchant, the system can prioritize responses based on the merchant's predefined intent model to output the correct answer. Alternatively, it can obtain historical conversations and / or browsing data from other merchants on the platform (e.g., similar products) or with other platforms, and use the aforementioned prioritization model to determine the priority of the intent model or the response information, thereby prioritizing the responses.

[0172] For example, considering both scenarios with and without historical behavioral data, the sorting can be based on the priority of two stages:

[0173] Phase 1: Prioritize users based on their historical consultation data with this merchant. For example, frequently asked questions can be sorted as follows: Size > Product Attributes > Activity Q&A > FAQ. If a user has such historical behavior data with this merchant, the answers will be output according to this sorting. This method of sorting based on users' historical behavior data makes the answers that users receive more likely to match the consumer's questions and makes it easier to get answers that satisfy them.

[0174] The second stage: If the user has no historical behavior data, the answers are sorted and output according to the skill priority set by the merchant. Therefore, in scenarios without historical behavior data, the sorting is based on the merchant's customized skill priority, which is a merchant-specific sorting method rather than a uniform and mandatory rule, thus providing more accurate answers to the user's questions.

[0175] Optionally, based on the sorted responses, output the target response information, including:

[0176] For any response information, determine the response type corresponding to the response information; the response type is used to indicate at least one of the following: whether to modify the response information before output, whether to add recommendation information to the user after the response information, and whether to output the added recommendation information after concatenation;

[0177] Based on the sorting results and the determined response type, output the target response information.

[0178] In this embodiment, the answer type includes at least one of the following: direct output answer type, answer conversion type, answer appending type, and answer concatenation type. Based on the answer type, the sorted multiple response information is directly output. The sorted multiple response information serves as the target response information, and / or the sorted response information is edited to output the target response information. The editing operation includes: answer replacement, appending recommended answers, and answer concatenation, such as whether to modify the response information before output, whether to append recommended information to the user after the response information, and whether to concatenate the appended recommended information before output.

[0179] Therefore, based on different types, the output answers can be controlled: directly output or edited. Editing can include rewriting, replacing, recommending, etc. This can meet the business requirements of merchants and platforms while avoiding user reading experience problems such as message overload, time sequence errors, and mechanical feeling, thereby improving user satisfaction.

[0180] Optionally, the response type corresponding to the response information is determined, including at least one of the following:

[0181] Based on whether the response information contains preset words, determine whether to modify the response information before outputting it;

[0182] Based on the context information corresponding to the response information, determine whether to add recommendation information after the response information is output;

[0183] If multiple recommended answers are added, the added recommended information will be concatenated and output.

[0184] In this embodiment of the application, for any one of the sorted response messages, the following steps are performed:

[0185] If the response information corresponding to the current order is compliant information (i.e., it does not contain preset words, such as sensitive words and / or contains useless information), then the answer type of the response information corresponding to the current order is determined to be the direct output answer type, which indicates that the direct output of the response information corresponding to the current order is supported; the compliant information is used to indicate information that does not contain sensitive words and / or does not contain useless information.

[0186] If the response information corresponding to the current order contains sensitive words and / or useless information, then the answer type of the response information corresponding to the current order is determined to be an answer conversion type, indicating that answer replacement is supported for the response information corresponding to the current order. Answer replacement can be performed by replacing non-compliant answers or by replacing useless information (e.g., please refer to the details page).

[0187] If, based on the context information corresponding to the response information in the current order, it is determined that a recommended answer will be added after the output of the response information in the current order, then the answer type of the response information in the current order is determined to be an answer addition type, which indicates that it is supported to add a recommended answer to the response information in the current order.

[0188] For the answer appending type, if there are multiple recommended answers appended, then the response information corresponding to the current order is determined to be an answer concatenation type, which indicates that it supports concatenating the recommended answers appended to the response information corresponding to the current order.

[0189] For example, after a skill triggers an answer, some answer types are preprocessed. For instance, answers containing sensitive content are modified with fallback text to avoid significant public opinion risks. Alternatively, if an answer contains multiple sub-cards, they are merged to prevent consumers from receiving multiple card messages. See Table 2 for details.

[0190] Table 2

[0191]

[0192] It should be noted that the above answer types are merely examples, and the answer types and corresponding sorting methods based on answer types can be expanded, allowing for flexibility in specific configurations.

[0193] Optionally, based on the sorted responses, output the target response information, including:

[0194] Based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the time interval in the time limit condition; and / or, based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the number of recommendations in the recommendation flow limit condition.

[0195] In this embodiment, multiple sorted responses can be controlled based on constraints to output the target response. Both time-limit and recommendation-based traffic-limiting conditions can be configured with on / off switches for merchants, who can then enable or disable them. Merchants can also configure time intervals, recommendation numbers, etc., to better balance the requirements of the platform and the merchant. The answer output type can include text, images, cards, instructions / events, etc., without specific limitations.

[0196] Enabling time-limit restrictions allows responses to be output sequentially at predefined time intervals, preventing issues like disordered timing. Enabling recommendation rate limiting allows responses to be output sequentially according to a predefined number of recommendations, preventing message overload. For example, a maximum of five cards can be recommended, each with a response. The response information can include recommendations such as similar questions, products, promotional activities, and key selling points.

[0197] Optionally, based on the sorted responses, output the target response information, including:

[0198] Based on the sorted responses, the target response is output according to the determined response type and constraints.

[0199] The restrictions include time-limited restrictions and / or recommended rate-limiting conditions.

[0200] Specifically, based on whether the response information contains preset words, it is determined whether the response information should be modified before output, and the modified or unmodified response information is taken as the target response information;

[0201] If, based on the context information corresponding to the response information, it is determined that a recommended answer should be added after the response information is output, then the recommended information is added to the user after the response information, and the response information is used as the target response information.

[0202] If there are multiple additional recommended answers, then the additional recommended information will be concatenated and output, and the additional recommended information will be used as the target response information.

[0203] Output the target response information sequentially according to the time intervals that meet the business requirements and time constraints; and / or,

[0204] Based on the recommended rate limiting conditions that meet the business requirements, the target response information is output sequentially until the number of recommendations in the recommended rate limiting conditions is reached.

[0205] Therefore, combining business requirement control with answer type-based control ensures that the response information output by the above control methods not only meets business requirements, but also meets personalized needs based on the management and control of response messages (rewriting, replacing, recommending, etc.). Furthermore, based on multiple capabilities, the output answers meet the scenarios, improving accuracy and thus enhancing user experience.

[0206] Specifically, it not only meets business requirements but also has control capabilities, such as message replacement, message appending recommendation (identifying the scenario based on the context and determining whether to append a recommended answer based on the scenario; this recommended answer can be selected from the sorted answers to be output), and message splicing (for the appending recommendation scenario: splicing multiple skill answers into a single card to avoid message overload).

[0207] Optionally, the method further includes:

[0208] If the user receives the next input, stop outputting the target response for the current input and start searching for the target response for the next input.

[0209] In this embodiment of the application, when interacting with the user, once a new question is presented on the customer service conversation page or once new input information is obtained from the user's current conversation with the merchant, the output of the answer to the question in the current round is immediately stopped, and the search for the answer to the new input information begins, thus avoiding problems such as time sequence disorder.

[0210] Optionally, the method further includes:

[0211] Obtain the configuration information input by the merchant, which includes at least one of the following: time limit conditions, recommendation traffic limiting conditions, and intent model strategy.

[0212] In this embodiment of the application, in order to improve the conversion rate of merchant inquiries, the question-and-answer system can determine whether to add a recommendation message based on the user's contextual intent, and recommend suitable products, coupons or text content based on the identified scenario, audience, and answer reply. Merchants can also decide to enable or disable the corresponding functions.

[0213] For example, see Figure 5 The diagram shows the merchant configuration interface. Merchants can enable / disable the "Recommendations" feature by clicking the on / off button on the configuration interface, allowing them to recommend suitable products, coupons, or text content based on the identified scenario, target audience, and answer responses. Once enabled, the question-and-answer system can obtain the merchant's configuration information, such as the enabled features. Merchants can also input settings like "the time interval between each response" and "the priority of the predefined intent model" on the configuration interface. The configuration content is not specifically limited and can be expanded. Then, based on the configuration information, the system can control the output of the aforementioned response information.

[0214] Therefore, this application constructs an intelligent customer service Q&A control center in the new retail field by combining multi-engine parallel scheduling capabilities with historical merchant data, and can support the expansion of business rules to make the robot respond to user input information more accurate and compliant with the solution.

[0215] Optionally, embodiments of this application also provide a specific solution based on an information processing method.

[0216] Figure 6 This is a flowchart illustrating another information processing method provided in an embodiment of this application; it can be applied to e-commerce platforms, travel platforms, food delivery platforms, entertainment / audio-visual platforms, payment platforms, etc., and the process includes:

[0217] S601. Obtain user input information, process the input information through multiple intent models respectively, and obtain multiple response information;

[0218] S602. Locate the user's historical sessions and sort the multiple response messages according to the found historical sessions corresponding to the user;

[0219] S603. Output the target response information based on the sorted response information.

[0220] The customer service chat page is not only with the merchant, but may also be with the entire platform or across platforms.

[0221] For example, taking a payment platform as an example, the system obtains the user's input information from the customer service chat page of the application corresponding to the payment platform. For instance, if the user inquires about a bill, the system determines the user's intent based on the bill information and calls multiple intent models to process the bill information, obtaining multiple response information corresponding to the intent models. For example, the time the bill occurred, the payee of the bill, whether the address of the payee needs to be provided, and whether any rights protection steps are required. The system then uses the user's historical chat history on the customer service chat page of the payment platform to sort the above response information and determine the target response information to be output. This ensures that the user receives the most appropriate answer to the question during the consultation process, thereby improving the user's satisfaction with the payment platform.

[0222] In food delivery platforms, the system obtains user input information from the customer service chat page of the corresponding application. For example, if a user inquires about which food delivery restaurants have promotions, the system determines the user's intent based on the inquiry and calls multiple intent models to process the inquiry, obtaining multiple response messages corresponding to these intent models. These responses may include a list of restaurants with promotions, the details of the promotions, the type of food delivery (e.g., noodles, rice), and whether a link to the relevant restaurant is needed. By utilizing the user's historical chat history on the food delivery platform's customer service chat page, the system sorts these response messages to determine the target response message for output. This ensures that the user receives the most appropriate answer to their question during the inquiry process, thereby improving their satisfaction with the food delivery platform.

[0223] The information processing method provided in this application can acquire user input information, process the input information through multiple intent models to obtain multiple response messages; further, it searches for the user's corresponding historical conversations and sorts the multiple response messages according to the found historical conversations; further, it outputs a target response message based on the sorted multiple response messages. This method is applicable to different personalized scenarios, eliminates the need for mandatory rule-based sorting, and avoids inconsistent responses due to simultaneous triggering of answers. It also avoids outputting confusing response messages due to different users' varying opinions on priority, thus preventing maintenance personnel from having a basis for maintenance. Therefore, it is suitable for personalized scenarios, can accurately output answers to user inquiries, thereby meeting user needs and improving the user's service experience.

[0224] Optionally, the multiple response messages are sorted according to the historical conversations corresponding to the user, including:

[0225] Based on the user's historical sessions, the priority of each intent model is determined according to the intent ranking model.

[0226] Based on the priority of each intent model, the multiple response messages corresponding to the multiple intent models are sorted.

[0227] Optionally, the method further includes:

[0228] Obtain historical sessions corresponding to at least one user, and extract at least one round of dialogue from the historical sessions of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models;

[0229] For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue.

[0230] Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical conversations. The network model is then trained to obtain the intent ranking model.

[0231] The operation information here can be user replies or other information, such as the operation information of the user clicking a link to jump to another page.

[0232] Specifically, at least one user’s historical conversation is obtained, and at least one round of dialogue is extracted from the historical conversation of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models;

[0233] Based on the context information corresponding to each round of dialogue, determine the intent model corresponding to the actual response information used by the user and the order in which the actual response information is used from the multiple response information corresponding to the user input information in each round of dialogue.

[0234] Based on the order of the intent models corresponding to the user's input information in each round of dialogue, the priority of the intent models corresponding to each user's historical sessions is determined.

[0235] The network model is trained by using each user's historical sessions and the corresponding intent model priority as training samples to obtain the intent ranking model.

[0236] Optionally, the multiple response messages are sorted according to the historical conversations corresponding to the user, including:

[0237] Input the user's historical conversations and multiple response messages into a predefined response sorting model to determine the priority of the multiple response messages;

[0238] The multiple response messages are sorted according to their priority. If at least two response messages have the same priority, the at least two response messages are sorted according to a predefined intent model strategy.

[0239] The intent model strategy is used to represent the priority of each pre-set intent model.

[0240] Optionally, the method further includes:

[0241] If the user's historical sessions are not found, the predefined intent model strategy corresponding to the platform is obtained;

[0242] According to the intent model strategy, the multiple response messages are sorted to obtain multiple sorted response messages.

[0243] Optionally, based on the sorted responses, output the target response information, including:

[0244] For any response information, determine the response type corresponding to the response information; the response type is used to indicate at least one of the following: whether to modify the response information before output, whether to add recommendation information to the user after the response information, and whether to output the added recommendation information after concatenation;

[0245] Based on the sorting results and the determined response type, output the target response information.

[0246] Optionally, the response type corresponding to the response information is determined, including at least one of the following:

[0247] Based on whether the response information contains preset words, determine whether to modify the response information before outputting it;

[0248] Based on the context information corresponding to the response information, determine whether to add recommendation information after the response information is output;

[0249] If multiple recommended answers are added, the added recommended information will be concatenated and output.

[0250] Optionally, based on the sorted responses, output the target response information, including:

[0251] Based on the sorted multiple responses, and according to the time intervals specified in the time limit conditions, at least a portion of the responses are sequentially output as the target response; and / or,

[0252] Based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the number of recommendations in the recommendation flow limiting conditions;

[0253] The method further includes:

[0254] If the user receives the next input, stop outputting the target response for the current input and start searching for the target response for the next input.

[0255] Optionally, the method further includes:

[0256] Obtain the configuration information input by the merchant on the platform, wherein the configuration information includes at least one of the following: time limit conditions, recommendation traffic limiting conditions, and intent model strategy;

[0257] Based on the configuration information, a processing strategy for multiple response messages corresponding to multiple intent models is determined to output the target response message.

[0258] The implementation principle and technical effects of the above embodiments are the same as those described above. Figure 3 The corresponding subordinate scheme implementation is similar and will not be described in detail here.

[0259] Corresponding to the above information processing method, embodiments of this application provide an information processing apparatus. Figure 7 This is a schematic diagram of the structure of an information processing device provided in an embodiment of this application. The device includes:

[0260] The first processing module 701 is used to obtain the user's input information in the current conversation with the merchant, process the input information through multiple intent models respectively, and obtain the response information corresponding to each intent model;

[0261] The second processing module 702 is used to sort multiple response information corresponding to multiple intent models according to the historical conversation between the user and the merchant;

[0262] The third processing module 703 is used to output the target response information based on the sorted multiple response information.

[0263] The information processing apparatus provided in this application embodiment can be used to perform the above-described... Figures 1 to 5 The technical solutions of the embodiments shown are similar in principle and in effect, and will not be described again here.

[0264] For example, embodiments of this application also provide an information processing apparatus. Figure 8 This is a schematic diagram of another information processing apparatus provided in an embodiment of this application. The apparatus includes:

[0265] The fourth processing module 801 is used to acquire user input information, process the input information through multiple intent models respectively, and obtain multiple response information;

[0266] The fifth processing module 802 is used to search for the user's historical sessions and sort the multiple response messages according to the searched historical sessions.

[0267] The sixth processing module 803 is used to output the target response information based on the sorted multiple response information.

[0268] The information processing apparatus provided in this application embodiment can be used to perform the above-described... Figures 1 to 6 The technical solutions of the embodiments shown are similar in principle and in effect, and will not be described again here.

[0269] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device in this embodiment may include:

[0270] At least one processor 901; and

[0271] Memory 902 communicatively connected to the at least one processor;

[0272] The memory 902 stores instructions that can be executed by the at least one processor 901 to cause the electronic device to perform the method as described in any of the above embodiments.

[0273] Optionally, the memory 902 can be either standalone or integrated with the processor 901. Optionally, the memory 902 can be connected to the processor 901 via bus 903.

[0274] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0275] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method described in any of the foregoing embodiments.

[0276] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the methods described in any of the foregoing embodiments.

[0277] The technical solution of this application involves the collection, storage, use, processing, transmission, provision, and disclosure of various information of users and merchants, which all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0278] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0279] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application.

[0280] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.

[0281] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0282] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0283] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0284] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0285] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0286] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An information processing method, characterized in that, The method includes: The system obtains the user's input information in the current conversation with the merchant, processes the input information through multiple intent models, and obtains the response information corresponding to each intent model. The user's historical conversations with the merchant are input into a preset intent ranking model to determine the priority of each intent model; Based on the priority of each intent model, multiple response messages are sorted. Based on the sorted responses, output the target response. The method further includes: Obtain the historical sessions of at least one user, and extract multi-turn dialogues from the historical sessions of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models; For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue. Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical conversations. The network model is then trained to obtain the intent ranking model.

2. The method according to claim 1, characterized in that, Based on the user's historical conversations with the merchant, the response information corresponding to multiple intent models is sorted, including: Input the user's historical conversations with the merchant and multiple response messages into a predefined response sorting model to determine the priority of the multiple response messages; The multiple response messages are sorted according to their priority. If at least two response messages have the same priority, the at least two response messages are sorted according to the merchant's intent model strategy. The intent model strategy is used to represent the priority of each pre-set intent model.

3. The method according to any one of claims 1-2, characterized in that, The method further includes: If no historical conversation between the user and the merchant is found, then the merchant's intent model strategy is obtained; According to the intent model strategy, the multiple response messages are sorted to obtain multiple sorted response messages.

4. The method according to any one of claims 1-2, characterized in that, Based on the sorted responses, output the target response, including: For any response information, determine the response type corresponding to the response information; the response type is used to indicate at least one of the following: whether to modify the response information before output, whether to add recommendation information to the user after the response information, and whether to output the added recommendation information after concatenation; Based on the sorting results and the determined response type, output the target response information.

5. The method according to claim 4, characterized in that, Determining the response type corresponding to the response information includes at least one of the following: Based on whether the response information contains preset words, determine whether to modify the response information before outputting it; Based on the context information corresponding to the response information, determine whether to add recommendation information after the response information is output; If multiple recommended answers are added, the added recommended information will be concatenated and output.

6. The method according to any one of claims 1-2, characterized in that, Based on the sorted responses, output the target response, including: Based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the time interval in the time limit condition; and / or, based on the sorted multiple response information, at least a portion of the response information is output as the target response information according to the number of recommendations in the recommendation flow limit condition; The method further includes: if the user's next input information is received, stopping the output of the target response information corresponding to the current input information and starting the search for the target response information corresponding to the next input information.

7. The method according to claim 6, characterized in that, The method further includes: Obtain the configuration information input by the merchant, which includes at least one of the following: time limit conditions, recommendation traffic limiting conditions, and intent model strategy.

8. An information processing method, characterized in that, The method includes: The user's input information is obtained, and the input information is processed through multiple intent models to obtain multiple response information; Find the user's historical sessions, input the historical sessions into a preset intent ranking model, and determine the priority of each intent model; The multiple response messages are sorted according to the priority of each intent model; Based on the sorted responses, output the target response. The method further includes: Obtain the historical sessions of at least one user, and extract multi-turn dialogues from the historical sessions of at least one user, wherein each round of dialogue includes user input information and multiple response information corresponding to the user input information determined by multiple intent models; For any given round of dialogue, the priority of multiple intent models in that round of dialogue is determined based on the user's actions following multiple responses in that round of dialogue. Training samples are constructed based on the priorities of multiple intent models corresponding to each round of dialogue and historical conversations. The network model is then trained to obtain the intent ranking model.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, cause the electronic device to perform the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the method as described in any one of claims 1-8.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images