Session processing method and related equipment

Abstract

The invention discloses a session processing method and related equipment. The method comprises the following steps: selecting a plurality of initial to-be-replied sessions similar to a target session content from a plurality of candidate to-be-replied sessions based on a similar problem clustering operation for the target session content; constructing semantic analysis prompt information according to the target session content and the initial to-be-replied session; based on semantic analysis prompt information, calculating semantic relevancy between the target session content and the initial to-be-replied session; determining a plurality of target similar to-be-replied sessions of target session contents from the initial to-be-replied sessions according to the semantic relevancy; and carrying out aggregation processing on the plurality of target similar to-be-replied sessions to obtain a similar session cluster so as to carry out batch reply processing on the plurality of target similar to-be-replied sessions in the similar session cluster. According to the method and the device, sessions with similar intentions can be identified more accurately, the clustering effect is optimized, the similar sessions obtained by clustering are replied in batches, and the reply efficiency can be greatly improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and particularly relates to a session processing method and related devices. Background Art

[0002] With the rapid development of the Internet, the business scales of industries such as e-commerce and gaming have been continuously expanding, and the volume of customer inquiries has also increased explosively. Especially during peak periods such as e-commerce promotions, new product launches, or emergencies, the problem of facing a large number of customer inquiries with high traffic needs to be addressed.

[0003] In the current related technologies, the customer service mode mainly relies on human agents to answer customer questions one by one through online chat and other means. However, in the face of a huge volume of consultation needs, there are many limitations in human customer service. On the one hand, the number of human customer service agents is limited, making it difficult to meet the consultation needs during peak periods, resulting in long waiting times for customers and poor user experience. On the other hand, the work efficiency of human customer service agents is limited. Each customer service agent can only handle one customer's problem at the same time and it is difficult to handle multiple consultations simultaneously. This causes a large number of problems to queue up, further reducing customer satisfaction. Summary of the Invention

[0004] Embodiments of this application provide a session processing method and related devices; the related devices may include a session processing device, an electronic device, a computer-readable storage medium, and a computer program product. This application can more accurately identify sessions with similar intents, optimize the clustering effect, and then batch reply to the similar sessions obtained by clustering, which can greatly improve the reply efficiency.

[0005] Embodiments of this application provide a session processing method, including:

[0006] Based on a similar question clustering operation for the target session content, select multiple initial pending reply sessions similar to the target session content from multiple candidate pending reply sessions;

[0007] According to the target session content and the initial pending reply sessions, construct semantic analysis prompt information for the initial pending reply sessions;

[0008] Based on the semantic analysis prompt information, calculate the semantic correlation degree between the target session content and the initial pending reply sessions;

[0009] According to the semantic correlation degree, determine multiple target similar pending reply sessions of the target session content from the respective initial pending reply sessions;

[0010] Perform an aggregation process on the multiple target similar pending reply sessions to obtain a similar session cluster, so as to perform a batch reply process on the multiple target similar pending reply sessions in the similar session cluster.

[0011] Correspondingly, an embodiment of the present application provides a session processing device, including:

[0012] A selection unit, configured to select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions based on a similar question clustering operation for the target session content;

[0013] A prompt construction unit, configured to construct semantic analysis prompt information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions;

[0014] A calculation unit, configured to calculate the semantic relevance between the target session content and the initial to-be-replied sessions based on the semantic analysis prompt information;

[0015] A determination unit, configured to determine multiple target similar to-be-replied sessions of the target session content from the respective initial to-be-replied sessions according to the semantic relevance;

[0016] An aggregation unit, configured to perform an aggregation process on the multiple target similar to-be-replied sessions to obtain a similar session cluster, so as to perform a batch reply process on the multiple target similar to-be-replied sessions in the similar session cluster.

[0017] Optionally, in some embodiments of the present application, the target session content is the reply content in the processed session;

[0018] The prompt construction unit may include a retrieval subunit, a prompt construction subunit, and a combination subunit, as follows:

[0019] The retrieval subunit is configured to retrieve historical similar session content corresponding to the target session content in a historical session database;

[0020] The prompt construction subunit is configured to construct initial prompt information according to the historical similar session content;

[0021] The combination subunit is configured to perform a combination process on the initial prompt information and the session content of each initial to-be-replied session respectively to obtain semantic analysis prompt information for each initial to-be-replied session.

[0022] Optionally, in some embodiments of the present application, the session processing device further includes a session content separation unit, as follows:

[0023] The session content separation unit is configured to perform a content separation process on the target session corresponding to the target session content to obtain the session historical background message, the current question message, and the reply content of the current question message of the target session;

[0024] The prompt construction subunit can specifically be used to fuse and process the session history background message, the current question message of the target session, and the historical similar session content to obtain an initial prompt message.

[0025] Optionally, in some embodiments of the present application, the target session content is an unanswered question message;

[0026] The prompt construction unit may include a fusion processing subunit and a combination processing subunit, as follows:

[0027] The fusion processing subunit is used to fuse and process the target session content and its session history background message in the corresponding session to obtain an initial prompt message;

[0028] The combination processing subunit is used to respectively combine the initial prompt message with the session content of each initial session to be replied to obtain a semantic analysis prompt message for each initial session to be replied to.

[0029] Optionally, in some embodiments of the present application, the selection unit may include a first display subunit, a second display subunit, and a selection subunit, as follows:

[0030] The first display subunit is used to display a session list page, the session list page includes a first session list area and a second session list area, the first session list area includes candidate sessions with an interaction degree higher than a preset interaction degree, and the second session list area includes candidate sessions with an interaction degree not higher than the preset interaction degree;

[0031] The second display subunit is used to, in response to a selection operation on a target session in the first session list area, display a session page corresponding to the target session, and the session page includes at least one session content of the target session;

[0032] The selection subunit is used to, in response to a similar question clustering operation on the target session content in the session page, select multiple initial sessions to be replied to that are similar to the target session content from multiple candidate sessions to be replied to, and the target session content is any session content in the session page.

[0033] Optionally, in some embodiments of the present application, the session processing device may further include a display unit and a sending unit, as follows:

[0034] The display unit is used to display a session cluster page, and the session cluster page includes a session content display area of multiple target similar sessions to be replied to in the similar session cluster and a reply area for the similar session cluster;

[0035] A sending unit, configured to, in response to a content reply operation on the content in the reply area, send the reply content corresponding to the content reply operation to each target similar pending reply session in the session cluster page.

[0036] Optionally, in some embodiments of the present application, the session cluster page includes deletion controls corresponding to each target similar pending reply session in the similar session cluster;

[0037] The session processing device may further include a deletion unit, as follows:

[0038] The deletion unit is configured to, in response to a trigger operation on a target deletion control, delete the target similar pending reply session corresponding to the target deletion control from the similar session cluster, and remove the session content display area corresponding to the target deletion control from the session cluster page, so as to update the session cluster page.

[0039] Optionally, in some embodiments of the present application, the selection unit may include a session determination subunit and a session selection subunit, as follows:

[0040] The session determination subunit is configured to, based on a similar question clustering operation for target session content, determine multiple candidate pending reply sessions according to an unprocessed session queue and currently generated real-time sessions, where the unprocessed session queue includes session identification information of unprocessed sessions within a historical time period;

[0041] The session selection subunit is configured to select multiple initial pending reply sessions similar to the target session from the respective candidate pending reply sessions according to the content matching degree between each candidate pending reply session and the target session content.

[0042] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be fully or partially implemented by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit including the function of the module or unit.

[0043] An electronic device provided in an embodiment of the present application includes a processor and a memory, the memory stores multiple instructions, and the processor loads the instructions to execute the steps in the session processing method provided in the embodiment of the present application.

[0044] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps in the session processing method provided by the embodiment of the present application are implemented.

[0045] In addition, an embodiment of the present application further provides a computer program product, including a computer program or instruction, and when the computer program or instruction is executed by a processor, the steps in the session processing method provided by the embodiment of the present application are implemented.

[0046] An embodiment of the present application provides a session processing method and related devices, which can select a plurality of initial to-be-replied sessions similar to the target session content from a plurality of candidate to-be-replied sessions based on a clustering operation of similar questions for the target session content; construct semantic analysis prompt information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions; calculate the semantic relevance between the target session content and the initial to-be-replied sessions based on the semantic analysis prompt information; determine a plurality of target similar to-be-replied sessions of the target session content from the respective initial to-be-replied sessions according to the semantic relevance; perform an aggregation process on the plurality of target similar to-be-replied sessions to obtain a similar session cluster, so as to perform a batch reply process on the plurality of target similar to-be-replied sessions in the similar session cluster.

[0047] The present application can actively cluster similar to-be-replied sessions based on a similar question clustering operation. Through active clustering, it can respond to user questions more flexibly and improve the real-time performance of session processing. Specifically, during the clustering process, the present application first performs a rough screening on the candidate to-be-replied sessions to filter out irrelevant sessions, obtaining the initial to-be-replied sessions after rough screening. Subsequently, through the construction of semantic analysis prompt information, deep semantic matching is achieved, more accurately identifying sessions with similar intentions, optimizing the clustering effect, ensuring that similar questions are correctly clustered, and then batch replying to the clustered similar sessions, which can greatly improve the reply efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0049] Figure 1a is a schematic diagram of the scenario of the session processing method provided by the embodiment of the present application;

[0050] Figure 1b is a flowchart of the session processing method provided by the embodiment of the present application;

[0051] Figure 1c It is a schematic diagram of a page of the session processing method provided by an embodiment of the present application;

[0052] Figure 1d It is another schematic diagram of a page of the session processing method provided by an embodiment of the present application;

[0053] Figure 1e It is another schematic diagram of a page of the session processing method provided by an embodiment of the present application;

[0054] Figure 1f It is another schematic diagram of a page of the session processing method provided by an embodiment of the present application;

[0055] Figure 1g It is another flowchart of the session processing method provided by an embodiment of the present application;

[0056] Figure 1h It is an explanatory diagram of the session processing method provided by an embodiment of the present application;

[0057] Figure 1i It is another explanatory diagram of the session processing method provided by an embodiment of the present application;

[0058] Figure 1j It is another flowchart of the session processing method provided by an embodiment of the present application;

[0059] Figure 2 It is another flowchart of the session processing method provided by an embodiment of the present application;

[0060] Figure 3 It is a schematic structural diagram of the session processing device provided by an embodiment of the present application;

[0061] Figure 4 It is a schematic structural diagram of the electronic device provided by an embodiment of the present application. Detailed implementation manners

[0062] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0063] An embodiment of the present application provides a session processing method and related devices. The related devices may include a session processing device, an electronic device, a computer-readable storage medium, and a computer program product. The session processing device may be specifically integrated in the electronic device, and the electronic device may be a device such as a terminal or a server.

[0064] It can be understood that the session processing method in this embodiment can be executed on a terminal, on a server, or jointly by a terminal and a server. The above examples should not be construed as limitations on this application.

[0065] As Figure 1a shown, taking the joint execution of the session processing method by a terminal and a server as an example. The session processing system provided by the embodiments of this application includes a terminal 10, a server 11, etc.; the terminal 10 and the server 11 are connected through a network, for example, through a wired or wireless network connection, etc., where the session processing device can be integrated in the server.

[0066] Among them, the server 11 can be used to: based on the clustering operation of similar questions for the target session content in the terminal 10, select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions; according to the target session content and the initial to-be-replied sessions, construct semantic analysis prompt information for the initial to-be-replied sessions; based on the semantic analysis prompt information, calculate the semantic correlation degree between the target session content and the initial to-be-replied sessions; according to the semantic correlation degree, determine multiple target similar to-be-replied sessions of the target session content from each of the initial to-be-replied sessions; perform aggregation processing on the multiple target similar to-be-replied sessions to obtain a similar session cluster, so as to perform batch reply processing on the multiple target similar to-be-replied sessions in the similar session cluster. Among them, the server 11 can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0067] Among them, the terminal 10 can be used to: display a session list page, and in response to a selection operation on a target session in the session list page, display a session page corresponding to the target session, where the session page includes at least one piece of session content of the target session; in response to a similar question clustering operation on the target session content in the session page, trigger the server 11 to select multiple target similar to-be-replied sessions of the target session content from multiple candidate to-be-replied sessions, so as to obtain a similar session cluster. The terminal 10 can also receive the similar session cluster sent by the server 11, and display a session cluster page, where the session cluster page includes a session content display area of multiple target similar to-be-replied sessions in the similar session cluster, and a reply area for the similar session cluster; in response to a content reply operation on the reply area, send the reply content corresponding to the content reply operation to each target similar to-be-replied session in the session cluster page. Among them, the terminal 10 can include a mobile phone, a vehicle-mounted terminal, an aircraft, a tablet computer, a laptop computer, or a personal computer (PC, Personal Computer), etc. A client can also be set on the terminal 10, and the client can be an application client or a browser client, etc.

[0068] The steps such as session processing in the above-mentioned server 11 can also be executed by the terminal 10.

[0069] The following will be described in detail respectively. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments.

[0070] This embodiment will be described from the perspective of a session processing device, which can be specifically integrated in an electronic device, and the electronic device can be a device such as a server or a terminal.

[0071] It can be understood that in the specific implementation manner of this application, when it comes to data such as user information, when the above embodiments of this application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0072] As Figure 1b shown, the specific process of this session processing method can be as follows:

[0073] 101. Based on a similar question clustering operation on target session content, select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions.

[0074] Among them, in this embodiment, the target conversation content serves as a trigger source for actively clustering similar questions, and it can be any selected conversation message. For example, the target conversation content can specifically be the reply content in the processed conversation or the unanswered question message. This embodiment does not limit this. Among them, the processed conversation is specifically the replied conversation. This embodiment can be applied to the customer service scenario, and the active clustering is specifically a clustering method actively triggered by the customer service staff, using the selected question or answer as the trigger source to retrieve and cluster similar questions in real time.

[0075] Specifically, in the conversation page, the target conversation content is displayed. A trigger operation is performed on the target conversation content. This trigger operation can be a right-click operation. Based on this trigger operation, a similar question clustering trigger control for the target conversation content can be displayed; the similar question clustering operation can specifically be a trigger operation on this similar question clustering trigger control, such as a click operation.

[0076] Optionally, in this embodiment, the step of "selecting multiple initial to-be-replied conversations similar to the target conversation content from multiple candidate to-be-replied conversations" may include:

[0077] Extract the content semantic feature information of the target conversation content;

[0078] Extract the question semantic feature information of the candidate to-be-replied conversations;

[0079] According to the semantic matching degree between the content semantic feature information and the question semantic feature information of each candidate to-be-replied conversation, select multiple initial to-be-replied conversations of the target conversation content from each candidate to-be-replied conversation.

[0080] Among them, the candidate to-be-replied conversation can specifically be a customer consultation conversation that needs to be replied but has not yet received an actual reply.

[0081] Among them, the content semantic feature information of the target conversation content and the question semantic feature information of the candidate reply conversation to be selected can be extracted through a content matching model. The content matching model can be a neural network model, and the neural network model can be LSTM (Long Short-Term Memory), RNN (Recurrent Neural Network), GRU (Gate Recurrent Unit), transformer, BERT (Bidirectional Encoder Representations from Transformers), etc. However, it should be understood that the neural network model in this embodiment is not limited to the several types listed above.

[0082] Among them, the semantic matching degree can be determined by the feature distance between the content semantic feature information of the target conversation content and the question semantic feature information of the candidate reply conversation to be selected. The greater the feature distance, the smaller the semantic matching degree; conversely, the smaller the feature distance, the greater the semantic matching degree.

[0083] In some embodiments, the candidate reply conversations with a semantic matching degree higher than a preset value can be selected as the initial reply conversations similar to the target conversation content. In other embodiments, the candidate reply conversations can be sorted according to the semantic matching degree, for example, sorted from largest to smallest according to the size of the semantic matching degree to obtain the sorted candidate reply conversations, and the first n candidate reply conversations in the sorted candidate reply conversations are used as the initial reply conversations similar to the target conversation content.

[0084] Optionally, in this embodiment, the step of "selecting multiple initial reply conversations similar to the target conversation content from multiple candidate reply conversations based on the similar question clustering operation for the target conversation content" may include:

[0085] Based on the similar question clustering operation for the target conversation content, multiple candidate reply conversations are determined according to the unprocessed conversation queue and the currently generated real-time conversation. The unprocessed conversation queue includes the conversation identification information of the unprocessed conversations within the historical time period;

[0086] According to the content matching degree between each candidate reply conversation and the target conversation content, multiple initial reply conversations similar to the target conversation are selected from each candidate reply conversation.

[0087] Among them, specifically, "current" in the currently generated real-time conversation refers to the time period from after the historical time period to the current moment.

[0088] Among them, the unprocessed session queue is used to store the session identification information of unprocessed sessions within a historical time period. The unprocessed sessions can specifically be unresponded sessions, and the session identification information is a unique identifier used to distinguish sessions.

[0089] Among them, specifically, the content matching degree between the candidate sessions to be replied and the target session content can be calculated through cosine similarity, etc.

[0090] Specifically, the step of "determining multiple candidate sessions to be replied according to the unprocessed session queue and the currently generated real-time session" may include:

[0091] Pull the session identification information of unprocessed sessions within the historical time period from the unprocessed session queue;

[0092] Obtain the real-time session identification information of the currently generated real-time session;

[0093] Determine candidate sessions to be replied according to the session identification information of unprocessed sessions within the historical time period and the real-time session identification information.

[0094] Among them, in this embodiment, all unresponded sessions at the current moment and newly generated user questions can be used as candidate sessions to be replied.

[0095] Optionally, in this embodiment, the step of "selecting multiple initial sessions to be replied similar to the target session content from multiple candidate sessions to be replied based on the similar question clustering operation for the target session content" may include:

[0096] Display a session list page, where the session list page includes a first session list area and a second session list area. The first session list area includes candidate sessions with an interaction degree higher than a preset interaction degree, and the second session list area includes candidate sessions with an interaction degree not higher than the preset interaction degree;

[0097] In response to the selection operation on the target session in the first session list area, display the session page corresponding to the target session, where the session page includes at least one piece of session content of the target session;

[0098] In response to the similar question clustering operation on the target session content in the session page, select multiple initial sessions to be replied similar to the target session content from multiple candidate sessions to be replied, and the target session content is any session content in the session page.

[0099] Among them, the preset interaction degree can be set according to the actual situation. The interaction degree of a conversation can be determined based on the frequency of the conversation and the duration of the conversation. The frequency of the conversation represents the number of times the conversation occurs within a preset time, and the duration of the conversation reflects the depth and degree of engagement of each conversation. Specifically, the conversations in the first conversation list area are active conversations, and the conversations in the second conversation list area can be regarded as silent conversations.

[0100] Among them, the target conversation is any candidate conversation in the first conversation list area. The selection operation for the target conversation in the first conversation list area can specifically be a click operation, etc.

[0101] In a specific scenario, as Figure 1c shown, it is a schematic diagram for selecting an active clustering trigger conversation. The customer service steward can, in the conversation chat interface of the customer service platform, select a well - handled excellent reply or a player message in a newly - emerging conversation with a high probability in the "active conversation" list as the trigger source for active clustering. In some embodiments, it is also possible to enter the openid (open user identifier), contact person remarks, or nickname, etc. in the search bar in the conversation chat interface to search for the corresponding conversation, and then select the conversation message in the conversation as the trigger source for active clustering.

[0102] After selecting the trigger conversation, the steward can right - click on the conversation message (i.e., the above - mentioned target conversation content) to display relevant controls for the conversation message on the conversation page, such as buttons like "multiple selection", "copy", "reply", "withdraw", "batch process similar problems", etc. Then click the "batch process similar problems" button to initiate the retrieval and clustering for similar problems. Among them, the "batch process similar problems" button is specifically the similar - problem clustering trigger control in the above - mentioned embodiment. Here, there are two cases: one is to select the message replied by the steward as the trigger source, and the other is to directly select the message of the player user as the trigger source, as shown in Figure 1d and Figure 1e shown respectively. Figure 1d For selecting the steward's reply for active clustering, Figure 1e For selecting the player's message for active clustering. After clicking the "batch process similar problems" button to trigger, the system background can perform in - depth semantic matching on the player messages to be replied in all active conversations based on the trigger source, and accurately identify conversations with similar intentions.

[0103] 102. Construct semantic analysis prompt information for the initial conversation to be replied based on the target conversation content and the initial conversation to be replied.

[0104] Optionally, in this embodiment, the target conversation content is the reply content in the processed conversation;

[0105] The step of "constructing semantic analysis prompt information for the initial session to be replied according to the target session content and the initial session to be replied" may include:

[0106] Retrieving historical similar session content corresponding to the target session content in the historical session database;

[0107] Constructing initial prompt information according to the historical similar session content;

[0108] Combining the initial prompt information with the session content of each initial session to be replied respectively to obtain semantic analysis prompt information for each initial session to be replied.

[0109] Among them, the historical session database may include historical processed session content, and the session content here may be a question or a reply. The historical session database is equivalent to an external knowledge base. In this embodiment, through the Retrieval-Augmented Generation (RAG) technology, the target session content can be converted into a retrieval query (Query) to find similar session content to the target session content in the historical session database.

[0110] Among them, retrieval-augmented generation is a method that combines generative models and retrieval techniques. By retrieving relevant information from an external knowledge base, it enhances the quality and accuracy of the answers of the generative model and is applicable to tasks that require real-time or knowledge-rich answers.

[0111] Among them, the process of retrieving historical similar session content may specifically include: extracting the content feature information of the target session content, respectively extracting the content feature information of each historical session content in the historical session database, and then calculating the similarity between the content feature information of the target session content and the content feature information of the historical session content. Specifically, cosine similarity, etc. can be used for similarity calculation, so as to retrieve historical similar session content from the historical session database according to the calculated similarity. Specifically, in some embodiments, the historical session content with a similarity greater than the preset similarity may be used as the historical similar session content of the target session content; in other embodiments, sorting may be performed according to the similarity size, such as sorting from large to small, to obtain a historical session content sequence, and the first n historical session contents in the historical session content sequence may be used as the historical similar session content of the target session content.

[0112] Among them, the combination process of the initial prompt information and the session content of the initial session to be replied may specifically be splicing the initial prompt information and the session content of the initial session to be replied, etc.

[0113] Optionally, in this embodiment, before the step of "constructing an initial prompt message according to the historical similar conversation content", the following may also be included:

[0114] Perform content separation processing on the target conversation corresponding to the target conversation content to obtain the conversation historical background message, the current problem message, and the reply content of the current problem message of the target conversation;

[0115] The step of "constructing an initial prompt message according to the historical similar conversation content" may include:

[0116] Fuse the conversation historical background message, the current problem message of the target conversation, and the historical similar conversation content to obtain an initial prompt message.

[0117] Among them, the target conversation is the conversation where the target conversation content is located. The content separation processing of the target conversation specifically means dividing the conversation content in the target conversation.

[0118] Among them, the conversation historical background message is specifically the conversation above the target conversation content; the conversation historical background message can be the conversation content in the target conversation whose message release time is before the message release time of the current problem message. Specifically, the conversation historical background message is the other conversation content in the target conversation except the target conversation content and the current problem message.

[0119] Among them, the current problem message can specifically be the latest question in the target conversation, that is, the question with the latest message release time in the target conversation. Here, the target conversation content can specifically be the latest reply in the target conversation. That is to say, the target conversation content is the reply content of the current problem message.

[0120] Among them, the fusion processing of the conversation historical background message, the current problem message of the target conversation, and the historical similar conversation content can specifically be splicing processing, etc. It can also be to obtain a preset prompt template and fill the conversation historical background message, the current problem message of the target conversation, and the historical similar conversation content into the corresponding positions in the preset prompt template to obtain an initial prompt message.

[0121] Optionally, in this embodiment, the target conversation content is an unanswered question message;

[0122] The step of "constructing a semantic analysis prompt message for the initial to-be-replied conversation according to the target conversation content and the initial to-be-replied conversation" may include:

[0123] Fuse the target conversation content and its conversation historical background message in the corresponding conversation to obtain an initial prompt message;

[0124] Combined process the initial prompt information with the conversation content of each initial reply-to-be session respectively to obtain semantic analysis prompt information for each initial reply-to-be session.

[0125] Among them, the conversation historical background message is specifically the conversation context above the target conversation content. The conversation historical background message can be the conversation content in the target conversation whose message publishing time is before the message publishing time of the target conversation content. Specifically, the conversation historical background message is the other conversation content in the target conversation except the target conversation content.

[0126] Among them, if the trigger source is a user question rather than the reply content of the customer service, the target conversation content can be fused with its conversation context above. This fusion process can be a splicing process, or the target conversation content and its conversation context above can be filled into a preset prompt template to obtain the initial prompt information.

[0127] Among them, the combined process of the initial prompt information and the conversation content of the initial reply-to-be session can specifically be a splicing process of the initial prompt information and the conversation content of the initial reply-to-be session, etc.

[0128] 103. Calculate the semantic relevance between the target conversation content and the initial reply-to-be session based on the semantic analysis prompt information.

[0129] Among them, for the semantic relevance between the target conversation content and the initial reply-to-be session, it can be calculated through cosine similarity, etc. This embodiment does not limit this.

[0130] Among them, the semantic relevance between the target conversation content and the initial reply-to-be session can be calculated through a large language model. The large language model is specifically a huge neural network model based on deep learning, trained on a large amount of text data, and has powerful language understanding, generation, and discrimination capabilities.

[0131] Optionally, in this embodiment, the step of "calculate the semantic relevance between the target conversation content and the initial reply-to-be session based on the semantic analysis prompt information" may include:

[0132] Extract the content feature information of the target conversation content;

[0133] Extract the content feature information of the conversation content of the initial reply-to-be session;

[0134] Based on the semantic analysis prompt information, adjust the content feature information of the target conversation content to obtain the adjusted content feature information of the target conversation content;

[0135] Adjust the content feature information of the initial to-be-replied conversation based on the semantic analysis prompt information to obtain the adjusted content feature information of the initial to-be-replied conversation;

[0136] Calculate the semantic relevance between the target conversation content and the initial to-be-replied conversation according to the adjusted content feature information of the target conversation content and the adjusted content feature information of the initial to-be-replied conversation.

[0137] Among them, the adjustment of the content feature information can specifically be to perform weighted processing on the vectors of certain words or sentences according to the semantic focus of the semantic analysis prompt information, so as to calculate the semantic relevance more accurately.

[0138] Among them, the feature distance between the adjusted content feature information of the target conversation content and the adjusted content feature information of the initial to-be-replied conversation can be calculated, and according to this feature distance, the semantic relevance between the target conversation content and the initial to-be-replied conversation can be determined.

[0139] 104. Determine multiple target similar to-be-replied conversations of the target conversation content from the respective initial to-be-replied conversations according to the semantic relevance.

[0140] In some embodiments, the initial to-be-replied conversations with a semantic relevance greater than a preset value can be determined as the target similar to-be-replied conversations of the target conversation content, and this preset value can be set according to the actual situation.

[0141] In other embodiments, the respective initial to-be-replied conversations can be arranged according to the magnitude of the semantic similarity, specifically arranged from large to small according to the semantic similarity to obtain a list of initial to-be-replied conversations, and the first n initial to-be-replied conversations in the list of initial to-be-replied conversations are determined as the target similar to-be-replied conversations of the target conversation content.

[0142] 105. Aggregate the multiple target similar to-be-replied conversations to obtain a similar conversation cluster for batch reply processing of the multiple target similar to-be-replied conversations in the similar conversation cluster.

[0143] Among them, the aggregation processing of the multiple target similar to-be-replied conversations can specifically be to combine the multiple target similar to-be-replied conversations and summarize them together to obtain a similar conversation cluster.

[0144] Optionally, in this embodiment, after the step of "aggregate the multiple target similar to-be-replied conversations to obtain a similar conversation cluster", it may further include:

[0145] Display a conversation cluster page, where the conversation cluster page includes a conversation content display area of multiple target similar to-be-replied conversations in the similar conversation cluster and a reply area for the similar conversation cluster;

[0146] In response to a content reply operation on the content in the reply area, send the reply content corresponding to the content reply operation to each target similar pending reply session in the session cluster page.

[0147] Among them, the session content display area is used to display the session content of the corresponding session, and it can support the up and down scrolling view processing of the session content in the corresponding session.

[0148] Among them, the reply area is used to uniformly reply to the sessions corresponding to the session content display areas of each session in the session cluster page. The content reply operation on the reply area can specifically include editing the reply content in the reply area and sending the edited reply content.

[0149] Among them, in this embodiment, multiple aggregated target similar pending reply sessions can be provided to the front-end for display for batch reply. In some specific scenarios, if the number of target similar pending reply sessions is large, a part of the target similar pending reply sessions can be displayed on the session cluster page, and the other target similar pending reply sessions can be viewed in a paging manner; the content reply operation in the reply area can be only for the target similar pending reply sessions displayed on the current session cluster page, or for all target similar pending reply sessions in the similar session cluster. This embodiment does not limit this.

[0150] Among them, through the content reply operation on the reply area, the corresponding reply content can be uniformly sent to each target similar pending reply session in the session cluster page, realizing batch reply processing.

[0151] In a specific embodiment, when the customer service steward follows Figure 1d and Figure 1e selects the trigger source, the system will automatically complete the matching of similar sessions and display the clustering result in the front-end interface, specifically as Figure 1f shown. Figure 1f The page shown is the session cluster page, which includes a reply area and the session content display areas of four target similar pending reply sessions. The customer service steward can enter the reply content in the reply area and then click the "Send" button to send the reply content to these four target similar pending reply sessions.

[0152] By introducing active clustering technology, this application can actively cluster unanswered user questions, group similar questions together for the butler to process in batches. The system allows the butler to trigger clustering actively at any time, significantly improving real-time performance, enabling quick responses to hot issues or urgent needs, and reducing the waiting time for players. At the same time, the butler can select key questions as the trigger points for clustering according to the actual situation, flexibly prioritize and batch process high-priority questions, improving the flexibility of the system in emergencies and enhancing service efficiency and quality. In addition, by leveraging advanced natural language processing technology, the system has the ability of deep semantic understanding, can more accurately identify questions with similar intentions, optimize the clustering results, and reduce the screening pressure. For the diverse expressions of users for the same question, the semantic recognition ability of the system can ensure that similar questions are correctly clustered, guaranteeing the comprehensiveness of clustering. More importantly, the system utilizes the powerful generalization ability of the large language model, can effectively handle new question types or expression ways that have not been seen before, showing a high degree of adaptability and meeting the changing needs of users.

[0153] Specifically, after batch reply, the conversation content of the processed target similar to-be-replied conversations can also be archived in the historical conversation database.

[0154] Optionally, in this embodiment, the conversation processing method may further include:

[0155] Return the conversation identification information of other initial to-be-replied conversations except the target similar to-be-replied conversations to the unprocessed conversation queue;

[0156] Return the conversation identification information of other candidate to-be-replied conversations except the initial to-be-replied conversations to the unprocessed conversation queue.

[0157] Optionally, in this embodiment, the conversation cluster page includes deletion controls corresponding to each target similar to-be-replied conversation in the similar conversation cluster; the conversation processing method further includes:

[0158] In response to the triggering operation on the target deletion control, delete the target similar to-be-replied conversation corresponding to the target deletion control from the similar conversation cluster, and remove the conversation content display area corresponding to the target deletion control from the conversation cluster page to update the conversation cluster page.

[0159] Among them, the triggering operation on the target deletion control may specifically be a click operation on the target deletion control, and the target deletion control may be any deletion control on the conversation cluster page. As Figure 1f shown, the circled icon with a cross in the upper right corner of each conversation content display area is the deletion control.

[0160] Among them, in this embodiment, irrelevant conversations can be actively eliminated by triggering the target deletion control, and batch responses can be made for the remaining similar questions, thereby improving work efficiency.

[0161] In a specific scenario, for the customer service system of large online games, due to the large number of players, and there may be a large number of players asking similar or identical questions during specific time periods such as game version updates or the launch of popular activities. From the perspective of customer service, this application proposes an efficient method for processing similar questions based on active clustering, which can actively cluster all unprocessed player conversations for batch response through clustering, and can improve the efficiency of customer service staff in processing similar questions and improve the service experience of users.

[0162] Under the function of active clustering provided by this application, the customer service steward can, according to the actual situation, actively select excellent replies that have been processed or player questions in new conversations as the trigger source for clustering, and initiate real-time retrieval and clustering of similar questions. This application integrates advanced large language models and retrieval-enhanced generation technologies to perform deep semantic matching between the trigger source and all un-answered player questions, and accurately identify questions with similar intentions. Therefore, it can handle various ways of asking questions by players, including different wordings, word orders, and expression styles. Due to its strong generalization ability, this method is not only applicable to the current customer service scenario, but can also be extended to other fields that need to process similar text information and has the ability to handle new types of questions.

[0163] Specifically, this application focuses on realizing the efficient processing of similar questions based on active clustering. By integrating large language models and retrieval-enhanced generation technologies, intelligent customer service question clustering and batch processing are realized. The specific technical implementation logic is as Figure 1g shown and is specifically described as follows:

[0164] First, it is the trigger of active clustering and the construction of candidate conversations to be replied: The steward needs to determine the conversation for which active clustering is to be performed, and use the conversation message in this conversation as the trigger source to trigger active clustering. When the steward triggers active clustering, the system will automatically use all un-answered conversations at the current moment and newly generated player questions as the candidate conversation set. For the triggered conversation (that is, the conversation corresponding to the target conversation content), the conversation content of this conversation can be separated into three parts, such as Figure 1hThe schematic diagram of trigger source session separation shown. These three parts are session context (specifically the session history background message in the above embodiment, which can also be simply referred to as session context), player message (specifically the current question message in the above embodiment), and butler reply. The session context can be used to provide context information to help the model understand the background of the question. The player message is the latest message content of the current player, denoted as q. The butler reply is the reply content to the latest message of the current player, denoted as A. It should be noted that if the trigger source is the player message, the trigger session only needs to be separated into two parts - session context and player message, without the butler reply part.

[0165] Then, calculate the similarity between the candidate reply session and the trigger session to efficiently filter out candidate reply sessions irrelevant to the trigger session through session rough ranking, and obtain the initial reply session after screening. Specifically, text embedding technology can be used to represent the candidate reply session. Specifically, a text embedding model (such as the Bert model, etc.) can be used to map the text into a high-dimensional vector space, making semantically similar texts closer in the vector space. First, for each candidate reply session, first encode the text content of its last player into a high-dimensional vector through the model, denoted as E c , where: E c ∈R d , representing the d-dimensional vector of the session content of the candidate reply session, and d is the dimension of the embedding space, generally 768 dimensions. Similarly, for the trigger session (i.e., the session corresponding to the target session content), its text content is also encoded into a high-dimensional vector through the same model, denoted as E t , where: E t ∈R d . In this way, the text representation vectors of the candidate reply session and the trigger session are obtained. Then, judge whether the candidate reply session is relevant to the trigger session by calculating the similarity between these two vectors. Specifically, the cosine similarity can be used to calculate the similarity S(E t ,E c ), and the calculation formula is shown in formula (1):

[0166]

[0167] Among them, the value range of the cosine similarity is [-1, 1], and the closer the value is to 1, the more similar the semantics of the two sessions are. In the filtering stage, the system calculates the cosine similarity between each candidate reply session and the trigger session, and sets the screening threshold of the similarity to 0.6. Only when the similarity S(E t ,E c) When it is >0.6, the candidate sessions to be replied will be retained and enter the next round of fine ranking calculation by the large model. Sessions that do not reach the threshold will be temporarily rolled back to the unprocessed queue and wait for subsequent processing. Through the above filtering mechanism, the system can significantly reduce the number of candidate sessions to be replied entering the fine ranking stage in the first stage, thereby reducing the computational complexity in the subsequent stage. Let N represent the number of original candidate sessions to be replied, and M represent the number of sessions retained after threshold filtering (obviously M≤N). Since the computational complexity of cosine similarity calculation is O(d), the overall computational complexity is reduced from the original O(N×d) to O(M×d), where M is usually much smaller than N.

[0168] Reference Figure 1i , the sessions can be roughly ranked according to the calculated cosine similarity. The candidate sessions to be replied are arranged in descending order of similarity to obtain a list of sessions to be replied. Then, the sessions with a threshold greater than 0.6 in the list of sessions to be replied are retained, and other sessions with a threshold not greater than 0.6 are deleted to obtain an updated list of sessions to be replied.

[0169] After the rough ranking of the sessions is completed, the construction of the prompt information and the generation of retrieval enhancement need to be carried out. The system constructs a Prompt prompt and combines the semantic understanding ability of the large language model (LLM) to efficiently screen and judge the candidate sessions to be replied. For different trigger sources, the present application designs two processing methods to ensure the accuracy and flexibility of the judgment.

[0170] The first case is when the customer service steward selects the already replied answer A as the trigger source. The system will, based on the retrieval enhancement generation technology, convert answer A into a retrieval query (Query) and search for sessions similar to the content of answer A from the historical session database. The calculation method used in the retrieval process is also to use cosine similarity. The threshold set in the present application can be 0.8 here. To ensure the operation efficiency, only the top-5 similar sessions can be considered. These retrieved sessions are called "historical similar session contents", which provide important semantic references for the current judgment. Subsequently, the system integrates the "session context" of the trigger session, the player's message q, and the retrieved "historical similar session contents", and combines them with each candidate session to be replied in the list of sessions to be replied one by one to generate multiple complete Prompts (i.e., the semantic analysis prompt information in the above embodiments). These Prompts are used as the input of the large language model to help the model better judge the semantic relevance between the trigger session and the candidate sessions. Specifically, the Prompt includes the background of the trigger session, the message sent by the player, the content of the historical similar session, and the content of the session to be replied. Finally, through the semantic analysis of the model, an association score is generated for each candidate session to be replied.

[0171] The second case is when the customer service steward selects the message q of the current player as the trigger source, and the processing logic of the system will be simplified. In this case, the system directly integrates the "conversation context" of the trigger conversation with the player's message q, and then combines them with each candidate reply conversation in the list of candidate reply conversations to be generated one by one to generate the corresponding Prompt. Compared with the first case, this scenario does not involve the retrieval of historical similar conversations, and the construction of the Prompt is more direct. However, it can still rely on the context information of the trigger conversation to help the large language model accurately judge the semantic relevance between the trigger conversation and the candidate conversation. In this way, the system can construct an accurate semantic context according to different trigger sources to provide support for the subsequent judgment of the large model, ensuring the effectiveness and flexibility of the judgment.

[0172] Specifically, in the fine-tuning stage, in order to reduce the online inference cost and save computing time, this application can adopt a locally trained large language model (with 7 billion parameters) as an alternative to the large GPT (Generative Pre-trained Transformer) series of models. This model can significantly reduce resource consumption and response latency while ensuring semantic understanding and reasoning capabilities. When the input Prompt is sent to this large language model, the model analyzes the semantic relationship between the candidate reply conversation and the trigger conversation through a deep semantic understanding and matching algorithm, and classifies the results into three categories, namely "similar", "unable to judge", and "dissimilar". The "similar" conversations are specifically homogeneous conversations, and the "dissimilar" conversations are specifically heterogeneous conversations. For the conversations classified as "similar", the system will push them to the front-end display page in batches for the customer service steward to view and operate. The steward can directly reply to these "similar" questions in batches, significantly improving the processing efficiency. After processing, these conversations will be archived in the historical conversation database as resources for subsequent retrieval and reference, thus continuously enriching the historical data foundation of the system. For the conversations classified as "unable to judge" or "dissimilar", the system will dynamically update the database and temporarily store them in the unprocessed conversation queue. These questions will wait to be triggered by other trigger mechanisms for semantic clustering in the future, or be processed by the customer service steward through manual intervention.

[0173] The above entire active clustering process can be referred to Figure 1j , Figure 1j which shows the complete link from the steward triggering the active clustering function to Prompt construction, model inference, result classification, and subsequent operations.

[0174] Through the active clustering technology, this application solves the problem of the traditional customer service system's inefficient handling of similar issues. The system uses large language models and retrieval-augmented generation technology. When a session is triggered, it performs a rough ranking through Embedding feature embedding technology to filter out irrelevant sessions. Subsequently, through the retrieval of historical sessions and the construction of Prompts, it achieves deep semantic matching. Through this solution, the system can quickly screen out "similar" questions and push them to the front-end display page for the butler to reply in batches. Experimental data shows that compared with the manual processing method, the active clustering technology has significantly improved the processing efficiency of similar questions, and the accuracy of similar questions has reached about 95%. In addition, the processed sessions will be archived in the historical session database to support subsequent retrieval and clustering, continuously optimizing the system performance.

[0175] This application not only improves the customer service efficiency and user experience, but also significantly reduces the workload of the butler, and has strong industry adaptability and scalability. Specifically, it can be applied to the customer service system of large online games, and can also be applied to the customer service system of e-commerce, etc.

[0176] As can be seen from the above, this embodiment can perform a clustering operation on similar questions for the target session content, select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions; construct semantic analysis prompt information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions; calculate the semantic relevance between the target session content and the initial to-be-replied sessions based on the semantic analysis prompt information; determine multiple target similar to-be-replied sessions of the target session content from the respective initial to-be-replied sessions according to the semantic relevance; perform an aggregation process on the multiple target similar to-be-replied sessions to obtain a similar session cluster, so as to perform a batch reply process on the multiple target similar to-be-replied sessions in the similar session cluster.

[0177] This application can actively cluster similar to-be-replied sessions based on a similar question clustering operation. Through active clustering, it can respond to user questions more flexibly and improve the real-time performance of session processing. Specifically, during the clustering process, this application first performs a rough screening on the candidate to-be-replied sessions to filter out irrelevant sessions, obtaining the initial to-be-replied sessions after rough screening. Subsequently, through the construction of semantic analysis prompt information, it achieves deep semantic matching, more accurately identifies sessions with similar intentions, optimizes the clustering effect, ensures that similar questions are correctly clustered, and then performs a batch reply on the clustered similar sessions, which can greatly improve the reply efficiency.

[0178] According to the method described in the previous embodiment, the following will take the specific integration of the session processing device in the terminal as an example for further detailed description.

[0179] An embodiment of this application provides a session processing method, such asFigure 2 As shown in Figure 2 , the specific process of this session processing method can be as follows:

[0180] 201. Based on the clustering operation of similar questions for the target session content, the terminal selects multiple initial candidate sessions to be replied that are similar to the target session content from multiple candidate sessions to be replied.

[0181] Among them, in this embodiment, the target session content serves as the trigger source for active clustering of similar questions, and it can be any selected session message. For example, the target session content can specifically be the reply content in the processed session or the un-answered question message, and this embodiment does not limit this.

[0182] Optionally, in this embodiment, the step of "selecting multiple initial candidate sessions to be replied that are similar to the target session content from multiple candidate sessions to be replied based on the clustering operation of similar questions for the target session content" may include:

[0183] Display a session list page, where the session list page includes a first session list area and a second session list area. The first session list area includes candidate sessions with an interaction degree higher than a preset interaction degree, and the second session list area includes candidate sessions with an interaction degree not higher than the preset interaction degree;

[0184] In response to the selection operation on the target session in the first session list area, display the session page corresponding to the target session, where the session page includes at least one piece of session content of the target session;

[0185] In response to the clustering operation of similar questions for the target session content in the session page, select multiple initial candidate sessions to be replied that are similar to the target session content, where the target session content is any session content in the session page.

[0186] Among them, the preset interaction degree can be set according to the actual situation. The interaction degree of a session can be determined based on the frequency of the session and the duration of the session. The frequency of the session represents the number of times the session occurs within a preset time, and the duration of the session reflects the depth and degree of investment of each session. Specifically, the sessions in the first session list area are active sessions, and the sessions in the second session list area can be regarded as silent sessions.

[0187] Among them, the target session is any candidate session in the first session list area. The selection operation on the target session in the first session list area can specifically be a click operation, etc.

[0188] 202. The terminal constructs semantic analysis prompt information for the initial candidate sessions to be replied according to the target session content and the initial candidate sessions to be replied.

[0189] Optionally, in this embodiment, the target session content is the reply content in the processed session;

[0190] The step of "constructing a semantic analysis prompt message for the initial session to be replied according to the target session content and the initial session to be replied" may include:

[0191] Retrieving the historical similar session content corresponding to the target session content in the historical session database;

[0192] Constructing an initial prompt message according to the historical similar session content;

[0193] Combining the initial prompt message with the session content of each initial session to be replied respectively to obtain a semantic analysis prompt message for each initial session to be replied.

[0194] Among them, the historical session database may include the session content that has been processed historically. Here, the session content may be a question or a reply. The historical session database is equivalent to an external knowledge base. In this embodiment, through the retrieval-augmented generation technology, the target session content can be converted into a retrieval query (Query) to find the session content similar to the target session content in the historical session database.

[0195] Among them, the combination processing of the initial prompt message and the session content of the initial session to be replied may specifically be splicing processing of the initial prompt message and the session content of the initial session to be replied, etc.

[0196] Optionally, in this embodiment, before the step of "constructing an initial prompt message according to the historical similar session content", it may further include:

[0197] Performing content separation processing on the target session corresponding to the target session content to obtain the session historical background message, the current problem message, and the reply content of the current problem message of the target session;

[0198] The step of "constructing an initial prompt message according to the historical similar session content" may include:

[0199] Fusing the session historical background message, the current problem message of the target session, and the historical similar session content to obtain an initial prompt message.

[0200] Among them, the session historical background message is specifically the session context above the target session content. The session historical background message may be other session content in the target session corresponding to the target session content except the target session content and the current problem message.

[0201] Among them, the target conversation content can specifically be the latest reply in the target conversation, and the current question message can specifically be the latest question in the target conversation.

[0202] Among them, the fusion processing of the conversation historical background message, the current question message of the target conversation, and the historical similar conversation content can specifically be splicing processing, etc., or it can be to obtain a preset prompt template, and fill the conversation historical background message, the current question message of the target conversation, and the historical similar conversation content into the corresponding positions in the preset prompt template to obtain the initial prompt information.

[0203] Optionally, in this embodiment, the target conversation content is an unanswered question message;

[0204] The step of "constructing semantic analysis prompt information for the initial to-be-replied conversation according to the target conversation content and the initial to-be-replied conversation" may include:

[0205] Fuse the target conversation content and its conversation historical background message in the corresponding conversation to obtain the initial prompt information;

[0206] Combine the initial prompt information with the conversation content of each initial to-be-replied conversation respectively to obtain semantic analysis prompt information for each initial to-be-replied conversation.

[0207] Among them, the conversation historical background message is specifically the conversation context above the target conversation content, and the conversation historical background message can be other conversation content in the target conversation corresponding to the target conversation content except the target conversation content.

[0208] Among them, if the trigger source is a user question rather than the reply content of the customer service, then the target conversation content can be fused with its conversation context above. This fusion processing can be splicing processing, or the target conversation content and its conversation context above can be filled into the preset prompt template to obtain the initial prompt information.

[0209] 203. The terminal calculates the semantic relevance between the target conversation content and the initial to-be-replied conversation based on the semantic analysis prompt information.

[0210] Among them, the semantic relevance between the target conversation content and the initial to-be-replied conversation can be calculated through a large language model. The large language model is specifically a huge neural network model based on deep learning, trained on a large amount of text data, and has powerful language understanding, generation, and discrimination capabilities.

[0211] 204. The terminal determines multiple target similar to-be-replied conversations of the target conversation content from the various initial to-be-replied conversations according to the semantic relevance.

[0212] 205. The terminal performs an aggregation process on the multiple target similar to-be-replied sessions to obtain a similar session cluster.

[0213] 206. The terminal displays a session cluster page, where the session cluster page includes a session content display area for multiple target similar to-be-replied sessions in the similar session cluster, and a reply area for the similar session cluster.

[0214] Among them, in this embodiment, the aggregated multiple target similar to-be-replied sessions can be provided to the front end for display for batch reply. In some specific scenarios, if the number of target similar to-be-replied sessions is large, a part of the target similar to-be-replied sessions can be displayed on the session cluster page, and the other target similar to-be-replied sessions can be viewed in a paging manner; the content reply operation in the reply area can be for only the target similar to-be-replied sessions displayed on the current session cluster page, or for all target similar to-be-replied sessions in the similar session cluster. This embodiment does not limit this.

[0215] 207. In response to the content reply operation on the reply area, the terminal sends the reply content corresponding to the content reply operation to each target similar to-be-replied session in the session cluster page.

[0216] By introducing the active clustering technology, this application can actively cluster the un-replied user questions, group similar questions together for the butler to process in batches. The system allows the butler to trigger clustering actively at any time, significantly improving the real-time performance, being able to quickly respond to hot issues or urgent needs, and reducing the waiting time of players. At the same time, the butler can select key questions as the trigger points for clustering according to the actual situation, flexibly prioritize and batch process high-priority questions, improving the flexibility of the system in emergencies and enhancing the service efficiency and quality. With the help of advanced natural language processing technology, the system has the ability of deep semantic understanding, can more accurately identify questions with similar intentions, optimizes the clustering results, and reduces the screening pressure. For the diverse expressions of players for the same question, the semantic recognition ability of the system can ensure that similar questions are correctly clustered, ensuring the comprehensiveness of clustering. More importantly, the system utilizes the powerful generalization ability of the large language model, can effectively handle new question types or expression methods that have not been seen before, showing a high degree of adaptability and meeting the ever-changing needs of users.

[0217] As can be seen from the above, in this embodiment, the terminal can select multiple initial candidate reply conversations similar to the target conversation content from multiple candidate reply conversations based on the clustering operation of similar questions for the target conversation content; construct semantic analysis prompt information for the initial candidate reply conversations according to the target conversation content and the initial candidate reply conversations; calculate the semantic relevance between the target conversation content and the initial candidate reply conversations based on the semantic analysis prompt information; determine multiple target similar reply conversations of the target conversation content from the initial candidate reply conversations according to the semantic relevance; perform aggregation processing on the multiple target similar reply conversations to obtain a similar conversation cluster; display a conversation cluster page, where the conversation cluster page includes a conversation content display area for multiple target similar reply conversations in the similar conversation cluster and a reply area for the similar conversation cluster; and in response to a content reply operation on the reply area, send the reply content corresponding to the content reply operation to each target similar reply conversation in the conversation cluster page.

[0218] This application can actively cluster similar candidate reply conversations based on the clustering operation of similar questions. Through active clustering, it can respond to user questions more flexibly and improve the real-time performance of conversation processing. Specifically, during the clustering process, this application first performs a rough screening on the candidate reply conversations to filter out irrelevant conversations, obtaining the initial candidate reply conversations after rough screening. Subsequently, through the construction of semantic analysis prompt information, it realizes deep semantic matching, more accurately identifies conversations with similar intentions, optimizes the clustering effect, ensures that similar questions are correctly clustered, and then performs batch replies to the clustered similar conversations, which can greatly improve the reply efficiency.

[0219] To better implement the above method, an embodiment of this application also provides a conversation processing device, as Figure 3 shown. The conversation processing device may include a selection unit 301, a prompt construction unit 302, a calculation unit 303, a determination unit 304, and an aggregation unit 305, as follows:

[0220] (1) Selection unit 301;

[0221] The selection unit is configured to select multiple initial candidate reply conversations similar to the target conversation content from multiple candidate reply conversations based on the clustering operation of similar questions for the target conversation content.

[0222] Optionally, in some embodiments of this application, the selection unit may include a first display subunit, a second display subunit, and a selection subunit, as follows:

[0223] The first display subunit is configured to display a conversation list page, which includes a first conversation list area and a second conversation list area. The first conversation list area includes candidate conversations with an interaction degree higher than a preset interaction degree, and the second conversation list area includes candidate conversations with an interaction degree not higher than the preset interaction degree;

[0224] The second display subunit is configured to, in response to a selection operation on a target conversation in the first conversation list area, display a conversation page corresponding to the target conversation, where the conversation page includes at least one piece of conversation content of the target conversation;

[0225] The selection subunit is configured to, in response to a similar question clustering operation on the target conversation content in the conversation page, select multiple initial to-be-replied conversations similar to the target conversation content from multiple candidate to-be-replied conversations, where the target conversation content is any conversation content in the conversation page.

[0226] Optionally, in some embodiments of the present application, the selection unit may include a conversation determination subunit and a conversation selection subunit, as follows:

[0227] The conversation determination subunit is configured to, based on a similar question clustering operation on the target conversation content, determine multiple candidate to-be-replied conversations according to an unprocessed conversation queue and currently generated real-time conversations. The unprocessed conversation queue includes conversation identification information of unprocessed conversations within a historical time period;

[0228] The conversation selection subunit is configured to select multiple initial to-be-replied conversations similar to the target conversation from the respective candidate to-be-replied conversations according to the content matching degree between each candidate to-be-replied conversation and the target conversation content.

[0229] (2) The prompt construction unit 302;

[0230] The prompt construction unit is configured to construct semantic analysis prompt information for the initial to-be-replied conversations according to the target conversation content and the initial to-be-replied conversations.

[0231] Optionally, in some embodiments of the present application, the target conversation content is the reply content in the processed conversations;

[0232] The prompt construction unit may include a retrieval subunit, a prompt construction subunit, and a combination subunit, as follows:

[0233] The retrieval subunit is configured to retrieve historical similar conversation content corresponding to the target conversation content in a historical conversation database;

[0234] The prompt construction subunit is configured to construct initial prompt information according to the historical similar conversation content;

[0235] A combinatorial subunit for combinatorially processing the initial prompt information with the conversation content of each initial to-be-replied conversation respectively to obtain semantic analysis prompt information for each initial to-be-replied conversation.

[0236] Optionally, in some embodiments of the present application, the conversation processing device further includes a conversation content separation unit, as follows:

[0237] The conversation content separation unit is configured to perform content separation processing on the target conversation corresponding to the target conversation content to obtain the conversation historical background message, the current problem message, and the reply content of the current problem message of the target conversation;

[0238] The prompt construction subunit may specifically be configured to fuse the conversation historical background message, the current problem message of the target conversation, and the historical similar conversation content to obtain initial prompt information.

[0239] Optionally, in some embodiments of the present application, the target conversation content is an unanswered question message;

[0240] The prompt construction unit may include a fusion processing subunit and a combination processing subunit, as follows:

[0241] The fusion processing subunit is configured to fuse the target conversation content and its conversation historical background message in the corresponding conversation to obtain initial prompt information;

[0242] The combination processing subunit is configured to combinatorially process the initial prompt information with the conversation content of each initial to-be-replied conversation respectively to obtain semantic analysis prompt information for each initial to-be-replied conversation.

[0243] (3) Calculation unit 303;

[0244] The calculation unit is configured to calculate the semantic relevance between the target conversation content and the initial to-be-replied conversation based on the semantic analysis prompt information.

[0245] (4) Determination unit 304;

[0246] The determination unit is configured to determine multiple target similar to-be-replied conversations of the target conversation content from the respective initial to-be-replied conversations according to the semantic relevance.

[0247] (5) Aggregation unit 305;

[0248] The aggregation unit is configured to perform aggregation processing on the multiple target similar to-be-replied conversations to obtain a similar conversation cluster for batch reply processing of the multiple target similar to-be-replied conversations in the similar conversation cluster.

[0249] Optionally, in some embodiments of the present application, the session processing device may further include a display unit and a sending unit, as follows:

[0250] The display unit is configured to display a session cluster page, where the session cluster page includes a session content display area of multiple target similar to-be-replied sessions in the similar session cluster, and a reply area for the similar session cluster;

[0251] The sending unit is configured to, in response to a content reply operation on the reply area, send the reply content corresponding to the content reply operation to each target similar to-be-replied session in the session cluster page.

[0252] Optionally, in some embodiments of the present application, the session cluster page includes a deletion control corresponding to each target similar to-be-replied session in the similar session cluster;

[0253] The session processing device may further include a deletion unit, as follows:

[0254] The deletion unit is configured to, in response to a trigger operation on a target deletion control, delete the target similar to-be-replied session corresponding to the target deletion control from the similar session cluster, and remove the session content display area corresponding to the target deletion control from the session cluster page to update the session cluster page.

[0255] As can be seen from the above, in this embodiment, the selection unit 301 can select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions based on a similar question clustering operation for the target session content; the prompt construction unit 302 constructs semantic analysis prompt information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions; the calculation unit 303 calculates the semantic relevance between the target session content and the initial to-be-replied sessions based on the semantic analysis prompt information; the determination unit 304 determines multiple target similar to-be-replied sessions of the target session content from the respective initial to-be-replied sessions according to the semantic relevance; and the aggregation unit 305 performs an aggregation process on the multiple target similar to-be-replied sessions to obtain a similar session cluster, so as to perform a batch reply process on the multiple target similar to-be-replied sessions in the similar session cluster.

[0256] This application can actively cluster similar sessions to be replied based on similar problem clustering operations. Through active clustering, it can respond to user questions more flexibly and improve the real-time performance of session processing. Specifically, during the clustering process, this application first roughly screens the candidate sessions to be replied to filter out irrelevant sessions, obtaining the initial sessions to be replied after rough screening. Subsequently, through the construction of semantic analysis prompt information, deep semantic matching is achieved, more accurately identifying sessions with similar intentions, optimizing the clustering effect, ensuring that similar problems are correctly clustered, and then batch replying to the clustered similar sessions, which can greatly improve the reply efficiency.

[0257] An embodiment of this application also provides an electronic device, such as Figure 4 shown, which shows the structural schematic diagram of the electronic device involved in the embodiment of this application. This electronic device can be a terminal or a server, etc. Specifically:

[0258] This electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art can understand that Figure 4 the structure of the electronic device shown in

[0259] does not constitute a limitation on the electronic device, and it may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements. Among them:

[0260] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the electronic device. In addition, the memory 402 can include high-speed random access memory and can also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices. Correspondingly, the memory 402 can also include a memory controller to provide the processor 401 with access to the memory 402.

[0261] The electronic device further includes a power supply 403 for supplying power to each component. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to implement functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 403 can also include any components such as one or more DC or AC power supplies, a recharge system, a power failure detection circuit, a power converter or inverter, and a power status indicator.

[0262] The electronic device may further include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0263] Although not shown, the electronic device may further include a display unit, etc., which will not be elaborated here. Specifically, in this embodiment, the processor 401 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402 to implement various functions as follows:

[0264] Based on the clustering operation of similar questions for the target session content, select multiple initial candidate reply sessions similar to the target session content from multiple candidate reply sessions; construct semantic analysis prompt information for the initial candidate reply sessions according to the target session content and the initial candidate reply sessions; calculate the semantic relevance between the target session content and the initial candidate reply sessions based on the semantic analysis prompt information; determine multiple target similar reply sessions of the target session content from the respective initial candidate reply sessions according to the semantic relevance; perform aggregation processing on the multiple target similar reply sessions to obtain a similar session cluster, so as to perform batch reply processing on the multiple target similar reply sessions in the similar session cluster.

[0265] For the specific implementation of each of the above operations, reference may be made to the previous embodiments, which will not be elaborated herein.

[0266] As can be seen from the above, in this embodiment, based on the clustering operation of similar questions for the target conversation content, multiple initial candidate replies similar to the target conversation content can be selected from multiple candidate replies to be replied; according to the target conversation content and the initial candidate replies, semantic analysis prompt information for the initial candidate replies can be constructed; based on the semantic analysis prompt information, the semantic relevance between the target conversation content and the initial candidate replies can be calculated; according to the semantic relevance, multiple target similar candidate replies of the target conversation content can be determined from the respective initial candidate replies; and the multiple target similar candidate replies can be aggregated to obtain a similar conversation cluster, so as to perform batch reply processing on the multiple target similar candidate replies in the similar conversation cluster.

[0267] This application can actively cluster similar candidate replies to be replied based on the clustering operation of similar questions. Through active clustering, the user questions can be responded to more flexibly, and the real-time performance of conversation processing can be improved. Specifically, during the clustering process, this application first performs a rough screening on the candidate replies to be replied to filter out irrelevant conversations, obtaining the initial candidate replies after rough screening. Subsequently, through the construction of semantic analysis prompt information, deep semantic matching is achieved, and conversations with similar intents can be more accurately identified, optimizing the clustering effect and ensuring that similar questions are correctly clustered. Furthermore, batch replies are made to the clustered similar conversations, which can greatly improve the reply efficiency.

[0268] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions or by controlling related hardware through instructions. The instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0269] Therefore, an embodiment of this application provides a computer-readable storage medium, which stores multiple instructions that can be loaded by a processor to execute the steps in any one of the conversation processing methods provided by the embodiments of this application. For example, the instructions can execute the following steps:

[0270] Based on the clustering operation of similar questions for the target session content, select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions; construct semantic analysis prompt information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions; calculate the semantic relevance between the target session content and the initial to-be-replied sessions based on the semantic analysis prompt information; determine multiple target similar to-be-replied sessions of the target session content from the respective initial to-be-replied sessions according to the semantic relevance; perform an aggregation process on the multiple target similar to-be-replied sessions to obtain a similar session cluster, so as to perform a batch reply process on the multiple target similar to-be-replied sessions in the similar session cluster.

[0271] For the specific implementation of each of the above operations, reference may be made to the previous embodiments and will not be elaborated here.

[0272] Among them, the computer-readable storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc.

[0273] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the session processing methods provided in the embodiments of the present application, the beneficial effects achievable by any of the session processing methods provided in the embodiments of the present application can be achieved. For details, reference may be made to the previous embodiments and will not be elaborated here.

[0274] According to one aspect of the present application, there is provided a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementation manners of the above session processing aspect.

[0275] The above has introduced in detail a session processing method and related devices provided in the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A session processing method, characterized in that including: selecting, from multiple candidate to-be-replied conversations, multiple initial to-be-replied conversations similar to the target conversation content based on a similar question clustering operation for the target conversation content; constructing, according to the target conversation content and the initial to-be-replied conversations, semantic analysis prompt information for the initial to-be-replied conversations; calculating, based on the semantic analysis prompt information, the semantic relevance between the target conversation content and the initial to-be-replied conversations; determining, according to the semantic relevance, multiple target similar to-be-replied conversations of the target conversation content from the respective initial to-be-replied conversations; performing an aggregation process on the multiple target similar to-be-replied conversations to obtain a similar conversation cluster, so as to perform a batch reply process on the multiple target similar to-be-replied conversations in the similar conversation cluster.

2. The method according to claim 1, wherein The target conversation content is the reply content in the processed conversation; The constructing, according to the target conversation content and the initial to-be-replied conversations, semantic analysis prompt information for the initial to-be-replied conversations includes: retrieving, in a historical conversation database, historical similar conversation content corresponding to the target conversation content; constructing initial prompt information according to the historical similar conversation content; performing a combination process on the initial prompt information and the conversation content of each initial to-be-replied conversation respectively to obtain semantic analysis prompt information for each initial to-be-replied conversation.

3. The method according to claim 2, wherein Before the constructing initial prompt information according to the historical similar conversation content, it further includes: performing a content separation process on the target conversation corresponding to the target conversation content to obtain the conversation historical background message, the current question message, and the reply content of the current question message of the target conversation; The constructing initial prompt information according to the historical similar conversation content includes: performing a fusion process on the conversation historical background message, the current question message of the target conversation, and the historical similar conversation content to obtain initial prompt information.

4. The method according to claim 1, characterized in that The target conversation content is an unanswered question message; The constructing, according to the target conversation content and the initial to-be-replied conversations, semantic analysis prompt information for the initial to-be-replied conversations includes: performing a fusion process on the target conversation content and its conversation historical background message in the corresponding conversation to obtain initial prompt information; performing a combination process on the initial prompt information and the conversation content of each initial to-be-replied conversation respectively to obtain semantic analysis prompt information for each initial to-be-replied conversation.

5. The method according to claim 1, wherein The selecting, from multiple candidate to-be-replied conversations, multiple initial to-be-replied conversations similar to the target conversation content based on a similar question clustering operation for the target conversation content includes: displaying a conversation list page, where the conversation list page includes a first conversation list area and a second conversation list area, the first conversation list area includes candidate conversations with an interaction degree higher than a preset interaction degree, and the second conversation list area includes candidate conversations with an interaction degree not higher than the preset interaction degree; responding to a selection operation on a target conversation in the first conversation list area, and displaying a conversation page corresponding to the target conversation, where the conversation page includes at least one piece of conversation content of the target conversation; In response to a clustering operation on similar questions of the target session content in the session page, select multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions, where the target session content is any session content in the session page.

6. The method according to claim 1, wherein After aggregating and processing the multiple target similar to-be-replied sessions to obtain a similar session cluster, it further includes: Display a session cluster page, where the session cluster page includes a session content display area for multiple target similar to-be-replied sessions in the similar session cluster and a reply area for the similar session cluster; In response to a content reply operation on the reply area, send the reply content corresponding to the content reply operation to each target similar to-be-replied session in the session cluster page.

7. The method according to claim 6, wherein The session cluster page includes a deletion control corresponding to each target similar to-be-replied session in the similar session cluster; the method further includes: In response to a trigger operation on a target deletion control, delete the target similar to-be-replied session corresponding to the target deletion control from the similar session cluster, and remove the session content display area corresponding to the target deletion control from the session cluster page to update the session cluster page.

8. The method according to claim 1, characterized in that, The step of selecting multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions based on a clustering operation on similar questions of the target session content includes: Based on a clustering operation on similar questions of the target session content, determine multiple candidate to-be-replied sessions according to an unprocessed session queue and currently generated real-time sessions, where the unprocessed session queue includes session identification information of unprocessed sessions within a historical time period; Select multiple initial to-be-replied sessions similar to the target session from each of the candidate to-be-replied sessions according to the content matching degree between each candidate to-be-replied session and the target session content.

9. A session processing device, characterized in that, It includes: A selection unit for selecting multiple initial to-be-replied sessions similar to the target session content from multiple candidate to-be-replied sessions based on a clustering operation on similar questions of the target session content; A hint construction unit for constructing semantic analysis hint information for the initial to-be-replied sessions according to the target session content and the initial to-be-replied sessions; A calculation unit for calculating the semantic correlation degree between the target session content and the initial to-be-replied sessions based on the semantic analysis hint information; A determination unit for determining multiple target similar to-be-replied sessions of the target session content from each of the initial to-be-replied sessions according to the semantic correlation degree; An aggregation unit for aggregating and processing the multiple target similar to-be-replied sessions to obtain a similar session cluster for batch reply processing of the multiple target similar to-be-replied sessions in the similar session cluster.

10. An electronic device, characterized in that, It includes a memory and a processor; the memory stores an application program, and the processor is used to run the application program in the memory to execute the operations in the session processing method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores multiple instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the session processing method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instruction is executed by a processor, it implements the steps in the session processing method according to any one of claims 1 to 8.

Images