A customer service information generation method and system based on historical conversations

Abstract

The present application discloses a method and system for generating customer service information based on historical conversations. The method includes: obtaining historical conversation data; performing a vectorization operation on the historical conversation data to obtain a first target corpus; clustering the question sentences included in the first target corpus according to vector similarity to obtain corpus clusters; determining the corpus clusters with a cluster size reaching a preset size threshold as target clusters, and determining the target question sentences corresponding to the target clusters; determining the conversation context of the target question sentences in the first target corpus, and generating target reply information based on the conversation context; obtaining the current customer service conversation data, and performing a vectorization operation on the customer service conversation data to obtain a second target corpus; determining the target question sentences and target reply information corresponding to the second target corpus, and generating customer service information. The present application greatly improves the AI configuration efficiency of the system and can accurately judge the user's intention, improving the accuracy of customer service information.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular, to a method for generating customer service information based on historical conversations, a method for generating customer service information based on classification scenarios, and corresponding systems. Background Art

[0002] With the development of artificial intelligence technology, the automatic question answering module has gradually become an important part of the system for information retrieval and human-computer interaction. In existing systems, the automatic question answering module usually adopts the methods of keyword recognition and text similarity comparison. After querying for the answers to similar questions in the historical conversation, the answers are fed back to the user.

[0003] However, the processing methods of the above automatic question answering module have the following problems: 1. Querying questions and answers based on keyword matching is difficult to apply in multi-language scenarios, and a large amount of human cost is required to label keywords in minority languages; 2. Sentences with relatively high or low text similarity may all express very different meanings; 3. For the user questions queried by the methods of keyword recognition and text similarity comparison, the question accuracy is relatively low, and due to the failure to combine the conversation context, the aggregated answer information is prone to deviation.

[0004] In addition, when the automatic question answering module calls a generative large model to generate answer information, there are situations where the model misidentifies the user's intention, resulting in the generated answer information not matching the user's question and failing to meet the needs of buyer users; or, due to the hallucination problem of the model, the answer information has an obvious deviation from the actual product information, misleading the buyer users.

[0005] Therefore, how to improve the accuracy of the system in generating customer service information is a difficult problem that urgently needs to be solved in the current e-commerce field.

[0006] Other technical problems related to this application will be further elaborated later. The above content is only used to assist in understanding the technical solution of this application, and it does not mean that all of the above content is prior art. Summary of the Invention

[0007] The main purpose of the present invention is to provide a method and system for generating customer service information based on historical conversations, which can accurately judge the user's intention by combining historical conversation data and user preferences in multi-language scenarios and automatically generate customer service information with a high degree of matching, thereby greatly improving the accuracy of customer service information. In addition, this application also provides a method and system for generating customer service information based on classification scenarios, which can, on the basis of accurately identifying the user's intention, perform scenario classification operations on the user's intention, thereby reducing the hallucination impact of the model and greatly improving the accuracy of customer service information.

[0008] To achieve the above object, the present application proposes a customer service information generation method and system based on historical conversations, and the method includes:

[0009] Step S1: Obtain historical conversation data, where the historical conversation data represents the conversation content between a buyer user and a customer service account;

[0010] Step S2: Perform a vectorization operation on the historical conversation data to obtain a first target corpus;

[0011] Step S3: Cluster the question sentences included in the first target corpus according to vector similarity to obtain corpus clusters;

[0012] Step S4: Determine the corpus clusters with a cluster size reaching a preset size threshold as target clusters, and determine the target question sentences corresponding to the target clusters;

[0013] Step S5: Determine the conversation context of the target question sentences in the first target corpus, and generate target reply information corresponding to the target question sentences based on the conversation context;

[0014] Step S6: Obtain current customer service conversation data, and perform a vectorization operation on the customer service conversation data to obtain a second target corpus;

[0015] Step S7: Determine the target question sentences and target reply information corresponding to the second target corpus, and generate customer service information based on the target question sentences and the target reply information.

[0016] Other features and technical effects of the present application are described in the following part of the specification. The technical problem-solving ideas and related product design solutions of the present application are as follows:

[0017] Currently, the automatic question-answering module of the system usually adopts keyword recognition and text similarity comparison methods to query similar questions and question answers in historical conversations. However, the above processing methods of the automatic question-answering module have the following problems:

[0018] 1. Querying questions and answers based on keyword matching is difficult to apply in a multilingual scenario, and it requires a large amount of human cost to label keywords in minority languages. For example, cross-border e-commerce sellers need to label keywords in various languages when dealing with consultation information from customers in multiple countries, and the personnel performing the labeling operation need to be highly proficient in that language to ensure the accuracy of the labeling operation. Therefore, the requirement for human cost is relatively high.

[0019] 2. Sentences with high or low text similarity may express very different meanings. For example, "I won't eat anymore" and "I can't eat anymore" have high text similarity but express very different meanings; another example is "I want to buy an anti-hair loss shampoo" and "Do you have any anti-hair loss shampoo", which have low text similarity but express very similar meanings. Therefore, it is difficult to ensure the accuracy of the reply information by querying similar questions and question replies based on text similarity.

[0020] 3. For user questions queried by means of keyword recognition and text similarity comparison, the question accuracy is low, and the similar questions and question replies queried do not combine the conversation context, so it is easy to cause deviations in aggregating multiple question replies to obtain the final reply.

[0021] The applicant found that by performing vectorization operations on the conversation data of the system based on the Embedding encoding model, the conversation data in each language is converted into vector representations in the same vector space (i.e., the target corpus), and then the target corpus is analyzed, thus avoiding the situation where cross-border e-commerce sellers need to perform corpus annotation operations according to each language respectively, and reducing the labor cost. Moreover, since the original word order structure is broken after the conversation data is converted into the target corpus through vectorization operations, the semantic expression is not limited by the word order structure. Therefore, for sentences with high text similarity but very different expressed meanings, or sentences with low text similarity but very similar expressed meanings, the system can also accurately distinguish and identify them.

[0022] Furthermore, the applicant also found that the low accuracy of the customer service information generated by the system is largely due to the inability to accurately judge the user's intention in the conversation scenario, and thus it is difficult to locate the user's question sentence.

[0023] On this basis, the applicant proposed that using the community discovery algorithm to cluster the question sentences included in the target corpus to obtain multiple corpus clusters, and determining the target question sentences corresponding to the corpus clusters with a larger cluster size can greatly improve the AI configuration efficiency of the system and reduce the labor cost.

[0024] It can be understood that since the corpus clusters correspond to the question sentences, the corpus clusters with a larger cluster size indicate that the occurrence frequency of some question sentences is higher, and e-commerce sellers need to provide corresponding reply information for this part of the question sentences for the system to call the AI model to generate customer service information. Moreover, for question sentences with a lower occurrence frequency, seller users can choose to postpone providing the corresponding reply information, so as to apply human resources to the questions that users are more concerned about.

[0025] Further, the system obtains the current session data, performs a vectorization operation on the current session data, and then determines the target question statement corresponding to the current session data; queries the session context associated with the target question statement in the historical session data, and determines the reply information corresponding to the target question statement based on the session context; finally, the system generates the customer service information based on the reply information and displays the customer service information to the user, so as to accurately identify the user's intention by combining the session context and the target question statement, determine the corresponding target question statement and reply information according to the user's intention, and make the finally generated customer service information highly accurate.

[0026] In this way, the customer service information generation method based on historical sessions provided by the present application, when applied to a multilingual scenario, by vectorizing the session data of the system and clustering the question statements of the session data using the community discovery algorithm, can greatly improve the AI configuration efficiency of the system, avoid strong manual intervention, and reduce labor costs; and accurately judge the user's intention by combining the target question statement of the session context, and automatically generate customer service information with a high matching degree, thereby greatly improving the accuracy of the customer service information.

[0027] Further, the present application also provides a customer service information generation method based on a classification scenario, which is applied to the AI module of the system. The method includes:

[0028] Step M1: Obtain the current session data based on a preset memory window;

[0029] Step M2: Perform a vectorization operation on the current session data to obtain the target corpus;

[0030] Step M3: Invoke the AI model to analyze the target corpus, and classify the current session data into a preset scenario according to the analysis result. The preset scenario includes a first scenario, a second scenario, and a third scenario. The first scenario generates customer service information based on preset configuration information, the second scenario generates customer service information based on a preset knowledge base, and the third scenario generates customer service information based on a preset information collection program;

[0031] Step M4: Generate the customer service information corresponding to the current session data based on the preset scenario.

[0032] Other features and technical effects of the present application are described in the later part of the specification. The technical problem-solving ideas and related product design solutions of the present application are as follows:

[0033] The existing intelligent customer service system for cross-border e-commerce generates customer service information using an intention recognition method. Usually, an AI model that supports multiple languages is trained. The AI model recognizes the user's question information and generates reply information related to the user's question information.

[0034] However, the above-mentioned method for generating customer service information has the following defects:

[0035] 1. The model is prone to misidentifying the user's intention, resulting in the generated reply information not matching the user's question.

[0036] 2. The model has a hallucination problem, and the reply information generated based on hallucination speculation has a significant deviation from the actual product information, misleading the buyer users.

[0037] 3. A general model needs to generate customer service information for user questions in multiple scenarios. Therefore, the training process requires a large amount of corpus, resulting in low model training efficiency and poor training effect.

[0038] The applicant found that the main reasons for the model's misidentification of the user's intention include that the model fails to analyze the user's question in combination with the context of the conversation data. Due to missing some information sent by the user, the model's recognition result of the user's intention is deviated.

[0039] Then, since the AI model is essentially a reasoning machine, when generating reply information, it tends to generate information with a strong correlation and high probability with the input information. And because the current e-commerce customer service system uses a general AI model to generate customer service information for user questions in multiple scenarios, the information inferred by the AI model based on the user's question is too much and too miscellaneous, resulting in the AI model may finally generate reply information that does not match the actual product information due to hallucination (that is, overusing the imagination ability).

[0040] In addition, since a general AI model needs to generate customer service information for user questions in multiple scenarios, it is necessary to provide the general AI model with training corpus in multiple scenarios; and in cross-border e-commerce business, it is necessary to provide the corresponding training corpus in each language respectively, resulting in low model training efficiency and poor training effect.

[0041] On this basis, the applicant proposes to obtain multi-round conversation data of the current conversation data based on a preset memory window, so that when the model identifies the user's intention for the user's question, it can combine the conversation context related to the user's question, thereby improving the accuracy of user intention recognition.

[0042] Furthermore, the applicant proposes to classify the user's question by scenario, and then adopt corresponding reply strategies for different scenarios; and only call the model to generate reply information for some of the scenarios, and distinguish the way of generating reply information by the model according to the scenario type, which can effectively reduce the negative impact caused by the model's hallucination and greatly improve the accuracy of the model's generated customer service information.

[0043] Specifically, this application classifies user questions into a first scenario, a second scenario, and a third scenario. First, the first scenario does not involve complex information and only requires a response to fixed terms or a response after querying status information. Therefore, this application uses a classification model to classify user questions in the first scenario and generates customer service information containing preset configuration information based on the classification results. Second, the second scenario involves complex information. This application uses a generative large model to match the user questions in the second scenario with a preset knowledge base, and thus generates corresponding customer service information based on the preset knowledge base; among them, the preset knowledge base uses historical conversation data as training corpus to generate corresponding target question statements and target reply information. Finally, the third scenario is an information collection and feedback scenario, which collects information provided by users based on a preset information collection program, including unpacking videos or user complaints, and then executes corresponding processing procedures (such as refund and return) or transfers to a human customer service.

[0044] It can be understood that in the above processing scenarios for user questions, only the first scenario and the second scenario need to call the model, and the first scenario only requires the model to provide pre-configured information, and the difficulty of model training and model generation of customer service information is very low; while the second scenario requires the model to answer user questions based on a preset knowledge base to solve complex problems of user questions, with strong targetability and relatively less training corpus required for the model.

[0045] In this way, based on accurately identifying the user's intention, this application can further perform scenario classification operations on the user's intention, adopt corresponding reply strategies for different scenarios, only call the model to generate reply information for some of the scenarios, and distinguish the way of generating reply information by the model according to the scenario type, thereby reducing the impact of model hallucinations, greatly improving the accuracy of customer service information, and improving model training efficiency and reducing model training costs.

[0046] This application also provides a server, which includes a memory and a processor. The system in this application is stored in the memory, and the processor can run the operation instructions of each method step of this application.

[0047] This application also provides a computer device, which includes a memory and a processor. The system in this application is stored in the memory, and the processor can run the operation instructions of each method step of this application. Brief Description of the Drawings

[0048] The drawings are used to provide a further understanding of this application and do not constitute a limitation to this application; the content shown in the drawings can be the actual data of the embodiments and belongs to the protection scope of this application.

[0049] Figure 1 It is a schematic diagram of the process of generating customer service information by the customer service system after clustering question statements in an embodiment of this application.

[0050] Figure 2 This is a schematic flowchart of a method for generating customer service information based on historical conversations in an embodiment of the present application.

[0051] Figure 3 In an embodiment of the present application, this is a schematic diagram of the process of a customer service system generating customer service information based on classifying preset scenarios.

[0052] Figure 4 This is a schematic flowchart of a method for generating customer service information based on classified scenarios in an embodiment of the present application. Detailed implementation manners

[0053] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the following further elaborates on the embodiments of the present application in detail through specific implementation manners in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] Figure 1 This is a schematic diagram of the process of an e-commerce customer service system (hereinafter simply referred to as the system) in an embodiment of the present application generating customer service information after clustering question statements. As Figure 1 shown, the AI module of the system obtains the historical conversation data of the customer service module and performs a vectorization operation on the historical conversation data to obtain a first target corpus. Then, the AI module uses a community discovery algorithm to cluster the user question statements in the first target corpus to obtain multiple corpus clusters. Among them, the corpus cluster with a larger cluster size represents the user question statements with a higher occurrence frequency. Next, the AI module selects the corpus cluster with a larger cluster size and determines a target question statement from this corpus cluster, that is, the occurrence frequency of this target question statement is higher and it has higher value. Then, the AI module searches for the corresponding context conversation data for this target question statement (such as after a buyer user previously consulted this type of question and the customer service staff gave an answer), and generates a target reply information corresponding to this target question statement based on this context conversation data. When the system receives the consultation information from the buyer user, the system performs a vectorization operation on the current conversation data to obtain a second target corpus; and, the AI module matches the second target corpus with the target question statement, and then determines the target reply information; finally, the AI module generates a customer service information for replying to the buyer user based on this target reply information.

[0055] The e-commerce customer service system of the present application can be a subsystem of an e-commerce ERP system, an e-commerce platform system or other software systems, or can also be an independent customer service system.

[0056] Figure 2 Shows a method for generating customer service information based on historical conversations in an embodiment of the present application, which mainly includes the following steps S1 to S7.

[0057] Step S1: Obtain historical conversation data, where the historical conversation data characterizes the conversation content between the buyer user and the customer service account.

[0058] Specifically, the system obtains the historical conversation data of the customer service module, that is, the historical conversation data characterizes the conversation content between the buyer user and the customer service account, including the question statements of the buyer user and the reply statements of the customer service staff.

[0059] Step S2: Perform a vectorization operation on the historical conversation data to obtain a first target corpus.

[0060] Specifically, the system uses an AI model to perform a vectorization operation on the historical conversation data, that is, to convert the conversation data in each language in the customer service module into a vector representation in the same vector space, so that the semantic expression of the historical conversation data is not limited by the word order structure.

[0061] Step S3: Cluster the question statements included in the first target corpus according to the vector similarity to obtain corpus clusters.

[0062] Specifically, the AI module uses a community discovery algorithm to cluster the user question statements of the first target corpus according to the vector similarity to obtain multiple corpus clusters; the multiple corpus clusters characterize various inquiry statements proposed by the buyer user.

[0063] Step S4: Determine the corpus clusters whose cluster sizes reach a preset size threshold as target clusters, and determine the target question statements corresponding to the target clusters.

[0064] It can be understood that since the corpus clusters correspond to the question statements, the corpus clusters with larger cluster sizes characterize that the occurrence frequencies of some question statements are higher, that is, these user question statements have higher value.

[0065] Therefore, in this embodiment, the target question statements are determined according to the corpus clusters with larger cluster sizes, and the corresponding target reply information is preferentially generated for the target question statements for the system to call the AI model to generate customer service information. For the question statements with lower occurrence frequencies, the seller user can choose to defer providing the corresponding reply information, so as to apply the AI configuration resources to the questions that the users are more concerned about.

[0066] Step S5: Determine the conversation context of the target question statement in the first target corpus, and generate target reply information corresponding to the target question statement based on the conversation context.

[0067] Specifically, after the system determines the target question statements with a higher occurrence frequency, it searches for the conversation context related to the target question statements in the first target corpus, extracts the reply statements of the customer service staff from the conversation context, and then generates the corresponding target reply information based on the reply statements.

[0068] Step S6: Obtain the current customer service conversation data, and perform a vectorization operation on the customer service conversation data to obtain a second target corpus.

[0069] Specifically, when the system receives the consultation information from the buyer user, the system performs a vectorization operation on the current conversation data, so that the current conversation data is converted into a one-dimensional vector representation within the vector space where the first target corpus is located, that is, the current conversation data is converted into a second target corpus.

[0070] Step S7: Determine the target question statements and target reply information corresponding to the second target corpus, and generate customer service information based on the target question statements and the target reply information.

[0071] Specifically, the system calls the AI model to analyze the user intention of the above-mentioned second target corpus, determines the target question statements and target reply information corresponding to the user intention, and then generates the customer service information for replying to the buyer user by combining the target question statements and the target reply information.

[0072] As a feasible implementation manner, the system performs a word frequency analysis on the second target corpus. For the functions or requirements repeatedly mentioned by the buyer user, the content related to the function or requirement is emphasized in the generated customer service information. For example, if the buyer user mentions the camera function of the mobile phone many times in the conversation, the content about the camera function of the mobile phone will occupy a large proportion in the generated customer service information.

[0073] Further, in an embodiment, step S2 includes the following steps S2.1 and S2.2.

[0074] Step S2.1: Based on the historical conversation data, obtain the product conversation data related to the product information;

[0075] Step S2.2: Perform a vectorization operation on the product conversation data to obtain a first target corpus.

[0076] Specifically, the customer service conversation data of the system usually contains a large amount of meaningless content, including the information content such as greetings, chatting, and expressing emotions sent by the buyer user. In order to improve the quality of the training corpus for training the AI module, improve the training efficiency of the AI model and improve the training effect, this embodiment performs a screening operation on the historical conversation data of the customer service module, only obtains the product conversation data related to the product information, and performs a vectorization operation on the product conversation data to obtain the first target corpus for configuring the AI model.

[0077] Further, on the basis of the above embodiments, step S2.2 includes the following steps S2.21 and S2.22.

[0078] Step S2.21: Perform a filtering operation on the commodity conversation data based on the similarity of the conversation subject and conversation data;

[0079] Step S2.22: Perform a vectorization operation on the filtered commodity conversation data to obtain a first target corpus.

[0080] Specifically, in the customer service conversation data of the system, the customer service module will send an automatic reply to the buyer user in the case where the manual customer service cannot handle it, the AI model cannot recognize and answer, or a specific statement is received. It can be understood that the frequency of automatic replies usually is relatively high and has no practical significance for AI configuration. In this embodiment, a filtering operation is performed based on the similarity of the conversation subject and conversation data, that is, highly repetitive statements sent by the customer service account are determined as automatic replies, and the relevant conversation data of the automatic replies is filtered. Finally, a vectorization operation is performed on the filtered commodity conversation data, thereby obtaining a higher-quality first target corpus.

[0081] Further, in one embodiment, step S5 includes the following steps S5.1 and S5.2.

[0082] Step S5.1: In the first target corpus, determine the conversation context of the target question statement for the conversation data whose time interval from the target question statement is within a preset time range;

[0083] Step S5.2: Generate target reply information corresponding to the target question statement based on the conversation context.

[0084] Specifically, the system searches in the first target corpus for conversation data corresponding to the target question statement within a preset time range. For example, the system takes the conversation data with a time interval of 6 hours above and below the target question statement, and then determines the customer service reply statement corresponding to the target question statement within this conversation data, and aggregates the customer service reply statements to generate target reply information.

[0085] In this way, this embodiment limits that when the system configures target reply information for the target question statement, the target reply information needs to be generated based on the context within the historical conversation scenario, so as to ensure the integrity and accuracy of the target reply information to the greatest extent.

[0086] Further, on the basis of the above embodiments, step S5.2 includes the following steps S5.21 to S5.23.

[0087] Step S5.21: Among multiple conversation scenarios in the first target corpus, determine corresponding candidate response information based on the conversation context;

[0088] Step S5.22: If it is detected that there are multiple candidate response information, push the multiple candidate response information to the user confirmation interface;

[0089] Step S5.23: In response to the first confirmation instruction from the user confirmation interface, determine the candidate response information corresponding to the first confirmation instruction as the target response information.

[0090] In this embodiment, the system searches in the first target corpus and finds multiple conversation scenarios corresponding to the target question statement. The conversation scenario contains the conversation context related to the target question statement; moreover, the system aggregates multiple candidate response information obtained from the multiple conversation scenarios, and the multiple candidate response information is not completely consistent, and even expresses mutually exclusive and opposing meanings. On this basis, the system pushes the multiple candidate response information to the user confirmation interface, and the user confirmation interface is operated and managed by the buyer user. When the buyer user interacts with the user confirmation interface and triggers the first confirmation instruction to select the corresponding candidate response information, the system determines the candidate response information corresponding to the first confirmation instruction as the target response information.

[0091] In this way, in this embodiment, if the system aggregates multiple candidate response information, the decision-making power for the candidate response information is given to the buyer user, so that the customer service information generated by the system conforms to the original intention of the buyer user to the greatest extent.

[0092] Further, on the basis of the above embodiment, after step S5.21, when it is detected that there are multiple candidate response information, the method further includes the following steps S5.24 and S5.25.

[0093] Step S5.24: Detect the occurrence frequency of the multiple candidate response information, and perform a sorting operation on the multiple candidate response information based on the occurrence frequency;

[0094] Step S5.25: Determine the target response information based on the sorting result.

[0095] As a feasible implementation manner, when the system aggregates multiple candidate response information and the buyer user cannot select and confirm the multiple candidate response information immediately, the system detects the occurrence frequency of the multiple candidate response information, and determines the candidate response information with a higher occurrence frequency as the target response information. For example, if the system aggregates response information A, response information B, and response information C, and response information B appears the most times in the historical conversation data, the system determines that the seller's customer service staff tends to approve of response information B and determines response information B as the target response information.

[0096] Further, in one embodiment, after step S7, the method further includes the following steps S7.1 and S7.2.

[0097] Step S7.1: Push the customer service information to the user confirmation interface;

[0098] Step S7.2: In response to a second confirmation instruction from the user confirmation interface, send the customer service information to the buyer user.

[0099] Specifically, in this embodiment, after the system automatically generates the customer service information corresponding to the current session data, it pushes the customer service information to the user confirmation interface operated and managed by the seller user, and only after obtaining the confirmation of the seller user, it is sent to the buyer user. In this way, this embodiment gives the seller user the right to decide on the reply to the buyer user, so that the customer service information for replying to the buyer user can respect the wishes of the seller user to the greatest extent.

[0100] As a feasible implementation manner, when the system detects that there is a preset authorization agreement between the system and the seller user account, the system directly sends the automatically generated customer service information to the buyer user without the need for the seller user to confirm, thereby improving the reply efficiency of the customer service information.

[0101] Further, in one embodiment, after step S6, the method further includes the following steps S7.3 to S7.5.

[0102] Step S7.3: When the target question statement and the target reply information corresponding to the second target corpus cannot be detected, generate a preset question statement corresponding to the second target corpus;

[0103] Step S7.4: Mark the preset question statement; and,

[0104] Step S7.5: Send a preset reply statement to the buyer user, or transfer the current customer session to a manual reply.

[0105] In this embodiment, the system cannot find the target question statement and the target reply information that match the second target corpus, that is, the user question statement included in the second target corpus is relatively niche, and the AI module has not configured the corresponding reply information for this user question statement; at this time, the system analyzes the second target corpus, generates the corresponding preset question statement, and performs the marking operation. In the subsequent operation process, if the customer service staff replies to a similar question with a marked preset question statement, the system will establish an association relationship between the reply content provided by the customer service staff and the preset question statement, and use it as training corpus to further train the AI model of the customer service module.

[0106] Thus, in this embodiment, for user questions that the AI model of the system cannot answer immediately, they are recorded in a marked form. After the customer service staff answers similar questions subsequently, the user questions and the customer service staff's responses are used as new training materials to train the AI model, enabling the AI model to be continuously iteratively optimized.

[0107] Further, in one embodiment, step S7 includes steps S7.6 to S7.8 as follows.

[0108] Step S7.6: Detect the number of characters in the user statement in the second target corpus, and determine a memory window based on the number of characters. The memory window includes multiple rounds of conversations in the second target corpus, and the number of characters in the user statement is negatively correlated with the size of the memory window.

[0109] Step S7.7: Analyze the conversation data included in the memory window based on a preset attenuation factor, where the preset attenuation factor characterizes the weight change rule corresponding to each round of conversation in the memory window.

[0110] Step S7.8: Determine a target question statement and target reply information corresponding to the memory window according to the analysis result, and generate customer service information based on the target question statement and the target reply information.

[0111] The applicant found that when the customer service module of the system receives information sent by buyer users, due to the different communication habits of each buyer user, the system cannot completely and accurately identify the user's intention by using a fixed memory window in combination with the conversation context.

[0112] For example, the communication habit of buyer user A is: "I want a mobile phone with top-notch chip performance and good camera performance" (the first line of information), "The budget for the mobile phone is below 6000 yuan" (the second line of information).

[0113] The communication habit of buyer user B is: "I want a mobile phone" (the first line of information), "With top-notch chip performance" (the second line of information), "Good camera performance" (the third line of information), "Mobile phone budget" (the fourth line of information), "Below 6000 yuan" (the fifth line of information).

[0114] It can be seen that the above-mentioned Buyer A is used to sending the requirements after editing them completely in one sentence, while Buyer B is used to sending the requirements separately after distributing them in multiple sentences. If the customer service module uses a memory window of a fixed size in combination with the conversation context, such as a memory window including three rounds of conversation in the context, the customer service module will miss some information when identifying the intention of Buyer B (information outside the scope of three rounds of conversation will not be recognized by the customer service module), resulting in a deviation in the generated customer service information for replying to Buyer B. In addition, if the customer service module uses a larger memory window, such as a memory window including thirty rounds of conversation in the context, it will lead to a large amount of data processed by the system, imposing a greater burden on the system operation, and causing an obvious delay in the buyer's reply to the customer service information, seriously affecting the user experience.

[0115] On this basis, the applicant proposes to adjust the memory window according to the number of characters in the user's statement, so as to ensure that the customer service module can completely and accurately identify the intention of the buyer user on the premise of avoiding overburdening the system.

[0116] For example, the customer service module uses an initial memory window including three rounds of conversation in the context. However, when the customer service module detects that the number of characters in a single statement of Buyer B is less than 10 characters, the customer service module automatically adjusts the memory window to include five rounds of conversation in the context, or eight rounds of conversation in the context, so that the system can completely and accurately identify the intention of Buyer B on the premise of not processing too much data.

[0117] Furthermore, the applicant finds that when the customer service module uses a larger memory window, since the memory window includes more rounds of conversation, it is easy for the AI model to have difficulty grasping the key points of the conversation, resulting in a deviation in the AI model's recognition of the user's intention. Therefore, the applicant proposes that the customer service module analyzes the conversation data of the memory window based on a preset attenuation factor, that is, assigns different weights to each round of conversation included in the memory window according to the preset attenuation factor, so that the recognition result of the AI model for the user's intention can be closer to the current user's needs.

[0118] For example, assume that the preset attenuation factor is 0.8, the weight of the current round of conversation information is 1, the weight of the previous round is 0.8, and the weight of the round before that is 0.64, and so on. When the AI model processes the user's intention and generates an answer, weighted processing is performed according to the weight of the information, so that the recent conversation information has a greater impact on the current decision of the AI model.

[0119] In this way, the customer service module of this embodiment adjusts the size of the memory window according to the number of characters in the user's statement, and when the memory window includes more rounds of conversation, assigns weights to each round of conversation based on the preset attenuation factor, so as to ensure that the AI model can completely obtain the question information sent by the user and accurately analyze the user's intention, making the finally generated customer service information highly accurate.

[0120] As a feasible implementation, the customer service module marks session information such as the current user intention, consultation type, and user feedback for the conversation content included in the above memory window, and displays the session information on the background visualization interface, so that the customer service staff of the buyer user can understand the session situation in real time and optimize and upgrade the AI model.

[0121] As a feasible implementation, the customer service module configures anomaly detection rules. For example, when the system cannot accurately understand the buyer user's question for a long time (the confidence level of intention recognition is lower than a certain threshold) or the number of dialogue turns is too many but no effective result is achieved (such as the product is not successfully recommended or the problem is not solved, resulting in the user asking repeatedly), an anomaly warning is triggered. For example, if after 5 consecutive rounds of conversation, the confidence level of the system's recognition of the user intention is lower than 30%, a warning notice is sent to the customer service staff, indicating that manual intervention may be required.

[0122] In addition, the customer service module monitors the change of the user's mood. When it detects that the user is excited or dissatisfied (for example, words or sentence patterns expressing dissatisfaction appear), the current session is marked, and the dialogue strategy is adjusted in a timely manner to give priority to handling the user's emotional problems: such as adjusting the dialogue style of the customer service module, providing solutions, or arranging for a human customer service to respond first.

[0123] In one embodiment, the customer service information is generated based on the system calling the AI model. There is a hallucination problem in the AI model, resulting in an obvious deviation between the reply information and the actual product information; the present application also provides a method for generating customer service information based on classification scenarios.

[0124] Figure 3 It shows a schematic diagram of the process of the system generating customer service information based on classifying preset scenarios in an embodiment of the present application. As Figure 3 shown, the AI module of the system obtains the current session data of the customer service module based on a memory window of a preset size, and performs a vectorization operation on the current session data to obtain the target corpus. Then, the AI module performs user intention recognition on the target corpus, and classifies the current session into the first scenario, the second scenario, and the third scenario according to the recognition result. Among them, the first scenario does not involve complex information, and the target reply information is determined based on the preset configuration information; the second scenario involves complex information, and a generative large model is used to determine the target reply information based on the preset knowledge base; the third scenario is an information collection and feedback scenario, and the target reply information is determined based on the preset information collection program. Finally, the system generates customer service information containing the above target reply information and sends the customer service information to the user.

[0125] Figure 4 It shows a method for generating customer service information based on classification scenarios in an embodiment of the present application. This method mainly includes the following steps M1 to step M7.

[0126] Step M1: Obtain the current session data based on a preset memory window.

[0127] Specifically, the preset memory window is used to define the scope of the AI model's understanding of the session data. The system obtains the current session data of the customer service module based on the preset memory window size. For example, the system obtains the content of the previous three rounds of conversations in the current session of the customer service module.

[0128] Step M2: Perform a vectorization operation on the current session data to obtain the target corpus.

[0129] Specifically, the system uses the AI model to perform a vectorization operation on the current session data, that is, to convert the session data in each language in the customer service module into vector representations in the same vector space, so that the semantic expression of the current session data is not limited by the word order structure.

[0130] Step M3: Invoke the AI model to analyze the target corpus, and classify the current session data into preset scenarios according to the analysis results. The preset scenarios include the first scenario, the second scenario, and the third scenario. The first scenario generates customer service information based on preset configuration information, the second scenario generates customer service information based on a preset knowledge base, and the third scenario generates customer service information based on a preset information collection program.

[0131] Specifically, this application classifies user questions into the first scenario, the second scenario, and the third scenario. First, the first scenario does not involve complex information and only requires responding with fixed terms or querying status information and then responding. Therefore, this application uses a classification model to classify user questions in the first scenario and generates customer service information containing preset configuration information according to the classification results. Second, the second scenario involves complex information. This application uses a generative large model to match the user questions in the second scenario with a preset knowledge base, so as to generate corresponding customer service information based on the preset knowledge base; among them, the preset knowledge base uses historical session data as training corpus to generate corresponding target question statements and target reply information. Finally, the third scenario is an information collection and feedback scenario, which collects information provided by users based on a preset information collection program, including unboxing videos or user complaints, and then executes corresponding processing procedures (such as refund and return) or transfers to a human customer service.

[0132] Step M4: Generate customer service information corresponding to the current session data based on the preset scenario.

[0133] Specifically, the system generates customer service information containing target reply information using corresponding processing methods based on the preset scenario corresponding to the target corpus, and sends the customer service information to the user.

[0134] Furthermore, in one embodiment, the first scenario includes chat interaction, activity information, process consultation, and logistics inquiry.

[0135] Specifically, when a buyer sends a greeting message, information unrelated to the product or order, inquires about promotional information, order status, or logistics status in the customer service window, there is no need to train and call a specific generative large model to process it, as it does not involve complex information processing. Instead, you only need to call the classification model to reply to the user based on the pre-configured information.

[0136] Furthermore, in one embodiment, the second scenario includes product consultation.

[0137] Specifically, when a buyer asks questions about product-related content in the customer service window, including product recommendations, product parameters, product usage tutorials, etc., due to the complex information processing involved, the system calls the generative big model to identify the question statements contained in the current session data, and retrieves the target reply information corresponding to the current session data in the preset knowledge base, and replies to the user based on the target reply information.

[0138] Further, in one embodiment, the third scenario includes after-sales processing.

[0139] Specifically, when buyers raise after-sales demands such as returns, exchanges, refunds, or complaints at the customer service window, there is no need to train and call specific generative large models for processing, as it does not involve complex information processing and cannot meet user needs at the information communication level. On this basis, the system starts the preset information collection program, and asks users to "send the type of product problem", "send the order number", or "send the unboxing video", and executes the corresponding after-sales processing procedures based on user feedback.

[0140] Furthermore, in one embodiment, the method further includes the following steps M3.1 to M3.4.

[0141] Step M3.1: Cluster the question sentences in the historical conversation data to obtain corpus clusters;

[0142] Step M3.2: Determine the corpus cluster whose cluster size reaches a preset size threshold as the target cluster, and determine the target question sentence corresponding to the target cluster;

[0143] Step M3.3: Determine the target answer information corresponding to the target question sentence in the historical conversation data;

[0144] Step M3.4: Generate a preset knowledge base based on the target question sentence and the target answer information.

[0145] Specifically, in this embodiment, after clustering the question statements in the historical conversation data, the system selects the question statements with higher occurrence frequencies and representativeness as the target question statements, determines the target reply information corresponding to the target question statements in the historical conversation data, and then constructs a preset knowledge base based on the target question statements and the target reply information.

[0146] Further, in one embodiment, when the current conversation data corresponds to the third scenario, step M4 includes the following steps M4.1 and M4.2.

[0147] Step M4.1: Determine the preset prompt statement corresponding to the current conversation data and send the preset prompt statement to the user.

[0148] Step M4.2: In response to the target information sent by the user and corresponding to the preset prompt statement, execute the after-sales processing program associated with the preset prompt statement.

[0149] Specifically, when the system determines that the current conversation data corresponds to the third scenario, it further determines whether the current conversation data belongs to sub-scenarios such as "replace goods", "refund", "complaint", etc., and sends the preset prompt statement corresponding to the sub-scenario to the user. After the user provides feedback corresponding to the preset prompt statement, the system executes the after-sales processing program required by the user.

[0150] For example, when the user proposes "the goods are damaged and need to be returned", the system sends the preset prompt statement "Please provide an unpacking video", and initiates the return after-sales process after receiving the unpacking video provided by the user. Another example, when the user makes a complaint, the system sends the preset prompt statement "Describe the specific complaint type and provide complaint evidence", and initiates the complaint handling process after receiving the complaint information and complaint evidence provided by the user.

[0151] Further, on the basis of the above embodiments, after step M4.4, the method further includes the following step M4.3.

[0152] Step M4.3: If, in the preset number of dialogue turns, the target information corresponding to the preset prompt statement is not detected, or if it is detected that the current conversation indicates that the user has negative emotions, transfer to the artificial customer service.

[0153] Specifically, after the system sends the preset prompt statement to the user, if the user fails to provide the target information corresponding to the preset prompt statement in multiple rounds of conversation, it may be that the system misidentifies the user's intention, or the system cannot recognize the information currently provided by the user, and the artificial customer service needs to intervene to determine whether to initiate the corresponding after-sales processing program.

[0154] For example, if the system determines that the current session is a return scenario and sends the user a preset prompt statement "Please provide an unpacking video", and the user fails to provide an unpacking video within three rounds of conversation, it may be that the system's scenario determination is incorrect (i.e., the user does not intend to return the goods and therefore will not provide an unpacking video), or the user has provided other evidence that can corroborate the damage of the goods, but the system fails to recognize it. In this case, a human customer service needs to intervene.

[0155] In addition, when the system recognizes that the user has negative emotions, such as the user expressing anger, dissatisfaction or disappointment in the conversation, the system automatically transfers the call to a human customer service.

[0156] Further, in one embodiment, when the current session data corresponds to the second scenario, step M4 includes steps M4.4 to M4.7 as follows.

[0157] Step M4.4: Obtain the basic information, historical purchase behavior data, and browsing behavior data of the user account as basic data, and generate a user profile corresponding to the user account based on the basic data;

[0158] Step M4.5: Adjust the weights of the knowledge items included in the initial retrieval strategy based on the user profile to obtain a target retrieval strategy corresponding to the user account;

[0159] Step M4.6: Based on the target detection strategy, retrieve the target reply information corresponding to the current session data in the preset knowledge base, and generate customer service information based on the target reply information;

[0160] Step M4.7: Analyze the feedback behavior of the user account for the customer service information, and adjust the target retrieval strategy according to the analysis result. The feedback behavior includes the questioning behavior, purchase behavior, and customer service rating behavior for the product information included in the customer service information.

[0161] The applicant found that even though the customer service module of the system can accurately identify user information after receiving the information sent by the buyer user and generate corresponding customer service information based on the preset knowledge base, the customer service information cannot fully meet the user's information needs. The reason is that the preset knowledge base stores highly standardized target question statements and target reply information. Therefore, the customer service information generated by the system based on the preset knowledge base also tends to be in a standardized mode and fails to provide personalized customer service information in combination with the actual situation of the user, resulting in the user needing to ask multiple questions to meet the requirements, which seriously affects the efficiency of the customer service process.

[0162] The applicant proposed that for specific users (such as VIP users) or user questions involving complex information processing (such as asking for product recommendations), the system collects user data and builds user portraits, and then adjusts the search strategy of the corresponding preset knowledge base according to the user portrait, so that the final generated customer service information is more in line with the user's actual situation, thereby being able to meet user needs to a greater extent. The essential reason is that the customer service information originally generated is a collection of user questions and preset knowledge bases, but user questions in most cases fail to fully reflect the user's actual situation or user needs, so the system cannot provide personalized customer service information; when the system builds a user portrait, the system can combine the user needs / user characteristics contained in the user portrait that are not reflected in the user's questions into the search strategy (that is, the system takes into account the user's "unspoken needs"), so that the final generated customer service information has personalized characteristics.

[0163] The user data collection process includes:

[0164] Collect basic information about users, such as age, gender, and region. Basic information can help the system gain a preliminary understanding of the user's background and possible preferences. For example, users of different ages may have different needs and concerns about products. Young users may pay more attention to the fashion of products, while older users may pay more attention to the practicality of products.

[0165] Integrate the user's historical purchase behavior data, including the types of goods purchased, brands, price ranges, purchase frequency, etc. For example, a user who frequently purchases high-end electronic products will be marked in his profile as having high requirements for the quality and performance of electronic products and strong purchasing power.

[0166] Collect users' browsing behavior data, such as the product pages they browse, the duration of stay, search keywords, etc. If users frequently browse products in a certain category, it means that they may have a high interest in that category. For example, if users frequently browse sports equipment pages, the system can infer that they are interested in sports-related products.

[0167] The feature extraction and user profile generation process includes:

[0168] The system extracts features from the collected user data. For example, it classifies and codes the categories of goods purchased by users, and converts regional information into feature vectors related to culture and consumption habits; and then integrates these feature vectors to generate user portraits through machine learning algorithms, such as clustering algorithms and principal component analysis.

[0169] Generate target search strategies based on user portraits, including:

[0170] Adjust the weights of the knowledge items included in the initial detection strategy according to the key features in the user profile to generate a target retrieval strategy. For example, if the user profile shows that the user is a beginner photography enthusiast, the target retrieval strategy focuses on basic photography equipment, photography effect descriptions, and photography tutorials; if the user profile shows that the user is a senior photographer, the target retrieval strategy focuses on high-end photography equipment and the specific parameters of high-end photography equipment. Another example is that the target retrieval strategy assigns higher weights to knowledge related to the brands or product types that the user often purchases. When the user asks a question about skin care, the system preferentially retrieves knowledge related to the ingredients, usage methods, and suitable skin types of the skin care products of the brands that the user often purchases or related brands.

[0171] In addition, since there is a possibility of the AI module of the system to recognize the probability of the user's intention, especially for user questions involving complex information, the user further proposes to adjust the target retrieval strategy based on the feedback behavior of the user account for the customer service information, so that in the case of deviations in the previous customer service information, the subsequent generated customer service information can be adjusted in time to meet the user's needs.

[0172] Specifically, the user feedback behavior includes follow-up questions, purchase behavior, and customer service rating behavior. When the user asks similar questions about the same product included in the customer service information multiple times, or when the user does not generate a purchase behavior after multiple rounds of dialogue consultations about the product, or when the user's satisfaction rating for the customer service information is low, it indicates that the customer service information generated by the current target retrieval strategy fails to meet the user's needs well, and it is necessary to further adjust the target retrieval strategy or transfer to a human customer service.

[0173] In this way, based on the setting of the preset knowledge base, the accuracy of the customer service information is ensured through the preset knowledge base, and the probability of hallucinations generated by the AI model is reduced to a great extent; moreover, the system further combines the user profile to adjust the retrieval strategy of the corresponding preset knowledge base, so that the retrieval results obtained based on the target retrieval strategy can better fit the actual situation and user needs of the user, achieving the purpose of personalizing the customer service information; finally, the system also sets a trigger condition for continuously optimizing the target retrieval strategy, that is, the user's feedback behavior, so as to further improve the communication efficiency between the user and the customer service system.

[0174] Further, on the basis of the above embodiments, after step M4.4, the method further includes the following step M4.8.

[0175] Step M4.8: Determine the language style corresponding to the customer service information based on the user profile and the target corpus.

[0176] Specifically, the system analyzes the user's personality traits in real time based on the user profile and the target corpus, and then adjusts the interaction style according to the user's personality traits to achieve more empathetic and effective communication, thereby meeting the user's personalized interaction needs and providing emotional value to the user.

[0177] For example, when the user asks about the product specifications and uses technical language, the system provides detailed technical explanations, including data sheets and data comparison analyses; when the user communicates in an oral style, the system provides simple and direct explanations, sends customer service messages with a more life-like style, and even uses local slang.

[0178] The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent transformation made under the inventive concept of the present application by using the content of the specification and drawings of the present application, or any direct / indirect application in the relevant technical fields, is included in the patent protection scope of the present application.

Claims

1. A method for generating customer service information based on historical conversations, characterized in that: The AI ​​module applied to the system, the method is used to generate customer service information in a multilingual scenario, the method does not perform small language keyword annotation during AI configuration, and the method includes: Step S1: Acquire historical session data, where the historical session data represents the content of the session between the buyer user and the customer service account; Step S2: performing a vectorization operation on the historical conversation data based on the Embedding encoding model to obtain a first target corpus; Step S3: using a community discovery algorithm and according to vector similarity, clustering the question sentences contained in the first target corpus to obtain corpus clusters; Step S4: determining a corpus cluster whose cluster size reaches a preset size threshold as a target cluster, and determining a target question sentence corresponding to the target cluster; Step S5: determining the conversation context of the target question sentence in the first target corpus, generating target answer information corresponding to the target question sentence based on the conversation context, and generating a knowledge base based on the target question sentence and the target answer information; Step S6: obtaining current customer service conversation data, and performing a vectorization operation on the current customer service conversation data to obtain a second target corpus; Step S7: calling the AI ​​model to analyze the user intent of the second target corpus, and classifying the current customer service session data into preset scenarios according to the analysis results, wherein the preset scenarios include a first scenario, a second scenario, and a third scenario, wherein the first scenario includes chat interaction, activity information, process consultation, and logistics inquiry, the second scenario includes product consultation, and the third scenario includes after-sales processing; the first scenario generates customer service information based on preset configuration information, the second scenario generates customer service information based on the knowledge base, and the third scenario generates customer service information based on a preset information collection program; in the second scenario, determining the target question statement and target answer information corresponding to the user intent from the knowledge base, and the AI ​​model generates customer service information based on the target question statement and the target answer information corresponding to the user intent; in the first scenario, the preset configuration information is provided by the AI ​​model; When analyzing the user intent of the second target corpus, a word frequency analysis is performed on the second target corpus, and for functions or requirements that are mentioned multiple times by the buyer user, content related to the function or requirement is displayed more prominently when generating customer service information; When the current customer service session data corresponds to the second scenario, the method further includes: Step M4.4: Obtain basic information, historical purchase behavior data, and browsing behavior data of the user account as basic data, and generate a user profile corresponding to the user account based on the basic data; Step M4.5: adjusting the weights of the knowledge items included in the initial search strategy based on the user portrait to obtain a target search strategy corresponding to the user account; Step M4.6: based on the target retrieval strategy, searching the knowledge base for target reply information corresponding to the current customer service session data, and generating customer service information based on the target reply information; Step M4.7: Analyze the feedback behavior of the user account on the customer service information, and adjust the target search strategy according to the analysis result, wherein the feedback behavior includes questioning behavior, purchasing behavior, and customer service rating behavior for the product information included in the customer service information; Step M4.8: Based on the user portrait and the second target corpus, determine the language style corresponding to the customer service information.

2. The method according to claim 1, characterized in that Step S2 includes: Step S2.1: Based on the historical session data, obtain commodity session data related to commodity information; Step S2.2: performing a vectorization operation on the commodity conversation data to obtain a first target corpus.

3. The method according to claim 2, characterized in that Step S2.2 includes: Step S2.21: performing a filtering operation on the commodity session data based on the similarity between the session subject and the session data; Step S2.22: Perform a vectorization operation on the filtered commodity conversation data to obtain a first target corpus.

4. The method according to claim 1, characterized in that Step S5 includes: Step S5.1: In the first target corpus, determining the conversation data whose time interval with the target question sentence is within a preset time range as the conversation context of the target question sentence; Step S5.2: Generate target answer information corresponding to the target question sentence based on the conversation context.

5. The method according to claim 4, characterized in that Step S5.2 includes: Step S5.21: determining corresponding candidate answer information in a plurality of conversation scenarios in the first target corpus based on the conversation context; Step S5.22: If it is detected that there are multiple candidate answer information, the multiple candidate answer information is pushed to the user confirmation interface; Step S5.23: In response to a first confirmation instruction from the user confirmation interface, determine that candidate reply information corresponding to the first confirmation instruction is the target reply information.

6. The method according to claim 5, characterized in that After step S5.21, when it is detected that there are multiple candidate reply information, the method further includes: Step S5.24: Detecting the occurrence frequencies of the plurality of candidate answer information, and performing a sorting operation on the plurality of candidate answer information based on the occurrence frequencies; Step S5.25: Determine target reply information based on the sorting result.

7. The method according to claim 1, characterized in that After step S7, the method further comprises: Step S7.1: Push the customer service information to the user confirmation interface; Step S7.2: In response to a second confirmation instruction from the user confirmation interface, the customer service information is sent to the buyer user.

8. The method according to claim 1, characterized in that After step S6, the method further comprises: Step S7.3: when the target question sentence and target answer information corresponding to the second target corpus cannot be detected, generating a preset question sentence corresponding to the second target corpus; Step S7.4: marking the preset question statement; and, Step S7.5: Send a preset reply statement to the buyer user, or transfer the current customer conversation to a manual reply.

9. The method according to claim 1, characterized in that Step S7 includes: Step S7.6: Detecting the number of characters in the user's sentence in the second target corpus, and determining a memory window based on the number of characters, wherein the memory window includes multiple rounds of conversations of the second target corpus, and the number of characters in the user's sentence is negatively correlated with the size of the memory window; Step S7.7: analyzing the conversation data contained in the memory window based on a preset attenuation factor, wherein the preset attenuation factor represents a weight change rule corresponding to each round of conversation in the memory window; Step S7.8: Determine the target question sentence and target answer information corresponding to the memory window according to the analysis result, and generate customer service information based on the target question sentence and the target answer information.

10. A customer service information generation system based on historical conversations, characterized in that: The system includes an AI module, and the AI ​​module is used to execute the operation instructions included in the method for generating customer service information based on historical conversations as described in any one of claims 1-9.

Images