Session processing method, device and system, medium and product

By determining the semantic similarity and response mode of the session problem and the knowledge base, and combining the question-and-answer matching model, the response is automatically determined, the problem of a single response mode in intelligent customer service session processing is solved, the reply quality and efficiency are improved, and the manual quality inspection steps are reduced .

CN119048106APending Publication Date: 2024-11-29BEIJING JINGDONG YUANSHENG TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411179319.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing intelligent customer service session processing methods rely on a single fixed response mode, resulting in the need to improve the response efficiency and quality and require manual quality inspection.

Method used

By determining the semantic similarity of the knowledge base between the session problem and the preset knowledge base, selecting the target response mode, and using the question-answer matching model to obtain the response matching degree, automatically determine the reference session response, and replace the manual quality inspection step.

Benefits of technology

It improves the ability of conversation processing methods to cope with different descriptive text problems, improves the quality and efficiency of responses, and reduces manual intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048106A_ABST
    Figure CN119048106A_ABST
Patent Text Reader

Abstract

The invention discloses a session processing method, device and system, a medium and a product. The method comprises the steps of determining knowledge base semantic similarity according to a session question and a preset knowledge base in response to the received session question of a to-be-responded user sent by a session platform; determining a target response mode corresponding to the session question according to the knowledge base semantic similarity, and determining a reference session response corresponding to the session question by adopting the target response mode; inputting the session question and the reference session response into a pre-trained question and answer matching model to obtain an output response matching degree; and sending the reference session response and the response matching degree to the session platform, so that the session platform takes the reference session response as a target session response corresponding to the session question under the condition that the response matching degree meets a matching degree condition. And meanwhile, the reply efficiency and the reply quality are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a session processing method, device, system, medium and product. Background Art

[0002] Intelligent customer service is a technical means for industry applications developed based on large-scale knowledge processing. It is used to help improve the efficiency and quality of conversation responses, thereby reducing the manual maintenance costs of customer relationships.

[0003] In the process of implementing the present invention, it is found that there are at least the following technical problems in the prior art:

[0004] At present, the conversation processing method of intelligent customer service relies on a single fixed response mode, and manual customer service is still required to conduct quality inspection on the response content. Therefore, the response efficiency and quality need to be further improved. Summary of the invention

[0005] The embodiments of the present invention provide a conversation processing method, device, system, medium and product to solve the problem that the traditional conversation processing method relies on a single response mode, while improving the response efficiency and response quality.

[0006] According to an embodiment of the present invention, a session processing method is provided, which is applied to a session processing platform. The method includes:

[0007] In response to receiving a conversation question from a user to be answered sent by the conversation platform, determining a semantic similarity of the knowledge base according to the conversation question and a preset knowledge base;

[0008] Determining a target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and determining a reference conversation answer corresponding to the conversation question by adopting the target answer mode;

[0009] Inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree;

[0010] sending the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as the target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition;

[0011] Wherein, the target response mode is a knowledge base response mode or a general model response mode.

[0012] According to another embodiment of the present invention, a session processing method is provided, which is applied to a session platform. The method includes:

[0013] In response to detecting a conversation question input by a user to be answered, sending the conversation question to a conversation processing platform;

[0014] Receiving a reference session response and a response matching degree sent by the session processing platform;

[0015] In the case where the answer matching degree satisfies the matching degree condition, taking the reference conversation answer as the target conversation answer corresponding to the conversation question;

[0016] The reference conversation answer is determined by the conversation processing platform according to the conversation question using a target answer mode, the target answer mode is determined according to the semantic similarity between the conversation question and a knowledge base corresponding to a preset knowledge base, and the target answer mode is a knowledge base answer mode or a general model answer mode;

[0017] The answer matching degree is obtained by the conversation processing platform inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model.

[0018] According to another embodiment of the present invention, a session processing device is provided, which is applied to a session processing platform. The device includes:

[0019] A knowledge base semantic similarity determination module, configured to determine the knowledge base semantic similarity according to the conversation question and a preset knowledge base in response to receiving a conversation question of a user to be answered sent by the conversation platform;

[0020] A reference conversation answer determination module, used to determine a target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and determine a reference conversation answer corresponding to the conversation question using the target answer mode;

[0021] An answer matching degree determination module is used to input the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree;

[0022] an answer matching degree sending module, configured to send the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as a target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition;

[0023] Wherein, the target response mode is a knowledge base response mode or a general model response mode.

[0024] According to another embodiment of the present invention, a session processing device is provided, which is applied to a session platform, and the device includes:

[0025] A session question sending module, configured to send the session question to the session processing platform in response to detecting the session question input by the user to be answered;

[0026] A response matching degree receiving module, used for receiving the reference session response and the response matching degree sent by the session processing platform;

[0027] a target conversation answer determination module, configured to use the reference conversation answer as a target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition;

[0028] The reference conversation answer is determined by the conversation processing platform according to the conversation question using a target answer mode, the target answer mode is determined according to the semantic similarity between the conversation question and a knowledge base corresponding to a preset knowledge base, and the target answer mode is a knowledge base answer mode or a general model answer mode;

[0029] The answer matching degree is obtained by the conversation processing platform inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model.

[0030] According to another embodiment of the present invention, there is provided a session processing system, the session processing system comprising: a session processing platform and a session platform in communication connection;

[0031] The session processing platform is used to execute the session processing method applied to the session processing platform described in any embodiment of the present invention;

[0032] The session platform is used to execute the session processing method applied to the session platform described in any embodiment of the present invention.

[0033] According to another embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the session processing method described in any embodiment of the present invention when executed.

[0034] According to another embodiment of the present invention, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the session processing method described in any embodiment of the present invention is implemented.

[0035] The technical solution of the embodiment of the present invention solves the problem that the traditional conversation processing method relies on a single response mode by determining the knowledge base semantic similarity between the conversation question and the preset knowledge base, and determines the target response mode corresponding to the conversation question based on the knowledge base semantic similarity, and adopts the target response mode to determine the reference conversation response corresponding to the conversation question, thereby improving the ability of the conversation processing method to deal with conversation questions with different description texts, and thus improving the quality of the conversation response. In addition, the embodiment of the present invention obtains the answer matching degree between the reference conversation response and the conversation question through the question-answer matching model, so that the conversation platform uses the reference conversation response as the target conversation response corresponding to the conversation question when the answer matching degree meets the matching degree condition, replacing the step of manual customer service to perform quality inspection on the answer content, while ensuring the answer quality, it also improves the efficiency of the conversation response.

[0036] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0038] Figure 1 A flow chart of a session processing method provided by an embodiment of the present invention;

[0039] Figure 2 A flowchart of another session processing method provided by an embodiment of the present invention;

[0040] Figure 3 A flowchart of a specific example of a method for determining a reference session response provided by an embodiment of the present invention;

[0041] Figure 4 A flow chart of a session processing method provided by an embodiment of the present invention;

[0042] Figure 5 A schematic diagram of a conversation group provided by an embodiment of the present invention;

[0043] Figure 6 A schematic diagram of the structure of a session processing device provided by an embodiment of the present invention;

[0044] Figure 7A schematic diagram of the structure of a session processing device provided by an embodiment of the present invention;

[0045] Figure 8 A structural diagram of a session processing system provided by an embodiment of the present invention;

[0046] Fig. 9 A functional architecture diagram of a session processing system provided by an embodiment of the present invention;

[0047] Fig.10 An interactive flow chart of a session processing system provided by an embodiment of the present invention;

[0048] Fig.11 A functional schematic diagram of active reach of a conversation platform provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0050] It should be noted that the terms "first", "second", "original", "target", "reference", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0051] Figure 1 This is a flow chart of a conversation processing method provided by an embodiment of the present invention. This embodiment is applicable to the case of answering conversation questions raised by users in a conversation platform. The method can be executed by a conversation processing device. The conversation processing device can be implemented in the form of hardware and / or software. The conversation processing device can be configured in a conversation processing platform. Figure 1 As shown, the method includes:

[0052] S110 : In response to receiving a conversation question of a user to be answered sent by the conversation platform, determine the semantic similarity of the knowledge base according to the conversation question and a preset knowledge base.

[0053] Specifically, the conversation platform is an application platform with conversation functions, which is used to receive conversation questions sent by the client of the user to be answered, and send the conversation questions to the conversation processing platform. Exemplarily, the platform type of the conversation platform can be a social platform, a consumer platform, or a game platform, etc. The platform type of the conversation platform is not limited here, and can be customized according to actual needs.

[0054] For example, the problem scenario of a conversation question can be a business problem for a specific field, such as "How long is the shelf life of a product?" The problem scenario of a conversation question can also be the content of a chat conversation, such as "The weather is nice today." The problem scenario of a conversation question is not limited here, and can be customized according to actual needs.

[0055] Specifically, knowledge base semantic similarity can be used to characterize the possibility of a conversation question being retrieved from a preset knowledge base.

[0056] In a specific embodiment, determining the knowledge base semantic similarity based on the conversation question and the preset knowledge base includes: inputting the conversation question into a semantic similarity model to obtain the output knowledge base semantic similarity. In this embodiment, the semantic similarity model is constructed based on the preset knowledge base.

[0057] In another specific embodiment, determining the semantic similarity of the knowledge base according to the conversation question and the preset knowledge base includes: inputting the conversation question and the preset knowledge base into a semantic similarity model to obtain the output semantic similarity of the knowledge base.

[0058] In another specific embodiment, the semantic similarity of the knowledge base is determined based on the conversation question and the preset knowledge base, including: obtaining text features corresponding to the conversation question, obtaining a set of candidate questions from the preset knowledge base based on the text features using an inverted index method, and determining the semantic similarity of the knowledge base based on the number of candidate questions in the candidate question set.

[0059] Exemplarily, feature extraction operations such as word segmentation, part-of-speech tagging, keyword extraction, and keyword position tagging are performed on the conversation question to obtain text features of the conversation question.

[0060] Specifically, the number of questions is positively correlated with the semantic similarity of the knowledge base. The more questions there are, the higher the semantic similarity of the knowledge base is. Conversely, the fewer questions there are, the lower the semantic similarity of the knowledge base is. The specific mapping relationship between the number of questions and the semantic similarity of the knowledge base can be determined by a mapping list or a mapping function, which is not limited here.

[0061] On the basis of the above embodiments, specifically, the method also includes: inputting the conversation question and the historical conversation data corresponding to the conversation question into a pre-trained question intention prediction model to obtain the output target question intention; determining the business attributes corresponding to the user to be answered based on the user identifier corresponding to the user to be answered; and obtaining a preset knowledge base from the original knowledge base based on the business attributes and the target question intention.

[0062] Among them, the intent is objectively descriptive and unique, and the annotation standard is easy to understand. Different intents in the same target field should not have overlapping, inclusion, and other relationships, and should be independent. For example, the target problem intent includes but is not limited to claims, performance, and billing, etc. The division method of the target problem intent is not limited here, and can be customized according to actual needs.

[0063] Exemplarily, the model architecture of the question intent prediction model may be an ALBERT model, and the model architecture of the question intent prediction model is not limited herein.

[0064] The user identification may be sent by the conversation platform to the conversation processing platform when sending the conversation question. For example, the user identification may be composed of words, numbers, letters, and special characters, and the identification form of the user identification is not limited here.

[0065] The attribute mapping list records the service attributes corresponding to at least one user identification. By querying the attribute mapping list according to the user identification, the service attributes corresponding to the user to be answered are obtained.

[0066] For example, based on the customer source, the business attributes may be supply chain, international business, and domestic business, etc. Based on the customer service, the business attributes may be pre-sales, after-sales, and technical, etc. Based on the product business, the business attributes may be home appliances, communications, and clothing, etc. The classification method corresponding to the business attributes is not limited here, and the business attributes include attribute values ​​corresponding to at least one classification dimension.

[0067] In this embodiment, the original knowledge base is divided according to two dimensions, namely, business attributes and target question intent, to obtain at least two preset knowledge bases, and a preset knowledge base corresponding to the business attributes and target question intent is selected from the original knowledge base.

[0068] The advantage of this setting is that it solves the problem of too much data in the preset knowledge base, thereby further improving the efficiency of conversation responses.

[0069] S120. Determine a target response pattern corresponding to the conversation question based on the semantic similarity of the knowledge base, and use the target response pattern to determine a reference conversation response corresponding to the conversation question.

[0070] In this embodiment, the target response mode is a knowledge base response mode or a general model response mode. The knowledge base response mode is used to represent that the reference conversation response is extracted from a preset knowledge base, and the general model response mode is used to represent that the reference conversation response is generated by a general domain language model.

[0071] In a specific embodiment, the target response mode corresponding to the conversation question is determined based on the semantic similarity of the knowledge base, including: when the semantic similarity of the knowledge base is greater than a first semantic similarity threshold, the target response mode corresponding to the conversation question is set to the knowledge base response mode; when the semantic similarity of the knowledge base is less than or equal to the first semantic similarity threshold, the target response mode corresponding to the conversation question is set to the general model response mode.

[0072] Exemplarily, the first semantic similarity threshold may be 0.6. The specific parameter value of the first semantic similarity threshold is not limited here and can be customized according to actual needs.

[0073] In a specific embodiment, a target response mode is adopted to determine a reference conversation response corresponding to a conversation question, including: when the target response mode is a knowledge base response mode, obtaining the semantic similarity corresponding to the conversation question and at least two knowledge base questions in a preset knowledge base, and taking the knowledge base response corresponding to the knowledge base question with the highest semantic similarity as the reference conversation response corresponding to the conversation question.

[0074] In another specific embodiment, a target answering mode is adopted to determine a reference conversation answer corresponding to a conversation question, including: when the target answering mode is a general model answering mode, obtaining a user portrait corresponding to the user to be answered and historical conversation data corresponding to the conversation question; inputting the user portrait, conversation question and historical conversation data into a general domain language model to obtain an output reference conversation answer corresponding to the conversation question.

[0075] Exemplarily, the user portrait includes but is not limited to the user's age, occupation, product category, order service, etc. The dimensions of the user portrait are not limited here and can be customized according to actual needs.

[0076] Specifically, the historical conversation data is used to characterize the conversation content of a historical time period or a preset number of historical conversation contents. For example, the historical time period may be one hour, the preset number may be 10, and the historical conversation content is the conversation content before the current moment. The content form of the conversation content in the historical conversation data includes but is not limited to text, video, audio, and pictures.

[0077] In this embodiment, the general domain language model is used to generate reference conversation responses corresponding to conversation questions based on user portraits, conversation questions, and historical conversation data. Among them, the general domain language model represents a machine learning model with a large parameter scale and complexity. The general domain language model requires a large amount of computing resources and storage space, often using distributed computing and special hardware technology, and has strong generalization and expression capabilities. The model architecture of the general domain language model used here is not limited, and can be customized according to actual needs.

[0078] S130: Input the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree.

[0079] Specifically, the answer matching degree is used to characterize the matching degree between the conversation question and the reference conversation answer.

[0080] In a specific embodiment, the method further includes: a training method for the question-answer matching model. Specifically, the training question-answer pair is input into the untrained question-answer matching model to obtain the output predicted matching degree; the loss function is determined according to the predicted matching degree and the standard matching degree corresponding to the training question-answer pair; the model parameters of the question-answer matching model are adjusted according to the loss function until the loss function converges, and the question-answer matching model in the current iteration process is used as the trained question-answer matching model.

[0081] The training problem pair includes a training problem and a training answer. The function type of the loss function includes but is not limited to the square loss function, logarithmic loss function, exponential loss function, mean square error loss function, logistic regression loss function, Huber loss function, cross entropy loss function, and Kullback-Leibler divergence loss function, etc. There is no limitation on the function type of the loss function here, and it can be customized according to actual needs.

[0082] S140: Send the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as the target conversation answer corresponding to the conversation question when the answer matching degree satisfies the matching degree condition.

[0083] Specifically, the expression of the answer matching degree is a numerical value or a yes or no. In a specific embodiment, when the expression of the answer matching degree is a numerical value, the matching degree condition is that the answer matching degree is greater than the matching degree threshold; when the expression of the answer matching degree is a yes or no, the matching degree condition is that the answer matching degree is a successful match. Among them, by way of example, the matching degree threshold can be 0.9, and the matching degree threshold is not limited here, and can be customized according to actual needs.

[0084] The technical solution of this embodiment solves the problem that the traditional conversation processing method relies on a single response mode by determining the knowledge base semantic similarity between the conversation question and the preset knowledge base, and determines the target response mode corresponding to the conversation question based on the knowledge base semantic similarity, and adopts the target response mode to determine the reference conversation response corresponding to the conversation question, thereby improving the ability of the conversation processing method to deal with conversation questions with different description texts, and thus improving the quality of the conversation response. In addition, the answer matching degree between the reference conversation response and the conversation question is obtained through the question-answer matching model, so that the conversation platform uses the reference conversation response as the target conversation response corresponding to the conversation question when the answer matching degree meets the matching degree condition, replacing the step of manual customer service to perform quality inspection on the answer content, while ensuring the answer quality, it also improves the efficiency of the conversation response.

[0085] Figure 2 This is a flowchart of another session processing method provided by an embodiment of the present invention. This embodiment further refines the "knowledge base response mode" in the above embodiment. In this embodiment, the knowledge base response mode is a question-and-answer knowledge base response mode or a document knowledge base response mode. Figure 2 As shown, the method includes:

[0086] S210: In response to receiving a conversation question of a user to be answered sent by the conversation platform, determine the semantic similarity of the knowledge base according to the conversation question and a preset knowledge base.

[0087] In this embodiment, the preset knowledge base includes a question and answer knowledge base and a document knowledge base.

[0088] The question-and-answer knowledge base is composed of multiple question-and-answer pairs, and the question-and-answer pair includes a preset knowledge base question and a knowledge base answer corresponding to the preset knowledge base question. For example, the preset knowledge base question in the question-and-answer pair may be "Business is not doing well recently", and the knowledge base answer corresponding to the preset knowledge base question may be "Our staff can do some publicity for you, which may help you increase your revenue".

[0089] The document knowledge base is composed of multiple documents, and the types corresponding to different documents can be the same or different. For example, according to the form of the document, the document type can be a report document, such as a market research report or a business progress report, or a paper document, such as an academic paper, a graduation thesis or a scientific paper, or a manual document, such as an operation manual, a user manual or a product manual. The document type is not limited here, and can be collected and sorted according to actual needs.

[0090] Based on the above embodiment, the method further includes: extracting question-answer pairs from historical conversation records to obtain at least one pair of question-answer pairs, and constructing a question-answer knowledge base based on the at least one pair of question-answer pairs. The historical conversation records are real conversation contents between the user and the manual customer service.

[0091] In a specific embodiment, question-answer pairs are extracted from historical conversation records to obtain at least one pair of question-answer pairs, including: inputting historical conversation records and prompts into a general domain language model to obtain at least one pair of question-answer pairs as output.

[0092] Specifically, the prompt is used to prompt the general domain language model to extract one or more question-answer pairs from the historical conversation records. Exemplarily, the prompt may be "Extract question-answer pairs from the following conversation content. Question-answer pairs include questions asked by users and corresponding answers from customer service personnel. There may be no question-answer pairs, or there may be multiple question-answer pairs. **Questions asked by users** refer to questions that users ask customer service personnel for help when they encounter unknown problems. They are not ordinary conversations, nor are they questions asked by customer service personnel. #Question-answer pair output format: [\{"question":"[the question you extracted]","answer":"[the answer you extracted]"\}]. The content of the conversation is:". The prompt is not limited here and can be customized according to actual needs.

[0093] The general domain language model in this embodiment and the general domain language model used in the general model response mode in the above embodiments may be the same general domain language model or different general domain language models.

[0094] On the basis of the above embodiments, specifically, before constructing a question-answer knowledge base based on at least one question-answer pair, the method also includes: for each question-answer pair, inputting the question-answer pair into a question-answer matching model to obtain an output answer matching degree; and screening at least one question-answer pair based on at least one answer matching degree.

[0095] The question-answer matching model is used to evaluate the matching degree between the question and the answer in the question-answer pair. Exemplarily, the answer matching degree is expressed in the form of a numerical value or yes or no.

[0096] In a specific embodiment, at least one question-answer pair is screened according to at least one answer matching degree, including: when the answer matching degree is expressed in the form of "yes or no", the question-answer pair whose answer matching degree is a successful match is used as the screened question-answer pair.

[0097] In another specific embodiment, at least one question-answer pair is screened according to at least one answer matching degree, including: when the answer matching degree is expressed in a numerical form, the question-answer pairs greater than the matching degree threshold are used as screened question-answer pairs; or, at least one question-answer pair is sorted according to at least one answer matching degree, and the screened question-answer pairs are obtained from the sorted results according to a preset selection number.

[0098] Exemplarily, the matching degree threshold may be 0.8, and the preset number of selections may be 100. There is no limitation on the matching degree threshold and the preset number of selections, and they may be customized according to actual needs.

[0099] The advantage of this setting is that it can improve the matching quality of question-answer pairs in the question-answer knowledge base, which in turn helps to improve the response quality of the conversation processing method.

[0100] In this embodiment, the answer matching model used for quality inspection of the extracted question and answer pairs and the answer matching model used in S130 in the above embodiment may be the same answer matching model or different answer matching models.

[0101] S220 , determining whether the knowledge base semantic similarity is greater than a first semantic similarity threshold, if yes, executing S240 , if no, executing S230 .

[0102] S230, setting the target response mode corresponding to the conversation question to the general model response mode, and executing S270.

[0103] S240 , determining whether the knowledge base semantic similarity is greater than a second semantic similarity threshold, if yes, executing S250 , if no, executing S260 .

[0104] In this embodiment, the second semantic similarity threshold is greater than the first semantic similarity threshold. Exemplarily, the first semantic similarity threshold is 0.6, and the second semantic similarity threshold is 0.9. The specific parameter values ​​of the first semantic similarity threshold and the second semantic similarity threshold are not limited here, and can be customized according to actual needs.

[0105] S250, setting the target answering mode corresponding to the conversation question to the question-answering knowledge base answering mode, and executing S270.

[0106] Specifically, the question-answering knowledge base response pattern is used to characterize the reference conversation response retrieved from the question-answering knowledge base.

[0107] S260. Set the target answer mode corresponding to the conversation question to the document knowledge base answer mode, and execute S270.

[0108] Specifically, the document knowledge base answer model is used to characterize the reference conversation question and answer extracted from the document knowledge base.

[0109] S270: Use the target response mode to determine a reference conversation response corresponding to the conversation question.

[0110] In a specific embodiment, a target answer mode is adopted to determine a reference conversation answer corresponding to a conversation question, including: when the target answer mode is a question-and-answer knowledge base answer mode, obtaining text similarities corresponding to the conversation question and at least two preset knowledge base questions in a question-and-answer knowledge base; filtering a target knowledge base question from at least two preset knowledge base questions based on a text filtering condition and at least two text similarities; and using the knowledge base answer corresponding to the target knowledge base question in the question-and-answer knowledge base as a reference conversation answer corresponding to the conversation question.

[0111] Specifically, text similarity is used to characterize the similarity between the text content of the conversation question and the preset knowledge base question.

[0112] In a specific embodiment, obtaining text similarities between a conversation question and at least two preset knowledge base questions in a question-and-answer knowledge base includes: obtaining a conversation question vector corresponding to the conversation question, obtaining, for each preset knowledge base question, a preset knowledge base question vector corresponding to the preset knowledge base question, and determining, based on a vector distance between the conversation question vector and the preset knowledge base question vector, the text similarities corresponding to the conversation question and the preset knowledge base question.

[0113] Exemplarily, the method for calculating the vector distance includes, but is not limited to, a cosine distance algorithm, an Euclidean distance algorithm, or a Manhattan distance algorithm, etc. The method for calculating the vector distance is not limited here, and can be customized according to actual needs.

[0114] In a specific embodiment, based on the text screening condition and at least two text similarities, the target knowledge base question is screened from at least two preset knowledge base questions, including: taking the preset knowledge base question whose text similarity is greater than the text similarity threshold as the target knowledge base question.

[0115] In another specific embodiment, a target knowledge base question is selected from at least two preset knowledge base questions based on a text screening condition and at least two text similarities, including: taking at least one preset knowledge base question whose text similarity is greater than a text similarity threshold as a reference knowledge base question, and taking the reference knowledge base question corresponding to the maximum text similarity as the target knowledge base question, or randomly taking a reference knowledge base question as the target knowledge base question.

[0116] Exemplarily, the text similarity threshold is 0.9. The text similarity threshold is not limited here and can be customized according to actual needs.

[0117] In a specific embodiment, the target response mode is adopted to determine the reference conversation response corresponding to the conversation question, and further includes: when there is no target knowledge base question that meets the text screening condition, the fallback response is used as the reference conversation response corresponding to the conversation question. Exemplarily, the fallback response can be "unable to retrieve suitable response content from the knowledge base".

[0118] In another specific embodiment, a target answer mode is adopted to determine a reference conversation answer corresponding to a conversation question, and further includes: when there is no target knowledge base question that satisfies a text screening condition, obtaining semantic similarities corresponding to the conversation question and at least two preset knowledge base questions respectively; screening a target knowledge base question from at least two preset knowledge base questions according to the semantic screening condition and at least two semantic similarities; and using a knowledge base answer corresponding to the target knowledge base question in a question-and-answer knowledge base as a reference conversation answer corresponding to the conversation question.

[0119] Specifically, there is no target knowledge base question that meets the text screening condition, indicating that there is no preset knowledge base question in the question-and-answer knowledge base whose text similarity is greater than the text similarity threshold.

[0120] Specifically, semantic similarity is used to characterize the similarity between the meanings of conversation questions and questions in the preset knowledge base.

[0121] In a specific embodiment, obtaining the semantic similarities corresponding to the conversation question and at least two preset knowledge base questions respectively includes: for each preset knowledge base question, inputting the conversation question and the preset knowledge base question into a semantic similarity model to obtain output semantic similarities.

[0122] The semantic similarity model in this embodiment and the semantic similarity model used to obtain the semantic similarity of the knowledge base can be the same semantic similarity model or different semantic similarity models.

[0123] In a specific embodiment, based on the semantic screening condition and at least two semantic similarities, the target knowledge base question is screened from at least two preset knowledge base questions, including: taking the preset knowledge base question whose semantic similarity is greater than the semantic similarity threshold as the target knowledge base question.

[0124] In another specific embodiment, a target knowledge base question is selected from at least two preset knowledge base questions based on semantic screening conditions and at least two semantic similarities, including: taking at least one preset knowledge base question whose semantic similarity is greater than a semantic similarity threshold as a reference knowledge base question, taking the reference knowledge base question corresponding to the maximum semantic similarity as the target knowledge base question, or randomly taking a reference knowledge base question as the target knowledge base question.

[0125] Exemplarily, the semantic similarity threshold is 0.9. The semantic similarity threshold is not limited here and can be customized according to actual needs.

[0126] On the basis of the above embodiment, specifically, a target response mode is adopted to determine a reference conversation response corresponding to the conversation question, and also includes: when there is no target knowledge base question that meets the semantic screening conditions, a fallback response is used as a reference conversation response corresponding to the conversation question.

[0127] In another specific embodiment, a target answering mode is adopted to determine a reference conversation answer corresponding to the conversation question, including: when the target answering mode is a document knowledge base answering mode, obtaining a user portrait corresponding to the user to be answered and historical conversation data corresponding to the conversation question; inputting the user portrait, conversation question, historical conversation data and document knowledge base into a pre-trained document answering model to obtain an output reference conversation answer corresponding to the conversation question.

[0128] Specifically, the document answer model is used to extract reference conversation answers corresponding to conversation questions from the document knowledge base based on user portraits, conversation questions and historical conversation data.

[0129] Figure 3 A flowchart of a specific example of a method for determining a reference conversation response provided by an embodiment of the present invention. Specifically, in response to detecting a conversation question, it is determined whether the knowledge base semantic similarity between the conversation question and the preset knowledge base is less than or equal to a first semantic similarity threshold (P1). If yes, a general model response mode is adopted to generate a reference conversation response corresponding to the conversation question. If not, it is further determined whether the knowledge base semantic similarity is greater than a second semantic similarity threshold (P2). If yes, the target response mode corresponding to the conversation question is set to the question-and-answer knowledge base response mode. Specifically, it is determined whether there is a preset knowledge base question in the question-and-answer knowledge base whose text similarity with the conversation question is greater than the text similarity threshold (T1). If yes, the knowledge base response corresponding to the retrieved preset knowledge base question is used as a reference conversation response. If no, it is further determined whether there is a preset knowledge base question in the question-and-answer knowledge base whose semantic similarity with the conversation question is greater than the semantic similarity threshold (T2). If yes, the knowledge base response corresponding to the retrieved preset knowledge base question is used as a reference conversation response. If no, the fallback response is used as a reference conversation response.

[0130] If the knowledge base semantic similarity is greater than the first semantic similarity threshold and less than or equal to the second semantic similarity threshold, the target response mode corresponding to the conversation question is set to the document knowledge base response mode, and the document response model is used to extract the reference conversation response from the document knowledge base.

[0131] S280: Input the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree.

[0132] S290: Send the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as the target conversation answer corresponding to the conversation question when the answer matching degree satisfies the matching degree condition.

[0133] S280-S290 in this embodiment are similar to those in the above embodiment. Figure 1 The S130 - S140 shown correspond to the same or similar ones and are not described in detail in this embodiment.

[0134] The technical solution of this embodiment, by setting the knowledge base response mode to the question-and-answer knowledge base response mode or the document knowledge base response mode, sets the target response mode corresponding to the conversation question to the question-and-answer knowledge base response mode when the knowledge base semantic similarity is greater than a second semantic similarity threshold, and sets the target response mode corresponding to the conversation question to the document knowledge base response mode when the knowledge base semantic similarity is less than or equal to the second semantic similarity threshold. This solves the problem of poor response quality caused by a single knowledge base response mode, improves the granularity of the knowledge base response mode, and further improves the quality of conversation responses.

[0135] Figure 4 This is a flow chart of a conversation processing method provided by an embodiment of the present invention. This embodiment is applicable to the case of answering conversation questions raised by users in a conversation platform. The method can be executed by a conversation processing device. The conversation processing device can be implemented in the form of hardware and / or software. The conversation processing device can be configured in the conversation platform. Figure 4 As shown, the method includes:

[0136] S310: In response to detecting a conversation question input by a waiting user, sending the conversation question to a conversation processing platform.

[0137] Specifically, the conversation platform is an application platform with conversation functions, which is used to receive conversation questions sent by the client of the user to be answered, and send the conversation questions to the conversation processing platform. Exemplarily, the platform type of the conversation platform can be a social platform, a consumer platform, or a game platform, etc. The platform type of the conversation platform is not limited here, and can be customized according to actual needs.

[0138] For example, the problem scenario of a conversation question can be a business problem for a specific field, such as "How long is the shelf life of a product?" The problem scenario of a conversation question can also be the content of a chat conversation, such as "The weather is nice today." The problem scenario of a conversation question is not limited here, and can be customized according to actual needs.

[0139] S320: Receive a reference session response and a response matching degree sent by the session processing platform.

[0140] In this embodiment, the reference conversation response is determined by the conversation processing platform according to the conversation question using a target response mode. The target response mode is determined based on the semantic similarity between the conversation question and the knowledge base corresponding to the preset knowledge base. The target response mode is a knowledge base response mode or a general model response mode.

[0141] Specifically, knowledge base semantic similarity can be used to characterize the possibility of a conversation question being retrieved from a preset knowledge base.

[0142] In this embodiment, the answer matching degree is obtained by the conversation processing platform inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model.

[0143] The method of obtaining the reference session response and the response matching degree in S320 in this embodiment is similar to that in the above embodiment. Figure 1 S110-S130 of the session processing platform shown in the figure are the same or similar, or the same as those in the above embodiment. Figure 2 S210 - S280 of the session processing platform shown are identical or similar to each other, and are not described in detail in this embodiment.

[0144] S330: When the answer matching degree satisfies the matching degree condition, the reference conversation answer is used as the target conversation answer corresponding to the conversation question.

[0145] Specifically, the expression of the answer matching degree is a numerical value or a yes or no. In a specific embodiment, when the expression of the answer matching degree is a numerical value, the matching degree condition is that the answer matching degree is greater than the matching degree threshold; when the expression of the answer matching degree is a yes or no, the matching degree condition is that the answer matching degree is a successful match. Among them, by way of example, the matching degree threshold can be 0.9, and the matching degree threshold is not limited here, and can be customized according to actual needs.

[0146] Based on the above embodiment, specifically, the method further includes: when the answer matching degree does not meet the matching condition, using the manual conversation answer input by the manual customer service as the target conversation answer corresponding to the conversation question.

[0147] The technical solution of this embodiment, by receiving the reference conversation answer and the answer matching degree sent by the conversation processing platform, wherein the reference conversation answer is determined by the conversation processing platform using the target answer mode according to the conversation question, solves the problem that the traditional conversation processing method relies on a single answer mode, improves the ability of the conversation processing method to deal with conversation questions with different description texts, and thus improves the quality of the conversation answer. And when the answer matching degree meets the matching degree condition, the conversation platform uses the reference conversation answer as the target conversation answer corresponding to the conversation question, replacing the step of manual customer service to perform quality inspection on the answer content, while ensuring the answer quality, it also improves the efficiency of the conversation answer.

[0148] On the basis of the above embodiments, specifically, the method further includes: in response to detecting the information to be pushed sent by the conversation processing platform, obtaining the information source corresponding to the information to be pushed; when the information source is a promotion source, performing an approval operation on the information to be pushed, and if the approval result obtained is approval passed, outputting the information to be pushed to the conversation window corresponding to the user to be responded to; when the information source is a non-promotion source, outputting the information to be pushed to the conversation window corresponding to the user to be responded to.

[0149] Specifically, the information source can be used to characterize the platform attributes of the third-party push platform corresponding to the information to be pushed. For example, the information to be pushed from a promotional source can be activity information, discount information, rights and interests information, or promotional material information, etc., and the information to be pushed from a non-promotional source can be bill notification information, settlement notification information, or renewal notification information, etc.

[0150] Exemplarily, the information content of the information to be pushed includes but is not limited to text, pictures, videos or audio, etc., and the information content of the information to be pushed is not limited here.

[0151] In a specific embodiment, when the conversation window corresponding to the user to be answered belongs to the conversation group window, the information to be pushed also includes a user field, and accordingly, the information to be pushed is output to the conversation window corresponding to the user to be answered, including: when the user field is empty, the information to be pushed is output to the conversation window corresponding to the user to be answered; when the user field is not empty, prompt information is generated according to the user field, and the prompt information and the information to be pushed are output to the conversation window corresponding to the user to be answered. Exemplarily, the prompt information can be "@User A".

[0152] The advantage of this setting is that it provides a processing mechanism for actively triggering sessions for the session processing method, broadens the scenarios of the session processing method, and screens the push information by information source, thereby ensuring the security of the session platform.

[0153] On the basis of the above embodiments, specifically, the method also includes: in response to detecting a group creation instruction, obtaining at least one target user that has a binding relationship with the customer service robot from at least one preset user corresponding to the group creation instruction; determining the conversation group information corresponding to the group creation instruction according to the user identifier corresponding to the at least one target user, and establishing a conversation group according to the conversation group information; wherein the conversation group includes users to be responded to, and the conversation group information includes at least one of a group name, a group announcement, a group note information and a group message.

[0154] Specifically, the group creation instruction may be generated by the conversation platform in response to detecting a group creation function triggered by a client. The at least one preset user corresponding to the group creation instruction includes the user of the client that triggers the group creation function and other users who have a friend relationship with the user. The target user is used to represent the preset user in the conversation group who has a binding relationship with the customer service robot.

[0155] In a specific embodiment, the user group mapping list records at least one set of user identification sets and the session group information corresponding to each set of user identification sets. The user identification set includes at least one user identification. The user group mapping list is queried according to the user identification corresponding to at least one target user to obtain the session group information corresponding to the group creation instruction.

[0156] Figure 5 A schematic diagram of a conversation group provided by an embodiment of the present invention. Specifically, Figure 5 The "xx Company Communication Group" in the upper left corner indicates the group name. Figure 5 The "xxxx" on the right side indicates a group announcement. Figure 5 The middle "@Everyone, this group is a communication group for xx company. Regarding service quality issues, welcome to communicate and feedback" indicates a group message. Group notes are in Figure 5 Not shown in FIG.

[0157] The advantage of such a setting is that the session processing method is not only applicable to 1V1 session scenarios, but also to 1V-multiple group session scenarios, thereby helping to improve the efficiency of establishing session groups.

[0158] It should be noted that in the technical solution of the present invention, the collection, use, storage, sharing and transfer of user personal information involved are in compliance with the provisions of relevant laws and regulations, and it is necessary to inform the user and obtain the user's consent or authorization. When applicable, the user's personal information is de-identified and / or anonymized and / or encrypted.

[0159] The following is an embodiment of a session processing device provided in an embodiment of the present invention. The device and the session processing method of the above-mentioned embodiment belong to the same inventive concept. For details not described in detail in the embodiment of the session processing device, reference can be made to the contents about the session processing method in the above-mentioned embodiment.

[0160] Figure 6 The structure diagram of a session processing device provided by an embodiment of the present invention is applied to a session processing platform. Figure 6 As shown, the device includes: a knowledge base semantic similarity determination module 410, a reference conversation response determination module 420, a response matching degree determination module 430 and a response matching degree sending module 440.

[0161] The knowledge base semantic similarity determination module 410 is used to determine the knowledge base semantic similarity according to the conversation question and the preset knowledge base in response to receiving the conversation question of the user to be answered sent by the conversation platform;

[0162] A reference conversation answer determination module 420 is used to determine a target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and to determine a reference conversation answer corresponding to the conversation question using the target answer mode;

[0163] The answer matching degree determination module 430 is used to input the conversation question and the reference conversation answer into the pre-trained question-answer matching model to obtain the output answer matching degree;

[0164] The answer matching sending module 440 is used to send the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as the target conversation answer corresponding to the conversation question when the answer matching degree meets the matching degree condition;

[0165] Among them, the target response mode is the knowledge base response mode or the general model response mode.

[0166] The technical solution of this embodiment solves the problem that the traditional conversation processing method relies on a single response mode by determining the knowledge base semantic similarity between the conversation question and the preset knowledge base, and determines the target response mode corresponding to the conversation question based on the knowledge base semantic similarity, and adopts the target response mode to determine the reference conversation response corresponding to the conversation question, thereby improving the ability of the conversation processing method to deal with conversation questions with different description texts, and thus improving the quality of the conversation response. In addition, the answer matching degree between the reference conversation response and the conversation question is obtained through the question-answer matching model, so that the conversation platform uses the reference conversation response as the target conversation response corresponding to the conversation question when the answer matching degree meets the matching degree condition, replacing the step of manual customer service to perform quality inspection on the answer content, while ensuring the answer quality, it also improves the efficiency of the conversation response.

[0167] In a specific embodiment, the reference session response determination module 420 includes:

[0168] a knowledge base answer mode determination unit, configured to set the target answer mode corresponding to the conversation question to the knowledge base answer mode when the knowledge base semantic similarity is greater than a first semantic similarity threshold;

[0169] The general model answer mode determination unit is used to set the target answer mode corresponding to the conversation question to the general model answer mode when the knowledge base semantic similarity is less than or equal to the first semantic similarity threshold.

[0170] In a specific embodiment, the preset knowledge base includes a question-answer knowledge base and a document knowledge base, and the knowledge base response mode is a question-answer knowledge base response mode or a document knowledge base response mode; accordingly, the knowledge base response mode determination unit is specifically used to:

[0171] When the knowledge base semantic similarity is greater than a second semantic similarity threshold, setting the target answer mode corresponding to the conversation question to the question-and-answer knowledge base answer mode;

[0172] When the knowledge base semantic similarity is less than or equal to the second semantic similarity threshold, the target answer mode corresponding to the conversation question is set to the document knowledge base answer mode.

[0173] In a specific embodiment, the reference session response determination module 420 includes:

[0174] A text similarity acquisition unit, used for acquiring text similarities respectively corresponding to the conversation question and at least two preset knowledge base questions in the question and answer knowledge base when the target answer mode is the question and answer knowledge base answer mode;

[0175] A first target knowledge base question screening unit, configured to screen a target knowledge base question from at least two preset knowledge base questions according to a text screening condition and at least two text similarities;

[0176] The first reference conversation answer determination unit is used to use the knowledge base answer corresponding to the target knowledge base question in the question and answer knowledge base as the reference conversation answer corresponding to the conversation question.

[0177] In a specific embodiment, the reference session response determination module 420 further includes:

[0178] A semantic similarity acquisition unit, used for acquiring semantic similarities corresponding to the conversation question and at least two preset knowledge base questions respectively when there is no target knowledge base question satisfying the text screening condition;

[0179] A second target knowledge base question screening unit, configured to screen a target knowledge base question from at least two preset knowledge base questions according to a semantic screening condition and at least two semantic similarities;

[0180] The second reference conversation answer determination unit is used to use the knowledge base answer corresponding to the target knowledge base question in the question and answer knowledge base as the reference conversation answer corresponding to the conversation question.

[0181] In a specific embodiment, the reference session response determination module 420 includes:

[0182] A third reference conversation answer determination unit, configured to obtain a user portrait corresponding to the user to be answered and historical conversation data corresponding to the conversation question when the target answer mode is the document knowledge base answer mode;

[0183] The user portrait, conversation questions, historical conversation data and document knowledge base are input into the pre-trained document answer model to obtain the output reference conversation answer corresponding to the conversation question.

[0184] In a specific embodiment, the device further comprises:

[0185] A preset knowledge base acquisition module is used to input the conversation question and the historical conversation data corresponding to the conversation question into the pre-trained question intention prediction model to obtain the output target question intention;

[0186] Determine the service attribute corresponding to the user to be answered according to the user identifier corresponding to the user to be answered;

[0187] According to business attributes and target question intent, a preset knowledge base is obtained from the original knowledge base.

[0188] The session processing device provided in the embodiment of the present invention can execute the session processing method applied to the session processing platform provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0189] Figure 7 The structure diagram of a session processing device provided by an embodiment of the present invention is applied to a session platform. Figure 7 As shown, the device includes: a conversation question sending module 510, a response matching degree receiving module 520 and a target conversation response determining module 530.

[0190] The session question sending module 510 is used to send the session question to the session processing platform in response to detecting the session question input by the waiting user;

[0191] The response matching degree receiving module 520 is used to receive the reference session response and the response matching degree sent by the session processing platform;

[0192] A target conversation answer determination module 530 is used to use the reference conversation answer as a target conversation answer corresponding to the conversation question when the answer matching degree satisfies the matching degree condition;

[0193] The reference conversation answer is determined by the conversation processing platform according to the conversation question using a target answer mode, the target answer mode is determined according to the semantic similarity between the conversation question and the knowledge base corresponding to the preset knowledge base, and the target answer mode is a knowledge base answer mode or a general model answer mode;

[0194] The answer matching degree is obtained by the conversation processing platform by inputting the conversation question and reference conversation answer into the pre-trained question-answer matching model.

[0195] The technical solution of this embodiment solves the problem that the traditional conversation processing method relies on a single response mode by determining the knowledge base semantic similarity between the conversation question and the preset knowledge base, and determines the target response mode corresponding to the conversation question based on the knowledge base semantic similarity, and adopts the target response mode to determine the reference conversation response corresponding to the conversation question, thereby improving the ability of the conversation processing method to deal with conversation questions with different description texts, and thus improving the quality of the conversation response. In addition, the answer matching degree between the reference conversation response and the conversation question is obtained through the question-answer matching model, so that the conversation platform uses the reference conversation response as the target conversation response corresponding to the conversation question when the answer matching degree meets the matching degree condition, replacing the step of manual customer service to perform quality inspection on the answer content, while ensuring the answer quality, it also improves the efficiency of the conversation response.

[0196] In a specific embodiment, the device further comprises:

[0197] The information output module to be pushed is used to obtain the information source corresponding to the information to be pushed in response to detecting the information to be pushed sent by the session processing platform;

[0198] When the information source is a promotion source, the information to be pushed is approved, and when the approval result is approved, the information to be pushed is output to the conversation window corresponding to the user to be answered;

[0199] When the information source is a non-promotion source, the information to be pushed is output to the conversation window corresponding to the user to be answered.

[0200] In a specific embodiment, the device further comprises:

[0201] A conversation group establishment module, configured to, in response to detecting a group establishment instruction, obtain at least one target user that has a binding relationship with the customer service robot from at least one preset user corresponding to the group establishment instruction;

[0202] Determine, according to the user identifier corresponding to at least one target user, the conversation group information corresponding to the group establishment instruction, and establish the conversation group according to the conversation group information;

[0203] The conversation group includes users to be answered, and the conversation group information includes at least one of a group name, a group announcement, a group note information, and a group message.

[0204] The session processing device provided in the embodiment of the present invention can execute the session processing method applied to the session platform provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0205] Figure 8 This is a structural diagram of a session processing system provided by an embodiment of the present invention. The session processing system in this embodiment provides services for the session processing method in the above embodiment.

[0206] like Figure 8 As shown, the session processing system 600 includes a session processing platform 610 and a session platform 620 that are communicatively connected; wherein the session processing platform 610 is used to execute the session processing method applied to the session processing platform 610 in any of the above embodiments of the present invention, and the session platform 620 is used to execute the session processing method applied to the session platform 620 in any of the above embodiments of the present invention.

[0207] On the basis of the above embodiment, specifically, the conversation processing platform 610 includes a customer service robot 611 that is communicatively connected to the conversation platform 620 and a conversation response platform 612 that is communicatively connected to the customer service robot 611; wherein the conversation response platform 612 is used to determine the semantic similarity of the knowledge base according to the conversation question sent by the customer service robot 611 and the preset knowledge base, determine the target response mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and use the target response mode to determine the reference conversation response corresponding to the conversation question, input the conversation question and the reference conversation response into a pre-trained question-answer matching model to obtain an output answer matching degree; the customer service robot 611 is used to send the conversation question sent by the conversation platform 620 to the conversation response platform 612, receive the reference conversation response sent by the conversation response platform 612 and the answer matching degree corresponding to the reference conversation response, and send the reference conversation response and the answer matching degree to the conversation platform 620, so that the conversation platform 620 uses the reference conversation response as the target conversation response corresponding to the conversation question when the answer matching degree meets the matching degree condition.

[0208] Specifically, the conversation response platform 612 includes a knowledge base and a language model, wherein the knowledge base can be used to support model training of the language model.

[0209] Fig. 9A functional architecture diagram of a session processing system provided by an embodiment of the present invention is shown in FIG. Fig. 9 The language model and knowledge base in the conversation processing system 600 constitute the conversation response platform 612. Specifically, Fig. 9 The customer service robot in the communication is connected with the conversation platform, so that the active contact function and passive reply function in the group interaction scenario and the 1V1 interaction scenario are realized in the conversation platform, wherein the active contact function represents sending the information to be pushed to the conversation window corresponding to the user to be answered, and the passive reply function represents sending the target conversation answer to the conversation window corresponding to the user to be answered.

[0210] Fig. 9 The knowledge base in the repository stores question-and-answer knowledge base, document knowledge base, and user portraits. The input data of the speech model includes but is not limited to text, images, videos, and audio. The language model is trained and deployed with the support of the knowledge base. The language model is deployed with a semantic model for predicting semantic similarity and a question-and-answer matching model.

[0211] Fig.10 The interactive flow chart of a conversation processing system provided by an embodiment of the present invention is as follows: specifically, the conversation platform monitors whether the client has a conversation question input. If so, the user identification, conversation question, historical conversation data, and message creation time are sent to the conversation processing platform through an interface call. The conversation processing platform uses a language model and a knowledge base to send the user identification, reference conversation response, response matching degree, and message creation time to the conversation platform through an interface. The conversation platform determines whether the response matching degree is a successful match. If so, the reference conversation response is output to the client as the target conversation response. If not, a manual response is made.

[0212] On the basis of the above embodiments, specifically, the conversation processing system 600 also includes a third-party push platform 630 that is respectively communicated with the customer service robot 611, and is used to send the information to be pushed to the customer service robot 611; wherein the customer service robot 611 is also used to send the information to be pushed to the conversation platform 620; the conversation platform 620 is also used to obtain the information source corresponding to the information to be pushed, and according to the information source, output the information to be pushed to the conversation window corresponding to the user to be answered.

[0213] Fig.11A functional schematic diagram of active contact of a conversation platform provided by an embodiment of the present invention. Specifically, push platform A and push platform B send the information to be pushed to the customer service robot, and the customer robot sends the information to be pushed to the conversation platform through the active contact function. The conversation platform determines whether to approve the information according to the information source corresponding to the information to be pushed. If the approval is carried out and passed, the information to be pushed is output to the conversation window corresponding to the user to be answered. If the approval is carried out but not passed, the information to be pushed is deleted. If the approval is not carried out, the information to be pushed is output to the conversation window corresponding to the user to be answered.

[0214] The computer programs for implementing the session processing method of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that when the computer programs are executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer programs may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0215] In the context of the present application, a computer readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, device or equipment. A computer readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer readable storage medium may be a machine readable storage medium. Examples of machine readable storage media may include electrical connections based on at least one line, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM), flash memories, optical fibers, portable compact disc read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0216] To provide interaction with a user, the systems and techniques described herein may be implemented on a terminal device having: a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball), through which the user can provide input to the terminal device. Other types of devices may also provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0217] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a Local Area Network (LAN), a Wide Area Network (WAN), a blockchain network, and the Internet.

[0218] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and virtual private servers (VPS) services.

[0219] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0220] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A session processing method, characterized in that: Applied to session processing platforms, including: In response to receiving a conversation question from a user to be answered sent by the conversation platform, determining a semantic similarity of the knowledge base according to the conversation question and a preset knowledge base; Determining a target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and determining a reference conversation answer corresponding to the conversation question by adopting the target answer mode; Inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree; sending the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as a target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition; Wherein, the target response mode is a knowledge base response mode or a general model response mode.

2. The method according to claim 1, characterized in that Determining the target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base includes: When the knowledge base semantic similarity is greater than a first semantic similarity threshold, setting the target answer mode corresponding to the conversation question to a knowledge base answer mode; When the semantic similarity of the knowledge base is less than or equal to the first semantic similarity threshold, the target answer mode corresponding to the conversation question is set to a general model answer mode.

3. The method according to claim 2, characterized in that The preset knowledge base includes a question-answer knowledge base and a document knowledge base, and the knowledge base response mode is a question-answer knowledge base response mode or a document knowledge base response mode; Accordingly, setting the target answer mode corresponding to the conversation question to the knowledge base answer mode includes: When the knowledge base semantic similarity is greater than a second semantic similarity threshold, setting the target answer mode corresponding to the conversation question to a question-and-answer knowledge base answer mode; When the knowledge base semantic similarity is less than or equal to the second semantic similarity threshold, the target answer mode corresponding to the conversation question is set to the document knowledge base answer mode.

4. The method according to claim 3, characterized in that The step of adopting the target response mode to determine a reference conversation response corresponding to the conversation question includes: When the target answering mode is a question-and-answer knowledge base answering mode, obtaining text similarities respectively corresponding to the conversation question and at least two preset knowledge base questions in the question-and-answer knowledge base; Filtering a target knowledge base question from the at least two preset knowledge base questions according to a text filtering condition and at least two text similarities; The knowledge base answer corresponding to the target knowledge base question in the question-answer knowledge base is used as a reference conversation answer corresponding to the conversation question.

5. The method according to claim 4, characterized in that The step of adopting the target answer mode to determine a reference conversation answer corresponding to the conversation question further includes: When there is no target knowledge base question that meets the text screening condition, obtaining semantic similarities respectively corresponding to the conversation question and at least two preset knowledge base questions; Filtering a target knowledge base question from the at least two preset knowledge base questions according to a semantic filtering condition and at least two semantic similarities; The knowledge base answer corresponding to the target knowledge base question in the question-answer knowledge base is used as a reference conversation answer corresponding to the conversation question.

6. The method according to claim 3, characterized in that The step of adopting the target response mode to determine a reference conversation response corresponding to the conversation question includes: When the target answering mode is the document knowledge base answering mode, obtaining a user portrait corresponding to the user to be answered and historical conversation data corresponding to the conversation question; The user portrait, the conversation question, the historical conversation data and the document knowledge base are input into a pre-trained document response model to obtain an output reference conversation response corresponding to the conversation question.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Inputting the conversation question and the historical conversation data corresponding to the conversation question into a pre-trained question intention prediction model to obtain an output target question intention; Determining the service attribute corresponding to the user to be answered according to the user identifier corresponding to the user to be answered; According to the business attributes and the target question intention, a preset knowledge base is obtained from the original knowledge base.

8. A session processing method, characterized in that: Applied to conversational platforms, including: In response to detecting a conversation question input by a user to be answered, sending the conversation question to a conversation processing platform; Receiving a reference session response and a response matching degree sent by the session processing platform; In the case where the answer matching degree satisfies the matching degree condition, taking the reference conversation answer as the target conversation answer corresponding to the conversation question; The reference conversation answer is determined by the conversation processing platform according to the conversation question using a target answer mode, the target answer mode is determined according to the semantic similarity between the conversation question and a knowledge base corresponding to a preset knowledge base, and the target answer mode is a knowledge base answer mode or a general model answer mode; The answer matching degree is obtained by the conversation processing platform inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model.

9. The method according to claim 8, characterized in that The method further comprises: In response to detecting the information to be pushed sent by the session processing platform, obtaining an information source corresponding to the information to be pushed; When the information source is a promotion source, performing an approval operation on the information to be pushed, and if the approval result obtained is approval passed, outputting the information to be pushed to a conversation window corresponding to the user to be answered; When the information source is a non-promotion source, the information to be pushed is output to a conversation window corresponding to the user to be responded to.

10. The method according to claim 8, characterized in that The method further comprises: In response to detecting a group creation instruction, obtaining at least one target user that has a binding relationship with the customer service robot from at least one preset user corresponding to the group creation instruction; Determining conversation group information corresponding to the group creation instruction according to user identifiers corresponding to at least one target user, and establishing a conversation group according to the conversation group information; The conversation group includes the waiting user, and the conversation group information includes at least one of a group name, a group announcement, a group note information, and a group message.

11. A conversation processing device, characterized in that: Applied to session processing platforms, including: A knowledge base semantic similarity determination module, configured to determine the knowledge base semantic similarity according to the conversation question and a preset knowledge base in response to receiving a conversation question of a user to be answered sent by the conversation platform; A reference conversation answer determination module, used to determine a target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and determine a reference conversation answer corresponding to the conversation question using the target answer mode; An answer matching degree determination module is used to input the conversation question and the reference conversation answer into a pre-trained question-answer matching model to obtain an output answer matching degree; an answer matching degree sending module, configured to send the reference conversation answer and the answer matching degree to the conversation platform, so that the conversation platform uses the reference conversation answer as a target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition; Wherein, the target response mode is a knowledge base response mode or a general model response mode.

12. A conversation processing device, characterized in that: Applied to conversational platforms, including: A session question sending module, configured to send the session question to the session processing platform in response to detecting the session question input by the user to be answered; A response matching degree receiving module, used for receiving the reference session response and the response matching degree sent by the session processing platform; a target conversation answer determination module, configured to use the reference conversation answer as a target conversation answer corresponding to the conversation question if the answer matching degree satisfies a matching degree condition; The reference conversation answer is determined by the conversation processing platform according to the conversation question using a target answer mode, the target answer mode is determined according to the semantic similarity between the conversation question and a knowledge base corresponding to a preset knowledge base, and the target answer mode is a knowledge base answer mode or a general model answer mode; The answer matching degree is obtained by the conversation processing platform inputting the conversation question and the reference conversation answer into a pre-trained question-answer matching model.

13. A session processing system, characterized in that: The session processing system comprises: a session processing platform and a session platform in communication connection; The session processing platform is used to execute the session processing method described in any one of claims 1 to 7, and the session platform is used to execute the session processing method described in any one of claims 8 to 10.

14. The session processing system according to claim 13, characterized in that: The conversation processing platform includes a customer service robot communicatively connected to the conversation platform and a conversation answering platform communicatively connected to the customer service robot; The conversation answering platform is used to determine the semantic similarity of the knowledge base according to the conversation question sent by the customer service robot and the preset knowledge base, determine the target answer mode corresponding to the conversation question according to the semantic similarity of the knowledge base, and use the target answer mode to determine the reference conversation answer corresponding to the conversation question, input the conversation question and the reference conversation answer into the pre-trained question-answer matching model, and obtain the output answer matching degree; The customer service robot is used to send the conversation questions sent by the conversation platform to the conversation response platform, and send the reference conversation responses and response matching degrees received from the conversation response platform to the conversation platform.

15. The session processing system according to claim 14, characterized in that: The conversation processing system also includes a third-party push platform that is communicatively connected to the customer service robot and is used to send the information to be pushed to the customer service robot; Wherein, the customer service robot is further used to send the information to be pushed to the conversation platform; The conversation platform is further used to obtain the information source corresponding to the information to be pushed, and output the information to be pushed to the conversation window corresponding to the user to be responded to according to the information source.

16. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the session processing method according to any one of claims 1 to 7 or the session processing method according to any one of claims 8 to 10 when executed.

17. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the session processing method according to any one of claims 1 to 7 or the session processing method according to any one of claims 8 to 10.