Message interaction method and device based on AI customer service robot

Abstract

The invention discloses a message interaction method and device based on an AI customer service robot. The method comprises the following steps: classifying received user messages to obtain messages which need to be replied and messages which do not need to be replied; extracting a question of the user from the message needing to be replied, and retrieving related content from a knowledge base based on the extracted question; performing context analysis on the message needing to be replied, and reordering related contents retrieved from the knowledge base on the basis of a context analysis result; and providing the reordered related content to an AI large model, and obtaining reply content through the AI large model so as to respond to the message needing to be replied. The technical problem of inaccurate reply of the customer service robot in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a message interaction method and device based on an AI customer service robot. Background Art

[0002] In the prior art, when users encounter problems, they usually get help from manual customer service. Although this method can solve most problems, it still has many disadvantages. First, the cost of manual customer service is high, especially when 24-hour service is required, the operating cost increases significantly. In addition, the response speed of manual customer service is slow, especially during peak hours, which can easily cause users to wait too long and affect the user experience. On the other hand, the working hours of manual customer service are limited and it is impossible to provide round-the-clock service, while AI customer service can reduce customer service pressure by automating the handling of common problems. However, when dealing with complex problems, the existing AI customer service system often needs to rely on the intervention of manual customer service, which makes it impossible to completely replace the role of manual customer service, and its processing accuracy and flexibility are still insufficient in complex scenarios.

[0003] Existing intelligent customer service technologies based on natural language processing and machine learning are highly efficient in solving simple problems, but they are still difficult to handle when users raise complex cross-domain and cross-file questions. In particular, for automatic correction of user assignments, existing technologies have difficulty in identifying cross-file reference errors or correcting multi-layer logical errors. Human customer service has limited efficiency and is easily affected by emotions, while AI customer service performs poorly in handling complex tasks, and its overall work efficiency and accuracy need to be improved.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present invention provide a message interaction method and device based on an AI customer service robot, so as to at least solve the technical problem of inaccurate responses of customer service robots in the prior art.

[0006] According to one aspect of an embodiment of the present invention, a message interaction method based on an AI customer service robot is provided, comprising: classifying received user messages to obtain messages that require a reply and messages that do not require a reply; extracting user questions from the messages that require a reply, and retrieving relevant content from a knowledge base based on the extracted questions; performing context analysis on the messages that require a reply, and reordering the relevant content retrieved from the knowledge base based on the results of the context analysis; providing the reordered relevant content to an AI big model, and obtaining reply content through the AI ​​big model to respond to the messages that require a reply.

[0007] According to another aspect of an embodiment of the present invention, a message interaction device based on an AI customer service robot is also provided, including: a receiving module, configured to classify received user messages to obtain messages that require replies and messages that do not require replies; a retrieval module, configured to extract user questions from the messages that require replies, and retrieve relevant content from a knowledge base based on the extracted questions; an analysis module, configured to perform context analysis on the messages that require replies, and reorder the relevant content retrieved from the knowledge base based on the results of the context analysis; a reply module, configured to provide the reordered relevant content to an AI big model, and obtain reply content through the AI ​​big model to respond to the messages that require replies.

[0008] In an embodiment of the present invention, received user messages are classified to obtain messages that need to be replied and messages that do not need to be replied; user questions are extracted from the messages that need to be replied, and relevant content is retrieved from the knowledge base based on the extracted questions; context analysis is performed on the messages that need to be replied, and the relevant content retrieved from the knowledge base is reordered based on the results of the context analysis; the reordered relevant content is provided to the AI ​​big model, and the reply content is obtained through the AI ​​big model to respond to the messages that need to be replied. The above scheme solves the technical problem of inaccurate replies by customer service robots in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0010] Figure 1 It is a flow chart of a message interaction method based on an AI customer service robot according to an embodiment of the present application;

[0011] Figure 2 is a flowchart of another message interaction method based on an AI customer service robot according to an embodiment of the present application;

[0012] Figure 3 is a flowchart of message classification according to an embodiment of the present application;

[0013] Figure 4 is a flowchart of a method for extracting questions according to an embodiment of the present application;

[0014] Figure 5 is a flowchart of a method for searching a knowledge base according to an embodiment of the present application;

[0015] Figure 6 is a flowchart of a reordering method according to an embodiment of the present application;

[0016] Figure 7 is a flowchart of a method for context analysis according to an embodiment of the present application;

[0017] Figure 8 It is a structural schematic diagram of a message interaction device based on an AI customer service robot according to an embodiment of the present application;

[0018] Fig. 9 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0019] 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.

[0020] It should be noted that the terms "first", "second", 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 interchanged 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 that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] According to an embodiment of the present invention, a method embodiment of a message interaction method based on an AI customer service robot is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0022] Figure 1 is a message interaction method based on an AI customer service robot according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0023] Step S102: Classify the received user messages to obtain messages that need a reply and messages that do not need a reply.

[0024] For example, a trained NLP classification model is used to identify the intent of the user message; based on the result of the intent identification, it is determined whether the user message needs a reply, so as to classify the user message.

[0025] Step S104: extracting the user's question from the message that needs a reply, and retrieving relevant content from the knowledge base based on the extracted question.

[0026] First, a plurality of knowledge documents are obtained, the content of each of the plurality of knowledge documents is parsed, and the parsed knowledge document content is converted into a corresponding feature vector representation; and the feature vector representation corresponding to the parsed knowledge document content is stored in a knowledge base.

[0027] Next, the user's question is extracted from the message that needs to be replied. The message that needs to be replied is semantically parsed using semantic parsing technology to obtain a parsed semantic structure; based on the parsed semantic structure, the core content of the message that needs to be replied is extracted to obtain the extracted question.

[0028] Then, relevant content is retrieved from the knowledge base based on the extracted question. The extracted question is converted into a feature vector representation; and relevant content similar to the feature vector representation is retrieved from the knowledge base based on the feature vector representation.

[0029] Step S106: performing context analysis on the message that needs to be replied, and reordering the relevant content retrieved from the knowledge base based on the result of the context analysis.

[0030] For example, context analysis is performed on the message that needs to be replied to, semantic features are extracted and corresponding semantic vectors are generated, and based on the semantic vectors, similarity is calculated between the retrieved related content and the semantic vectors; priority rules are set according to context information, user historical preferences, timeliness and importance of the content; and based on the similarity calculation results and the priority rules, the retrieved related content is reordered.

[0031] Step S108, providing the reordered related content to the AI ​​big model, and obtaining reply content through the AI ​​big model to respond to the message that needs a reply.

[0032] This application also provides another message interaction method based on AI customer service robot, such as Figure 2 As shown, the method comprises the following steps:

[0033] Step S202: message classification.

[0034] The received user messages are initially processed by the NLP model to determine the type of message and whether further replies are required. Message classification is not limited to the judgment of grammatical structure, but also involves operations such as semantic analysis and intent recognition. The specific implementation steps are as follows: Figure 3 As shown, the following steps are included:

[0035] Step S2022: message reception and preprocessing.

[0036] When a user sends a message to the AI ​​customer service system, the system needs to perform preprocessing because the user message may contain spelling errors, special characters, or non-standard language expressions. First, the system identifies and corrects spelling errors in user messages through a predefined dictionary or a deep learning model based on character sequences. Filter out extra special characters in the message (such as extra punctuation marks) to avoid interference with subsequent analysis. For some languages ​​without spaces, word segmentation is performed to divide the message content into recognizable phrases or phrases.

[0037] Step S2024, intent recognition.

[0038] The preprocessed messages are processed by the intent recognition module. This module is based on deep learning models (such as BERT, GPT, etc.) to identify the user's intent from the user's message. The specific steps include the following:

[0039] Use a pre-trained deep learning model to convert user messages into high-dimensional word vector representations. The model generates a vector representation of the message through contextual semantics in order to capture the user's hidden intentions. The word vectors are input into multiple predefined intent categories, such as query, complaint, feedback, etc., through classification models (such as SVM, Softmax, etc.). In this way, the system determines the core intent of the message and classifies it according to the intent results. Based on the results of intent recognition, the AI ​​customer service system classifies user messages and determines whether the message needs to be replied: For messages that require further feedback such as inquiries and questions, the system classifies them as messages that need to be replied and enters the next processing flow. For messages that do not require a reply, such as non-critical messages such as confirmation messages, the system automatically classifies them as messages that do not require a reply, and does not perform subsequent processing to reduce resource usage.

[0040] Step S204, question extraction.

[0041] After completing the message classification, the system then processes the messages that need to be replied to. Through the question extraction module, the key questions in the user's message are accurately extracted for knowledge base retrieval in the subsequent steps. The implementation technology of the question extraction link relies on semantic parsing and context analysis to ensure that the system understands the specific questions raised by the user. Figure 4 As shown, the following steps may be included:

[0042] Step S2042, semantic analysis.

[0043] By parsing the syntactic structure in the user's message, the user's core question is extracted. For example, first, a semantic tree is constructed. The user message is segmented and a semantic tree is generated based on grammatical rules. The semantic tree can show the subject-predicate-object structure in the sentence, helping the system to identify the main idea expressed by the user. For example, for the sentence "I want to query this week's courses", the semantic tree will use "query" as the core verb and "this week's courses" as the query object. After that, syntactic analysis is performed. The key to semantic parsing is the recognition and analysis of syntactic structure. The system extracts the core information of the user's sentence through a dependency-based model (such as Stanford Parser).

[0044] Step S2044, context-assisted extraction.

[0045] Semantic parsing can identify the core issues in a sentence, but sometimes syntactic analysis alone is not enough to accurately capture the user's full intent, especially when multiple sentences or contexts are involved. Therefore, it is necessary to further adjust the extraction results by combining the user's historical conversation records and the context information of the current session.

[0046] First, historical records are associated. By analyzing the user's historical conversation records, the background information of the user's current question is captured. For example, if a user mentions the same question in multiple conversations, the real focus of the user's question can be confirmed by combining historical records. Next, context association is performed. When the sentence raised by the user is relatively short or unclear (such as "help me check this"), the system will use context analysis technology to associate the object or question mentioned by the user before, so as to accurately extract the current question.

[0047] Step S2046, problem vectorization.

[0048] After extracting the core content of the user's question, the question is converted into a vector representation for subsequent knowledge base retrieval. The conversion process relies on a deep learning model, such as using a pre-trained BERT model to convert the question into a word vector.

[0049] Step S206, searching the knowledge base.

[0050] After the question is extracted, the answer related to the user's question is retrieved from the knowledge base. The knowledge base is the core component of the customer service system, storing a large amount of documents, FAQs and support information for the system to retrieve and feedback. Figure 5 As shown, searching the knowledge base includes the following steps:

[0051] Step S2062, knowledge base construction.

[0052] In order to support efficient knowledge retrieval, a knowledge base is first constructed. The data in the knowledge base can include product manuals, FAQs, technical support documents, etc. These documents are preprocessed and stored in the form of feature vectors. Specifically, the knowledge documents are first parsed. The system parses the documents in the knowledge base and extracts the key information in the documents. Each document will go through steps such as word segmentation and semantic analysis to generate a set of feature vectors representing the main content of the document. Next, the feature vector is stored, and the parsed document feature vector will be stored in the database for subsequent similarity calculation and matching.

[0053] Step S2064, retrieve the question vector.

[0054] After the user question is converted into a vector representation, the most relevant content to the question is retrieved from the knowledge base. By calculating the similarity between the user question vector and the document vector in the knowledge base, the document content most relevant to the user question is found. The similarity is usually calculated using methods such as cosine similarity and Euclidean distance, and the similarity between the two is measured by comparing the angle or distance between the vectors. For complex questions, this application not only matches a single document, but also matches different parts of multiple documents based on the multi-level structure of the question. For example, a user question may involve multiple sub-questions. In this case, the answers to each sub-question are matched separately in the knowledge base.

[0055] Step S2066, search optimization.

[0056] In order to improve the accuracy and efficiency of retrieval, this application adopts a variety of optimization strategies, such as weight-based retrieval, where documents in the knowledge base are assigned different weights based on the importance of the content and the frequency of update. When searching, the system will give priority to documents with higher weights. As users ask questions and the system continues to use, the documents in the knowledge base will be continuously updated to ensure the timeliness and accuracy of the retrieval results.

[0057] Step S208, reordering.

[0058] After retrieving relevant content, the results are further optimized through the re-ranking module. The re-ranking process involves considerations of multiple dimensions to ensure that the returned results meet the user's expectations. Figure 6 As shown, the specific steps of reordering include the following:

[0059] Step S2082, similarity reordering.

[0060] First, a preliminary re-ranking is performed based on the similarity between the user question and the retrieved documents. The relevance of each document is determined based on the similarity between the user question vector and the retrieved document vector. Documents with higher similarity will be prioritized.

[0061] Step S2084, context relevance analysis.

[0062] On the basis of similarity re-ranking, the ranking results are further optimized by combining the context information of the current conversation. Context relevance analysis is to ensure that the system can not only match relevant answers, but also return results that are more in line with the user's current needs based on the context of the conversation. Specifically, a context vector is generated first. By performing context analysis on the user's entire conversation, a semantic vector corresponding to the context is generated. The context vector is generated by analyzing the user's previous interaction history and the semantic clues of the current conversation to capture the implicit intention in the conversation. Next, similarity and context are combined for analysis. The vector representation of the user's current question is combined with the context vector, and a new score for the relevance to each document is obtained through weighted calculation. The combination of context and questions helps the system better understand the user's current needs. For example, if the user has previously asked a question about "course fees", the system will prioritize returning documents related to course fees based on this context, rather than other content that may only be similar but not relevant.

[0063] Step S2086, setting the user's historical preference weight.

[0064] During the re-sorting process, the system will also base it on the user's historical preferences. Preferences can be dynamically adjusted based on the user's previous questions, usage habits, and common questions in the past. For example, if a user often asks questions related to a certain type of product, the relevant documents of this type of product will be given priority during sorting. Specifically, the system builds a user's interest profile based on the user's historical question data. The system analyzes information such as the frequency and type of the user's questions, and maps the user's historical behavior to a preference weight. The weight is not fixed, but is dynamically adjusted based on the user's real-time behavior. For example, when a user has recently frequently asked questions about a newly released product, the system will temporarily increase the weight of documents related to the new product in the sorting.

[0065] Step S2088, setting timeliness and importance priorities.

[0066] In addition to the user's historical preferences, the re-ranking is also based on the timeliness and importance of the document content. Especially in technical support questions, answers with strong timeliness are often more effective. Specifically, the content update frequency analysis is first performed. The update time of each document is tracked to ensure that newer content is displayed first in the sorting. Then, for some key information (such as emergency announcements or important policy updates), specific priority rules are set to make this information have a greater weight in the sorting than general content.

[0067] Through the above steps, the present application comprehensively considers the retrieved content in multiple dimensions such as similarity, contextual relevance, user preference, timeliness and importance, and optimizes the ranking results.

[0068] Step S210, context analysis.

[0069] Context analysis methods such as Figure 7 As shown, the following steps are included:

[0070] Step S2102, conversation history tracking.

[0071] Track the entire conversation history of the user to ensure the semantic relevance of multiple rounds of conversations. For example, when a user expresses a complex question in multiple consecutive messages, the conversation history can be tracked to ensure that the semantic relevance before and after is not ignored. Specifically, a conversation ID is assigned. The system generates a unique conversation ID for each conversation to distinguish different users and different conversation processes, ensuring that previous conversations are recorded and analyzed in the same conversation process. Then, cache the conversation history. The historical conversation content of each conversation is cached in memory, and the corresponding semantic vector representation is generated.

[0072] Step S2104: semantic feature extraction.

[0073] The core of context analysis is to extract key semantic features from multiple rounds of user conversations and convert them into feature vectors as an important basis for subsequent processing. The system processes the user's continuous conversations through a dynamic semantic model (such as a model based on the Transformer architecture) to extract implicit semantic features. The dynamic model can adjust the results of semantic understanding according to the context of the conversation, ensuring that the system can take into account the content of previous conversations when processing subsequent questions. The semantic information of each historical conversation will be converted into a semantic vector and stored in the system's temporary memory. The semantic vector of the current user's question will be merged with the semantic vector of the historical conversation to ensure that the system processes the question based on the overall semantic structure of the context.

[0074] Step S2106, context semantic association.

[0075] After extracting the user's contextual semantic features, the semantic vector of the current question is fused with the context vector. By calculating the similarity between the current question vector and the context vector, the two are fused into a new feature vector. The fused feature vector can better express the semantic background of the user's current question, ensuring that the system understands the question more accurately. In multi-round dialogue scenarios, the context of the user's current question is identified through semantic association. For example, when a user uses pronouns (such as "this", "mentioned last time") in a continuous dialogue, the specific object or question referred to by the user is understood through context analysis.

[0076] Step S212, large model reply.

[0077] After message classification, question extraction, knowledge base retrieval, reordering and context analysis, the system finally needs to generate specific reply content through the AI ​​big model. The big model reply method includes the following steps:

[0078] First, the feature vector is input into the large model.

[0079] The feature vector obtained through context analysis and knowledge base retrieval is input into the AI ​​big model. As the input of the big model, the feature vector helps the model understand the semantic structure of the user's question and generate matching reply content. Before entering the big model, the feature vector is standardized to ensure that the numerical range of the input vector meets the expectations of the model. At the same time, standardization can also improve the computational efficiency and accuracy of the model. For scenarios involving multiple rounds of dialogue, the system will input the context vector together with the current question vector into the big model to ensure that the generated reply content can maintain consistency with the previous dialogue.

[0080] Next, the big model generates the response content.

[0081] The large model generates matching responses based on the input feature vector. The model performs complex semantic reasoning and generation through deep neural networks (such as the Transformer architecture). The large model infers the core intent of the user's question based on the semantic information in the input vector and generates corresponding responses. For example, when a user asks about the purchase steps of a product, the model can infer the specific terms that the user is concerned about and generate relevant responses. The large model generates natural language responses based on the inferred semantic results through the language generation module.

[0082] In order to improve the user experience, the AI ​​big model can also generate diverse and personalized responses based on questions from different users. For example, when generating responses, the big model can randomize the selection of certain words or phrases to ensure that the same question does not always return exactly the same response. This diversity generation not only improves the user experience, but also makes the response content more vivid and closer to human conversation habits. In addition, the system combines the user's historical preferences and contextual information to generate more personalized responses. For example, when the system recognizes that a user often uses short inquiries, the AI ​​model will automatically generate concise and direct responses to match the user's communication style.

[0083] Among them, the training process of the large model is as follows:

[0084] First, the primary embedding vector group of the speech recognition text and the auxiliary embedding vector group of the manual recognition text are extracted respectively. Specifically, the speech recognition text generates the primary embedding vector group through the model, and these vector groups represent the characteristics of the text in the multidimensional semantic space. Similarly, the manual recognition text generates the auxiliary embedding vector group to represent its semantic information. Next, according to the preset expansion rank criterion, the primary embedding vector group and the auxiliary embedding vector group are respectively multi-dimensionally upgraded.

[0085] Subsequently, the two vector groups are feature coupled and fused using a multi-channel convolution kernel to obtain a coupled primary embedding group and an auxiliary embedding group. Next, a nonlinear activation function is applied to activate the coupled vector group to generate an activated primary embedding group and an auxiliary embedding group. Next, a normalization operation is performed to ensure the consistency of the data scale of the activated vector group in different dimensions.

[0086] Next, a high-order matrix factorization is performed on the normalized primary and auxiliary embedding groups. These embedding groups are expanded along the expansion dimension to form multiple sub-matrices. Singular value decomposition is performed on each sub-matrix to extract the core subspace representation and factor matrix. The core subspace representation condenses the key information in the matrix, while the factor matrix captures the local structural features.

[0087] Finally, the principal component differences between the primary and auxiliary embedding groups are calculated, and the key differences between the speech recognition text and the human recognition text are identified through similarity analysis and principal component difference measurement. Based on these differences, the loss function of the model is updated, and finally a trained model is obtained.

[0088] This method combines high-order matrix decomposition with singular value decomposition to fully extract and retain the structural information between speech recognition text and manual recognition text, thereby improving the training effect of the model.

[0089] Step S214, return to the large model reply content

[0090] When the big model generates the final reply content, the system returns it to the user, completing the entire message interaction process. Before returning it to the user, the system will properly format the reply content generated by the big model to ensure the display effect of the reply content.

[0091] This application also provides a message interaction device based on an AI customer service robot, such as Figure 8 As shown, it includes: a receiving module 82, which is configured to classify the received user messages to obtain messages that need to be replied and messages that do not need to be replied; a retrieval module 84, which is configured to extract the user's questions from the messages that need to be replied, and retrieve relevant content from the knowledge base based on the extracted questions; an analysis module 86, which is configured to perform context analysis on the messages that need to be replied, and reorder the relevant content retrieved from the knowledge base based on the results of the context analysis; a reply module 88, which is configured to provide the reordered relevant content to the AI ​​big model, and obtain reply content through the AI ​​big model to respond to the messages that need to be replied.

[0092] It should be noted that the above embodiment is based on the message interaction device of the AI ​​customer service robot, and only uses the division of the above functional modules as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the message interaction device based on the AI ​​customer service robot provided in the above embodiment and the message interaction method embodiment based on the AI ​​customer service robot belong to the same concept. The specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0093] Fig. 9 FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Fig. 9 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0094] like Fig. 9 As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0095] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

[0096] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A message interaction method based on an AI customer service robot, characterized in that: include: Classify the received user messages to obtain the messages that need to be replied and the messages that do not need to be replied; Extracting the user's question from the message that needs a reply, and retrieving relevant content from a knowledge base based on the extracted question; Performing context analysis on the message requiring reply, and reordering relevant content retrieved from the knowledge base based on the result of the context analysis; The reordered related content is provided to the AI ​​big model, and the reply content is obtained through the AI ​​big model to respond to the message that needs a reply.

2. The method according to claim 1, characterized in that Categorize received user messages, including: Using the trained NLP classification model to identify the intent of the user message; Based on the result of intent recognition, it is determined whether the user message needs a reply, so as to classify the user message.

3. The method according to claim 1, characterized in that: Extracting the user's question from the message that needs to be replied to includes: Using semantic parsing technology to perform semantic parsing on the message that needs to be replied to, to obtain a parsed semantic structure; Based on the parsed semantic structure, the core content of the message that needs to be replied is extracted to obtain the extracted question.

4. The method according to claim 1, characterized in that: Retrieve relevant content from the knowledge base based on the extracted questions, including: Convert the extracted questions into feature vector representation; Based on the feature vector representation, relevant content similar to the feature vector representation is retrieved from the knowledge base.

5. The method according to claim 4, characterized in that Before retrieving relevant content from the knowledge base based on the extracted question, the method further includes: Acquire multiple knowledge documents, parse the content of each of the multiple knowledge documents, and convert the parsed knowledge document content into a corresponding feature vector representation; The feature vector representation corresponding to the parsed knowledge document content is stored in the knowledge base.

6. The method according to claim 1, characterized in that Performing context analysis on the message requiring reply, and reordering relevant content retrieved from the knowledge base based on the result of the context analysis, including: Performing context analysis on the message to be replied, extracting semantic features and generating corresponding semantic vectors, and calculating similarity between the retrieved relevant content and the semantic vector based on the semantic vector; Set priority rules based on contextual information, user historical preferences, timeliness and importance of content; Based on the similarity calculation result and the priority rule, the retrieved related contents are reordered.

7. A message interaction device based on an AI customer service robot, characterized in that: include: A receiving module is configured to classify received user messages to obtain messages that need to be replied to and messages that do not need to be replied to; A retrieval module is configured to extract the user's question from the message that needs to be replied, and retrieve relevant content from the knowledge base based on the extracted question; An analysis module, configured to perform context analysis on the message requiring a reply, and reorder the relevant content retrieved from the knowledge base based on the result of the context analysis; The reply module is configured to provide the reordered related content to the AI ​​big model, and obtain the reply content through the AI ​​big model to respond to the message that needs a reply.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images