Dialogue processing method and device based on AI large model and storage medium

Abstract

The invention discloses a dialogue processing method and device based on an AI large model and a storage medium. The method comprises the following steps: in response to a received user question, acquiring a role attribute of a virtual object interacting with a user, and dynamically generating a knowledge graph of a field corresponding to the role attribute based on the role attribute; the method comprises the following steps: searching knowledge nodes related to user questions from a plurality of knowledge sources by using a knowledge graph, generating a plurality of potential reply candidates, and performing semantic understanding and context matching on the reply candidates through an AI large model to determine candidate replies most related to the user questions; and performing stylization processing conforming to the language style and the expression mode of the virtual object on the most relevant candidate reply by using a deep learning model, and feeding back the stylized candidate reply to the user as final reply information through the virtual object. According to the method and the device, the technical problem of low accuracy and individuation of virtual object answering 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 dialogue processing method, device and storage medium based on an AI big model. Background Art

[0002] Some virtual object dialogue systems usually rely on preset rules or limited knowledge bases when answering user questions, resulting in insufficient accuracy and flexibility in their answers. When users ask cross-domain or complex questions, the answers of virtual objects often fail to meet user expectations. In addition, the answers of virtual objects usually lack personalization, and it is difficult to adjust the language style and expression according to the specific attributes of different virtual characters, resulting in poor user experience. Therefore, there is an urgent need for a dialogue processing method based on an AI large model, which can dynamically generate a knowledge graph of related fields based on the role attributes of virtual objects, retrieve relevant information from multiple knowledge sources, and stylize the answers through a deep learning model, thereby improving the accuracy and personalization of virtual object answers.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present invention provide a dialogue processing method, device and storage medium based on an AI big model, so as to at least solve the technical problems of low accuracy and personalization of virtual object answers in the prior art.

[0005] According to one aspect of an embodiment of the present invention, a dialogue processing method based on an AI big model is provided, including: in response to receiving a user question, obtaining role attributes of a virtual object interacting with the user, and dynamically generating a knowledge graph corresponding to the field of the role attributes based on the role attributes; using the knowledge graph to retrieve knowledge nodes related to the user question from multiple knowledge sources, generating multiple potential answer candidates, and performing semantic understanding and context matching on the answer candidates through the AI ​​big model to determine the candidate answer that is most relevant to the user question; using a deep learning model to stylize the most relevant candidate answer to match the language style and expression of the virtual object, and feeding back the stylized candidate answer as final answer information to the user through the virtual object.

[0006] According to another aspect of an embodiment of the present invention, a dialogue processing device based on an AI big model is also provided, including: a graph generation module, configured to obtain role attributes of a virtual object interacting with a user in response to receiving a user question, and dynamically generate a knowledge graph corresponding to the field of the role attributes based on the role attributes; a reply generation module, configured to use the knowledge graph to retrieve knowledge nodes related to the user question from multiple knowledge sources, generate multiple potential reply candidates, and perform semantic understanding and context matching on the reply candidates through the AI ​​big model to determine the candidate reply most relevant to the user question; a style processing module, configured to use a deep learning model to stylize the most relevant candidate reply to conform to the language style and expression of the virtual object, and feed back the stylized candidate reply as final reply information to the user through the virtual object.

[0007] In an embodiment of the present invention, in response to receiving a user question, the role attributes of the virtual object interacting with the user are obtained, and a knowledge graph corresponding to the field of the role attributes is dynamically generated based on the role attributes; the knowledge graph is used to retrieve knowledge nodes related to the user question from multiple knowledge sources to generate multiple potential answer candidates, and the answer candidates are semantically understood and context-matched by the AI ​​big model to determine the candidate answer that is most relevant to the user question; the most relevant candidate answer is stylized to conform to the language style and expression of the virtual object using a deep learning model, and the stylized candidate answer is fed back to the user as the final answer information through the virtual object. The above scheme solves the technical problems of low accuracy and personalization of virtual object answers in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] 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:

[0009] Figure 1 is a flow chart of a conversation processing method based on an AI big model according to an embodiment of the present invention;

[0010] Figure 2 is a flowchart of another conversation processing method based on an AI big model according to an embodiment of the present invention;

[0011] Figure 3 is a flow chart of a method for obtaining virtual object role attributes according to an embodiment of the present invention;

[0012] Figure 4is a flow chart of a method for dynamically generating a knowledge graph corresponding to a field of role attributes according to an embodiment of the present invention;

[0013] Figure 5 is a flow chart of a method for generating multiple potential answer candidates according to an embodiment of the present invention;

[0014] Figure 6 is a flow chart of a method for determining a candidate answer most relevant to a user's question according to an embodiment of the present invention;

[0015] Figure 7 is a flow chart of a stylization processing method according to an embodiment of the present invention;

[0016] Figure 8 is a structural schematic diagram of a dialogue processing device based on an AI large model according to an embodiment of the present invention;

[0017] 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

[0018] 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.

[0019] 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.

[0020] According to an embodiment of the present invention, a method embodiment of a dialogue processing method based on an AI big model 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 can be executed in an order different from that shown here.

[0021] Figure 1 is a dialogue processing method based on an AI big model according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0022] Step S102, in response to receiving a user question, obtaining role attributes of a virtual object interacting with the user, and dynamically generating a knowledge graph corresponding to the field of the role attributes based on the role attributes.

[0023] First, the role attributes of the virtual object that interacts with the user are obtained. A virtual object is a virtual character or person that interacts with the user, with predefined role attributes and language style, used to simulate real-life conversations or other interactive scenarios. Role attributes are specific features or identity information of the virtual object, including but not limited to the professional background, identity positioning, and personality characteristics of the virtual object. Role attributes are used to dynamically generate a knowledge graph related to the role during the conversation and guide the stylized processing of the answer.

[0024] Next, a knowledge graph corresponding to the field of the role attributes is dynamically generated based on the role attributes. A knowledge graph is a structured form of knowledge representation, which represents concepts and their relationships through nodes and edges. In the present invention, a knowledge graph is dynamically generated based on the role attributes of a virtual object, covers the knowledge of the field to which the role belongs, and is expanded based on the user's historical interaction information and the current dialogue context, and is used to retrieve relevant information from multiple knowledge sources. In some embodiments, the generation process of the knowledge graph may be: analyzing the role attributes of the virtual object to determine the professional field corresponding to the role attributes; extracting core concepts and relationship nodes related to the professional field from multiple predefined knowledge bases or online data sources, and generating an initial framework of the knowledge graph based on the core concepts and relationship nodes; dynamically expanding the knowledge graph based on the user's historical interaction information and the current dialogue context, so as to expand the relevant content and semantic relationships of the knowledge graph.

[0025] Step S104, using the knowledge graph to retrieve knowledge nodes related to the user's question from multiple knowledge sources, generating multiple potential answer candidates, and performing semantic understanding and context matching on the answer candidates through the AI ​​big model to determine the candidate answer that is most relevant to the user's question.

[0026] First, multiple potential answer candidates are generated. For example, when the length of the user's question is greater than a preset length threshold, the user's question is segmented to obtain multiple sub-question information; keywords and semantic features are extracted from each of the multiple sub-question information to generate sub-parsed data; the sub-parsed data of each sub-question information is fused to generate the preliminary semantic parsing data; based on the semantic parsing data, a node search is performed in the knowledge graph to identify knowledge nodes related to the user's question and generate a candidate node set; the relevance of the candidate node set is calculated, and qualified knowledge nodes are screened out based on a preset relevance threshold to obtain a screened knowledge node set; based on the screened knowledge node set, multiple preliminary answer candidates are constructed, and semantic consistency checks and optimizations are performed on the multiple preliminary answer candidates to obtain the multiple potential answer candidates.

[0027] Next, the most relevant candidate responses are determined. For example, the candidate responses are preliminarily semantically parsed using a natural language processing algorithm to identify the potential semantic relevance between the candidate responses and the user's question to generate semantic matching data; based on the semantic matching data, contextual matching data is generated in combination with the contextual information of the user's question; based on the contextual matching data, the candidate responses are optimized using the deep semantic network of the AI ​​big model to generate optimized response data, and based on the optimized response data, the candidate responses that are most relevant to the user's question are determined.

[0028] Step S106, using a deep learning model to stylize the most relevant candidate response to match the language style and expression of the virtual object, and feeding back the stylized candidate response as final response information to the user through the virtual object.

[0029] For example, based on the most relevant candidate reply, the semantic features of the most relevant candidate reply are analyzed through the deep learning model to generate preliminary style feature data; based on the preliminary style feature data, combined with the language style library of the virtual object, style data matching the style of the virtual object is generated; based on the matching style data, the language style of the most relevant candidate reply is adjusted to generate the stylized candidate reply.

[0030] The present application embodiment provides another dialogue processing method based on the AI ​​big model, such as Figure 2 As shown, the method comprises the following steps:

[0031] Step S202, responding to the user's question and obtaining the virtual object role attributes.

[0032] like Figure 3As shown, the methods for obtaining role attributes include:

[0033] Step S2022, receiving user questions.

[0034] First, the user's question is received and converted into a format suitable for subsequent analysis through preprocessing, including noise removal, word segmentation, and stop word removal.

[0035] Step S2024, identifying role attributes.

[0036] After receiving the user's question, the role attributes of the virtual object that interacts with it are identified. Role attributes can include information such as role name, gender, professional field, personality characteristics, etc., to provide basic data for subsequent generation of knowledge graphs.

[0037] Step S204, dynamically generate a knowledge graph corresponding to the domain of the role attributes.

[0038] like Figure 4 As shown, the following steps are included:

[0039] Step S2042, analyzing role attributes.

[0040] Obtain the role attributes of the virtual object and determine the corresponding professional field, or directly obtain the professional field of the virtual object. For example, if the role is a medical guide, the professional field is "medical".

[0041] Step S2044, extracting core concepts and relationships.

[0042] The system accesses multiple predefined knowledge bases and online data sources. These knowledge bases can include publicly available databases, industry-specific documents, structured data tables containing professional knowledge, and local professional databases. Online data sources can include real-time updated industry reports, academic papers, and other relevant Internet resources. Each knowledge base and data source will be mapped to a unified data interface within the system to ensure consistency and accessibility of the data structure.

[0043] The system regularly crawls the latest field-related data from online data sources through web crawlers or API interfaces. At the same time, the system regularly updates the content in the predefined knowledge base to ensure the timeliness of its content. After the data is crawled, preprocessing operations are performed.

[0044] After the preprocessing is completed, the captured data is analyzed by a natural language processing model based on deep learning. First, the input data is matched with the key terms and corpus in the field to identify the core concepts related to the professional field. This embodiment uses a word vector model based on domain knowledge to ensure that the system can capture the contextual relationships and semantic associations between concepts. After extracting the core concepts, the relationship between the concepts is further analyzed. By using a knowledge graph, the present application can associate the identified core concepts with relationship nodes. Specifically, the present application uses a graph neural network to model the relationship between concepts, construct a concept graph, and identify key relationship nodes. These relationship nodes can include "causal relationships", "subordinate relationships", and "dependencies", etc.

[0045] After extracting core concepts and identifying relationship nodes, a knowledge graph related to the professional field is generated. The knowledge graph consists of concept nodes and relationship nodes, which intuitively shows the relationship between the extracted concepts. The knowledge graph can not only serve as a visualization tool for domain knowledge, but also serve as the basis for reasoning and decision-making in subsequent intelligent applications. To ensure that the system's knowledge base and data sources are always up to date, the system performs regular automatic dynamic updates. New data is continuously added to the knowledge base and processed through the aforementioned preprocessing, concept extraction, and relationship node construction steps. In order to improve the accuracy of the system, the system will optimize its concept extraction and relationship construction algorithms based on a feedback mechanism, so that the extracted core concepts and relationships are more in line with the actual situation in the professional field.

[0046] Step S2046, dynamically expand the knowledge graph.

[0047] Combine the user's historical interaction information and the current conversation context to dynamically expand the knowledge graph to ensure that the graph contains content and semantic relationships related to the current conversation. Monitor and record the historical data generated by the user's interaction with the virtual object in real time. This interaction information includes the user's conversation content, question type, topic of interest, and the response results previously generated by the system. In the current conversation context, the user's input is processed through semantic analysis technology. This processing step includes the application of the natural language understanding module to identify the user's current questions or core topics of interest, and generate a dynamic conversation context model based on the context information of the current conversation.

[0048] In order to ensure the continuity of the conversation, the current conversation context is associated with the historical interaction information to identify related topics or concepts that the user may have discussed before. On this basis, the existing knowledge graph is dynamically expanded in combination with the user's historical interaction information and the current conversation context. Specifically, new concepts and relationship nodes related to the current conversation are retrieved from a predefined knowledge base or online data source, and these new concepts are associated with the nodes in the existing graph through the knowledge graph expansion module. During the expansion process, a graph neural network based on deep learning is used to analyze the semantic relationship between the newly added nodes and the existing nodes to ensure that the expanded knowledge graph maintains semantic consistency and logic. After the expansion is completed, the relevant content and semantic relationship of the knowledge graph are further optimized to ensure that it can accurately reflect the user's interests and conversation needs. Through the dynamic expansion method, the present invention can continuously update the knowledge graph according to the user's historical interaction information and the current conversation context, ensuring that the knowledge graph always maintains a high degree of relevance to the user's needs, thereby achieving more intelligent and personalized conversation interaction.

[0049] Step S206, retrieving knowledge nodes related to the user's question and generating multiple potential answer candidates.

[0050] like Figure 5 As shown, the method comprises the following steps:

[0051] Step S2062: keyword extraction.

[0052] Perform natural language processing analysis on user questions to extract keywords and semantic features related to the user questions. Pre-trained language models can be used to capture the deep semantic information of user questions to ensure that the complex meaning behind user intentions can be understood. The extracted keywords can include specific objects, behaviors, attributes, and other words, while semantic features help the system better identify the context and implicit information of the questions. Based on these keywords and semantic features, preliminary semantic parsing data is generated.

[0053] The training process of the language model is as follows: 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 through the language model. For the speech recognition text, the model generates its feature representation, namely the primary embedding vector group, which captures the characteristics of the speech text in the multidimensional semantic space. Similarly, for the manual recognition text, the model generates an auxiliary embedding vector group to characterize its semantic information. The present application effectively enhances the model's ability to capture different types of text features by extracting the embedding vector group of the speech recognition text and the manual recognition text, and performing multi-dimensional order-up extension. Then, according to the preset extension rank criterion, the primary embedding vector group and the auxiliary embedding vector group are respectively multi-dimensionally upgraded. Subsequently, a multi-channel convolution kernel is used to couple and fuse the features of the vector group to obtain the coupled primary embedding group and the auxiliary embedding group. The present application uses a multi-channel convolution kernel for feature coupling and fusion processing, so that text features from different sources can be efficiently combined, thereby improving the semantic understanding depth of the model. Next, a nonlinear activation function is applied to obtain the activated primary embedding group and the auxiliary embedding group. Finally, the activated primary embedding group and auxiliary embedding group are normalized to ensure the scale consistency of the data. The nonlinear activation function further optimizes the expressiveness of the embedding vector, while the normalization operation ensures the scale consistency of the feature data at each level, improves the training stability and generalization ability of the model, and thus improves the overall language recognition accuracy.

[0054] Subsequently, the primary feature matrix and the auxiliary feature matrix are respectively subjected to high-order matrix decomposition. First, these matrices are expanded along the direction of the extended dimension to form multiple sub-matrices. Each sub-matrix is ​​subjected to singular value decomposition to extract the core subspace representation and the factor matrix. The present application effectively extracts the key information in the matrix by performing high-order matrix decomposition on the primary feature matrix and the auxiliary feature matrix, and retains its local structural characteristics through singular value decomposition. The core subspace representation condenses the key information of the matrix, and the factor matrix captures the local structure of its features. Next, the principal component difference between the primary feature matrix and the auxiliary feature matrix is ​​calculated. The core subspace representation condenses the important information, and the factor matrix captures the subtle structural changes of the features. By measuring the similarity of the core subspace representation and the difference of the principal components in the factor matrix, the subtle changes between the two sets of data can be accurately captured. The principal component difference is further calculated, and the subtle differences between the speech recognition text and the manual recognition text can be accurately identified through similarity analysis and principal component difference measurement. Finally, the loss function is updated based on the difference between the two to obtain a trained language model. This helps optimize the loss function, thereby improving the model's training results and the final language recognition accuracy.

[0055] Step S2064, node search.

[0056] Perform node search operations in the knowledge graph. Use keywords and semantic features as search parameters to find content related to the user's question by matching existing knowledge nodes. For example, use a graph traversal algorithm or a query algorithm based on a graph database, such as the SPARQL query language, to identify and extract a set of nodes that are closely related to the user's question. Preferably, nodes directly related to the core concept of the question are searched first, but nodes indirectly related to the core concept can also be searched through relationship chains to ensure the breadth and completeness of the search results.

[0057] Step S2066, correlation calculation.

[0058] After generating a preliminary set of candidate nodes, the relevance between the nodes and the user's question is calculated. The relevance is calculated based on the semantic similarity of the nodes, the contextual relevance, and the importance of the nodes in the graph. After that, a relevance score is assigned to each candidate node based on the relevance index and compared with the preset relevance threshold. The knowledge nodes with scores exceeding the threshold are selected as the final set of candidate nodes to ensure the accuracy and relevance of the answer.

[0059] Step S2068, construct preliminary response candidates.

[0060] Based on the screened knowledge node set, preliminary answer candidates are constructed. A short sentence or keyword phrase is generated based on the content of the knowledge node as a preliminary version of the potential answer. For example, it can be based on template matching, sentence reconstruction and other technologies, or it can be generated by using a generative language model (such as GPT, etc.) for natural language generation. For each candidate knowledge node, multiple answer options are generated based on its contextual relationship to ensure that different aspects of the user's question are covered.

[0061] Step S208, determining the candidate answer that is most relevant to the user's question.

[0062] like Figure 6 As shown, the method comprises the following steps:

[0063] Step S2082, preliminary semantic analysis.

[0064] First, a preliminary semantic analysis is performed on the candidate answers. By using natural language processing technology and semantic analysis models, the key semantic elements in the candidate answers are identified and compared with the semantic content of the user's question. By detecting potential semantic relevance, a set of semantic matching data is generated to reflect the degree of relevance between the candidate answers and the user's question at the semantic level.

[0065] Step S2084, context information fusion.

[0066] In order to ensure that the candidate answer is not only relevant to the user's question itself, but also consistent with the context of the entire conversation, the context information of the user's question is combined with the preliminary semantic analysis results. By analyzing the user's previous questions and conversation history, context matching data is generated to reflect the adaptability of the candidate answer in the current conversation context. The context fusion step helps to improve the coherence of the answer so that it can flow smoothly with the entire conversation.

[0067] Step S2086, deep semantic network optimization.

[0068] The candidate responses are further optimized through the deep semantic network of the AI ​​big model. Based on the context matching data, the deep semantic network comprehensively optimizes the semantics, logic and language expression of the response, ensuring that the final response is not only the most semantically consistent with the user's question, but also achieves the expected effect in terms of language fluency and logical consistency.

[0069] Step S210: stylization processing and final feedback.

[0070] like Figure 7 As shown, the method comprises the following steps:

[0071] Step S2102: style feature analysis.

[0072] First, the style features of the candidate responses are analyzed. Through the deep learning model, the semantic features, language structure and expression of the candidate responses are identified. Combined with the corpus of virtual objects or the previously trained style model, the potential style features in the response are analyzed, such as sentence complexity, emotional color, and formality of expression. The style feature matrix composed of style features is subjected to dimensionality reduction processing. Specifically, the style feature matrix is ​​obtained, and the high-dimensional space similarity between each pair of feature points in the feature matrix is ​​calculated in the high-dimensional space using the Gaussian kernel function, where the Gaussian kernel function is a function used to measure the similarity between two data points in the high-dimensional space. The high-dimensional space similarity can be obtained based on the local density weight, the number of Gaussian kernel functions, the weight of the Gaussian kernel function, and the bandwidth parameter, where the local density weight reflects the density of the feature points in their local neighborhood, and the bandwidth parameter is used to control the scale and influence range of the Gaussian kernel function. Then, each feature point in the high-dimensional space corresponding to the feature matrix is ​​randomly mapped to the low-dimensional space; the low-dimensional space similarity between each pair of feature points in the feature matrix is ​​calculated in the low-dimensional space using T distribution, wherein T distribution is a probability distribution model suitable for small sample data and is used to deal with data dimensionality reduction problems. Finally, based on the high-dimensional space similarity and the low-dimensional space similarity, the position of each feature point in the low-dimensional space is adjusted to reduce the dimension of the feature matrix. For example, for each pair of feature point pairs in the high-dimensional space similarity matrix and the low-dimensional space similarity matrix, the difference between the feature point pairs is calculated to obtain a difference matrix; based on the difference matrix, the difference between the feature point pairs is quantified using a divergence function to generate a divergence value, wherein the divergence function is used to measure the difference between two data distributions. Based on the divergence value, a divergence matrix is ​​generated, wherein the divergence matrix is ​​used to represent the position change of each feature point in the high-dimensional space and the low-dimensional space. Afterwards, the divergence matrix is ​​used to construct an objective function that represents the difference between the similarity in high-dimensional space and the similarity in low-dimensional space. By minimizing the objective function, the position of each feature point in the low-dimensional space is adjusted to obtain the adjusted feature point position as the feature matrix after dimensionality reduction.

[0073] Step S2104, generating style data.

[0074] On the basis of style feature analysis, the virtual object's language style library is further combined to generate style data that matches the virtual object's style. The language style library contains the virtual object's personalized language templates and specific expression habits, such as word choice, tone adjustment, grammatical structure, etc. By identifying the virtual object's personality characteristics, such as whether it is funny, formal or friendly, as well as common expressions and language habits, this style information is integrated into the candidate replies to ensure that the generated replies are consistent with the virtual object's image and language style.

[0075] Step S2106, adjusting the language style.

[0076] Based on the generated style data, the candidate responses are finely adjusted in terms of language style. The semantic content of the responses is modified and reprocessed through the language model without changing the meaning, ensuring that the language expression matches the expected style of the virtual object. For example, the response can be made more consistent with the personality characteristics of the virtual object by adjusting the sentence structure, modal particles, and expression habits, thereby generating stylized candidate responses.

[0077] Step S2108: Feedback to the user.

[0078] After completing the language style adjustment, the stylized candidate response is fed back to the user as the final response through a virtual object.

[0079] The present application also provides a dialogue processing device based on an AI large model, such as Figure 8 As shown, it includes: a graph generation module 82, which is configured to obtain the role attributes of the virtual object interacting with the user in response to receiving a user question, and dynamically generate a knowledge graph corresponding to the field of the role attributes based on the role attributes; a reply generation module 84, which is configured to use the knowledge graph to retrieve knowledge nodes related to the user question from multiple knowledge sources, generate multiple potential reply candidates, and perform semantic understanding and context matching on the reply candidates through the AI ​​big model to determine the candidate reply that is most relevant to the user question; a style processing module 86, which is configured to use a deep learning model to stylize the most relevant candidate reply to conform to the language style and expression of the virtual object, and feed back the stylized candidate reply as the final reply information to the user through the virtual object.

[0080] It should be noted that the dialogue processing device based on the AI ​​big model provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the dialogue processing device based on the AI ​​big model provided in the above embodiment and the dialogue processing method embodiment based on the AI ​​big model belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0081] 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.

[0082] like Fig. 9As 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.

[0083] 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.

[0084] The above is only a preferred implementation 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 dialogue processing method based on an AI large model, characterized in that: include: In response to receiving a user question, obtaining a role attribute of a virtual object interacting with the user, and dynamically generating a knowledge graph of a field corresponding to the role attribute based on the role attribute; The knowledge graph is used to retrieve knowledge nodes related to the user's question from multiple knowledge sources to generate multiple potential answer candidates, and the answer candidates are semantically understood and context-matched by the AI ​​big model to determine the candidate answer that is most relevant to the user's question; The most relevant candidate responses are stylized using a deep learning model to match the language style and expression of the virtual object, and the stylized candidate responses are fed back to the user as final response information through the virtual object.

2. The method according to claim 1, characterized in that Dynamically generating a knowledge graph corresponding to a field of the role attribute based on the role attribute includes: Analyzing the role attributes of the virtual object to determine the professional field corresponding to the role attributes; Extracting core concepts and relationship nodes related to the professional field from a plurality of predefined knowledge bases or online data sources, and generating an initial framework of the knowledge graph based on the core concepts and relationship nodes; The knowledge graph is dynamically expanded based on the user's historical interaction information and the current conversation context, so as to expand the relevant content and semantic relationships of the knowledge graph.

3. The method according to claim 1, characterized in that The knowledge graph is used to retrieve knowledge nodes related to the user's question from multiple knowledge sources to generate multiple potential answer candidates, including: Extracting keywords and semantic features from the user's question, and generating preliminary semantic parsing data based on the keywords and semantic features; Based on the semantic parsing data, a node search is performed in the knowledge graph to identify knowledge nodes related to the user's question and generate a candidate node set; Calculating the relevance of the candidate node set, filtering out qualified knowledge nodes based on a preset relevance threshold, and obtaining a filtered knowledge node set; Based on the screened knowledge node set, a plurality of preliminary answer candidates are constructed, and semantic consistency check and optimization are performed on the plurality of preliminary answer candidates to obtain the plurality of potential answer candidates.

4. The method according to claim 3, characterized in that Extracting keywords and semantic features from the user's question, and generating preliminary semantic parsing data based on the keywords and semantic features, including: When the length of the user's question is greater than a preset length threshold, the user's question is segmented to obtain a plurality of sub-question information; Extracting keywords and semantic features from each of the plurality of sub-question information to generate sub-parsed data; The sub-parsed data of each sub-question information is integrated to generate the preliminary semantic parsed data.

5. The method according to claim 1, characterized in that The AI ​​big model performs semantic understanding and context matching on the answer candidates to determine the candidate answer that is most relevant to the user's question, including: Performing preliminary semantic analysis on the answer candidate by using a natural language processing algorithm to identify potential semantic relevance between the answer candidate and the user's question, so as to generate semantic matching data; Generate context matching data based on the semantic matching data and in combination with context information of the user's question; Based on the context matching data, the answer candidates are optimized through the deep semantic network of the AI ​​big model to generate optimized answer data, and based on the optimized answer data, the candidate answer most relevant to the user's question is determined.

6. The method according to claim 1, characterized in that The deep learning model is used to perform stylized processing on the most relevant candidate answer to match the language style and expression of the virtual object, including: Based on the most relevant candidate answers, analyzing semantic features of the most relevant candidate answers through the deep learning model to generate preliminary style feature data; Based on the preliminary style feature data and in combination with the language style library of the virtual object, generating style data matching the style of the virtual object; Based on the matching style data, the language style of the most relevant candidate response is adjusted to generate the stylized candidate response.

7. A dialogue processing device based on an AI large model, characterized in that: include: A graph generation module is configured to obtain, in response to receiving a user question, a role attribute of a virtual object interacting with the user, and dynamically generate a knowledge graph of a field corresponding to the role attribute based on the role attribute; A reply generation module is configured to use the knowledge graph to retrieve knowledge nodes related to the user's question from multiple knowledge sources, generate multiple potential reply candidates, and perform semantic understanding and context matching on the reply candidates through the AI ​​big model to determine the candidate reply most relevant to the user's question; The style processing module is configured to use a deep learning model to stylize the most relevant candidate response to conform to the language style and expression of the virtual object, and feed back the stylized candidate response as final response information to the user through the virtual object.

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