AI interactive digital human dialogue method
Patent Information
- Application Number
- CN202611114993.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-10-30
AI Technical Summary
[0003]本发明的主要目的是提出一种AI交互式数字人对话方法，旨在解决如何准确且快速的进行对话信息的检索与匹配并输出结果的技术问题
[0042]本发明的上述技术方案中，该AI交互式数字人对话方法包括以下步骤：创建AI数字人，并设置AI数字人的初始参数；获取标准文件，将标准文件进行分类整合形成标准化知识库，并对标准化知识库进行向量化处理，与AI数字人进行绑定；基于所述标准化知识库进行用户提问文本的相似性检索，并提取关键信息，生成结合层级上下文的标准回答。本发明解决了如何准确且快速的进行对话信息的检索与匹配并输出结果的技术问题。
Smart Images
Figure CN123012103A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent interaction technology, and in particular to an AI-interactive digital human dialogue method. Background Technology
[0002] Interactive digital human dialogue is a form of human-computer interaction based on artificial intelligence technology. It achieves real-time two-way communication with users by simulating human speech and dialogue logic. A specific knowledge base forms the basis for dialogue with the digital human, while answer retrieval is a prerequisite. Before the answer retrieval process, information related to the question is collected, and then matched with this information. Finally, the answer is output based on the matching results. However, if the information collected before matching the question with the answer is flawed or incomplete, errors or omissions may occur during information matching, adversely affecting the accuracy of the output answer. Therefore, there is an urgent need to propose an AI-based interactive digital human dialogue method to solve the technical problem of how to accurately and quickly retrieve and match dialogue information and output results. Summary of the Invention
[0003] The main objective of this invention is to propose an AI-interactive digital human dialogue method, which aims to solve the technical problem of how to accurately and quickly retrieve and match dialogue information and output results.
[0004] To achieve the above objectives, the present invention provides an AI interactive digital human dialogue method, wherein the AI interactive digital human dialogue method includes the following steps:
[0005] S1. Create an AI digital human and set its initial parameters;
[0006] S2. Obtain standard documents, classify and integrate the standard documents to form a standardized knowledge base, and vectorize the standardized knowledge base to bind it with the AI digital human;
[0007] S3. Based on the standardized knowledge base, perform similarity retrieval of user question texts, extract key information, and generate standard answers that combine hierarchical context.
[0008] One preferred embodiment, step S1, specifically comprises:
[0009] Based on the interaction of the target user, set the initial parameters of the AI digital human; the initial parameters include the AI digital human's name, avatar, and prompt word template;
[0010] Identify the application areas of AI digital humans and define their interaction modes and functional boundaries; the interaction modes include voice, text, or multimodal.
[0011] In one preferred embodiment, step S1 packages the initial parameters using the CreateAgent function and constructs an interactive AI digital human; specifically:
[0012]
[0013] in, To build an interactive AI digital human, For name, As the profile picture This is a prompt word.
[0014] One preferred embodiment, step S2, specifically comprises:
[0015] S21. Based on the application areas of AI digital humans, acquire and match standard documents, and classify and integrate the matched standard documents to form a standardized knowledge base;
[0016] S22. Perform word segmentation on the text in the standardized knowledge base;
[0017] S23. Perform vector transformation on the standard knowledge base to generate dense vectors and sparse vectors, and associate and store the dense vectors and sparse vectors with the original text to construct a searchable multimodal vector standardized knowledge base.
[0018] S24. Call the Bind function to bind the multimodal vector standardized knowledge base to the AI digital human.
[0019] One preferred embodiment is that step S22 involves word segmentation of the text in the standardized knowledge base, specifically as follows:
[0020] The text in the standardized knowledge base is segmented into Chinese words using a text splitter. The default segmentation parameters are: chunk size=512, and chunk overlap=20 for each segment. Large files are split into parent and child segments.
[0021] One preferred embodiment, step S3, specifically includes:
[0022] S31. Generate a query vector, receive user query text, and encode the user query text in parallel into dense vector and sparse vector;
[0023] S32. Vector retrieval: Input the dense vector and sparse vector into a standardized knowledge base, perform hybrid similarity retrieval, and obtain a retrieval structure set; the retrieval structure set includes subtext content, similarity distance, and hierarchical index identifier;
[0024] S33. Extract the parent text content corresponding to the sub-text content based on the hierarchical index identifier;
[0025] S34. A standard answer generated by AI digital human combined with hierarchical context.
[0026] In one preferred embodiment, in step S31, the user's question text is encoded in parallel into dense vectors and sparse vectors using the BGE-M3 model.
[0027] In one preferred embodiment, step S32 inputs the dense vector and sparse vector into a standardized knowledge base and calls the hybrid_search interface to perform hybrid similarity retrieval; the score of the similarity retrieval is:
[0028]
[0029] in, The score for similarity retrieval. For dynamic weighting coefficients, , The dense vector of the user's question text. This is a dense vector for retrieving documents from a standardized knowledge base. The coefficient representation of the user query, i.e., the term set. This is the inverse document frequency calculation formula. For terms In the current document word frequency in To adjust the parameters.
[0030] In one preferred embodiment, the hierarchical index identifier includes the parent text UUID and the paragraph position offset;
[0031] Based on the parent text UUID, match the parent text file; extract the complete paragraph containing related subtext based on the paragraph offset; the following is:
[0032]
[0033] in, This is the offset of the parent text. This is the offset of the subtext. To preset the paragraph length threshold, This is the context window compensation coefficient.
[0034] In one preferred embodiment, after step S3, the method further includes:
[0035] S4. Render the structured question and answer content into a standardized document output based on the custom user layout and / or template-driven approach;
[0036] The structure of the template is extracted as follows:
[0037]
[0038]
[0039] The standardized document output is as follows:
[0040]
[0041] in, Extracting structure for calling template The structure extracted from the template file, For the uploaded template, For custom user layout rules, This is the final set of merged typesetting rules. A collection of user-generated question texts. To standardize document output.
[0042] The AI interactive digital human dialogue method described above includes the following steps: creating an AI digital human and setting its initial parameters; acquiring standard files, classifying and integrating these files to form a standardized knowledge base, vectorizing the standardized knowledge base, and binding it to the AI digital human; performing similarity retrieval of user question texts based on the standardized knowledge base, extracting key information, and generating a standard answer that combines hierarchical context. This invention solves the technical problem of how to accurately and quickly retrieve and match dialogue information and output results.
[0043] In this invention, by creating an AI digital human and setting initial parameters, standard documents are classified and integrated to form a standardized knowledge base that can be used for intelligent answers. This knowledge base is then bound to the AI digital human to improve the quality and reliability of the answers. Based on the rapid similarity retrieval of the standard knowledge base, the vector most similar to the target Xiangxiang is found, and key information is extracted. This avoids the problem of inaccurate or incomplete output answers due to mismatched information, thus forming the most suitable answer.
[0044] In this invention, templates can be created to enable flexible layout of user-defined styles, thereby standardizing document output and meeting the needs of multiple scenarios such as exporting, archiving, and delivery. This helps to solve resource and cost issues while ensuring the accuracy and efficiency of answer output, thus enabling the integration of the pending knowledge base into the AI interactive digital human dialogue. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0046] Figure 1 This is a schematic diagram of an AI interactive digital human dialogue method according to an embodiment of the present invention.
[0047] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] It should be noted that all directional indicators (such as up, down, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.
[0050] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.
[0051] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0052] See Figure 1 According to one aspect of the present invention, the present invention provides an AI interactive digital human dialogue method, wherein the AI interactive digital human dialogue method includes the following steps:
[0053] S1. Create an AI digital human and set its initial parameters;
[0054] S2. Obtain standard documents, classify and integrate the standard documents to form a standardized knowledge base, and vectorize the standardized knowledge base to bind it with the AI digital human;
[0055] S3. Based on the standardized knowledge base, perform similarity retrieval of user question texts, extract key information, and generate standard answers that combine hierarchical context.
[0056] Specifically, in this embodiment, step S1 is as follows:
[0057] Based on the interaction of the target user, set the initial parameters of the AI digital human; the initial parameters include the AI digital human's name, avatar, and prompt word template; design a behavior pattern that meets the needs of the scenario.
[0058] The application areas of the AI digital human are determined, and the interaction modes and functional boundaries of the AI digital human are defined. The interaction modes include voice, text, or multimodal. The application areas include customer service, education, medical care, etc. This invention does not impose specific limitations and can be set as needed. The functional boundaries specifically include whether multilingual support, emotion recognition, etc. are required. This invention does not impose specific limitations and can be set as needed. Step S1 constructs a virtual character with human language and interaction capabilities and can simulate real human characteristics, forming an AI digital human with identity, image, and semantic guidance capabilities.
[0059] Specifically, in this embodiment, step S1 packages the initial parameters using the CreateAgent function and constructs an interactive AI digital human; specifically:
[0060]
[0061] in, To build an interactive AI digital human, For name, As the profile picture This is a prompt word.
[0062] Specifically, in this embodiment, step S2 is as follows:
[0063] S21. Based on the application areas of the AI digital human, acquire and match standard documents, and classify and integrate the matched standard documents to form a standardized knowledge base. Depending on the specific application area set up in the AI digital human, relevant standard documents can be collected from national or local standard platforms or the company's internal knowledge base. For example, engineering project specifications and standards include bidding documents, contract terms, construction project standards and quotas, etc. Different standard documents can be imported depending on the application area. The matched documents are then screened to determine those that conform to the specific application area, and these standard documents are classified and integrated to form a standardized knowledge base. The application area of the AI digital human refers to the virtual human generated using artificial intelligence technology. The scenarios in which characters perform tasks or provide services in an anthropomorphic manner in different fields are pre-defined and usually categorized according to different industry sectors, such as enterprise services, education, healthcare, financial consulting, or legal Q&A. These are implemented using conventional technologies, which will not be specifically elaborated upon in this invention. In knowledge-based question answering, professional knowledge-based question answering typically involves the process of efficiently and accurately acquiring the required knowledge from massive amounts of professional data. Through the acquisition of the required knowledge, the answers to user questions can be determined, and based on this, they can be summarized, analyzed, and output. The standardized knowledge base includes the storage of high-dimensional vector data. In the standardized knowledge base, storage specifically refers to the persistent preservation and management mechanism of high-dimensional vector data and related metadata.
[0064] S22. Perform word segmentation on the text in the standardized knowledge base; specifically:
[0065] The Recursive Character Text Splitter is used to perform Chinese word segmentation on text in a standardized knowledge base. The default segmentation parameters are: chunk_size=512, and chunk_overlap=20 for repeated content in each segment. The value of chunk_size can be adjusted according to the actual situation. A reasonable value can improve retrieval efficiency and reduce storage and computing costs. Large files are segmented using a parent-child splitting method, with child texts segmented in a normal mode and the content of the parent text retained. A recursive segmentation algorithm is used to perform multi-level semantic segmentation on the input text, generating a hierarchical text structure containing parent-child relationships. Child text paragraphs are overlapped within a set length range, while parent text retains its complete upper and lower bounds. Text information; the recursive segmentation algorithm adopts: a preset segmentation length threshold of 512 character units; an overlap area of 20 character units between adjacent text paragraphs; and hierarchical semantic preservation of long texts through recursive iteration; for example: assuming there is a long text (e.g., a 2000-word Chinese article), parent segmentation: the entire article is divided into larger blocks (e.g., each block is about 1000 characters) by paragraphs or chapters; child segmentation: each parent block is further recursively segmented according to chunk_size=512 and chunk_overlap=20: first child block: characters [0-511] (512 characters in total), second child block: starting from character 512-20=492, covering [492-1003] (512 characters in total, of which the first 20 characters overlap with the first block), repeating until the end of the parent block; parent block preservation: each child block is stored in association with the original content of the parent block to ensure context integrity;
[0066] S23. Perform vector transformation on the standard knowledge base to generate dense vectors and sparse vectors, and associate and store the dense vectors and sparse vectors with the original text to construct a searchable multimodal vector standardized knowledge base; perform vector transformation using the encoding method of the BGE-M3 model, generating dense vectors and sparse vectors simultaneously during the transformation, and use a multi-encoding fusion model to vectorize text paragraphs, simultaneously generating dimension-compressed dense vectors and feature-explicitly expressed sparse vectors; wherein, dense vectors are used for semantic similarity calculation, and sparse vectors are used for keyword matching retrieval; the multi-encoding fusion model adopts a dual-channel encoding architecture based on an attention mechanism; generates high-dimensional semantic representations of more than 768 dimensions through the dense vector channel; generates interpretable feature weight distributions through the sparse vector channel; for example: input text: encode each sub-block, dense vector: generated through the multi-layer attention mechanism of the multi-encoding fusion model, for example, a 768-dimensional floating-point array: [0. 23,−0.45,1.2,…](Dimension=768); Sparse vectors: generated by algorithms such as term frequency or BM25, for example: {(1234:0.8,5678:0.5,…)} (index-value pairs with non-zero weights); Dense vector dimensions: higher dimensions (e.g., 1024 dimensions) indicate stronger representation capabilities, but higher storage and computation costs; lower dimensions (e.g., 256 dimensions) indicate higher efficiency, but may lose semantic details; Sparse vector weights: high sparsity (few non-zero values) indicates fast retrieval speed, but may overlook… Slightly low-frequency important words; low sparsity (more non-zero values) indicates that more information is retained, but the computational complexity increases; the dense vector representation, sparse vector representation and corresponding original text are associated and stored to construct a searchable multimodal vector database; a storage structure containing a triple mapping relationship is constructed: a joint index of dense vectors and sparse vectors; a bidirectional pointer association between vector data and original text paragraphs; a hierarchical mapping between child paragraphs and parent contexts; the storage logic is to use dense vectors for similarity search and sparse vectors for fast keyword filtering.
[0067] S24. Call the Bind function to bind the multimodal vector standardized knowledge base to the AI digital human; the bound AI digital human is:
[0068]
[0069] in, This refers to the collection of documents uploaded from the standardized knowledge base, including PDFs, Word documents, and text files. For AI digital humans that are not bound to a system, To bind an AI digital human with a multimodal standardized knowledge base, this AI digital human possesses a specific knowledge context.
[0070] Specifically, in this embodiment, step S3 is as follows:
[0071] S31. Generate a query vector, receive user query text, and encode the user query text in parallel into dense vector and sparse vector;
[0072] S32. Vector retrieval: Input the dense vector and sparse vector into a standardized knowledge base, perform hybrid similarity retrieval, and obtain a retrieval structure set; the retrieval structure set includes subtext content, similarity distance, and hierarchical index identifier;
[0073] S33. Extract the parent text content corresponding to the sub-text content based on the hierarchical index identifier;
[0074] S34. A standard answer generated by AI digital human combined with hierarchical context.
[0075] Specifically, in this embodiment, in step S31, the user's question text is encoded in parallel into dense vectors and sparse vectors using the BGE-M3 model.
[0076] Specifically, in this embodiment, step S32 inputs the dense vector and sparse vector into a standardized knowledge base and calls the hybrid_search interface to perform hybrid similarity retrieval; the similarity retrieval score is:
[0077]
[0078] in, The score for similarity retrieval. For dynamic weighting coefficients, , The dense vector of the user's question text. This is a dense vector for retrieving documents from a standardized knowledge base. The coefficient representation of the user query, i.e., the term set. This is the inverse document frequency calculation formula. For terms In the current document word frequency in To adjust the parameters, The length of the current document.
[0079] Specifically, in this embodiment, the hierarchical index identifier includes the parent text UUID and the paragraph position offset;
[0080] Based on the parent text UUID, match the parent text file; extract the complete paragraph containing related subtext based on the paragraph offset; the following is:
[0081]
[0082] in, This is the offset of the parent text. This is the offset of the subtext. To preset the paragraph length threshold, This is the context window compensation coefficient.
[0083] Specifically, in this embodiment, when the construction unit encounters problems during the construction of asphalt pavement on a highway, the user asks through the AI interactive digital human dialogue method described in this invention: "How to control the temperature of the mixture and prevent rainwater penetration when paving asphalt pavement during the rainy season?" Bimodal encoding: Input text: User question text (length L=28 words).
[0084] BGE-M3 model output: Dense vector: (L2 norm = 14.3); Sparse vector: contains keywords such as "rainy season", "asphalt paving", "temperature control", and "rainwater infiltration", with 10 dimensions. 5 ;
[0085] Hybrid Search: Call the `hybrid_search` interface with the following parameters: =14.3 / 14.3+8.2=0.64, =1- =0.36, because the norm of dense vectors is high, focusing on semantic matching;
[0086] Search formula: BM25 parameters , ;in, This represents the coefficients of the documents, i.e., word segmentation of all documents in the knowledge base. This represents the coefficients used in the user query, i.e., the set of terms.
[0087] Search results: Subtext fragment (Top-1): "During the rainy season, the temperature of the mixture at the factory should be ≥160℃, and a waterproof tarpaulin should be covered immediately after paving", Distance: 0.78 Hierarchical index: (JTG_F40-2023, offset=2456);
[0088] Based on subtext offset Locate the parent text, where... , Specifically:
[0089]
[0090] The parent text content is from the "Special Climate Construction Requirements" section of the "Technical Specification for Construction of Asphalt Pavement on Highways JTG_F40-2023": During the rainy season, vehicles transporting asphalt mixtures must be equipped with insulation layers, and the temperature upon arrival at the site must not be lower than 155℃; mobile rain shelters should be provided at the paving site, and compaction should be completed within 30 minutes after paving; in case of moderate to heavy rain, construction should be stopped immediately, and the paved sections should be waterproofed.
[0091] AI-generated digital human response: According to the "JTG_F40-2023" regulations, the following measures must be taken for asphalt paving during the rainy season: Temperature control: The temperature of the mixture leaving the factory must be ≥160℃, and the temperature upon arrival at the site must be ≥155℃ (using refrigerated transport vehicles); the paver must be preheated to above 130℃ to reduce temperature loss; Rain protection measures: immediately cover the paved area with waterproof tarpaulin, and complete compaction within 30 minutes; mobile rain shelters must be provided on site, and real-time rain protection must be provided by monitoring radar images; construction must be suspended during moderate to heavy rain, and paved sections must be covered with waterproof geotextile.
[0092] Specifically, in this embodiment, after step S3, the method further includes:
[0093] S4. Render the structured question and answer content into a standardized document output based on the custom user layout and / or template-driven approach;
[0094] The structure of the template is extracted as follows:
[0095]
[0096]
[0097] The standardized document output is as follows:
[0098]
[0099] in, Extracting structure for calling template The structure extracted from the template file, For the uploaded template, Customizable user-defined formatting rules, including font, paragraph locking, heading formatting, etc. This is the final set of merged typesetting rules. A collection of user-generated question texts. To standardize document output.
[0100] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.
Claims
1. An AI-interactive digital human dialogue method, characterized in that, Includes the following steps: S1. Create an AI digital human and set its initial parameters; S2. Obtain standard documents, classify and integrate the standard documents to form a standardized knowledge base, and vectorize the standardized knowledge base to bind it with the AI digital human; S3. Based on the standardized knowledge base, perform similarity retrieval of user question texts, extract key information, and generate standard answers that combine hierarchical context.
2. The AI interactive digital human dialogue method according to claim 1, characterized in that, Step S1 specifically includes: Based on the interaction of the target user, set the initial parameters of the AI digital human; the initial parameters include the AI digital human's name, avatar, and prompt word template; Identify the application areas of AI digital humans and define their interaction modes and functional boundaries; the interaction modes include voice, text, or multimodal.
3. The AI interactive digital human dialogue method according to claim 2, characterized in that, Step S1 involves packaging the initial parameters using the CreateAgent function and constructing an interactive AI digital human; specifically: in, To build an interactive AI digital human, For name, As the profile picture This is a prompt word.
4. The AI interactive digital human dialogue method according to any one of claims 1-3, characterized in that, Step S2 specifically includes: S21. Based on the application areas of AI digital humans, acquire and match standard documents, and classify and integrate the matched standard documents to form a standardized knowledge base; S22. Perform word segmentation on the text in the standardized knowledge base; S23. Perform vector transformation on the standard knowledge base to generate dense vectors and sparse vectors, and associate and store the dense vectors and sparse vectors with the original text to construct a searchable multimodal vector standardized knowledge base. S24. Call the Bind function to bind the multimodal vector standardized knowledge base to the AI digital human.
5. The AI interactive digital human dialogue method according to claim 4, characterized in that, Step S22 involves word segmentation of the text in the standardized knowledge base, specifically as follows: The text in the standardized knowledge base is segmented into Chinese words using a text splitter. The default segmentation parameters are: chunk size=512, and chunk overlap=20 for each segment. Large files are split into parent and child segments.
6. The AI interactive digital human dialogue method according to any one of claims 1-3, characterized in that, Step S3 specifically includes: S31. Generate a query vector, receive user query text, and encode the user query text in parallel into dense vector and sparse vector; S32. Vector retrieval: Input the dense vector and sparse vector into a standardized knowledge base, perform hybrid similarity retrieval, and obtain a retrieval structure set; the retrieval structure set includes subtext content, similarity distance, and hierarchical index identifier; S33. Extract the parent text content corresponding to the sub-text content based on the hierarchical index identifier; S34. A standard answer generated by AI digital human combined with hierarchical context.
7. The AI interactive digital human dialogue method according to claim 6, characterized in that, In step S31, the user's question text is encoded in parallel into dense vectors and sparse vectors using the BGE-M3 model.
8. The AI interactive digital human dialogue method according to claim 6, characterized in that, Step S32 inputs the dense vector and sparse vector into the standardized knowledge base and calls the hybrid_search interface to perform hybrid similarity retrieval; the score of the similarity retrieval is: in, The score for similarity retrieval. For dynamic weighting coefficients, , The dense vector of the user's question text. This is a dense vector for retrieving documents from a standardized knowledge base. The coefficient representation of the user query, i.e., the term set. This is the inverse document frequency calculation formula. For terms In the current document word frequency in To adjust the parameters.
9. The AI interactive digital human dialogue method according to claim 6, characterized in that, The hierarchical index identifier includes the parent text UUID and the paragraph position offset; Based on the parent text UUID, match the parent text file; extract the complete paragraph containing related subtext based on the paragraph offset; the following is: in, This is the offset of the parent text. This is the offset of the subtext. To preset the paragraph length threshold, This is the context window compensation coefficient.
10. The AI interactive digital human dialogue method according to claim 8, characterized in that, After step S3, the method further includes: S4. Render the structured question and answer content into a standardized document output based on the custom user layout and / or template-driven approach; The structure of the template is extracted as follows: The standardized document output is as follows: in, Extracting structure for calling template The structure extracted from the template file, For the uploaded template, For custom user layout rules, This is the final set of merged typesetting rules. A collection of user-generated question texts. To standardize document output.