Digital human live broadcast interaction method and system, electronic device and storage medium

Abstract

The application relates to a digital person live broadcast interaction method and system, electronic equipment and a storage medium. The method comprises the following steps: in the process of digital person live broadcast, extracting a plurality of historical comment information in a preset time period from a historical comment pool, and extracting a plurality of commodity portrait information corresponding to commodities from a commodity knowledge base; the commodity portrait information comprises commodity explanation rhetoric corresponding to the commodities; through a large language model, context understanding and context adaptation are performed on the plurality of historical comment information and the plurality of commodity portrait information, and a target intention of an audience is recognized; target commodity portrait information matched with the target intention is determined from the plurality of commodity portrait information, so that a target commodity explanation rhetoric is extracted from the target commodity portrait information; and the target commodity explanation rhetoric is used to generate and voice output a target reply text. The scheme provided by the application can meet the demand of the audience for specific commodity links, sizes and live broadcast room specific rhetoric.

Description

Technical Field

[0001] This application relates to the field of digital human live streaming technology, and in particular to digital human live streaming interaction methods, systems, electronic devices and storage media. Background Technology

[0002] Digital human live streaming, leveraging technologies such as AI and real-time rendering, reconstructs the live streaming ecosystem from business, operations, technology, and brand dimensions, enabling cost reduction and efficiency improvement, stable output, full-domain coverage, and IP accumulation, while providing the industry with a new path for compliant, controllable, and innovative growth.

[0003] The interactive methods of digital human live streaming in related technologies mainly rely on keyword triggers or general large models to generate responses. The basic process is to collect a single comment from the audience, and then the AI ​​generates a reply text based on semantics, which is then converted into voice output, forming a single-round question-and-answer mode. However, in e-commerce live streaming scenarios, this mode often fails to simulate the on-the-spot adaptability of top-notch live streamers and cannot meet the audience's needs for specific product links, sizes, and live streaming-specific language. For example, digital human live streamers in related technologies have difficulty understanding ellipses and referential words in comments (such as "that link," "this size," "how much does it cost," "can I wear it," etc.), leading to errors in intent recognition and resulting in "answers that are not relevant to the question." Summary of the Invention

[0004] To address or partially address the problems existing in related technologies, this application provides a digital human live streaming interaction method, system, electronic device, and storage medium that can meet the audience's needs for specific product links, sizes, and live streaming-specific scripts.

[0005] The first aspect of this application provides a digital human live streaming interaction method, including: During the digital human live stream, multiple historical comment information within a preset time period is extracted from the historical comment pool, and multiple product profile information corresponding to products is extracted from the product knowledge base; the product profile information includes the product explanation script for the corresponding product; By using a large language model to perform contextual understanding and contextual adaptation on multiple historical review information and multiple product profile information, the target intent of the audience can be identified. From the multiple product profile information, determine the target product profile information that matches the target intent, so as to extract the target product explanation script from the target product profile information; Using the script used to explain the target product, generate and output the target response text via voice.

[0006] In one embodiment, multiple historical comment information entries are stored in the historical comment pool in the form of a first structured field, and multiple product profile information entries are stored in the product knowledge base in the form of a second structured field; before performing contextual understanding and contextual adaptation on the multiple historical comment information entries and multiple product profile information entries using a large language model to identify the audience's target intent, the method further includes: Based on the first structured field and the second structured field, entity disambiguation is performed on user entities in the multiple historical comment information and product entities in the multiple product profile information.

[0007] In one embodiment, the plurality of products includes the currently explained product and related products filtered through pre-processing logic. The historical audience information includes at least one of username, comment content, and posting time. The product profile information further includes a product attribute index corresponding to the product, which includes at least one of product ID, link number, name, specifications, material, price, and logistics. The step of using a large language model to perform contextual understanding and contextual adaptation on the plurality of historical comment information and the plurality of product profile information to identify the audience's target intent includes: By using a large language model, multiple historical comment messages are sorted according to their sending time to construct a continuous interaction context; The large language model extracts multiple keywords from the continuous interaction context, and maps each keyword to a corresponding standardized entity. The large language model is used to match each standardized entity with each product attribute index in the product profile information to obtain a recognition result that includes the target intent of the audience.

[0008] In one embodiment, before mapping the plurality of keywords to corresponding standardized entities, the method further includes: When there are numeric keywords among the multiple keywords, the numeric keywords are format-matched and context-aligned to convert them into corresponding semantic keywords.

[0009] In one embodiment, the identification result further includes the confidence level of the target intent; before determining the target product profile information matching the target intent from the plurality of product profile information, the method further includes: Determine whether the confidence level is greater than or equal to a preset confidence threshold; If the confidence level is less than the preset confidence threshold, a follow-up questioning logic is triggered; the follow-up questioning logic is used to guide the audience to output corresponding valid auxiliary information. The target intent of the audience is re-identified by the large language model based on the effective auxiliary information until the confidence level is greater than or equal to the preset confidence threshold. Then, the step of determining the target product profile information that matches the target intent from multiple product profile information is executed.

[0010] In one embodiment, generating and outputting the target response text via voice using the target product explanation script includes: Obtain a preset prompt word template; the preset prompt word template is configured with live broadcast script rules, which include at least one of script repetition rules, address preference rules, and negative behavior blocking rules. The script repetition rules are used to standardize the expression of reply sentences, the address preference rules are used to unify the interactive titles in the live broadcast room, and the negative behavior blocking rules are used to filter external information, sensitive words, and illegal traffic redirection information. The target product description script is injected into the preset prompt word template to obtain the initial response text; The initial response text is subjected to compliance verification, and the initial response text that passes the compliance verification is determined as the target response text that meets the live broadcast style and compliance requirements; The target response text is output via voice.

[0011] In one embodiment, the voice output of the target response text includes: The target response text is converted into a target response voice that matches the digital human persona, and a target response action that matches the target response text is determined. While controlling the digital human to perform the target response action, the target response voice is output.

[0012] A second aspect of this application provides a digital human live streaming interactive system, comprising: The data acquisition module is used to extract multiple historical comment information within a preset time period from the historical comment pool and to extract product profile information corresponding to multiple products from the product knowledge base during the digital human live broadcast; the product profile information includes the product explanation script for the corresponding product; The intent recognition module is used to perform contextual understanding and contextual adaptation on multiple historical comment information and multiple product profile information through a large language model to identify the target intent of the audience. The script matching module is used to determine the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information; The response output module is used to generate and output the target response text via voice using the description script for the target product.

[0013] A third aspect of this application provides an electronic device, comprising: Processor; and A memory that stores executable code, which, when executed by the processor, causes the processor to perform the method described above.

[0014] A fourth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.

[0015] The technical solution provided in this application may include the following beneficial results: The solution provided in this application, during digital human live streaming, extracts multiple historical comment information within a preset time period from a historical comment pool, and extracts multiple product profile information corresponding to multiple products from a product knowledge base; the product profile information includes the corresponding product explanation script; through a large language model, it performs contextual understanding and context adaptation on multiple historical comment information and multiple product profile information to identify the audience's target intent; it determines the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information; using the target product explanation script, it generates and outputs the target reply text via voice. By introducing multiple historical comment information within a preset time period, this application upgrades from a "single-turn question-and-answer mode" to a "multi-turn dialogue management mode," enabling the large language model to have "memory" capabilities, ensuring that replies always fit the live streaming process and the audience's interactive context, which can greatly improve the accuracy of intent recognition, avoid the phenomenon of "answering the wrong question," and each product has a corresponding product explanation script, thus meeting the audience's needs for specific product links, sizes, and live streaming-specific scripts.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0018] Figure 1 This is a flowchart illustrating the digital human live-streaming interaction method according to an embodiment of this application; Figure 2 This is another schematic flowchart illustrating the digital human live streaming interaction method in an embodiment of this application; Figure 3 This is a schematic diagram of the framework of a digital human live streaming interactive system shown in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a digital human live streaming interactive system shown in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0019] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0020] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0021] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] The interactive methods of digital human live streaming in related technologies mainly rely on keyword triggers or general large models to generate responses. The basic process is to collect a single comment from the audience, and then the AI ​​generates a reply text based on semantics, which is then converted into voice output, forming a single-round question-and-answer mode. However, in e-commerce live streaming scenarios, this mode often fails to simulate the on-the-spot adaptability of top-notch live streamers and cannot accurately identify the audience's needs for specific product links, sizes, and live streaming-specific language. For example, digital human live streamers in related technologies have difficulty understanding ellipses and referential words in comments (such as "that link," "this size," "how much does it cost," "can I wear it," etc.), leading to incorrect intent recognition and the phenomenon of "answering the wrong question."

[0023] To address the aforementioned issues, this application provides a digital human live-streaming interaction method. By introducing multiple historical comment information within a preset time period, it upgrades from a "single-turn question-and-answer mode" to a "multi-turn dialogue management mode," enabling the large language model to possess "memory" capabilities. This ensures that responses always align with the live-streaming process and the audience's interactive context, significantly improving the accuracy of intent recognition and avoiding irrelevant responses. Furthermore, each product has corresponding product description scripts, thus meeting the audience's needs for specific product links, sizes, and live-streaming-specific scripts.

[0024] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a flowchart illustrating the digital human live streaming interaction method in an embodiment of this application.

[0026] See Figure 1 The digital human live streaming interaction method of this application may include: S110, during the live streaming of digital humans, extracts multiple historical comment information within a preset time period from the historical comment pool, and extracts product profile information corresponding to multiple products from the product knowledge base; the product profile information includes the product explanation script for the corresponding product.

[0027] In this embodiment, it can be applied to a digital human live streaming interactive system (hereinafter referred to as the "interaction system"). The interaction system is a closed-loop response system that integrates "real-time context + product knowledge base + large model Prompt (prompt word) project". Its core relies on the NLP (Natural Language Processing) capabilities of large language model (LLM), combined with multi-dimensional real-time information enhancement, customized optimization of live streaming scenarios, knowledge base adaptation and real-time interaction rules. At the same time, it links with the digital human rendering engine to complete the oral output. The whole process has a response time of seconds (e.g., 5~10 seconds), which is adapted to the high concurrency, conversational and fragmented comment characteristics of live streaming rooms.

[0028] The interactive system's product knowledge base contains product profile information for each product in the live stream. This product profile information includes pre-configured product description text (chat_content) for each product in the live stream, and its fields also include a link number (or hanging link number), name, etc. For example, the product profile information for the currently described product (present_product_info) might include: Name: High-end Mulberry Silk Commuter Shirt; Link Number: Link 3; Product description text (chat_content): "Dear ladies, this shirt is 100% mulberry silk, it feels as smooth as clouds! Wearing it will make you the most elegant person in the office…".

[0029] During the digital human live stream, the interactive system can retrieve multiple historical comments from the historical comment pool within a preset time period. The preset time period can be approximately N minutes (N being a positive integer). These multiple historical comments can cover at least one historical comment from multiple viewers. In other words, within these N minutes (e.g., 5 minutes), multiple viewers participate in the comments, and each viewer can send one or more comments. Simultaneously, the interactive system can extract product profile information for each product in the live stream from the product knowledge base. This product profile information includes the corresponding product description text (chat_content).

[0030] S120 uses a large language model to perform contextual understanding and contextual adaptation on multiple historical comment information and multiple product profile information to identify the target intent of the audience.

[0031] The comment stream in a live stream is highly concurrent and fragmented. The interaction system can preprocess multiple historical comment messages from nearly N minutes to provide high-quality input data for a large language model. In its implementation, the interaction system can remove invalid comments such as garbled text, pure symbols, repetitive content, and advertisements from multiple historical comment messages; it can also standardize and correct colloquial, typo-prone, and abbreviated historical comments, for example, correcting "Is link 1 the one we just talked about?" to "Is link 1 the product we just discussed?", and correcting "How long is the warranty for this thing?" to "How long is the warranty period for this product?"; and it sorts the massive influx of historical comments by interaction popularity (likes / follow-up questions), core needs (product inquiries / after-sales service / benefits), and frequency, prioritizing responses to the questions most viewers care about, avoiding meaningless, scattered replies.

[0032] The interactive system can call the large language model, and then input the preprocessed historical comment information of nearly N minutes and the product profile information of each product in the live room into the large language model. In order to use the powerful natural language processing capabilities of the large language model, the system can perform contextual understanding and context adaptation on the historical comment information of nearly N minutes and the product profile information of each product in the live room, and identify the target intent of each viewer. The specific dimensions of the large language model's analysis can include: core intent judgment, such as determining whether the viewer is inquiring about products (links, specifications, prices, functions), having after-sales questions (warranty, express delivery, returns), asking about benefits (discounts, full reductions, gifts), engaging in casual conversation (praise / criticism), expressing negative doubts (product problems / high prices), or making a directive request (link / explain again); key information extraction: extracting core keywords from historical comments, such as "link 1", "warranty", "X code", "free shipping", etc., to pinpoint the core focus of the response; and livestream context adaptation: combining the current livestream progress (e.g., which link is being explained, what function has just been introduced, what benefits are being offered), the livestream product library (product parameters, prices, benefits for the corresponding link), and brand communication guidelines (prohibited words, essential selling points) to avoid "answering the wrong question" (e.g., if viewer A asks about link 1, the large language model will not reply with information about link 2). For example, suppose the preprocessed historical comment information is "Is link 1 the product we just talked about?", the big language model identifies that the core intent is to confirm the product link, the key information includes link 1, the product we just talked about, and the live broadcast context is that the live broadcast room is currently promoting link 1 / link 1 has just been explained, and finally the target intent is to confirm link 1 and give a positive answer.

[0033] S130, determine the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information.

[0034] If the target intent is directed to any product, such as the intent to find a product, and the historical review information is "Which link is the one we just saw?" or "Do you have the red one?", and the intent is to locate a specific product in the shopping cart, then the interactive system can determine the product profile information of the product that the target intent is directed to from multiple product profile information as the target product profile information, so as to directly extract the target product description text (chat_content) from the target product profile information.

[0035] This application embodiment is also configured with multiple globally universal phrases. If the target intent does not point to any product, such as a social intent, the historical comment information is "The anchor is really beautiful" or "Check-in", there is no core product entity, and the intent is to conduct emotional interaction, then the interaction system can determine the target globally universal phrase that matches the target intent from multiple globally universal phrases.

[0036] S140 uses the target product description script to generate and output the target response text via voice.

[0037] The interactive system can generate target response text based on the target product's description (chat_content). For example, if the target product's description (chat_content) is, "Dear ladies, this shirt is 100% mulberry silk, it feels as smooth as clouds! Wearing it will make you the most elegant person in the office...", then the target response text could be, "Some ladies have asked about the comfort of the fabric. This shirt is 100% mulberry silk, it feels as smooth as clouds, wearing it next to the skin is a true pleasure! We recommend hand washing it in cold water and laying it flat to better protect its elegant feel."

[0038] Interactive systems can utilize TTS (Text-To-Speech) technology to convert the target response text into the target response speech for output.

[0039] As can be seen from this example, the solution provided in this application upgrades from a "single-turn Q&A mode" to a "multi-turn dialogue management mode" by introducing multiple historical comment information within a preset time period. This enables the large language model to have "memory" capabilities, ensuring that responses always fit the live broadcast process and the context of audience interaction. This can greatly improve the accuracy of intent recognition, avoid the phenomenon of "answering the wrong question," and each product has corresponding product description scripts, thus meeting the audience's needs for specific product links, sizes, and live broadcast-specific scripts.

[0040] Figure 2 This is another flowchart illustrating the digital human live streaming interaction method shown in this application.

[0041] See Figure 2 The digital human live streaming interaction method of this application may include: S210, during the digital human live stream, extracts multiple historical comment information within a preset time period from the historical comment pool, and extracts product profile information corresponding to multiple products from the product knowledge base; the product profile information includes the product explanation script for the corresponding product.

[0042] This step can be found in the description in S110, and will not be repeated here.

[0043] S220 uses a large language model to perform contextual understanding and adaptation on multiple historical review information and multiple product profile information to identify the target intent of the audience.

[0044] This step can be found in the description in S120, and will not be repeated here.

[0045] In one embodiment, multiple historical review information entries are stored in a historical review pool in the form of a first structured field, and multiple product profile information entries are stored in a product knowledge base in the form of a second structured field. Before performing contextual understanding and contextual adaptation on the multiple historical review information entries and multiple product profile information entries using a large language model to identify the audience's target intent, the method may further include: Based on the first and second structured fields, entity disambiguation is performed on user entities in multiple historical comment information and product entities in multiple product profile information.

[0046] Furthermore, digital human response technology in related technologies still has the following shortcomings in complex e-commerce interaction environments: User identification interference in responses: General large models are prone to mistaking audience usernames (such as "Nike" or "a certain brand") for product brands, leading to confusion in response logic.

[0047] To address this, this application embodiment utilizes structured field design and physical isolation mechanisms to enforce the distinction between entities with the same name (such as the username "Nike" and the product brand "Nike"), thus preventing confusion during the intent recognition stage. Specifically, this application embodiment stores user entities (such as usernames) from multiple historical comment information entries and product entities (such as product brands) from multiple product profile information entries in different structured fields. The user entities from multiple historical comment information entries are stored in the `comment_history` field (the first structured field), which records the comment content and identity identifiers of viewers in the live stream. The product entities from multiple product profile information entries are stored in the `product_info` field (the second structured field), which stores the structured data of the products in the live stream (such as product ID, brand, specifications, price, etc.). Through this operation, the `comment_history` field and the `product_info` field correspond to different physical data sources or data tables, achieving complete isolation at the data storage level. This ensures that the two types of information do not physically overlap, enabling the forced distinction between user entities and product entities during the intent recognition stage and preventing confusion due to the same name.

[0048] In its implementation, the interactive system can perform entity disambiguation between user entities in multiple historical comment records and product entities in multiple product profile records based on the `comment_history` and `product_info` fields: if an entity originates from the `comment_history` field, it is determined to be a user entity (e.g., the username "Nike"); if the entity originates from the `product_info` field, it is determined to be a product entity (e.g., the product brand "Nike"). This method effectively distinguishes between the username "Nike" and the product brand "Nike," avoiding entity recognition errors due to identical names and further improving the accuracy of intent recognition.

[0049] In one embodiment, the multiple products include the currently explained product and related products filtered through pre-processing logic. Historical audience information includes at least one of username, comment content, and posting time. Product profile information also includes a product attribute index for the corresponding product, which includes at least one of product ID, link number, name, specifications, material, price, and logistics. By using a large language model to perform contextual understanding and adaptation on multiple historical comment information and multiple product profile information, the target intent of the audience can be identified, which may include: By sorting multiple historical comments according to their sending time using a large language model, a continuous interaction context is constructed. Multiple keywords are extracted from the continuous interaction context using the large language model, and each keyword is mapped to a corresponding standardized entity. The large language model is then used to match each standardized entity with the product attribute indexes in each product profile information to obtain an identification result containing the audience's target intent.

[0050] Furthermore, digital human response technology in related technologies still has the following shortcomings in complex e-commerce interaction environments: Poor product relevance: Unable to effectively identify "link number" mentioned by the audience, and lacks support for product recommendations outside of the explanation and diverse wording, resulting in dull responses.

[0051] The product knowledge base provided in this application provides a comprehensive database of product profiles. This database supports both the use of the current product profile by a large language model and the use of product profiles for products not currently being explained. Furthermore, each product profile is configured with diverse product description scripts, supporting product recommendations in different scenarios and avoiding monotonous responses. Specifically, the comprehensive database includes the currently explained product and related products (or non-currently explained products) filtered by pre-processing logic. The second structured field (product_info field) can cover two forms: a present_product_info field configured for the currently explained product, and a product_info field configured for related products (i.e., non-currently explained products) filtered by pre-processing logic. Each product profile can include a product attribute index, which may include, but is not limited to: product ID, link number, name, specifications (or size), material, price, and logistics information.

[0052] Each historical viewer's information can include username, comment content, and sending time. The large language model can assign multiple historical comment messages from nearly N minutes to the corresponding viewer's historical comment information set based on the username. Then, the historical comment information set of each viewer is sorted according to its sending time to construct a continuous interaction context for each viewer. The continuous interaction context is a temporal context chain of [username-comment content-sending time]. For example, user A's continuous interaction context could be: 10:05 "Can I wear this red one?" → 10:06 "I am 165cm and 50kg".

[0053] Furthermore, the large language model can transform each viewer's continuous interaction context into a semantic vector and label the context interaction type: initial question (without prior relevant comments), follow-up question (adding height and weight), repeated confirmation (asking the same question multiple times), etc.

[0054] Large language models can extract multiple keywords from each viewer's continuous interaction context (or semantic vector) and map these keywords to corresponding standardized entities. Standardized entities include, but are not limited to, height and weight, color, size, material, price, and link number. For example, assuming keywords such as "red," "can I wear it?", "165cm, 50kg," "is it genuine leather?", and "which link number" are extracted from user A's continuous interaction context (or semantic vector), the large language model can map these keywords to corresponding standardized entities. For example, "red" → mapped to [color: red]; "can I wear it?" → mapped to [size]; "165cm, 50kg" → mapped to [height and weight: 165 / 50]; "is it genuine leather?" → mapped to [material: genuine leather]; "which link number" → mapped to [link number].

[0055] The large language model can accurately match these standardized entities with each product attribute index in all product profile information, determine the product and attribute dimensions corresponding to each standardized entity, and then combine the contextual interaction type (initial question / follow-up question / repeated confirmation), keyword semantics, and product attribute matching results to accurately identify the target intent of each audience member. For example, Example 1: Standardized entity is [height and weight: 165 / 50] + [size] → Match the size index (or specification index) of the corresponding product in all products (the currently explained product and related products filtered by the pre-processing logic) → Determine the target intent is size / specification intent, pointing to product XX; Example 2: Standardized entity is [material: genuine leather] → Match the material index of the corresponding product in all products → Determine the target intent is material / efficacy intent, pointing to product YY; Example 3: Standardized entity is [link number] + [color: red] → Match the link number index of the corresponding product in all products → Determine the target intent is product search intent, pointing to product ZZ (link number 3, red version).

[0056] In one embodiment, before mapping the multiple keywords to their corresponding standardized entities, the method may further include: When multiple keywords contain numeric keywords, the numeric keywords are formatted and context-aligned to convert them into corresponding semantic keywords.

[0057] Furthermore, digital human response technology in related technologies still has the following shortcomings in complex e-commerce interaction environments: Difficulty in identifying intent in specific vertical fields: For purely numerical comments such as "160, 60" in the clothing category, the general large model cannot combine the context to identify them as height and weight inquiries rather than price or quantity.

[0058] To address this, this application embodiment establishes a digital intent conversion matrix to automatically map numbers to sizes, link numbers, or parameters based on context (such as clothing categories). In specific implementation, after the large language model extracts multiple keywords, the interactive system can determine whether there are numerical keywords (such as "160, 60") among the multiple keywords. When numerical keywords (such as "160, 60") are present, the interactive system can perform format matching on the numerical keywords, for example, identifying a specific combination pattern of "number A + separator + number B," and can perform context alignment on the numerical keywords. For example, when the "clothing category" tag is true and historical review information contains size keywords, the digital intent conversion matrix forces the numerical keywords to be converted into corresponding semantic keywords, for example, forcibly mapping the first number to "height (cm)" and the second number to "weight (kg / jin)."

[0059] S230, determine whether the confidence level of the target intent is greater than or equal to the preset confidence threshold; if not, proceed to step S240; if yes, proceed to step S260.

[0060] This application embodiment introduces multiple historical comment information over N minutes, which can greatly improve the accuracy of intent recognition, but there is also a certain probability of inaccurate intent recognition. To avoid forced answers that lead to logical confusion in responses, this application embodiment provides a dynamic follow-up questioning mechanism. Specifically, the recognition result output by the large language model includes the target intent and the confidence level of the target intent. The confidence level is used to characterize the recognition accuracy of the large language model for the target intent. The interactive system can first compare the confidence level with a preset confidence threshold (such as 70%) to determine whether to trigger the dynamic follow-up questioning mechanism based on the comparison result. For example, if the confidence level is less than the preset confidence threshold, then proceed to step S240 and trigger the dynamic follow-up questioning mechanism; if the confidence level is greater than or equal to the preset confidence threshold, then proceed to step S260 without triggering the dynamic follow-up questioning mechanism.

[0061] Furthermore, the embodiments of this application can calibrate the confidence level based on the contextual interaction type (initial question / follow-up question / repeated confirmation). Specifically, the interaction system can calculate the matching degree between the standardized entity and the product attribute index. If the matching degree is ≥ the first matching threshold (e.g., 0.8), and the contextual interaction type is the initial question, then the calibrated confidence level is ≥ the first confidence threshold (e.g., 90%), and proceed to step S260 without triggering the dynamic follow-up question mechanism. If the second matching threshold (e.g., 0.5) is ≤ the matching degree < the first matching threshold (e.g., 0.8), and the contextual interaction type is the follow-up question, then the second confidence threshold (e.g., 70%) is ≤ the calibrated confidence level < the first confidence threshold (e.g., 90%), and proceed to step S260 without triggering the dynamic follow-up question mechanism. If the matching degree is < the second matching threshold (e.g., 0.5), then the confidence level is < the second confidence threshold (e.g., 70%), and proceed to step S240 to trigger the dynamic follow-up question mechanism.

[0062] S240 triggers the follow-up questioning logic; the follow-up questioning logic is used to guide the audience to output corresponding effective auxiliary information.

[0063] When the confidence level is less than the preset confidence threshold, it indicates that the large language model has not accurately identified the audience's true intention. In other words, the accuracy of the target intention output by the large language model is not high. To avoid forced answers that lead to confusing response logic, the interactive system can trigger follow-up questions (such as "Does Baozi want to see which link, or do you want to ask about the size?" / "Does Baozi want to ask about the size of this red one?"). Follow-up questions are used to guide the audience to provide more effective auxiliary information.

[0064] S250, the target intent of the audience is identified again based on effective auxiliary information through the large language model, and the process returns to step S230.

[0065] The large language model can re-identify the audience's target intent based on effective auxiliary information, and return to step S230 to determine whether the confidence level of the current target intent is greater than or equal to the preset confidence threshold. If it is still less than the preset confidence threshold, then steps S240 to S250 are executed again until the confidence level of the target intent is greater than or equal to the preset confidence threshold, and then step S260 is entered.

[0066] S260, determine the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information.

[0067] This step can be found in the description in S130, and will not be repeated here.

[0068] S270 uses the target product description script to generate and output the target response text via voice.

[0069] This step can be found in the description in S140, and will not be repeated here.

[0070] In one embodiment, using a script to explain the target product and generating and outputting the target response text via voice may include: Obtain a preset prompt word template; the preset prompt word template is configured with live broadcast script rules, which include at least one of the following: script repetition rules, address preference rules, and negative behavior blocking rules. The script repetition rules are used to standardize the expression of reply sentences, the address preference rules are used to unify the interactive titles in the live broadcast room, and the negative behavior blocking rules are used to filter external information, sensitive words, and illegal traffic-driving information; inject the target product explanation script into the preset prompt word template to obtain the initial reply text; perform compliance verification on the initial reply text, and determine the initial reply text that passes the compliance verification as the target reply text that conforms to the live broadcast style and compliance requirements; output the target reply text by voice.

[0071] This application embodiment can pre-configure a preset prompt template that matches the style of the live broadcast room, so that when the digital human replies to the audience, it will no longer use a general cold tone, but will imitate the style of this template (such as friendly / professional / funny), so that the AI ​​reply style is consistent with the tone of the live broadcast room.

[0072] The preset prompt template (Prompt) contains live streaming script rules, which may include, but are not limited to: script repetition rules, address preference rules, and negative behavior blocking rules. Among them, script repetition rules are used to standardize the expression of reply sentences (such as "A baby asked..."), address preference rules are used to unify the interactive address in the live streaming room (such as "baby"), and negative behavior blocking rules are used to filter external information, sensitive words, and illegal traffic redirection information.

[0073] For example, the preset prompt template could be: You are the intelligent assistant in the XX live stream, and you must strictly imitate the following preset script style when replying to viewers: Pre-set dialogue script reference: {chat_content} Audience question: {user_question} Response rules: 1. Repetition of the original message; 2. Rules for addressing the recipient; 3. Rules for blocking negative behavior.

[0074] The interactive system can inject the target product description text (chat_content) into the preset text reference of the preset prompt template (Prompt) to obtain the initial response text.

[0075] Taking the current product (present_product_info) as an example: Name: High-end Mulberry Silk Commuter Shirt; Link Number: Link 3; Target product description (chat_content): "Dear ladies, this shirt is 100% mulberry silk, it feels as smooth as clouds! Wearing it will make you the most elegant person in the office..."; The initial response text could be: "Someone asked if the fabric is comfortable (paraphrasing the question). Ladies, this shirt is 100% mulberry silk, it feels as smooth as clouds (sales pitch), wearing it is simply a pleasure! We recommend hand washing in cold water and laying it flat to better protect its elegant feel (style alignment)."

[0076] To ensure the security and compliance of the final output response text, the interactive system can perform compliance verification on the initial response text. If the initial response text passes the compliance verification, it can be designated as the target response text that conforms to the live broadcast style and compliance requirements for voice output. If the initial response text fails the compliance verification, it can be adjusted to pass the compliance verification and then designated as the target response text that conforms to the live broadcast style and compliance requirements for voice output.

[0077] In one embodiment, the voice output target response text may include: The target response text is converted into a target response voice that matches the digital human persona, and a target response action that matches the target response text is determined; the digital human is controlled to output the target response voice while performing the target response action.

[0078] The digital human-driven terminal can interact in real time based on the generated reply text and action commands. In its implementation, the interaction system utilizes TTS (Text-to-Speech) technology to convert the target reply text into a target reply voice. This voice voice matches the digital human's established personality in terms of tone, speed, and intonation; for example, a lively tone with a slightly faster speed, or a professional tone with a steady speed. Simultaneously, the interaction system can determine the target reply action that matches the text. This action can include at least one of lip movements, body language, or facial expressions. For instance, the digital human can nod or raise a hand while replying to avoid slurred speech or stiff movements, thus enhancing the realism of the interaction. After the target reply voice and action are rendered synchronously, they are output directly in real-time in the live stream, allowing viewers to see the digital human's precise response to their comments.

[0079] Furthermore, embodiments of this application can update the historical comment pool. For example, the current interaction (including audience comments and digital human replies) can be recorded in the historical comment pool. The updated historical data will serve as the context for the next round of responses, providing a basis for subsequent decisions.

[0080] To facilitate understanding of the solutions provided in this application, the following examples are provided in the embodiments of this application: Example 1: Figure 3 This is a schematic diagram of the interactive system architecture. The specific steps are as follows: 1. Multi-source data aggregation (1) Context extraction: Real-time capture of multiple historical comment information (comment_history) from the live broadcast room over the past N minutes, including username, comment content and sending time.

[0081] (2) Product profile extraction: includes the currently presented product (present_product_info) and related products (product_info) filtered by the preceding logic. The fields cover product ID, link number, name, specifications, material, price, logistics and product description (chat_content).

[0082] 2. Core Prompt Response Strategy Generation (1) Entity decoupling: By using structured fields, user entities (such as username) and product entities (such as product brand) are forcibly distinguished to avoid confusion with the same name.

[0083] (2) Logical mapping: Establish a digital intent conversion matrix to automatically map numbers to sizes, link numbers or parameters according to context (such as clothing categories).

[0084] (3) Script alignment: Inject the script for explaining the target product (chat_content) into the Prompt to make the AI ​​response style consistent with the tone of the live broadcast room.

[0085] 3. Compliance and Behavioral Constraints (1) Behavioral guidelines: Set rules for repeating phrases (such as “Baby asked…”), rules for addressing (such as “Baby”), and rules for blocking negative behaviors (such as prohibiting mentioning external information, sensitive words, and illegal traffic-driving information).

[0086] (2) Dynamic follow-up questioning: When the confidence level of the target's intention is lower than the preset confidence threshold, the follow-up questioning logic is automatically triggered instead of forcing an answer.

[0087] 4. Implementation and Feedback The digital human-driven terminal interacts in real time with the generated response text and action commands, and updates the historical comment pool.

[0088] Example 2: Examples of target response texts corresponding to different target intentions 1. Size / Specification Intent Viewers commenting "Can I wear this?" or "160, 60" are trying to get suggestions on the appropriate SKU (Stock Keeping Unit). Example of the target response text: 1a) General version: "Baby, you are 160cm tall and weigh 60kg, so size M is just right for you. Our sizes are accurate." 1b) Concise version: "Size M fits your baby well, just order the corresponding SKU." 1c) Low-confidence follow-up question version (such as only asking "Can I wear this?", without height and weight): "Honey, please tell me your height and weight, and I will help you choose the most suitable size."

[0089] 2. Material / Function Intent Viewers commenting "Is it genuine leather?" or "Can I use it if I have allergies?" are trying to confirm the quality of the product. Example of the target response text: 2a) Material: "Don't worry, baby, this is made of top-grain genuine leather. The quality is guaranteed. We will compensate you tenfold if it's fake." 2b) Efficacy category: "This product does not contain any allergenic ingredients, so it is safe for those with sensitive skin or allergies to use."

[0090] 3. Price / Activity Intent Viewers commenting "How much?" or "Is it buy one get one free?" are trying to confirm the promotional offer. Example of the target response text: 3a) Price category: "Baby, this product is priced at 99 yuan for the live stream exclusive, which is 30 yuan cheaper than the regular price. It's a super deal." 3b) Promotional items: "Yes, baby! Buy one get one free now, and the free gift will be sent with your order. This is a special offer!"

[0091] 4. Purpose of sourcing goods Viewers commenting "Which link is the one from before?" or "Are there any in red?" are trying to locate specific items in their shopping carts.

[0092] Example of the target response text: 4a) Link positioning type: "Honey, the one I just explained is in link number 1. Just click on the shopping cart to buy it." 4b) Color / Style: "Yes, yes! The red one is in link number 1. Just select the red SKU when you place your order."

[0093] 5. Order expediting / logistics intent Viewers commenting "Why hasn't it been shipped yet?" or "When will it arrive?" are trying to find out the after-sales progress. Example of the target response text: Order expediting messages: "Don't worry, honey, we'll ship within 48 hours of you placing your order. Shipping is free. We'll update you with the logistics information once the goods are shipped." Logistics: "Honey, XX Express will arrive in 3-5 days. You can check the logistics progress in real time on the order page."

[0094] 6. Social Intent Viewers commenting "The streamer is so pretty" or "checking in" are trying to engage in emotional interaction. Example of the target response text: The compliment was responded with: "Thank you for the compliment, sweetheart. Sending love! If you like it, you can stay a little longer!" The person who checked in responded: "Welcome to check in, baby! There are tons of perks in the live stream today, don't miss it!"

[0095] As can be seen from this example, the solution provided in this application includes the following technical advantages: (1) The reply mechanism based on the dual index of "link number-product information" solves the problem that the general large model in related technologies cannot recognize the logic of the live broadcast room link, and realizes the accurate mapping from digital instructions to product details.

[0096] (2) Dynamic Prompt architecture that integrates historical comment context: through<comment_history> (Historical commentary information) and<product_info> Cross-comparison of (product profile information) enables intelligent processing of complex audience needs (such as comparing two pieces of clothing or asking follow-up questions).

[0097] (3) "Digital intent" parsing algorithm in e-commerce vertical field: It has been specially optimized to recognize the semantics of numbers, SKUs and link numbers in e-commerce scenarios, reducing the response errors caused by product category characteristics.

[0098] Therefore, the solution provided in this application can significantly improve the professionalism, human-likeness, and conversion efficiency of digital human anchors' interactions with viewers: (1) Extremely high accuracy of intent recognition: By introducing the historical comment context of nearly N minutes, the large language model has the ability to "memorize" and can accurately identify ellipsis and pronouns (such as "that link" and "this size").

[0099] (2) Deeply adapted to e-commerce business logic: For the first time, the linkage of "product ID-link number-product explanation script" has been realized, enabling the digital human to guide the audience to view the shopping cart link like a real anchor.

[0100] (3) Highly humanized operation: The mandatory requirement to repeat the questions and use specific nicknames greatly enhances the interactive atmosphere of the live broadcast room and reduces the feeling of being machine-like.

[0101] (4) Safe, compliant and controllable: Built-in e-commerce platform red line warning, automatically filtering external information, sensitive words and illegal traffic-driving information.

[0102] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a digital human live streaming interactive system, electronic device, and corresponding embodiments.

[0103] Figure 4 This is a schematic diagram of the structure of a digital human live streaming interactive system shown in an embodiment of this application.

[0104] See Figure 4 This application provides a digital human live streaming interactive system, which may include: The data acquisition module 410 is used to extract multiple historical comment information within a preset time period from the historical comment pool and to extract product profile information corresponding to multiple products from the product knowledge base during the digital human live broadcast; the product profile information includes the product explanation script for the corresponding product; The intent recognition module 420 is used to perform contextual understanding and contextual adaptation on multiple historical comment information and multiple product profile information through a large language model to identify the target intent of the audience. The script matching module 430 is used to determine the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information. The response output module 440 is used to generate and output the target response text via voice using the target product explanation script.

[0105] In one embodiment, multiple historical review information entries are stored in a historical review pool in the form of a first structured field, and multiple product profile information entries are stored in a product knowledge base in the form of a second structured field; before the intent recognition module 420, the system may further include: The entity disambiguation module is used to perform entity disambiguation between user entities in multiple historical review information and product entities in multiple product profile information based on the first structured field and the second structured field.

[0106] In one embodiment, the multiple products include the currently explained product and related products filtered through pre-processing logic. Historical audience information includes at least one of username, comment content, and sending time. The product profile information also includes a product attribute index for the corresponding product, which includes at least one of product ID, link number, name, specifications, material, price, and logistics. The intent recognition module 420 may include: The context modeling submodule is used to sort multiple historical comment information according to the sending time using a large language model to construct a continuous interactive context; The keyword mapping submodule is used to extract multiple keywords from the continuous interaction context through a large language model, and map the multiple keywords to their corresponding standardized entities. The product attribute matching submodule is used to match each standardized entity with the product attribute indexes in each product profile information through a large language model, so as to obtain the recognition result containing the target intent of the audience.

[0107] In one implementation, before the keyword mapping submodule performs the task of mapping multiple keywords to their corresponding normalized entities, the keyword mapping submodule may also be used to: When multiple keywords contain numeric keywords, the numeric keywords are formatted and context-aligned to convert them into corresponding semantic keywords.

[0108] In one embodiment, the identification result also includes a confidence level of the target intent; prior to the speech matching module 430, the system may further include: The confidence level determination module is used to determine whether the confidence level is greater than or equal to a preset confidence threshold. The dynamic follow-up questioning module is used to trigger follow-up questioning logic if the confidence level is less than a preset confidence threshold; the follow-up questioning logic is used to guide the audience to output corresponding effective auxiliary information. The re-identification module is used to re-identify the audience's target intent based on effective auxiliary information through a large language model until the confidence level is greater than or equal to a preset confidence threshold, at which point the step of executing the speech matching module 430 is determined.

[0109] In one embodiment, the response output module 440 may include: The template acquisition submodule is used to acquire preset prompt word templates. The preset prompt word templates are configured with live broadcast script rules, which include at least one of the following: script repetition rules, address preference rules, and negative behavior blocking rules. Script repetition rules are used to standardize the expression of reply sentences, address preference rules are used to unify the interactive titles in the live broadcast room, and negative behavior blocking rules are used to filter external information, sensitive words, and illegal traffic redirection information. The script injection submodule is used to inject the script explaining the target product into a preset prompt template to obtain the initial response text; The compliance verification submodule is used to verify the compliance of the initial response text and determine the initial response text that passes the compliance verification as the target response text that conforms to the live broadcast style and compliance requirements. The voice output submodule is used to output the target reply text via voice.

[0110] In one embodiment, the voice output submodule may include: The interactive information determination unit is used to convert the target reply text into a target reply voice that matches the digital human persona, and to determine the target reply action that matches the target reply text. The interactive information execution unit is used to control the digital human to output the target response voice while performing the target response action.

[0111] As can be seen from this example, the solution provided in this application upgrades from a "single-turn Q&A mode" to a "multi-turn dialogue management mode" by introducing multiple historical comment information within a preset time period. This enables the large language model to have "memory" capabilities, ensuring that responses always fit the live broadcast process and the context of audience interaction. This can greatly improve the accuracy of intent recognition, avoid the phenomenon of "answering the wrong question," and each product has corresponding product description scripts, thus meeting the audience's needs for specific product links, sizes, and live broadcast-specific scripts.

[0112] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0113] Figure 5 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.

[0114] See Figure 5 The electronic device 500 includes a memory 510 and a processor 520.

[0115] The processor 520 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0116] Memory 510 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 520 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 510 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 510 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0117] The memory 510 stores executable code, which, when processed by the processor 520, can cause the processor 520 to execute part or all of the methods described above.

[0118] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0119] Alternatively, this application may be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) that, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0120] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described above.

[0121] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A digital human live streaming interaction method, characterized in that, include: During the digital human live stream, multiple historical comment information within a preset time period is extracted from the historical comment pool, and multiple product profile information corresponding to products is extracted from the product knowledge base; the product profile information includes the product explanation script for the corresponding product; By using a large language model to perform contextual understanding and contextual adaptation on multiple historical review information and multiple product profile information, the target intent of the audience can be identified. From the multiple product profile information, determine the target product profile information that matches the target intent, so as to extract the target product explanation script from the target product profile information; Using the script used to explain the target product, generate and output the target response text via voice.

2. The method according to claim 1, characterized in that, Multiple historical review information entries are stored in the historical review pool in the form of a first structured field, and multiple product profile information entries are stored in the product knowledge base in the form of a second structured field; before performing contextual understanding and contextual adaptation on the multiple historical review information entries and multiple product profile information entries using a large language model to identify the audience's target intent, the method further includes: Based on the first structured field and the second structured field, entity disambiguation is performed on user entities in the multiple historical comment information and product entities in the multiple product profile information.

3. The method according to claim 1, characterized in that, The multiple products include the currently explained product and related products filtered through pre-processing logic. The historical audience information includes at least one of username, comment content, and posting time. The product profile information also includes a product attribute index for the corresponding product, which includes at least one of product ID, link number, name, specifications, material, price, and logistics. The step of using a large language model to perform contextual understanding and contextual adaptation on the multiple historical comment information and the multiple product profile information to identify the audience's target intent includes: By using a large language model, multiple historical comment messages are sorted according to their sending time to construct a continuous interaction context; The large language model extracts multiple keywords from the continuous interaction context, and maps each keyword to a corresponding standardized entity. The large language model is used to match each standardized entity with each product attribute index in the product profile information to obtain a recognition result that includes the target intent of the audience.

4. The method according to claim 3, characterized in that, Before mapping the multiple keywords to their corresponding standardized entities, the method further includes: When there are numeric keywords among the multiple keywords, the numeric keywords are format-matched and context-aligned to convert them into corresponding semantic keywords.

5. The method according to claim 3, characterized in that, The identification result also includes the confidence level of the target intent; before determining the target product profile information matching the target intent from the multiple product profile information, the method further includes: Determine whether the confidence level is greater than or equal to a preset confidence threshold; If the confidence level is less than the preset confidence threshold, a follow-up questioning logic is triggered; the follow-up questioning logic is used to guide the audience to output corresponding valid auxiliary information. The target intent of the audience is re-identified by the large language model based on the effective auxiliary information until the confidence level is greater than or equal to the preset confidence threshold. Then, the step of determining the target product profile information that matches the target intent from multiple product profile information is executed.

6. The method according to claim 1, characterized in that, The step of generating and outputting the target response text via voice using the script for explaining the target product includes: Obtain a preset prompt word template; the preset prompt word template is configured with live broadcast script rules, which include at least one of script repetition rules, address preference rules, and negative behavior blocking rules. The script repetition rules are used to standardize the expression of reply sentences, the address preference rules are used to unify the interactive address in the live broadcast room, and the negative behavior blocking rules are used to filter external information, sensitive words, and illegal traffic redirection information. The target product description script is injected into the preset prompt word template to obtain the initial response text; The initial response text is subjected to compliance verification, and the initial response text that passes the compliance verification is determined as the target response text that meets the live broadcast style and compliance requirements; The target response text is output via voice.

7. The method according to claim 6, characterized in that, The voice output of the target response text includes: The target response text is converted into a target response voice that matches the digital human persona, and a target response action that matches the target response text is determined. While controlling the digital human to perform the target response action, the target response voice is output.

8. A digital human live streaming interactive system, characterized in that, include: The data acquisition module is used to extract multiple historical comment information within a preset time period from the historical comment pool and to extract product profile information corresponding to multiple products from the product knowledge base during the digital human live broadcast; the product profile information includes the product explanation script for the corresponding product; The intent recognition module is used to perform contextual understanding and contextual adaptation on multiple historical comment information and multiple product profile information through a large language model to identify the target intent of the audience. The script matching module is used to determine the target product profile information that matches the target intent from multiple product profile information, so as to extract the target product explanation script from the target product profile information; The response output module is used to generate and output the target response text via voice using the description script for the target product.

9. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method as described in any one of claims 1-7.

Images