A generative engine content optimization method, apparatus and device

Abstract

The application discloses a generative engine content optimization method, device and equipment. By reconstructing the target content into a structured content unit including interrelated question part and content part, and publishing to a source platform for retrieval by a generative engine, the organization form of the target content is adapted to the retrieval and reference logic of the generative engine, thereby improving the probability of the target content being retrieved and referenced by the generative engine. Meanwhile, the reference feedback information based on the answer of the generative engine adjusts and re-executes at least one of the reconstruction mode and the publishing mode, so that the content optimization can continue to iterate according to the reference situation of the generative engine, the deviation between the content optimization mode and the actual reference preference of the generative engine is reduced, and the reference rate of the target content in the answer of the generative engine is further improved.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and content optimization technology, and in particular to a generative engine content optimization method, apparatus and device. Background Technology

[0002] With the increasing popularity of generative engines, the way users obtain information is changing. After a user submits a query to a generative engine, the engine extracts information from various sources such as web pages, technical documents, industry reports, and user reviews, and integrates them to form a comprehensive answer that is directly presented to the user. In this process, users often no longer browse through the original sources one by one, but instead directly adopt the answer provided by the generative engine.

[0003] Under the aforementioned information acquisition methods, the main challenge for content providers is how to make their content more easily retrieved by generative engines and cited in their generated answers. However, existing content optimization techniques are ill-suited to the content retrieval and citation logic of generative engines, resulting in content that, even after publication, is difficult for generative engines to effectively retrieve and cite, leading to a low citation rate of the content in generative engine answers.

[0004] Therefore, how to increase the probability of content being retrieved and cited by generative engines has become an urgent problem to be solved in this field. Summary of the Invention

[0005] To address this issue, this application provides a method, apparatus, and device for optimizing content in generative engines, thereby resolving the problem that existing content optimization methods are difficult to adapt to the retrieval and citation logic of generative engines, resulting in low content citation rates.

[0006] Firstly, this application provides a method for optimizing content in a generative engine, the method comprising:

[0007] Retrieve the target content to be retrieved and referenced by the generative engine;

[0008] The target content is reconstructed into multiple structured content units; wherein each structured content unit includes an interrelated question section and a content section supporting the question section, the content section being determined from the target content and corresponding to the question section;

[0009] The structured content unit is published to the information source platform so that the generative engine can obtain the structured content unit from the information source platform;

[0010] Obtain the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer;

[0011] Based on the reference feedback information, at least one of the reconstruction method of the structured content unit and the publishing method of the structured content unit is adjusted;

[0012] If the reconstruction method of the structured content unit is adjusted, the reconstruction is re-executed based on the adjusted reconstruction method;

[0013] If the publishing method of the structured content unit is adjusted, the publishing is re-executed based on the adjusted publishing method.

[0014] Secondly, this application provides a generative engine content optimization apparatus, the apparatus comprising:

[0015] The acquisition unit is used to acquire the target content to be retrieved and referenced by the generative engine.

[0016] A content reconstruction unit is used to reconstruct the target content into multiple structured content units; wherein each structured content unit includes an interrelated question part and a content part supporting the question part, the content part being determined from the target content and corresponding to the question part;

[0017] A content publishing unit is used to publish the structured content unit to the information source platform so that the generative engine can obtain the structured content unit from the information source platform;

[0018] The feedback acquisition unit is used to acquire the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer;

[0019] An adjustment unit is configured to adjust at least one of the reconstruction method and the publishing method of the structured content unit based on the reference feedback information; when the reconstruction method is adjusted, the content reconstruction unit is triggered to re-execute the reconstruction based on the adjusted reconstruction method; when the publishing method is adjusted, the content publishing unit is triggered to re-execute the publishing based on the adjusted publishing method.

[0020] Thirdly, this application provides a computer device including a processor and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the executable instructions to implement the steps of the generative engine content optimization method as described above.

[0021] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the generative engine content optimization method described above.

[0022] The generative engine content optimization method provided in this application reconstructs the target content into structured content units that include interrelated question parts and content parts, and publishes them to the information source platform for the generative engine to retrieve. This adapts the organization of the target content to the generative engine's content retrieval and citation logic, thereby increasing the probability that the target content will be retrieved and cited by the generative engine. At the same time, since at least one of the reconstruction method and the publication method is adjusted and re-executed based on the citation feedback information of the generative engine's answers, the content optimization can continuously iterate with the citation situation of the generative engine, reducing the deviation between the content optimization method and the actual citation preference of the generative engine, and further improving the citation rate of the target content in the generative engine's answers. Attached Figure Description

[0023] To more clearly illustrate the embodiments of this application or the technical solutions in 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 merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a generative engine content optimization process provided in an embodiment of this application;

[0025] Figure 2 This is a schematic diagram illustrating the specific process of content optimization for a generative engine provided in the embodiments of this application;

[0026] Figure 3 This is a schematic diagram of the structure of a generative engine content optimization device provided in an embodiment of this application;

[0027] Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of this application. Detailed Implementation

[0028] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] The applicant discovered that optimizing content using existing Search Engine Optimization (SEO) methods typically relies on keyword stuffing, link exchanges, and page structure adjustments. These methods are designed for search engines that present results to users in a ranked manner, aiming to rank the pages higher in search results. However, generative engines, when responding to user queries, do not present retrieved pages directly as a ranked list. Instead, they extract information from multiple sources, including web pages, technical documents, industry reports, and user reviews, and then integrate this information to form an answer. Therefore, optimization methods relying on keyword density and page ranking are unlikely to ensure that content is effectively retrieved and referenced during the information integration and answer formation process by generative engines.

[0030] The applicant further discovered that if existing generative marketing models (such as models that optimize advertising bidding strategies through generative technology) are used to optimize the ad delivery process, they focus on optimizing the bidding decisions in the ad delivery process, without adapting the organization of the content itself to the retrieval and referencing logic of generative engines. Moreover, such models also have problems such as data bias, generated content that does not match the facts (i.e., illusions), and high computational resource consumption, making it difficult to increase the probability of the content being cited by generative engines.

[0031] The applicant also found that if a single text format is used to process the content, when the original content is a non-text format such as an image, PDF document, or video, or when the content processed by the generative engine is long text (such as text with a length of about 200,000 tokens), the existing method is difficult to parse and optimize these formats of content. There are problems with insufficient dynamic response and insufficient multimodal adaptation, which makes it impossible to include this part of the content in the optimization scope and difficult to adapt in real time with the data update cycle of the generative engine.

[0032] Based on the above findings, this application provides a method, apparatus, and device for optimizing content in a generative engine, in order to reduce the deviation between the content optimization method and the actual citation preferences of the generative engine, and to increase the citation rate of the target content in the answers of the generative engine.

[0033] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1:

[0035] This application provides a method for optimizing content in a generative engine. Figure 1 This is a schematic diagram of a generative engine content optimization process provided in an embodiment of this application. (Refer to...) Figure 1 The method includes:

[0036] S101: Obtain the target content to be retrieved and referenced by the generative engine.

[0037] In this application, the generative engine content optimization method is applied to a computer device, which can be a smart terminal, such as a personal computer or tablet computer, or a server, such as an application server or a business server.

[0038] A generative engine is a system that can retrieve information from one or more sources and integrate it to generate an answer based on a user's query. For example, a generative engine can be a conversational question-answering system built on a Large Language Model (LLM), or an intelligent question-answering platform with information retrieval and content generation capabilities. When responding to a user query, a generative engine typically retrieves relevant content fragments from accessible sources and uses these fragments as the basis for generating an answer, referencing or paraphrasing them within the response. The sources accessible to a generative engine are diverse, including web pages, technical documents, industry reports, and user reviews.

[0039] Therefore, in this application, target content refers to the content that the content provider hopes will be retrieved by the generative engine and cited in its answers. The source of this target content is diverse; for example, it can be existing technical blog posts, product introduction articles, technical white papers, or industry reports. The topics covered by the target content can also be diverse; for example, it can be a functional description of a technology, or an explanation of whether a technology can be used in a certain scenario.

[0040] In one possible implementation, obtaining the target content includes: obtaining original content; when the original content is non-text modal content, performing cross-modal conversion on the original content to obtain text content corresponding to the original content, and determining the text content as the target content; wherein, the non-text modal content includes at least one of image content, audio content, and video content.

[0041] Original content refers to content provided by the content provider that has not yet undergone modal identification and transformation. The modal form of original content is diverse, and it can be text modal content, such as a text article; or it can be non-text modal content, such as an image, an audio clip, a video clip, or a PDF document.

[0042] Cross-modal conversion refers to the process of converting non-textual content into text content. Different processing methods can be used for different modalities of the original content. For example, when the original content is an image, text can be extracted using Optical Character Recognition (OCR), or a text description can be generated from the image content to obtain the corresponding text content. When the original content is audio, speech can be converted into text using Automatic Speech Recognition (ASR). When the original content is video, speech can be extracted from the video and converted into text using ASR, or subtitles, bullet comments, and other text information can be extracted to obtain the corresponding text content. When the original content is a PDF document, text can be extracted through document parsing.

[0043] Furthermore, cross-modal conversion can be achieved using a multimodal parsing suite, which supports the conversion of various content formats into text content. This multimodal parsing suite supports a wide variety of content formats, such as short video comments, live stream interactive text, and video subtitles.

[0044] In some possible implementations, for knowledge graph-style content, entity relationship extraction can be used to convert the entities and relationships between them into text content. For long text content (e.g., text with approximately 200,000 tokens), it can be segmented through long text parsing and then converted into text content. Here, a token refers to the basic unit obtained after text segmentation; a token can be a character, a word, or a subword.

[0045] This approach allows for the conversion of original content—whether it's non-textual modal content like images, audio, video, or PDF documents, or content in the form of short video comments, live stream interactive words, knowledge graphs, or long text—into text content through cross-modal transformation and incorporated into subsequent optimization processes. This expands the modal range of content that can be optimized, enabling more content formats to be retrieved and referenced by the generative engine. When the original content itself is textual, it can be directly identified as the target content without the need for cross-modal transformation.

[0046] S102: Reconstruct the target content into multiple structured content units; wherein each structured content unit includes an interrelated question part and a content part supporting the question part, the content part being determined from the target content and corresponding to the question part.

[0047] A structured content unit refers to a content fragment with a specific organizational form obtained after reconstructing target content. The process of reconstructing target content into structured content units can also be called modular knowledge distillation, which involves reconstructing technical blog posts, long articles, and other content into a collection of multiple question-and-answer format content units. Each structured content unit includes an interrelated question section and a content section. The question section is the part expressed in question form that guides the content of the structured content unit, such as the question "Can a certain technology be used in industrial scenarios?" The content section is the part determined from the target content that responds to the question section; the content section corresponds to the question section, meaning it is the response to the question posed in the question section.

[0048] It should be noted that there are various ways to reconstruct target content into structured content units. For example, based on the different aspects covered by the target content, it can be broken down into multiple content fragments. For each content fragment, a corresponding question can be generated, and this question can be used as the question part, while the content fragment can be used as the content part, thus forming a structured content unit. Alternatively, multiple questions can be predetermined, and for each question, a corresponding content fragment can be selected from the target content as the content part.

[0049] By reconstructing the target content into structured content units that include question and content sections, the organization of the target content is adapted to the logic of users asking questions to the generative engine, and the generative engine retrieving and integrating answers based on these questions. This makes it easier for the generative engine to retrieve content fragments related to the query and to reference the target content.

[0050] In one possible implementation, reconstructing the target content into multiple structured content units includes:

[0051] From the multi-source data, target multi-source data associated with the target content are filtered out; wherein, the multi-source data includes at least one of the generative engine's search logs, e-commerce review data, and social media data;

[0052] Based on the target multi-source data, the predicted query information is determined;

[0053] Based on the predicted question information, the question portion in the structured content unit is determined, and the content portion corresponding to the question portion is determined from the target content.

[0054] Multi-source data refers to data from multiple different sources, including at least one of the following: search logs from generative engines, e-commerce review data, and social media data. Search logs are records generated when users submit queries on generative engines, revealing the actual queries users submit, such as frequently requested queries. E-commerce review data consists of user evaluations of products on e-commerce platforms, revealing the issues users are concerned about. Social media data refers to content posted by users on social media platforms, such as topics and discussions, revealing the relevant needs users are discussing. Target multi-source data refers to data filtered from multi-source data that is relevant to the target content. There are various ways to filter target multi-source data. For example, based on keywords related to the target content, data containing or semantically related to those keywords can be filtered from multi-source data as target multi-source data. Alternatively, the similarity between multi-source data and the target content can be determined, and multi-source data whose similarity meets preset conditions can be identified as target multi-source data.

[0055] Predictive question information refers to information about the questions users might ask the generative engine, predicted based on target multi-source data. For example, predictive question information could be a specific question a user might ask, or a set of keywords that might be involved in a user's question.

[0056] After determining the predicted question information, the question portion of the structured content unit can be determined based on this information. For example, the question represented by the predicted question information can be used directly as the question portion, or a question can be generated based on the keywords represented by the predicted question information. After determining the question portion, the corresponding content segment is selected from the target content and used as the content portion corresponding to that question portion.

[0057] By determining the predicted query information based on multi-source data, and accordingly determining the question part of the structured content unit, the question part of the structured content unit can closely resemble the questions that users might actually ask the generative engine. As a result, when users submit relevant queries, the structured content unit is more easily retrieved and referenced by the generative engine.

[0058] In some possible implementations, a corpus can be constructed based on the aforementioned multi-source data, and the corpus can be updated as the multi-source data is updated, so that the corpus can dynamically change with the changes in data sources.

[0059] In one possible implementation, determining the predictive query information based on the target multi-source data includes at least one of the following methods:

[0060] Method 1: Cluster the target multi-source data to obtain associated word clusters, and determine the predicted question information based on the associated word clusters; wherein, the associated word clusters are a set of semantically related words.

[0061] In this context, a word cluster refers to a set of semantically related words; that is, multiple words within a word cluster are semantically related to each other. Clustering target multi-source data refers to grouping semantically similar content in the target multi-source data into one category. Clustering can be implemented using various clustering algorithms. For example, the K-means clustering algorithm can be used to divide the words in the target multi-source data into several clusters according to their semantic similarity, with each cluster being a word cluster. Specifically, the K-means clustering algorithm can represent the words in the target multi-source data as vectors, pre-determine the number of clusters K, select K initial cluster centers from the word vectors, and assign the remaining word vectors to the clusters corresponding to their nearest cluster centers. Then, the cluster centers of each cluster are updated based on the mean of the word vectors in each cluster. This process of division and updating is repeated until the cluster centers no longer change significantly, thus obtaining K word clusters. Alternatively, hierarchical clustering and other clustering algorithms can also be used. After obtaining the related word clusters, predictive question information can be generated based on the words in the related word clusters. For example, words in a related word cluster can be combined into a question as predictive question information.

[0062] Method 2: Determine the predicted question information based on at least one of the occurrence frequency of question data and the co-occurrence relationship of question data in the target multi-source data.

[0063] In this context, "question data" refers to data representing user questions within the target multi-source data. The frequency of question data occurrence refers to the number of times a particular question appears in the target multi-source data; the co-occurrence relationship of question data refers to the relationship between multiple questions appearing together in the target multi-source data. For example, frequently occurring question data can be identified as predicted question information because a higher frequency indicates a greater likelihood that the question was asked by a user. Similarly, based on the co-occurrence relationship of question data, frequently occurring questions can be grouped together to determine predicted question information.

[0064] By determining the predicted question information through the above method, we can predict the questions that users may ask from the data actually generated by users, making the determined question part closer to the actual questions asked by users, thereby increasing the probability of structured content units being retrieved and cited.

[0065] In one possible implementation, each of the structured content units includes an evidence section and a conclusion section; wherein the evidence section contains evidence data to support the conclusion section, and the conclusion section is a conclusion given regarding the question section.

[0066] The conclusion section provides a definitive answer to the question posed in the question section. For example, regarding the question "Can a certain technology be used in industrial scenarios?", the conclusion could be "The technology can be used in industrial scenarios." The evidence section contains supporting data for the conclusion. Evidence data includes data that supports the conclusion, such as experimental data, application cases, and cited sources. By further dividing the content section into evidence and conclusion sections, the structured content unit adopts a three-part organizational form consisting of the question section, evidence section, and conclusion section. This allows the content section to provide both a conclusion on the question and supporting evidence, making it easier for generative engines to identify and cite relevant, well-founded content during retrieval.

[0067] In some possible implementations, when reconstructing structured content units, structured tags can be added to these units to make it easier for the generative engine to parse their composition. For example, JSON-LD (JavaScript Object Notation for Linked Data) can be used to mark the question, evidence, and conclusion sections of the structured content unit, making these sections clearly distinguishable in terms of data format.

[0068] S103: Publish the structured content unit to the information source platform so that the generative engine can obtain the structured content unit from the information source platform.

[0069] A source platform refers to a platform capable of publishing content that can be retrieved by generative engines. Source platforms are diverse; for example, they can be platforms for publishing technical white papers and patent documents, media articles, or Q&A content. After a structured content unit is published to a source platform, the generative engine can retrieve information from the source platform when responding to queries, thus obtaining the structured content unit and referencing it when generating an answer.

[0070] When publishing structured content units, you can publish them to one source platform or multiple source platforms. For example, you can publish the same structured content unit to multiple different types of source platforms to increase its probability of being retrieved by generative engines.

[0071] In some possible implementations, different types of structured content units can be published to different types of information source platforms. For example, structured content units with technical descriptions as their content part can be published to information source platforms used for publishing technical white papers, patent documents, and media articles; structured content units with more detailed and long-tailed question parts (e.g., the question part is "Can a certain technology be used in industrial scenarios") can be published to information source platforms used for publishing question-and-answer content.

[0072] In some possible implementations, the publication of structured content units can also be performed through dynamic data interfaces. For example, structured content units can be synchronized to the source platform via an Application Programming Interface (API), enabling the content on the source platform to be updated according to a preset synchronization cycle, thereby allowing the published structured content units to be updated along with the target content or multi-source data.

[0073] S104: Obtain the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer.

[0074] After structured content units are published to the information source platform, queries can be submitted to the generative engine, and the engine can generate answers to those queries. These queries can be related to the target content, such as queries constructed based on the question portion of the structured content units.

[0075] In this context, citation feedback information refers to information determined by the generative engine's response that reflects how structured content units are cited by the generative engine. There are various ways to determine citation feedback information; for example, the content of the response can be compared with the content of the structured content unit, and the comparison result can be used to determine whether and how the structured content unit has been cited.

[0076] In one possible implementation, the citation feedback information includes at least one of the following: citation identifier information indicating whether the structured content unit is cited in the answer, citation fragment information in the answer referencing the structured content unit, and citation position information indicating whether the structured content unit is cited in the answer. The citation identifier information indicates whether the structured content unit is cited in the answer; for example, it can be represented by an identifier value, where a first value indicates citation, and a second value indicates no citation. The citation fragment information refers to a specific segment in the answer that references the structured content unit; for example, if a paragraph in the answer originates from a part of a structured content unit, then that paragraph is the citation fragment information. The citation position information indicates the position in the answer where the structured content unit is cited; for example, whether the citation fragment is located at the beginning, middle, or end of the answer.

[0077] By obtaining the aforementioned citation feedback information, we can reflect the citation status of structured content units from multiple dimensions, such as whether they are cited, which fragments are cited, and the citation location, providing a basis for subsequent adjustments.

[0078] S105: Based on the reference feedback information, adjust at least one of the reconstruction method of the structured content unit and the publishing method of the structured content unit;

[0079] If the reconstruction method of the structured content unit is adjusted, the reconstruction is re-executed based on the adjusted reconstruction method;

[0080] If the publishing method of the structured content unit is adjusted, the publishing is re-executed based on the adjusted publishing method.

[0081] Reconstruction method refers to the approach used when reconstructing target content into structured content units, such as the method for determining the problem section or the method for determining the content section. Publishing method refers to the approach used when publishing the structured content units to the information source platform, such as the information source platform to which the content is published.

[0082] When making adjustments based on citation feedback, at least one of the reconstruction and publishing methods can be adjusted according to the citation status reflected in the feedback. For example, when the citation feedback indicates that the structured content unit is not cited or the citation status is unsatisfactory, the reconstruction method can be adjusted to make the reconstructed structured content unit easier to cite. Similarly, when the citation feedback indicates that the structured content unit published on some source platforms is cited well while the citation status on others is poor, the publishing method can be adjusted to ensure that the structured content unit is subsequently published to the source platforms with better citation status.

[0083] In some possible implementations, since the retrieval and generation rules adopted by the generative engine may change over time, the responses from the generative engine can be continuously obtained and citation feedback information can be determined, thereby continuously adjusting the reconstruction and publishing methods to adapt them to the current retrieval and generation rules of the generative engine.

[0084] In one possible implementation, the publishing method includes the information source platforms to which the structured content unit is published; adjusting the publishing method of the structured content unit based on the citation feedback information includes: determining the information source weight of each information source platform based on the citation feedback information; and determining the target information source platform for subsequent publishing of the structured content unit based on the information source weight.

[0085] In this context, source weight refers to the weight value used to characterize the degree to which structured content units published on a source platform are referenced by the generative engine. A higher source weight indicates a higher degree of citation for the structured content units published on that source platform. There are various ways to determine source weight. For example, the number or proportion of citations for structured content units published on each source platform can be statistically analyzed, and the corresponding source weight can be determined based on the number or proportion of citations; the higher the number or proportion of citations, the higher the source weight. Since the citation status of the generative engine may change over time, the source weight can be dynamically updated based on the updates to citation feedback information, thereby achieving dynamic management of the source weight.

[0086] In this context, the target source platform refers to the source platform used to subsequently publish structured content units. There are various ways to determine the target source platform based on source weight. For example, one or more source platforms with high source weight can be designated as target source platforms, thus ensuring that subsequent structured content units are published on these more frequently cited platforms. This approach tilts subsequent content publishing towards more frequently cited source platforms, thereby increasing the probability that the structured content units will be cited by the generative engine.

[0087] In one possible implementation, adjusting the reconstruction method of the structured content unit based on the reference feedback information includes: when the reference feedback information indicates that the structured content unit is not referenced by the answer, adjusting at least one of the question part and the content part in the structured content unit to obtain an updated structured content unit, and publishing the updated structured content unit to the information source platform.

[0088] For example, when feedback indicates that a structured content unit has not been cited in an answer, the question section of that unit can be adjusted to better reflect the user's actual question; the content section can also be adjusted, such as adding evidence data or revising the conclusion; or both sections can be adjusted simultaneously. The resulting updated structured content unit is then published to the information source platform. This method allows uncited structured content units to be republished after adjustment, thereby increasing their probability of being cited.

[0089] If the reconstruction method is adjusted, the reconstruction steps of S102 are re-executed based on the adjusted reconstruction method to obtain the reconstructed structured content units, and subsequent steps such as publishing and feedback acquisition can continue to be executed.

[0090] If the publishing method is adjusted, the publishing steps of S103 are re-executed based on the adjusted publishing method to publish the structured content units to the source platform determined after the adjustment.

[0091] By adjusting at least one of the reconstruction and publishing methods based on citation feedback information and re-executing the corresponding steps, content optimization forms a closed loop, which can continuously iterate with the citation situation of the generative engine, reduce the deviation between the content optimization method and the actual citation preference of the generative engine, and thus further improve the citation rate of the target content in the generative engine's answers.

[0092] In some possible implementations, a Mixture of Experts (MoE) architecture can be used when processing structured content units. This architecture includes multiple expert modules and a gating module. Each expert module is used to process one type of content or perform one type of processing, while the gating module determines which expert modules(s) should handle the input content. For example, different expert modules can be assigned to different modalities of content, and the gating module can allocate the content to the appropriate expert module based on its modality. This hybrid expert architecture reduces the computational load required when processing multiple types of content by only invoking a subset of expert modules for a given input, rather than all modules.

[0093] In some possible implementations, when the model involved in the reconstruction or release method needs to be updated, federated learning can be used for model updates. Federated learning refers to a method where, without centrally aggregating the original data from all parties, each party trains its model locally based on its local data, and then updates and aggregates the trained model parameters to obtain the updated model. Updating the model through federated learning can be completed without centrally aggregating the original data from all parties.

[0094] The generative engine content optimization method provided in this embodiment reconstructs the target content into structured content units that include interrelated question parts and content parts, and publishes them to the information source platform. This adapts the organization of the target content to the retrieval and citation logic of the generative engine, increasing the probability of the target content being retrieved and cited. At the same time, based on citation feedback information, at least one of the reconstruction method and the publication method is adjusted and re-executed, enabling continuous iteration of content optimization and further improving the citation rate of the target content in the generative engine's answers.

[0095] Example 2:

[0096] Based on the above embodiments, the content optimization method for generative engines provided in this application will be described below through a specific embodiment. Figure 2 This is a schematic diagram illustrating a specific generative engine content optimization process provided in an embodiment of this application. (Refer to...) Figure 2 The process includes:

[0097] S201: Obtain the original content. When the original content is non-text modal content, perform cross-modal conversion on the original content to obtain text content and determine it as the target content; when the original content is text modal content, determine the original content as the target content.

[0098] Cross-modal conversion can be achieved through a multimodal parsing suite, supporting the conversion of various content formats, such as images, audio, video, PDF documents, short video bullet comments, live interactive words, knowledge graphs, and long texts, into text content.

[0099] S202: Select target multi-source data that is relevant to the target content from multi-source data including at least one of generative engine search logs, e-commerce review data, and social media data.

[0100] S203: Cluster the target multi-source data to obtain associated word clusters, and determine the predicted question information based on the associated word clusters.

[0101] S204: Determine the question section based on the predicted question information, determine the content section corresponding to the question section from the target content, and form a structured content unit by combining the question section and the content section; wherein, the content section includes the evidence section and the conclusion section.

[0102] S205: Publish the structured content units to the information source platform.

[0103] S206: Obtain the response generated by the generative engine for the query, and determine the reference feedback information of the structured content units referenced based on the response.

[0104] S207: Determine whether the reference feedback information indicates that the structured content unit is referenced. If not, proceed to S208; if yes, proceed to S209.

[0105] S208: Adjust at least one of the problem part and the content part in the structured content unit to obtain an updated structured content unit, and return to execute S205.

[0106] S209: Determine the source weight of each source platform based on the citation feedback information, and determine the target source platform for subsequent release of structured content units based on the source weight.

[0107] The generative engine content optimization method provided in this embodiment performs cross-modal conversion on the original non-textual modality, determines the predicted question information based on multi-source data to reconstruct structured content units, adjusts the question part and content part based on citation feedback information, and determines the target information source platform. This enables content optimization to cover content of multiple modalities and continuously iterate with the citation situation, thereby improving the citation rate of the target content in the generative engine's answers.

[0108] Example 3:

[0109] Based on the same inventive concept, embodiments of this application provide a generative engine content optimization device. Figure 3 This is a schematic diagram of the structure of a generative engine content optimization device provided in an embodiment of this application, with reference to... Figure 3 The device includes:

[0110] The acquisition unit 31 is used to acquire the target content to be retrieved and referenced by the generative engine.

[0111] The content reconstruction unit 32 is used to reconstruct the target content into multiple structured content units; wherein each structured content unit includes an interrelated question part and a content part supporting the question part, and the content part is determined from the target content and corresponds to the question part;

[0112] Content publishing unit 33 is used to publish the structured content unit to the information source platform so that the generative engine can obtain the structured content unit from the information source platform;

[0113] Feedback acquisition unit 34 is used to acquire the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer;

[0114] The adjustment unit 35 is used to adjust at least one of the reconstruction method and the publishing method of the structured content unit based on the reference feedback information; when the reconstruction method is adjusted, the content reconstruction unit is triggered to re-execute the reconstruction based on the adjusted reconstruction method; when the publishing method is adjusted, the content publishing unit is triggered to re-execute the publishing based on the adjusted publishing method.

[0115] In this embodiment, the generative engine content optimization device is presented in the form of a functional unit. Here, a unit refers to an application-specific integrated circuit (ASIC), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0116] The further functional descriptions of each of the above units are the same as those in the corresponding embodiments described above, and will not be repeated here. The generative engine content optimization device provided in this embodiment, through the cooperation of each unit, reconstructs the target content into structured content units and publishes them to the information source platform. Furthermore, it adjusts and re-executes based on citation feedback information, thereby improving the citation rate of the target content in the generative engine's answers.

[0117] Example 4:

[0118] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of this application, such as... Figure 4As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0119] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0120] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0121] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device as shown by a landing page for an app. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0122] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0123] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 4 Taking the example of a connection between China and Israel via a bus.

[0124] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0125] Example 5:

[0126] Based on the above embodiments, this application also provides a computer-readable storage medium storing a computer program executable by a processor. When the program runs on the processor, it causes the processor to perform the following steps:

[0127] Retrieve the target content to be retrieved and referenced by the generative engine;

[0128] The target content is reconstructed into multiple structured content units; wherein each structured content unit includes an interrelated question section and a content section supporting the question section, the content section being determined from the target content and corresponding to the question section;

[0129] The structured content unit is published to the information source platform so that the generative engine can obtain the structured content unit from the information source platform;

[0130] Obtain the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer;

[0131] Based on the reference feedback information, at least one of the reconstruction method of the structured content unit and the publishing method of the structured content unit is adjusted;

[0132] If the reconstruction method of the structured content unit is adjusted, the reconstruction is re-executed based on the adjusted reconstruction method;

[0133] If the publishing method of the structured content unit is adjusted, the publishing is re-executed based on the adjusted publishing method.

[0134] Since the principle of the computer-readable storage medium in solving the problem is similar to that of the generative engine content optimization method, the implementation of the computer-readable storage medium can be found in the embodiments of the method, and repeated details will not be repeated.

[0135] The storage media described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memory.

[0136] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this application can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0137] Although this application has been described in detail above with general descriptions and specific embodiments, some modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of this application fall within the scope of protection claimed in this application.

Claims

1. A content optimization method for a generative engine, characterized in that, The method includes: Retrieve the target content to be retrieved and referenced by the generative engine; The target content is reconstructed into multiple structured content units; wherein each structured content unit includes an interrelated question section and a content section supporting the question section, the content section being determined from the target content and corresponding to the question section; The structured content unit is published to the information source platform so that the generative engine can obtain the structured content unit from the information source platform; Obtain the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer; Based on the reference feedback information, at least one of the reconstruction method of the structured content unit and the publishing method of the structured content unit is adjusted; If the reconstruction method of the structured content unit is adjusted, the reconstruction is re-executed based on the adjusted reconstruction method; If the publishing method of the structured content unit is adjusted, the publishing is re-executed based on the adjusted publishing method.

2. The method according to claim 1, characterized in that, The step of reconstructing the target content into multiple structured content units includes: From the multi-source data, target multi-source data associated with the target content are filtered out; wherein, the multi-source data includes at least one of the generative engine's search logs, e-commerce review data, and social media data; Based on the target multi-source data, the predicted query information is determined; Based on the predicted question information, the question portion in the structured content unit is determined, and the content portion corresponding to the question portion is determined from the target content.

3. The method according to claim 2, characterized in that, The determination of the predicted query information based on the target multi-source data includes at least one of the following methods: Clustering is performed on the target multi-source data to obtain associated word clusters, and the predicted question information is determined based on the associated word clusters; wherein, the associated word clusters are a set of semantically related words; The predicted question information is determined based on at least one of the occurrence frequency of question data in the target multi-source data and the co-occurrence relationship of question data.

4. The method according to any one of claims 1-3, characterized in that, In each of the structured content units, the content portion includes an evidence portion and a conclusion portion; wherein, the evidence portion contains evidence data to support the conclusion portion, and the conclusion portion is the conclusion given in response to the question portion.

5. The method according to any one of claims 1-3, characterized in that, The citation feedback information includes at least one of the following: citation identifier information indicating whether the structured content unit is cited by the answer, citation fragment information in the answer citing the structured content unit, and citation position information of the structured content unit being cited in the answer.

6. The method according to any one of claims 1-3, characterized in that, The publishing method includes the information source platform to which the structured content unit is published; The adjustment of the publishing method of the structured content unit based on the reference feedback information includes: Based on the aforementioned reference feedback information, the source weights of each source platform are determined; Based on the aforementioned source weights, the target source platforms for subsequently publishing the structured content units are determined.

7. The method according to any one of claims 1-3, characterized in that, The acquisition of target content includes: Get the original content; When the original content is non-text modal content, cross-modal conversion is performed on the original content to obtain the text content corresponding to the original content, and the text content is determined as the target content; The non-text modal content includes at least one of image content, audio content, and video content.

8. The method according to any one of claims 1-3, characterized in that, The adjustment of the reconstruction method of the structured content unit based on the reference feedback information includes: When the citation feedback information indicates that the structured content unit is not cited by the answer, at least one of the question part and the content part in the structured content unit is adjusted to obtain an updated structured content unit, and the updated structured content unit is published to the information source platform.

9. A content optimization device for a generative engine, characterized in that, The device includes: The acquisition unit is used to acquire the target content to be retrieved and referenced by the generative engine. A content reconstruction unit is used to reconstruct the target content into multiple structured content units; wherein each structured content unit includes an interrelated question part and a content part supporting the question part, the content part being determined from the target content and corresponding to the question part; A content publishing unit is used to publish the structured content unit to the information source platform so that the generative engine can obtain the structured content unit from the information source platform; The feedback acquisition unit is used to acquire the answer generated by the generative engine for the query, and determine the reference feedback information of the structured content unit being referenced by the generative engine based on the answer; An adjustment unit is configured to adjust at least one of the reconstruction method and the publishing method of the structured content unit based on the reference feedback information; when the reconstruction method is adjusted, the content reconstruction unit is triggered to re-execute the reconstruction based on the adjusted reconstruction method; when the publishing method is adjusted, the content publishing unit is triggered to re-execute the publishing based on the adjusted publishing method.

10. A computer device, characterized in that, It includes a processor and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the executable instructions to implement the generative engine content optimization method as described in any one of claims 1-8.

Images