A content adaptive distribution and ranking dynamic monitoring feedback method for a generative intelligent engine

Abstract

The application is suitable for the technical field of information processing and data monitoring, and provides a content self-adaptive distribution and ranking dynamic monitoring feedback method for a generative intelligent engine, which is executed by a content distribution monitoring server and includes the following steps: S1, obtaining to-be-distributed content and a target problem scene set, dividing the to-be-distributed content into multiple content segments, and configuring segment identifiers and content type identifiers of the content segments; the application first divides the to-be-distributed content into multiple content segments which can independently participate in backtracking matching, then generates segment feature loads containing segment identifiers, content type identifiers, segment feature vectors, segment check codes and version identifiers for the content segments, and further establishes segment node mapping records after the target distribution node completes publishing, so that an association relationship which can be verified among answer texts, content segments, publishing nodes and publishing resource addresses is formed.

Description

Technical Field

[0001] This invention belongs to the field of information processing and data monitoring technology, and in particular relates to a method for adaptive content distribution and dynamic ranking monitoring and feedback for generative intelligent engines. Background Technology

[0002] As generative intelligent engines are increasingly used in information retrieval, content question answering, and technical document querying, users can obtain generated answers from multiple sources simply by asking natural language queries. These generated answers typically include the corresponding textual conclusions, source URLs, and the order in which the content appears. Existing content distribution methods usually publish complete content to multiple content nodes according to a fixed template, and then evaluate the content's effectiveness based on page views, link clicks, or the occurrence of a single answer. This approach is ill-suited to the dynamic answer processing characteristics of generative intelligent engines, which generate answers based on different question expressions, different sampling times, and different content sources.

[0003] In practical applications, the same content to be distributed may be broken down, paraphrased, or published to multiple content nodes. The answer text returned by the generative intelligent engine may only reference part of the technical content, or only show the source address without presenting the original content in its entirety. When there is a lack of corresponding records between content fragments and publishing nodes, the monitoring end has difficulty in accurately determining which content fragment and which publishing node the specific text in the generated answer comes from. Consequently, it is impossible to judge the actual contribution of different distribution nodes to the generated answer result. At the same time, the generative intelligent engine may generate different answers for the same question scenario under different query expressions, different sampling times, and different engine conditions. If only a single generated answer is evaluated, it is easy to mistake an accidental reference for a stable reference and a single non-reference for content invalidation, resulting in insufficient accuracy of the feedback results.

[0004] Furthermore, existing content adjustment methods typically modify and resend content uniformly based on whether the entire page is referenced or whether the content appears. They lack a retrospective analysis mechanism for specific content fragments, specific publishing nodes, and specific answer positions. When a content fragment is never included in the answer generation process, or although it is referenced, it appears in a later position and has low cross-engine stability, the existing methods cannot specifically determine the content objects and distribution nodes that need adjustment. This can easily lead to problems such as increased duplicate content transmission, invalid nodes continuing to receive data, and low feedback adjustment efficiency. Summary of the Invention

[0005] This invention provides a method for adaptive content distribution and dynamic ranking monitoring and feedback for generative intelligent engines, aiming to solve the problems of accuracy and stability in identifying the citation status of generated answers, ranking content contributions, and adjusting subsequent distribution.

[0006] This invention is implemented as follows: a content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines, executed by a content distribution monitoring server, includes the following steps:

[0007] S1. Obtain the content to be distributed and the target problem scenario set, divide the content to be distributed into multiple content fragments, and configure the fragment identifier and content type identifier of each content fragment;

[0008] S2. Extract the content features of each content segment, form a segment feature vector corresponding to each content segment, and generate a segment feature payload containing the segment feature vector;

[0009] S3. Based on the target problem scenario set, content type identifier, and node interface description data of multiple distribution nodes, determine at least one target distribution node corresponding to each content fragment, convert the content fragment containing the fragment feature payload into a node adaptation data packet, and send it to the corresponding target distribution node.

[0010] S4. Receive the published resource address returned by the target distribution node, and establish a fragment node mapping record between the content fragment, the target distribution node, and the published resource address;

[0011] S5. Establish a query expression group corresponding to the target problem scenario set, and input the query expression group into multiple generative intelligent engines at multiple sampling times to obtain generated answer data;

[0012] S6. Extract answer evidence units from the generated answer data, perform address backtracking matching and content backtracking matching between the answer evidence units and the fragment feature payload and fragment node mapping records, and record the valid reference status of content fragment records that meet the matching conditions and the repeated occurrence conditions.

[0013] S7. Determine the fragment contribution ranking value of each content fragment based on the effective reference status, and generate the ranking monitoring result of the content to be distributed;

[0014] S8. Determine the feedback correction parameters for each target distribution node based on the valid reference status with address backtracking matching results and the ranking monitoring results, adjust the next round of node adaptation data packets according to the feedback correction parameters, and send them to the corresponding target distribution node.

[0015] Preferably, dividing the content to be distributed into multiple content segments includes: identifying the title level position, semantic topic change position, parameter description position, condition limitation position, and conclusion description position in the content to be distributed; and dividing the content to be distributed into multiple content segments that can independently participate in backtracking matching based on the identification results.

[0016] The content type identifiers include technical principle type identifiers, parameter limitation type identifiers, application scenario type identifiers, comparison conclusion type identifiers, processing step type identifiers, and data basis type identifiers.

[0017] Preferably, the content features include entity features, relationship features between entities, parameter combination features, and conclusion expression features;

[0018] The step of forming a segment feature vector corresponding to each of the content segments includes: extracting the content features from each of the content segments, and standardizing the synonym expressions, parameter units and conditional expressions in the content features to obtain the segment feature vector;

[0019] The step of generating a fragment feature payload containing fragment feature vectors includes: configuring a version identifier for each content fragment, generating a fragment checksum based on the fragment identifier, fragment feature vector, and version identifier, wherein the fragment feature payload is composed of the fragment identifier, content type identifier, fragment feature vector, fragment checksum, and version identifier;

[0020] The segment verification code is used to determine the version status of the content segment and the backtracking matching object, and the segment feature vector is used to perform content backtracking matching with the answer evidence unit.

[0021] Preferably, the node interface description data includes a node type identifier, a data field template, a data length limit, a structured markup format, a content publishing interface address, and a publishing result return field;

[0022] The process of generating the node adaptation data packet includes: configuring the content fragment containing the fragment feature payload into the title field, summary field, body paragraph field, parameter field, and source field according to the data field template corresponding to the target distribution node; encapsulating the configured field data according to the structured markup format; validating the encapsulated field data according to the data length limit value, and forming the node adaptation data packet after the verification passes.

[0023] The fragment node mapping record includes a fragment identifier, a fragment checksum, a version identifier, a target distribution node identifier, and a normalized publishing address after address normalization processing.

[0024] Preferably, the query expression group includes a baseline query expression, a semantically equivalent query expression, a conditional constraint query expression, and an object comparison query expression corresponding to the same target problem scenario;

[0025] The content distribution monitoring server configures query identifiers for each query expression and inputs the query expression group into multiple generative intelligent engines through a query interface according to a preset sampling period.

[0026] The generated answer data includes answer text, source address, answer location data, generative intelligent engine identifier, sampling time identifier, and query identifier, to form an answer data set across generative intelligent engines and sampling times under the same target question scenario.

[0027] Preferably, the answer evidence unit includes an answer text fragment, a source address, an answer paragraph number, a citation sequence number, answer text features, a generative intelligent engine identifier, a sampling time identifier, and a query identifier;

[0028] The answer location data is determined based on the answer paragraph number and the citation arrangement number;

[0029] The address backtracking matching includes: performing address normalization processing on the reference source address, deleting the dynamic query field in the reference source address, unifying the address protocol identifier and domain name character format, extracting the resource path identifier, and forming a normalized reference address; performing the same address normalization processing on the published resource address in the fragment node mapping record to form a normalized published address; and comparing the normalized reference address with the normalized published address to obtain an address matching value.

[0030] Preferably, the content backtracking matching includes: extracting answer entity features, answer relationship features, answer parameter features, and answer conclusion features from the answer text fragment to form an answer feature vector; obtaining the fragment feature vector from the fragment feature payload; calculating the similarity between the answer feature vector and the fragment feature vector to obtain a content matching value; and determining a backtracking matching confidence value based on the address matching value and the content matching value.

[0031] When the content matching value reaches the preset content matching threshold, and the corresponding content fragment forms a matching result in the generated answer data at at least two sampling times, or forms a matching result in the generated answer data of at least two generative intelligent engines, the content fragment is recorded as a valid reference fragment.

[0032] When the address matching value corresponding to the valid reference fragment reaches the preset address matching threshold, the valid reference fragment is recorded as a node traceable valid reference fragment;

[0033] The valid reference fragment is used to determine the fragment contribution ranking value, and the node can backtrack the valid reference fragment to determine the feedback correction parameters of the target distribution node.

[0034] Preferably, the fragment contribution ranking value is determined according to the following formula:

[0035]

[0036] in, For the first The contribution ranking value of each content fragment. For the first Normalized value of effective citation frequency for each content fragment For the first Normalized values ​​of the answer positions for each content segment For the first The consistency value of a content fragment across the generative intelligent engine. For the first The duration value of a content segment across sampling times. For the first The average confidence score of the backtracking match for each content fragment. , , , and Let be the weight parameters, and satisfy:

[0037]

[0038] The normalized value of effective citation frequency is determined based on the ratio of the number of effective citations of the corresponding content fragment to the total number of sampled answers; the normalized value of answer position is determined based on the answer paragraph number and citation arrangement number of the corresponding effective citation fragment; the consistency value across generative intelligent engines is determined based on the ratio of the number of generative intelligent engines that form effective citation states to the total number of generative intelligent engines participating in sampling; and the persistence value across sampling times is determined based on the ratio of the number of sampling times that form effective citation states to the total number of sampling times.

[0039] Preferably, for those capable of locating the first [location] through address backtracking matching. The nodes of the target distribution nodes can trace back to valid reference fragments, respectively determining the th node before and after this round of distribution. The content fragment is relative to the first The node-backtrackable fragment contribution value of each target distribution node;

[0040] The first The feedback correction parameters for each target distribution node are determined according to the following formula:

[0041]

[0042] in, For the first Feedback correction parameters for each target distribution node For the first The content fragment and the first Distribution correspondence identifiers between target distribution nodes This is the [number]th distribution after this round. The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. The first in this round of distribution The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. For the first The normalized value of the amount of duplicate content data received by each target distribution node. This is the data suppression coefficient for duplicate content. For distribution to the first The number of content fragments per target distribution node;

[0043] The contribution value of the node's traceable fragment is determined based on the valid reference fragment with address matching results.

[0044] Preferably, adjusting the next round of node adaptation data packets based on feedback correction parameters includes:

[0045] When the feedback correction parameter of the target distribution node is greater than the first correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is increased, and the node interface format selection parameter corresponding to the target distribution node is retained.

[0046] When the feedback correction parameter of the target distribution node is less than the second correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is reduced, and the target distribution node corresponding to the content segment is re-determined based on the feedback correction parameters of other target distribution nodes.

[0047] When the contribution ranking value of the same content fragment is lower than the preset contribution threshold in multiple consecutive sampling periods, the answer text fragment with the highest content matching value is selected from the generated answer data corresponding to the content fragment, the answer feature vector corresponding to the answer text fragment is extracted, and the content adjustment object is determined according to the difference between the answer feature vector and the fragment feature vector corresponding to the content fragment. When there is an answer text fragment that forms a valid reference state with the content fragment, the answer feature vector is extracted from the answer text fragment that forms a valid reference state first.

[0048] The content fragment is updated according to the content adjustment object, and the fragment feature payload and node adaptation data packet corresponding to the updated content fragment are regenerated. The regenerated node adaptation data packet is sent to the newly determined target distribution node, and the fragment node mapping record is updated according to the published resource address returned by the newly determined target distribution node.

[0049] Compared with related technologies, the content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines provided by this invention has the following beneficial effects:

[0050] 1. This invention improves the accuracy of backtracking the source of generated answers by establishing a correspondence between content fragments, fragment feature payloads, target distribution nodes, and publishing resource addresses. Existing content distribution methods typically publish and monitor complete pages or articles. When the same content is published by multiple nodes or a portion of the text is extracted by a generative intelligent engine to form an answer, it is difficult to accurately determine the specific source of the content in the answer. This invention first divides the content to be distributed into multiple content fragments that can independently participate in backtracking matching. Then, it generates fragment feature payloads for each content fragment, including fragment identifier, content type identifier, fragment feature vector, fragment checksum, and version identifier. After the target distribution node completes the publication, it further establishes a fragment node mapping record, forming a verifiable association between the answer text, content fragments, publishing nodes, and publishing resource addresses. Thus, in subsequent monitoring, it is possible to accurately determine the specific content fragments cited or paraphrased in the generated answer and their corresponding publishing sources, reducing backtracking errors caused by repeated publication by multiple nodes or changes in content versions.

[0051] 2. This invention improves the reliability of dynamic answer monitoring results from generative intelligent engines by combining address backtracking matching and content backtracking matching. Answers generated by generative intelligent engines may display source addresses, merely paraphrase content conclusions, or exhibit fluctuations in results at different sampling times. Relying solely on source addresses or text similarity alone is insufficient to fully reflect the content citation status. This invention compares the standardized citation addresses in the generated answers with the standardized publication addresses in the fragment node mapping records to determine the specific target distribution node. Furthermore, it calculates the similarity between the answer feature vector and the fragment feature vector to determine the degree of association between the generated answer text and the original content fragment. It further distinguishes between valid citation fragments and node-trackable valid citation fragments based on recurrence conditions. This processing method can identify answer results that do not explicitly display source addresses but actually cite relevant content, and it provides clear address location criteria when adjusting specific distribution nodes, thereby improving the stability of monitoring results and the accuracy of feedback control.

[0052] 3. This invention uses fragment contribution ranking values ​​to form a dynamic ranking monitoring method for generated answer results, reducing evaluation bias caused by a single monitoring indicator. This invention does not evaluate results solely based on whether a content fragment appears, but comprehensively considers effective citation frequency, answer appearance position, consistency across generative intelligent engines, persistence across sampling times, and backtracking matching reliability to determine the fragment contribution ranking value for each content fragment. Citation frequency characterizes the degree to which a content fragment is used in the generated answer; answer position characterizes the visibility order of the corresponding content in the generated answer; cross-engine consistency and persistence across time periods reflect the stability of the answer results; and backtracking matching reliability reflects the reliability of the monitoring results. Through the combined evaluation of these indicators, the impact of changes in single query results, accidental citations, or source identification errors on the ranking monitoring results can be reduced, making the generated ranking monitoring results more suitable for controlling subsequent content distribution actions.

[0053] 4. This invention controls the sending process of the next round of node adaptation data packets by using feedback correction parameters, reducing invalid content transmission and improving the targeting of distribution adjustments. When determining the feedback correction parameters for the target distribution node, this invention only uses the traceable valid reference fragments of nodes that can be located to the specific target distribution node through address backtracking matching as the basis for node contribution. It also combines the changes in fragment contribution before and after this round of distribution and the amount of duplicate content data for feedback correction. When the target distribution node generates a high and stable content contribution, the sending priority of the corresponding content fragment of that node is increased; when the target distribution node's contribution is low or the amount of duplicate content data is high, the sending priority of that node is reduced and a new target distribution node is selected; when the content fragment contribution is insufficient for a long period of time, the content fragment is further adjusted according to the corresponding text features in the generated answer and redistributed. This avoids continuously sending the same content to low-contribution nodes, reduces the data processing burden caused by repeated transmission, and improves the accuracy of determining the subsequent distribution path and content adjustment object.

[0054] 5. This invention ensures the consistency of monitored objects during multiple rounds of feedback by updating the content version status and publication address mapping records. After a content fragment is adjusted and redistributed, the text content, feature vector, and publication address of the original content fragment may change. If the original matching object is still used for answer monitoring, it is easy to mix the reference results of different versions, resulting in distortion of the fragment contribution ranking value and node feedback correction parameters. This invention determines the version status of the content fragment through version identifier and fragment check code, and updates the fragment node mapping record according to the newly published resource address returned by the target distribution node after the updated node adaptation data packet is resent. This ensures that the monitored object, content version, and publication node corresponding to each round of sampling remain distinguishable. Therefore, this invention can support multi-round continuous monitoring and feedback distribution processing, and improve the accuracy of data statistics and the ability to reproduce results during long-term operation. Attached Figure Description

[0055] Figure 1 This is an overall flowchart of the content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to the present invention.

[0056] Figure 2 This is a schematic diagram of the feedback adjustment processing logic of the present invention;

[0057] Figure 3 This is a flowchart illustrating the backtracking matching and valid reference determination logic of the present invention.

[0058] Figure 4 This is a diagram of the content distribution monitoring server architecture for this invention.

[0059] Figure 5 A trend graph showing the change of the ranking value of the fragment contributing to the present invention with the sampling time. Detailed Implementation

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0061] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0062] Example

[0063] A preferred embodiment of the content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines provided by this invention is as follows: Figures 1 to 5 As shown: A content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines, executed by a content distribution monitoring server, is characterized by including the following steps:

[0064] S1. Obtain the set of content to be distributed and target problem scenarios, divide the content to be distributed into multiple content fragments, and configure the fragment identifier and content type identifier for each content fragment;

[0065] S2. Extract the content features of each content segment, form the segment feature vector corresponding to each content segment, and generate the segment feature payload containing the segment feature vector;

[0066] S3. Based on the target problem scenario set, content type identifier, and node interface description data of multiple distribution nodes, determine at least one target distribution node corresponding to each content fragment, convert the content fragment containing fragment feature payload into a node adaptation data packet, and send it to the corresponding target distribution node.

[0067] S4. Receive the published resource address returned by the target distribution node, and establish a fragment node mapping record between the content fragment, the target distribution node, and the published resource address;

[0068] S5. Establish a set of query expressions corresponding to the target problem scenario set, and input the query expression sets into multiple generative intelligent engines at multiple sampling times to obtain generated answer data;

[0069] S6. Extract answer evidence units from the generated answer data, perform address backtracking and content backtracking matching between the answer evidence units and the fragment feature payload and fragment node mapping records, and record the effective reference status of content fragment records that meet the matching conditions and the recurrence conditions.

[0070] S7. Determine the fragment contribution ranking value of each content fragment based on the effective citation status, and generate the ranking monitoring results of the content to be distributed;

[0071] S8. Determine the feedback correction parameters for each target distribution node based on the valid reference status with address backtracking matching results and the ranking monitoring results. Adjust the next round of node adaptation data packets according to the feedback correction parameters and send them to the corresponding target distribution nodes.

[0072] In this embodiment, the content distribution monitoring server establishes data communication connections with multiple distribution nodes and multiple generative intelligent engines. After obtaining the content to be distributed and the target problem scenario set, it processes the content to be distributed into multiple content fragments with fragment identifiers and content type identifiers, generates corresponding fragment feature payloads, and forms node adaptation data packets based on node interface description data, which are then sent to the target distribution node. The server receives the published resource address returned by the target distribution node and establishes a fragment node mapping record. At multiple sampling times, it sends query expression groups to multiple generative intelligent engines to obtain generated answer data, extracts answer evidence units for address backtracking matching and content backtracking matching, determines the effective reference status, fragment contribution ranking value, and ranking monitoring results, and then determines feedback correction parameters based on the effective reference status that can be located to a specific target distribution node. It adjusts the sending object and sending priority of the next round of node adaptation data packets, thereby forming a closed-loop processing process of content distribution, answer monitoring, source backtracking, and feedback redistribution.

[0073] In a further preferred embodiment of the present invention, the content to be distributed is divided into multiple content segments, including: identifying the title level position, semantic topic change position, parameter description position, condition limitation position and conclusion description position in the content to be distributed; and dividing the content to be distributed into multiple content segments that can independently participate in backtracking matching based on the identification results.

[0074] Content type identifiers include technical principle type identifiers, parameter limitation type identifiers, application scenario type identifiers, comparison conclusion type identifiers, processing step type identifiers, and data basis type identifiers.

[0075] In this embodiment, the content distribution monitoring server identifies the title level position, semantic topic change position, parameter description position, condition limitation position, and conclusion description position in the content to be distributed, and forms content segments accordingly. Each content segment contains at least an independently identifiable technical object, processing relationship, condition information, or result information. When a segmented content lacks a referred object or the processing relationship is incomplete, the content segment is merged with adjacent related segments. The server configures content type identifiers based on the content expressed by the segments. Among them, the technical principle type identifier corresponds to the function relationship or implementation process, the parameter limitation type identifier corresponds to the parameter and value conditions, the application scenario type identifier corresponds to the usage environment, the comparison conclusion type identifier corresponds to the difference result, the processing step type identifier corresponds to the operation process, and the data basis type identifier corresponds to the detection data or verification result.

[0076] In a further preferred embodiment of the present invention, the content features include entity features, relationship features between entities, parameter combination features, and conclusion expression features;

[0077] Forming segment feature vectors corresponding to each content segment includes: extracting content features from each content segment, standardizing synonyms, parameter units, and conditional expressions in the content features, and obtaining segment feature vectors;

[0078] Generate a fragment feature payload containing fragment feature vectors, including: configuring a version identifier for each content fragment, generating a fragment checksum based on the fragment identifier, fragment feature vector, and version identifier, and the fragment feature payload consisting of the fragment identifier, content type identifier, fragment feature vector, fragment checksum, and version identifier;

[0079] The fragment checksum is used to determine the version status of the content fragment and the backtracking matching object, while the fragment feature vector is used to perform content backtracking matching with the answer evidence unit.

[0080] In this embodiment, the content distribution monitoring server extracts entity features, relation features, parameter combination features, and conclusion expression features from each content fragment. Synonymous expressions with the same meaning are converted into a unified standard expression, parameter values ​​and units are converted into preset unit formats, and conditional expressions are converted into unified condition fields. The server encodes the entity feature fields, relation feature fields, parameter combination feature fields, and conclusion expression feature fields in a preset order, converting each field into a corresponding numerical feature item. Fields for which no corresponding feature is extracted are configured with zero values, and the numerical feature items form a fragment feature vector. The server configures a version identifier for each content fragment and generates a fragment verification code based on the fragment identifier, fragment feature vector, and version identifier. When a content fragment is updated, the version identifier and fragment verification code are updated synchronously to distinguish the monitoring results corresponding to different versions.

[0081] In a further preferred embodiment of the present invention, the node interface description data includes a node type identifier, a data field template, a data length limit, a structured markup format, a content publishing interface address, and a publishing result return field;

[0082] The process of generating a node adaptation data packet includes: configuring content fragments containing fragment feature payloads into the title field, summary field, body paragraph field, parameter field, and source field according to the data field template corresponding to the target distribution node; encapsulating the configured field data according to the structured markup format; validating the encapsulated field data according to the data length limit; and forming a node adaptation data packet after the validation passes.

[0083] The fragment node mapping record includes fragment identifier, fragment checksum, version identifier, target distribution node identifier, and normalized publishing address after address normalization.

[0084] In this embodiment, the node interface description data includes a node type identifier, a data field template, a data length limit, a structured markup format, a content publishing interface address, and a publishing result return field. The content distribution monitoring server, based on the target distribution node's data field template, configures the content fragment containing the fragment feature payload to the corresponding fields, encapsulates it according to the structured markup format, and performs length verification to form a node adaptation data packet. After the target distribution node completes publishing, it returns the publishing resource address. The server removes the tracking parameters and session parameters from the publishing resource address, unifies the HTTP and HTTPS address representations and domain name character formats, extracts the page resource path, and converts the redirect address into the final landing address to obtain a standardized publishing address. The server associates and saves the fragment identifier, fragment verification code, version identifier, target distribution node identifier, and standardized publishing address as a fragment node mapping record.

[0085] In a further preferred embodiment of the present invention, the query expression group includes a baseline query expression, a semantically equivalent query expression, a conditional constraint query expression, and an object comparison query expression corresponding to the same target problem scenario;

[0086] The content distribution monitoring server configures query identifiers for each query expression and inputs the query expression groups into multiple generative intelligent engines through the query interface according to the preset sampling period;

[0087] The generated answer data includes answer text, source URL, answer location data, generative intelligent engine identifier, sampling time identifier, and query identifier, to form a set of answer data across generative intelligent engines and across sampling times for the same target question scenario.

[0088] In this embodiment, the content distribution monitoring server establishes a query expression group for the same target problem scenario. The query expression group includes a baseline query expression, a semantically equivalent query expression, a conditional constraint query expression, and an object comparison query expression, and configures a query identifier for each query expression. The server inputs the query expression group into multiple generative intelligent engines at multiple sampling times according to a preset sampling period to obtain the answer text, the reference source address, the answer location data, the generative intelligent engine identifier, the sampling time identifier, and the query identifier. The server aggregates the generated answer data according to the target problem scenario identifier and the query identifier, so that the results returned by different generative intelligent engines under the same target problem scenario and the results returned at different sampling times can be compared accordingly.

[0089] In a further preferred embodiment of the present invention, the answer evidence unit includes an answer text fragment, a source address, an answer paragraph number, a citation sequence number, an answer text feature, a generative intelligent engine identifier, a sampling time identifier, and a query identifier;

[0090] The answer location data is determined based on the paragraph number and the citation sequence number.

[0091] Address backtracking matching includes: normalizing the source address, removing dynamic query fields from the source address, unifying the address protocol identifier and domain name character format, extracting resource path identifiers, and forming a normalized reference address; performing the same address normalization on the published resource addresses in the fragment node mapping record to form a normalized published address; and comparing the normalized reference address with the normalized published address to obtain the address matching value.

[0092] In this embodiment, the content distribution monitoring server extracts answer text fragments, citation source addresses, answer paragraph numbers, citation order numbers, answer text features, generative intelligent engine identifiers, sampling time identifiers, and query identifiers from the generated answer data to form answer evidence units. The answer paragraph numbers and citation order numbers are used to characterize the order in which the answer text fragments appear. The server performs the same normalization processing on the citation source address as on the published resource address to obtain a normalized citation address, and compares it with the normalized published address. When the domain name identifier and resource path identifier of the normalized citation address and the normalized published address are both consistent, the address matching value is determined as the first matching value. When the domain name identifier is consistent and the resource path identifier is partially consistent, the address matching value is determined as the second matching value, which is less than the first matching value. When neither the domain name identifier nor the resource path identifier is consistent, the address matching value is determined as zero.

[0093] In a further preferred embodiment of the present invention, content backtracking matching includes: extracting answer entity features, answer relationship features, answer parameter features, and answer conclusion features from the answer text fragment to form an answer feature vector; obtaining the fragment feature vector from the fragment feature payload; calculating the similarity between the answer feature vector and the fragment feature vector to obtain a content matching value; and determining a backtracking matching confidence value based on the address matching value and the content matching value.

[0094] When the content matching value reaches the preset content matching threshold, and the corresponding content fragment forms a matching result in the generated answer data at at least two sampling times, or forms a matching result in the generated answer data of at least two generative intelligent engines, the content fragment is recorded as a valid reference fragment.

[0095] When the address matching value corresponding to the valid reference fragment reaches the preset address matching threshold, the valid reference fragment is recorded as a node traceable valid reference fragment;

[0096] Valid reference fragments are used to determine the fragment contribution ranking value, and the node can trace back to valid reference fragments to determine the feedback correction parameters of the target distribution node.

[0097] In this embodiment, the content distribution monitoring server extracts answer entity features, answer relationship features, answer parameter features, and answer conclusion features from the answer text fragment. It forms an answer feature vector according to the same field order as the fragment feature vector and calculates its similarity with the fragment feature vector in the fragment feature payload to obtain a content matching value. The backtracking matching confidence value is the sum of the address matching value and the content matching value calculated according to a preset weight, where the preset weight sum is 1. When the answer evidence unit does not contain a source address, the address matching value is configured to zero. When the content matching value reaches a preset content matching threshold, and the corresponding content fragment satisfies the condition of repeated occurrence across sampling times or across generative intelligent engines, it is recorded as a valid reference fragment. When the address matching value of a valid reference fragment reaches a preset address matching threshold, it is recorded as a node-backtrackable valid reference fragment. The former is used for sorting calculations, and the latter is used for node feedback calculations.

[0098] In a further preferred embodiment of the present invention, the fragment contribution ranking value is determined according to the following formula:

[0099]

[0100] in, For the first The contribution ranking value of each content fragment. For the first Normalized value of effective citation frequency for each content fragment For the first Normalized values ​​of the answer positions for each content segment For the first The consistency value of a content fragment across the generative intelligent engine. For the first The duration value of a content segment across sampling times. For the first The average confidence score of the backtracking match for each content fragment. , , , and Let be the weight parameters, and satisfy:

[0101]

[0102] The normalized value of effective citation frequency is determined based on the ratio of the number of effective citations of the corresponding content fragment to the total number of sampled answers; the normalized value of answer position is determined based on the answer paragraph number and citation sequence number of the corresponding effective citation fragment; the consistency value across generative intelligent engines is determined based on the ratio of the number of generative intelligent engines that form effective citation states to the total number of generative intelligent engines participating in sampling; and the persistence value across sampling times is determined based on the ratio of the number of sampling times that form effective citation states to the total number of sampling times.

[0103] In this embodiment, the normalized value of effective citation frequency is the ratio of the number of times the corresponding content fragment forms an effective citation state to the total number of sampled answers; the normalized value of answer position is determined by subtracting one from the sum of the answer paragraph number and the citation arrangement number corresponding to the effective citation fragment as the single answer position value, and taking the arithmetic mean of multiple single answer position values ​​as the answer position normalized value; the consistency across generative intelligent engines is the ratio of the number of generative intelligent engines forming an effective citation state to the total number of generative intelligent engines participating in the sampling; the persistence across sampling time is the ratio of the number of sampling time moments forming an effective citation state to the total number of sampling time moments; the average backtracking matching confidence value is the arithmetic mean of the backtracking matching confidence values ​​corresponding to each answer evidence unit of the corresponding effective citation fragment; the same weight parameters and normalization rules are used in the same round of ranking calculation.

[0104] In a further preferred embodiment of the present invention, for those capable of locating the first [location] through address backtracking matching The nodes of the target distribution nodes can trace back to valid reference fragments, respectively determining the th node before and after this round of distribution. The content fragment is relative to the first The node-backtrackable fragment contribution value of each target distribution node;

[0105] No. The feedback correction parameters for each target distribution node are determined according to the following formula:

[0106]

[0107] in, For the first Feedback correction parameters for each target distribution node For the first The content fragment and the first Distribution correspondence identifiers between target distribution nodes This is the [number]th distribution after this round. The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. The first in this round of distribution The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. For the first The normalized value of the amount of duplicate content data received by each target distribution node. This is the data suppression coefficient for duplicate content. For distribution to the first The number of content fragments per target distribution node;

[0108] The contribution value of a node's traceable fragment is determined based on a valid reference fragment with address matching results.

[0109] In this embodiment, the content fragments participating in the feedback correction parameter calculation are limited to valid reference fragments that can be traced back to the node with address matching results; when the content fragment has been sent to the first... When a target distribution node is identified and an address matching result is formed for that node, its relationship with the first target distribution node is determined. Each target distribution node has a distribution correspondence; the node's contribution value before and after this round of distribution can be traced back to the fragment contribution value, which is determined based on the answer data set corresponding to the same target problem scenario, the same number of query expressions, the same number of generative intelligent engines, and the same number of sampling times; the normalized value of the duplicate content data volume is the ratio of the duplicate content data volume with the same fragment check code and unchanged version identifier sent to the corresponding target distribution node within a feedback calculation cycle to the total amount of data sent.

[0110] In a further preferred embodiment of the present invention, adjusting the next round of node adaptation data packets based on feedback correction parameters includes:

[0111] When the feedback correction parameter of the target distribution node is greater than the first correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is increased, and the node interface format selection parameter corresponding to the target distribution node is retained.

[0112] When the feedback correction parameter of the target distribution node is less than the second correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is reduced, and the target distribution node corresponding to the content segment is re-determined based on the feedback correction parameters of other target distribution nodes.

[0113] When the contribution ranking value of the same content fragment is lower than the preset contribution threshold in multiple consecutive sampling periods, the answer text fragment with the highest content matching value is selected from the generated answer data corresponding to the content fragment, the answer feature vector corresponding to the answer text fragment is extracted, and the content adjustment object is determined according to the difference between the answer feature vector and the fragment feature vector corresponding to the content fragment. Among them, when there is an answer text fragment that forms a valid reference state with the content fragment, the answer feature vector is extracted from the answer text fragment that forms a valid reference state first.

[0114] The content fragment is updated according to the content adjustment object. The fragment feature payload and node adaptation data packet corresponding to the updated content fragment are regenerated. The regenerated node adaptation data packet is sent to the newly determined target distribution node. The fragment node mapping record is updated according to the published resource address returned by the newly determined target distribution node.

[0115] In this embodiment, the content distribution monitoring server compares the feedback correction parameters with a preset correction threshold. When the feedback correction parameters are greater than the first correction threshold, the sending priority of the corresponding content segment is increased while retaining the corresponding node interface format selection parameters. When the feedback correction parameters are less than the second correction threshold, the sending priority of the corresponding content segment is decreased and the target distribution node is re-determined based on the feedback correction parameters of other target distribution nodes. When the segment contribution ranking value of a content segment is lower than the preset contribution threshold in multiple consecutive sampling periods, the server selects the answer text segment with the highest content matching value from the corresponding generated answer data. When a valid reference segment exists, the valid reference segment is selected first. Based on the difference between its answer feature vector and segment feature vector, the server determines the entity features, relationship features, parameter combination features, or conclusion expression features to be supplemented or adjusted. The server updates the content segment accordingly, regenerates the segment feature payload and node adaptation data packet, completes the retransmission, and updates the segment node mapping record.

[0116] In summary, this invention is executed by a content distribution monitoring server. After obtaining the content to be distributed and the target problem scenario set, the server first divides the content to be distributed into multiple content segments based on the position of title level, semantic topic, parameter description, condition limitation and conclusion description, and configures segment identifier and content type identifier for each content segment. Subsequently, the server extracts entity features, relation features, parameter combination features and conclusion expression features from each content segment, and after standardization processing, forms a segment feature vector. Combined with the version identifier, a segment verification code is generated, thereby forming the segment feature payload.

[0117] Based on the target problem scenario, content type identifier, and node interface description data of the distribution node, the server determines the target distribution node corresponding to each content fragment, converts the content fragment containing the fragment feature payload into a node adaptation data packet and sends it to the corresponding target distribution node. After the target distribution node completes the publication, it returns the publication resource address. The server performs normalization processing on the publication resource address and establishes a fragment node mapping record between the fragment identifier, version identifier, target distribution node, and normalized publication address.

[0118] During the monitoring phase, the server establishes a query expression group corresponding to the target problem scenario and inputs the query expression group into multiple generative intelligent engines at multiple sampling times to obtain generated answer data. The server extracts answer evidence units from the generated answer data, performs address backtracking matching on the source address, and performs content backtracking matching on the answer feature vector formed by the answer text fragment and the fragment feature vector to determine the valid reference fragment and the node that can be traced back to the valid reference fragment.

[0119] The server calculates the contribution ranking value of each content fragment based on the valid reference status and generates ranking monitoring results. Based on the valid reference status with address backtracking matching results, it determines the feedback correction parameters for each target distribution node. When the feedback correction parameter is high, the sending priority of the corresponding node is increased; when the feedback correction parameter is low, the sending priority of the corresponding node is decreased and a new target distribution node is selected. For content fragments with low contribution ranking values ​​over multiple consecutive sampling periods, the server updates the content fragment based on the difference between the answer feature vector and the fragment feature vector, regenerates the fragment feature payload and node adaptation data packet, and updates the fragment node mapping record, thus forming a closed-loop processing process of content adaptive distribution, dynamic ranking monitoring, and feedback adjustment.

[0120] It is worth noting that the circuits, electronic components, and modules involved in this invention are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The content protected by this invention does not involve improvements to the software and methods.

[0121] It should be understood that the disclosed apparatus can be implemented in other ways, given the several embodiments provided in this application. For example, the apparatus embodiments described above are merely illustrative; the division of units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or communication connections shown or discussed may be through some interfaces; the indirect coupling or communication connections between devices or units may be telecommunications or other forms.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A content adaptive distribution and ranking dynamic monitoring feedback method for generative intelligent engine, executed by a content distribution monitoring server, characterized in that, Includes the following steps: S1. Obtain the content to be distributed and the target problem scenario set, divide the content to be distributed into multiple content fragments, and configure the fragment identifier and content type identifier of each content fragment; S2. Extract the content features of each content segment, form a segment feature vector corresponding to each content segment, and generate a segment feature payload containing the segment feature vector; S3. Based on the target problem scenario set, content type identifier, and node interface description data of multiple distribution nodes, determine at least one target distribution node corresponding to each content fragment, convert the content fragment containing the fragment feature payload into a node adaptation data packet, and send it to the corresponding target distribution node. S4. Receive the published resource address returned by the target distribution node, and establish a fragment node mapping record between the content fragment, the target distribution node, and the published resource address; S5. Establish a query expression group corresponding to the target problem scenario set, and input the query expression group into multiple generative intelligent engines at multiple sampling times to obtain generated answer data; S6. Extract answer evidence units from the generated answer data, perform address backtracking matching and content backtracking matching between the answer evidence units and the fragment feature payload and fragment node mapping records, and record the valid reference status of content fragment records that meet the matching conditions and the repeated occurrence conditions. S7. Determine the fragment contribution ranking value of each content fragment based on the effective reference status, and generate the ranking monitoring result of the content to be distributed; S8. Determine the feedback correction parameters for each target distribution node based on the valid reference status with address backtracking matching results and the ranking monitoring results, adjust the next round of node adaptation data packets according to the feedback correction parameters, and send them to the corresponding target distribution node.

2. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 1, characterized in that: The step of dividing the content to be distributed into multiple content segments includes: identifying the title level position, semantic topic change position, parameter description position, condition limitation position, and conclusion description position in the content to be distributed; and dividing the content to be distributed into multiple content segments that can independently participate in backtracking matching based on the identification results. The content type identifiers include technical principle type identifiers, parameter limitation type identifiers, application scenario type identifiers, comparison conclusion type identifiers, processing step type identifiers, and data basis type identifiers.

3. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 1, characterized in that: The content features include entity features, relationship features between entities, parameter combination features, and conclusion expression features; The step of forming a segment feature vector corresponding to each of the content segments includes: extracting the content features from each of the content segments, and standardizing the synonym expressions, parameter units and conditional expressions in the content features to obtain the segment feature vector; The step of generating a fragment feature payload containing fragment feature vectors includes: configuring a version identifier for each content fragment, generating a fragment checksum based on the fragment identifier, fragment feature vector, and version identifier, wherein the fragment feature payload is composed of the fragment identifier, content type identifier, fragment feature vector, fragment checksum, and version identifier; The segment verification code is used to determine the version status of the content segment and the backtracking matching object, and the segment feature vector is used to perform content backtracking matching with the answer evidence unit.

4. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 3, characterized in that: The node interface description data includes node type identifier, data field template, data length limit, structured markup format, content publishing interface address, and publishing result return field; The process of generating the node adaptation data packet includes: configuring the content fragment containing the fragment feature payload into the title field, summary field, body paragraph field, parameter field, and source field according to the data field template corresponding to the target distribution node; encapsulating the configured field data according to the structured markup format; validating the encapsulated field data according to the data length limit value, and forming the node adaptation data packet after the verification passes. The fragment node mapping record includes a fragment identifier, a fragment checksum, a version identifier, a target distribution node identifier, and a normalized publishing address after address normalization processing.

5. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 1, characterized in that: The query expression group includes a baseline query expression, a semantically equivalent query expression, a conditional constraint query expression, and an object comparison query expression corresponding to the same target problem scenario; The content distribution monitoring server configures query identifiers for each query expression and inputs the query expression group into multiple generative intelligent engines through a query interface according to a preset sampling period. The generated answer data includes answer text, source address, answer location data, generative intelligent engine identifier, sampling time identifier, and query identifier, to form an answer data set across generative intelligent engines and sampling times under the same target question scenario.

6. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 5, characterized in that: The answer evidence unit includes an answer text fragment, a source address, an answer paragraph number, a citation sequence number, answer text features, a generative intelligent engine identifier, a sampling time identifier, and a query identifier; The answer location data is determined based on the answer paragraph number and the citation arrangement number; The address backtracking matching includes: performing address normalization processing on the reference source address, deleting the dynamic query field in the reference source address, unifying the address protocol identifier and domain name character format, extracting the resource path identifier, and forming a normalized reference address; performing the same address normalization processing on the published resource address in the fragment node mapping record to form a normalized published address; and comparing the normalized reference address with the normalized published address to obtain an address matching value.

7. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 6, characterized in that: The content backtracking matching includes: extracting answer entity features, answer relationship features, answer parameter features, and answer conclusion features from the answer text fragment to form an answer feature vector; obtaining the fragment feature vector from the fragment feature payload; calculating the similarity between the answer feature vector and the fragment feature vector to obtain a content matching value; and determining a backtracking matching confidence value based on the address matching value and the content matching value. When the content matching value reaches the preset content matching threshold, and the corresponding content fragment forms a matching result in the generated answer data at at least two sampling times, or forms a matching result in the generated answer data of at least two generative intelligent engines, the content fragment is recorded as a valid reference fragment. When the address matching value corresponding to the valid reference fragment reaches the preset address matching threshold, the valid reference fragment is recorded as a node traceable valid reference fragment; The valid reference fragment is used to determine the fragment contribution ranking value, and the node can backtrack the valid reference fragment to determine the feedback correction parameters of the target distribution node.

8. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 7, characterized in that: The contribution ranking value of the fragment is determined according to the following formula: in, For the first The contribution ranking value of each content fragment. For the first Normalized value of effective citation frequency for each content fragment For the first Normalized values ​​of the answer positions for each content segment For the first The consistency value of a content fragment across the generative intelligent engine. For the first The duration value of a content segment across sampling times. For the first The average confidence score of the backtracking match for each content fragment. , , , and Let be the weight parameters, and satisfy: The normalized value of effective citation frequency is determined based on the ratio of the number of effective citations of the corresponding content fragment to the total number of sampled answers; the normalized value of answer position is determined based on the answer paragraph number and citation arrangement number of the corresponding effective citation fragment; the consistency value across generative intelligent engines is determined based on the ratio of the number of generative intelligent engines that form effective citation states to the total number of generative intelligent engines participating in sampling; and the persistence value across sampling times is determined based on the ratio of the number of sampling times that form effective citation states to the total number of sampling times.

9. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 8, characterized in that: For those who can locate the first [address] through address backtracking and matching The nodes of the target distribution nodes can trace back to valid reference fragments, respectively determining the th node before and after this round of distribution. The content fragment is relative to the first The node-backtrackable fragment contribution value of each target distribution node; The first The feedback correction parameters for each target distribution node are determined according to the following formula: in, For the first Feedback correction parameters for each target distribution node For the first The content fragment and the first Distribution correspondence identifiers between target distribution nodes This is the [number]th distribution after this round. The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. The first in this round of distribution The content fragment is relative to the first Each target distribution node can trace back the contribution value of a fragment. For the first The normalized value of the amount of duplicate content data received by each target distribution node. This is the coefficient for suppressing duplicate content data. For distribution to the The number of content fragments per target distribution node; The contribution value of the node's traceable fragment is determined based on the valid reference fragment with address matching results.

10. The content adaptive distribution and ranking dynamic monitoring and feedback method for generative intelligent engines according to claim 9, characterized in that: The step of adjusting the next round of node adaptation data packets based on feedback correction parameters includes: When the feedback correction parameter of the target distribution node is greater than the first correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is increased, and the node interface format selection parameter corresponding to the target distribution node is retained. When the feedback correction parameter of the target distribution node is less than the second correction threshold, the sending priority of the content segment corresponding to the target distribution node in the next round of distribution is reduced, and the target distribution node corresponding to the content segment is re-determined based on the feedback correction parameters of other target distribution nodes. When the contribution ranking value of the same content fragment is lower than the preset contribution threshold in multiple consecutive sampling periods, the answer text fragment with the highest content matching value is selected from the generated answer data corresponding to the content fragment, the answer feature vector corresponding to the answer text fragment is extracted, and the content adjustment object is determined according to the difference between the answer feature vector and the fragment feature vector corresponding to the content fragment. When there is an answer text fragment that forms a valid reference state with the content fragment, the answer feature vector is extracted from the answer text fragment that forms a valid reference state first. The content fragment is updated according to the content adjustment object, and the fragment feature payload and node adaptation data packet corresponding to the updated content fragment are regenerated. The regenerated node adaptation data packet is sent to the newly determined target distribution node, and the fragment node mapping record is updated according to the published resource address returned by the newly determined target distribution node.

Images