Aigc content organization distribution method and system

Abstract

The application discloses an AIGC content organization distribution method and system, and belongs to the technical field of content organization distribution; the method specifically comprises the following steps: collecting an AIGC content evolution data set and a source content change record; based on the AIGC content evolution data set, a content evolution correlation relationship is constructed to obtain a content evolution relationship set; based on the content evolution relationship set and the source content change record, a change semantic inheritance knowledge graph is constructed, residual amount reservation type inheritance matching is performed, a key semantic inheritance relationship set and a content set affected by changes are obtained; based on the key semantic inheritance relationship set and the content set affected by changes, change influence driven content organization distribution updating is performed to obtain an updated content organization relationship set and a content distribution disposal result; the application improves the accuracy of AIGC content organization distribution and enhances the stability of AIGC content organization distribution.

Description

Technical Field

[0001] This invention belongs to the field of content organization and distribution, specifically relating to AIGC content organization and distribution methods and systems. Background Technology

[0002] In the process of organizing and distributing AIGC content, the system needs to organize internally generated text, Q&A, promotional materials, training materials, customer service scripts, and content displayed on multiple platforms. It then retrieves, reuses, and distributes this content based on business scenarios, target audiences, and distribution platforms. AIGC content is typically generated repeatedly around the same business theme, undergoing partial rewriting, summary compression, style adjustments, and cross-platform adaptation to form a content system where source content, derived content, and distributed content coexist. These contents not only exhibit thematic similarity but also semantic continuity, extending from source content to derived content and from derived content to distributed content.

[0003] In existing technologies, AIGC content organization and distribution methods are typically based on tag classification, topic categorization, keyword matching, or overall semantic similarity to archive, retrieve, and call AIGC content. However, this approach primarily relies on the overall similarity of content for organization, making it difficult to identify whether derived and distributed content specifically inherits the changed key semantics from the source content. When product descriptions, business rules, scope of application, or core expressions in the source content are adjusted, some derived and distributed content, while still thematically similar to the source content, may no longer be compatible with the new source content due to outdated semantics. Existing methods struggle to accurately distinguish such affected content from other content on the same topic, resulting in the continued retrieval, reuse, or distribution of invalid or incompatible AIGC content, leading to poor accuracy and stability in AIGC content organization and distribution. Summary of the Invention

[0004] The purpose of this invention is to provide an AIGC content organization and distribution method and system. By constructing a change impact propagation graph based on key semantic inheritance relationships and introducing a residual retention weak inheritance propagation mechanism, the method achieves semantic impact tracking and hierarchical updates across the generation link, thereby improving the accuracy and stability of AIGC content organization and distribution.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] On one hand, the present invention provides an AIGC content organization and distribution method, including:

[0007] Collect AIGC content evolution dataset and source content change records. The AIGC content evolution dataset includes source content, derived content, distributed content, generation records, and adaptation records.

[0008] Based on the AIGC content evolution dataset, we construct content evolution relationships to obtain a content evolution relationship set;

[0009] Based on the content evolution relationship set and source content change records, a change semantic inheritance knowledge graph is constructed, and residual retention inheritance matching is performed to obtain the key semantic inheritance relationship set and the content set affected by the change.

[0010] Based on the key semantic inheritance relationship set and the content set affected by the change, the content organization and distribution are updated driven by the change, resulting in the updated content organization relationship set and content distribution processing results.

[0011] Specifically, based on the AIGC content evolution dataset, content evolution relationships are constructed to obtain a content evolution relationship set, including:

[0012] Based on the generation records, the generation association between the source content and the derived content, as well as the regeneration association between each derived content, are established to obtain the generation relationship set;

[0013] Based on the adaptation record, establish cross-platform adaptation relationships between derived content and distributed content to obtain a set of adaptation relationships;

[0014] Based on the generation relation set and the adaptation relation set, the evolution chain of source content, derived content and distributed content is organized to obtain the content evolution relation set.

[0015] Specifically, based on the content evolution relationship set and source content change records, a change semantic inheritance knowledge graph is constructed, and residual retention inheritance matching is performed to obtain the key semantic inheritance relationship set and the content set affected by the change, including:

[0016] Based on the source content change record, extract the content fragments before and after the change corresponding to the source content, and perform semantic difference recognition to obtain the set of key semantic units of change and the corresponding set of semantic change degree;

[0017] Based on the content evolution relationship set, the content objects that have downstream evolution associations with each key semantic unit of change are identified, and the inheriting semantics are extracted to obtain the downstream associated content object set and the candidate inheriting semantic unit set.

[0018] Based on the set of key semantic units for change, the set of downstream related content objects, and the set of candidate semantic units for inheritance, a knowledge graph of semantic inheritance for change is constructed.

[0019] Based on the knowledge graph of semantic connection, directional semantic connection metric learning is performed to obtain the source-to-source connection matching cost;

[0020] Based on the source-to-end matching cost and semantic change degree set, we perform end-to-end quality diversion and residual retention semi-constrained transmission matching, and weak explicit end-to-end aggregation to obtain the key semantic inheritance relationship set and the content set affected by changes.

[0021] Specifically, based on the set of key semantic units of change, the set of downstream related content objects, and the set of candidate inheriting semantic units, a knowledge graph of change semantic inheritance is constructed, including:

[0022] Based on the set of key semantic units for change, the set of candidate inheriting semantic units and the set of downstream related content objects, node mapping and semantic content attribution are performed to obtain the set of semantic nodes, the set of content nodes and the set of semantic content attribution relationships.

[0023] Based on the content evolution relationship set, content evolution edges are established between each content node to obtain the content evolution edge set;

[0024] Based on the semantic content attribution relation set, content inclusion relation edges are established to obtain the content inclusion relation edge set;

[0025] Based on the source content change records, establish source-side semantic change relationship edges to obtain a set of semantic change relationship edges;

[0026] Based on semantic node set, content node set, content evolution edge set, content inclusion relation edge set, and semantic change relation edge set, a change semantic inheritance knowledge graph is constructed.

[0027] Specifically, based on the semantic connection knowledge graph, directional semantic connection metric learning is performed to obtain the source-to-source connection matching cost, including:

[0028] Based on the semantic knowledge graph of change and the content evolution relationship set, the content evolution direction is determined, and the semantic connection relationship between each key semantic unit of change and each candidate semantic unit of connection is identified along the content evolution direction to obtain the positive connection sample set.

[0029] Based on the forward acceptance sample set, the semantic correspondence between the candidate acceptance semantic units and the corresponding change key semantic units is determined as the reverse pseudo acceptance sample, thus obtaining the reverse pseudo acceptance sample set.

[0030] Based on the positive succession sample set, candidate succession semantic units that are on the same topic as each change key semantic unit but do not form a semantic succession relationship are screened to obtain the non-succession sample set on the same topic;

[0031] Based on the positive acceptance sample set, the reverse pseudo acceptance sample set, and the non-acceptance sample set of the same topic, the sample acceptance cost ranking is determined, and a directional acceptance measurement target is constructed.

[0032] Based on the directional acceptance metric objective, metric learning is performed on the set of change key semantic units and the set of candidate acceptance semantic units to obtain the source-to-source acceptance matching cost.

[0033] Specifically, based on the source-to-receive matching cost and the set of semantic change degrees, reception quality splitting and residual retention semi-constrained transmission matching are performed, followed by weak explicit reception aggregation to obtain a set of key semantic inheritance relationships and a set of content affected by changes, including:

[0034] Based on the source-to-source matching cost and the set of semantic change degrees, the direct acceptance quality and acceptance residual quality corresponding to each key semantic unit of change are determined.

[0035] Based on the candidate acceptance semantic unit set, direct acceptance quality and source-to-acceptance matching cost, a set of acceptance and receiving objects containing residual acceptance objects is constructed, and semi-constrained optimal transmission matching is performed to obtain the direct acceptance matching result and the residual to be aggregated.

[0036] Based on the quality of the inherited residual and the residual to be aggregated, the downstream content evolution edge is determined, and weak explicit inheritance and aggregation are carried out along the downstream content evolution edge to obtain the residual inheritance result.

[0037] Based on the direct acceptance matching results and residual acceptance results, the impact of acceptance is summarized to obtain the key semantic inheritance relationship set and the content affected by the change set.

[0038] Specifically, based on the quality of the inherited residual and the residual to be aggregated, the downstream content evolution edge is determined, and weak explicit inheritance and aggregation are performed along the downstream content evolution edge to obtain the residual inheritance result, including:

[0039] Based on the quality of the received residual and the residual to be aggregated, determine the final received residual corresponding to each key semantic unit of change;

[0040] Based on the semantic knowledge graph of change, downstream content evolution edges are determined, and candidate inheritance and propagation paths corresponding to each key semantic unit of change are generated.

[0041] Based on the source-to-end matching cost, weak explicit acceptance nodes in each candidate acceptance propagation path are identified, the continuous distribution of weak explicit acceptance nodes is determined, and effective acceptance propagation paths are selected.

[0042] Based on the effective transmission path, the final residual data is allocated at the path level and aggregated at the content level to obtain the residual data transfer result.

[0043] Specifically, based on the key semantic inheritance relationship set and the content set affected by the change, a change-driven content organization and distribution update is performed to obtain the updated content organization relationship set and content distribution processing results, including:

[0044] Based on the set of key semantic inheritance relationships and the set of content affected by changes, a propagation graph of the impact of content changes is constructed.

[0045] Based on the content change impact propagation graph, a restart random walk impact propagation is performed to obtain the impact propagation score corresponding to each content affected by the change.

[0046] Based on the impact propagation score, the impact level of each affected content is determined, resulting in a set of affected content levels.

[0047] Based on the affected content hierarchy set and the content evolution relationship set, the organization and distribution relationship is updated hierarchically and the distribution and disposal status is mapped to obtain the updated content organization relationship set and content distribution and disposal results.

[0048] Specifically, based on the affected content hierarchy set and content evolution relationship set, the organizational distribution relationship is updated hierarchically and the distribution and disposal status is mapped, resulting in the updated content organizational relationship set and content distribution and disposal results, including:

[0049] Based on the affected content hierarchy set, each affected content is divided into high-impact content, medium-impact content, and low-impact content;

[0050] Based on the content evolution relationship set, the content organization relationship and distribution relationship corresponding to high-impact content, medium-impact content and low-impact content are extracted to obtain the organizational relationship set to be updated and the distribution relationship set to be adjusted.

[0051] The set of organizational relationships to be updated and the set of distribution relationships to be adjusted are updated hierarchically, and the distribution handling status corresponding to each content affected by the change is determined to obtain a set of distribution handling statuses.

[0052] Based on the distribution and disposal status set, the status mapping and result aggregation are performed on the hierarchically updated set of organizational relationships to be updated and the set of distribution relationships to be adjusted to obtain the updated set of content organizational relationships and content distribution and disposal results.

[0053] On the other hand, the present invention provides an AIGC content organization and distribution system, comprising:

[0054] The data acquisition module is used to collect AIGC content evolution datasets and source content change records;

[0055] The evolutionary relationship module, based on the AIGC content evolution dataset, constructs content evolutionary relationships to obtain a content evolutionary relationship set;

[0056] The graph construction module constructs a semantic inheritance knowledge graph based on the content evolution relationship set and source content change records, and performs residual retention inheritance matching to obtain the key semantic inheritance relationship set and the content set affected by the change.

[0057] The organization and distribution module updates the content organization and distribution based on the key semantic inheritance relationship set and the content set affected by the change, and obtains the updated content organization relationship set and content distribution processing results.

[0058] Compared with existing technologies, the beneficial effects of this invention include: constructing a semantic inheritance knowledge graph based on the content evolution relationship set and source content change records; and constructing a directional inheritance metric target by combining a positive inheritance sample set, a reverse pseudo-inheritance sample set, and a non-inheritance sample set of the same topic. This enables the semantic inheritance relationship between source content, derived content, and distributed content to be identified along the content evolution direction, reducing the problem of misjudgment of content of the same topic caused by relying solely on tags, topics, or overall semantic similarity, and improving the accuracy of identifying content affected by source content changes. Through inheritance quality diversion and residual retention semi-constrained transmission matching, inheritance quality that has not formed an explicit strong match is retained as residual to be aggregated, and weak explicit inheritance aggregation is carried out along the downstream content evolution edge. This allows the semantic inheritance relationships that exist scattered after summary compression, partial rewriting, and cross-end adaptation to continue to be identified, reducing the probability of truly affected content being missed and invalid content continuing to be retrieved, reused, or distributed, and enhancing the stability of AIGC content organization and distribution. Attached Figure Description

[0059] Figure 1 A flowchart of the AIGC content organization and distribution method provided by the present invention;

[0060] Figure 2 This is a structural diagram of the AIGC content organization and distribution system provided by the present invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0062] Example 1:

[0063] See Figure 1 This embodiment provides a method for organizing and distributing AIGC content, including the following specific steps:

[0064] Step S1: Collect AIGC content evolution dataset and source content change records.

[0065] In this embodiment, the AIGC content evolution dataset includes source content data, derived content data, distributed content data, generation records, and adaptation records.

[0066] Collect source content data. Source content is the basic content that enterprises initially use to generate subsequent AIGC content. For each piece of source content, save the source content identifier, source content text, business theme, creation time and version identifier.

[0067] Collect derived content data. Derived content is content generated from source content through AIGC. For each piece of derived content, save the derived content identifier, the derived content text, the identifier of the directly generated source content, the generation task type, and the generation time.

[0068] Collect and distribute content data. The distributed content is based on derived content and is prepared to be sent to different terminals or channels. For each piece of distributed content, save the distributed content identifier, the distributed content text, the derived content identifier of the directly adapted source, the distribution channel, and the formation time.

[0069] Collect and generate records. These records are used to indicate which upstream content generated a particular piece of content. For each generation record, save the upstream content identifier, the downstream generated content identifier, the generation type, and the generation time.

[0070] Collect adaptation records. Adaptation records are used to explain which derived content a distributed content was obtained through cross-platform adaptation. For each adaptation record, save the derived content identifier, the distributed content identifier, the adaptation type, and the adaptation time.

[0071] Collect source content change records. Source content change records are records generated when source content is updated. For each source content change record, save the source content identifier that changed, the source content before the change, the source content after the change, and the change time.

[0072] The number of days for backtracking the source content creation time is determined based on the average cycle of content generation for the enterprise, and is calculated using the median of the enterprise's content generation cycle over the past three months. For example, in this application embodiment, the preferred number of backtracking days is 1 to 7 days.

[0073] The interval between the time of derivative content generation and the time of source content formation is determined based on the average time taken to generate AIGC content for the enterprise. It is calculated using the average time taken to generate AIGC content for the enterprise in the past three months. For example, in the embodiments of this application, the interval is preferably 0.5 to 2 hours.

[0074] The interval between the time of content distribution and the time of derivative content generation is determined based on the average time taken for cross-platform content adaptation of the enterprise. It is calculated using the average time taken for content adaptation over the past three months. For example, the interval in this application embodiment is preferably 0.1 to 1 hour.

[0075] The interval between the time of source content change and the time of initial formation of source content is determined based on the average cycle of enterprise business rule updates, and is calculated using the median of the enterprise's business rule update cycle over the past three months. For example, in the embodiments of this application, the interval is preferably 7 to 30 days.

[0076] For example, a source content is collected, the source content is identified as S1, the main text of the source content is the enterprise's internal general business rules text, the business subject is internal business specifications, the creation time is the number of days before the current system time, and the version is identified as V1.0.

[0077] A derived content is generated based on the source content. The derived content is identified as D1. The main text of the derived content is a concise summary of the source content. The source content is directly generated and identified as S1. The generation task type is summary generation, and the generation time is the above-mentioned interval after the source content was formed.

[0078] Based on this derived content, a distribution content is generated, with the distribution content identifier being F1. The main text of the distribution content is the platform-adapted text of the derived content, the derived content from which direct adaptation is sourced is identified as D1, the distribution channel is the enterprise's internal office platform, and the generation time is the aforementioned interval after the time the derived content was generated.

[0079] The corresponding generation record is as follows: the upstream content is identified as S1, the downstream generated content is identified as D1, the generation type is summary generation, and the generation time is the same as the derived content generation time.

[0080] The corresponding adaptation record is as follows: the derived content identifier is D1, the distribution content identifier is F1, the adaptation type is character compression, and the adaptation time is consistent with the time when the distribution content was formed.

[0081] If the source content S1 changes after the aforementioned interval following its initial formation time, a source content change record is generated. The source content that has changed is identified as S1. Before the change, the source content is the initial business rule text, and after the change, the source content is the updated business rule text. The change time is the source content update time recorded by the system.

[0082] Step S2: Based on the AIGC content evolution dataset, construct content evolution relationships to obtain a content evolution relationship set.

[0083] The specific steps of step S2 are as follows:

[0084] Step S201: Based on the generation record, establish the generation association between the source content and the derived content, as well as the regeneration association between each derived content, to obtain the generation relationship set.

[0085] In this embodiment, all generated records in the AIGC content evolution dataset are traversed, and the upstream content identifier, downstream generated content identifier, generation type, and generation time contained in each generated record are extracted.

[0086] The validity threshold for the generation record time is determined based on the longest valid period of the enterprise's content generation chain. It is calculated by taking 1.2 times the maximum time taken for all generation chains of the enterprise from source content to final derived content in the past three months. For example, in this application embodiment, the validity threshold for the generation record time is preferably 7 to 30 days.

[0087] Filter out generated records with empty upstream content identifiers; filter out generated records with empty downstream content identifiers; filter out generated records whose generation time is earlier than the current time minus the validity threshold of the generated record time; establish a directed generation association for each remaining valid generated record, with the direction of the generation association pointing from the content corresponding to the upstream content identifier to the content corresponding to the downstream content identifier.

[0088] All established directed generation associations are deduplicated. The deduplication rule is that generation associations with the same upstream content identifier, downstream generated content identifier, and generation type are considered duplicate associations, and the one with the latest generation time is retained.

[0089] All deduplicated valid directed generated associations are uniformly collected to obtain a generated relation set.

[0090] Step S202: Based on the adaptation record, establish cross-platform adaptation associations between derived content and distributed content to obtain an adaptation relationship set.

[0091] In this embodiment, all adaptation records in the AIGC content evolution dataset are traversed, and the derived content identifier, distribution content identifier, adaptation type, and adaptation time contained in each adaptation record are extracted.

[0092] The validity threshold for adaptation record time is determined based on the longest valid period of the enterprise's cross-platform content adaptation links. It is calculated by taking 1.2 times the maximum time taken by all adaptation links of the enterprise in the past three months from derived content to final distributed content. For example, the validity threshold for adaptation record time in this application embodiment is preferably 3 to 15 days.

[0093] Filter out adaptation records with empty derived content identifiers; filter out adaptation records with empty distribution content identifiers; filter out adaptation records whose adaptation time is earlier than the current time minus the adaptation record time validity threshold; establish a directed cross-platform adaptation association for each remaining valid adaptation record, with the direction of the adaptation association pointing from the content corresponding to the derived content identifier to the content corresponding to the distribution content identifier.

[0094] All established directed cross-platform adaptation associations are deduplicated. The deduplication rule is that adaptation associations with the same derived content identifier, distribution content identifier, and adaptation type are considered duplicate associations, and the one with the latest adaptation time is retained.

[0095] All deduplicated valid directed cross-platform adaptation relationships are uniformly collected to obtain an adaptation relationship set.

[0096] Step S203: Based on the generation relation set and the adaptation relation set, organize the evolution chain of source content, derived content and distributed content to obtain the content evolution relation set.

[0097] In this embodiment, the starting and ending nodes of all directed generation associations in the generation relation set are extracted, and the starting and ending nodes of all directed cross-end adaptation associations in the adaptation relation set are extracted. All the above nodes are merged and deduplicated to obtain the set of all content nodes participating in content evolution.

[0098] The maximum length threshold of the evolution chain is determined based on the average level of the enterprise's AIGC content evolution. It is calculated by adding 1 to the maximum number of levels of all complete content evolution chains in the past three months. For example, in this application embodiment, the maximum length threshold of the evolution chain is preferably 3 to 6 levels.

[0099] Each source content node in the content node set is used as the starting node of the evolution chain. A depth-first traversal is performed along the direction of the directed association. During the traversal, the cumulative length of the current evolution chain is recorded in real time. When the cumulative length of the evolution chain of a certain branch reaches the maximum length threshold of the evolution chain, the traversal of that branch is stopped immediately.

[0100] When traversing to any derived content node, in addition to continuing to traverse downwards along the directed generation association corresponding to that node, it also traverses downwards along the directed cross-end adaptation association corresponding to that node to the corresponding distribution content node.

[0101] All complete evolutionary chains obtained through traversal are deduplicated. The deduplication rule is that a chain containing a completely identical sequence of nodes is considered a duplicate chain, and the chain whose last node was generated the latest is retained.

[0102] All deduplicated complete evolutionary chains are uniformly collected to obtain a content evolutionary relationship set.

[0103] Step S3: Based on the content evolution relationship set and source content change records, construct a change semantic inheritance knowledge graph, and perform residual retention inheritance matching to obtain the key semantic inheritance relationship set and the content set affected by the change.

[0104] The specific steps of step S3 are as follows:

[0105] Step S301: Based on the source content change record, extract the content fragments before and after the change corresponding to the source content, and perform semantic difference recognition to obtain the set of key semantic units of change and the corresponding set of semantic change degree.

[0106] In this embodiment, the source content before and after the change is extracted from the source content change record. The source content before and after the change is then segmented into semantic units. The segmentation rule is to first segment the source content into sentences according to periods, semicolons, question marks, and exclamation marks. Long sentences containing parallel components or multiple action relationships are then split into clauses. Each short text that can express a relatively complete business meaning is taken as a semantic unit, resulting in the set of semantic units before and after the change.

[0107] A threshold calibration sample set was prepared based on historical data from the AIGC content evolution dataset and source content change records. When preparing the threshold calibration sample set, historical content was extracted hierarchically according to business theme, content type, generation type, adaptation type, and content evolution level to avoid the calibration sample set containing only a single business scenario or a single generation method.

[0108] Each extracted semantic unit before change, semantic unit after change, and downstream inheriting semantic unit is manually labeled. During labeling, each semantic unit is broken down into business subject, business action, constraints, scope of application, and exceptions. If the business subject or business action changes, or if changes in constraints, scope of application, or exceptions lead to changes in business meaning or execution result, the corresponding semantic unit is labeled as a key change semantic unit. If the business subject and business action remain consistent, and at least two of the constraints, scope of application, and exceptions remain consistent or form equivalent expressions, and there are no semantic conflicts, the corresponding downstream semantic unit is labeled as a direct inheriting semantic unit. If the business subject and business action remain consistent, and at least one of the remaining semantic elements remains consistent or forms a general expression, and the semantic unit is located on the corresponding content evolution path, the corresponding downstream semantic unit is labeled as a weakly explicit inheriting semantic unit. If the business subject or business action changes, or there are conflicting constraints, the corresponding semantic unit is labeled as a non-inheriting semantic unit.

[0109] Each calibration sample is independently labeled by two labelers. When the labeling results from the two labelers are consistent, the sample is considered a valid calibration sample. If the labeling results are inconsistent, a third labeler reviews the sample based on the content before and after the change, the generated records, and the adaptation records, and the review result is used as the final labeling result. The reviewed samples are then divided into training samples, validation samples, and test samples in a 7:2:1 ratio.

[0110] When dealing with subsequent threshold determinations, a unified calibration is performed based on the threshold calibration sample set. Specifically, during calibration, the corresponding indicator values ​​are arranged from smallest to largest, and the median values ​​of adjacent indicator values ​​are used as candidate thresholds. The number of misjudged samples and the number of missed samples are counted for each candidate threshold. Based on the actual handling costs caused by misjudgments and missed judgments in the corresponding business scenarios, misjudgment cost weights and missed judgment cost weights are set. The number of misjudged samples is multiplied by the misjudgment cost weight, and the number of missed samples is multiplied by the missed judgment cost weight. The two products are then added together to obtain the calibration cost corresponding to the candidate threshold. The candidate threshold with the lowest calibration cost is selected as the target threshold. When multiple candidate thresholds have the same calibration cost, for scenarios involving business rules, applicable scope, and compliant content, the candidate threshold with fewer missed samples is selected; for scenarios involving promotional content and general display content, the candidate threshold with fewer misjudged samples is selected.

[0111] A pre-trained bidirectional semantic coding model is used to vectorize and encode each semantic unit. The semantic coding model includes a 12-layer encoder, a hidden layer dimension of 768, and 12 attention heads.

[0112] When fine-tuning the semantic coding model, semantic unit pairs are divided into semantic equivalence classes, local variation classes, and key variation classes, and a semantic relation classification layer is set at the model output. During training, on the one hand, the classification loss is calculated based on the difference between the class probabilities output by the semantic relation classification layer and the manually labeled classes; on the other hand, semantically equivalent samples are selected as positive samples, and samples with the same topic but labeled as local or key variations are selected as hard negative samples. The contrastive loss is calculated based on the requirement that the distance between positive samples should be less than the distance between hard negative samples. The classification loss and the contrastive loss are added together to obtain the training loss of the semantic coding model.

[0113] When initializing word vectors for enterprise-specific terms, a basic vocabulary is first used to segment the terminology. If an enterprise-specific term can be segmented into two or more basic words, the word vectors corresponding to each basic word are extracted, and the average value of the corresponding components of each word vector is calculated. This average vector is then used as the initial word vector for the enterprise-specific term. If an enterprise-specific term cannot be segmented from the basic vocabulary, the 20 most frequently occurring context words that co-occur with the terminology are extracted from historical business texts. The average value of the word vectors corresponding to these 20 context words is calculated, and this average vector is then used as the initial word vector for the enterprise-specific term. During model training, the word vectors of the enterprise-specific terminology are updated synchronously with the parameters of the last four encoder layers of the model. The learning rate for the enterprise-specific terminology word vectors is set to 0.5 times the learning rate of the main model.

[0114] During model training, the first 8 encoder layers are frozen, while the last 4 encoder layers, the semantic relation classification layer, and the enterprise-specific vocabulary word vectors are updated. Each training batch includes 16 pairs of semantic units, where the difference in the number of samples for semantic equivalence classes, local variation classes, and key variation classes does not exceed 1. The initial learning rate is set to 0.00002. After each training epoch, the validation loss and semantic relation classification accuracy are calculated based on the validation samples. When the decrease in validation loss is less than 0.001 for 5 consecutive training epochs, and the increase in semantic relation classification accuracy is less than 0.001, model training is stopped, resulting in the fine-tuned semantic encoding model.

[0115] The modified semantic coding model is used to encode the set of semantic units before and after the change, respectively, to obtain the set of semantic vectors before and after the change.

[0116] For each semantic vector before the change, cosine similarity is calculated between it and each semantic vector after the change. When calculating cosine similarity, the sum of the products of the corresponding components in the two semantic vectors is calculated first, and then the square root of the sum of the squares of the components in the two semantic vectors is calculated. The sum of the component products is divided by the product of the two square roots to obtain the corresponding semantic similarity.

[0117] Select the maximum value from the semantic similarity between each semantic unit before the change and all semantic units after the change, and subtract the maximum value from 1 to obtain the degree of semantic change of the corresponding semantic unit before the change.

[0118] Based on the semantic change degree corresponding to key and non-key semantic change units in the threshold calibration sample set, the semantic change threshold is determined according to the unified threshold calibration method described above. Different business scenarios are calibrated using verification samples corresponding to those scenarios. When the number of valid verification samples for a certain business scenario is less than 300 sets, that business scenario is merged with the business scenario that is closest in terms of business subject, content type, and change type for threshold calibration. For example, in this embodiment, the semantic change threshold is preferably between 0.25 and 0.40.

[0119] The semantic units before the change whose semantic change degree is greater than or equal to the semantic change threshold are identified as key semantic units for change. At the same time, the semantic units after the change with which they have the greatest semantic similarity are recorded to form source semantic change pairs.

[0120] All source semantic change pairs are uniformly collected to obtain a set of key semantic units of change. The degree of semantic change corresponding to each key semantic unit of change is uniformly collected to obtain a set of semantic change degree.

[0121] Step S302: Based on the content evolution relationship set, identify the content objects that have downstream evolution associations with each change key semantic unit, and extract the inheriting semantics to obtain the downstream associated content object set and the candidate inheriting semantic unit set.

[0122] In this embodiment, the source content node to which each key semantic unit of change belongs is located from the content evolution relationship set. The corresponding source content node is used as the starting node, and a breadth-first traversal is performed layer by layer along the content evolution relationship downstream.

[0123] During traversal, the content identifier, content type, access level, and generation time of each access node are recorded. Traversal stops when a content node is deleted, expired, or archived. Content nodes are excluded if their generation time is earlier than the current version's creation time of the corresponding source content. Content nodes are also excluded if their business theme differs from that of the corresponding source content.

[0124] The upper limit of traversal depth is determined based on the hierarchical distribution of each complete content evolution chain in the content evolution relationship set. First, the number of content nodes contained in each complete content evolution chain is counted. Then, the maximum number of content nodes is selected, and this maximum number of content nodes is incremented by 1 to obtain the upper limit of traversal depth. For example, in the embodiments of this application, the upper limit of traversal depth is preferably 3 to 6 layers.

[0125] Collect the derived content nodes and distributed content nodes that were not excluded during the traversal, and deduplicate them according to their generation time and content identifier to obtain a set of downstream related content objects.

[0126] Perform text preprocessing on each content object in the downstream related content object set, removing whitespace characters and format marks, splitting table content by cell, splitting list content by list item, and retaining complete code blocks for code snippets.

[0127] The same semantic unit segmentation rules as in step S301 are used to segment the preprocessed content objects into semantic units. Semantic units shorter than 5 characters that cannot express a complete business meaning are merged with the preceding or following semantic unit. Semantic units longer than 200 characters that contain more than two business actions or constraints are further segmented according to the business actions or constraints.

[0128] Remove semantic units that contain only stop words, numbers, or punctuation marks, and remove semantic units that do not share a common business entity, business action, or constraint with the corresponding key change semantic units.

[0129] When calibrating the frequency threshold of general semantic units, the proportion of each semantic unit in the downstream associated content objects is used as an indicator value. Semantic units manually labeled as general expressions and non-general expressions are used as calibration samples. The frequency threshold of general semantic units is determined according to the unified threshold calibration method in step S301. Semantic units with an occurrence proportion greater than or equal to the frequency threshold of general semantic units are identified as general semantic units and removed. For example, in the embodiments of this application, the frequency threshold of general semantic units is preferably 0.85 to 0.95.

[0130] When performing semantic deduplication on candidate semantic units, the semantic similarity between any two candidate semantic units is calculated. Based on the manually labeled pairs of semantic units as repetitive and non-repetitive expressions, a semantic deduplication threshold is determined according to the unified threshold calibration method in step S301. Semantic units with a semantic similarity greater than or equal to the semantic deduplication threshold are identified as repetitive semantic units, and semantic units with earlier generation times and more complete business elements are retained. For example, in this embodiment, the semantic deduplication threshold is preferably between 0.90 and 0.97.

[0131] The semantic units that have been filtered and deduplicated are uniformly collected to obtain a candidate set of semantic units.

[0132] Step S303: Based on the set of key semantic units for change, the set of downstream related content objects, and the set of candidate semantic units for inheritance, construct a knowledge graph of semantic inheritance for change.

[0133] The specific steps of step S303 are as follows:

[0134] Step S3031: Based on the set of key semantic units for change, the set of candidate inheriting semantic units and the set of downstream related content objects, perform node mapping and semantic content attribution labeling to obtain the set of semantic nodes, the set of content nodes and the set of semantic content attribution relationships.

[0135] In this embodiment, each pre-change semantic unit in the set of key semantic units is mapped to a pre-change semantic node, and each post-change semantic unit is mapped to a post-change semantic node. A unique semantic node identifier is assigned to each changed semantic node, and its corresponding text content, semantic vector, source content identifier, start and end positions in the source content, and semantic importance score are recorded. The semantic importance score is calculated using the TF-IDF algorithm.

[0136] Each candidate semantic unit in the candidate semantic unit set is mapped to a candidate semantic node. A unique semantic node identifier is assigned to each candidate semantic node, and its corresponding text content, semantic vector, the identifier of the content object to which it belongs, the start and end positions in the content object, and semantic importance score are recorded.

[0137] Each content object in the downstream associated content object set is mapped to a content node. A unique content node identifier is assigned to each content node, and its content type, main text content, business theme, generation time, version number, number of visits, usage frequency, and status information are recorded.

[0138] Map each content object to a business topic as a business topic node, assign a unique topic node identifier to each business topic node, and record its topic name, topic description, and number of associated content.

[0139] Establish the attribution relationship between semantic nodes and their respective content nodes, and record the semantic node identifier, content node identifier, and the relative position of the semantic unit in the content.

[0140] Establish the attribution relationship between content nodes and their respective business theme nodes, and record the content node identifier and theme node identifier.

[0141] All changed semantic nodes and candidate successor semantic nodes are uniformly collected to obtain a semantic node set.

[0142] All content nodes are uniformly collected to obtain a content node set.

[0143] The attribution relationships between all semantic nodes and content nodes, and between content nodes and business theme nodes, are uniformly collected to obtain a semantic content attribution relationship set.

[0144] Step S3032: Based on the content evolution relationship set, establish content evolution edges between each content node to obtain the content evolution edge set.

[0145] In existing technologies, content evolution edges typically employ unweighted or fixed-weight methods, with weight values ​​uniformly set to 1 or 0.5. In AIGC content scenarios, such as enterprise-level AIGC content, content is constantly updated and iterated as business evolves. Content generated more recently is usually based on the latest version of the source content, exhibiting a higher semantic continuity with the source content; conversely, content generated earlier is more likely to be based on an older version of the source content, resulting in a lower semantic continuity with the current source content. Fixed weights cannot reflect this temporal difference, leading to biases in identifying the continuity relationships of historical content and a high misjudgment rate.

[0146] To address the aforementioned issues, this solution introduces an exponential decay function to calculate the time decay weight of content evolution edges. The weight is calculated based on the difference between the generation time of the content evolution edge and the current time; the smaller the difference, the greater the weight. The decay coefficient is determined based on the average effective period of the enterprise's content; the faster the content is updated, the greater the decay coefficient. This method ensures that content evolution edges with more recent generation times have higher weights, more accurately reflecting the true strength of semantic connection. After the improvement, the correlation between the weight of content evolution edges and the actual semantic connection strength is significantly enhanced, and the accuracy of identifying historical content connection relationships is significantly improved.

[0147] In this embodiment, all content evolution edges in the content evolution relationship set are traversed, and the upstream node identifier, downstream node identifier, evolution type, and evolution time of each content evolution edge are extracted.

[0148] Content evolution edges whose upstream and downstream node identifiers both exist in the content node set are selected. The selected content evolution edges are deduplicated. The deduplication rule is that evolution edges whose upstream node identifier, downstream node identifier, and evolution type are all the same are considered duplicate edges, and the one with the latest evolution time is retained.

[0149] For each retained content evolution edge, a time decay weight is calculated. The time decay weight is calculated using an exponential decay function. The decay coefficient is determined based on the average effective period of the enterprise content. For example, in the embodiments of this application, the decay coefficient is preferably 0.01 to 0.05.

[0150] Assign a unique edge identifier to each content evolution edge and record its upstream node identifier, downstream node identifier, evolution type, evolution time, and time decay weight.

[0151] All processed content evolution edges are uniformly collected to obtain the content evolution edge set.

[0152] For example, if the current time is June 1, 2025, then after filtering, 3 valid content evolution edges are obtained.

[0153] The first evolutionary edge has an upstream node of S1 and a downstream node of D1. The evolution type is digest generation, and the evolution time is May 2, 2025, which is 30 days later than the current time.

[0154] The second evolutionary edge has D1 as the upstream node and D3 as the downstream node. The evolution type is question-and-answer generation, and the evolution time is May 5, 2025, which is 27 days later than the current time.

[0155] The third evolutionary edge has D3 as its upstream node and F1 as its downstream node. Its evolution type is character compression, and its evolution time is May 10, 2025, which is 22 days later than the current time.

[0156] The attenuation coefficient is set to 0.03.

[0157] To calculate the time decay weight of the first evolution edge, first calculate the product of 0.03 and 30, then calculate the negative power of the natural exponent. The result is approximately 0.407.

[0158] To calculate the time decay weight of the second evolution edge, first calculate the product of 0.03 and 27, then calculate the negative power of the natural exponent. The result is approximately 0.445.

[0159] To calculate the time decay weight of the third evolution edge, first calculate the product of 0.03 and 22, then calculate the negative power of the natural exponent. The result is approximately 0.517.

[0160] These three weighted evolution edges are grouped together to obtain the content evolution edge set.

[0161] Step S3033: Based on the semantic content attribution relation set, establish content inclusion relation edges to obtain the content inclusion relation edge set.

[0162] In this embodiment, the attribution relationships between all semantic nodes and content nodes in the semantic content attribution relationship set are traversed, and a directed content inclusion relationship edge is established for each attribution relationship, with the direction of the edge pointing from the content node to the semantic node it contains.

[0163] For each content containing relation edges, calculate the semantic importance weight. The semantic importance weight is equal to the semantic importance score of the semantic node divided by the sum of the semantic importance scores of all semantic nodes in its content node.

[0164] Assign a unique edge identifier to each content-containing relation edge, and record its content node identifier, semantic node identifier, the start position and end position of the semantic unit in the content, and the semantic importance weight.

[0165] All content containing relationship edges are uniformly collected to obtain the content containing relationship edge set.

[0166] Step S3034: Based on the source content change record, establish source-side semantic change relationship edges to obtain the semantic change relationship edge set.

[0167] In this embodiment, all source semantic change pairs in the set of key semantic units are traversed, and the semantic node identifier before the change, the semantic node identifier after the change, and the corresponding degree of semantic change for each source semantic change pair are extracted.

[0168] For each source semantic change pair, establish a directed source-side semantic change relationship edge, with the edge pointing from the semantic node before the change to the semantic node after the change.

[0169] Calculate the change impact weight for each source-side semantic change relation edge. The change impact weight is equal to the corresponding semantic change degree.

[0170] Assign a unique edge identifier to each source-side semantic change relationship edge, and record its semantic node identifier before the change, semantic node identifier after the change, degree of semantic change, and weight of the change impact.

[0171] All source-side semantic change relation edges are uniformly collected to obtain the semantic change relation edge set.

[0172] Step S3035: Construct a semantic change inheritance knowledge graph based on the semantic node set, content node set, content evolution edge set, content inclusion relationship edge set, and semantic change relationship edge set.

[0173] In this embodiment, the semantic node set, content node set, and business topic node set are merged to obtain the node set of the change semantic inheritance knowledge graph. Each node contains its type identifier and all attribute information.

[0174] By merging the content evolution edge set, the content inclusion relation edge set, and the semantic change relation edge set, we obtain the edge set of the semantic change inheritance knowledge graph. Each edge contains its type identifier and all attribute information.

[0175] A directed weighted change semantic inheritance knowledge graph is constructed using an attribute graph model. A primary key index based on node identifier is established for the node set, an approximate nearest neighbor index based on semantic vector is established for semantic nodes, and a composite index based on business theme and generation time is established for content nodes.

[0176] Perform consistency checks on the knowledge graph to ensure that the starting and ending points of all edges exist in the node set, that the attribute information of all nodes is complete, and that the weight values ​​of all edges are within the range of 0 to 1.

[0177] The validated knowledge graph is persistently stored to obtain the modified semantic inheritance knowledge graph.

[0178] Step S304: Based on the change semantic connection knowledge graph, perform directional semantic connection metric learning to obtain the source-to-source connection matching cost.

[0179] The specific steps of step S304 are as follows:

[0180] Step S3041: Based on the change semantic inheritance knowledge graph and content evolution relationship set, determine the content evolution direction, and identify the semantic inheritance relationship between each change key semantic unit and each candidate inheritance semantic unit along the content evolution direction to obtain the positive inheritance sample set.

[0181] In this embodiment, the content evolution direction is determined based on the content evolution relationship set as starting from the source content and pointing to the derived content, and then from the derived content to subsequent derived content or distributed content.

[0182] Traverse the semantic knowledge graph of change along the direction of content evolution. For each key semantic unit of change, locate its source content node and extract candidate semantic units from each downstream content node of that source content node.

[0183] Calculate the basic semantic similarity between the key semantic unit and each candidate inheriting semantic unit. The calculation method for the basic semantic similarity is the same as the semantic similarity calculation method in step S301.

[0184] Extract business entities, business actions, constraints, scope of application, and exceptions from the key semantic units and candidate successor semantic units respectively, and unify the business elements with different names but marked as synonyms through the enterprise glossary in the two semantic units.

[0185] The total number of business elements contained in the key semantic unit of change is counted, and the number of business elements that are the same or equivalent to those in the candidate accepting semantic unit and the key semantic unit of change is counted. The number of the same or equivalent business elements is divided by the total number of business elements contained in the key semantic unit of change to obtain the business entity overlap.

[0186] Based on the basic semantic similarity of directly receiving semantic unit pairs and non-receiving semantic unit pairs in the threshold calibration sample set, the basic semantic similarity threshold is determined according to the unified threshold calibration method in step S301. Based on the overlap of business entities corresponding to the same calibration samples, the business entity overlap threshold is determined according to the same method. For business rules and applicable scope scenarios, the missed judgment cost weight is determined based on the average processing cost caused by the continued distribution of invalid content. For general promotional content scenarios, the misjudgment cost weight is determined based on the average manual review cost caused by the erroneous interception of normal content. For example, in the embodiments of this application, the basic semantic similarity threshold is preferably 0.50 to 0.70, and the business entity overlap threshold is preferably 0.30 to 0.50.

[0187] Semantic unit pairs with a basic semantic similarity greater than or equal to the basic semantic similarity threshold and a business entity overlap greater than or equal to the business entity overlap threshold are identified as positive acceptance samples.

[0188] The positive acceptance samples are deduplicated. Samples with the same change key semantic unit identifier and candidate acceptance semantic unit identifier are identified as duplicate samples, and only one is retained to obtain the positive acceptance sample set.

[0189] Step S3042: Based on the forward acceptance sample set, the semantic correspondence between the candidate acceptance semantic units and the corresponding change key semantic units is determined as the reverse pseudo acceptance sample, thus obtaining the reverse pseudo acceptance sample set.

[0190] In this embodiment, the positive acceptance sample set is traversed, and the change key semantic unit and candidate acceptance semantic unit corresponding to each positive acceptance sample are extracted.

[0191] The semantic order in the forward receiving sample is reversed to form a reverse semantic pair from the candidate receiving semantic unit to the changed key semantic unit.

[0192] Verify the content evolution direction between the content node to which the candidate inheriting semantic unit belongs and the source content node to which the change key semantic unit belongs. When the content node to which the candidate inheriting semantic unit belongs is downstream of the source content node, and the direction of the reverse semantic pair is opposite to the content evolution direction, retain the reverse semantic pair.

[0193] Exclude reverse semantic pairs where the candidate successor semantic unit was generated earlier than the corresponding source content, and exclude candidate successor semantic pairs where the candidate itself is a reverse semantic pair of the source content.

[0194] The remaining reverse semantic pairs are identified as reverse pseudo-acceptance samples, and the corresponding forward acceptance sample identifier is recorded for each reverse pseudo-acceptance sample, thus obtaining the reverse pseudo-acceptance sample set.

[0195] Step S3043: Based on the positive inheritance sample set, select candidate inheritance semantic units that are on the same topic as each change key semantic unit but have not formed a semantic inheritance relationship, and obtain the non-inheritance sample set on the same topic.

[0196] In this embodiment, the business theme to which each change key semantic unit belongs is extracted, and candidate undertaking semantic units belonging to the same business theme are selected from the candidate undertaking semantic unit set.

[0197] Candidate acceptance semantic units that are already included in the positive acceptance sample set are excluded from the screening results. Candidate acceptance semantic units that have the same business subject and the same business action as the corresponding change key semantic unit are also excluded.

[0198] When determining the proportion of core business elements, the importance of each business element is calculated based on its inverse document frequency in historical source content and the results of manual importance labeling. Business elements are then ranked from highest to lowest importance, and selections are made sequentially starting from the first element in the ranking until the sum of the importance of the selected business elements reaches 0.70 of the sum of the importance of all business elements. The selected business elements are then identified as core business elements. Candidate semantic units containing the same core business elements are excluded or modified.

[0199] When calibrating the upper limit of similarity for non-adhering samples on the same topic, the basic semantic similarity of directly adhering semantic unit pairs and non-adhering semantic unit pairs on the same topic in the sample set is calibrated based on a threshold, and the upper limit of non-adhering similarity on the same topic is determined according to the unified threshold calibration method in step S301. Candidate adhering semantic units with a basic semantic similarity greater than or equal to the upper limit of non-adhering similarity on the same topic are excluded to avoid the same semantic unit being simultaneously labeled as a positive adhering sample and a non-adhering sample on the same topic. For example, in the embodiments of this application, the upper limit of non-adhering similarity on the same topic is preferably 0.75 to 0.85.

[0200] The remaining candidate accepting semantic units are combined with the corresponding change key semantic units to form a non-accepting sample of the same theme, and the business theme and content evolution path corresponding to the two semantic units are recorded to obtain a set of non-accepting samples of the same theme.

[0201] Step S3044: Based on the positive acceptance sample set, the reverse pseudo acceptance sample set, and the non-acceptance sample set of the same topic, determine the ranking of sample acceptance costs and construct a directional acceptance metric target.

[0202] In this embodiment, the acceptance cost corresponding to the positive acceptance sample is set to be lower than the acceptance cost corresponding to the non-acceptance sample of the same topic, and the acceptance cost corresponding to the non-acceptance sample of the same topic is set to be lower than the acceptance cost corresponding to the reverse pseudo-acceptance sample, thus obtaining the sample acceptance cost ranking.

[0203] When constructing the training sample set, a key semantic unit of change is used as the source-side anchor point. One candidate semantic unit corresponding to this source-side anchor point is selected from the positive receiving sample set as a positive sample. The candidate semantic unit with the highest basic semantic similarity to this source-side anchor point is selected from the non-receiving sample set of the same topic as a difficult-to-bear sample of the same topic. Finally, one semantic unit formed by the reverse of the corresponding positive sample is selected from the reverse pseudo-receiving sample set as a reverse difficult-to-bear sample. If multiple positive samples correspond to the same source-side anchor point, they are arranged in order of increasing content evolution level to form the training sample set. If a source-side anchor point lacks a difficult-to-bear sample of the same topic, it is supplemented by selecting the non-receiving semantic unit with the highest basic semantic similarity from the same business topic and content type.

[0204] For each training sample group, calculate the positive acceptance cost between the source anchor and the positive sample, the non-acceptance cost between the source anchor and the same-topic hard-to-bear sample, and the reverse pseudo-acceptance cost between the source anchor and the reverse hard-to-bear sample. First, calculate the sum of the positive acceptance cost and the first margin coefficient, then subtract the same-topic non-acceptance cost. If the result is greater than 0, the result is determined as the first loss term; if the result is less than or equal to 0, the first loss term is determined as 0. Next, calculate the sum of the positive acceptance cost and the second margin coefficient, and subtract the reverse pseudo-acceptance cost. If the result is greater than 0, the result is determined as the second loss term; if the result is less than or equal to 0, the second loss term is determined as 0. Add the first loss term and the second loss term to obtain the directional acceptance loss for that training sample group.

[0205] The first and second interval coefficients are jointly calibrated based on validation samples. The candidate values ​​for the first interval coefficient are set to 0.05 to 0.40, increasing in increments of 0.05. The candidate values ​​for the second interval coefficient are set to 0.10 to 0.60, increasing in increments of 0.05, ensuring that the second interval coefficient is always greater than the first interval coefficient. Model training is performed on each set of candidate interval coefficients. The accuracy rates for forward acceptance recognition, reverse pseudo-acceptance recognition, and non-acceptance recognition of the same topic in the validation samples are statistically analyzed. The candidate interval coefficient corresponding to the highest average of these three recognition accuracies is determined as the target interval coefficient. When the average values ​​of multiple sets of candidate interval coefficients are the same, the set with the smaller validation loss is selected. For example, in this embodiment, the first interval coefficient is preferably 0.15 to 0.25, and the second interval coefficient is preferably 0.30 to 0.40.

[0206] The directional acceptance loss of each training sample group is averaged to obtain the directional acceptance metric target.

[0207] Step S3045: Based on the directional acceptance metric objective, perform metric learning on the change key semantic unit set and the candidate acceptance semantic unit set to obtain the source-direction acceptance matching cost.

[0208] In this embodiment, a source projection matrix and a receiving projection matrix are set. The source projection matrix is ​​used to project the semantic vector of the key semantic unit of change, and the receiving projection matrix is ​​used to project the semantic vector of the candidate receiving semantic unit.

[0209] The number of rows and columns of both the source projection matrix and the receiving projection matrix are set to 768. During initialization, the diagonal elements of the two projection matrices are set to 1, and the remaining elements are set to random numbers greater than or equal to -0.01 and less than or equal to 0.01. This preserves the original semantic vector structure during the initial stage of model training and allows the source-side semantics and receiving-side semantics to gradually form different projection directions.

[0210] Each training batch selects 32 different source-side anchors from the positive acceptance sample set, and configures one positive sample, one topic-related hard-to-reject sample, and one reverse hard-to-reject sample for each source-side anchor. When the number of samples of a certain type is insufficient, samples of that type are selected cyclically from high to low based on basic semantic similarity, but the same semantic unit appears as a hard-to-reject sample no more than twice in the same training batch.

[0211] The semantic vectors of the source-side anchor points are projected using the source-direction projection matrix, and the semantic vectors of the forward samples, the same-topic hard negative samples, and the reverse hard negative samples are projected using the successor-direction projection matrix.

[0212] Normalize each semantic vector after projection. During normalization, first calculate the square of each component in the semantic vector, then sum the squared values ​​and calculate the square root of the sum. Divide each component in the semantic vector by this square root to obtain the normalized semantic vector. When the square root is less than 0.000001, set the square root to 0.000001 to avoid division by zero.

[0213] The source-to-receiver similarity is obtained by summing the products of the source-side anchor point normalized vector and the corresponding components in each receiver-side normalized vector. The source-to-receiver similarity is then subtracted from 1 to obtain the source-to-receiver matching cost for the corresponding semantic unit pair.

[0214] Based on the directional acceptance metric target determined in step S3044, calculate the directional acceptance loss corresponding to the current training batch.

[0215] Based on the directional acceptance loss, the sum of the squared values ​​of each element in the source projection matrix and the acceptance projection matrix is ​​calculated separately. This sum is multiplied by the regularization coefficient and added to the directional acceptance loss to obtain the total loss of the current training batch. The candidate values ​​of the regularization coefficient are set to 0.0001, 0.0002, 0.0005, and 0.001, and validation training is performed for each. The candidate value corresponding to the lowest validation loss is selected as the target regularization coefficient.

[0216] The backpropagation method is used to determine the direction and magnitude of change of the total loss relative to each element in the source and receiving projection matrices. Each magnitude is multiplied by the current learning rate, and then the product is subtracted from the current value of the corresponding matrix element to complete one projection matrix update. After the update, the sum of the absolute values ​​of the gradients of the two projection matrices is calculated. When the sum of the absolute values ​​of the gradients is greater than 5, all gradients are proportionally reduced according to the ratio of 5 to the sum of the absolute values ​​of the gradients.

[0217] The initial learning rate is set to 0.001. Every 10 rounds of training, the current learning rate is multiplied by 0.5 to obtain the learning rate for the next stage.

[0218] After each training round, the total validation loss, forward acceptance recognition accuracy, non-acceptance recognition accuracy for the same topic, and reverse pseudo-acceptance recognition accuracy are calculated based on the validation samples. Training stops when the adjacent changes in the total validation loss are all less than 0.001 over 10 consecutive training rounds, and the adjacent improvements in the average recognition accuracy of the three types of samples are all less than 0.001. When the number of training rounds reaches 300, training stops regardless of whether the above conditions are met, and the source projection matrix and acceptance projection matrix corresponding to the round with the lowest total validation loss are selected.

[0219] Based on the source projection matrix and the receiving projection matrix after training, projection, normalization and matching cost calculation are performed on each change key semantic unit and each candidate receiving semantic unit. The obtained matching costs are arranged in a matrix according to the identifiers of the change key semantic units and the candidate receiving semantic units to obtain the source receiving matching cost.

[0220] Step S305: Based on the source-to-end matching cost and semantic change degree set, perform end-to-end quality diversion and residual retention semi-constrained transmission matching, and perform weak explicit end-to-end aggregation to obtain the key semantic inheritance relationship set and the content set affected by changes.

[0221] The specific steps of step S305 are as follows:

[0222] Step S3051: Based on the source-to-source matching cost and the set of semantic change degrees, determine the direct acceptance quality and acceptance residual quality corresponding to each key semantic unit of change.

[0223] In this embodiment, the semantic change degree of each semantic change degree in the semantic change degree set is normalized.

[0224] During normalization, the maximum and minimum semantic change levels are first determined from the set of semantic change levels. For each key semantic unit of change, the difference between its semantic change level and the minimum semantic change level is calculated first, and then the difference between the maximum and minimum semantic change levels is calculated. Dividing the former difference by the latter difference yields the initial semantic quality of the corresponding key semantic unit of change. When the maximum and minimum semantic change levels are equal, the initial semantic quality of each key semantic unit of change is set to 1.

[0225] For each change key semantic unit, the minimum value is selected from all source-to-acceptance matching costs corresponding to it to obtain the minimum acceptance cost.

[0226] The average value of all source-direction matching costs is calculated, and the square of the difference between each source-direction matching cost and the average value is calculated. The sum of these squares is then divided by the number of source-direction matching costs, and the square root of the result is calculated to obtain the standard deviation of the source-direction matching costs. This standard deviation is compared to 0.10, and the larger value is taken. This larger value is then compared to 0.50, and the smaller value is taken to obtain the temperature coefficient.

[0227] When calculating the direct acceptance confidence level, first divide the minimum acceptance cost by the temperature coefficient to obtain the cost scaling value. Then calculate the natural index value corresponding to the cost scaling value. Divide 1 by the sum of 1 and the natural index value to obtain the direct acceptance confidence level. The smaller the minimum acceptance cost, the higher the direct acceptance confidence level; the larger the minimum acceptance cost, the lower the direct acceptance confidence level.

[0228] Multiplying the initial semantic quality by the direct acceptance confidence yields the direct acceptance quality of the corresponding changed key semantic unit.

[0229] Subtracting the direct acceptance quality from the initial semantic quality yields the acceptance residual quality of the corresponding changed key semantic unit.

[0230] The direct acceptance quality and acceptance residual quality corresponding to each key semantic unit of change are uniformly collected.

[0231] Step S3052: Based on the candidate acceptance semantic unit set, direct acceptance quality and source-to-southbound acceptance matching cost, construct an acceptance receiving object set containing residual acceptance objects, and perform semi-constrained optimal transmission matching to obtain the direct acceptance matching result and the residual to be aggregated.

[0232] In this embodiment, each candidate semantic unit in the candidate accepting semantic unit set is determined as the actual accepting and receiving object, and one residual accepting object is set for each change key semantic unit.

[0233] The actual receiving objects and the residual receiving objects are merged to obtain the receiving object set.

[0234] A corresponding source-to-receive matching cost is applied to the actual receiving object. A residual receiving cost is set for the residual receiving object, and the residual receiving cost is calibrated based on the verification samples. Specifically, all actual receiving matching costs are arranged from smallest to largest, and the median value of adjacent costs is used as a candidate residual receiving cost. Semi-constrained transmission matching is performed on each, and the number of directly received samples that are mistakenly transmitted to the residual receiving object and the number of non-received samples that are mistakenly transmitted to the actual receiving object are counted. The candidate value with the smallest calibration cost is selected as the residual receiving cost according to the unified threshold calibration method in step S301. For example, in this embodiment, the residual receiving cost is preferably 0.60 to 0.80.

[0235] An extended acceptance cost matrix is ​​constructed based on the source-to-receiver matching cost corresponding to the actual receiving object and the residual receiving cost corresponding to the residual receiving object.

[0236] A receiving capacity is set for each actual receiving object. The receiving capacity is determined based on the semantic importance of the candidate receiving semantic unit and the valid status of its content object. First, the semantic importance of the candidate receiving semantic unit is divided by the sum of the semantic importance of all candidate receiving semantic units to obtain the capacity ratio. Then, the sum of all directly received quality is multiplied by the capacity ratio to obtain the receiving capacity of the corresponding actual receiving object. There is no upper limit to the receiving capacity for residual receiving objects.

[0237] The objective of semi-constrained optimal transmission is to minimize the sum of the product of the quality of each modified key semantic unit transmitted to each receiving object and its corresponding receiving cost, while avoiding excessive concentration of transmission quality through an entropy regularization term. To calculate the entropy regularization term, the natural logarithm of each transmission quality value greater than 0 is calculated. The transmission quality is then multiplied by its natural logarithm, and the sum of all products is taken, with the negative value. This result is then multiplied by the entropy regularization coefficient. Finally, the sum of the transmission costs is added to the entropy regularization term to obtain the transmission target value.

[0238] The candidate values ​​for the entropy regularization coefficient are set from 0.01 to 0.10, increasing in increments of 0.01. Each candidate entropy regularization coefficient is used to solve for the validation samples. The direct matching accuracy and the identification accuracy of the remaining data to be aggregated are statistically analyzed in the validation samples. The candidate value corresponding to the highest average of these two accuracies is determined as the target entropy regularization coefficient. When multiple candidate values ​​have the same average value, the candidate value with the smaller transmission target value is selected.

[0239] The Sinkhorn iterative method is used to solve the semi-constrained optimal transport relationship. During initialization, all row and column scaling factors are set to 1, and a transport kernel is generated based on the extended acceptance cost matrix and entropy regularization coefficient. When generating the transport kernel, each acceptance cost is first divided by the entropy regularization coefficient, then the negative of the result is taken, and the corresponding natural exponent value is calculated.

[0240] In each iteration, the provisional output quality corresponding to each changed key semantic unit is first calculated based on the current column scaling factor. Then, the direct acceptance quality of the changed key semantic unit is divided by the provisional output quality to update the corresponding row scaling factor. Subsequently, the provisional reception quality corresponding to each actual receiving object is calculated based on the updated row scaling factor. When the provisional reception quality of a certain actual receiving object is greater than its reception capacity, the reception capacity is divided by the provisional reception quality to update the corresponding column scaling factor. When the provisional reception quality is less than or equal to the reception capacity, the corresponding column scaling factor remains unchanged. The column scaling factor of the residual receiving object is always kept at 1.

[0241] The transmission quality matrix is ​​calculated based on the updated row scaling factor, transmission kernel, and column scaling factor. The maximum absolute difference between the sum of the transmission quality of each row and the corresponding directly received quality is calculated, along with the maximum change in the corresponding element in the transmission quality matrix obtained from two consecutive iterations. Iteration stops when the maximum absolute difference is less than 0.000001, the maximum change in the matrix element is less than 0.000001, and this condition is met for five consecutive iterations. Iteration stops when the number of iterations reaches 500, and the result of the last iteration is used.

[0242] The quality transmitted to the actual receiving object in the transmission quality matrix is ​​determined as the direct receiving matching result, and the quality transmitted to the residual receiving object is determined as the residual to be aggregated.

[0243] Step S3053: Based on the quality of the inherited residual quantity and the residual quantity to be aggregated, determine the downstream content evolution edge, and perform weak explicit acceptance and aggregation along the downstream content evolution edge to obtain the residual quantity acceptance result.

[0244] The specific steps of step S3053 are as follows:

[0245] Step S30531: Based on the quality of the received residual quantity and the residual quantity to be aggregated, determine the final received residual quantity corresponding to each change key semantic unit.

[0246] In this embodiment, for each change key semantic unit, its corresponding residual quality and the residual to be aggregated are extracted.

[0247] The final received residual quantity is obtained by adding the received residual quantity to the residual quantity to be aggregated, which is the key semantic unit of the change.

[0248] When the final residual capacity is greater than the initial semantic quality of the corresponding key semantic unit of change, the initial semantic quality is determined as the final residual capacity to avoid the residual capacity quality exceeding the total semantic quality that the key semantic unit of change can be allocated.

[0249] The final residual quantities corresponding to each key semantic unit of change are uniformly collected.

[0250] Step S30532: Based on the change semantic inheritance knowledge graph, determine the downstream content evolution edge and generate candidate inheritance propagation paths corresponding to each change key semantic unit.

[0251] In this embodiment, the source content node to which each key semantic unit of change belongs is located from the semantic knowledge graph of change.

[0252] Starting from the source content node, perform a depth-first traversal along the direction of the content evolution edge, and record the content node identifiers, content evolution edge identifiers, and node order encountered during the traversal.

[0253] When a content node that has already obtained direct connection quality in the direct connection matching result appears in the traversal path, the path before that node is preserved, and the expansion along that node is stopped.

[0254] When the path length reaches the maximum evolution chain length threshold determined in step S203, the expansion stops.

[0255] For paths with identical content node sequences, deduplication is performed, and paths with larger time decay weight products of content evolution edges are retained to obtain candidate inheritance and propagation paths corresponding to each key semantic unit of change.

[0256] Step S30533: Based on the source-to-end acceptance matching cost, identify the weak explicit acceptance nodes in each candidate acceptance propagation path, determine the continuous distribution of the weak explicit acceptance nodes, and screen the effective acceptance propagation path.

[0257] In this embodiment, for each candidate acceptance propagation path, the candidate acceptance semantic units contained in each content node are extracted according to the order of the content nodes in the path.

[0258] Based on the source-direction acceptance matching cost corresponding to weak explicit acceptance semantic units and non-acceptance semantic units in the threshold calibration sample set, the weak explicit acceptance threshold is determined according to the unified threshold calibration method in step S301. For example, in this embodiment, the weak explicit acceptance threshold is preferably between 0.40 and 0.60.

[0259] Candidate acceptance semantic units whose source-to-acceptance matching cost is less than or equal to the weak explicit acceptance threshold and which have not obtained direct acceptance quality are identified as weak explicit acceptance semantic units, and their content nodes are identified as weak explicit acceptance nodes.

[0260] The continuous distribution length of weakly explicit accepting nodes represents the number of content nodes that consecutively contain weakly explicit accepting semantic units in the same candidate accepting propagation path, arranged in the order of content nodes. If there is a direct content evolution edge between two adjacent content nodes, and both content nodes contain at least one weakly explicit accepting semantic unit, then the two content nodes are counted in the same continuous segment. If a content node does not contain a weakly explicit accepting semantic unit, or there is no direct content evolution edge between adjacent content nodes, then the current continuous segment is terminated at that position, and the counting starts again from the subsequent weakly explicit accepting nodes. The maximum value among the number of content nodes contained in each continuous segment is determined as the continuous distribution length of weakly explicit accepting nodes for that candidate accepting propagation path.

[0261] Based on the effective weak explicit acceptance paths and invalid weak explicit acceptance paths in the threshold calibration sample set, integers from 1 to the maximum length of the candidate acceptance propagation path are sequentially used as candidate continuous distribution length thresholds. The number of misjudged paths and the number of missed paths are counted respectively, and the continuous distribution length threshold is determined according to the unified threshold calibration method in step S301. For example, in the embodiments of this application, the continuous distribution length threshold is preferably 1 to 3.

[0262] Candidate acceptance propagation paths with a continuous distribution length greater than or equal to the continuous distribution length threshold are identified as valid acceptance propagation paths.

[0263] Step S30534: Based on the effective transmission path, perform path-level allocation and content-level aggregation on the final transmission residual to obtain the residual transmission result.

[0264] In this embodiment, for each valid acceptance propagation path, the source-direction acceptance matching cost corresponding to each weak explicit acceptance node in the path is extracted.

[0265] Subtract the source-direction acceptance matching cost corresponding to each weak explicit acceptance node from 1 to obtain the acceptance score corresponding to each weak explicit acceptance node.

[0266] When calculating the path convergence score, first subtract the acceptance score of each weak explicit accepting node from 1 to obtain the non-acceptance ratio of the corresponding node. Then, multiply the non-acceptance ratios of each weak explicit accepting node in the same path in turn to obtain the non-acceptance ratio of the path. Finally, subtract the non-acceptance ratio of the path from 1 to obtain the weak explicit accepting convergence score of the effective accepting propagation path.

[0267] For all valid acceptance and propagation paths corresponding to the same key semantic unit of change, calculate the sum of the weak explicit acceptance convergence scores of each path.

[0268] When the sum of the weak explicit acceptance convergence scores is greater than 0, the weak explicit acceptance convergence scores of each effective acceptance propagation path are divided by this sum to obtain the residual allocation ratio of the corresponding path. Then, the final acceptance residual is multiplied by the residual allocation ratio to obtain the residual quality of the corresponding path allocation.

[0269] When the total score of weak explicit acceptance convergence is equal to 0, the final acceptance residual will be distributed equally according to the number of effective acceptance propagation paths.

[0270] For each content node in a valid transmission path, calculate the node's local contribution. During calculation, sum the transmission scores corresponding to each weakly explicit transmission semantic unit contained in the content node, then multiply the sum of the transmission scores by the time decay weight corresponding to the incoming edge of the content node to obtain the local contribution of that content node. For the starting node of the path, set the time decay weight to 1.

[0271] Calculate the sum of the local contributions of all content nodes in the same path. When the sum of local contributions is greater than 0, divide the local contribution of each content node by the sum of local contributions to obtain the content-level residual allocation ratio. Then multiply the residual quality allocated to the path by the content-level residual allocation ratio to obtain the residual carrying quality corresponding to each content node.

[0272] When the sum of local contributions equals 0, the residual quality of the path allocation is distributed evenly according to the number of content nodes in the path.

[0273] The residual acceptance quality obtained from different effective acceptance and propagation paths for the same content node is added together to obtain the residual acceptance result for each content node.

[0274] Step S3054: Based on the direct acceptance matching results and residual acceptance results, summarize the acceptance impact to obtain the key semantic inheritance relationship set and the content set affected by the change.

[0275] In this embodiment, for each downstream associated content object, the direct acceptance quality corresponding to each candidate acceptance semantic unit contained therein is summarized to obtain the total direct acceptance quality of the corresponding downstream associated content object.

[0276] The total residual quality is obtained by summing the residual quality of the downstream related content object in the residual quality transfer results.

[0277] The total quality of direct acceptance is added to the total quality of residual acceptance to obtain the total acceptance weight of the downstream related content object for the corresponding change key semantic unit.

[0278] Based on the content objects and modified key semantic units in the threshold calibration sample set that are manually labeled as having key semantic inheritance relationships and not having key semantic inheritance relationships, the total inheritance weight corresponding to each sample is extracted, and the key semantic inheritance threshold is determined according to the unified threshold calibration method in step S301. For example, in this embodiment, the key semantic inheritance threshold is preferably 0.10 to 0.30.

[0279] Content objects whose total inheritance weight is greater than or equal to the key semantic inheritance threshold are identified as key semantic inheritance relationships with the changed key semantic units. All key semantic inheritance relationships are then uniformly collected to obtain a key semantic inheritance relationship set.

[0280] For each downstream related content object, the total weight of all its corresponding change key semantic units is summarized to obtain the content impact value.

[0281] When calibrating the threshold for affected content, historical content objects are manually labeled according to whether distribution needs to be stopped, reorganized, or manually reviewed after the source content is changed. Content objects requiring any of the above actions are labeled as affected content, while content objects requiring no adjustment are labeled as unaffected content. Based on the content impact values ​​corresponding to the two types of content objects, the threshold for affected content is determined according to the unified threshold calibration method in step S301. For example, in this embodiment, the threshold for affected content is preferably between 0.20 and 0.40.

[0282] Downstream related content objects whose content impact value is greater than or equal to the threshold of affected content are identified as content affected by the change, and all content affected by the change is uniformly collected to obtain the set of content affected by the change.

[0283] Step S4: Based on the key semantic inheritance relationship set and the content set affected by the change, perform a change-driven content organization and distribution update to obtain the updated content organization relationship set and content distribution processing results.

[0284] The specific steps of step S4 are as follows:

[0285] Step S401: Based on the set of key semantic inheritance relationships and the set of content affected by the changes, construct a propagation diagram of the impact of content changes.

[0286] In this embodiment, all changed key semantic nodes in the key semantic inheritance relationship set are extracted, and all changed content nodes in the content set affected by the changes are extracted.

[0287] By merging and deduplicating the key semantic nodes and the content nodes affected by the changes, we obtain the node set of the content change impact propagation graph.

[0288] Traverse all key semantic inheritance relationships in the key semantic inheritance relationship set, and establish a directed semantic inheritance edge for each inheritance relationship. The direction of the edge is from the changed key semantic node to the content node affected by the change, and the weight of the edge is equal to the weight of the key semantic inheritance relationship.

[0289] Traverse all content evolution edges in the content evolution relationship set, filter out content evolution edges whose two ends both belong to the node set, retain their original direction and time decay weight, and use them as influence propagation edges between content.

[0290] By merging all semantic inheritance edges and content-related influence propagation edges, we obtain the edge set of the content change influence propagation graph.

[0291] Based on the set of nodes and the set of edges, a directed weighted graph of the propagation of the impact of content changes is constructed.

[0292] Step S402: Based on the content change impact propagation graph, perform a restart random walk impact propagation to obtain the impact propagation score corresponding to each affected content.

[0293] In this embodiment, the nodes in the content change impact propagation graph are sorted, and a transition matrix is ​​established based on the node sorting.

[0294] For a node with outgoing edges, first calculate the sum of the weights of all outgoing edges of the node, then divide the weight of each outgoing edge by the sum to obtain the transition probability of the node propagating along each outgoing edge.

[0295] For dangling nodes without outgoing edges, their transition probability is not set to 0. Instead, all propagation quality corresponding to the dangling node is redistributed to the key semantic nodes according to their weights in the initial influence vector. If the initial influence vector contains three key semantic nodes with corresponding weights of 0.50, 0.30, and 0.20, the transition probabilities of the dangling node will point to these three key semantic nodes, and will be set to 0.50, 0.30, and 0.20 respectively.

[0296] When setting the initial impact vector, the semantic change degree corresponding to each key semantic node is extracted. The sum of the semantic change degrees of all key semantic nodes is calculated, and the semantic change degree of each key semantic node is divided by this sum to obtain the initial impact weight for each key semantic node. The initial impact weight of non-key semantic nodes is set to 0. When the sum of the semantic change degrees of all key semantic nodes equals 0, the initial impact quality is evenly distributed among the key semantic nodes.

[0297] When calibrating the restart probability, values ​​between 0.05 and 0.50, increasing by 0.05, are sequentially used as candidate restart probabilities. These probabilities are calculated for each historical impact propagation sample, and the content impact level is determined based on the impact propagation score. The calculated impact level is compared with the manually labeled impact level. The number of misclassifications of high-impact content, recombined content, and low-impact content is counted. The candidate restart probability with the lowest calibration cost is selected according to the unified threshold calibration method in step S301. For example, in this embodiment, the restart probability is preferably between 0.10 and 0.30.

[0298] Use the initial influence vector as the current influence vector for the first iteration.

[0299] In each iteration, the current influence vector is propagated according to the transition matrix to obtain the graph propagation vector. The restart probability is subtracted from 1 to obtain the persistence propagation ratio. Each component of the graph propagation vector is multiplied by the persistence propagation ratio to obtain the persistence propagation part.

[0300] The restart component is obtained by multiplying each component in the initial influence vector by the restart probability.

[0301] The corresponding components in the continuous propagation part and the restart part are added together to obtain the influence vector for the next iteration.

[0302] After each iteration, the absolute difference between the corresponding components of the next iteration's influence vector and the current influence vector is calculated, and all absolute differences are summed to obtain the iteration change. Candidate values ​​for the convergence threshold are set to 0.00001, 0.0001, 0.001, and 0.01, respectively. Verification calculations are performed on each, and the candidate value with the largest value when the influence level identification result no longer changes is selected as the target convergence threshold to reduce invalid iterations. For example, in this embodiment, the convergence threshold is preferably between 0.0001 and 0.001.

[0303] The iteration stops when the change in iteration is less than the convergence threshold, and this condition is met for 5 consecutive iterations. The iteration stops when the number of iterations reaches 1000, and the influence vector from the last iteration is used.

[0304] Extract the vector components corresponding to each content node affected by the change when the iteration stops, and use them as the impact propagation score for each content affected by the change.

[0305] Step S403: Based on the impact propagation score, determine the impact level of each affected content to obtain the set of affected content levels.

[0306] In this embodiment, an impact level calibration sample is prepared based on historical source content change records and corresponding actual content handling records.

[0307] Content affected by a change must be immediately discontinued after the source content is changed, or continued distribution would lead to errors in business rules, scope of application, or compliance requirements, and is therefore marked as high-impact content. Content affected by a change needs to be regenerated, partially rewritten, or re-adapted before it can be distributed again, and is therefore marked as medium-impact content. Content affected by a change does not require modification, or its semantics do not affect the actual business expression and distribution purpose, and is therefore marked as low-impact content.

[0308] Extract the impact propagation score corresponding to the labeled samples at each impact level, and sort the impact propagation scores from smallest to largest.

[0309] Two different candidate thresholds are selected sequentially. The larger candidate threshold is designated as the high-impact threshold, and the smaller candidate threshold is designated as the medium-impact threshold. For each set of candidate thresholds, samples with an impact propagation score greater than or equal to the high-impact threshold are classified as high-impact content, samples with an impact propagation score greater than or equal to the medium-impact threshold but less than the high-impact threshold are classified as medium-impact content, and samples with an impact propagation score less than the medium-impact threshold are classified as low-impact content.

[0310] The number of high-impact content incorrectly classified as medium or low impact, the number of medium-impact content incorrectly classified as high or low impact, and the number of low-impact content incorrectly classified as high or medium impact were counted separately. Based on historical processing records, the average misclassification cost caused by missed high-impact content, the average reorganization cost caused by misclassification of medium-impact content, and the average manual review cost caused by misclassification of low-impact content were calculated. The number of misclassifications for each category was multiplied by the corresponding average cost, and the products were summed to obtain the hierarchical calibration cost corresponding to the candidate threshold group. The candidate threshold group with the minimum hierarchical calibration cost was selected as the target high-impact threshold and the target medium-impact threshold.

[0311] For business rules, applicable scope, and compliant content scenarios, when the hierarchical calibration cost corresponding to multiple candidate thresholds is the same, the candidate threshold group with fewer missed high-impact content is selected. For general publicity and display content scenarios, the candidate threshold group with fewer misjudged low-impact content is selected. For example, in the embodiments of this application, the high-impact threshold is preferably 0.60 to 0.80, and the medium-impact threshold is preferably 0.20 to 0.40.

[0312] For each piece of content affected by the change, its impact propagation score is compared with the high-impact threshold and the medium-impact threshold.

[0313] When the impact propagation score is greater than or equal to the high impact threshold, the corresponding content is identified as a high impact level.

[0314] When the impact propagation score is greater than or equal to the medium impact threshold and less than the high impact threshold, the corresponding content is determined to be at the medium impact level.

[0315] When the impact propagation score is less than the medium impact threshold, the corresponding content is determined to be at the low impact level.

[0316] The affected content and its corresponding impact level are uniformly collected to obtain the affected content level set.

[0317] Step S404: Based on the affected content hierarchy set and content evolution relationship set, perform hierarchical updates of organization and distribution relationships and mapping of distribution and disposal status to obtain the updated content organization relationship set and content distribution and disposal results.

[0318] The specific steps of step S404 are as follows:

[0319] Step S4041: Based on the affected content hierarchy set, divide each affected content into high-impact content, medium-impact content, and low-impact content.

[0320] In this embodiment, all content affected by the changes in the affected content hierarchy set is traversed.

[0321] Content marked as high-impact level is uniformly collected to obtain a high-impact content set.

[0322] Content marked as medium impact level is uniformly grouped to obtain a medium impact content set.

[0323] Content marked as low impact level is uniformly grouped to obtain a low impact content set.

[0324] Step S4042: Based on the content evolution relationship set, extract the content organization relationship and distribution association relationship corresponding to high-impact content, medium-impact content and low-impact content to obtain the organization relationship set to be updated and the distribution relationship set to be adjusted.

[0325] In this embodiment, all content in the high-impact content set, medium-impact content set, and low-impact content set is traversed.

[0326] Extract all content organization relationships corresponding to each piece of content from the content evolution relationship set. The content organization relationships include generation relationships, derivation relationships, and adaptation relationships.

[0327] Extract all distribution relationships corresponding to each piece of content from the content distribution system. These relationships include the relationship between content and distribution channels, and the relationship between content and user groups.

[0328] All extracted content organization relationships are merged and deduplicated to obtain the set of organization relationships to be updated.

[0329] All extracted distribution relationships are merged and deduplicated to obtain the set of distribution relationships to be adjusted.

[0330] Step S4043: Perform hierarchical updates on the set of organizational relationships to be updated and the set of distribution relationships to be adjusted, and determine the distribution handling status corresponding to each content affected by the change, thereby obtaining a distribution handling status set.

[0331] In this embodiment, for high-impact content, all valid distribution associations in the distribution relationship set to be adjusted are cancelled, while its historical distribution records are retained.

[0332] In the set of organizational relationships to be updated, retain historical organizational relationships with high-impact content, but mark them as invalid.

[0333] Set the distribution status of high-impact content to "stop distribution".

[0334] For content affected by this, mark its distribution status as paused in the distribution relationship set to be adjusted.

[0335] In the set of organizational relationships to be updated, retain all organizational relationships that affect the content and mark them as pending updates.

[0336] Set the distribution status of the affected content to reorganized distribution.

[0337] For low-impact content, all valid distribution relationships in the distribution relationship set to be adjusted are kept unchanged.

[0338] Keep all organizational relationships in the set of organizational relationships to be updated unchanged.

[0339] Set the distribution status of low-impact content to continue distribution.

[0340] The distribution and handling status of all content affected by the changes are collected and unified to obtain a distribution and handling status set.

[0341] Step S4044: Based on the distribution and disposal status set, perform state mapping and result aggregation on the hierarchically updated organizational relationship set to be updated and the distribution relationship set to be adjusted to obtain the updated content organizational relationship set and content distribution and disposal results.

[0342] In this embodiment, the distribution and disposal status of each content in the distribution and disposal status set is mapped to the corresponding content in the organizational relationship set to be updated and the distribution relationship set to be adjusted.

[0343] All mapped content organization relationships are uniformly collected to obtain an updated content organization relationship set, which includes the validity status, update status, and historical organization relationships of each piece of content.

[0344] All distribution relationships after mapping are uniformly collected to obtain content distribution processing results, which include continuing to distribute content sets, reorganizing and distributing content sets, and stopping the distribution of content sets.

[0345] Example 2:

[0346] See Figure 2 This embodiment provides an AIGC content organization and distribution system, including: a data acquisition module, an evolutionary relationship module, a graph construction module, and an organization and distribution module.

[0347] The data acquisition module is used to collect AIGC content evolution datasets and source content change records.

[0348] The evolutionary relationship module constructs content evolution associations based on the AIGC content evolution dataset, thereby obtaining a content evolution relationship set.

[0349] The graph construction module constructs a change semantic inheritance knowledge graph based on the content evolution relationship set and source content change records, and performs residual retention inheritance matching to obtain the key semantic inheritance relationship set and the content set affected by the change.

[0350] The organization and distribution module performs content organization and distribution updates driven by changes based on the key semantic inheritance relationship set and the content set affected by changes, and obtains the updated content organization relationship set and content distribution processing results.

[0351] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0352] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0353] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AIGC content organization and distribution method, characterized in that, include: Collect AIGC content evolution dataset and source content change records. The AIGC content evolution dataset includes source content, derived content, distributed content, generation records, and adaptation records. Based on the AIGC content evolution dataset, we construct content evolution relationships to obtain a content evolution relationship set; Based on the content evolution relationship set and source content change records, a change semantic inheritance knowledge graph is constructed, and residual retention inheritance matching is performed to obtain the key semantic inheritance relationship set and the content set affected by the change. Based on the key semantic inheritance relationship set and the content set affected by the change, the content organization and distribution are updated driven by the change, resulting in the updated content organization relationship set and content distribution processing results.

2. The AIGC content organization and distribution method according to claim 1, characterized in that, Based on the AIGC content evolution dataset, a content evolution relationship set is constructed, including: Based on the generation records, the generation association between the source content and the derived content, as well as the regeneration association between each derived content, are established to obtain the generation relationship set; Based on the adaptation record, establish cross-platform adaptation relationships between derived content and distributed content to obtain a set of adaptation relationships; Based on the generation relation set and the adaptation relation set, the evolution chain of source content, derived content and distributed content is organized to obtain the content evolution relation set.

3. The AIGC content organization and distribution method according to claim 1, characterized in that, Based on the content evolution relationship set and source content change records, a change semantic inheritance knowledge graph is constructed, and residual retention inheritance matching is performed to obtain a key semantic inheritance relationship set and a content set affected by changes, including: Based on the source content change record, extract the content fragments before and after the change corresponding to the source content, and perform semantic difference recognition to obtain the set of key semantic units of change and the corresponding set of semantic change degree; Based on the content evolution relationship set, the content objects that have downstream evolution associations with each key semantic unit of change are identified, and the inheriting semantics are extracted to obtain the downstream associated content object set and the candidate inheriting semantic unit set. Based on the set of key semantic units for change, the set of downstream related content objects, and the set of candidate semantic units for inheritance, a knowledge graph of semantic inheritance for change is constructed. Based on the knowledge graph of semantic connection, directional semantic connection metric learning is performed to obtain the source-to-source connection matching cost; Based on the source-to-end matching cost and semantic change degree set, we perform end-to-end quality diversion and residual retention semi-constrained transmission matching, and weak explicit end-to-end aggregation to obtain the key semantic inheritance relationship set and the content set affected by changes.

4. The AIGC content organization and distribution method according to claim 3, characterized in that, The construction of a change semantic inheritance knowledge graph based on the change key semantic unit set, the downstream related content object set, and the candidate inheritance semantic unit set includes: Based on the set of key semantic units for change, the set of candidate inheriting semantic units and the set of downstream related content objects, node mapping and semantic content attribution are performed to obtain the set of semantic nodes, the set of content nodes and the set of semantic content attribution relationships. Based on the content evolution relationship set, content evolution edges are established between each content node to obtain the content evolution edge set; Based on the semantic content attribution relation set, content inclusion relation edges are established to obtain the content inclusion relation edge set; Based on the source content change records, establish source-side semantic change relationship edges to obtain a set of semantic change relationship edges; Based on semantic node set, content node set, content evolution edge set, content inclusion relation edge set, and semantic change relation edge set, a change semantic inheritance knowledge graph is constructed.

5. The AIGC content organization and distribution method according to claim 3, characterized in that, The method of performing directional semantic connection metric learning based on the modified semantic connection knowledge graph to obtain the source-to-source connection matching cost includes: Based on the semantic knowledge graph of change and the content evolution relationship set, the content evolution direction is determined, and the semantic connection relationship between each key semantic unit of change and each candidate semantic unit of connection is identified along the content evolution direction to obtain the positive connection sample set. Based on the forward acceptance sample set, the semantic correspondence between the candidate acceptance semantic units and the corresponding change key semantic units is determined as the reverse pseudo acceptance sample, thus obtaining the reverse pseudo acceptance sample set. Based on the positive succession sample set, candidate succession semantic units that are on the same topic as each change key semantic unit but do not form a semantic succession relationship are screened to obtain the non-succession sample set on the same topic; Based on the positive acceptance sample set, the reverse pseudo acceptance sample set, and the non-acceptance sample set of the same topic, the sample acceptance cost ranking is determined, and a directional acceptance measurement target is constructed. Based on the directional acceptance metric objective, metric learning is performed on the set of change key semantic units and the set of candidate acceptance semantic units to obtain the source-to-source acceptance matching cost.

6. The AIGC content organization and distribution method according to claim 3, characterized in that, The process involves source-to-source matching cost and semantic change degree set, followed by quality-based traffic splitting and residual-preservation semi-constrained transmission matching, and weak explicit aggregation to obtain a key semantic inheritance relationship set and a set of content affected by changes, including: Based on the source-to-source matching cost and the set of semantic change degrees, the direct acceptance quality and acceptance residual quality corresponding to each key semantic unit of change are determined. Based on the candidate acceptance semantic unit set, direct acceptance quality and source-to-acceptance matching cost, a set of acceptance and receiving objects containing residual acceptance objects is constructed, and semi-constrained optimal transmission matching is performed to obtain the direct acceptance matching result and the residual to be aggregated. Based on the quality of the inherited residual and the residual to be aggregated, the downstream content evolution edge is determined, and weak explicit inheritance and aggregation are carried out along the downstream content evolution edge to obtain the residual inheritance result. Based on the direct acceptance matching results and residual acceptance results, the impact of acceptance is summarized to obtain the key semantic inheritance relationship set and the content affected by the change set.

7. The AIGC content organization and distribution method according to claim 6, characterized in that, The process of determining the downstream content evolution edge based on the quality of the inherited residual and the residual to be aggregated, and performing weak explicit acceptance and aggregation along the downstream content evolution edge to obtain the residual acceptance result includes: Based on the quality of the received residual and the residual to be aggregated, determine the final received residual corresponding to each key semantic unit of change; Based on the semantic knowledge graph of change, downstream content evolution edges are determined, and candidate inheritance and propagation paths corresponding to each key semantic unit of change are generated. Based on the source-to-end matching cost, weak explicit acceptance nodes in each candidate acceptance propagation path are identified, the continuous distribution of weak explicit acceptance nodes is determined, and effective acceptance propagation paths are selected. Based on the effective transmission path, the final residual data is allocated at the path level and aggregated at the content level to obtain the residual data transfer result.

8. The AIGC content organization and distribution method according to claim 1, characterized in that, The process of updating the content organization and distribution based on the key semantic inheritance relationship set and the content set affected by the change, driven by the change, yields the updated content organization relationship set and content distribution processing results, including: Based on the set of key semantic inheritance relationships and the set of content affected by changes, a propagation graph of the impact of content changes is constructed. Based on the content change impact propagation graph, a restart random walk impact propagation is performed to obtain the impact propagation score corresponding to each content affected by the change. Based on the impact propagation score, the impact level of each affected content is determined, resulting in a set of affected content levels. Based on the affected content hierarchy set and the content evolution relationship set, the organization and distribution relationship is updated hierarchically and the distribution and disposal status is mapped to obtain the updated content organization relationship set and content distribution and disposal results.

9. The AIGC content organization and distribution method according to claim 8, characterized in that, The process involves updating the organizational distribution relationship hierarchically and mapping the distribution and handling status based on the affected content hierarchy set and content evolution relationship set, resulting in an updated content organizational relationship set and content distribution and handling results, including: Based on the affected content hierarchy set, each affected content is divided into high-impact content, medium-impact content, and low-impact content; Based on the content evolution relationship set, the content organization relationship and distribution relationship corresponding to high-impact content, medium-impact content and low-impact content are extracted to obtain the organizational relationship set to be updated and the distribution relationship set to be adjusted. The set of organizational relationships to be updated and the set of distribution relationships to be adjusted are updated hierarchically, and the distribution handling status corresponding to each content affected by the change is determined to obtain a set of distribution handling statuses. Based on the distribution and disposal status set, the status mapping and result aggregation are performed on the hierarchically updated set of organizational relationships to be updated and the set of distribution relationships to be adjusted to obtain the updated set of content organizational relationships and content distribution and disposal results.

10. An AIGC content organization and distribution system, used to implement the AIGC content organization and distribution method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect AIGC content evolution datasets and source content change records; The evolutionary relationship module, based on the AIGC content evolution dataset, constructs content evolutionary relationships to obtain a content evolutionary relationship set; The graph construction module constructs a semantic inheritance knowledge graph based on the content evolution relationship set and source content change records, and performs residual retention inheritance matching to obtain the key semantic inheritance relationship set and the content set affected by the change. The organization and distribution module updates the content organization and distribution based on the key semantic inheritance relationship set and the content set affected by the change, and obtains the updated content organization relationship set and content distribution processing results.

Images