Enterprise internet short video interactive information search method

Abstract

The application discloses an enterprise Internet short video interactive information search method, and belongs to the short video information retrieval field. The method collects short video interactive information, time stamps, account identifiers and user historical behavior data, constructs a heterogeneous graph, calculates a real-time credibility score through a graph attention network, reduces the weight of low-credibility interactive information, maps interactive text to an enterprise knowledge graph semantic space through a cross-domain semantic mapping matrix trained through comparative learning to complete entity alignment, generates a time slice level index unit in combination with a time axis relationship, dynamically weights and sorts long-tail short videos in combination with content semantic benchmark scores and incremental confidence scores, and the application narrows the semantic gap between interactive text and enterprise professional knowledge, alleviates the long-tail video exposure unfairness problem, realizes real-time authenticity discrimination in the search link, and improves the search result accuracy and credibility.

Description

Technical Field

[0001] This invention relates to the field of short video information retrieval technology, specifically to a method for searching interactive information via short videos on the internet for enterprises, falling under the category of enterprise internet information retrieval and artificial intelligence semantic analysis technology. Background Technology

[0002] Short video platforms typically build inverted indexes for interactive text such as bullet comments and comments. After users enter keywords, the system returns search results through keyword matching and a ranking model based on collaborative filtering features such as the number of likes and comments. The ranking process of recommendation systems generally relies heavily on the richness of historical behavioral data. Existing anti-fraud technologies are usually deployed in the review process after content is published, and post-processing of comments is carried out by rule engines or classification models.

[0003] The closest existing technical solution to this invention is a short video interactive information retrieval and ranking method based on keyword matching and collaborative filtering feature fusion. This method performs word segmentation on the bullet comments and comment text of short videos and establishes an inverted index. After receiving a user query request, it retrieves candidate interactive information and its corresponding videos from the inverted index through keyword matching. Based on collaborative filtering features such as the historical number of likes, comments, and playback completion rate of candidate videos, it calculates a ranking score and returns the search results list in descending order of the ranking score. For interactive information published by comment accounts marked as abnormal by the platform's content review system, it is deleted from the index after the offline review period ends.

[0004] The existing technical solutions have the following shortcomings: First, the keyword matching and recall module only performs surface text matching and does not perform cross-domain semantic understanding of interactive text. When interactive text contains special colloquial expressions such as irony or derogatory connotations, there is a semantic gap between the matching results and the user's true search intent, resulting in low search accuracy. Second, the ranking module relies heavily on collaborative filtering features such as the number of likes and comments. Enterprise long-tail professional short videos naturally have the problem of sparse historical behavioral data. Even if the semantic quality of such videos is excellent, they will be systematically suppressed in the ranking due to the lack of behavioral feedback, resulting in unfair exposure. Third, the anti-fraud mechanism is deployed in the offline review stage after the content is published, which has obvious time lag. Noise signals such as fake clicks and fake comments have already participated in the scoring calculation of the search ranking stage before the review is completed, causing the objectivity and credibility of the search results to be continuously polluted during the review period. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing enterprise internet short video interactive information search methods lack cross-domain semantic understanding capabilities in keyword matching, resulting in a semantic gap between colloquial interactive text and structured professional knowledge of enterprises. The ranking process heavily relies on historical behavioral data, leading to unfair exposure of short videos from long-tail enterprises due to data sparsity. Furthermore, anti-fraud mechanisms are deployed only in the post-audit stage, causing false interactive signals to pollute the search and ranking results before the audit is completed.

[0006] To address the aforementioned technical problems, this invention provides a method for enterprise internet short video interactive information search, comprising:

[0007] Collect interactive information corresponding to short videos, as well as the sending timestamp, sending account identifier, and user historical behavior data corresponding to the interactive information. The interactive information includes bullet screen text and comment text.

[0008] A heterogeneous graph is constructed based on the sending timestamp, sending account identifier, user historical behavior data, and text content of interactive information. The heterogeneous graph includes user nodes, video nodes, and interactive information nodes. The heterogeneous graph is input into a graph attention network to aggregate multi-layer neighbor node information to obtain the authenticity and credibility score of each interactive message.

[0009] The authenticity and credibility score is compared with a preset threshold. For interactive information with an authenticity and credibility score lower than the preset threshold, a reduced retrieval ranking contribution weight is calculated based on the authenticity and credibility score. The reduced retrieval ranking contribution weight is then used in the retrieval index building stage and the ranking stage.

[0010] The text content of the interactive information filtered by the authenticity and credibility score is semantically embedded to obtain the interactive text semantic embedding vector. The interactive text semantic embedding vector is mapped to the semantic space of the enterprise knowledge graph through a cross-domain semantic mapping matrix obtained in advance based on contrastive learning to obtain the mapping vector. The nearest neighbor search algorithm is used to determine the knowledge graph entities in the semantic space of the enterprise knowledge graph that have a similarity to the mapping vector that meet the preset similarity conditions, and these entities are used as knowledge graph entities aligned with the interactive information.

[0011] Based on the alignment relationship between the sending timestamp of the interactive information and the time segments obtained by dividing the audio transcribed text of the corresponding short video along the time axis, the aligned knowledge graph entities and the semantic features of the interactive information are bound to the corresponding short video time segments to generate triples in which time segments, knowledge graph entities, and interactive semantic features are interconnected, which serve as retrieval index units.

[0012] Determine whether the target short video is a long-tail short video with sparse historical behavior data. If so, calculate the similarity between the multimodal content representation vector of the target short video and the semantic vector of the associated enterprise knowledge graph entity to obtain a semantic quality benchmark score. Obtain the average similarity between the semantic vector of the interaction information generated by the target short video with an authenticity and credibility score higher than the preset threshold and the aligned semantic vector of the knowledge graph entity to obtain an incremental confidence score. Calculate the dynamic fusion weight of the benchmark score and the incremental confidence score based on the number of interaction information generated by the target short video with an authenticity and credibility score higher than the preset threshold. Sum the semantic quality benchmark score and the incremental confidence score according to the dynamic fusion weight to obtain the ranking score of the target short video.

[0013] The system receives natural language query requests from enterprise users and, based on the retrieval index unit and the sorting list including the sorting score, returns fragment-level search results that match the natural language query request.

[0014] The beneficial effects of this invention are as follows:

[0015] 1. By using graph attention networks to identify the authenticity of interactive information in real time, and by proactively downgrading low-credibility interactive information during the index building and ranking stages, compared with the existing post-audit mechanism, it can suppress the pollution of ranking results by noise signals such as fake traffic and fake comments from the source during the retrieval process, thereby improving the objectivity and credibility of search results.

[0016] 2. By combining the cross-domain semantic mapping matrix trained through comparative learning with the language model encoder fine-tuned by pragmatic intent, it can identify special colloquial expressions such as irony and positive / negative connotations and accurately align them to entities in the enterprise knowledge graph. At the same time, it can achieve fine-grained time-segment level indexing by leveraging time axis alignment relationships, narrowing the semantic gap between colloquial interactive text and enterprise structured professional knowledge, and improving the fine-grained usability of search results.

[0017] 3. By integrating the content semantic benchmark score and the incremental confidence score of sparse high-confidence interaction signals into a dynamic weighted ranking mechanism, the short videos of long-tail enterprises can still obtain a fair ranking opportunity based on the content semantic quality in the stage of sparse historical behavioral data, and gradually transition to interaction signal dominance as real interaction signals accumulate, thus alleviating the problem of unfair exposure of long-tail videos.

[0018] 4. The authenticity and credibility score is reused across the three stages of retrieval weight reduction, semantic alignment input filtering, and long-tail cold start incremental confidence score calculation, forming a cross-module collaborative technical closed loop, which improves the overall consistency and engineering efficiency of the system. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating the overall system architecture of the present invention.

[0020] Figure 2 This is a flowchart of the interactive information authenticity verification and retrieval weight reduction process of the present invention;

[0021] Figure 3 This is a flowchart of the semantic quality assessment process for long-tail cold start in this invention. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] System Overall Architecture

[0024] The enterprise internet short video interactive information search method provided in this embodiment consists of a data acquisition and preprocessing layer, an interactive information authenticity verification module, a cross-domain semantic alignment and fine-grained time segment anchoring module, a long-tail cold start semantic quality assessment module, a retrieval index and online ranking module, and an enterprise knowledge graph. The data acquisition and preprocessing layer obtains video frame data, audio transcribed text, bullet screen and comment text, as well as their sending timestamps and user account identifiers through the open interface of the short video platform. Its output is sent to the interactive information authenticity verification module and the cross-domain semantic alignment and time segment anchoring module, respectively. The credibility score and weighting coefficient output by the authenticity verification module are simultaneously fed back to the cross-domain semantic alignment module and the long-tail cold start semantic quality assessment module. The cross-domain semantic alignment module and the enterprise knowledge graph module perform entity alignment interaction. The time segment-entity-semantic triple output by the cross-domain semantic alignment module and the ranking score output by the long-tail cold start semantic quality assessment module are jointly incorporated into the retrieval index and online ranking module. The retrieval index and online ranking module receives natural language query requests from enterprise users and outputs a list of segment-level search results.

[0025] Verification of the authenticity of interactive information and demotion of search results

[0026] Collect the sending timestamps, sending account identifiers and historical behavior sequences of each bullet screen comment and interactive information in short videos, interactive text content, and user social relationship data associated with the interactive information to construct a heterogeneous graph. The node set includes user nodes, video nodes and interactive information nodes, and the edge set includes interaction edges between users and videos, social relationship edges between users, and belonging edges between interactive information and its video.

[0027] An initial feature vector is extracted for each interactive information node. The initial feature vector integrates three types of sub-features: spatiotemporal distribution features, specifically the time difference between the sending time of the interactive information and the video posting time, and the statistical measure of the time interval between two adjacent sending of interactive information by the same account; text semantic features, which are semantic vectors obtained by encoding the interactive text through a pre-trained language model; and user historical behavior features, specifically the registration duration of the sending account, the frequency of historical posts, and the entropy of the diversity of historical interactive content.

[0028] The initial feature vectors are input into a multi-layer graph attention network, which aggregates neighbor node information layer by layer to update node representations. Layer nodes The feature update formula is:

[0029] ;

[0030] in, Indicates the first Layer nodes The feature representation vector, Represents a node The set of neighboring nodes in a heterogeneous graph Indicates the first Layer nodes Its neighboring nodes Attention weight coefficients between them Indicates the first The learnable weight matrix of the layer, This represents a non-linear activation function.

[0031] The attention weighting coefficient is calculated as follows:

[0032] ;

[0033] in, This represents a learnable attention vector. This represents the concatenation operation of two vectors. This represents a linear activation function with leakage rectification.

[0034] Will pass The final feature representation of the interactive information node after layer aggregation is the output layer consisting of a fully connected layer and a Sigmoid activation function. The authenticity and credibility score of this interactive information is calculated as follows:

[0035] ;

[0036] in, Represents interactive information nodes The authenticity and credibility score ranges from [0,1]. A higher score indicates that the interaction information is more likely to be expressed by a genuine user. This represents the weight matrix of the output layer. This represents the bias term of the output layer. Indicates the first Layer nodes The feature representation vector.

[0037] credibility score Compared with a preset threshold, the retrieval ranking contribution weight is calculated for interactive information below the threshold after weight reduction:

[0038] ;

[0039] in, Indicates interactive information The reduced weight when participating in search ranking This indicates the overparameter of the weight reduction intensity. The larger the value, the stronger the suppression effect of low-credibility interaction information in the ranking. This weight is continuously used in the retrieval index construction stage and the subsequent ranking stage. In the model training stage, a small number of manually labeled real interaction samples and fake interaction samples are used as supervision signals. The parameters of the graph attention network are trained end-to-end through the cross-entropy loss function, and the model is incrementally updated with newly labeled data at a preset period to adapt to the continuous evolution of fake interaction behavior patterns.

[0040] Cross-domain semantic alignment and fine-grained temporal segment anchoring

[0041] The product entities, technical term entities, and their relationship descriptions in the enterprise knowledge graph, as well as the interactive text filtered for authenticity and the audio transcription of the corresponding videos, are vectorized using a fine-tuned large language model encoder to obtain their respective semantic embedding representations. A cross-domain contrastive learning training sample set is constructed. Semantically related interactive texts within the same time segment are paired with enterprise knowledge graph entities to form positive sample pairs, while randomly sampled semantically unrelated interactive texts are paired with knowledge graph entities to form negative sample pairs. Based on this sample set, the cross-domain semantic mapping matrix is ​​trained using the following contrastive learning loss function:

[0042] ;

[0043] in, The semantic embedding vector representing interactive text. Indicates and Semantically relevant positive sample knowledge graph entity embedding vectors Indicates the first The negative sample knowledge graph entity embedding vectors were sampled in total. One negative sample, This represents the cross-domain semantic mapping matrix to be learned. Represents the cosine similarity function. This represents the temperature coefficient, used to control the smoothness of the similarity distribution.

[0044] During the fine-tuning phase of the large language model, a sample dataset labeled with true sentiment polarity and pragmatic intent (including special expression types such as irony and positive / negative connotations) is introduced to supervise the encoder's fine-tuning. This ensures that the encoder's output semantic vectors reflect the deep semantics corrected for pragmatic intent, rather than the surface literal semantics. The trained mapping matrix is ​​then used to refine this process. The semantic embedding vectors applied to the cleaned interactive text are mapped to the semantic space of the enterprise knowledge graph. The nearest neighbor search method is used to find entities in the entity embedding set of the enterprise knowledge graph that meet the preset similarity conditions, thus aligning the interactive text with the knowledge graph entities. Combined with the sending timestamp carried by the interactive text itself, it is aligned and matched with each time segment after the video and audio transcribed text is divided according to the time axis. The obtained aligned entity labels and interactive semantic features are bound to the corresponding video time segments to generate time segment-entity-interaction semantic triplets, which serve as index units for subsequent segment-level retrieval.

[0045] Semantic quality assessment of long-tailed cold start by incorporating sparse belief signals

[0046] Visual feature vectors from video frames and semantic feature vectors from audio-transcribed text of short videos from long-tail enterprises are extracted. These two feature vectors are then fused to obtain the multimodal content representation vector of the video. The similarity between this vector and the semantic vector of the enterprise knowledge graph entity associated with the video is calculated to obtain the semantic quality benchmark score:

[0047] ;

[0048] in, This represents the semantic quality benchmark score of the video. This indicates the first element in the enterprise knowledge graph associated with the video. The semantic embedding vectors of each entity are associated with each other. Individual entities.

[0049] Determine whether the video has generated interactive information that has been confirmed as highly credible by the authenticity verification module. If so, calculate the average similarity between the semantic vectors of these interactive information and the semantic vectors of the aligned knowledge graph entities, and use this as the incremental confidence score.

[0050] ;

[0051] in, Indicates the incremental confidence score. This indicates the number of highly credible interactive messages generated by the video. Indicates the first The authenticity and credibility score of each interactive message Indicates the first The semantic embedding vector of each interactive message, This represents the semantic vector of the knowledge graph entity corresponding to the interaction information after alignment.

[0052] Based on the amount of highly credible interactive information accumulated in this video Calculate the dynamic fusion weights of the baseline score and the incremental confidence score:

[0053] ;

[0054] in, Indicates the amount of interactive information. The relevant dynamic weighting function, This represents the decay rate hyperparameter, used to control how quickly the weights decay as the amount of interactive information increases. The larger, The smaller the value, the more the ranking and scoring gradually shifts from being dominated by content semantics to being dominated by interaction signals as real interaction signals accumulate.

[0055] By combining the baseline score and the incremental confidence score, the ranking score for this long-tail video is obtained:

[0056] ;

[0057] If the video does not contain any highly credible interactive information, then directly... The calculated ranking score for this video will be included in the global search ranking list along with non-long-tail videos, and will participate in subsequent online search ranking processes.

[0058] Specific application examples

[0059] Taking a short product demonstration video released by an industrial valve manufacturer as an example, the implementation process of this invention is illustrated. After the short video was released, it received several bullet comments and messages. One of the comments was a colloquial expression of affirmation regarding the valve's sealing performance. At the same time, there were several comments with similar content and very short intervals between posts, which appeared to be machine-generated. The system first constructs a heterogeneous graph containing the sending account, the short video, and various interactive information according to the aforementioned authenticity verification process. It extracts the spatiotemporal distribution features, textual semantic features, and user historical behavior features of each interactive information node. After aggregation by a multi-layer graph attention network, it outputs the authenticity and credibility score of each interactive information. The interactive information suspected of being machine-generated receives a lower authenticity and credibility score and is calculated with a lower retrieval ranking contribution weight according to the aforementioned weighting formula, thus effectively suppressing its role in subsequent retrieval ranking. The colloquial comments expressing affirmation retain their original weight for subsequent processing.

[0060] Subsequently, the system maps the semantic embedding vector of the colloquial comment to the semantic space of the enterprise knowledge graph through a cross-domain semantic mapping matrix. Nearest neighbor search is used to align the comment to the corresponding technical term entity "sealing performance" in the enterprise knowledge graph. Combining the comment's sending timestamp with the timeline alignment of the short video's audio transcription, the system binds the entity and its interactive semantic features to the specific time segment demonstrating the sealing structure in the short video, generating a corresponding time segment-entity-interaction semantic triplet and writing it into the retrieval index. Due to the short video's short publication time and sparse historical behavior data, the system classifies it as a long-tail short video and extracts its semantic features. The system calculates the semantic quality baseline score by using the multimodal content representation vector of the short video and its similarity to entities in the associated knowledge graph. It then calculates the incremental confidence score by combining the generated high-confidence interaction information and calculates the dynamic fusion weight based on the amount of accumulated high-confidence interaction information. Finally, it obtains the ranking score of the short video and includes it in the global retrieval ranking list. When an enterprise user enters a natural language query request related to valve sealing performance, the system performs a nearest neighbor search in the vector database, recalls the aforementioned time segment-entity-interaction semantic triplet, and returns segment-level retrieval results including video identifier, corresponding time segment interval, and matching entity by combining the ranking list.

[0061] Alternative implementation methods

[0062] In the process of verifying the authenticity of interactive information, as an alternative implementation method, a sequence self-attention model based on Transformer can be used instead of graph attention network. The historical interaction sequence of the same user and the interaction information sequence of the same video are encoded as sequence features respectively. The time sequence abnormal patterns are captured by the self-attention mechanism and combined with the text semantic features for fusion classification, and the authenticity credibility score is output to reduce the engineering complexity of heterogeneous graph construction and maintenance.

[0063] In the long-tail cold start semantic quality assessment stage, as an alternative implementation method, instead of using a fixed form of exponentially decaying dynamic weight function, features such as the amount of interactive information and the account's historical activity level can be used as inputs to train a lightweight weight prediction network. This network directly outputs the fusion weights of the baseline score and the incremental confidence score, and adaptively learns the fusion strategy for different categories and enterprise scenarios through a data-driven approach.

[0064] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for enterprise internet short video interactive information search, characterized in that, include: Collect interactive information corresponding to short videos, as well as the sending timestamp, sending account identifier, and user historical behavior data corresponding to the interactive information. The interactive information includes bullet screen text and comment text. A heterogeneous graph is constructed based on the sending timestamp, sending account identifier, user historical behavior data, and text content of interactive information. The heterogeneous graph includes user nodes, video nodes, and interactive information nodes. The heterogeneous graph is input into a graph attention network to aggregate multi-layer neighbor node information to obtain the authenticity and credibility score of each interactive message. The authenticity and credibility score is compared with a preset threshold. For interactive information with an authenticity and credibility score lower than the preset threshold, a reduced retrieval ranking contribution weight is calculated based on the authenticity and credibility score. The reduced retrieval ranking contribution weight is then used in the retrieval index building stage and the ranking stage. The text content of the interactive information filtered by the authenticity and credibility score is semantically embedded to obtain the interactive text semantic embedding vector. The interactive text semantic embedding vector is mapped to the semantic space of the enterprise knowledge graph through a cross-domain semantic mapping matrix obtained in advance based on contrastive learning to obtain the mapping vector. The nearest neighbor search algorithm is used to determine the knowledge graph entities in the semantic space of the enterprise knowledge graph that have a similarity to the mapping vector that meet the preset similarity conditions, and these entities are used as knowledge graph entities aligned with the interactive information. Based on the alignment relationship between the sending timestamp of the interactive information and the time segments obtained by dividing the audio transcribed text of the corresponding short video along the time axis, the aligned knowledge graph entities and the semantic features of the interactive information are bound to the corresponding short video time segments to generate triples in which time segments, knowledge graph entities, and interactive semantic features are interconnected, which serve as retrieval index units. Determine whether the target short video is a long-tail short video with sparse historical behavior data. If so, calculate the similarity between the multimodal content representation vector of the target short video and the semantic vector of the associated enterprise knowledge graph entity to obtain a semantic quality benchmark score. Obtain the average similarity between the semantic vector of the interaction information generated by the target short video with a authenticity credibility score higher than the preset threshold and the aligned semantic vector of the knowledge graph entity to obtain an incremental confidence score. Based on the number of interaction information generated by the target short video with a authenticity credibility score higher than the preset threshold, calculate the dynamic fusion weight of the semantic quality benchmark score and the incremental confidence score. Perform a weighted sum of the semantic quality benchmark score and the incremental confidence score based on the dynamic fusion weight to obtain the ranking score of the target short video. The system receives natural language query requests from enterprise users and, based on the retrieval index unit and the sorting list including the sorting score, returns fragment-level retrieval results that match the natural language query request. The fragment-level retrieval results include video identifiers, time segment intervals, and matching entities.

2. The method according to claim 1, characterized in that, When constructing the heterogeneous graph, an initial feature vector is extracted for each interactive information node. The initial feature vector integrates spatiotemporal distribution features, textual semantic features, and user historical behavior features. The spatiotemporal distribution features include the time difference between the sending time of the interactive information and the posting time of the corresponding short video, and the statistical measure of the time interval between two adjacent sending of interactive information by the same sending account. The textual semantic features are semantic vectors obtained by encoding the text content of the interactive information through a pre-trained language model. The user historical behavior features include the registration duration of the sending account, the frequency of historical posts, and the entropy of the diversity of historical interactive content.

3. The method according to claim 2, characterized in that, In the graph attention network, the first... Layer nodes The features are updated in the following way: ； in, Indicates the first Layer nodes The feature representation vector, Represents a node The set of neighboring nodes in the heterogeneous graph, Indicates the first Layer nodes Its neighboring nodes Attention weight coefficients between them Indicates the first The learnable weight matrix of the layer, Represents a nonlinear activation function; Will pass The feature representation input of the layered interactive information node is the output layer composed of a fully connected layer and a sigmoid activation function, which yields the authenticity and credibility score: ； in, Represents interactive information nodes The authenticity and credibility score ranges from [0,1]. This represents the weight matrix of the output layer. This represents the bias term of the output layer. Indicates the first Layer nodes The feature representation vector.

4. The method according to claim 1, characterized in that, The reduced retrieval ranking contribution weight is calculated in the following way: ； in, Indicates interactive information The reduced weight when participating in search ranking Indicates interactive information The authenticity and credibility score, This represents the hyperparameter of the weight reduction intensity.

5. The method according to claim 1, characterized in that, The cross-domain semantic mapping matrix is ​​trained as follows: Semantically related interactive information texts within the same time segment are paired with enterprise knowledge graph entities to form positive sample pairs; randomly sampled semantically unrelated interactive information texts are paired with knowledge graph entities to form negative sample pairs. Based on these positive and negative sample pairs, the cross-domain semantic mapping matrix is ​​trained using a contrastive learning loss function, which is: ； in, A semantic embedding vector representing interactive information text. Indicates and Semantically relevant positive sample knowledge graph entity embedding vectors Indicates the first The negative sample knowledge graph entity embedding vectors were sampled in total. One negative sample, This represents the cross-domain semantic mapping matrix to be learned. Represents the cosine similarity function. This represents the temperature coefficient.

6. The method according to claim 1, characterized in that, Before obtaining the semantic embedding vector of the interactive text, the method further includes: using a sample dataset labeled with sentiment polarity and pragmatic intent labels to perform supervised fine-tuning on the language model encoder used for semantic embedding. The pragmatic intent labels include ironic expression type labels and positive / negative expression type labels, so that the semantic embedding vector output by the fine-tuned language model encoder reflects the semantics after pragmatic intent correction.

7. The method according to claim 1, characterized in that, The semantic quality benchmark score is calculated in the following manner: ； in, This represents the multimodal content representation vector of the target short video. The first element in the enterprise knowledge graph associated with the target short video is... The semantic embedding vectors of each entity are associated with each other. One entity; The incremental confidence score is calculated as follows: ； in, This indicates the number of interactive messages generated by the target short video whose authenticity and credibility score is higher than the preset threshold. Indicates the first The authenticity and credibility score of each interactive message Indicates the first The semantic embedding vector of each interactive message, Indicates the relationship with the first Semantic vectors of knowledge graph entities aligned with interactive information.

8. The method according to claim 7, characterized in that, The dynamic fusion weights are calculated in the following manner: ； in, The decay rate hyperparameter is represented by the weighted sum of the semantic quality baseline score and the incremental confidence score based on the dynamic fusion weights, resulting in the ranking score, which is calculated as follows: 。 9. The method according to claim 1, characterized in that, The authenticity and credibility score is used simultaneously to calculate the contribution weight of the reduced retrieval ranking, as a basis for filtering interactive information before semantic embedding, and as a weighting coefficient in the calculation of the incremental confidence score, so as to realize the reuse of the authenticity identification results in the three stages of retrieval reduction, semantic alignment input filtering, and long-tail cold start incremental confidence score calculation.

10. The method according to claim 1, characterized in that, The graph attention network is replaced by a Transformer-based sequence self-attention model. The historical interaction information sequence of the same sending account and the interaction information sequence of the same short video are encoded as sequence features respectively. The self-attention mechanism captures temporal anomaly patterns and combines them with text semantic features for fusion classification to obtain the authenticity and credibility score.

Images