Live content detection method and device, electronic equipment, storage medium and program product

Abstract

A live content detection method and device, electronic equipment, storage medium and program product are disclosed. The method comprises: obtaining target content data associated with live content from a live data stream, determining content feature data of the target content data, determining a live scene corresponding to the live content, determining a suspected violation type corresponding to the live content according to the content feature data, the live scene and a violation type feature corresponding to the live scene, and generating a violation type detection result based on the suspected violation type; determining a target violation semantic feature according to the violation type detection result and the target content data, providing the target violation semantic feature to a pre-trained semantic classification model to obtain a target semantic classification result, and training a machine learning model based on sample violation semantic features and corresponding expected violation categories to improve the accuracy of live content detection.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method, apparatus, electronic device, storage medium, and program product for detecting live streaming content. Background Technology

[0002] In the digital age, live streaming, with its advantages of immediacy, interactivity, and immersion, has developed into a core medium for information dissemination, commercial monetization, and social entertainment.

[0003] Current technologies primarily rely on manual review for live stream content detection. However, with high concurrency, 24 / 7 streaming volumes, manual review struggles to achieve comprehensive coverage, resulting in high false negative rates. Furthermore, inconsistencies in reviewer experience, cognitive abilities, and subjective judgment standards can lead to inconsistent review criteria, affecting the fairness and consistency of detection results. Additionally, manual review is inefficient and slow to respond, failing to meet the demands of real-time and large-scale detection. Therefore, a new live stream content detection method is needed to achieve intelligent and standardized detection of live stream content. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, storage medium, and program product for detecting live streaming content, in order to solve the problems of high false negative rate, inconsistent review standards, and delayed response in related technologies that rely on manual detection of live streaming content.

[0005] According to one aspect of the present invention, a method for detecting live streaming content is provided, the method comprising: Obtain target content data associated with the live stream from the live stream data stream, determine the content feature data of the target content data; determine the live stream scene corresponding to the live stream content, determine the suspected violation type corresponding to the live stream content based on the content feature data, the live stream scene and the violation type feature corresponding to the live stream scene, and generate a violation type detection result based on the suspected violation type; Based on the violation type detection result and the target content data, the target violation semantic features are determined, and the target violation semantic features are provided to the pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

[0006] According to another aspect of the present invention, a live streaming content detection device is provided, the device comprising: The violation type detection module is used to obtain target content data associated with the live content from the live data stream, determine the content feature data of the target content data, determine the live scene corresponding to the live content, determine the suspected violation type corresponding to the live content based on the content feature data, the live scene and the violation type feature corresponding to the live scene, and generate a violation type detection result based on the suspected violation type. The violation semantic classification module is used to determine the target violation semantic features based on the violation type detection result and the target content data, and provide the target violation semantic features to a pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the live content detection method according to any embodiment of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the live content detection method according to any embodiment of the present invention.

[0009] According to another aspect of the present invention, embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements the live content detection method as described in any of the embodiments of this disclosure.

[0010] The technical solution of this invention firstly involves obtaining target content data associated with the live stream from the live stream data stream, determining the content feature data of the target content data, determining the live stream scene corresponding to the live stream content, determining the suspected violation type corresponding to the live stream content based on the content feature data, the live stream scene, and the violation type features corresponding to the live stream scene, and generating a violation type detection result based on the suspected violation type. This approach achieves preliminary automated identification of violating content through content feature data, and infers the suspected violation type by combining the live stream scene context and the violation type features corresponding to the scene, significantly improving the accuracy and scene adaptability of violation determination, laying the foundation for subsequent refined analysis. Next, the target violation semantic features are determined based on the violation type detection results and the target content data. These features are then provided to a pre-trained semantic classification model to obtain a target semantic classification result. The semantic classification model is trained on a machine learning model based on the sample violation semantic features and their corresponding expected violation categories. This semantic classification model can achieve high-precision semantic classification results, enabling rapid interception of violation semantics, ensuring the real-time nature and controllability of violation detection, improving the depth and accuracy of violation detection, and possessing good scalability and adaptability. It can dynamically adjust judgment strategies for different live streaming scenarios, reducing false positives and false negatives, and ensuring the security of live streaming content and user experience.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a live content detection method provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of a live content detection method provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the structure of a live content detection device according to Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device that implements the live content detection method of this invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0015] It should be noted that the terms "first," "second," "target," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0018] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0019] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution, based on the prompt message.

[0020] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0021] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0022] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0023] Example 1 Figure 1 This is a flowchart of a live streaming content detection method provided in Embodiment 1 of the present invention. This embodiment is applicable to detecting whether live streaming content violates regulations. The method can be executed by a live streaming content detection device, which can be implemented in hardware and / or software, optionally through an electronic device such as a mobile terminal, PC, or server. Figure 1 As shown, the method may specifically include: S110. Obtain target content data associated with the live content from the live data stream, determine the content feature data of the target content data, determine the live scene corresponding to the live content, determine the suspected violation type corresponding to the live content based on the content feature data, the live scene, and the violation type feature corresponding to the live scene, and generate a violation type detection result based on the suspected violation type.

[0024] In this embodiment of the invention, a live data stream can be understood as a data stream continuously generated during the live stream that contains information related to the live stream content. For example, a live data stream may include, but is not limited to, various data stream types such as video streams, audio streams, and text-based bullet comment streams. Target content data can be data extracted from the live data stream that is directly related to the live stream content and can be used for subsequent violation detection and analysis, such as live video clips, audio content, bullet comment text, and comment content. Content feature data can be data used to reflect the semantic attribute characteristics of the target content data. A live stream scenario can be understood as the specific scenario type presented in the live stream, such as a game live stream scenario, an entertainment performance live stream scenario, an e-commerce live stream scenario, or an educational live stream scenario. Violation type features can be features corresponding to the live stream scenario that can reflect the characteristics of violation types in that scenario. For example, in an e-commerce live stream scenario, the characteristics of false advertising violations may include exaggerated statements about product efficacy and false promises. A suspected violation type refers to the type of violation that may exist in the live stream content after determining that there is a semantic violation, combined with content feature data, the current live stream scenario, and the violation type features corresponding to that scenario, preliminarily judged. For example, a suspected false advertising violation. Violation type detection results refer to standardized detection records generated based on suspected violation types. The violation type detection results may include, but are not limited to, various information such as suspected violation type, detection confidence level, occurrence timestamp, and the target audience of the violation content.

[0025] Optionally, appropriate feature extraction methods can be adopted according to the type of live data stream to obtain target content data associated with the live content from the live data stream and generate corresponding content feature data of the target content data. The content feature data may include, but is not limited to, at least one of sentiment feature data, tone feature data, and intent feature data, to achieve a comprehensive characterization of the semantic dimensions of the target content data.

[0026] Taking the text bullet screen stream as an example, the target content data can be the bullet screen text data generated in real time. First, Apache Kafka can be used to capture 5,000 bullet screen and comment data per second and store them in a standardized JSON format. The stored fields can include various information such as data identification, timestamp, text content, etc., to obtain the target content data associated with the live content. Then, the context semantic parsing technology is used to preprocess the target content data, and at the same time, the pre-trained model of Bidirectional Encoder Representations from Transformers (BERT) is called to complete word segmentation and semantic encoding, generating a 768-dimensional vector representation. On this basis, the emotional features, semantic features and intention features are further extracted, and finally 1500-dimensional content feature data is generated. Among them, in the process of extracting content feature data, the emotional feature analysis can be realized based on the word embedding cosine similarity, the tone features can be determined by various end-of-sentence punctuation marks and tone keywords such as "ma", "!", and the intention features can be analyzed by combining the sentence structure and verb phrases.

[0027] On the basis of the above solution, optionally, before determining the violation semantic detection result of the target content data according to multiple violation semantic conditions in the content feature data and the logical rule set, it further includes: obtaining the content constraint text associated with the live content, and constructing multiple structured semantic representations associated with the live content according to the content constraint text. The structured semantic representation includes multiple entities associated with the live content, the attribute information of the entities, and the association relationship between the entities; determining the logical relationship between multiple entities according to the structured semantic representation, taking the entities as nodes and the logical relationship between the entities as edges, constructing a content constraint semantic graph, and generating a logical rule set according to the content constraint semantic graph. The logical rule set includes multiple violation semantic conditions, and the violation semantic conditions are used to detect whether there is any violation content in the live content. Constructing structured semantic representations and semantic graphs based on content constraint texts and automatically generating logical rule sets make the violation detection rules more in line with the actual business context and enhance the semantic coverage and dynamic adaptation ability of the rules.

[0028] Content constraint text can be understood as pre-defined, restrictive text related to the live stream content, used to regulate it. Structured semantic representation can be understood as transforming unstructured content constraint text into a semantic form with a clear semantic structure, clearly reflecting the entities and relationships related to the live stream content. Structured semantic representation can include, but is not limited to, multiple entities associated with the live stream content, entity attribute information, and relationships between entities. An entity can be understood as a specific object related to the live stream content, such as the live stream subject, live stream object, and interactive features of the live stream content. Entity attribute information can be understood as information describing the inherent characteristics of an entity, such as the name, brand, price, and specifications of a product. Relationships between entities refer to the logical connections between different entities. Logical relationships can be understood as logically consistent relationships between multiple entities, derived from structured semantic representation, such as causal relationships, subordinate relationships, and conditional relationships. A content constraint semantic graph can be a graphical structure built with entities as nodes and logical relationships between entities as edges, used to visually display the entities and their logical relationships in the content constraint text, clearly presenting the semantic framework of the content constraints.

[0029] Specifically, text crawling technology can be used to obtain content constraint text associated with live streaming content, resulting in a content constraint text set; then, a pre-trained natural language processing model is used to perform word segmentation and syntactic analysis on the content constraint text set to identify the grammatical structure of the content constraint text sentences and preliminarily determine a candidate entity list; in response to the detection of an event that the candidate entity list contains noun phrases or verb phrases, key entities such as subject, object, and interaction features are extracted using a named entity recognition algorithm to obtain an entity set.

[0030] Based on the entity set, dependency parsing techniques can be used to parse the relationships between entities, generating relation triples containing subject, object, and interaction features, thus obtaining a semantic relationship set. Furthermore, the entity and attribute information in the semantic relationship set are mapped through a preset attribute template to generate structured data containing entity annotations and attribute standards, resulting in a structured semantic representation.

[0031] Optionally, to improve the completeness of the structured semantic representation, knowledge graph construction technology can be used to link the entity and attribute information in the structured semantic representation to a preset content-constrained text database, generating a queryable content-constrained text semantic network. If the entity in the semantic network has missing attribute information, supplementary information can be extracted from the content-constrained text set through a text completion algorithm, and the semantic network can be updated to obtain a complete structured semantic representation.

[0032] For example, after obtaining the content constraint text associated with the live stream content, entity extraction techniques can be used to identify the core entities in the content constraint text. The content constraint text can be stored in JSON format. Named Entity Recognition (NER) is performed using various entity recognition models pre-trained with BERT, such as the BERT-BiLSTM-CRF model, which combines Transformer-based bidirectional encoder representation, Bidirectional Long Short-Term Memory (BiLSTM), and Conditional Random Field (CRF). The content constraint text is input into the entity recognition model, which splits each content constraint text into sentences, identifies entities such as subjects, objects, and interaction features, and outputs them. Then, a structured semantic representation is generated based on the above recognition results. This representation is stored in the Neo4j graph database using Resource Description Framework (RDF) triples such as subject-interaction feature-object, and finally, JavaScript Object Notation for Linked Data can be exported. A structured semantic representation in JSON-LD format, which may include entity attribute information and association weights.

[0033] In the process of identifying entities and the relationships between entities, the importance of an entity in the content-constrained text can be determined by statistically analyzing the frequency of its occurrence, and the PageRank algorithm can be used to analyze the strength of the relationships between interactive feature entities.

[0034] After obtaining the structured semantic representation, the relationships between entities can be extracted from it. A pre-trained BERT model is then used for relation mapping to identify various logical relationships among multiple entities, such as causal, conditional, and constraint relationships, resulting in a set of logical relationships. For example, relation mapping techniques can be used to analyze the structured semantic representation to determine the logical relationships between entities. A graph construction algorithm is used to embed triples into a 100-dimensional vector space, calculating the Euclidean distance between the subject and interaction features within an entity. For instance, if the Euclidean distance between subject A and interaction feature B is 0.3, it can be inferred that there is a causal relationship: "Subject A executes interaction feature B, leading to event C." For conditional relationships, sentence patterns such as "if…then…" in the triples can be identified. For constraint relationships, keywords such as "prohibited" and "must" can be extracted. Dependency parsing is then used to determine the scope of keyword modification, generating a logical closed loop.

[0035] For sets of logical relationships, various graph construction algorithms, such as Translating Embeddings for Modeling Multi-relational Data (TransE), can be used to generate a graph structure containing semantic nodes and logical loops, with entities as nodes and logical relationships between entities as edges. Then, various community detection algorithms, such as the Louvain algorithm, are used to optimize the graph structure, constructing a content-constrained semantic graph. Subsequently, logical rules are extracted from the content-constrained semantic graph using rule derivation techniques to generate an initial set of logical rules. If there are incomplete logical rules in the initial set, supplementary information can be extracted from the content-constrained text set using text completion algorithms to update the initial set of logical rules, resulting in a complete set. Further, regularized matching can be used to transform the complete set of logical rules into machine-executable rules containing conditions for judging violations of semantic rules, resulting in an executable rule set. The Drools engine is then invoked, and knowledge graph query technology is used to extract the violation semantic conditions required for real-time detection and judgment from the executable rule set, generating a detection condition set. Finally, the SPARQL protocol and RDF Query Language (SPARQL) were used to query the content constraint text database based on the detection condition set. This verified the completeness of the detection condition set, ensuring that no key violation live streaming scenarios were missed, and yielded the final logical rule set used for detection. The violation semantic conditions include violation judgment conditions and post-violation processing operations.

[0036] Based on the above scheme, optionally, after generating a violation type detection result based on the suspected violation type, the method further includes: determining a violation analysis object based on the violation type detection result and the target content data; obtaining object attribute data of the violation analysis object; determining context information associated with the violation type detection result based on the object attribute data; generating semantic supplementary data based on the context information; constructing a structured extended dataset based on the target content data, the object attribute data, and the semantic supplementary data; and updating at least some of the violation semantic conditions in the logical rule set based on the structured extended dataset. By introducing object attributes and context to generate semantic supplementary data and updating the logical rule set accordingly, the detection rules can be continuously optimized as the scenario evolves, improving detection timeliness and accuracy.

[0037] The violation analysis object can be an object that requires further analysis of violations, determined based on the violation type detection results and target content data, such as the product corresponding to the violation content or the user who posted the violation content. Object attribute data can be data describing the inherent characteristics of the violation analysis object, such as the manufacturer information of the product or its quality inspection report. Semantic supplementary data can be data generated based on contextual information associated with the violation type detection results to supplement and improve the semantic information of the target content data, such as detailed descriptions of the live broadcast scene when the violation occurred or user interaction history. The structured extended dataset can be an integrated extended dataset constructed from the target content data, object attribute data, and semantic supplementary data. The structured extended dataset can be used to update the violation semantic conditions in the logical rule set.

[0038] Specifically, semantic parsing technology can be used to segment the target content data, extract semantic feature vectors from the target content data, and obtain a semantic feature set. Based on the violation type detection results, core features related to violations can be selected from the semantic feature set to determine the violation analysis object. The object attribute data of the violation analysis object can be obtained through a preset data interface. The context information associated with the violation type detection results can be analyzed in combination with the object attribute data. Semantic supplementary data can be generated based on the context information to improve the semantic description of the violation scenario. The target content data, object attribute data, and semantic supplementary data can be structurally integrated to construct a structured extended dataset. Based on this dataset, the adaptability of the violation semantic conditions in the logical rule set can be analyzed. Violation semantic rules with deviations or omissions can be updated to improve the detection accuracy of the logical rule set.

[0039] S120. Determine the target violation semantic features based on the violation type detection result and the target content data, and provide the target violation semantic features to the pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

[0040] The target violation semantic features can be understood as semantic features selected from the target content data that accurately reflect the core information of the violation. The semantic classification model can be built based on machine learning techniques. It is trained using sample violation semantic features and their corresponding expected violation categories, and can be used to classify the input target violation semantic features to determine their respective violation categories. Sample violation semantic features can be obtained by collecting a large number of live content samples with known violation categories and extracting features. The expected violation category can be the violation category to which the features of the sample violation semantic features belong, pre-labeled during the training of the semantic classification model. The target semantic classification result can be the judgment result of the semantic classification model on the violation category to which the target violation semantic features belong.

[0041] Optionally, speech-to-text technology and feature extraction technology can be used on the target content data to obtain various target violation semantic features such as text features and audio features. Then, the target violation semantic features are input into a pre-trained semantic classification model for semantic analysis to generate target semantic classification results.

[0042] Specifically, initial violation semantic features can be extracted from the target content data based on the violation type detection results, and the initial violation semantic features can be dimensionality reduced to obtain the target violation semantic features.

[0043] Optionally, based on the above scheme, after obtaining the target semantic classification result, the method further includes: determining violation evidence records based on the target semantic classification result and the target content data; generating a violation semantic detection report of the live content based on the violation evidence records; extracting key semantic fragments from the target content data based on the violation semantic detection report; determining a structured violation description based on the key semantic fragments and the violation evidence records; and generating a violation content detection report of the live content based on the structured violation description and multiple violation semantic conditions in the logical rule set.

[0044] Among them, violation evidence records refer to a set of relevant records that can prove the existence of violations in the live broadcast content, determined based on the target semantic classification results and target content data. For example, violation evidence records may include, but are not limited to, screenshots of violating text content, violating audio clips, violating video clips, and their corresponding timestamps. A violation semantic detection report can be understood as a report generated after detecting violation semantic information in the live broadcast content, and may include information such as the violation semantic content and violation type. Key semantic fragments can be fragments extracted from the target content data that contain core violation semantic information, and can intuitively reflect the violation information of the live broadcast content. Structured violation descriptions can be descriptions of the violation situation in the live broadcast content according to a preset structured format based on key semantic fragments and violation evidence records, making the violation information more organized, easier to understand, and easier to review. A violation content detection report can be a report generated based on the target semantic classification results and target content data, used to reflect the violation situation of the live broadcast content.

[0045] After obtaining the target semantic classification result, the target violation semantic features can be accurately extracted from the target content data by combining the violation type detection result. These features are then input into the semantic classification model to obtain the target semantic classification result. Compliance analysis is performed according to preset hierarchical review rules. If the confidence level of the classification result meets the compliance threshold, the compliance category is determined and a compliance classification result is generated. Next, a feature subset related to the violation semantic category is extracted from the compliance classification result. A text similarity algorithm is used to calculate the matching degree between the feature subset and the preset violation feature template corresponding to the semantic category, obtaining the violation matching result. Then, knowledge graph query technology is used to obtain contextual information related to the target semantic classification result from a pre-built compliance knowledge base, generating a supplementary description of the violation content. Regular expression technology is used to verify the matching degree between the supplementary description of the violation content and the expected violation category pattern. If the matching degree meets the corresponding threshold, violation evidence records are determined based on the supplementary description of the violation content.

[0046] Based on the above solution, optionally, generating a violation semantic detection report for the live stream content based on the violation evidence records includes: extracting key violation features from the violation evidence records; verifying the key violation features based on multiple violation semantic conditions in the logical rule set; and generating a violation semantic detection report based on the verification results of the key violation features, context information, and the violation evidence records. By extracting key violation features from the violation evidence records, combining them with logical rule verification, and then integrating context information to generate the detection report, the accuracy, interpretability, and evidentiary support of the report are improved.

[0047] Key violation features can be extracted from violation evidence records and highlight the core characteristics of the violation. Verification results can be derived by examining the key violation features based on multiple violation semantic conditions in a logical rule set. These results confirm whether the key violation features meet the violation judgment criteria and can include various types of results, such as meeting or not meeting the violation criteria. Contextual information can be understood as information associated with the key violation features that helps in understanding the environment and background in which the violation features occur.

[0048] Specifically, key violation features can be extracted from violation evidence records. Rule verification technology is used to verify the key violation features based on multiple violation semantic conditions in the logical rule set. A violation semantic detection report is then generated based on the verification results of the key violation features, context information, and violation evidence records.

[0049] Taking a game live streaming scenario as an example, the target content data can be the audio stream of a game live stream and the corresponding bullet screen text. The violation type detection result is "suspected derogatory violation". For the target content data, the audio stream can be converted into a text sequence through speech-to-text technology, while retaining intonation features, generating feature vectors corresponding to the text sequence and audio features. The audio features can include pitch changes and speech rate, and the text features can be obtained by extracting word vectors to form a content feature set. The content feature set has 500 dimensions, including 200-dimensional text features and 300-dimensional audio features. Next, a pre-trained semantic classification model is used to analyze the content feature set. The model can be based on the BERT architecture. After inputting the content feature set, it can output its corresponding target semantic classification result, such as "negative evaluation" or "neutral expression". Through the hierarchical review rules, a compliance threshold can be set. If it is lower than the compliance threshold, manual review is required. If the classification result meets the compliance threshold, the compliance category can be directly determined as "definite violation" as its corresponding compliance classification result. Next, keyword features and sudden volume increase features are extracted from the compliance classification results to form a feature subset. The cosine similarity algorithm is used to calculate the matching degree with the violation feature template. If the matching degree meets the corresponding threshold, it is determined to be a violation match. Contextual information is obtained from the compliance knowledge base using knowledge graph query technology. It is found that the user has made similar derogatory remarks in 3 game live streams in the past 7 days, so a supplementary description of the violation content is generated: "User XXX has continuously made derogatory remarks about the streamer in game live streams." Regular expressions are used to verify the matching of this supplementary description of the violation content with the expected violation category pattern of "derogatory + streamer." If the matching degree exceeds the corresponding matching threshold, a violation evidence record is determined, including the violation text ID, the corresponding audio segment, the publisher ID, and the knowledge graph query result. After generating a violation semantic detection report based on the violation evidence record, text segments and corresponding audio segments can be extracted from the target content data as key semantic segments.

[0050] Optionally, rule verification technology can be used to perform secondary verification on key violation features in violation evidence records based on multiple violation semantic conditions in the logical rule set. For example, check the consistency of keyword frequency and audio features. After confirming that the key violation features meet the violation judgment criteria, a violation semantic detection report can be generated by combining the verification results, related context information and violation evidence records.

[0051] Optionally, based on the violation evidence records, the violation types obtained by integrating the violation semantic detection report of the live broadcast content can be linked to the violation analysis objects of the live broadcast content using business association technology to generate an extended review basis dataset.

[0052] Specifically, semantic parsing techniques can be used to segment the live stream content, extract semantic feature vectors from the text, obtain a semantic feature set, and then use clustering algorithms to group the semantic feature vectors in the semantic feature set, determine the semantic category of each group, and obtain the semantic classification result. For example, if the live stream content is a 100-word viewer comment, a BERT-based pre-trained model can be used to segment and semantically encode the text, splitting it into a sequence of words and generating a semantic vector for each word, ultimately generating a 150-dimensional semantic feature vector set to capture the contextual relationships between words and ensure the accuracy of semantic representation. When grouping the semantic feature set using clustering algorithms, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm can be used to perform density clustering of the semantic feature vectors, dividing them into multiple semantic categories based on the cosine distance between vectors, such as "positive evaluation," "neutral feedback," and "negative complaints," and marking outliers as noise, dynamically identifying the semantic patterns of diverse live stream content.

[0053] Optionally, historical user interaction records, such as comment frequency and likes, can be extracted from the database to construct corresponding user profile data. Association rule mining techniques, such as the Apriori algorithm, are used to match the correlation between semantic classification results and user profile data. If the correlation meets a preset correlation threshold, valid profile features are determined, resulting in a valid profile feature set. For example, suppose a user is classified as "negative complaints," and their profile data shows they sent 30 comments in the past 7 days, 20 of which contain negative words. Correlation detection reveals a correlation confidence of 0.8 between the semantic classification result "negative complaints" and the user profile data "high-frequency comments." Assuming a correlation threshold of 0.75, "high-frequency comments" are determined to be a valid profile feature.

[0054] After obtaining an effective profile feature set, knowledge graph query technology can be used to obtain contextual information related to semantic categories from a knowledge base containing information such as interaction features, live streaming scenarios, and related nodes of violation types. This generates semantic supplementary data for semantic features, and extracts key feature data related to violation type classification. Regular expressions are used to verify the matching degree between the key feature data and the preset violation feature template. If the matching degree meets the preset matching degree threshold, the potential violation content is determined, and the violation content judgment result is obtained. Then, a text similarity algorithm, such as Sentence Bidirectional Encoder Representations from Transformers (Sentence-BERT), is used to calculate the semantic consistency between the potential violation content and the known violation samples. If the consistency meets the preset threshold, an extended review basis dataset is generated.

[0055] For example, contextual information can be obtained for the "negative comments" category, such as users frequently posting comments late at night and often targeting information like game live streams. Semantic supplementary descriptions can be generated, including information such as user activity time and topic preferences, forming more comprehensive semantic supplementary data and enriching the contextual information. When extracting key feature data related to violation types, high-frequency negative words in the "negative comments" category can be focused on. Regular expressions can be used to detect templates. If a comment matches the "derogatory" template and the matching degree exceeds a preset matching degree threshold, it is determined to be a potential violation.

[0056] Among them, the extended review basis dataset can be a structured dataset containing various data such as potential violation text IDs, semantic categories, and similarity scores, which can provide rich samples for subsequent model training and significantly improve the automated review capability.

[0057] Optionally, by expanding the dataset used for review, word segmentation technology can be employed to preprocess the text of live stream content segments, extracting keyword sets to obtain preliminary text feature sets. For example, assuming a live stream segment contains a viewer comment "The streamer sings well, the atmosphere is great," a BERT-based word segmentation tool can be used to split it into keywords such as "streamer," "singing," "good," "atmosphere," and "great," generating a preliminary text feature set. This preliminary text feature set contains the semantic vector of each word, preserving the semantic integrity of the text and providing a reliable foundation for subsequent analysis. Further, the K-means clustering algorithm can be used to group the keywords, determining the semantic theme of each group and obtaining semantic theme classification results. For example, assuming there are 1000 keyword feature vectors, and K is set to 4, the clustering algorithm can classify the keywords into four categories: "positive evaluation," "neutral description," "negative sentiment," and "irrelevant chatter." This can include various classification methods, such as grouping "good" and "great" as positive evaluations, to clearly distinguish semantic tendencies and facilitate subsequent feature mining.

[0058] Optionally, user profile data can be obtained based on semantic topic classification results. Association rule mining techniques can then be used to match the correspondence between semantic topics and user profile features. If the matching degree meets a preset threshold, valid user profile features are determined, resulting in a valid profile feature set. For example, if a user recently commented 20 times in a music live stream, with 15 comments containing positive words, the Apriori algorithm can be used to mine association rules. If the confidence score between the "positive evaluation" topic and "high-frequency interaction" exceeds a threshold, it is confirmed as a valid profile feature, accurately depicting the user's behavior pattern. Next, using the valid profile feature set, knowledge graph query technology is employed to obtain contextual information related to the semantic topic from a pre-built compliance knowledge base, generating supplementary features for the semantic topic, resulting in a supplementary feature set. For example, if the pre-built knowledge base contains nodes of various types such as "user behavior," "live stream type," and "compliance," querying the "positive evaluation" topic reveals that users are mostly active in evening music live streams, a supplementary feature set is generated that includes user activity time, preferred topics, etc., enriching the contextual information. From the supplementary feature set, regular expression technology is used to verify the matching degree between the features and the preset violation template. If the matching degree meets the preset corresponding threshold, the potential violation segment is determined and a candidate set of violation segments is obtained.

[0059] Furthermore, using a candidate set of violation fragments, a text similarity algorithm such as Sentence-BERT is employed to calculate the semantic consistency between the candidate fragments and known violation samples. If the consistency meets a preset threshold, variant samples are generated, resulting in a variant sample set. For the variant sample set, attribute fusion technology is used to integrate the semantic features of the variant samples and the violation information in the violation semantic detection report, generating optimized compliance judgment conditions. Assuming the fused dataset contains the semantic categories and user profiles of 500 comments, a structured optimized compliance judgment condition set containing text IDs and similarity scores can be formed, improving automated review capabilities.

[0060] Optionally, feature extraction technology can be used to perform semantic analysis on the live stream content to generate a semantic feature set. This set can then be combined with user profile data and attribute fusion technology to generate a comprehensive feature set. For the comprehensive feature set, a random forest algorithm can be used for classification to identify potential violation segments and obtain violation classification results. If the violation classification results meet a preset threshold, association analysis technology is used to calculate the business relevance between the live stream content and known violation samples, generating a variant sample set. Key semantic features are extracted from the variant sample set, and a cosine similarity algorithm is used to verify the semantic consistency between the key semantic features and the compliance judgment conditions, determining the matching degree. For the matching degree results, knowledge graph query technology is applied to obtain contextual information related to the potential violation segments from a pre-built knowledge base, generating supplementary compliance evidence. Finally, based on the supplementary compliance evidence, regular expression technology is used to verify the matching degree between the potential violation segments and the preset compliance template, obtaining optimized compliance judgment conditions.

[0061] Optionally, potential violations can be obtained from the violation semantic detection report, and preliminary screening can be performed using optimized compliance judgment conditions to determine the extraction range of key semantic segments, resulting in a segment extraction set. Specifically, when obtaining potential violations from the violation semantic detection report, the report can first be structured and parsed. The report typically includes text transcription, audio features, and interaction data of the live stream content. This data is preprocessed to form a preliminary violation probability score. For example, assuming a violation semantic detection report shows a probability of 0.8 for a game live stream segment with a "violent language" tag, preliminary screening can be performed using optimized compliance judgment conditions. These conditions include a keyword blacklist and semantic threshold settings; for example, a violation word frequency exceeding 5% is marked as high-risk. Through screening, the extraction range of key semantic segments is determined, such as the dialogue section from the 10th to 15th minute of the live stream, from which text sequences containing "aggressive words" are extracted, ultimately resulting in a segment extraction set. This set is stored in JSON format, with each segment carrying a start timestamp and semantic tag, providing a foundation for subsequent processing.

[0062] Optionally, for the fragment extraction set of key semantic segments, an attention mechanism can be used to extract the violation semantic content. By weighted attention to the input text sequence, the semantic segments in the fragment extraction set are input, and the semantic representation of the prominent violation part is output, generating a structured violation description containing violation evidence records, thus obtaining a violation description set.

[0063] Optionally, based on the violation description set, a knowledge graph constructed from multiple violation semantic conditions in the rule engine matching logic rule set can be used. A tree-like structure is then used to compare the violation description set with the violation semantic condition nodes of the graph rules, generating a final review result including attribute annotations and timestamps—that is, a violation content detection report corresponding to the live stream content. Furthermore, a multimodal review mechanism can be introduced to simultaneously query the visual elements of video frames corresponding to key semantic segments, supplementing them to the violation description set. Finally, the structured violation description, the matching results of the logic rule set, and violation evidence records are integrated to generate a standardized violation content detection report.

[0064] For example, for a fragment extraction set, an attention mechanism can be used to extract the semantic content of violations. By calculating the weight of each word in the input sequence, important parts are highlighted. The normalized exponential function softmax is used to perform a weighted summation of the query, key, and value vectors. When the input is a semantic fragment from the fragment extraction set, the attention mechanism assigns high weights to the violating words, and the output is a semantic representation vector highlighting the violating parts. For example, the dimension value of the violating parts in the vector is increased to above 0.9. Then, a structured violation description containing evidence retention records is generated, which may include the original text, weight distribution, and violation type labeling, forming a violation description set, which can be used as audit evidence to ensure the traceability of violation review.

[0065] The rule engine can integrate a machine learning feedback loop mechanism to dynamically adjust rule thresholds or supplement new violation semantic conditions based on historical matching results, adapt to constantly updated violation patterns, and improve the dynamic adaptability and accuracy of violation detection.

[0066] Based on the above scheme, optionally, after generating the violation content detection report for the live stream content, the method further includes: receiving target feedback data for the violation content detection report; generating optimization guidance information for the semantic classification model and the logical rule set based on the target feedback data; wherein the target feedback data includes at least one of accuracy data of the violation detection results and appeal information for the target content data. Optimization guidance is generated based on the feedback and appeal information to achieve closed-loop iteration of violation detection, continuously improving the accuracy and fairness of the semantic classification model and the logical rule set.

[0067] The target feedback data refers to the feedback information received in response to the violation detection report, which may include at least one of the accuracy data of the violation detection results and appeal information regarding the target content data. The optimization guidance information may be guidance information generated based on the target feedback data, used to optimize and improve the semantic classification model and logical rule set.

[0068] Specifically, the system can receive target feedback data regarding violation detection reports. This target feedback data may include various data such as the accuracy of violation detection results and appeal information regarding the target content data. Valid fields are extracted using data cleaning techniques to obtain a structured feedback set. Then, a clustering algorithm is used to group the accuracy data of the violation detection results and the appeal information regarding the target content data within this structured feedback set. The feature vectors of the target feedback data are used as input, and the grouped data clusters are output, resulting in a feedback classification set. For example, the feedback data can be divided into categories such as "high accuracy, high appeal," "low accuracy, low appeal," and "low accuracy, high appeal." "High accuracy, high appeal" may correspond to overly strict rules, while "low accuracy, low appeal" may correspond to model misjudgment.

[0069] Furthermore, based on the feedback classification set, a mapping relationship between it and the variant text extended dataset can be established through business association rules preset according to actual business needs. The business association rules are based on grouping features, and the output assigns category labels such as "rule optimization sample" and "model misjudgment sample" to the variant text extended dataset, thus obtaining the mapped sample set.

[0070] For example, target feedback data can be obtained. Suppose a live streaming platform needs to analyze feedback information. The target feedback data includes an accuracy data of 0.9 for violation detection results, indicating that 90% of the violation detection results are consistent with the results of manual review. The appeal information for the target content data in the target feedback data may include objection text, such as "This content has no violation, requesting review". Invalid data such as null values ​​or irrelevant comments can be removed through data cleaning techniques, and valid fields such as accuracy data, keywords and timestamps in the appeal text can be extracted to form a structured feedback set. Next, for the structured feedback set, the K-means clustering algorithm is used with the feature vector of the target feedback data as input. The feature vector can be various types of information, such as the word frequency vector of accuracy data and appeal information keywords. The output is a grouped data cluster. For example, if a data cluster contains data with high accuracy scores but high appeal rates, it indicates that the violation classification rules may be too strict. If another cluster contains data with low accuracy scores and low appeal rates, it indicates that the model has misjudged. Thus, the structured feedback set can be divided into multiple categories such as high accuracy and high appeal, low accuracy and low appeal, to obtain a feedback classification set. Finally, incremental learning samples can be mapped according to the feedback classification set and business association rules. For example, if the business association rule is set to allocate rule optimization labels for high accuracy and high appeal, and a data cluster shows an accuracy of 0.9 but an appeal rate of 20%, it can be mapped as a rule optimization sample, thereby generating a mapped sample set to provide a basis for subsequent model adjustments.

[0071] Based on the mapped sample set, model parameter adjustment indices can be extracted. Using the gradient descent algorithm with the loss function of the mapped sample set as input, the model parameters are iteratively adjusted to obtain an optimized model parameter set. If the adjustment indices of the optimized model parameter set meet a preset adjustment threshold, a rule base update technique is used, with the optimized model parameter set as input, to output rule modification instructions, resulting in a rule optimization set. Next, optimization directions are obtained from the rule optimization set, and knowledge graph embedding technology is used to transform them into optimization guidance information. The knowledge graph can use the rule optimization set as nodes, outputting a structured set of guidance information, such as prioritizing certain keywords. Finally, information distribution technology transforms the guidance information set into instructions, such as increasing the detection weight of certain words, forming a distribution instruction set, which is then sent to the semantic classification model for incremental learning and model updates.

[0072] For example, model parameter adjustment indicators are extracted from the mapping sample set. For instance, a loss function value of 0.3 is used to represent the deviation between the semantic classification model's prediction result and the actual result. The gradient descent algorithm is used to optimize the parameters. By iteratively adjusting the weights, the loss value is reduced to 0.1, generating a model parameter optimization set. The parameters in the model parameter optimization set can reflect the improved sensitivity of the semantic classification model to illegal content. If the adjustment indicator of the model parameter optimization set is a loss value below 0.15, which meets the preset adjustment indicator threshold, then the optimization direction is generated through rule base update technology. For example, by analyzing the model parameter optimization set through update technology, a modification instruction of "reducing the threshold of rule XX from 0.8 to 0.75" is output to form a rule optimization set to identify new illegal detection patterns.

[0073] Optionally, based on the optimization guidance information, misjudgment statistics and business scenario adaptability assessment records can be obtained. Data preprocessing techniques can be used to clean up invalid data such as null values ​​or irrelevant fields, and key feature fields can be extracted to obtain a structured feature set. The misjudgment statistics can be records of incorrect classification of live stream content, such as a live stream content being misjudged as violating regulations. The business scenario adaptability assessment records reflect the performance of violation review in specific business scenarios, such as the ability to identify emerging internet slang. Data preprocessing techniques can extract fields such as misjudgment rate, scenario tags, and semantic features.

[0074] Based on the structured feature set, an incremental learning algorithm can be used to gradually optimize and update the semantic classification model using small batches of data. If the misclassification statistics of the structured feature set exceed a preset threshold, the model parameters are adjusted accordingly to obtain an updated model parameter set. Based on the updated model parameter set, semantic analysis techniques are used to deepen semantic understanding and extract the deeper meaning of the text. If the semantic understanding results do not match the business scenario adaptability assessment, the semantic parsing rules are adjusted to form an optimized semantic rule set. Then, dynamic adjustment instructions are obtained from the optimized semantic rule set, and parameter fusion techniques are used to combine these instructions with the model parameter set to obtain a fused optimized parameter set. For the fused optimized parameter set, an adaptive optimization algorithm is used to configure the model to meet the business scenario requirements, generating the final configuration set.

[0075] Optionally, a content constraint semantic graph and a logical rule set can be obtained. An online learning algorithm is used to iteratively update the semantic graph nodes and rule weights to obtain an optimized content constraint semantic graph and logical rule set. For the optimized semantic graph and rule set, semantic analysis techniques are used to extract semantic features from the query interface. If the matching degree between the semantic features and the logical rule set corresponding to the business scenario is lower than a preset threshold, the rule weights are adjusted to obtain an updated logical rule set. For example, when receiving live stream bullet screen text through a query interface, features such as "the proportion of slang in bullet screens is 0.4" are extracted. If the business scenario logic requires "the proportion of slang in youth live streams is less than 0.3", and the matching degree is lower than the threshold of 0.8, the rule weights are adjusted, such as lowering the slang trigger threshold from 0.5 to 0.3, resulting in an updated logical rule set that better suits the scenario requirements.

[0076] Based on the updated logical rule set, new semantically related nodes are generated using knowledge graph expansion technology to obtain an expanded knowledge graph. For example, analyzing emerging internet slang in live streaming content generates a "internet meme" node, which is then linked to a "positive semantic" node to form the expanded knowledge graph. For the expanded knowledge graph, feedback analysis technology is used to extract deviation data in the review quality. If the deviation data exceeds a preset range, a feedback-driven optimization instruction set is generated. The query interface configuration is updated based on the optimization instruction set, and dynamic configuration instructions are generated through configuration distribution technology to achieve adaptive optimization of violation detection. Finally, semantic verification technology is used to check the matching between the configuration and the business scenario logic, ensuring that the final configuration set meets the actual detection requirements.

[0077] The technical solution of this invention firstly involves obtaining target content data associated with the live stream from the live stream data stream, determining the content feature data of the target content data, determining the live stream scene corresponding to the live stream content, determining the suspected violation type corresponding to the live stream content based on the content feature data, the live stream scene, and the violation type features corresponding to the live stream scene, and generating a violation type detection result based on the suspected violation type. This approach achieves preliminary automated identification of violating content through content feature data, and infers the suspected violation type by combining the live stream scene context and the violation type features corresponding to the scene, significantly improving the accuracy and scene adaptability of violation determination, laying the foundation for subsequent refined analysis. Next, the target violation semantic features are determined based on the violation type detection results and the target content data. These features are then provided to a pre-trained semantic classification model to obtain a target semantic classification result. The semantic classification model is trained on a machine learning model based on the sample violation semantic features and their corresponding expected violation categories. This semantic classification model can achieve high-precision semantic classification results, enabling rapid interception of violation semantics, ensuring the real-time nature and controllability of violation detection, improving the depth and accuracy of violation detection, and possessing good scalability and adaptability. It can dynamically adjust judgment strategies for different live streaming scenarios, reducing false positives and false negatives, and ensuring the security of live streaming content and user experience.

[0078] Example 2 Figure 2 This is a flowchart of a live streaming content detection method provided in Embodiment 2 of the present invention, further describing the implementation process of generating a violation content detection report of the live streaming content based on the target semantic classification result and the target content data. Specific implementation details can be found in the description of this embodiment. Technical features that are the same as or similar to those in the foregoing embodiments will not be repeated here. Figure 2 As shown, the method may specifically include: S210. Obtain target content data associated with the live content from the live data stream, determine the content feature data of the target content data, determine the live scene corresponding to the live content, determine the suspected violation type corresponding to the live content based on the content feature data, the live scene, and the violation type feature corresponding to the live scene, and generate a violation type detection result based on the suspected violation type.

[0079] S220. Determine the target violation semantic features based on the violation type detection result and the target content data, and provide the target violation semantic features to the pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

[0080] S230. Based on the target violation semantic features of the multiple target content data, determine multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records. The supplementary feature data includes interactive feature data associated with the variant text feature data and contextual information of the interactive feature data.

[0081] The target variant text sample refers to a sample selected from multiple initial variant text samples based on supplementary feature data, which can be effectively used for incremental training of the semantic classification model. It typically exhibits typicality or diversity in its violation semantic features. The initial variant text sample can be a text sample generated based on the target violation semantic features of multiple target content data, differing somewhat from the original target violation semantic features but still retaining the core information of the violation semantics. Supplementary feature data refers to data further supplemented on the variant text feature data to more comprehensively describe the features of the initial variant text sample. Supplementary feature data may include interaction feature data associated with the variant text feature data and the contextual information of the interaction feature data. Interaction feature data can be understood as feature data reflecting the interaction relationship between the initial variant text sample and other relevant data. Violation evidence records can be a set of relevant information determined based on the target semantic classification results and target content data, capable of directly or indirectly proving the existence of violation interaction features in the live broadcast content. Variant text feature data can be data used to reflect the feature attributes of the initial variant text sample after feature extraction.

[0082] Based on the above scheme, optionally, the step of determining multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records based on the target violation semantic features of the multiple target content data includes: generating multiple initial variant text samples based on the target violation semantic features of the multiple target content data, and determining variant text feature data for each initial variant text sample; determining supplementary feature data for the multiple initial variant text samples based on the variant text feature data of the multiple initial variant text samples; determining multiple target variant text samples from the multiple initial variant text samples based on the supplementary feature data of the multiple initial variant text samples, and determining violation evidence records for each target variant text sample.

[0083] Specifically, principal component analysis can be used to extract and reduce the semantic features of the target violation semantic features of multiple target content data, thereby reducing computational complexity while retaining more than 90% of the semantic information. Semantic feature vectors corresponding to multiple initial variant text samples are generated and determined as variant text feature data for each initial variant text sample.

[0084] Based on the above scheme, optionally, determining supplementary feature data for multiple initial variant text samples based on the variant text feature data of multiple initial variant text samples includes: dividing the multiple initial variant text samples into multiple variant sample groups corresponding to semantic categories based on the variant text feature data of the multiple initial variant text samples; for each variant sample group, determining supplementary feature data for multiple initial variant text samples in the variant sample group based on the semantic category corresponding to the variant sample group. By determining supplementary feature data according to semantic category grouping, feature extraction becomes more targeted, improving the efficiency and relevance of subsequent sample selection and model training.

[0085] The semantic category can be a category divided according to the semantic content of the initial variant text samples. Initial variant text samples of the same semantic category have high similarity in semantic expression. The variant sample group can be a sample set formed by grouping multiple initial variant text samples according to their semantic categories. Each variant sample group corresponds to a specific semantic category, and the initial variant text samples within the group are semantically related.

[0086] Specifically, clustering algorithms can be used to group the variant text feature data of multiple initial variant text samples. The semantic similarity of the samples determines the semantic category to which the multiple initial variant text samples belong, resulting in variant sample groups corresponding to multiple semantic categories. For example, K-means clustering can be used to calculate the groups. If the semantic categories are x, the number of cluster centers is set to x during the clustering process, and clustering is completed in n iterations, ensuring that each category contains approximately 200 initial variant text samples. Similarity can be determined based on the Euclidean distance between vectors for grouping.

[0087] Optionally, for each variant sample group, the interaction features of the corresponding semantic category can be extracted. The interaction features are matched with an evidence database containing multiple violation evidence records to determine their correlation. If the correlation meets a preset correlation threshold, the interaction feature is determined to be valid, and a set of valid interaction features is obtained. For the set of valid interaction features, knowledge graph query technology is used to obtain the context information related to the interaction feature from a pre-built knowledge base containing information such as the interaction feature, the live streaming scenario, and the violation type of the associated nodes, and to generate a supplementary description of the interaction feature, thereby determining the supplementary feature data of multiple initial variant text samples. For example, in the initial variant text sample corresponding to the semantic category of "malicious attack", there may be an interaction feature of "frequent posting of comments in a short period of time". If a sample sends 50 comments within 10 minutes, with an average of 20 characters per comment, and the proportion of "offensive" keywords in the comments reaches 30%, the correlation between the interaction feature and the violation evidence records of the "malicious attack" sample in the evidence database is matched. If the threshold is set that the keyword proportion exceeds 25%, the interaction feature is determined to be valid, and a valid interaction feature set is formed. Further, the context information corresponding to the interaction feature is obtained from the knowledge base to generate a supplementary description, and finally, supplementary feature data containing information such as comment frequency, keyword proportion, and live streaming scenario is formed.

[0088] Optionally, based on the above scheme, determining multiple target variant text samples from the multiple initial variant text samples according to supplementary feature data of the multiple initial variant text samples includes: extracting key feature data related to the corresponding semantic category from the supplementary feature data, and determining multiple target variant text samples from the multiple initial variant text samples according to the key feature data and a preset violation feature template corresponding to the semantic category. By combining the key features of the semantic category and the violation feature template to screen target variant samples, the selected samples are ensured to be typical and representative, thus optimizing the quality of training data.

[0089] Among them, key feature data can be feature data extracted from supplementary feature data that can highlight the core violation characteristics of semantic categories. Optionally, violation feature templates can be pre-set for different semantic categories to determine whether the initial variant text sample has violation features under that semantic category. The violation feature templates can contain information such as the typical manifestations and judgment criteria of the violation features under that semantic category.

[0090] Optionally, key feature data related to the corresponding semantic category can be extracted from the supplementary feature data. The matching degree between the key feature data and the preset violation feature template corresponding to the semantic category can be verified using regular expressions to determine whether there are any violation interaction features, thus obtaining a violation interaction feature judgment result. Based on the violation interaction feature judgment result, multiple target variant text samples are selected from multiple initial variant text samples. The violation interaction feature judgment result may include key information such as the violation interaction feature type, the violation interaction feature judgment confidence level, and the text content.

[0091] For example, the violation feature template of the semantic category of "malicious attack" can contain keywords such as "insulting", "threat", and "belittling", with a matching degree threshold of 80%. If the text content of an initial variant text sample can match the "belittling" pattern by regular expression detection and the matching degree reaches 85%, exceeding the corresponding matching degree threshold, then it is determined that there is a violation interaction feature and it is selected as the target variant text sample.

[0092] Optionally, a text similarity algorithm can be used to calculate the semantic consistency between the target variant text sample and the illegal evidence records in the evidence database, and to screen and determine the illegal evidence records for each target variant text sample.

[0093] S240. Construct a variant text extended dataset based on multiple target variant text samples, the supplementary feature data of the target variant text samples, and the violation evidence records of the target variant text samples. Perform incremental training on the semantic classification model based on the variant text extended dataset to update the semantic classification model.

[0094] Among them, the variant text extended dataset refers to a dataset jointly constructed by multiple target variant text samples, supplementary feature data of target variant text samples, and violation evidence records of target variant text samples, which is used for incremental training of semantic classification models.

[0095] Optionally, a variant text extended dataset can be constructed based on multiple target variant text samples, the supplementary feature data of the target variant text samples, and violation evidence records of the target variant text samples. The semantic classification model can then be incrementally trained based on this variant text extended dataset to update the semantic classification model. By generating variant text samples and constructing an extended dataset for incremental training of the model, the generalization ability of the semantic classification model to identify novel or variant violations is effectively improved.

[0096] Optionally, an evidence chain of illegal interaction features can be constructed based on multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records of the target variant text samples to obtain a variant text extended dataset; further, relevant training data can be extracted from the variant text extended dataset, the training data may contain semantic category labels and interaction features, and a deep learning algorithm can be used to incrementally train the semantic classification model and update the model parameters.

[0097] For example, if the cosine similarity between a target variant text sample and the illegal evidence record sample corresponding to "malicious attack" in the evidence database is 0.9, which exceeds the corresponding preset threshold of 0.85, then the two are confirmed to be semantically consistent, and the illegal evidence record is assigned to the target variant text sample to form a complete evidence chain containing information such as the target variant text sample, similarity score, and associated illegal evidence record.

[0098] Optionally, when incrementally training the semantic classification model, 10,000 relevant samples can be extracted from the variant text extended dataset. Each sample can contain corresponding semantic category labels and interaction feature data. Deep learning algorithms, such as Long Short-Term Memory (LSTM) networks, are then used to iteratively update the model parameters. The updated model improves the classification accuracy of the target content data of the live broadcast content. Finally, the classification results are verified through a preset review threshold to ensure the reliability of the semantic classification results.

[0099] To further ensure classification accuracy, optionally, rule-based validation techniques can be used to perform secondary validation of the target semantic classification results output by the updated model. If the classification results meet the preset classification threshold, the final target semantic classification result is generated. Through multi-stage validation, the accuracy and automation of illegal interaction feature detection are significantly improved.

[0100] The technical solution of this invention firstly determines multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records based on the target violation semantic features of multiple target content data. The supplementary feature data includes interaction feature data associated with the variant text feature data and contextual information of the interaction feature data. This accurately determines the core violation features, ensuring targeted screening of target variant text samples and improving the efficiency of subsequent data processing. The supplementary feature data covers interaction features and contextual information, enriching the sample dimensions and providing more comprehensive feature support for subsequent violation judgment, reducing the risk of missed judgment. Next, a variant text extended dataset is constructed based on the multiple target variant text samples, the supplementary feature data of the target variant text samples, and the violation evidence records of the target variant text samples. The semantic classification model is incrementally trained based on the variant text extended dataset to update the semantic classification model. The dataset is constructed by integrating multi-dimensional data, providing a high-quality data foundation for model training. The incremental training mode is used to update the model, saving training resources and time costs, and can quickly adapt to new violation scenarios.

[0101] Example 3 Figure 3 This is a schematic diagram of a live content detection device according to Embodiment 3 of the present invention. This device is used to execute the live content detection method provided in any of the above embodiments. This device and the live content detection methods of the above embodiments belong to the same inventive concept. Details not described in detail in the embodiments of the live content detection device can be referred to the embodiments of the live content detection methods described above. Figure 3 As shown, the device includes: a violation type detection module 310 and a violation semantic classification module 320.

[0102] The violation type detection module 310 is used to obtain target content data associated with the live content from the live data stream, determine the content feature data of the target content data, determine the live scene corresponding to the live content, determine the suspected violation type corresponding to the live content based on the content feature data, the live scene, and the violation type features corresponding to the live scene, and generate a violation type detection result based on the suspected violation type; the violation semantic classification module 320 is used to determine the target violation semantic features based on the violation type detection result and the target content data, provide the target violation semantic features to a pre-trained semantic classification model to obtain a target semantic classification result, wherein the semantic classification model is trained on a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

[0103] The technical solution of this invention firstly obtains target content data associated with the live stream from the live stream data stream through the violation type detection module 310, determines the content feature data of the target content data, determines the live stream scene corresponding to the live stream content, determines the suspected violation type corresponding to the live stream content based on the content feature data, the live stream scene, and the violation type features corresponding to the live stream scene, and generates a violation type detection result based on the suspected violation type. The content feature data enables preliminary automated identification of violation content, and the combination of the live stream scene context and the violation type features corresponding to the scene infers the suspected violation type, significantly improving the accuracy and scene adaptability of violation determination, laying the foundation for subsequent refined analysis. Next, the violation semantic classification module 320 determines the target violation semantic features based on the violation type detection results and the target content data, and provides the target violation semantic features to the pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories. Based on the semantic classification model, high-precision semantic classification results can be obtained, enabling rapid interception of violation semantics, ensuring the real-time performance and controllability of violation detection, improving the depth and accuracy of violation detection, and possessing good scalability and adaptability. It can dynamically adjust the judgment strategy for different live streaming scenarios, reduce the false positive and false negative rates, and ensure the security of live streaming content and user experience.

[0104] Based on the above solution, optionally, the violation type detection module 310 includes a violation semantic detection result determination submodule and a live streaming scene determination submodule. The violation semantic detection result determination submodule is used to determine the violation semantic detection result of the target content data based on the content feature data and multiple violation semantic conditions in the logical rule set; the live streaming scene determination submodule is used to determine the live streaming scene corresponding to the live streaming content in response to the violation semantic detection result indicating a semantic violation.

[0105] Optionally, based on the above scheme, the violation type detection module 310 further includes a structured semantic representation construction submodule and a logical rule set generation submodule. The structured semantic representation construction submodule is used to obtain content constraint text associated with the live stream content before determining the violation semantic detection result of the target content data based on the content feature data and multiple violation semantic conditions in the logical rule set. It then constructs multiple structured semantic representations associated with the live stream content based on the content constraint text. These structured semantic representations include multiple entities associated with the live stream content, attribute information of the entities, and the relationships between the entities. The logical rule set generation submodule is used to determine the logical relationships between the multiple entities based on the structured semantic representations. Using the entities as nodes and the logical relationships between the entities as edges, it constructs a content constraint semantic graph. It then generates a logical rule set based on the content constraint semantic graph. The logical rule set includes multiple violation semantic conditions used to detect whether violation content exists in the live stream content.

[0106] Optionally, based on the above scheme, the violation type detection module 310 further includes a key semantic fragment determination submodule and a violation content detection report generation submodule. The key semantic fragment determination submodule is used to, after obtaining the target semantic classification result, determine violation evidence records based on the target semantic classification result and the target content data, generate a violation semantic detection report for the live stream content based on the violation evidence records, and extract key semantic fragments from the target content data based on the violation semantic detection report. The violation content detection report generation submodule is used to determine a structured violation description based on the key semantic fragments and the violation evidence records, and generate a violation content detection report for the live stream content based on the structured violation description and multiple violation semantic conditions in the logical rule set.

[0107] Optionally, based on the above scheme, the violation type detection module 310 further includes a feedback optimization submodule. The feedback optimization submodule is used to receive target feedback data for the violation content detection report after the violation content detection report is generated, and to generate optimization guidance information based on the target feedback data and the semantic classification model and the logical rule set. The target feedback data includes at least one of the accuracy data of the violation detection result and appeal information regarding the target content data.

[0108] Optionally, based on the above scheme, the device further includes a target variant text sample determination module and a semantic classification model update module. The target variant text sample determination module is used to determine multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records based on the target violation semantic features of multiple target content data after determining the target semantic classification result of the live content according to the target violation semantic features and the semantic classification model. The supplementary feature data includes interaction feature data associated with the variant text feature data and context information of the interaction feature data. The semantic classification model update module is used to construct a variant text extended dataset based on the multiple target variant text samples, the supplementary feature data of the target variant text samples, and the violation evidence records of the target variant text samples, and incrementally train the semantic classification model based on the variant text extended dataset to update the semantic classification model.

[0109] Based on the above scheme, optionally, the target variant text sample determination module includes a variant text feature data determination submodule, a supplementary feature data determination submodule, and a violation evidence record determination submodule. Specifically, the variant text feature data determination submodule is used to generate multiple initial variant text samples based on the target violation semantic features of multiple target content data, and determine the variant text feature data of each initial variant text sample; the supplementary feature data determination submodule is used to determine supplementary feature data of multiple initial variant text samples based on the variant text feature data of multiple initial variant text samples; and the violation evidence record determination submodule is used to determine multiple target variant text samples from the multiple initial variant text samples based on the supplementary feature data of multiple initial variant text samples, and determine the violation evidence record of each target variant text sample.

[0110] Based on the above scheme, optionally, the supplementary feature data determination submodule includes a variant sample group division unit and a supplementary feature data determination unit. The variant sample group division unit is used to divide the multiple initial variant text samples into multiple variant sample groups corresponding to semantic categories based on the variant text feature data of the multiple initial variant text samples; the supplementary feature data determination unit is used to determine, for each variant sample group, supplementary feature data of the multiple initial variant text samples in the variant sample group based on the semantic category corresponding to the variant sample group.

[0111] Optionally, based on the above scheme, the violation evidence record determination submodule includes a target variant text sample determination unit. The target variant text sample determination unit is used to extract key feature data related to the corresponding semantic category from the supplementary feature data, and determine multiple target variant text samples from the multiple initial variant text samples based on the key feature data and a preset violation feature template corresponding to the semantic category.

[0112] Optionally, based on the above scheme, the device further includes a semantic supplementary data generation module and a logical rule set update module. The semantic supplementary data generation module is used to, after generating a violation type detection result based on the suspected violation type, determine a violation analysis object based on the violation type detection result and the target content data, obtain object attribute data of the violation analysis object, determine context information associated with the violation type detection result based on the object attribute data, and generate semantic supplementary data based on the context information. The logical rule set update module is used to construct a structured extended dataset based on the target content data, the object attribute data, and the semantic supplementary data, and update at least some of the violation semantic conditions in the logical rule set based on the structured extended dataset.

[0113] The live content detection device provided in this embodiment of the invention can execute the live content detection method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0114] Example 4 Figure 4 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0115] like Figure 4As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0116] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0117] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as live content detection methods.

[0118] In some embodiments, the live content detection method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the live content detection method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the live content detection method by any other suitable means (e.g., by means of firmware).

[0119] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0120] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0121] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0122] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0123] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0124] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0125] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from ROM 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of the embodiments of the present invention.

[0126] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0127] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting live streaming content, characterized in that, include: Obtain target content data associated with the live stream from the live stream data stream, determine the content feature data of the target content data, determine the live stream scene corresponding to the live stream content, determine the suspected violation type corresponding to the live stream content based on the content feature data, the live stream scene, and the violation type feature corresponding to the live stream scene, and generate a violation type detection result based on the suspected violation type. Based on the violation type detection result and the target content data, the target violation semantic features are determined, and the target violation semantic features are provided to the pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

2. The live streaming content detection method according to claim 1, characterized in that, Determining the live streaming scenario corresponding to the live streaming content includes: The violation semantic detection result of the target content data is determined based on the content feature data and multiple violation semantic conditions in the logical rule set; In response to the semantic violation detection result indicating the existence of a semantic violation, the live streaming scenario corresponding to the live streaming content is determined.

3. The live streaming content detection method according to claim 2, characterized in that, Before determining the violation semantic detection result of the target content data based on the content feature data and multiple violation semantic conditions in the logical rule set, the method further includes: Obtain the content constraint text associated with the live stream content, and construct multiple structured semantic representations associated with the live stream content based on the content constraint text. The structured semantic representations include multiple entities associated with the live stream content, attribute information of the entities, and the relationships between the entities. Based on the structured semantic representation, the logical relationships between multiple entities are determined. Using the entities as nodes and the logical relationships between the entities as edges, a content constraint semantic graph is constructed. A set of logical rules is generated based on the content constraint semantic graph. The set of logical rules includes multiple violation semantic conditions, which are used to detect whether there is any violation content in the live broadcast content.

4. The live streaming content detection method according to claim 2, characterized in that, After obtaining the target semantic classification result, the following is also included: Based on the target semantic classification results and the target content data, violation evidence records are determined, a violation semantic detection report of the live content is generated based on the violation evidence records, and key semantic fragments are extracted from the target content data based on the violation semantic detection report. A structured violation description is determined based on the key semantic fragments and the violation evidence records. A violation content detection report for the live broadcast content is generated based on the structured violation description and multiple violation semantic conditions in the logical rule set.

5. The live streaming content detection method according to claim 4, characterized in that, After generating the violation content detection report for the live stream content, the method further includes: The system receives target feedback data for the violation detection report and generates optimization guidance information based on the target feedback data, which includes at least one of the accuracy data of the violation detection results and appeal information for the target content data.

6. The live streaming content detection method according to claim 1, characterized in that, After determining the target semantic classification result of the live stream content based on the target violation semantic features and the semantic classification model, the method further includes: Based on the target violation semantic features of the multiple target content data, multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records are determined. The supplementary feature data includes interactive feature data associated with the variant text feature data and contextual information of the interactive feature data. A variant text extended dataset is constructed based on multiple target variant text samples, the supplementary feature data of the target variant text samples, and the violation evidence records of the target variant text samples. The semantic classification model is incrementally trained based on the variant text extended dataset to update the semantic classification model.

7. The live streaming content detection method according to claim 6, characterized in that, The step of determining multiple target variant text samples, supplementary feature data of the target variant text samples, and violation evidence records based on the target violation semantic features of multiple target content data includes: Multiple initial variant text samples are generated based on the target violation semantic features of multiple target content data, and variant text feature data of each initial variant text sample are determined respectively; Supplementary feature data for multiple initial variant text samples are determined based on the variant text feature data of multiple initial variant text samples; Multiple target variant text samples are determined from the multiple initial variant text samples based on supplementary feature data of the multiple initial variant text samples, and violation evidence records are determined for each target variant text sample.

8. The live streaming content detection method according to claim 7, characterized in that, The step of determining supplementary feature data for multiple initial variant text samples based on the variant text feature data of multiple initial variant text samples includes: Based on the variant text feature data of the multiple initial variant text samples, the multiple initial variant text samples are divided into multiple variant sample groups corresponding to multiple semantic categories; For each variant sample group, supplementary feature data of multiple initial variant text samples in the variant sample group are determined according to the semantic category corresponding to the variant sample group.

9. The live streaming content detection method according to claim 7, characterized in that, The step of determining multiple target variant text samples from the multiple initial variant text samples based on supplementary feature data of the multiple initial variant text samples includes: Extract key feature data related to the corresponding semantic category from the supplementary feature data, and determine multiple target variant text samples from the multiple initial variant text samples based on the key feature data and the preset violation feature template corresponding to the semantic category.

10. The live streaming content detection method according to claim 1, characterized in that, After generating the violation type detection result based on the suspected violation type, the method further includes: Based on the violation type detection result and the target content data, a violation analysis object is determined, object attribute data of the violation analysis object is obtained, context information associated with the violation type detection result is determined based on the object attribute data, and semantic supplementary data is generated based on the context information. A structured extended dataset is constructed based on the target content data, the object attribute data, and the semantic supplementary data. At least some of the violation semantic conditions in the logical rule set are updated based on the structured extended dataset.

11. A live streaming content detection device, characterized in that, include: The violation type detection module is used to obtain target content data associated with the live content from the live data stream, determine the content feature data of the target content data, determine the live scene corresponding to the live content, determine the suspected violation type corresponding to the live content based on the content feature data, the live scene and the violation type feature corresponding to the live scene, and generate a violation type detection result based on the suspected violation type. The violation semantic classification module is used to determine the target violation semantic features based on the violation type detection result and the target content data, and provide the target violation semantic features to a pre-trained semantic classification model to obtain the target semantic classification result. The semantic classification model is obtained by training a machine learning model based on the sample violation semantic features and their corresponding expected violation categories.

12. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the live content detection method as described in any one of claims 1-10.

13. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the live content detection method as described in any one of claims 1-10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the live content detection method as described in any one of claims 1-10.

Images