Text label generation method and device, equipment, medium and product

Abstract

The invention relates to a text label generation method and device, equipment, a medium and a product. The method comprises the following steps: performing preliminary prediction processing on a to-be-marked text by utilizing a preset student model to obtain an initial prediction vector; based on the initial prediction vector, calling a teacher model corresponding to the student model from the cloud to carry out residual calculation processing to obtain residual data; performing correction processing on the initial prediction vector by using the residual data to obtain at least one candidate tag; and performing anti-illusion regularization processing on the candidate tag to obtain a target tag corresponding to the to-be-marked text. By adopting the method, multi-label classification can be efficiently and accurately performed on massive text data.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium and product for generating text tags. Background Technology

[0002] In the rapidly developing content ecosystem and large-scale information processing field, the accurate, real-time, and multi-dimensional generation of tags for massive amounts of text data is a core challenge for applications such as text content management and intelligent analysis.

[0003] In traditional technologies, text tag generation methods mainly rely on manually defined keywords or regular expressions to match text content and generate fixed tags, or on using traditional machine learning models, such as support vector machines, random forests, or deep learning models, to perform classification tasks and generate tags by training the model with labeled data.

[0004] However, current text tag generation methods suffer from low tag generation efficiency. Summary of the Invention

[0005] Therefore, it is necessary to provide a text tag generation method, apparatus, device, medium, and product that can efficiently and accurately perform multi-tag classification on massive text data in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a method for generating text tags, including:

[0007] The text to be marked is initially predicted using a pre-defined student model to obtain an initial prediction vector.

[0008] Based on the initial prediction vector, the teacher model corresponding to the student model is called from the cloud to perform residual calculation and processing to obtain residual data;

[0009] The initial prediction vector is corrected using the residual data to obtain at least one candidate label;

[0010] The candidate tags are subjected to anti-illusion regularization to obtain the target tags corresponding to the text to be tagged.

[0011] In one embodiment, based on the initial prediction vector, the teacher model corresponding to the student model is retrieved from the cloud for residual calculation to obtain residual data, including:

[0012] The initial prediction vector is quantized and encoded to obtain the encoded initial prediction vector. The encoded initial prediction vector is then sent to the cloud, where the cloud can call the teacher model to perform forward propagation processing on the encoded initial prediction vector to obtain the correction vector.

[0013] The residual data is obtained by performing residual calculation based on the corrected vector and the initial prediction vector.

[0014] In one embodiment, the initial prediction vector is corrected using residual data to obtain at least one candidate label, including:

[0015] The residual data is dequantized to obtain the residual correction vector;

[0016] The residual correction vector is added to the initial prediction vector to obtain the target prediction vector;

[0017] The target prediction vector is subjected to category mapping and filtering to obtain at least one candidate label.

[0018] In one embodiment, the candidate tags are subjected to anti-illusion regularization to obtain the target tags corresponding to the text to be tagged, including:

[0019] For each candidate label, a causal verification process is performed to obtain the causal effect value corresponding to the candidate label;

[0020] If the causal effect value meets the preset verification range, the candidate tags are subjected to knowledge graph consistency verification to obtain the consistency verification result.

[0021] If the consistency check result is ambiguous, perform a correlation query on each candidate label to obtain the target label.

[0022] In one embodiment, the method further includes:

[0023] Obtain multimodal information;

[0024] Construct a multimodal hypergraph based on multimodal information;

[0025] The parameters of a pre-defined multimodal large model are adjusted using a multimodal hypergraph to obtain a teacher model;

[0026] The teacher model is semantically distilled to obtain the student model.

[0027] In one embodiment, the method further includes:

[0028] Acquire user feedback data and external knowledge data;

[0029] The student model is incrementally trained based on user feedback data and external knowledge data to obtain an updated student model.

[0030] Secondly, this application also provides a text label generation apparatus, comprising:

[0031] The preliminary prediction module is used to perform preliminary prediction processing on the text to be marked using a preset student model to obtain an initial prediction vector.

[0032] The residual calculation module is used to perform residual calculation processing based on the initial prediction vector by calling the teacher model corresponding to the student model from the cloud to obtain residual data.

[0033] The residual correction module is used to correct the initial prediction vector using residual data to obtain at least one candidate label;

[0034] The tag filtering module is used to perform anti-illusion regularization on candidate tags to obtain the target tags corresponding to the text to be tagged.

[0035] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the text tag generation method as described in the first aspect.

[0036] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the text tag generation method of the first aspect.

[0037] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the text tag generation method as described in the first aspect.

[0038] The aforementioned text label generation method, apparatus, device, medium, and product utilize a pre-deployed student model to perform preliminary prediction processing on resource-constrained terminals such as mobile devices. Meanwhile, a teacher model calculates residual data in the cloud and transmits this residual data back to the terminal to correct its initial prediction vector. This edge-cloud collaborative residual learning reduces data transmission volume, improves the interaction speed between the terminal and server, and achieves prediction accuracy approaching that of the teacher model, while maintaining a significantly lower overall inference latency than the ideal effect of pure cloud-based large-model inference. Furthermore, anti-illusion regularization is applied to the candidate labels jointly generated by the student and teacher models to enhance the logical consistency, credibility, and accuracy of the target labels, ensuring the efficiency and accuracy of the text label generation method. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is an application environment diagram of a text label generation method in one embodiment;

[0041] Figure 2 This is a flowchart illustrating a text tag generation method in one embodiment;

[0042] Figure 3 This is a flowchart illustrating the steps for obtaining residual data in one embodiment;

[0043] Figure 4 This is a flowchart illustrating the steps for obtaining the target label in one embodiment;

[0044] Figure 5 This is a flowchart illustrating a text label generation method in another embodiment;

[0045] Figure 6 This is a structural block diagram of a text label generation device in one embodiment;

[0046] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0049] The text tag generation method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located on the cloud or other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0050] In one exemplary embodiment, such as Figure 2 As shown, a text tag generation method is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 202 to 208. Wherein:

[0051] Step 202: Use the preset student model to perform preliminary prediction processing on the text to be marked to obtain the initial prediction vector.

[0052] The text to be labeled includes at least one of text data, image data, and video data. The preliminary prediction process may include: preprocessing the text to be labeled to obtain initial features.

[0053] For example, the preprocessing process may include: segmenting text data into words, dividing continuous text into independent units, followed by named entity recognition to extract entities with specific meanings from the text, providing structured information for subsequent semantic understanding; using a lightweight convolutional neural network, such as MobileNet-V3, to extract features from image data, quickly performing forward propagation on each image to extract representative visual features and converting them into fixed-dimensional feature vectors; and performing keyframe extraction on video data, processing frames with incremental information, and simultaneously using automatic speech recognition technology to transcribe the video's audio content into text data.

[0054] After completing the above feature extraction process, the features extracted from various types of data are fused together, and a lightweight initial prediction vector is calculated using a preset student model.

[0055] The student model is a miniature network trained by knowledge distillation, which requires minimal computation and is designed to perform preliminary analysis and prediction of the labeled text with extremely low latency and resource consumption.

[0056] Step 204: Based on the initial prediction vector, call the teacher model corresponding to the student model from the cloud to perform residual calculation processing to obtain residual data.

[0057] The cloud-based server deploys a more powerful teacher model with a larger number of parameters, corresponding to the student model. The student model is a lightweight version of the teacher model obtained through hypergraph semantic distillation. This teacher model performs deeper feature extraction on the text to be labeled, obtaining the complete feature vector of the text across all potential labels through forward propagation. The cloud server then calculates the residual data between the complete feature vector and the initial prediction vector output by the student model. The residual data has relatively low information entropy, which greatly reduces the amount of communication data between the edge and the cloud, thereby significantly reducing network transmission load and minimizing network latency for end-to-end inference.

[0058] Step 206: Use the residual data to correct the initial prediction vector to obtain at least one candidate label.

[0059] The terminal corrects the initial prediction vector based on the received lightweight residual data to obtain a prediction vector with accuracy close to that of the teacher model.

[0060] Step 208: Perform anti-illusion regularization on the candidate tags to obtain the target tags corresponding to the text to be tagged.

[0061] In some embodiments, the method further includes: obtaining an attention matrix generated by the student model when reasoning about the text to be labeled; by analyzing the attention matrix, it is possible to quantify which semantic units in the text to be labeled are considered the most critical information when the student model makes predictions; after extracting several semantic units with higher weights, performing context mapping to locate these semantic units back to the original text of the text to be labeled, so as to establish a direct association between the model's internal decision-making basis and the human-readable original text.

[0062] Using a pre-set natural language interpretation template, the complex reasoning process is transformed into natural language interpretation text. This natural language interpretation text, along with key metadata such as target tags, is encapsulated into an audit record and written into an audit queue. This queue serves as a reliable message bus, streaming data to the operations backend or monitoring system, allowing operations personnel to view the generation basis and knowledge support for each target tag in real time, thus achieving transparency and traceability of the tag generation process.

[0063] In the aforementioned text label generation method, a pre-deployed student model is used on resource-constrained terminals such as mobile devices to perform preliminary prediction processing on the text to be labeled. Meanwhile, a teacher model is used in the cloud to calculate residual data, which is then transmitted back to the terminal to correct the initial prediction vector. This edge-cloud collaborative residual learning reduces data transmission volume, improves the interaction speed between the terminal and the server, and achieves prediction accuracy approaching that of the teacher model, while the total inference latency is far lower than the ideal effect of pure cloud-based large model inference. Furthermore, anti-illusion regularization is applied to the candidate labels jointly generated by the student and teacher models to improve the logical consistency, credibility, and accuracy of the target labels, ensuring the efficiency and accuracy of the text label generation method.

[0064] In an exemplary embodiment, the method further includes: acquiring multimodal information; constructing a multimodal hypergraph based on the multimodal information; adjusting the parameters of a preset multimodal large model using the multimodal hypergraph to obtain a teacher model; and performing semantic distillation on the teacher model to obtain a student model.

[0065] Among them, multimodal information can be all text and image content on the Internet in the past 12 months, ensuring the timeliness and comprehensiveness of the data.

[0066] Before data import, a rigorous data cleaning process is required. This includes using fingerprinting technology to remove duplicate content, de-identifying data to protect user privacy, and filtering politically sensitive content based on pre-trained models and keyword lists to ensure the dataset's legality and compliance. After data cleaning, a high-quality seed label set can be efficiently constructed using an active learning strategy. The seed label set includes multiple labeled samples, each stored in a structured format, specifically including text ID, title, body text, OCR (Optical Character Recognition) results for image content, ASR (Automatic Speech Recognition) transcribed text for video content, label, and label confidence level.

[0067] To capture the complex relationships between multimodal elements within an article and its external context, each sample is abstracted as a hyperedge to connect a set of nodes representing different semantic levels, collectively forming a sparse hypergraph for training. These connected nodes include four types: First, explicit entities extracted through named entity recognition, representing subjects explicitly mentioned in the article; second, main topics in the implicit topic distribution extracted through a latent Dirichlet assignment model, representing the macro-semantic direction of the text; third, a 7-dimensional sentiment vector calculated by a deep learning model, representing the overall sentiment polarity of the text, including fine-grained dimensions such as positive, negative, and neutral; and fourth, event trigger words extracted from the text using dependency parsing techniques, used to capture the specific actions and events reported in the article.

[0068] The hypergraph is represented by a sparse tensor H, which can be expressed as:

[0069] H∈R^(|V|×|E|×k),

[0070] Where V is the set of all nodes, E is the set of all hyperedges, and k is the feature dimension of each node. In this embodiment, it is uniformly set to 768 dimensions to ensure the richness of features.

[0071] After constructing the hypergraph dataset, a large multimodal model, such as the MoonGPT-Pro multimodal model with up to 32 bytes of parameters, can be selected as the teacher model. Through low-rank adaptation, only the attention weight matrix of the model is decomposed into low rank and these newly added parameters are trained, significantly reducing computational resource consumption and memory usage. The input structure of the teacher model is: concatenating the features of all nodes associated with the hyperedges, supplemented by positional encoding to distinguish node order, and modal encoding to distinguish modal sources such as text, image, and sentiment. The output is a label distribution vector for all labels. The loss function can be a composite loss function including cross-entropy, hypergraph Laplacian regularization, and anti-illusion regularization terms.

[0072] For the teacher model, a knowledge distillation process is initiated to transfer the deep semantic capabilities of the teacher model to a lightweight student model. The transfer process may include using the last hidden state of the teacher model as a soft label to train the initial model. The goal is to minimize the distance between the last hidden state of the teacher model and the last hidden state of the student model, resulting in the distilled student model, which is then deployed to the terminal.

[0073] In one possible implementation, the method further includes: acquiring user feedback data and external knowledge data; and incrementally training the student model based on the user feedback data and external knowledge data to obtain an updated student model.

[0074] User feedback data can be collected by starting a periodic maintenance task at midnight every day, which retrieves behavioral data generated by all end users in the past 24 hours, including clicks, likes, and reports.

[0075] In some embodiments, a Dynamic Knowledge Fusion Graph (DKF-Graph) is constructed and initialized as a central knowledge base for dynamic knowledge. This graph adopts a multi-relation heterogeneous graph structure, and the node types are explicitly divided into three categories: event nodes, such as those from external knowledge sources like breaking news headlines and trending topics; entity nodes, which can include concrete objects such as companies, people, and products; and tag nodes, corresponding to existing classification concepts in the tag ontology tree. Edge types include: co-occurrence, causality, reference, and sentiment. For model initialization, an architecture combining R-GCN (Relation-Graph Convolutional Networks) and a temporal encoder is adopted, enabling it to handle heterogeneous multi-relation graphs and capture the characteristics of knowledge evolution over time. The initial parameters are directly derived from the semantically distilled student model, ensuring that the DKF-Graph inherits all the basic semantic capabilities of the student model from the very beginning.

[0076] During incremental training, a neighbor sampling strategy is employed, sampling and training only on local subgraphs associated with newly added event nodes or high-frequency interaction nodes. This ensures the model learns new knowledge quickly while maintaining computational efficiency. For weight updates, an efficient parameter fine-tuning technique is used, updating only a small number of parameter layers in the student model related to graph neural networks or knowledge integration, while freezing the main parameters. This approach not only effectively avoids catastrophic forgetting of existing knowledge but also enables rapid and accurate absorption of the latest event knowledge, allowing for real-time updates to the student model's weights to handle dynamic content.

[0077] In some embodiments, during incremental training, the frequency of occurrence of newly added event nodes or potential new tags is continuously monitored. When the frequency of occurrence of an uncollected tag exceeds 1,000 times within 24 hours, it indicates that the tag has formed a stable level of social attention, and the ontology tree automatic expansion process will be automatically triggered to ensure the integrity and timeliness of the tag knowledge base.

[0078] The automatic expansion process refers to submitting the tag to the operations staff for manual review. If the review is successful, the insertion operation of the ontology tree will be performed, inserting the new tag into the corresponding level of the ontology tree, assigning a globally unique ID, and incorporating the newly added tag dimension into the hypergraph's data structure definition. This ensures that the model can recognize and learn the new tag in the future, thus completing a closed-loop adaptive process from knowledge discovery, manual verification to model structure update.

[0079] In this embodiment, a hypergraph is used to structurally integrate multimodal information, and the teacher model is fine-tuned through a loss function. Finally, the deep semantic knowledge of the teacher model is distilled into the edge student model, enabling the student model to transfer the generalization ability, multi-task processing ability, and even complex reasoning logic of the teacher model. This enhances the robustness of the student model, achieving near-teacher model accuracy with a significantly reduced parameter size and a significantly improved inference speed. This effectively reduces deployment costs and extends to resource-constrained devices, providing powerful support for artificial intelligence in edge computing, real-time interaction, and other scenarios.

[0080] In one exemplary embodiment, based on Figure 2 The illustrated embodiments, such as Figure 3 As shown, based on the initial prediction vector, the teacher model corresponding to the student model is called from the cloud to perform residual calculation processing, obtaining residual data, including:

[0081] Step 302: Quantize and encode the initial prediction vector to obtain the encoded initial prediction vector, and send the encoded initial prediction vector to the cloud so that the cloud can call the teacher model to perform forward propagation processing on the encoded initial prediction vector to obtain the correction vector.

[0082] The initial prediction vector can be quantized and encoded using the PQ (Product Quantization) algorithm, which decomposes the original high-dimensional vector space into multiple low-dimensional subspaces and quantizes each subspace independently, thus achieving efficient vector compression. The encoded initial prediction vector can be packaged together with the unique identifier of the text to be labeled and sent to the cloud server via a mobile network.

[0083] In one possible implementation, the cloud server receives the encoded initial prediction vector and the unique identifier of the text to be labeled, calls the teacher model to perform forward propagation of the complete content of the text to be labeled, and obtains the complete correction vector of the text to be labeled on all potential labels.

[0084] In some embodiments, the correction vector precisely captures the deviation of the student model from the teacher model in prediction accuracy. To efficiently transmit this high-precision residual information back to the terminal, the correction vector is quantized into a compact representation, where each floating-point element is quantized as an 8-bit integer, and then concatenated into a fixed-length 128-bit data packet. This data packet is transmitted to the requesting terminal via a streaming mechanism.

[0085] Step 304: Perform residual calculation based on the correction vector and the initial prediction vector to obtain residual data.

[0086] In one possible implementation, the residual data is used to correct the initial prediction vector to obtain at least one candidate label, including: performing inverse quantization on the residual data to obtain a residual correction vector; performing an addition operation between the residual correction vector and the initial prediction vector to obtain a target prediction vector; and performing class mapping and filtering on the target prediction vector to obtain at least one candidate label.

[0087] In this process, after receiving the data packet from the cloud, the terminal performs a decoding process, dequantizing the 128-bit data packet to recover the residual correction vector that has a corrective effect. The residual correction vector is then added to the initial prediction vector to obtain the target prediction vector. The category mapping can use the Softmax function to convert the target prediction vector into a probability distribution to determine the confidence level of the text to be labeled belonging to each label, and select several labels with higher confidence levels as candidate labels.

[0088] In some embodiments, to ensure prediction quality, a threshold filtering mechanism can be used to remove all low-confidence labels with confidence levels below a preset threshold, thereby obtaining a preliminary set of high-quality candidate labels.

[0089] In this embodiment, the high-precision residuals of the cloud-based teacher model are used to correct the initial predictions of the terminal student model, thereby obtaining a target prediction vector with accuracy close to that of the teacher model. This allows the complex reasoning capabilities of the teacher model to be compiled into the efficient output of the student model, maintaining excellent performance while efficiently compressing and accelerating the model, thus improving the accuracy and reliability of the text label generation method.

[0090] In one exemplary embodiment, such as Figure 4 As shown, anti-illusion regularization is performed on the candidate tags to obtain the target tags corresponding to the text to be tagged, including:

[0091] Step 402: For each candidate label, perform causal verification to obtain the causal effect value corresponding to the candidate label.

[0092] Among them, causal verification can be performed by calling the label causal probe to verify the causal effect. By quantifying the average causal effect between the text to be labeled and the candidate labels, the causal effect value is obtained, thereby distinguishing between erroneous predictions based solely on statistical correlation and effective predictions that truly have strong semantic support.

[0093] Specifically, for each candidate label, the probe designs and simulates a series of counterfactual interventions. If the candidate label is highly dependent on irrelevant features, or its causal effect value is lower than the preset significance threshold, it is considered that the text to be labeled does not have sufficient causal support for the validity of the candidate label. The candidate label will be regarded as a hallucination or low confidence prediction and needs to be removed from the candidate labels.

[0094] Step 404: If the causal effect value meets the preset verification range, perform knowledge graph consistency verification on the candidate tags to obtain the consistency verification result.

[0095] Step 406: If the consistency check result is ambiguous, perform association query processing on each candidate label to obtain the target label.

[0096] In this process, candidate tags verified by causal probes are linked to the knowledge graph for consistency checks. For example, if the text to be tagged mentions "Apple," and the candidate tags include both "technology company" and "fruit," the consistency check result is ambiguous. In this case, the disambiguation module can be invoked for association query processing. Utilizing the structural information of the knowledge graph and the relationships between tags, a probability score is calculated for conflicting paths. The path with the highest probability and the most logically sound is retained, and the ambiguous candidate tags are eliminated, ultimately outputting accurate and logically consistent target tags.

[0097] In this embodiment, causal verification and disambiguation processing are performed on candidate tags, which can solve the ambiguity problem caused by semantic analysis and ensure that the final output target tag is logically consistent, thereby improving the reliability of the text tag generation method.

[0098] In one exemplary embodiment, such as Figure 5 As shown, a text tag generation method is provided, which can be applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 501 to 509. Wherein:

[0099] Step 501: Use the preset student model to perform preliminary prediction processing on the text to be marked to obtain the initial prediction vector.

[0100] Step 502: Quantize and encode the initial prediction vector to obtain the encoded initial prediction vector, and send the encoded initial prediction vector to the cloud so that the cloud can call the teacher model to perform forward propagation processing on the encoded initial prediction vector to obtain the correction vector.

[0101] Step 503: Perform residual calculation based on the correction vector and the initial prediction vector to obtain residual data.

[0102] Step 504: Perform inverse quantization on the residual data to obtain the residual correction vector.

[0103] Step 505: Add the residual correction vector to the initial prediction vector to obtain the target prediction vector.

[0104] Step 506: Perform category mapping and filtering on the target prediction vector to obtain at least one candidate label.

[0105] Step 507: For each candidate label, perform causal verification to obtain the causal effect value corresponding to the candidate label.

[0106] Step 508: If the causal effect value meets the preset verification range, perform knowledge graph consistency verification on the candidate tags to obtain the consistency verification result.

[0107] Step 509: If the consistency check result is ambiguous, perform association query processing on each candidate label to obtain the target label.

[0108] In one embodiment, the method further includes: acquiring multimodal information; constructing a multimodal hypergraph based on the multimodal information; adjusting the parameters of a preset multimodal large model using the multimodal hypergraph to obtain a teacher model; and performing semantic distillation on the teacher model to obtain a student model.

[0109] In one embodiment, the method further includes: acquiring user feedback data and external knowledge data; and incrementally training the student model based on the user feedback data and external knowledge data to obtain an updated student model.

[0110] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0111] Based on the same inventive concept, this application also provides a text tag generating apparatus for implementing the text tag generating method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more text tag generating apparatus embodiments provided below can be found in the limitations of the text tag generating method described above, and will not be repeated here.

[0112] In one exemplary embodiment, such as Figure 6 As shown, a text tag generation device is provided, including: a preliminary prediction module 602, a residual calculation module 604, a residual correction module 606, and a tag filtering module 608, wherein:

[0113] The preliminary prediction module 602 is used to perform preliminary prediction processing on the text to be marked using a preset student model to obtain an initial prediction vector.

[0114] The residual calculation module 604 is used to call the teacher model corresponding to the student model from the cloud based on the initial prediction vector to perform residual calculation processing and obtain residual data.

[0115] The residual correction module 606 is used to correct the initial prediction vector using residual data to obtain at least one candidate label;

[0116] The tag filtering module 608 is used to perform anti-illusion regularization on candidate tags to obtain the target tags corresponding to the text to be tagged.

[0117] In one embodiment, the residual calculation module 604 is further configured to perform quantization encoding on the initial prediction vector to obtain the encoded initial prediction vector, and send the encoded initial prediction vector to the cloud for the cloud to call the teacher model to perform forward propagation processing on the encoded initial prediction vector to obtain the correction vector; and perform residual calculation processing based on the correction vector and the initial prediction vector to obtain residual data.

[0118] In one embodiment, the residual correction module 606 is further configured to perform inverse quantization on the residual data to obtain a residual correction vector; add the residual correction vector to the initial prediction vector to obtain a target prediction vector; and perform category mapping and filtering on the target prediction vector to obtain at least one candidate label.

[0119] In one embodiment, the tag filtering module 608 is further configured to perform causal verification processing on each candidate tag to obtain the causal effect value corresponding to the candidate tag; if the causal effect value meets the preset verification range, perform knowledge graph consistency verification on the candidate tag to obtain the consistency verification result; if the consistency verification result is ambiguous, perform association query processing on each candidate tag to obtain the target tag.

[0120] In one embodiment, the device further includes a model generation module for acquiring multimodal information; constructing a multimodal hypergraph based on the multimodal information; adjusting the parameters of a preset multimodal large model using the multimodal hypergraph to obtain a teacher model; and performing semantic distillation on the teacher model to obtain a student model.

[0121] In one embodiment, the model generation module is further configured to acquire user feedback data and external knowledge data; and to incrementally train the student model based on the user feedback data and external knowledge data to obtain an updated student model.

[0122] Each module in the aforementioned text label generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0123] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a text tag generation method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0124] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0126] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0127] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0128] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0129] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0131] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating text tags, characterized in that, The method includes: The text to be marked is initially predicted using a pre-defined student model to obtain an initial prediction vector. Based on the initial prediction vector, the teacher model corresponding to the student model is called from the cloud to perform residual calculation processing to obtain residual data; The initial prediction vector is corrected using the residual data to obtain at least one candidate label; The candidate tags are subjected to anti-illusion regularization to obtain the target tags corresponding to the text to be labeled.

2. The method according to claim 1, characterized in that, Based on the initial prediction vector, the teacher model corresponding to the student model is retrieved from the cloud for residual calculation to obtain residual data, including: The initial prediction vector is quantized and encoded to obtain an encoded initial prediction vector, and the encoded initial prediction vector is sent to the cloud. The cloud then calls the teacher model to perform forward propagation processing on the encoded initial prediction vector to obtain a correction vector. The residual data is obtained by performing residual calculation based on the corrected vector and the initial prediction vector.

3. The method according to claim 1, characterized in that, The step of correcting the initial prediction vector using the residual data to obtain at least one candidate label includes: The residual data is dequantized to obtain the residual correction vector; The residual correction vector is added to the initial prediction vector to obtain the target prediction vector. The target prediction vector is subjected to category mapping and filtering to obtain at least one candidate label.

4. The method according to claim 1, characterized in that, The step of performing anti-illusion regularization on the candidate tags to obtain the target tags corresponding to the text to be labeled includes: For each candidate label, a causal verification process is performed on the candidate label to obtain the causal effect value corresponding to the candidate label; If the causal effect value meets the preset verification range, the candidate label is subjected to knowledge graph consistency verification to obtain the consistency verification result. If the consistency check result is ambiguous, an association query is performed on each of the candidate labels to obtain the target label.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain multimodal information; Based on the aforementioned multimodal information, a multimodal hypergraph is constructed; The teacher model is obtained by adjusting the parameters of a preset multimodal large model using the multimodal hypergraph. The teacher model is semantically distilled to obtain the student model.

6. The method according to any one of claims 1-4, characterized in that, The method further includes: Acquire user feedback data and external knowledge data; The student model is incrementally trained based on the user feedback data and the external knowledge data to obtain an updated student model.

7. A text tag generation device, characterized in that, The device includes: The preliminary prediction module is used to perform preliminary prediction processing on the text to be marked using a preset student model to obtain an initial prediction vector. The residual calculation module is used to call the teacher model corresponding to the student model from the cloud based on the initial prediction vector to perform residual calculation processing and obtain residual data. The residual correction module is used to correct the initial prediction vector using the residual data to obtain at least one candidate label; The tag filtering module is used to perform anti-illusion regularization processing on the candidate tags to obtain the target tags corresponding to the text to be labeled.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images