Online business knowledge graph annotation method and system based on transfer learning

Abstract

The embodiment of the application discloses an online business knowledge graph annotation method and system based on transfer learning. The method first calls initial business data set related to the target business field from the online business log database. The data set contains multi-type business interaction records and business entity description information in a preset historical time period. Then, the knowledge graph annotation processing based on semantic understanding ability transfer is performed. The source field pre-training model is adapted through transfer learning, and the initial annotation result is output. The active learning strategy based on uncertainty and information entropy is used to optimize the annotation result. Finally, the optimization result is pushed to the business service interaction interface. If the interactive response operation is detected, the learning parameters and the annotation strategy of the active learning strategy are updated. The method can realize efficient and accurate knowledge graph annotation and dynamically adapt to business changes.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an online business knowledge graph annotation method and system based on transfer learning. Background Technology

[0002] In today's digital age, knowledge graphs, as a knowledge representation and management tool, can present complex business information in a structured way, clearly showing the relationships between entities and providing enterprises and organizations with comprehensive and in-depth business insights. By constructing knowledge graphs, multi-source heterogeneous data can be integrated to uncover potential business value and support applications such as decision-making, intelligent search, and recommendation systems. Accurate knowledge graph annotation is a crucial step in ensuring the quality and usability of the knowledge graph. Knowledge graph annotation clarifies the boundaries of entities and the relationships between them, enabling the knowledge graph to accurately reflect the actual business situation and providing strong support for efficient business operations.

[0003] However, existing knowledge graph annotation technologies are insufficient in handling annotation uncertainty and cannot adjust annotation strategies in a timely manner based on user interaction, resulting in low-quality knowledge graphs that are difficult to adapt flexibly to dynamic changes in business. Summary of the Invention

[0004] This application provides an online business knowledge graph annotation method and system based on transfer learning.

[0005] This application provides, in one aspect, an online business knowledge graph annotation method based on transfer learning, applied to an online business knowledge graph annotation system. The method includes:

[0006] The system retrieves an initial business dataset related to the target business domain from the online business log database. The initial business dataset includes multiple types of business interaction records and corresponding business entity description information recorded within a preset historical time period.

[0007] The initial business dataset is subjected to knowledge graph annotation processing based on semantic understanding ability transfer. The semantic understanding model pre-trained in the source domain is adapted to the target business domain through the transfer learning strategy, and the initial annotation results containing entity boundary annotations and entity relationship annotations are output.

[0008] The initial annotation results are optimized using knowledge graph annotation through an active learning strategy to obtain optimized annotation results. The active learning strategy is as follows: the annotation uncertainty is quantified by calculating the ambiguity of entity recognition boundaries and the dispersion of probability distribution of relation extraction results in the initial annotation results; the uncertain annotation samples are prioritized and selective re-annotation is performed based on information entropy theory.

[0009] The optimized annotation results are pushed to the business service interaction interface that is connected to the online business log database. If an interactive response operation is detected in the business service interaction interface within a preset time period after the push, the learning parameters and annotation strategy of the active learning strategy are updated according to the interactive response operation.

[0010] One embodiment of this application provides an online business knowledge graph annotation system, including:

[0011] A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the described online business knowledge graph annotation methods based on transfer learning.

[0012] One embodiment of this application provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the steps of the online business knowledge graph annotation method based on transfer learning. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating an online business knowledge graph annotation method based on transfer learning, provided as an embodiment of this application.

[0015] Figure 2 This is a schematic diagram of the basic structure of an online business knowledge graph annotation system provided in an embodiment of this application.

[0016] Figure 3 This is a functional block diagram of an online business knowledge graph annotation device provided in an embodiment of this application.

[0017] Figure 4 This is a schematic diagram of the interactive environment for an online business knowledge graph annotation method based on transfer learning, as provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] Please see Figure 1 , Figure 1 This is a flowchart of an online business knowledge graph annotation method based on transfer learning provided in an embodiment of this application. The method can be executed by an online business knowledge graph annotation system, or by the online business knowledge graph annotation system and a server. The method may include steps 110-140.

[0020] It is understood that the specific types and parameter configurations of the neural network model algorithms, related network layers and modules involved in the embodiments of this application are described below in a non-limiting exemplary manner, and those skilled in the art can make adaptive adjustments in combination with the actual situation.

[0021] For a Bidirectional Long Short-Term Memory (BiLSTM) network, its input layer receives a preprocessed sequence of business text. The dimension of the input layer is related to the length of the business text sequence. For example, if the processed business text sequence has a fixed length of 200, then the dimension of the input layer is 200. The embedding layer maps each word in the input text sequence to a low-dimensional vector space. The dimension of the embedding layer can be set to 300, meaning that each word will be represented as a 300-dimensional vector.

[0022] The BiLSTM layer contains LSTM units in two directions: forward LSTM units and backward LSTM units. The number of LSTM units in each direction can be set to 128. These two LSTM units process the input sequence separately and then concatenate the results to capture the context information of the sequence. During training, the BiLSTM layer uses the stochastic gradient descent algorithm to update the parameters, and the learning rate can be set to 0.001 to control the step size of the parameter updates.

[0023] The Conditional Random Field (CRF) layer is used for sequence labeling, receiving the output of the BiLSTM layer as input. The main parameter of the CRF layer is the transition matrix, the size of which depends on the number of entity labels. For example, if there are 10 entity labels, the size of the transition matrix is ​​10×10, used to store the transition probabilities between different labels.

[0024] The entity recognition sublayer uses a multi-scale convolutional structure containing convolutional kernels of different scales. For example, three different kernel scales can be set: 3, 5, and 7. The number of kernels at each scale can be set to 64, meaning that each scale of convolution operation will generate 64 feature maps. The stride of the convolutional layer is set to 1, and the padding settings must ensure that the size of the feature maps after the convolution operation is the same as the input.

[0025] The attention mechanism introduced in the relation extraction sublayer can have an attention dimension of 128. This mechanism calculates weights for input features, allowing the model to focus more on important information. When calculating the attention weights, the softmax function is used to normalize the weights to a range of 0 to 1.

[0026] The event encoding layer of the intent recognition model encodes the event type, trigger timestamp, and associated UI element identifier of key interaction events. Event types can be encoded using one-hot encoding; for example, if there are five event types, each event type will be encoded as a 5-dimensional vector. Trigger timestamps can be normalized and mapped to a range of 0 to 1. Associated UI element identifiers can be mapped using an embedding layer with a dimension of 100.

[0027] The transfer feature adapter achieves semantic alignment by minimizing the feature distribution difference between the source and target domains. The maximum mean difference (MMD) algorithm can be used to calculate the feature distribution difference between the two domains. The kernel function in the MMD algorithm can be a Gaussian kernel function, and the kernel bandwidth can be set to 1.

[0028] The intent classification layer performs multi-class probability prediction on the transferred feature representation, using a fully connected layer for classification. The number of neurons in the fully connected layer can be set to 3, because there are intents for annotation correction, entity completion, and relation adjustment. Figure 3 There are several classifications. The activation function uses the softmax function to convert the output into a probability distribution.

[0029] In the reinforcement learning parameters of the domain-adaptive relabeling sub-model, the learning rate is set to 0.0005, and the discount factor is set to 0.9. The maximum number of iterations for adjusting the labeling strategy is set to 20, and the uncertainty reduction threshold is set to 0.01.

[0030] For time series analysis methods that migrate from the source domain, the Autoregressive Integral Moving Average (ARIMA) model can be used. The parameters of the ARIMA model can be set as (p, d, q), where p represents the order of the autoregressive term, which is set to 2; d represents the difference order, which is set to 1; and q represents the order of the moving average term, which is set to 1.

[0031] The pseudo-entity generation model transfers text generation capabilities from the source domain and can use a language model based on a recurrent neural network (RNN). The number of units in the RNN layer is set to 256, and the dimension of the input layer is related to the length of the original entity name. For example, if the maximum length of the original entity name is 50, then the dimension of the input layer is 50. The embedding layer dimension is set to 200, which is used to map the input characters to a vector space. The output layer uses the softmax function to generate the probability distribution of the pseudo-entity names.

[0032] Step 110: Retrieve the initial business dataset related to the target business domain from the online business log database; wherein, the initial business dataset contains multiple types of business interaction records and corresponding business entity description information recorded within a preset historical time period.

[0033] In the online business knowledge graph annotation scenario based on transfer learning, the target business domain is the talent service business domain. The online business log database is the core repository for storing business interaction information, and its internal data structure is specifically designed to efficiently store and manage various types of business data. To retrieve data related to the talent service business domain, data needs to be retrieved from this database according to the set query rules. The preset historical time period is set to ensure the timeliness and relevance of the data; for example, it can be set to the past year, and the data within this time period can reflect the current situation of the talent service business.

[0034] The diverse business interaction records cover various business activities in the talent service field, such as companies posting job information, job seekers submitting resumes, and companies conducting interviews and assessments of job seekers. These business interaction records are stored in the database in various formats, including text and numerical data, and contain rich business information. The corresponding business entity description information details the entities involved in the business interaction, such as the company's name, size, and industry, and the job seeker's name, education, and professional skills. By accessing this data, a comprehensive understanding of the operation of the talent service business can be achieved.

[0035] Step 120: Perform knowledge graph annotation processing based on semantic understanding ability transfer on the initial business dataset. Adapt the pre-trained semantic understanding model of the source domain to the target business domain through the transfer learning strategy, and output the initial annotation results including entity boundary annotations and entity relationship annotations.

[0036] After acquiring the initial business dataset, a transfer learning strategy is needed to apply the pre-trained semantic understanding model from the source domain to the target business domain. The pre-trained semantic understanding model from the source domain has already undergone extensive training in other related domains and possesses a certain level of semantic understanding capability. Through transfer learning, this existing knowledge and capability can be transferred to the talent service business domain, reducing the time and cost of training a model from scratch in the target domain.

[0037] When performing knowledge graph annotation, the model conducts in-depth analysis of the text information in the initial business dataset. First, it identifies the entities and determines their boundaries, identifying which text fragments belong to a single entity. For example, in job postings, "a well-known internet company" can be identified as a company entity, and the model needs to accurately determine the boundaries of this text fragment. Then, it analyzes the relationships between entities, such as the recruitment relationship between the company and job seekers, and the relationship between job seekers and their skills. Finally, it outputs initial annotation results containing entity boundary annotations and annotations of relationships between entities.

[0038] Step 121: Construct a domain adaptation model based on a combination architecture of bidirectional long short-term memory network and conditional random field. Use the feature extraction layer parameters of the semantic understanding model pre-trained in the source domain as the shared weight parameters of the domain adaptation model. Transfer the knowledge graph annotation capability by adding a domain-defined task adaptation layer. The task adaptation layer includes an entity recognition sub-layer and a relation extraction sub-layer. The entity recognition sub-layer uses a multi-scale convolutional structure to capture the boundary features of entities of different lengths. The relation extraction sub-layer introduces an attention mechanism to model the semantic dependency relationship between entity pairs.

[0039] When constructing the domain-adaptive model, a combined architecture of Bidirectional Long Short-Term Memory (BiLSTM) and Conditional Random Field (CRF) was chosen. BiLSTM can capture contextual information in sequence data, making it highly effective for processing text sequences. In text data related to talent services, contextual information is crucial for accurately identifying entities and relationships. For example, in recruitment information, the context can provide key information such as company requirements and job seeker qualifications. CRF, on the other hand, is used for sequence labeling, taking into account the contextual relationships between labels and improving labeling accuracy.

[0040] By using the feature extraction layer parameters of a semantic understanding model pre-trained in the source domain as shared weight parameters in the domain-adapted model, the knowledge and feature extraction capabilities of the source domain can be leveraged. By adding a domain-defined task adaptation layer, the knowledge graph annotation capabilities can be transferred. The task adaptation layer includes an entity recognition sublayer and a relation extraction sublayer.

[0041] The entity recognition sublayer employs a multi-scale convolutional structure, where convolutional kernels of different scales can capture the boundary features of entities of varying lengths. In talent services, entity lengths can vary significantly; for example, company names may be long, while skill names may be short. The multi-scale convolutional structure can adapt to this variation, improving the accuracy of entity recognition.

[0042] The relation extraction sublayer introduces an attention mechanism, which allows the model to focus more on the important semantic information between entity pairs. In talent service businesses, the relationships between entities are complex and diverse. The attention mechanism can help the model better model the semantic dependencies between entity pairs, such as the recruitment relationship between a company and a job seeker, which can be accurately identified through the attention mechanism.

[0043] Step 122: Perform text serialization processing on the business interaction records in the initial business dataset, generate a fixed-length business text sequence through sentence truncation and sequence completion operations, perform placeholder replacement processing on special symbols in the business text sequence, and perform enhanced encoding processing on business-specific terms based on an external business dictionary. The enhanced encoding processing preserves the semantic integrity of terms in the business scenario through the term context window expansion method.

[0044] After building the domain adaptation model, the business interaction records in the initial business dataset need to be preprocessed. First, text serialization is performed to convert the business interaction records into a computer-processable sequence. This step requires word segmentation of the text data, converting each word into a corresponding index to form a word index sequence.

[0045] By performing sentence truncation and sequence padding, the business text sequence is standardized to a fixed length. In talent service operations, the length of business interaction records can vary greatly. To facilitate model processing, it needs to be standardized to a fixed length. If the text sequence length exceeds the fixed length, sentence truncation is performed; if the length is insufficient, sequence padding is performed, typically using predefined padding symbols.

[0046] For special characters in business text sequences, placeholders should be used for replacement. Special characters may interfere with model training; replacing them with placeholders can avoid this interference. For example, punctuation marks, special characters, etc., can be replaced with uniform placeholders.

[0047] For business-specific terms, enhanced encoding is achieved using an external business dictionary. This dictionary contains specialized terms and vocabulary within the talent service business domain. By expanding the term's context window, the surrounding textual information is considered, preserving the semantic integrity of the term within the business context. For example, in job postings, the term "big data analyst" can be understood through context window expansion, revealing its meaning within the specific job requirements, thus enabling better encoding.

[0048] Step 123: Call the entity recognition sub-model of the domain adaptation model to perform entity boundary detection on the pre-encoded business text sequence, guide the model attention weight distribution through the domain entity type dictionary, generate an initial entity boundary candidate set, and perform boundary adjustment on the initial entity boundary candidate set based on the semantic coherence of the entity context.

[0049] After preprocessing the business interaction records, the entity recognition sub-model of the domain adaptation model is invoked to perform entity boundary detection on the pre-encoded business text sequence. The domain entity type dictionary contains type information of various entities in the talent service business domain, such as enterprises, job seekers, positions, and skills. This dictionary guides the distribution of attention weights in the model, making the model focus more on information related to entities.

[0050] When processing pre-encoded business text sequences, the model generates an initial entity boundary candidate set based on the input sequence information. This candidate set contains possible entity boundaries, but may contain some inaccuracies. For example, the model may misidentify some non-entity text fragments as entity boundaries.

[0051] Based on the semantic coherence of entity context, the initial candidate set of entity boundaries is adjusted. In talent services, the contextual information of entities is crucial for determining entity boundaries. For example, in a job posting, "A well-known internet company is recruiting big data analysts," the "well-known internet company" is an entity, and its contextual information indicates that it is the recruiting entity. Its boundary can be accurately determined through semantic coherence of context. If a candidate entity boundary is semantically incoherent in its context, it needs to be corrected to ensure the accuracy of the entity boundary.

[0052] Step 124: Activate the relation extraction sub-model of the domain adaptation model to perform contextual semantic association analysis on entity pairs in the entity boundary annotation results, construct a candidate relation pair set through the entity pair co-occurrence matrix, filter the candidate relation pair set using the relation pattern library of the source domain knowledge graph corresponding to the transfer learning strategy, perform relation type determination based on dependency parsing on the filtered candidate relation pairs, identify the subject-verb-object relationship and adverbial-head relationship of entity pairs in the sentence structure, and determine the entity association relationship annotation results by combining the relation type system of the target business domain.

[0053] After completing entity boundary annotation, the relation extraction sub-model of the domain adaptation model is launched. This sub-model performs contextual semantic association analysis on the entity pairs in the entity boundary annotation results. An entity pair co-occurrence matrix is ​​a statistical tool used to record the frequency of entity pairs appearing together in a sequence of business text. In talent service businesses, the co-occurrence frequency of entity pairs can reflect the potential relationship between them. For example, if companies and job seekers frequently appear together in recruitment information, it indicates that a recruitment relationship may exist between them.

[0054] A candidate relation pair set is constructed using an entity pair co-occurrence matrix, containing entity pairs that may have a relationship. This candidate relation pair set is then filtered using a relation pattern library from the source domain knowledge graph corresponding to the transfer learning strategy. The source domain knowledge graph's relation pattern library contains relation patterns between various entities in the source domain; through comparison and filtering, some candidate relation pairs that do not conform to the relation patterns can be eliminated.

[0055] The filtered candidate relation pairs are then subjected to relation type determination based on dependency parsing. Dependency parsing can identify subject-verb-object and adverbial-head relations in the sentence structure of entity pairs. In talent service business texts, dependency parsing can accurately determine the relationship type between entities. For example, in the sentence "Companies recruit job seekers," dependency parsing can determine that "company" is the subject, "job seeker" is the object, and "recruitment" is the verb, thus establishing the recruitment relationship between the company and the job seeker. Combined with the relation type system of the target business domain, the annotation results of entity associations are determined, clarifying the specific associations between entities.

[0056] Step 125: Merge the entity boundary annotation results and the inter-entity relationship annotation results, establish an internal index of the annotation results through the entity unique identifier, perform priority determination on the annotation items that conflict between entity boundaries and relationships, retain the first conflict item in the entity boundary annotation results with a higher confidence level than the inter-entity relationship annotation results, and perform secondary entity boundary verification on the second conflict item in the inter-entity relationship annotation results with a higher confidence level than the entity boundary annotation results, and generate an initial annotation result containing an entity unique identifier, entity type label, entity context description, and relationship type triplet.

[0057] After obtaining the entity boundary annotation results and the entity relationship annotation results, these two results need to be merged. An internal index of the annotation results is established by assigning a unique identifier to each entity. The unique identifier ensures that each entity is unique within the annotation results, facilitating subsequent querying and management.

[0058] During the fusion process, conflicting annotations between entity boundaries and relationships may arise. In such cases, a priority determination is performed. The confidence scores of the entity boundary annotations and the inter-entity relationship annotations are calculated, reflecting the reliability of the annotation results. The first conflicting entity boundary annotation with a higher confidence score than the inter-entity relationship annotation is retained; that is, if the confidence score of the entity boundary annotation is higher, then that entity boundary annotation is retained.

[0059] For the second conflict item in the entity relationship annotation results with a higher confidence level than the entity boundary annotation results, a secondary validation of the entity boundaries is performed. In talent service operations, the relationships between entities may affect the determination of entity boundaries. For example, if the relationships between entities indicate that a certain entity should contain more textual information, then a secondary validation of the entity boundaries is required to ensure the accuracy of the entity boundaries.

[0060] The final output generates an initial annotation result containing a unique entity identifier, an entity type label, an entity context description, and a relation type triplet. The entity type label clarifies the type of the entity, such as "company" or "job seeker." The entity context description provides specific information about the entity within the business text. The relation type triplet clarifies the relationship between the entities, such as (company, recruitment, job seeker).

[0061] Step 130: Optimize the initial annotation results using knowledge graph annotation through an active learning strategy to obtain optimized annotation results; wherein, the active learning strategy is: to quantify the annotation uncertainty by calculating the fuzziness of entity recognition boundaries and the dispersion of probability distribution of relation extraction results in the initial annotation results, and to prioritize and selectively re-annotate uncertain annotation samples based on information entropy theory.

[0062] The initial annotation results may contain some uncertainty. In order to improve the accuracy of the annotation, an active learning strategy based on uncertainty and information entropy is adopted for optimization. The core idea of ​​this strategy is to quantify the uncertainty of the annotation and prioritize the processing of samples with high uncertainty.

[0063] The ambiguity of entity recognition boundaries and the dispersion of the probability distribution of relation extraction results are calculated in the initial annotation results. The ambiguity of entity recognition boundaries reflects the degree of uncertainty regarding entity boundaries. For example, in text, the boundary of an entity may have multiple possible divisions; this uncertainty can be quantified by ambiguity. The dispersion of the probability distribution of relation extraction results reflects the reliability of the relation extraction results; the more dispersed the probability distribution, the higher the uncertainty of relation extraction.

[0064] In an exemplary embodiment, the ambiguity calculation revolves around the probability distribution characteristics of entity boundary annotations: First, the probability distribution curves of the left and right boundaries of each entity candidate region are extracted. These curves reflect the probability values ​​of different positions as entity boundaries. Based on this, the half-peak width *a* and peak offset *b* of each curve are calculated. The half-peak width *a* represents the width of the curve at half its peak value; a larger width indicates a more dispersed boundary probability distribution. The peak offset *b* represents the deviation between the peak position of the curve and the preset ideal boundary position; a larger deviation indicates higher uncertainty in the boundary position. Subsequently, the product of the half-peak width *a* and the peak offset *b* is used as the core calculation factor, combined with the entity type weight coefficient *c* to construct the boundary ambiguity entropy, thereby quantifying the ambiguity. The calculation logic formula can be simplified to: Boundary ambiguity entropy = c × (a × b), where the weight coefficient *c* is adaptively adjusted according to the importance level of the entity in the target business domain to ensure more accurate boundary ambiguity assessment for important entities.

[0065] In an exemplary embodiment, the dispersion of the probability distribution is quantified by the discrete entropy of the distribution. The calculation logic is as follows: First, obtain the probability distribution vector d of the entity relationship annotations. This vector contains the predicted probability values ​​corresponding to each candidate relationship type. Input the probability distribution vector d into a calculation model based on information entropy theory. The model first performs a logarithmic operation on each probability value in the vector, and then obtains the initial entropy value through weighted summation. To eliminate the calculation bias caused by the number of different relationship types, a normalization coefficient e is introduced to correct the initial entropy value. The normalization coefficient e is positively correlated with the total number of candidate relationship types. The final calculation logic of the discrete entropy of the distribution is: Discrete entropy of the distribution = e × [-Σ(di × ln(di))] (where di is the i-th probability value in the probability distribution vector d). This entropy value is positively correlated with the uniformity of the probability distribution. The larger the entropy value, the less obvious the model's judgment on the relationship type. The higher the dispersion, the stronger the uncertainty of the annotation results.

[0066] Based on information entropy theory, uncertain labeled samples are prioritized. Information entropy is an indicator of information uncertainty; the higher the information entropy, the higher the uncertainty and the higher the priority. Selective re-labeling is performed on high-priority uncertain labeled samples, i.e., these samples are re-labeled to improve the accuracy of the labeling. In this way, the labeling results of the knowledge graph are gradually optimized.

[0067] Step 131: Calculate the confidence score of each entity boundary annotation and the probability distribution vector of the entity association annotation in the initial annotation result, and mark the annotation samples with confidence scores lower than the preset score as samples to be optimized.

[0068] When using an active learning strategy for optimization, the first step is to calculate the confidence score of each entity boundary label and the probability distribution vector of the entity relationship labels in the initial annotation results. The confidence score of an entity boundary label indicates the reliability of that entity boundary label and is usually provided by the model during the annotation process. For example, when identifying entity boundaries, the model calculates the confidence score of each possible entity boundary based on the input text information and its own training parameters.

[0069] The probability distribution vector of entity relationship annotations reflects the likelihood of different relationship types. During the relationship extraction process, the model predicts the relationships between each pair of entities and provides the probability of different relationship types. For example, for the entity pair of "company" and "job seeker," the model might provide the probability of different relationship types such as "recruitment," "interview," and "rejection."

[0070] Set a preset score, and mark labeled samples with confidence scores below this preset score as samples to be optimized. These samples require further processing to improve labeling accuracy. The preset score needs to be adjusted according to the actual situation to ensure that the samples that truly need optimization are selected.

[0071] Step 132: Perform uncertainty quantification processing based on information entropy on the sample to be optimized, calculate the boundary fuzzy entropy of entity boundary annotation and the distribution discrete entropy of relation annotation, and construct a comprehensive uncertainty evaluation index.

[0072] For samples marked as requiring optimization, uncertainty quantization based on information entropy is performed. The boundary fuzzy entropy of entity boundary labels and the distribution discrete entropy of relation labels are calculated separately.

[0073] When calculating the boundary fuzzy entropy of entity boundary annotations, the probability distribution of the entity boundaries is analyzed. The probability distribution of entity boundaries reflects the likelihood of them existing as entity boundaries at different locations. By calculating the boundary fuzzy entropy, the degree of fuzziness of the entity boundaries can be measured. For example, if the probability distribution of entity boundaries is relatively dispersed, it indicates that the uncertainty of the boundary is high, and the boundary fuzzy entropy will also be large.

[0074] For the discrete entropy distribution of relation labels, we analyze the probability distribution vector of the relation labels. The probability distribution vector of relation labels reflects the likelihood of different relation types. The larger the discrete entropy distribution, the higher the uncertainty of the relation labels. For example, if the probability distribution of relation labels is relatively uniform, it indicates that the model does not have a clear bias in judging different relation types, and the uncertainty is high.

[0075] A comprehensive uncertainty assessment index is constructed, which takes into account both boundary fuzzy entropy and distribution discrete entropy. This comprehensive uncertainty assessment index can more comprehensively evaluate the uncertainty of samples and provide a more accurate basis for relabeling.

[0076] Step 1321: For the entity boundary annotation results, extract the left and right boundary probability distribution curves of each entity candidate region, calculate the half-peak width and peak offset of the left and right boundary probability distribution curves, and calculate the boundary fuzzy entropy based on the product of the half-peak width and peak offset.

[0077] When calculating the boundary fuzzy entropy of entity boundary annotations, the probability distribution curves of the left and right boundaries of each entity candidate region are extracted based on the entity boundary annotation results. The left and right boundary probability distribution curves reflect the probability of being an entity boundary at different locations.

[0078] Calculate the half-maximum width (HWW) of the probability distribution curves for the left and right boundaries. The HWW represents the width of the curve at half its peak value. A larger HWW indicates a more dispersed boundary probability distribution and higher uncertainty regarding the entity boundary. Simultaneously, calculate the peak offset, which represents the deviation between the curve peak position and the expected position. A larger peak offset indicates greater uncertainty regarding the entity boundary position.

[0079] The boundary fuzzy entropy is calculated by multiplying the half-peak width and the peak offset. This calculation method takes into account the dispersion of the boundary probability distribution and the uncertainty of the peak position, and can more accurately measure the fuzziness of the entity boundary.

[0080] Step 1322: For the entity association labeling results, input the relationship probability distribution vector into the information entropy model to obtain the discrete entropy of the relationship labeling. The discrete entropy value is positively correlated with the uniformity of the probability distribution.

[0081] For the annotation results of relationships between entities, the relationship probability distribution vector is input into the information entropy model. The information entropy model calculates the discrete entropy of the relationship annotation based on the relationship probability distribution vector.

[0082] The discrete entropy value is positively correlated with the uniformity of the probability distribution. If the probability distribution of a relationship is relatively uniform, it indicates that the model does not show a clear bias in judging different relationship types, resulting in higher uncertainty and a larger discrete entropy. For example, in the labeling of relationships between companies and job seekers, if the probabilities of relationship types such as recruitment, interview, and rejection are relatively close, it indicates that the uncertainty of the relationship labeling is high, and the discrete entropy is large.

[0083] By calculating the discrete entropy of the distribution, the uncertainty of the labeling of the relationship between entities can be quantified, providing an important reference for relabeling.

[0084] Step 1323: Construct a two-dimensional uncertainty assessment space containing boundary fuzzy entropy and distribution discrete entropy, and perform normalization processing on the entropy values ​​of the two dimensions so that the processed entropy values ​​are mapped to a preset numerical range.

[0085] After obtaining the boundary fuzzy entropy of entity boundary labels and the discrete distribution entropy of relation labels, a two-dimensional uncertainty assessment space containing these two entropy values ​​is constructed. The two-dimensional uncertainty assessment space can more intuitively display the uncertainty of the samples.

[0086] Since the boundary fuzzy entropy and the discrete distribution entropy may have different value ranges, normalization is performed on the entropy values ​​of both dimensions to facilitate subsequent processing. Normalization maps the entropy values ​​to a preset numerical range, such as [0, 1]. This normalization process ensures that the boundary fuzzy entropy and the discrete distribution entropy of different samples are compared within the same numerical range, facilitating the construction of a comprehensive uncertainty assessment index.

[0087] Step 1324: The normalized boundary fuzzy entropy and the distributed discrete entropy are fused by weighted summation to generate a comprehensive uncertainty assessment index; wherein, the weight coefficient of the boundary fuzzy entropy is adjusted according to the importance level of the entity type in the business domain.

[0088] After normalizing the entropy values, the normalized boundary fuzzy entropy and distributed discrete entropy are fused using a weighted summation method to generate a comprehensive uncertainty assessment index. The weighted summation method can adjust the importance of the boundary fuzzy entropy and distributed discrete entropy according to different needs.

[0089] The weighting coefficients of boundary fuzzy entropy are adjusted according to the importance level of entity types in the business domain. In talent services, different entity types may have different levels of importance. For example, enterprise entities may be more important than some less important skill entities. For entity types with higher importance, the weighting coefficients of their boundary fuzzy entropy are relatively larger, indicating a greater focus on the accuracy of boundary labeling for these entities. In this way, by comprehensively considering the uncertainty of entity boundary labeling and relationship labeling, an indicator that can comprehensively assess sample uncertainty is obtained.

[0090] Step 133: Sort the samples to be optimized in descending order according to the comprehensive uncertainty evaluation index, and select a preset number of samples with the highest ranking to form a candidate relabeling sample set.

[0091] After generating a comprehensive uncertainty assessment index, the samples to be optimized are sorted in descending order according to this index. Samples ranked higher indicate higher uncertainty and need to be processed first.

[0092] A candidate re-labeling sample set is formed by selecting a predetermined number of samples from the top-ranked samples. The predetermined number can be set according to actual conditions, for example, selecting the top 100 samples with the highest uncertainty. In this way, the samples that most need re-labeling are selected, improving the efficiency of labeling optimization.

[0093] Step 134: Call the domain adaptive relabeling sub-model in the transfer learning strategy to perform reinforcement learning-style labeling processing on the candidate relabeling sample set, and adjust the comprehensive uncertainty evaluation index of the samples through multi-round labeling strategy.

[0094] For the candidate re-labeled sample set, the domain-adaptive re-labeling sub-model in the transfer learning strategy is invoked for reinforcement learning-based labeling. The domain-adaptive re-labeling sub-model can adaptively adjust according to the characteristics of the samples.

[0095] During multiple rounds of annotation, the annotation strategy is continuously adjusted, and the model's annotation capabilities are optimized based on feedback from the samples. The core idea of ​​reinforcement learning is to guide model learning through a reward mechanism. In each round of annotation, the model is given corresponding rewards or penalties based on the annotation results of the samples and the actual situation. If the annotation results are accurate, a positive reward is given; if the annotation results are inaccurate, a negative reward is given.

[0096] By adjusting the annotation strategy through multiple rounds, the overall uncertainty evaluation index of the samples is reduced. As the number of annotation rounds increases, the model gradually learns more accurate annotation strategies, thereby improving the accuracy of annotation.

[0097] Step 1341: Initialize the reinforcement learning parameters of the domain adaptive relabeling sub-model, and set the maximum number of iterations for labeling strategy adjustment and the uncertainty reduction threshold.

[0098] When calling the domain adaptive relabeling sub-model for reinforcement learning labeling, the reinforcement learning parameters are first initialized. These parameters include the learning rate, discount factor, etc., which affect the learning speed and convergence of the model.

[0099] The learning rate controls the step size at which the model updates its parameters in each iteration. If the learning rate is too large, the model may skip the optimal solution; if the learning rate is too small, the model will learn very slowly. The discount factor is used to balance the importance of current rewards and future rewards.

[0100] To limit the number of iterations for adjusting the annotation strategy, we set a maximum number of iterations to avoid infinite model iterations and improve processing efficiency. Simultaneously, we set an uncertainty reduction threshold; when the overall uncertainty evaluation index of the samples falls below this threshold, the annotation effect is considered to have met certain requirements, and iteration can stop.

[0101] Step 1342: In each iteration, randomly select a batch of samples from the candidate re-labeled sample set and call the domain adaptation model to generate temporary labeling results for the current batch of samples.

[0102] After initializing the reinforcement learning parameters and setting relevant thresholds, the iteration process begins. Batch samples are randomly drawn from the candidate relabeled sample set, which ensures the randomness of the samples and avoids overfitting of the model to the set samples.

[0103] The domain adaptation model is invoked to label the samples in the current batch, generating temporary labeling results. The domain adaptation model has been trained and optimized in previous steps, enabling accurate labeling of the samples. These temporary labeling results represent the labeling of samples in the current iteration, providing a basis for subsequent strategy adjustments.

[0104] Step 1343: Calculate the difference between the temporary annotation results and the historical annotation results of the batch of samples, and adjust the reward function parameters of reinforcement learning based on the magnitude of the difference.

[0105] After obtaining the temporary annotation results for the current batch of samples, the difference between the temporary annotation results and the historical annotation results for the same batch of samples is calculated. The difference reflects the change between the current annotation results and previous annotation results.

[0106] The reward function parameters of reinforcement learning are adjusted based on the magnitude of the difference. A large difference indicates a significant improvement in the current annotation result, resulting in a higher reward; conversely, a small difference indicates only minor improvement, leading to a lower reward. By adjusting the reward function parameters, the model is guided towards learning more accurate annotations.

[0107] Step 1344: Update the labeling decision threshold of the domain adaptation model based on the output value of the reward function, so that the domain adaptation model adopts enhanced boundary judgment conditions in the uncertain region.

[0108] Based on the output value of the reward function, update the labeling decision threshold of the domain adaptation model. A high reward function output value indicates that the current labeling strategy is effective, and the labeling decision threshold can be appropriately increased to enable the model to use enhanced boundary judgment conditions in uncertain regions.

[0109] In talent services, uncertain areas may contain ambiguous entity boundaries and relationships. By employing enhanced boundary determination criteria, entity boundaries and relationship annotations can be more rigorously screened, improving annotation accuracy.

[0110] Step 1345: When the reduction of the comprehensive uncertainty evaluation index in X consecutive iterations is less than the preset magnitude, stop the reinforcement learning labeling process and output the optimized labeling sub-results of the current batch of samples, where X is a positive integer.

[0111] During multiple iterations, the reduction rate of the comprehensive uncertainty assessment index is monitored. If the reduction rate of the comprehensive uncertainty assessment index is less than the preset range for X consecutive iterations, it indicates that the model's improvement effect is not significant, and reinforcement learning annotation processing is stopped.

[0112] Output the optimized annotation sub-results for the current batch of samples. These results are more accurate annotations obtained after reinforcement learning annotation processing. The value of X can be set according to the actual situation, for example, to 5 rounds.

[0113] Step 135: Fuse the samples after reinforcement learning annotation with the high-confidence samples in the initial annotation results to generate optimized annotation results.

[0114] After obtaining the optimized annotation sub-results for the current batch of samples, the samples processed by reinforcement learning are fused with the high-confidence samples from the initial annotation results. The high-confidence samples are those with high reliability from the initial annotation results.

[0115] By fusion, optimized samples and high-confidence samples are combined to generate the final optimized annotation result, thereby improving the overall accuracy and reliability of the annotation.

[0116] Step 140: Push the optimized annotation results to the business service interaction interface that is connected to the online business log database. If an interactive response operation is detected in the business service interaction interface within a preset time period after the push, update the learning parameters and annotation strategy of the active learning strategy according to the interactive response operation.

[0117] After obtaining the optimized annotation results, they are pushed to the business service interaction interface, which communicates with the online business log database. The business service interaction interface is an important platform for users to interact with the knowledge graph, where users can view, edit, and analyze the optimized annotation results.

[0118] After pushing the optimized annotation results, monitor the interactive response of the business service interface within a preset time period. Interactive response includes user clicks, inputs, selections, and other behaviors, which reflect user feedback and needs regarding the annotation results.

[0119] If an interactive response is detected, the learning parameters and annotation strategy of the active learning strategy are updated based on these responses. For example, if a user modifies the annotation of an entity, it indicates that the annotation may be inaccurate, and the sample selection logic and model adaptation rules of the active learning strategy need to be adjusted to improve the accuracy of the annotation.

[0120] Step 141: Organize the optimized annotation results into structured display data according to business entity type and interaction scenario, and push it to the front-end rendering engine of the business service interaction interface through asynchronous data transmission protocol.

[0121] To better display the optimization annotation results on the business service interface, the results are categorized and organized according to business entity type and interaction scenario. Business entity types can include enterprises, job seekers, positions, skills, etc., while interaction scenarios can include recruitment, training, assessment, etc.

[0122] The categorized and organized results are transformed into structured data, which has a clear structure and format, facilitating processing by the front-end rendering engine. This structured data is then pushed to the front-end rendering engine of the business service interface via an asynchronous data transmission protocol. This asynchronous data transmission protocol allows data transmission without blocking the main thread, improving the user experience.

[0123] Front-end rendering engines can transform structured data into a visual interface, making it easier for users to view and interact with it. For example, company information can be displayed in a list format, and the relationship between job seekers and companies can be shown graphically.

[0124] Step 142: Start the interaction response monitoring thread to capture the user interaction behavior sequence generated by the business service interaction interface within a preset time period in real time. The user interaction behavior sequence includes mouse click trajectory, text input content and interface element selection operation.

[0125] After the optimized annotation results are pushed to the business service interaction interface, the interaction response monitoring thread is started. This thread captures the user interaction behavior sequence generated by the business service interaction interface within a preset time period in real time.

[0126] User interaction sequences include mouse clicks, text input, and selection of UI elements. Mouse clicks can reflect a user's focus; for example, clicking a company name indicates interest. Text input reflects specific user needs; for example, entering keywords to search. UI element selection reflects how users use different functions; for example, selecting filters to filter data. By capturing these interaction sequences, user feedback can be obtained.

[0127] Step 143: Perform semantic intent parsing processing on the user interaction behavior sequence, identify the annotation correction intent, entity supplementation intent or relationship adjustment intent corresponding to the interaction behavior through the intent recognition model in the transfer learning strategy, and generate a structured interaction intent description vector.

[0128] For the captured user interaction sequence, semantic intent parsing is performed. Using the intent recognition model in the transfer learning strategy, the intentions corresponding to the interaction behaviors—such as annotation correction intent, entity supplementation intent, or relationship adjustment intent—are identified.

[0129] The annotation correction intent indicates that the user wants to correct the annotation results. For example, a user finds that the annotation of a certain entity is inaccurate and wants to modify it. The entity supplementation intent indicates that the user wants to add new entity information. For example, a user finds that a certain company is missing information in the knowledge graph and wants to add the company's information. The relationship adjustment intent indicates that the user wants to adjust the relationship between entities. For example, a user finds that the relationship between a company and a job seeker is incorrectly labeled and wants to adjust it. Further, a structured interaction intent description vector is generated, which contains information such as intent type, target entity ID, and intent triggering scenario.

[0130] Step 1431: Extract key interaction events from the user interaction behavior sequence. The key interaction events include text input completion events, entity region selection events, and relationship line dragging events.

[0131] During semantic intent parsing, key interaction events are extracted from the user's interaction sequence. Key interaction events include text input completion events, entity region selection events, and relationship line dragging events, etc.

[0132] Text input completion events may indicate that a user has entered new information; for example, a user pressing Enter after typing keywords in a search box may indicate that the user wants to perform a search. Entity selection events may indicate that a user is interested in a specific entity; for example, clicking a company's icon may indicate that the user wants to view the company's detailed information. Relationship dragging events may indicate that a user wants to adjust the relationship between entities; for example, dragging the line between a company and a job seeker may indicate that the user wants to modify the relationship between them. By extracting key interaction events, user intent can be analyzed more accurately.

[0133] Step 1432: Input the event type, trigger timestamp, and associated UI element identifier of the key interactive event into the event encoding layer of the intent recognition model to generate an event feature sequence.

[0134] The extracted key interaction events are processed by the event encoding layer of the input intent recognition model, which includes the event type, trigger timestamp, and associated UI element identifiers. The event encoding layer encodes this information to generate an event feature sequence. Event types can be represented using different encodings; for example, a text input completion event can be represented by a predefined encoding value. The trigger timestamp records the time the event occurred, and the associated UI element identifier records the UI element associated with the event, such as the ID of a button. The event feature sequence contains the characteristic information of the key interaction events.

[0135] Step 1433: Invoke the transfer feature adapter of the intent recognition model to map the event feature sequence to a feature representation aligned with the semantic space of the target business domain. The transfer feature adapter achieves semantic alignment by minimizing the feature distribution difference between the source domain and the target domain.

[0136] The transfer feature adapter of the intent recognition model is invoked to map the event feature sequence. The transfer feature adapter maps the event feature sequence to a feature representation aligned with the semantic space of the target business domain by minimizing the feature distribution difference between the source and target domains.

[0137] In the talent service business, the source domain may be other related business domains, while the target domain is the talent service business domain. By using a transfer feature adapter, the model can better understand the meaning of event feature sequences in the target business domain, thereby improving the accuracy of intent recognition.

[0138] Step 1434: Perform multi-class probability prediction on the transferred feature representation through the intent classification layer of the intent recognition model, and output the probability distributions corresponding to the labeled correction intent, entity supplementation intent, and relationship adjustment intent.

[0139] The intent classification layer of the intent recognition model performs multi-class probability prediction on the transferred feature representation. The intent classification layer can determine the probability that the transferred feature representation belongs to the annotation correction intent, entity supplementation intent, or relation adjustment intent.

[0140] The output layer defines the probability distributions for annotation correction intent, entity supplementation intent, and relationship adjustment intent. These probability distributions reflect the likelihood of different intents. For example, for a given user interaction, the intent classification layer might output an annotation correction intent with a probability of 0.6, an entity supplementation intent with a probability of 0.2, and a relationship adjustment intent with a probability of 0.2.

[0141] Step 1435: Select the intent type corresponding to the highest probability value in the probability distribution as the dominant interaction intent, and generate an interaction intent description vector containing intent type label, target entity ID and intent triggering scenario description in combination with the interface element identifier in the event feature sequence.

[0142] The intent type corresponding to the highest probability value from the probability distribution is selected as the dominant interaction intent. Combined with the UI element identifiers in the event feature sequence, an interaction intent description vector is generated, containing an intent type label, target entity ID, and a description of the intent triggering scenario.

[0143] The intent type label clarifies the user's primary intent, such as annotation correction intent, entity supplementation intent, or relationship adjustment intent. The target entity ID records the entity IDs related to the intent, and the intent trigger scenario description records the specific scenario in which the intent occurs, such as annotation correction on a recruitment page. The interaction intent description vector clarifies the user's primary intent, as well as related entity and scenario information.

[0144] Step 144: Adjust the sample selection logic and model adaptation rules of the active learning strategy based on the interaction intent description vector; wherein, the sample selection logic prioritizes the selection of labeled samples in the same interaction scenario according to the intent type, and the model adaptation rules are implemented by dynamically adjusting the feature weight allocation strategy of the domain adaptation model.

[0145] Based on the generated interaction intent description vector, the sample selection logic and model adaptation rules of the active learning strategy are adjusted. The sample selection logic prioritizes labeled samples from similar interaction scenarios based on intent type.

[0146] For example, if the interaction intent is to correct annotations, samples with annotations from the same interaction scenario will be prioritized for re-annotation. In talent services, if a user corrects an annotation of a company in a recruitment scenario, the sample selection logic will prioritize re-annotating samples with company annotations from the recruitment scenario.

[0147] Model adaptation rules are implemented by dynamically adjusting the feature weight allocation strategy of the domain adaptation model. Based on the information in the interaction intent description vector, the domain adaptation model adjusts its focus on different features. For example, if the interaction intent involves job seekers' skill information, the model adaptation rule will increase the weight allocation of skill features, improving the model's ability to recognize skill information.

[0148] Step 1441: Initialize the strategy adjustment iteration counter, set the maximum number of iterations and the intent matching threshold, wherein the intent matching threshold is used to determine whether the currently adjusted strategy matches the interaction intent description vector.

[0149] When adjusting the sample selection logic and model adaptation rules of the active learning strategy, an iteration counter for strategy adjustment is initialized. This counter records the number of iterations to control the iteration process. A maximum number of iterations is set to avoid infinite iterations and improve processing efficiency. Simultaneously, an intent matching threshold is set to determine whether the currently adjusted strategy matches the interaction intent description vector. If the matching degree between the currently adjusted strategy and the interaction intent description vector is higher than the intent matching threshold, the strategy adjustment has achieved the expected effect, and iteration can stop; otherwise, iterative adjustment continues.

[0150] Step 1442: In each iteration, retrieve the historical interaction strategy library based on the intent type label in the interaction intent description vector to obtain the sample selection template and model adaptation template corresponding to the same type of intent.

[0151] During each iteration, the historical interaction strategy library is retrieved based on the intent type label in the interaction intent description vector. This library stores past interaction strategy information, including sample selection templates and model adaptation templates. Based on the intent type label, sample selection templates and model adaptation templates corresponding to the same type of intent are retrieved. The sample selection template specifies how to select labeled samples under a given intent, while the model adaptation template specifies how to adjust the feature weight allocation strategy of the domain adaptation model under a given intent. For example, if the intent type label is "label correction intent," the library retrieves sample selection templates and model adaptation templates related to label correction intents.

[0152] Step 1443: Calculate the similarity between the scene features in the sample selection template and the intent trigger scene description in the current interaction intent description vector, and select the scene features with a similarity higher than the preset similarity as the key matching features.

[0153] The similarity between the scene features in the sample selection template and the intent-triggered scene description in the current interaction intent description vector is calculated. Various methods can be used for similarity calculation, such as cosine similarity calculation. Scene features with similarity scores higher than a preset similarity score are selected as key matching features. Key matching features reflect the relevance between the sample selection template and the current interaction intent. For example, if the scene feature in the sample selection template is a recruitment scene, and the intent-triggered scene description in the current interaction intent description vector is also a recruitment scene, and the similarity score is higher than the preset similarity score, then the recruitment scene feature is the key matching feature.

[0154] Step 1444: Adjust the scene weight coefficient of the sample selection logic based on key matching features, so that labeled samples in the same interaction scenario can obtain higher priority weight in the selection process.

[0155] The scenario weight coefficients of the sample selection logic are adjusted based on key matching features. These coefficients determine the selection priority of labeled samples in different interaction scenarios. Labeled samples from interaction scenarios similar to the key matching features receive higher priority during the selection process. For example, if the key matching feature is a recruitment scenario, labeled samples from that scenario will receive higher priority and are more likely to be selected for re-labeling. By adjusting the scenario weight coefficients, the sample selection logic can be made more aligned with the user's interaction intent, improving the efficiency of labeling optimization.

[0156] Step 1445: Synchronously adjust the feature weight allocation strategy in the model adaptation rules, assign higher weight values ​​to entity features and relationship features associated with key matching features, and reduce the weight influence of non-associated features.

[0157] The feature weight allocation strategy in the model adaptation rules is adjusted synchronously. Based on the key matching features, higher weight values ​​are assigned to entity features and relationship features associated with the key matching features. In talent service businesses, if the key matching feature is the recruitment scenario, then entity features associated with the recruitment scenario, such as company recruitment needs and job seeker skill matching, as well as relationship features, such as the recruitment relationship between the company and the job seeker, will be assigned higher weight values. Simultaneously, the influence of unrelated features is reduced to avoid interference from unrelated features on the model. By adjusting the feature weight allocation strategy, the domain adaptation model can focus more on features related to user interaction intent, improving the model's accuracy.

[0158] Step 1446: Calculate the matching degree between the adjusted sample selection logic and model adaptation rules and the interaction intent description vector. If the matching degree is higher than the intent matching threshold or the maximum number of iterations is reached, stop the iteration and output the adjusted sample selection logic and model adaptation rules.

[0159] Calculate the matching degree between the adjusted sample selection logic and model adaptation rules and the interaction intent description vector. The matching degree can be calculated in various ways, such as comparing the relevance between the samples selected by the sample selection logic and the target entities in the interaction intent description vector, and the consistency between the model output after the model adaptation rules are adjusted and the interaction intent.

[0160] If the matching degree is higher than the intent matching threshold or the maximum number of iterations is reached, the iteration stops and the adjusted sample selection logic and model adaptation rules are output. After stopping the iteration, the adjusted sample selection logic and model adaptation rules will be applied to subsequent annotation processes to improve the accuracy and efficiency of annotation.

[0161] Step 145: Integrate the adjusted sample selection logic and model adaptation rules into the updated annotation strategy, and synchronously update the learning parameter storage sub-model of the active learning strategy; wherein, the subsequent annotation process adopts the updated sample selection logic and model adaptation rules.

[0162] The adjusted sample selection logic and model adaptation rules are integrated into an updated annotation strategy. This updated strategy comprehensively considers user interaction intent, enabling more accurate selection of labeled samples and adjustment of model adaptation rules. The learning parameter storage sub-model for the active learning strategy is updated synchronously, storing various learning parameters and rules of the active learning strategy. The updated annotation strategy is then written into the learning parameter storage sub-model, overriding the original sample selection logic and model adaptation rules. Subsequent annotation processing will adopt the updated sample selection logic and model adaptation rules, thereby improving annotation accuracy and efficiency and better meeting user needs.

[0163] In an optional embodiment, the method further includes:

[0164] Step 151: Based on the optimized annotation results, perform knowledge graph entity association evolution update processing. The dynamic change characteristics of the relationship between entities are captured through temporal association modeling. The processing first extracts entity pairs and corresponding relationship type labels from the optimized annotation results to construct an initial set of association relationships.

[0165] After obtaining the optimized annotation results, the knowledge graph entity association evolution update process is performed. Temporal association modeling is used to capture the dynamic changes in the relationships between entities. First, entity pairs and their corresponding relationship type labels are extracted from the optimized annotation results.

[0166] In talent service operations, entity pairs can be companies and job seekers, job seekers and skills, etc., and relationship type tags can be recruitment, mastery, etc. The extracted entity pairs and relationship type tags are used to construct an initial set of relationships, which records the relationships between entities at the current moment.

[0167] Step 152: Invoke the temporal association analysis model in the transfer learning strategy, and perform cross-time period association evolution analysis on the initial association set through time series analysis of the transfer source domain to identify the frequency changes of entity pair associations and the migration pattern of relationship types in different business cycles.

[0168] The temporal association analysis model in the transfer learning strategy is invoked. This model performs cross-period association evolution analysis on the initial set of associations using time series analysis methods from the source domain. In talent service operations, entity-to-entity relationships may change across different business cycles. For example, during peak recruitment seasons, the frequency of recruitment relationships between companies and job seekers may increase; during off-seasons, the frequency may decrease. Simultaneously, relationship types may also shift; for instance, the relationship between a company and a job seeker may change from a recruitment relationship to a collaborative relationship. Through cross-period association evolution analysis, the changes in the frequency of entity-to-entity relationships and the migration patterns of relationship types across different business cycles are identified, providing a basis for updating the knowledge graph.

[0169] Step 153: Construct a dynamic evaluation function for entity association strength based on the association evolution analysis results. The function generates an association strength descriptor by combining the duration of the association relationship, the periodic occurrence pattern, and the importance of the business scenario.

[0170] Based on the results of the association evolution analysis, a dynamic evaluation function for entity association strength is constructed. This function generates an association strength descriptor by combining the duration of the association relationship, the periodic occurrence pattern, and the importance of the business scenario.

[0171] The duration of a relationship reflects the stability of the connection between entities; the longer the duration, the stronger the connection is likely to be. Regular occurrence reflects the frequency and pattern of the relationship across different business cycles; relationships with high frequency and stable patterns are likely to be stronger. The importance of the business context reflects the significance of the relationship within the business; for example, in a recruitment scenario, the relationship between the company and key job seekers may be more important.

[0172] By comprehensively considering these factors, a correlation strength descriptor is generated, which can more accurately reflect the strength of the correlation between entities.

[0173] Step 154: Based on the association strength descriptor, update the weight attributes of existing entity associations in the knowledge graph, temporarily freeze associations with association strength lower than the preset strength, and prioritize newly discovered active associations.

[0174] The weight attributes of existing entity relationships in the knowledge graph are updated based on the association strength descriptor. The weight attributes of the relationships reflect the importance and strength of the relationships. Updating the weight attributes according to the association strength descriptor enables the knowledge graph to more accurately reflect the relationships between entities.

[0175] Relationships with a strength lower than a preset threshold are temporarily frozen. This temporary freezing marks these relationships as inactive, reducing interference with knowledge graph queries and analysis. Newly discovered active relationships are prioritized, increasing their prominence within the knowledge graph. Through these operations, the structure and quality of the knowledge graph are continuously optimized.

[0176] Step 155: Store the updated entity relationships and weight attributes in the relationship table of the knowledge graph database, and generate a relationship evolution update log simultaneously to record the change type, change time, and relationship strength change curve of the relationship.

[0177] The updated entity relationships and weight attributes are stored in the relationship table of the knowledge graph database. The relationship table is the core table in the knowledge graph database for storing entity relationships. Storing the updated information in this table ensures that the latest state of the knowledge graph is preserved.

[0178] Synchronously generate an association evolution update log, recording the type of change, time of change, and association strength change curve of the association relationships. The association evolution update log helps users understand the dynamic changes of association relationships in the knowledge graph, providing a basis for subsequent analysis and decision-making. For example, the association evolution update log can record the changes in the association strength between a company and a job seeker over a certain period, as well as changes in the relationship type.

[0179] In an optional embodiment, the method further includes:

[0180] Step 161: Perform semantic desensitization protection processing on the optimized annotation results and the updated knowledge graph. Sensitive information is hidden through entity semantic replacement and relation abstraction. The processing process uses a named entity recognition model to identify sensitive entity types containing personal identification information and business sensitive data from the optimized annotation results and generate a list of sensitive entities.

[0181] Semantic anonymization is performed on the optimized annotation results and the updated knowledge graph to protect sensitive information. Sensitive information is hidden through entity semantic replacement and relation abstraction. First, a named entity recognition model is used to identify sensitive entity types containing personal identification information and business-sensitive data from the optimized annotation results. In talent services, personal identification information could be a job seeker's name, ID number, etc., while business-sensitive data could be the company's trade secrets, job seekers' salary information, etc. A sensitive entity list is generated from the identified sensitive entity types. This list records the entity information that needs to be anonymized, providing a foundation for subsequent anonymization operations.

[0182] Step 162: Invoke the pseudo-entity generation model in the transfer learning strategy to generate semantically unrelated pseudo-entity names and descriptions for each entity in the sensitive entity list. The pseudo-entity generation model transfers the text generation capabilities of the source domain so that the pseudo-entity names are syntactically consistent with the original entities but do not contain real semantic information.

[0183] The pseudo-entity generation model in the transfer learning strategy is invoked to generate semantically unrelated pseudo-entity names and descriptions for each entity in the sensitive entity list. The pseudo-entity generation model leverages the text generation capabilities of the source domain to generate pseudo-entity names and descriptions that conform to grammatical structure but do not contain true semantic information. For example, for a job seeker's name, a pseudo-name in a similar format can be generated; for a company's trade secrets, a vague description can be generated. In this way, sensitive information is concealed without affecting the overall structure and relationships of the knowledge graph.

[0184] Step 163: Construct a mapping table between sensitive entities and pseudo entities, recording entity ID, original entity information, pseudo entity information, and desensitized timestamp. The mapping table is stored in an encrypted manner to prevent unauthorized access.

[0185] A mapping table between sensitive entities and pseudo-entities is constructed. This table records the entity ID, original entity information, pseudo-entity information, and de-identification timestamp. The entity ID is used to uniquely identify each entity, the original entity information records the original sensitive information, the pseudo-entity information records the generated pseudo-entity information, and the de-identification timestamp records the time of the de-identification operation.

[0186] To protect sensitive information, the mapping table is stored encrypted to prevent unauthorized access. Encrypted storage ensures that only authorized personnel can access and use the mapping table, thus improving the security of sensitive information.

[0187] Step 164: Perform a batch replacement operation on the entity nodes in the knowledge graph, replacing the name and description attributes of sensitive entity nodes with the corresponding pseudo-entity information, while keeping the entity type labels and relationship structure unchanged.

[0188] Perform batch replacement operations on entity nodes in the knowledge graph. Replace the names and description attributes of sensitive entity nodes with the corresponding pseudo-entity information. During the replacement process, maintain the entity type labels and relationship structure unchanged to ensure that the overall structure and relationships of the knowledge graph are not affected.

[0189] For example, the name and details of job seeker nodes can be replaced with pseudo-names and pseudo-descriptions, and the trade secrets of enterprise nodes can be replaced with pseudo-information. Through batch replacement operations, sensitive information can be concealed, protecting user privacy and enterprise trade secrets.

[0190] Step 165: Perform abstraction processing on the relation descriptions in the knowledge graph, replace the relation attribute values ​​containing specific business data with general business terms, generate a de-identified knowledge graph and store it in the de-identified database.

[0191] Abstraction is performed on the relation descriptions in the knowledge graph. Relationship attribute values ​​containing specific business data are replaced with general business terms. In talent service businesses, relation attribute values ​​may contain sensitive information such as specific salaries and contract terms. This information is replaced with general terms such as "reasonable salary" and "standard contract term." This relation abstraction further conceals sensitive information. A de-identified knowledge graph is generated and stored in a de-identified database. This database is specifically used to store the de-identified knowledge graph, ensuring effective protection of sensitive information. Simultaneously, the de-identified knowledge graph can provide users with certain knowledge services without disclosing sensitive information.

[0192] Please see Figure 2 The figure is a schematic diagram of the basic structure of an online business knowledge graph annotation system 200 provided in an embodiment of this application. The online business knowledge graph annotation system 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the online business knowledge graph annotation methods based on transfer learning.

[0193] Please see Figure 3 This application provides a functional block diagram of an online business knowledge graph annotation device, which includes:

[0194] The business data retrieval module is used to retrieve an initial business dataset related to the target business domain from the online business log database; wherein, the initial business dataset includes multiple types of business interaction records and corresponding business entity description information recorded within a preset historical time period;

[0195] The transfer learning annotation module is used to perform knowledge graph annotation processing based on semantic understanding ability transfer on the initial business dataset. It adapts the pre-trained semantic understanding model of the source domain to the target business domain through the transfer learning strategy and outputs the initial annotation results including entity boundary annotations and entity relationship annotations.

[0196] The annotation optimization processing module is used to optimize the knowledge graph annotation of the initial annotation results through an active learning strategy to obtain optimized annotation results. The active learning strategy is as follows: the annotation uncertainty is quantified by calculating the ambiguity of entity recognition boundaries and the dispersion of probability distribution of relation extraction results in the initial annotation results; the uncertain annotation samples are prioritized and selective re-annotation is performed based on information entropy theory.

[0197] The learning strategy update module is used to push the optimized annotation results to the business service interaction interface that is connected to the online business log database. If an interaction response operation is detected on the business service interaction interface within a preset time period after the push, the learning parameters and annotation strategy of the active learning strategy are updated according to the interaction response operation.

[0198] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0199] Please see Figure 4 This application's embodiments achieve efficient and accurate online business knowledge graph annotation. First, by retrieving an initial business dataset related to the target business domain from an online business log database, the relevance and comprehensiveness of the data source are ensured, enabling the knowledge graph to closely align with the target business needs.

[0200] Secondly, the knowledge graph annotation processing based on semantic understanding ability transfer utilizes transfer learning to adapt the pre-trained semantic understanding model in the source domain to the target business domain, avoiding the high cost and low efficiency of training the model from scratch in the target domain, and quickly and effectively outputting the initial annotation results, thus shortening the knowledge graph construction cycle.

[0201] Then, an active learning strategy based on uncertainty and information entropy is used to optimize the initial annotation results. By quantifying the annotation uncertainty and prioritizing and selectively re-annotating uncertain samples, the system can focus on the annotation parts with high uncertainty, thereby improving the accuracy and reliability of the annotation and enhancing the quality of the knowledge graph.

[0202] Finally, the optimized annotation results are pushed to the business service interaction interface, and the learning parameters and annotation strategies of the active learning strategy are updated according to the interaction response. This realizes the dynamic optimization and adaptive adjustment of knowledge graph annotation, enabling the knowledge graph to respond to business changes and user needs in a timely manner, and enhancing the application value and flexibility of the knowledge graph in actual business.

[0203] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the online business knowledge graph annotation system reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the online business knowledge graph annotation system to perform the aforementioned... Figure 1The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0204] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A method for annotating online business knowledge graphs based on transfer learning, characterized in that, The method includes: The system retrieves an initial business dataset related to the target business domain from the online business log database. The initial business dataset includes multiple types of business interaction records and corresponding business entity description information recorded within a preset historical time period. The initial business dataset is subjected to knowledge graph annotation processing based on semantic understanding ability transfer. The semantic understanding model pre-trained in the source domain is adapted to the target business domain through the transfer learning strategy, and the initial annotation results containing entity boundary annotations and entity relationship annotations are output. The initial annotation results are optimized using an active learning strategy to obtain optimized annotation results. This includes: calculating the confidence score of each entity boundary annotation and the probability distribution vector of the inter-entity relationship annotations in the initial annotation results; marking annotation samples with confidence scores lower than a preset score as samples to be optimized; performing uncertainty quantification processing based on information entropy on the samples to be optimized, calculating the boundary fuzziness entropy of entity boundary annotations and the distribution discrete entropy of relationship annotations, and constructing a comprehensive uncertainty evaluation index; sorting the samples to be optimized in descending order according to the comprehensive uncertainty evaluation index, and selecting a preset number of samples from the top of the sorted list to form a candidate re-annotation sample set; calling the domain adaptive re-annotation sub-model in the transfer learning strategy to perform reinforcement learning-based annotation processing on the candidate re-annotation sample set, adjusting and reducing the comprehensive uncertainty evaluation index of the samples through multiple rounds of annotation strategies; and fusing the samples after reinforcement learning-based annotation processing with the high-confidence samples in the initial annotation results to generate optimized annotation results. The active learning strategy involves quantifying the uncertainty of the initial annotation results by calculating the fuzziness of entity recognition boundaries and the discreteness of the probability distribution of relationship extraction results. To address uncertainty, the process involves performing information entropy-based uncertainty quantification on the samples to be optimized. This includes calculating the boundary fuzzy entropy of entity boundary annotations and the discrete entropy of relationship annotations, and constructing a comprehensive uncertainty assessment index. The process includes: extracting the left and right boundary probability distribution curves for each entity candidate region based on the entity boundary annotation results; calculating the half-peak width and peak offset of the left and right boundary probability distribution curves; calculating the boundary fuzzy entropy based on the product of the half-peak width and peak offset; inputting the relationship probability distribution vector into the information entropy model to obtain the discrete entropy of the relationship annotations, where the discrete entropy value is positively correlated with the uniformity of the probability distribution; constructing a two-dimensional uncertainty assessment space containing boundary fuzzy entropy and discrete entropy; normalizing the entropy values ​​of both dimensions to map the processed entropy values ​​to a preset numerical range; fusing the normalized boundary fuzzy entropy and discrete entropy through a weighted summation method to generate a comprehensive uncertainty assessment index; where the weight coefficient of the boundary fuzzy entropy is adjusted according to the importance level of the entity type in the business domain; prioritizing the uncertain annotation samples based on information entropy theory and performing selective re-annotation processing. The optimized annotation results are pushed to the business service interaction interface that is connected to the online business log database. If an interactive response operation is detected in the business service interaction interface within a preset time period after the push, the learning parameters and annotation strategy of the active learning strategy are updated according to the interactive response operation.

2. The method as described in claim 1, characterized in that, The process of performing knowledge graph annotation processing based on semantic understanding capability transfer on the initial business dataset involves adapting the pre-trained semantic understanding model from the source domain to the target business domain using a transfer learning strategy, and outputting initial annotation results including entity boundary annotations and annotations of relationships between entities, including: A domain adaptation model based on a combination architecture of bidirectional long short-term memory network and conditional random field is constructed. The feature extraction layer parameters of the semantic understanding model pre-trained in the source domain are used as the shared weight parameters of the domain adaptation model. The knowledge graph annotation capability is transferred by adding a domain-defined task adaptation layer. The task adaptation layer includes an entity recognition sublayer and a relation extraction sublayer. The entity recognition sublayer uses a multi-scale convolutional structure to capture the boundary features of entities of different lengths. The relation extraction sublayer introduces an attention mechanism to model the semantic dependency relationship between entity pairs. The business interaction records in the initial business dataset are processed by text serialization. A fixed-length business text sequence is generated by sentence truncation and sequence completion. Placeholder replacement is performed on special symbols in the business text sequence. Enhanced encoding processing based on an external business dictionary is performed on business-specific terms. The enhanced encoding processing preserves the semantic integrity of terms in the business scenario through the term context window expansion method. The entity recognition sub-model of the domain adaptation model is invoked to perform entity boundary detection on the pre-encoded business text sequence. The attention weight distribution of the model is guided by the domain entity type dictionary to generate an initial entity boundary candidate set. The boundary adjustment of the initial entity boundary candidate set is performed based on the semantic coherence of the entity context. The relation extraction sub-model of the domain adaptation model is activated to perform contextual semantic association analysis on entity pairs in the entity boundary annotation results. A candidate relation pair set is constructed through the entity pair co-occurrence matrix. The candidate relation pair set is filtered using the relation pattern library of the source domain knowledge graph corresponding to the transfer learning strategy. The relation type determination based on dependency parsing is performed on the filtered candidate relation pairs to identify the subject-verb-object relationship and the adverbial-head relationship of the entity pairs in the sentence structure. The entity association relationship annotation results are determined by combining the relation type system of the target business domain. By integrating the entity boundary annotation results and the inter-entity relationship annotation results, an internal index of the annotation results is established through the entity unique identifier. Priority determination is performed on the annotation items that conflict between entity boundaries and relationships. The first conflict item in the entity boundary annotation results with a confidence level higher than that in the inter-entity relationship annotation results is retained, and the second conflict item in the inter-entity relationship annotation results with a confidence level higher than that in the entity boundary annotation results is subjected to secondary entity boundary verification. An initial annotation result containing entity unique identifier, entity type label, entity context description, and relationship type triplet is generated.

3. The method as described in claim 1, characterized in that, The domain-adaptive relabeling sub-model in the transfer learning strategy is invoked to perform reinforcement learning-based labeling processing on the candidate relabeled sample set. This process involves adjusting and reducing the overall uncertainty evaluation index of the samples through multiple rounds of labeling strategies, including: Initialize the reinforcement learning parameters of the domain adaptive relabeling sub-model, and set the maximum number of iterations for labeling strategy adjustment and the uncertainty reduction threshold; In each iteration, a batch of samples is randomly selected from the candidate re-labeled sample set, and the domain adaptation model is called to generate temporary labeling results for the current batch of samples; Calculate the difference between the temporary annotation results and the historical annotation results of the batch of samples, and adjust the parameters of the reinforcement learning reward function based on the magnitude of the difference; The labeling decision threshold of the domain adaptation model is updated based on the output value of the reward function, so that the domain adaptation model adopts enhanced boundary judgment conditions in the region of uncertainty. When the reduction in the comprehensive uncertainty assessment index for X consecutive iterations is less than the preset magnitude, the reinforcement learning annotation process is stopped, and the optimized annotation sub-results for the current batch of samples are output, where X is a positive integer.

4. The method as described in claim 1, characterized in that, The step of pushing the optimized annotation results to the business service interaction interface connected to the online business log database, and if an interactive response operation is detected on the business service interaction interface within a preset time period after the push, updating the learning parameters and annotation strategy of the active learning strategy according to the interactive response operation, includes: The optimized annotation results are categorized and organized into structured display data according to business entity type and interaction scenario, and pushed to the front-end rendering engine of the business service interaction interface through an asynchronous data transmission protocol. Start the interaction response monitoring thread to capture the user interaction behavior sequence generated by the business service interaction interface within a preset time period in real time. The user interaction behavior sequence includes mouse click trajectory, text input content and interface element selection operation. The semantic intent parsing process is performed on the user interaction behavior sequence. The intent recognition model in the transfer learning strategy is used to identify the annotation correction intent, entity supplementation intent or relationship adjustment intent corresponding to the interaction behavior, and generate a structured interaction intent description vector. The sample selection logic and model adaptation rules of the active learning strategy are adjusted based on the interaction intent description vector; wherein, the sample selection logic prioritizes the selection of labeled samples in the same interaction scenario according to the intent type, and the model adaptation rules are implemented by dynamically adjusting the feature weight allocation strategy of the domain adaptation model. The adjusted sample selection logic and model adaptation rules are integrated into an updated annotation strategy, and the learning parameter storage sub-model of the active learning strategy is updated synchronously. The subsequent annotation process adopts the updated sample selection logic and model adaptation rules.

5. The method as described in claim 4, characterized in that, The step of performing semantic intent parsing processing on the user interaction behavior sequence involves identifying the annotation correction intent, entity supplementation intent, or relationship adjustment intent corresponding to the interaction behavior through the intent recognition model in the transfer learning strategy, and generating a structured interaction intent description vector, including: Extract key interaction events from the user interaction behavior sequence. The key interaction events include text input completion events, entity region selection events, and relationship line dragging events. Input the event type, trigger timestamp, and associated UI element identifiers of key interactive events into the event encoding layer of the intent recognition model to generate an event feature sequence; The transfer feature adapter of the intent recognition model is invoked to map the event feature sequence to a feature representation aligned with the semantic space of the target business domain. The transfer feature adapter achieves semantic alignment by minimizing the feature distribution difference between the source domain and the target domain. The intent classification layer of the intent recognition model performs multi-class probability prediction on the transferred feature representation and outputs the probability distributions corresponding to the labeled correction intent, entity supplementation intent, and relationship adjustment intent. The intent type corresponding to the highest probability value in the probability distribution is selected as the dominant interaction intent. An interaction intent description vector containing intent type label, target entity ID and intent triggering scenario description is generated by combining the interface element identifier in the event feature sequence.

6. The method as described in claim 4, characterized in that, The sample selection logic and model adaptation rules for adjusting the active learning strategy based on the interaction intent description vector include: the sample selection logic prioritizes labeled samples from similar interaction scenarios based on intent type; and the model adaptation rules are implemented by dynamically adjusting the feature weight allocation strategy of the domain-adaptive model, including: Initialize the strategy adjustment iteration counter, set the maximum number of iterations and the intent matching threshold, the intent matching threshold is used to determine whether the currently adjusted strategy matches the interaction intent description vector; During each iteration, the historical interaction strategy library is retrieved based on the intent type label in the interaction intent description vector to obtain the sample selection template and model adaptation template corresponding to the same type of intent. The similarity between the scene features in the sample selection template and the intent trigger scene description in the current interaction intent description vector is calculated, and scene features with similarity higher than the preset similarity are selected as key matching features. The scene weight coefficients of the sample selection logic are adjusted based on key matching features, so that labeled samples in the same interaction scenario can obtain higher priority weights in the selection process. The feature weight allocation strategy in the model adaptation rules is adjusted synchronously, and higher weight values ​​are assigned to entity features and relationship features associated with key matching features, while reducing the weight influence of non-associated features. Calculate the matching degree between the adjusted sample selection logic and model adaptation rules and the interaction intent description vector. If the matching degree is higher than the intent matching threshold or the maximum number of iterations is reached, stop the iteration and output the adjusted sample selection logic and model adaptation rules. The adjusted sample selection logic and model adaptation rules are integrated into an updated annotation strategy, and the learning parameter storage sub-model of the active learning strategy is updated synchronously. The subsequent annotation process uses the updated sample selection logic and model adaptation rules, including: Construct a validation set for updating the annotation strategy, and extract historical business datasets that are consistent with the interaction intent description vector scenario from the online business log database as validation samples; The adjusted sample selection logic is used to select a subset of labeled samples from the validation samples. The domain adaptation model is then invoked to perform labeling processing using the updated model adaptation rules, and validation labeling results are generated. Calculate the F1 score of the verification annotation results and the manual annotation benchmark. If the F1 score is higher than the preset performance threshold, the adjusted sample selection logic and model adaptation rules are deemed to be effective. When the determination is valid, the scenario priority configuration of the sample selection logic, the feature weight allocation strategy of the model adaptation rule, and the intent type association rule are packaged into a strategy update package. The policy update package is written into the learning parameter storage sub-model of the active learning policy through atomic transaction operations, overriding the original sample selection logic and model adaptation rules, and synchronously recording the policy update timestamp and associated interaction intent description vector; the updated policy configuration is loaded in the subsequent annotation process.

7. An online business knowledge graph annotation system, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the online business knowledge graph annotation method based on transfer learning as described in any one of claims 1-6.

8. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the online business knowledge graph annotation method based on transfer learning as described in any one of claims 1-6.

Images