An author name disambiguation model construction method and application and electronic device

Abstract

This invention provides a method, application, and electronic device for constructing an author name disambiguation model. First, a name block and an initial training set consisting of entity pair training samples are constructed. An author name disambiguation baseline model is trained based on the initial training set. The baseline model is then used to calculate the global feature importance vector for each entity pair training sample through feature attribution calculation. The score for each entity pair training sample is calculated based on the global feature importance vector and a non-empty indicator vector, and a training set is selected. The baseline model is then retrained using the selected training set to obtain the author name disambiguation model. The objective loss function of the author name disambiguation model includes classification loss and attribution prior regularization, improving model interpretability, reducing information loss due to metadata sparsity, suppressing model attention drift, and being compatible with various model structures. It has low engineering modification costs and can be widely applied to name disambiguation scenarios in various academic data processing systems.

Description

Technical Field

[0001] This invention relates to the fields of literature data processing, machine learning, information retrieval and academic knowledge services, and specifically to a method for constructing an author name disambiguation model, its application and electronic device. Background Technology

[0002] Author names are a fundamental aspect of academic data processing. The core of this process is to distinguish between different author entities with the same or different names in the literature, eliminate identity confusion caused by name ambiguity, and thus improve the accuracy of literature retrieval, results statistics, and other tasks.

[0003] Currently, mainstream author name disambiguation methods primarily extract textual and attribute features such as paper titles, author affiliations, journal / conference names, publication years, co-authors, and references. These features are then combined with similarity calculations, traditional classification, cluster analysis, or deep learning models to identify authors with the same name. For example, the Chinese patent application CN202011536139.X utilizes semantic and relational features of the paper text, improving disambiguation accuracy and handling massive amounts of text, demonstrating a certain level of advancement. However, this approach mainly achieves disambiguation through feature learning, deep learning, and cluster analysis, without involving feature attribution analysis of the trained model or using the model's interpretation results to guide training sample selection or impose prior constraints on the model training process. Therefore, when significant amounts of document metadata are missing, the efficiency of training sample utilization is low, and the key discrimination criteria within the model are difficult to reveal, resulting in insufficient model interpretability. Specifically, this manifests in the following ways: 1. Insufficient model interpretability: Most existing author name disambiguation models are black box models, which can only output disambiguation results but lack feature attribution analysis capabilities. This makes it impossible to quantify the contribution of metadata features such as author affiliation, journal, co-authors, and references to the discrimination results. The model's decision-making logic is not transparent and the discrimination criteria are difficult to reveal, resulting in high difficulty in model debugging, optimization, and problem tracing, and low debuggability and reliability.

[0004] 2. Sparse metadata leads to serious information loss: In actual academic data, the metadata of literature generally suffers from the problem of missing fields (such as some papers without author affiliations, old documents without complete references, and some records missing journal / conference information). Existing technologies have not designed optimization strategies for sparse metadata scenarios, resulting in insufficient utilization of key features. The mapping and processing of missing fields can easily cause the loss of effective information, which directly limits the upper limit of disambiguation performance.

[0005] In summary, existing author name disambiguation techniques generally lack the ability to explain the features attribution process of the model discrimination process, and fail to use the explanation results for training sample optimization and training constraints. In real-world scenarios where literature metadata is sparse and features are missing, there are problems such as insufficient utilization of key features, poor model interpretability, and limited prediction performance.

[0006] Therefore, there is an urgent need for a method that can incorporate feature attribution results into the training and prediction process of author name disambiguation models, so as to improve training stability and prediction accuracy while enhancing model interpretability. Summary of the Invention

[0007] The purpose of this invention is to provide a method, application, and electronic device for constructing an author name disambiguation model. The author name disambiguation model is trained in two stages, and the feature attribution results are transformed into effective constraints and sample selection criteria in the training process, thereby improving the transparency of the author name disambiguation model while improving training efficiency and disambiguation performance.

[0008] To achieve the above objectives, this technical solution provides a method for constructing an author name disambiguation model, including the following steps: S1: Obtain author name disambiguation data containing multiple name blocks, construct entity pair training samples within the same name block, and construct the initial training set using entity pair training samples labeled with author tags, where each entity pair training sample records two paper entities in the same name block. S2: Train the basic model based on the initial training set to obtain the author name disambiguation baseline model. Use the author name disambiguation baseline model to perform feature attribution calculation on each entity for the training samples to obtain the feature contribution of each entity to each feature in the training samples. Calculate the global feature importance vector based on the feature contribution of all entities to the features in the samples. S3: Sort the features according to the global feature importance vector to obtain the feature ranking weights. Construct a non-empty indicator vector for each entity pair training sample in the initial training dataset. Calculate the score of each entity pair training sample using the feature ranking weights and the non-empty indicator vectors. Select a preset number of entity pairs training samples from high to low scores to form the selection training set. S4: Retrain the author name disambiguation baseline model based on the selected training set to obtain the author name disambiguation model, where the objective loss function of the author name disambiguation model includes classification loss and attribution prior regularization term.

[0009] Secondly, this solution provides an application method for an author name disambiguation model, including the following steps: Obtain at least two papers to be disambiguated, construct name blocks based on the authors' names, and form entity pair samples from two papers belonging to the same name block, then extract the similarity feature vector of the entity pair samples. The similarity feature vector of each entity pair sample is input into the author name disambiguation model to output the entity pair prediction probability, where the entity pair prediction probability represents the probability that the current entity pair belongs to the same author; Disambiguation of authors in academic papers is performed based on entity pair prediction probabilities.

[0010] Thirdly, this solution provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the above-described method for constructing the author name disambiguation model, or the above-described method for applying the author name disambiguation model.

[0011] Compared with existing technologies, this technical solution has the following characteristics and beneficial effects: 1. This invention can provide feature-level interpretation of the author name disambiguation model, giving the contribution of metadata features such as author affiliation, title, journal, conference, co-authors, and references to the discrimination results, which helps to reveal the basis of the model's decision-making and significantly improves the interpretability and debuggability of the model.

[0012] 2. This invention no longer uses feature attribution results only for post-hoc analysis, but directly introduces the interpretation results into the training process, realizing the transformation from "interpretable" to "optimizable", and expanding the application depth of feature attribution technology in author name disambiguation tasks.

[0013] 3. This invention can prioritize the selection of training samples with non-empty key features based on the global feature importance vector, thereby increasing the information density of training samples under the same training budget and reducing information loss caused by metadata sparsity, field missing and NULL mapping, thus improving sample utilization efficiency.

[0014] 4. This invention constrains the author name disambiguation model to maintain stable attention to key features during retraining by using attribution prior regularization terms. This can suppress attention drift caused by noisy features or sparse fields during training, thereby improving training stability and generalization performance.

[0015] 5. This invention does not rely on a specific author name disambiguation model structure and is applicable to neural network models, tree models, and other supervised learning models capable of feature attribution. Therefore, it has good model compatibility and engineering implementation capability.

[0016] 6. This invention can be integrated into existing author name disambiguation systems without changing the existing feature engineering process, resulting in low engineering modification costs and easy deployment to academic databases, institutional knowledge bases, and scientific research management and knowledge service platforms. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the method for constructing an author name disambiguation model according to an embodiment of this application; Figure 2 This is a flowchart of an application method for an author name disambiguation model according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0019] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0020] Example 1: Method for constructing the author name disambiguation model: This solution specifically addresses the shortcomings or problems in existing author name disambiguation techniques, such as insufficient model interpretability, lack of explanatory guidance in training sample selection, information loss due to sparse metadata and NULL value mapping, and insufficient model training stability. It transforms feature attribution results into effective constraints and sample selection criteria in the training process, thereby improving training efficiency and disambiguation performance while increasing model transparency.

[0021] Specifically, such as Figure 1As shown, the method for constructing the author name disambiguation model provided in this solution includes the following steps: S1: Obtain author name disambiguation data containing multiple name blocks, construct entity pair training samples within the same name block, and construct the initial training set using entity pair training samples labeled with author tags, where each entity pair training sample records two paper entities in the same name block. S2: Train the basic model based on the initial training set to obtain the author name disambiguation baseline model. Use the author name disambiguation baseline model to perform feature attribution calculation on each entity for the training samples to obtain the feature contribution of each entity to each feature in the training samples. Calculate the global feature importance vector based on the feature contribution of all entities to the features in the samples. S3: Sort the features according to the global feature importance vector to obtain the feature ranking weights. Construct a non-empty indicator vector for each entity pair training sample in the initial training dataset. Calculate the score of each entity pair training sample using the feature ranking weights and the non-empty indicator vectors. Select a preset number of entity pairs training samples from high to low scores to form the selection training set. S4: Retrain the author name disambiguation baseline model based on the selected training set to obtain the author name disambiguation model, where the objective loss function of the author name disambiguation model includes classification loss and attribution prior regularization term.

[0022] This solution employs an innovative two-step training architecture. First, an author name disambiguation baseline model is trained under unconstrained conditions. Then, SHAP feature attribution is used to mine the true global feature importance of the data, thereby addressing the issues of model black box and unclear key features. Next, based on the feature attribution results, high-value samples with non-empty key features are selected, and attribution prior regularization terms are introduced to retrain the author name disambiguation baseline model. This not only improves the sample utilization rate in the case of sparse metadata, but also suppresses feature attention drift during training, ultimately simultaneously improving model interpretability, disambiguation accuracy, and training stability.

[0023] Regarding step S1: In the section "Obtaining Author Name Disambiguation Data Containing Multiple Name Blocks," a name block refers to the set of all paper entities corresponding to the same author with the same name string or a similar name string after normalization. That is, each name block records the set of all paper entities of the same author who have the same name string or a similar name string.

[0024] In practical applications, author name disambiguation data comes from academic search platforms, institutional knowledge bases, scientific research management systems, paper recommendation systems, bibliometric analysis systems, expert profiling systems, academic knowledge graph construction systems, publisher metadata databases, citation databases, or open academic data platforms. Each paper record typically includes at least one or more of the following: paper title, author list, author affiliation, journal or conference name, publication year, abstract, keywords, and reference information.

[0025] Since the same name string may correspond to different real authors in different papers, the obtained paper records first need to be coarsely grouped according to the name string to form name blocks. For example, all paper records corresponding to name strings such as "J. Wang", "WeiZhang", and "Z. Fang" can be organized into their respective name blocks.

[0026] In some embodiments, author name disambiguation data is represented as follows: ; Where b i Let represent the i-th name block, 1≤i≤M, where M is the total number of name blocks.

[0027] Furthermore, the paper entity set consists of at least one paper entity, and each paper entity includes at least one paper attribute feature, wherein the paper attribute feature includes at least one feature representing a paper.

[0028] Furthermore, the paper's attributes include title, author affiliation, co-authors, journal, conference, reference authors, reference titles, reference journals, and year.

[0029] In some embodiments, the paper entity is represented as: ; in For the j-th paper entity in the i-th name block, Let d be the attribute feature of the paper.

[0030] In the step of “constructing entity pair training samples within the same name block”, any two paper entities in the same name block are selected to form entity pair training samples. That is, the entity pair training samples are two different paper entities corresponding to the same name block.

[0031] In the step of “constructing an initial training set from entity pairs labeled with author tags”, the similarity feature vector of each entity pair training sample is extracted, and the similarity feature vectors of all entity pairs training samples and the corresponding author tags are summarized as the initial training set.

[0032] It should be noted that although the same entity pair corresponds to the same name block in the training samples, they do not necessarily belong to the same author (there may be cases where authors share the same name). Therefore, each entity pair in the training samples still needs to be manually labeled. This scheme uses binary labels to define author labels. If the two paper entities in the current entity pair in the training samples belong to the same author, the author label of the current entity pair in the training samples is labeled as 1; if the two paper entities in the current entity pair in the training samples do not belong to the same author, the author label of the current entity pair in the training samples is labeled as 0. Through the above construction, the name disambiguation problem can be transformed into a supervised classification task of "given two paper entities with the same name, determine whether they belong to the same real author".

[0033] In other words, the training samples consist of entity pairs formed by the j-th paper entity and the k-th paper entity corresponding to the i-th name block. , ), based on the author tag, define a binary category tag y, when = When, the author tag is set to 1; when ≠ At that time, the author tag was set to 0, where This represents the author tag for the j-th paper entity corresponding to the i-th name block. This represents the author tag for the k-th paper entity corresponding to the i-th name block.

[0034] In "Extracting the similarity feature vector of each entity to the training samples", the similarity of each dimension of the features of each entity to the two paper entities in the training samples is calculated to obtain the feature similarity. The feature similarity of all dimensions is summarized as the similarity feature vector of the entity to the training samples.

[0035] Furthermore, in order to form numerical feature vectors of uniform length, this scheme performs different feature similarity calculations for different types of features: For textual attribute features (such as author affiliation, title, journal, conference, co-author, reference author, reference title, reference journal, etc.), feature similarity is calculated using the Jaccard coefficient on the character n-gram set, where the character n-gram can be selected from 2-gram to 4-gram, and the Jaccard coefficient is defined as the ratio of the intersection size to the union size. For the feature representing the full name of co-authors, the Jaccard similarity of the full name set is used to calculate the feature similarity; For features representing the year field, the absolute difference is used to calculate feature similarity, and an upper threshold is set to indicate the temporal proximity of two paper entities.

[0036] As above, for entity pairs training samples ( , The formula for extracting the similarity feature vector of each entity to the training samples is expressed as: ; in This represents the feature similarity of the d-th dimension.

[0037] Therefore, the initial training set is represented as: ; in Let be the similarity feature vector of the nth entity to the training samples. This represents the author label of the nth entity pair in the training samples, where N represents the total number of entity pairs in the training samples.

[0038] It should be noted that missing fields frequently occur in the author name disambiguation data. For example, some papers lack author affiliation information, some older documents lack complete reference fields, and some records are missing conference or journal names. For missing fields, the relevant similarity features cannot be directly calculated; therefore, in this implementation, they are marked as NULL. During subsequent model training, NULL can be mapped to 0 or other default values, but this mapping does not indicate low similarity of the feature itself, but only that the feature is unobservable. Therefore, this invention will subsequently utilize the interpretation results to perform targeted screening of samples with missing key features.

[0039] Regarding step S2: Step S2 trains an author name disambiguation baseline model on the initial training set, then uses the feature attribution method to calculate the feature contribution of each sample and aggregates them into a global feature importance vector. This quantifies the influence of features such as author affiliation, journal, and co-authors on the disambiguation results, effectively solving the model black box problem and clarifying the key discrimination criteria. At the same time, it provides a reliable basis for subsequent high-value sample screening and model retraining attribution regularization constraints, fundamentally improving the defects of insufficient model interpretability and unclear key features.

[0040] In the step of “training the base model based on the initial training set to obtain the author name disambiguation baseline model”, the base model is selected from any one of the following: single hidden layer neural network, deep multilayer perceptron, gradient boosting tree, graph neural network, or deep model based on attention mechanism.

[0041] In addition, the target loss function of the author-name disambiguation baseline model during training is expressed as follows: ; in Let N represent the base model, N represent the total number of entity pairs in the initial training set, and L represent the classification loss function. Let be the similarity feature vector of the nth entity to the training samples. This represents the author label for the nth entity in the training samples. This represents the probability that they belong to the same author.

[0042] Furthermore, this scheme trains a basic model based on the initial training set to obtain an author name disambiguation baseline model. This author name disambiguation baseline model has basic discriminative ability. The author name disambiguation baseline model can take similarity feature vectors as input and input the probability of whether they belong to the same author.

[0043] In the step of "using the author name disambiguation baseline model to calculate the feature attribution of each entity pair to the training samples and obtain the feature contribution of each entity to each feature in the training samples," each entity pair to the training samples is input into the author name disambiguation baseline model, and the feature attribution algorithm is used to calculate the feature contribution of each feature to the prediction result of the author name disambiguation baseline model. This scheme identifies the key features that the author name disambiguation baseline model truly relies on in this way.

[0044] Furthermore, if the feature contribution value of a certain feature is positive, it means that the feature pushes the author name disambiguation baseline model toward predicting "the same author", otherwise it is the opposite. The larger the absolute value of the feature contribution value, the more significant the influence of the current feature on the prediction result.

[0045] In some embodiments, the feature attribution algorithm employs the SHAP algorithm. SHAP originates from the Shapley value concept in cooperative game theory, which can assign a relatively fair and consistent contribution value to each input feature. If the base model of the author name disambiguation baseline model is a tree model, then the feature attribution algorithm can use TreeSHAP; if the base model of the author name disambiguation baseline model is a neural network model, then the feature attribution algorithm can use DeepSHAP.

[0046] To enhance the comparison between samples from the same author and samples from different authors, this invention preferably uses entity pairs with opposite author labels as training samples as background samples. For example, when interpreting positive samples, negative samples are used as background samples; when interpreting negative samples, positive samples are used as background samples. This better highlights the key metadata features that truly distinguish between the two types of samples.

[0047] Furthermore, the attribution results of a single training sample are insufficient to directly guide the training process; therefore, it is necessary to aggregate the feature attribution vectors of all training samples. In the step of "calculating the global feature importance vector based on the feature contribution of all entities to the features in the sample," the average of the absolute values ​​of the feature contribution of each feature to all entities is taken as the global importance of the current feature, and the global importance vector is obtained by summing the global importance of all features.

[0048] The formula for calculating global importance is as follows: ; in To represent the nth entity pair of training samples The The feature contribution is calculated from each feature, where N is the total number of entity pairs in the training samples. For the first Global importance of each feature.

[0049] The global feature importance vector, obtained by summing the global importance of all features, is represented as follows: ; in Let d be the global importance of the d-th feature. This global feature importance vector reflects the degree to which each feature is most critical to the author name disambiguation baseline model prediction across the entire initial training set.

[0050] In a practical experiment, author affiliations, journals, full names of co-authors, and reference journals often have high importance. This aligns with the intuition of author name disambiguation, which suggests that the same author typically maintains relatively stable affiliation information, collaborations, and publication channels over a period of time. However, the metadata quality varies across different datasets, potentially altering the ranking of important features. For instance, if the affiliation field is generally missing in a dataset, the importance of title or reference features may increase. This invention leverages these automatically revealed differences in importance from the interpretation results to further guide the selection of training sets and the construction of attribution priors for subsequent secondary training.

[0051] Regarding step S3 of this plan: Since there are a large number of samples in the large-scale author name disambiguation data, and some of the key features of these samples may be missing, this invention uses attribution-guided training sample selection based on the global feature importance vector to obtain a more valuable selection training set.

[0052] In the step of "sorting features according to the global feature importance vector to obtain feature ranking weights", the features in the global feature importance vector are sorted in descending order to obtain feature ranking weights, where the feature with higher importance has a higher feature ranking weight.

[0053] In the step of “constructing a non-empty indicator vector for each entity pair training sample in the initial training dataset”, the non-empty indicator vector is used to indicate whether the current entity pair has an observable value for each feature in the training sample. If the current entity pair has an observable value for the training sample on that feature, the non-empty indicator vector of the current feature is defined as 1; if the current entity pair does not have an observable value for the training sample on that feature, the non-empty indicator vector of the current feature is defined as 0.

[0054] In the step of “calculating the score of each entity pair training sample using feature ranking weights and non-empty indicator vectors”, for each entity pair training sample, the feature ranking weight of each feature is calculated as the product of the value of the non-empty indicator vector as the feature score, and the feature scores of all features are summed to obtain the score of the current entity pair training sample.

[0055] The formula for calculating the score of an entity relative to the training samples is as follows: ; in It is the first Feature ranking weights for each feature For the first The value of feature in the non-empty indicator vector of the nth entity pair training sample, where d is the total number of features. For the nth entity pair training sample, This represents the score of the entity against the training samples.

[0056] It should be noted that this score essentially measures whether the entity pair training sample is complete in terms of important feature dimensions. If an entity pair training sample is not empty in several high-importance features, its score is high, indicating that the entity pair training sample is more valuable for learning key discrimination rules. If an entity pair training sample is large in number, but most of its high-importance features are empty, its score is low. Therefore, this invention further sorts all entity pair training sample scores from high to low based on the entity pair training sample scores, and selects the top K entity pair training samples from 1 to d entity pair training samples as the selection training set. Through the above process, the information density of the training samples can be significantly improved under a fixed training budget.

[0057] Compared to random sampling, the training set selection method of this invention prioritizes entity pairs with complete fields for the most important features. For example, if the attribution results indicate that the author's affiliation and co-authors are the two most critical features, then entity pairs with both affiliation and co-author non-null values ​​will be prioritized for inclusion in the training set selection. If some samples have many important fields but are missing key fields, their ranking will be lowered. Through this mechanism, the interpretation results are directly transformed into a training sample selection strategy for the training set selection.

[0058] Regarding step S4 of this plan: This step utilizes a selected training set to retrain the author name disambiguation baseline model, thereby obtaining the author name disambiguation model. Unlike traditional methods that rely solely on supervisory labels, this invention further incorporates an attribution prior regularization term into the objective loss function of the author name disambiguation baseline model to restrict the global feature attention pattern of the baseline model from deviating from the effective structure revealed by the baseline model.

[0059] Furthermore, the objective loss function of the author name disambiguation model is expressed as: ; ; in Here, N represents the total number of samples in the training set, and L represents the classification loss function. This indicates that the nth entity pair in the training set is selected as the training sample, y n This indicates selecting the author label for the nth entity pair in the training set from the training samples. For regular strength control parameters, This represents the probability that the characters belong to the same author. This indicates the attribution prior regularity term. This indicates the author name disambiguation model in the current training state. The feature contribution of each feature This indicates the author name disambiguation model in the current training state. The global importance of each feature, where d is the total number of features.

[0060] Furthermore, the author name disambiguation model is used to perform feature attribution calculations on each entity in the selected training set for each feature in the training samples, obtaining the feature contribution of each entity to each feature in the training samples. Based on the feature contributions of all entities to the features in the samples, the global importance of each feature is calculated. Based on the global importance and feature contribution, an attribution prior regularization term is constructed. The methods for calculating feature contribution and global importance are the same as before, so they will not be elaborated here.

[0061] It should be noted that if the author name disambiguation prediction in a certain training phase focuses excessively on noisy features or sparse fields, then... and The deviation between them will increase, thus increasing the attribution prior regularization term. The attribution prior regularization term increases and penalizes the model; if the author name disambiguation model continues to maintain a stable focus on key features, then the attribution prior regularization term... It is relatively small and has a weak impact on training.

[0062] This approach introduces classification loss and attribution prior regularization terms into the objective loss function of the author name disambiguation model. This requires the model to fit the training labels as closely as possible while maintaining a reasonable feature focus pattern. Even if the selected training set has a slightly different distribution from the initial training set, the author name disambiguation model will not deviate from the key discrimination rules revealed in the baseline stage. The attribution prior essentially encodes the global domain knowledge extracted from the interpretation results back into the training process, stabilizing training, reducing overfitting, and suppressing noise bias.

[0063] Example 2: Application of the Author Name Disambiguation Model: like Figure 2 As shown, this solution provides an application method for an author name disambiguation model, that is, this solution provides an author name disambiguation method, including the following steps: Obtain at least two papers to be disambiguated, construct name blocks based on the authors' names, and form entity pair samples from two papers belonging to the same name block, then extract the similarity feature vector of the entity pair samples. The similarity feature vector of each entity pair sample is input into the author name disambiguation model to output the entity pair prediction probability, where the entity pair prediction probability represents the probability that the current entity pair belongs to the same author; Disambiguation of authors in academic papers is performed based on entity pair prediction probabilities.

[0064] This second embodiment uses the author name disambiguation model constructed in the first embodiment, and other identical technical features will not be described again.

[0065] In the step of "Constructing name blocks based on author names in academic papers", the name strings of author names in academic papers are normalized, and academic papers with the same name string or similar name strings are grouped into the same name block.

[0066] In the step of "extracting the similarity feature vector of entity pair samples", the similarity of each dimension of the features of the two paper entities in the training samples of each entity pair is calculated to obtain the feature similarity. The feature similarity of all dimensions is summarized as the similarity feature vector of the entity pair.

[0067] In the step of "disambiguating the authors of papers based on the predicted probability of entity pairs", an entity relationship graph is constructed based on the predicted probability of entity pairs. The author entities within the name block are merged through threshold segmentation, connected component partitioning, hierarchical clustering, graph clustering or a combination thereof, to obtain the author name disambiguation result.

[0068] Furthermore, after obtaining the predicted probability of entity pairs, this scheme constructs edge relationships based on a preset threshold to form an entity graph within the name block. If the entity pair probability of two entity pair samples is greater than the preset threshold, an edge is connected between the corresponding nodes. Subsequently, connected component partitioning, hierarchical clustering, graph clustering, or other aggregation strategies can be used to group entities belonging to the same real author within the name block into a cluster. Finally, each cluster corresponds to a disambiguated author entity.

[0069] In practical systems, author name-based disambiguation results can be used for various upper-level operations. For example, in academic search platforms, papers belonging to the same author can be aggregated into a unified author homepage; in research management systems, the achievements of teachers or researchers can be more accurately counted; in academic knowledge graph construction systems, erroneous merging of different author entities and duplicate representations of the same author entity can be reduced; in collaborative network analysis, the authenticity of collaborative edge relationships can be improved; and in literature recommendation and expert discovery, recommendation bias caused by author identity confusion can be avoided.

[0070] In other words, the author name disambiguation model proposed in this scheme can be applied to author name disambiguation scenarios in academic search platforms, institutional knowledge bases, scientific research management systems, paper recommendation systems, bibliometric analysis systems, expert profiling systems, or academic knowledge graph construction systems. In some embodiments, the base model is selected as a single-hidden-layer neural network, and experiments are conducted using tests identical to those on the initial training set. Experimental results show that, after adopting this method, the model's accuracy, AUROC, and F1 scores on multiple datasets are all improved compared to the baseline author name disambiguation prediction model, indicating that the attribution-guided training process is effective.

[0071] In other embodiments, the present invention uses gradient boosting trees as the base model. Since tree models have a strong ability to handle sparse and missing values, the benefits of training set selection on certain datasets may be less than those of neural network models. However, the attribution results can still be used to explain the importance of different features and can serve as the basis for subsequent model comparison analysis. The global importance vector obtained using TreeSHAP can also participate in the selection of training sets and can also be used to construct attribution prior regularization terms. This embodiment further illustrates that the present invention is not limited to a specific model structure and has strong generalization applicability.

[0072] Furthermore, in real-world engineering environments, limitations often exist in training time, GPU memory, CPU resources, and attribution computation costs, making it impossible to use the entire initial training set for training. To address this, this invention preferably operates under a fixed training budget, for example, limiting the number of training samples in each initial training set to K. Under this condition, comparing random sampling, active learning sampling, traditional importance sampling, and the attribution-guided sampling of this invention can more fairly reflect the differences in the effectiveness of different training set selection strategies. Experiments show that, with the same K value, the attribution-guided sample selection preferred in this invention achieves higher information density and better model performance, demonstrating that the explanatory signal can indeed be transformed into more efficient sample utilization.

[0073] In summary, this invention, focusing on the author name disambiguation task, proposes a complete method chain encompassing baseline training and interpretation, attribution-guided sample selection, attribution prior retraining, and inference clustering output. This forms a closed-loop process from model interpretation to training optimization and then to predictive output. Compared to existing technologies, this invention not only clearly reveals the model's discrimination logic and answers the intrinsic basis of model decisions, but also deeply integrates the interpretation results of feature attribution into the entire model training process. This achieves a substantial breakthrough from "model interpretability" to "training optimization," effectively transforming interpretive value into practical utility for improving model performance. It possesses both outstanding theoretical innovation value and strong engineering application value.

[0074] Example 3 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, the memory 404 storing a computer program, and the processor 402 being configured to run the computer program to perform the steps in any of the above embodiments of the method for constructing or applying the author name disambiguation model.

[0075] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0076] The memory 404 may include a large-capacity memory 404 for data or instructions. The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0077] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the author name disambiguation model construction methods or author name disambiguation model application methods in the above embodiments.

[0078] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0079] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0080] Input / output device 408 is used to input or output information. In this embodiment, the input information may be author name disambiguation data, paper abstracts, etc., and the output information may be entity pair prediction probabilities, etc.

[0081] Optionally, in this embodiment, the processor 402 can be configured to execute the following steps of the author name disambiguation model construction method via a computer program: S1: Obtain author name disambiguation data containing multiple name blocks, construct entity pair training samples within the same name block, and construct the initial training set using entity pair training samples labeled with author tags, where each entity pair training sample records two paper entities in the same name block. S2: Train the basic model based on the initial training set to obtain the author name disambiguation baseline model. Use the author name disambiguation baseline model to perform feature attribution calculation on each entity for the training samples to obtain the feature contribution of each entity to each feature in the training samples. Calculate the global feature importance vector based on the feature contribution of all entities to the features in the samples. S3: Sort the features according to the global feature importance vector to obtain the feature ranking weights. Construct a non-empty indicator vector for each entity pair training sample in the initial training dataset. Calculate the score of each entity pair training sample using the feature ranking weights and the non-empty indicator vectors. Select a preset number of entity pairs training samples from high to low scores to form the selection training set. S4: Retrain the author name disambiguation baseline model based on the selected training set to obtain the author name disambiguation model, where the objective loss function of the author name disambiguation model includes classification loss and attribution prior regularization term.

[0082] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0083] Alternatively, the steps for applying the author name disambiguation model are as follows: Obtain at least two papers to be disambiguated, construct name blocks based on the authors' names, and form entity pair samples from two papers belonging to the same name block, then extract the similarity feature vector of the entity pair samples. The similarity feature vector of each entity pair sample is input into the author name disambiguation model to output the entity pair prediction probability, where the entity pair prediction probability represents the probability that the current entity pair belongs to the same author; Disambiguation of authors in academic papers is performed based on entity pair prediction probabilities.

[0084] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0085] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0086] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing an author name disambiguation model, characterized in that, Includes the following steps: S1: Obtain author name disambiguation data containing multiple name blocks, construct entity pair training samples within the same name block, and construct the initial training set using entity pair training samples labeled with author tags, where each entity pair training sample records two paper entities in the same name block. S2: Train the basic model based on the initial training set to obtain the author name disambiguation baseline model. Use the author name disambiguation baseline model to perform feature attribution calculation on each entity for the training samples to obtain the feature contribution of each entity to each feature in the training samples. Calculate the global feature importance vector based on the feature contribution of all entities to the features in the samples. S3: Sort the features according to the global feature importance vector to obtain the feature ranking weights. Construct a non-empty indicator vector for each entity pair training sample in the initial training dataset. Calculate the score of each entity pair training sample using the feature ranking weights and the non-empty indicator vectors. Select a preset number of entity pairs training samples from high to low scores to form the selection training set. S4: Retrain the author name disambiguation baseline model based on the selected training set to obtain the author name disambiguation model, where the objective loss function of the author name disambiguation model includes classification loss and attribution prior regularization term.

2. The method for constructing the author name disambiguation model according to claim 1, characterized in that, The name block refers to the set of all paper entities corresponding to the same name string or whose name strings are similar after normalization. The non-empty indicator vector is used to indicate whether the current entity has an observable value for each feature in the training sample. If the current entity has an observable value for the training sample on that feature, the non-empty indicator vector of the current feature is defined as 1; if the current entity does not have an observable value for the training sample on that feature, the non-empty indicator vector of the current feature is defined as 0.

3. The method for constructing the author name disambiguation model according to claim 1, characterized in that, Take any two paper entities from the same name block to form an entity pair training sample. Each paper entity includes at least one paper attribute feature, where the paper attribute feature includes at least one feature that represents a paper.

4. The method for constructing the author name disambiguation model according to claim 1, characterized in that, Extract the similarity feature vector of each entity to the training samples, and summarize the similarity feature vectors of all entities to the training samples and the corresponding author labels as the initial training set.

5. The method for constructing the author name disambiguation model according to claim 1, characterized in that, The global importance of each feature is calculated by taking the average of the absolute values ​​of the feature contributions of all entities to the sample. The global feature importance vector is obtained by summing the global importance of all features.

6. The method for constructing the author name disambiguation model according to claim 1, characterized in that, For each entity pair training sample, the feature score is calculated by multiplying the feature ranking weight of each feature with the value of the non-empty indicator vector. The feature scores of all features are then summed to obtain the score of the current entity pair training sample.

7. The method for constructing the author name disambiguation model according to claim 1, characterized in that, The objective loss function of the author name disambiguation model is expressed as: ； ； in Here, N represents the total number of samples in the training set, and L represents the classification loss function. This indicates that the nth entity pair in the training set is selected as the training sample, y n This indicates selecting the author label for the nth entity pair in the training set from the training samples. For regular strength control parameters, This represents the probability that the characters belong to the same author. This indicates the attribution prior regularity term. This indicates the author name disambiguation model in the current training state. The feature contribution of each feature This indicates the author name disambiguation model in the current training state. The global importance of each feature, where d is the total number of features.

8. The method for constructing the author name disambiguation model according to claim 1, characterized in that, The author name disambiguation model is used to perform feature attribution calculations on each entity in the selected training set for each training sample, and the feature contribution of each entity to each feature in the training sample is obtained. The global importance of each feature is calculated based on the feature contributions of all entities to the features in the sample. Attribution prior regularization terms are constructed based on global importance and feature contribution.

9. A method for applying an author name disambiguation model, characterized in that, Includes the following steps: Obtain at least two papers to be disambiguated, construct name blocks based on the authors' names, and form entity pair samples from two papers belonging to the same name block, then extract the similarity feature vector of the entity pair samples. The similarity feature vector of each entity pair sample is input into the author name disambiguation model to output the entity pair prediction probability, where the entity pair prediction probability represents the probability that the current entity pair belongs to the same author; Disambiguation of authors in academic papers is performed based on the predicted probabilities of entity pairs.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method for constructing the author name disambiguation model according to any one of claims 1 to 8, or the method for applying the author name disambiguation model according to claim 9.

Images