A hierarchical entity alignment method for college student academic activity data

Abstract

The application provides a college student academic activity data hierarchical entity alignment method, and belongs to the fields of artificial intelligence and data processing technology.The application carries out multi-level semantic enhancement entity alignment, context enhancement entity alignment and multi-view similarity fusion unknown entity alignment on the to-be-aligned entity and a standard entity library, obtains an aligned effective alignment pair set, a context enhancement alignment set and a new standard entity set, realizes accurate matching of college student academic data entities, detailed modeling of uncategorized entities and efficient processing of new entities not included in a candidate library, and thus the accuracy and standardization of college student academic activity data entity alignment are improved.The application can avoid unknown entity accumulation, reduce artificial maintenance cost and improve data processing efficiency.

Description

A hierarchical entity alignment method for college students' academic activity data Technical Field

[0001] This invention belongs to the field of artificial intelligence and data processing technology, and in particular relates to a hierarchical entity alignment method for college students' academic activity data. Background Technology

[0002] With the increasing richness of academic activities among modern university students, the amount of data related to these activities is constantly growing. This data can be categorized into two types: static basic information such as student ID, department, major, gender, year of study, and place of origin; and dynamic growth data that continuously increases over time, including competition awards, training experiences, work experience, university activities, personal honors, and certifications. The collection of this data is dynamic and fragmented, often resulting in different expressions for the same event, such as "University Student Innovation and Entrepreneurship Competition" and "University Innovation and Entrepreneurship Competition," leading to issues such as duplicate entity information and semantic ambiguity. This results in subsequent academic analysis and planning recommendations being based on inaccurate facts. Therefore, entity alignment is a key technology for breaking down data silos and ensuring data accuracy. It is an important foundation for constructing knowledge graphs based on static and dynamic multi-source data of university students, enabling data analysis of university student activities and academic growth planning.

[0003] When dealing with university students' academic data, which combines static and dynamic data, existing entity alignment techniques typically use simple labeling methods such as "unclassified" or "other" to handle instances with low confidence or those that fail to match existing entity types. This lacks a further modeling and alignment mechanism for such entities. As these unclassified entities accumulate, relying on manual rules or annotations for subsequent processing not only significantly increases system maintenance costs but also severely restricts the overall efficiency of data processing. Currently, the main technical shortcomings are as follows: First, the growth achievement data entered by university students generally suffers from inconsistent naming and short text lengths, making it difficult for existing entity alignment methods to accurately extract effective entity features. Second, current mainstream entity alignment methods often employ a dual-encoding strategy, encoding both the target alignment entity and candidate entities separately. However, this approach cannot simultaneously utilize the semantic knowledge of the candidate context. Third, as new data is continuously generated, the number of entities not included in the candidate database increases. Furthermore, existing autoregressive entity alignment methods are prone to copy bias, causing the model to incorrectly link entities to high-frequency generalized words, ultimately severely impacting the accuracy and standardization of entity alignment. Therefore, how to achieve efficient and accurate alignment of information entities related to college students' academic activities within the existing entity alignment framework has become an urgent problem to be solved in the current processing of entities to be aligned. Summary of the Invention

[0004] The purpose of this invention is to provide a hierarchical entity alignment method for university students' academic activity data. Through a multi-level semantically enhanced entity alignment module, a context-feature enhanced entity alignment module, and a multi-perspective similarity fusion unknown entity alignment module, it achieves accurate matching of entities in university students' academic data, detailed modeling of unclassified entities, and efficient processing of new entities not included in the candidate database. This improves the accuracy and standardization of entity alignment for university students' academic activity data, and accurately constructs a knowledge graph of university students' characteristics. It addresses the technical problems in existing technologies regarding entity alignment in university students' static basic information and dynamic growth data, such as insufficient capture of short text data features, low utilization of contextual semantic knowledge in the candidate database, and the tendency for copy bias in autoregressive alignment methods.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0006] A hierarchical entity alignment method for college students' academic activity data, the method comprising the following steps:

[0007] Step S1: Obtain the set of entities to be aligned and load the standard entity library;

[0008] Step S2: Semantically encode the entity set to be aligned and the standard entity library to obtain the entity feature set to be aligned and the standard entity library feature set;

[0009] Step S3: Perform multi-level semantic enhancement on the feature set of entities to be aligned to obtain multi-level semantically enhanced features;

[0010] Step S4: Calculate the similarity between the multi-level semantic enhancement features and the standard entities to construct a similarity matrix;

[0011] Step S5: Based on the similarity matrix, evaluate the alignment confidence by combining the preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the standard entity with the highest posterior probability is the alignment result of the entity to be aligned, and a set of effective alignment pairs is obtained; otherwise, if not aligned, perform the subsequent steps S6~S9 for context-enhanced alignment.

[0012] Step S6: Construct the context entity pair representation of the standard entity;

[0013] Step S7: Enhance the context entity pairs of the standard entity to obtain the enhanced features of the context entity pairs of the standard entity; construct the query vector of the entity to be aligned based on the multi-level semantic enhancement features of the entity to be aligned; after updating the query vector of the entity to be aligned, obtain the enhanced features of the entity to be aligned.

[0014] Step S8: Perform a linear transformation on the context entity pair enhancement features of the standard entity to obtain the standard entity fusion features;

[0015] Step S9: Calculate the cosine similarity between the enhanced features of the entity to be aligned and the fused features of the standard entity; calculate the conditional probability of the standard entity corresponding to the entity to be aligned based on the cosine similarity. If the maximum conditional probability is greater than or equal to the conditional threshold, the standard entity corresponding to the maximum conditional probability is the alignment result of the entity to be aligned, and the context-enhanced alignment set is obtained; otherwise, it is not aligned, and the subsequent alignment steps S10~S14 are performed.

[0016] Step S10: Extract auxiliary feature information, full name text information, and main name information of the entity from the unaligned entities to be aligned, and form a multi-view supplementary information set;

[0017] Step S11: Calculate the auxiliary feature information similarity matrix, the entity full name semantic similarity matrix, and the entity name backbone semantic similarity matrix for the auxiliary feature information; obtain a set of multi-view similarity matrices.

[0018] Step S12: Normalize, enhance row confidence, and suppress column centering of the matrices in the multi-view similarity matrix set to obtain the final optimized similarity matrix;

[0019] Step S13: Filter the final optimized similarity matrix, construct a similarity graph, and cluster the similarity graph to obtain a set of entity clusters;

[0020] Step S14: For each entity cluster in the entity cluster set, extract the common features of the entities within the cluster, generate standard entity information, form a new standard entity set, and write it into the standard entity library;

[0021] Step S15: Summarize the set of valid alignment pairs, the set of context-enhanced alignments, and the set of new standard entities for output alignment.

[0022] Further, step S2 includes the following steps:

[0023] Step S21: Encode the entities to be aligned using BERT to obtain the feature set of the entities to be aligned;

[0024] Step S22: Encode the standard entities using BERT to obtain the standard entity feature set.

[0025] Further, step S3 includes the following steps:

[0026] Step S31: Perform a convolution operation on the feature set of entities to be aligned to obtain the local semantic features of the entities to be aligned under the sliding window; the local semantic features of all sliding windows constitute the local semantic features of the entities to be aligned.

[0027] Step S32: The local semantic features of the entities to be aligned are used to obtain the query matrix, key matrix, and value matrix through linear projection;

[0028] Step S33: Calculate the semantic similarity between features by scaling the dot product of the query matrix and the key matrix, and then obtain the attention weight matrix by Softmax normalization;

[0029] Step S34: Use the attention weight matrix to weight the value matrix to obtain the semantic feature sequence after global association enhancement;

[0030] Step S35: Perform row-level aggregation on the attention weight matrix to obtain the weight factor corresponding to each semantic feature;

[0031] Step S36: Based on the weight factor, the local semantic features of the entity to be aligned under the sliding window are weighted to obtain the semantically enhanced local feature representation. All semantically enhanced local feature representations are concatenated to form a weighted feature sequence.

[0032] Step S37: Input the weighted feature sequence into the feedforward neural network to extract deep semantics, and unify the feature distribution through layer normalization to obtain preliminary semantic enhancement features;

[0033] Step S38: Perform max pooling and average pooling on the initial semantic enhancement features to concatenate them to obtain the final multi-level semantic enhancement features.

[0034] Further, step S6 includes the following steps:

[0035] Step S61: Extract standard entity context information, encode the standard entity context information using BERT, and obtain the standard entity context vector;

[0036] Step S62: Concatenate the standard entity feature representation with the standard entity context vector to construct the standard entity context entity pair.

[0037] Further, step S7 includes the following steps:

[0038] Step S71: Concatenate the multi-level semantic enhancement features of the entity to be aligned with the zero vector to obtain the query vector of the entity to be aligned;

[0039] Step S72: Use the context entity pairs of the standard entity as both the standard entity key vector and the standard entity value vector; all standard entity key vectors and standard entity value vectors constitute the standard entity key matrix and the standard entity value matrix, respectively;

[0040] Step S73: Multiply the query vector of the entity to be aligned by the transpose of the standard entity key matrix to obtain the semantic similarity matrix between the entity to be aligned and all standard entity context entity pairs;

[0041] Step S74: Based on the semantic similarity matrix, select the sample with the highest semantic matching degree. For each standard entity, extract the corresponding standard entity key matrix, standard entity value matrix, and semantic similarity matrix;

[0042] Step S75: Normalize the semantic similarity submatrix to obtain the entity pairs to be aligned. Pre-matching attention weights for context entity pairs of standard entities;

[0043] Step S76: Perform matrix multiplication between the pre-matched attention weights and the standard entity value submatrix to obtain the context entity pair enhancement features of the standard entity;

[0044] Step S77: Perform matrix multiplication between the transposed pre-matching attention weights and the query vector of the entity to be aligned, to obtain the updated entity to be aligned; extract the first... The dimension obtains the updated enhanced features of the entity to be aligned.

[0045] Further, step S10 includes the following steps: all unaligned entities to be aligned constitute an unknown unmatched entity set;

[0046] Step S101: Extract auxiliary information from the structured metadata of unknown unmatched entities and encode it into auxiliary feature information using BERT;

[0047] Step S102: Extract the full name text feature representation of the unknown unmatched entity to obtain the full name text information of the entity;

[0048] Step S103: The full name text information of the entity is encoded by BERT to obtain the entity name vector; the entity name vector is projected to obtain the entity name backbone information.

[0049] Further, step S11 includes the following steps:

[0050] Step S111: Calculate the similarity of auxiliary feature information based on the intersection and union of auxiliary feature information to obtain the auxiliary feature information similarity matrix;

[0051] Step S112: Encode the full name text information of the entity into a vector using BERT, calculate the cosine similarity, and obtain the semantic similarity matrix of the full name of the entity.

[0052] Step S113: Encode the core information of entity names into vectors using BERT, calculate the cosine similarity, and obtain the semantic similarity matrix of core names;

[0053] The auxiliary feature information similarity matrix, the entity full name semantic similarity matrix, and the backbone name semantic similarity matrix constitute a multi-view similarity matrix set.

[0054] Further, step S12 includes the following steps:

[0055] Step S121: Normalize the matrices in the multi-view similarity matrix set, and perform weighted fusion on the normalized multi-view similarity matrices to obtain the fused similarity matrix;

[0056] Step S122: Perform row confidence enhancement on the fused similarity matrix to obtain the row-enhanced intermediate matrix;

[0057] Step S123: Perform column center suppression on the row-enhanced intermediate matrix to obtain the final optimized similarity matrix.

[0058] Further, step S13 includes the following steps:

[0059] Step S131: Filter and retain the edges corresponding to the R neighbor nodes with the highest similarity in each row of the final optimized similarity matrix to obtain the Top-R adjacency edge set;

[0060] Step S132: Based on the edge similarity threshold, perform a second filtering on the Top-R adjacent edge set to filter out weakly associated edges and obtain the filtered edge set;

[0061] Step S133: Perform edge symmetry processing on the filtered edge set to construct a sparse adjacency matrix and a similarity graph;

[0062] Step S134: Cluster the similar graphs to obtain a set of entity clusters.

[0063] Further, step S14 includes the following steps:

[0064] Step S141: For each entity cluster in the entity cluster group, extract the common features of the entities within the cluster and generate standard entity information;

[0065] Step S142: Associate the generated standard entity information as a new standard entity set and write it into the standard entity library.

[0066] Compared with the prior art, the present invention has the following beneficial technical effects:

[0067] 1) Avoid feature generalization of few-sample entities: By using the original features in the coding design of each layer of interaction, the representation is not overly smoothed, and the feature distinction between few-sample entities and other entities is enhanced.

[0068] 2) Adapting to short text / inconsistent naming scenarios for college student data: Combining entity-context cross-coding and name standardization, we can accurately extract short text features and achieve entity alignment for different naming expressions.

[0069] 3) Mitigating copy bias and new entity mismatch: By optimizing the similarity matrix through row and column locality enhancement and constructing a symmetric similarity graph, the erroneous links of high-frequency generalized words are reduced, thereby improving the accuracy of unknown entity alignment.

[0070] 4) Solve the problem of coarse-grained labeling of low-confidence entities: refine the modeling of unmatched entities through hierarchical entity alignment and multi-view coding, avoid the accumulation of unknown entities, reduce manual maintenance costs, and improve data processing efficiency. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 is a flowchart of the hierarchical entity alignment method for college students' academic activity data according to the present invention.

[0073] Figure 2 is a diagram of the multi-module collaborative process architecture for entity alignment according to the present invention.

[0074] Figure 3 is a diagram of the multi-level semantic enhancement model structure of the present invention.

[0075] Figure 4 is a model structure diagram of entity-context cross-coding of the present invention.

[0076] Figure 5 is a structural diagram of the multi-view similarity fusion and row and column calibration optimization of the present invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] This invention proposes a hierarchical entity alignment method for college students' academic activity data, as shown in Figures 1-2. The method includes the following steps:

[0079] Step S1: Obtain the set of entities to be aligned and load the standard entity library.

[0080] The entities to be aligned include static basic information and dynamic growth information. Static basic information includes student ID, department, major, gender, year, and place of origin; dynamic growth information includes competition awards, training experience, work experience, student activities, personal honors, and qualifications.

[0081] The entity set to be aligned is represented as ,in, For the first Each entity to be aligned represents a single piece of static basic information or dynamic growth information (such as an award record, an activity experience, etc.). Indicates the number of entities to be aligned. The standard entity library represents this as... ,in, For the first Each standard entity represents a standardized entity in the standard entity library. Indicates the number of standard entities. The loaded standard entity library is a predefined canonical entity library.

[0082] Step S2: Perform semantic encoding on the entity set to be aligned and the standard entity library to obtain the entity feature set to be aligned and the standard entity library feature set.

[0083] Step S21: Encode the entities to be aligned using BERT to obtain the feature set of the entities to be aligned.

[0084] Each entity in the set to be aligned The input is encoded in BERT (Bidirectional Encoder Representations from Transformers) to obtain the first... The feature representation of the entities to be aligned:

[0085]

[0086] in, This represents the entity feature dimension, which is the output dimension of BERT. All entity feature representations to be aligned constitute the entity feature set to be aligned, represented as:

[0087] .

[0088] Step S22: Encode the standard entities using BERT to obtain the standard entity feature set.

[0089] Each standard entity in the standard entity library Input the same BERT model for encoding to obtain the first... A standard entity feature representation:

[0090]

[0091] Because they share the same BERT model, the dimensions of the standard entity features are consistent with the dimensions of the entity features to be aligned. All standard entity feature representations constitute the standard entity feature set:

[0092]

[0093] Step S3: Perform multi-level semantic enhancement on the feature set of entities to be aligned to obtain multi-level semantically enhanced features.

[0094] As shown in Figure 3, the following operation is performed on each entity feature representation in the entity feature set to be aligned:

[0095] Step S31: Perform a convolution operation on the feature set of entities to be aligned to obtain the local semantic features of the entities to be aligned under the sliding window; the local semantic features of all sliding windows constitute the local semantic features of the entities to be aligned.

[0096] No. The feature representation of an entity to be aligned The corresponding BERT-encoded word vector sequence is:

[0097]

[0098] in, For the first Vector representation of each word For the first The number of word segments for each entity text to be aligned, using a window size of [value missing]. Convolution operation on a sliding window: For a subsequence of the sliding window , Perform a linear affine transformation to obtain the intermediate eigenvectors:

[0099]

[0100] in, This represents the intermediate feature vector. The convolution weight matrix is... For bias vectors, The output dimension of the convolution. The intermediate feature vector... By using a nonlinear activation function, the first... The entity to be aligned is in the first... Local semantic features under a sliding window :

[0101]

[0102] in, This represents a non-linear activation function.

[0103] No. The local semantic features of all sliding windows of the entities to be aligned constitute the first... The local semantic features of the entities to be aligned are represented as follows: , This represents the total number of sliding windows.

[0104] The local semantic features of all entities to be aligned constitute the global semantic feature set of the entities to be aligned. , means as follows:

[0105]

[0106] in, This represents the total number of entities to be aligned.

[0107] Step S32: The local semantic features of the entities to be aligned are used to obtain the query matrix, key matrix, and value matrix through linear projection.

[0108] With the first Local semantic feature matrix of entities to be aligned For input, where, Indicates the first The number of semantic features of the entities to be aligned, i.e., the total number of sliding windows; and , For the first The number of words segmented from the entity text to be aligned. This represents the size of the convolution window.

[0109] With local semantic feature matrix As input, a single-head query matrix is ​​obtained through linear projection. Key matrix Value matrix The formula is:

[0110]

[0111]

[0112]

[0113] in, , , These are the query projection weight matrix, key projection weight matrix, and value projection weight matrix, respectively, and the resulting query matrix after projection. Key matrix Value matrix The dimension is , is the feature dimension after projection.

[0114] Step S33: Calculate the semantic similarity between features by scaling the dot product of the query matrix and the key matrix, and then obtain the attention weight matrix by Softmax normalization.

[0115] The semantic similarity between features is calculated by scaling the dot product, and then the attention weight matrix is ​​obtained by Softmax normalization. :

[0116]

[0117]

[0118] in, For the first A query-key similarity matrix of entities to be aligned, whose elements Indicates the first The first entity to be aligned The semantic feature and the first The semantic relevance of each semantic feature; This is a scaling factor used to avoid the vanishing Softmax gradient problem caused by high-dimensional features; For the first The attention weight matrix of the entities to be aligned, its elements Indicates the first The first entity to be aligned The semantic feature is related to the first Attention weights for each semantic feature; Softmax(⋅) is a normalization function used to convert similarity into probabilistic weights.

[0119] Step S34: Use the attention weight matrix to weight the value matrix to obtain the semantic feature sequence after global association enhancement.

[0120] Using attention weight matrix Log-value matrix Weighting is performed to obtain the semantic feature sequence after global association enhancement. :

[0121]

[0122] in, This is a sequence of semantic features enhanced by global association, where each semantic feature incorporates global semantic information from all semantic features of the entity to be aligned.

[0123] Step S35: Perform row-level aggregation on the attention weight matrix to obtain the weight factor corresponding to each semantic feature.

[0124] To further characterize the relative importance of different semantic features in global semantic modeling, the attention weight matrix was adjusted. Perform row-level aggregation to obtain the weight factor corresponding to each semantic feature.

[0125] Specifically, the The weight factors of each semantic feature are defined as the attention weight matrix. No. The aggregated result of the rows is represented as follows:

[0126]

[0127] in, For the first The first entity to be aligned The weight factor of each semantic feature represents the saliency of that semantic feature in the global semantic association; Indicates the first The first entity to be aligned The semantic feature is related to the first Attention weights for each semantic feature.

[0128] Step S36: Based on the weight factor, weight the local semantic features of the entity to be aligned under the sliding window to obtain semantically enhanced local feature representations. Concatenate all semantically enhanced local feature representations to form a weighted feature sequence.

[0129] Based on this weighting factor, for the first The entity to be aligned is in the first... Local semantic features under a window Weighted summation yields semantically enhanced local feature representations:

[0130]

[0131] in, Indicates the first The first entity to be aligned The local feature representation of the semantic feature; ⊙ represents element-wise multiplication, used for the semantic feature of the first element. The entity to be aligned is in the first... Local semantic features under a window Perform explicit weight modulation.

[0132] A weighted feature sequence is constructed by concatenating all semantically enhanced local feature representations. :

[0133]

[0134] in, Indicates the first The first entity to be aligned Local feature representation of a semantic feature.

[0135] Step S37: Input the weighted feature sequence into the feedforward neural network to extract deep semantics, and unify the feature distribution through layer normalization to obtain preliminary semantic enhancement features.

[0136] Weighted feature sequence The input feedforward neural network (FFN) extracts deep semantics, and the feature distribution is unified through layer normalization to obtain preliminary semantic enhancement features. :

[0137]

[0138] in, This represents a feedforward neural network, consisting of two fully connected layers and a ReLU activation function, used to enhance the non-linear expressive power of features; This is a layer normalization operation used to stabilize feature distribution.

[0139] Step S38: Perform max pooling and average pooling on the initial semantic enhancement features to concatenate them to obtain multi-level semantic enhancement features.

[0140] By modifying the initial semantic enhancement features Max pooling and average pooling are performed to extract key semantics and globally distributed semantics, which are then concatenated to obtain the final multi-level semantic enhancement features.

[0141]

[0142]

[0143]

[0144] in, Max-pooled features represent key semantic information within the features; This indicates taking the maximum value along the dimension of the feature sequence; For average pooling features, the global distribution information of the features is preserved; This indicates taking the mean along the feature sequence dimension; Indicates feature concatenation operation; For the first Multi-level semantic enhancement features of the entities to be aligned are used for subsequent similarity calculation with features of the standard entity library; Enhance feature dimensions for multi-level semantics.

[0145] Step S4: Calculate the similarity between the multi-level semantic enhancement features and the standard entities to construct a similarity matrix.

[0146] Let the size of the entity set to be aligned be... That is, including One entity to be aligned; the standard entity library is [size missing]. That is, including A set of standard entities. Based on the multi-level semantic enhancement features of the entities to be aligned, the semantic similarity between each pair of standard entities is calculated, and a similarity matrix is ​​constructed.

[0147] For the One entity to be aligned Its multi-level semantic enhancement features are The first in the standard entity library Standard Entities Its standard entity features are represented as Semantic similarity between multi-level semantic enhancement features and standard entity feature representations Defined as:

[0148]

[0149] in, Indicates the first The multi-level semantic enhancement features of the entities to be aligned and the first Semantic similarity represented by standard entity features; Indicates cosine similarity; Represents the vector dot product. The L2 norm of a vector. The closer the value is to 1, the more similar the two are semantically.

[0150] Based on the above similarity definition, a similarity matrix is ​​constructed between the entity set to be aligned and the standard entity library. Its elements are:

[0151]

[0152] in, , The number of entities to be aligned. The number of standard entities. The th element of the similarity matrix. The row corresponds to the first The entity to be aligned, the first Column corresponding to the first Each standard entity fully reflects the semantic similarity relationship between all entities to be aligned and the standard entities.

[0153] Step S5: Based on the similarity matrix The alignment confidence is evaluated by combining the preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the standard entity with the highest posterior probability is the alignment result of the entity to be aligned, and a set of effective alignment pairs is obtained; otherwise, no alignment is performed, and the subsequent steps S6~S9 are performed for context-enhanced alignment.

[0154] Define alignment confidence as the entity to be aligned Matched to standard entities The posterior probability is denoted as Simultaneously, a preset similarity threshold is established. .

[0155] Posterior probability Similarity The mapping relationship is as follows:

[0156]

[0157] in, This is a temperature coefficient used to adjust the concentration of the probability distribution. The smaller the value, the more concentrated the probability distribution is on entities with high similarity; The larger the value, the flatter the probability distribution; Indicates the first The multi-level semantic enhancement features of the entities to be aligned and the first Semantic similarity represented by standard entity features.

[0158] For entities to be aligned Traverse the similarity matrix The OK:

[0159] If a standard entity exists satisfy If the posterior probability is the highest, then the standard entity with the highest posterior probability is selected as the matching result, that is:

[0160]

[0161] in, Indicates the first One entity to be aligned The matching standard entity; In the standard entity set In, make the posterior probability Largest subscript The corresponding standard entity. These are marked as valid alignment pairs. All valid alignment pairs constitute the set of valid alignment pairs, denoted as . The corresponding relationship is as follows:

[0162]

[0163] If all Then determine the entity to be aligned. If no valid standard entity is matched, proceed to the subsequent context-enhanced entity alignment process.

[0164] Step S6: Construct the context entity pair representation of the standard entity.

[0165] Step S61: Extract standard entity context information, encode the standard entity context information using BERT, and obtain the standard entity context vector.

[0166] For standard entity library The first in Standard Entities Extracting contextual information from annotated corpora or related knowledge bases: specifically, extracting the contextual information of standard entities before and after their appearance in the text. The text fragments of the words serve as contextual information for the standard entity; For hyperparameters, take .

[0167] The standard entity context information is encoded into a vector form using the pre-trained encoder BERT, resulting in the standard entity context vector, as shown below:

[0168]

[0169] in, Indicates the first _ standard entity context vectors, where Dimensions encoded for context.

[0170] Step S62: Concatenate the standard entity feature representation with the standard entity context vector to construct the standard entity context entity pair.

[0171] The first Each standard entity feature represents , For entity feature dimension, and the first Standard entity context vectors By splicing, we obtain the first... Context entity pairs of standard entities :

[0172]

[0173] Step S7: Enhance the context entity pairs of the standard entity to obtain the enhanced features of the context entity pairs of the standard entity; construct the query vector of the entity to be aligned based on the multi-level semantic enhancement features of the entity to be aligned; after updating the query vector of the entity to be aligned, obtain the enhanced features of the entity to be aligned.

[0174] In this embodiment, as shown in Figure 4, the model structure diagram of entity-context cross-coding is presented.

[0175] Step S71: Concatenate the multi-level semantic enhancement features of the entity to be aligned with the zero vector to obtain the query vector of the entity to be aligned.

[0176] Let the entity to be aligned be Its multi-level semantic enhancement features are , This refers to the dimension of multi-level semantic enhancement features. The multi-level semantic enhancement features are then combined with the dimension... The zero vectors are concatenated to obtain the query vector of the entity to be aligned. :

[0177]

[0178] in, The dimension is The zero vector, This is a vector concatenation operation.

[0179] Step S72: Use the context entity pairs of the standard entity as both the standard entity key vector and the standard entity value vector; all standard entity key vectors and standard entity value vectors constitute the standard entity key matrix and the standard entity value matrix, respectively.

[0180] For standard entity library The first in Standard Entities , with its context entity pair At the same time, it serves as a standard entity key vector. Compared with standard entity value vector ,Right now:

[0181]

[0182] All standard entity key vectors and standard entity value vectors constitute the standard entity key matrix. With standard entity value matrix :

[0183]

[0184]

[0185] Step S73: Multiply the query vector of the entity to be aligned by the transpose of the standard entity key matrix to obtain the semantic similarity matrix between the entity to be aligned and all standard entity context entity pairs;

[0186] The query vector of the entity to be aligned With standard entity key matrix Perform transpose matrix multiplication to obtain the semantic similarity matrix between the entity to be aligned and all standard entity context entity pairs. :

[0187]

[0188] in, Standard entity key matrix The transpose of the matrix, This indicates the transpose. The transposed matrix has dimensions of 1. , and query vector dimension match.

[0189] Step S74: Based on the semantic similarity matrix, select the sample with the highest semantic matching degree. For each standard entity, extract the corresponding standard entity key matrix, standard entity value matrix, and semantic similarity matrix.

[0190] Based on semantic similarity matrix Filter out the ones with the highest semantic matching degree Each standard entity is synchronously retrieved. Standard entity key matrix and Standard entity value submatrix :

[0191]

[0192]

[0193] Simultaneously extract the semantic similarity matrix. Chinese correspondence semantic similarity submatrix .

[0194] Step S75: Normalize the semantic similarity submatrix to obtain the entity pairs to be aligned. Pre-matching attention weights for context entity pairs of standard entities.

[0195] For semantic similarity submatrix The entity to be aligned is obtained through Softmax normalization. right Pre-matching attention weights for context entity pairs of standard entities :

[0196]

[0197] in, This is the square root of the feature dimension, used to eliminate the interference of dimension on similarity.

[0198] Step S76: Perform matrix multiplication between the pre-matched attention weights and the standard entity value submatrix to obtain the context entity pair enhancement features of the standard entity.

[0199] Pre-matched attention weights and Standard entity value submatrix After performing matrix multiplication, we get Contextual entity pairs of standard entities enhance features :

[0200]

[0201] in, express The context entity pairs of the standard entity enhance the features, incorporating weighted information from the attention weights.

[0202] Step S77: Perform matrix multiplication between the transposed pre-matching attention weights and the query vector of the entity to be aligned, to obtain the updated entity to be aligned; extract the first... The dimension obtains the updated enhanced features of the entity to be aligned.

[0203] Pre-matched attention weights After transposition, the query vector of the entity to be aligned Perform matrix multiplication to obtain the updated entity to be aligned. :

[0204]

[0205] in, This indicates the previously extracted indivual.

[0206] Extract the updated entity to be aligned The former Dimension, which is the updated augmented feature of the entity to be aligned. :

[0207]

[0208] in, Indicates taking the first part of the vector Projection operations on dimensions.

[0209] Step S8: Perform a linear transformation on the context entity pair enhancement features of the standard entity to obtain the standard entity fusion features.

[0210] right Contextual entity pairs of standard entities enhance features (Same as step S76) Perform a linear transformation, then... Context entity pairs of dimensional standard entities enhance features Mapped to 3D features, to obtain Standard entity fusion features of dimensionality The formula is:

[0211]

[0212] in, Represents the weight matrix; Represents the bias vector; and These are learnable parameters used to enhance features based on contextual entity pairs of standard entities. The entity and context information are unified in dimension to adapt to the fusion standard.

[0213] Step S9: Calculate the cosine similarity between the enhanced features of the entity to be aligned and the fused features of the standard entity; calculate the conditional probability of the standard entity corresponding to the entity to be aligned based on the cosine similarity. If the maximum conditional probability is greater than or equal to the conditional threshold, the standard entity corresponding to the maximum conditional probability is the alignment result of the entity to be aligned, and the context-enhanced alignment set is obtained; otherwise, it is not aligned, and the subsequent alignment steps S10~S14 are performed.

[0214] Enhanced features of entities to be aligned based on dimensional unification Features of integration with standard entities The cosine similarity between the two is calculated to represent the semantic matching degree. The formula for cosine similarity is:

[0215]

[0216] in, Indicates the enhanced features of the entity to be aligned With the Standard entity fusion characteristics of a standard entity cosine similarity, for and Dot product operation; Let L2 be the L2 norm of the vector.

[0217] Entities to be aligned based on similarity calculation Conditional probability of corresponding standard entities :

[0218]

[0219] in, It is an exponential function; Indicates the enhanced features of the entity to be aligned With the Standard entity fusion characteristics of a standard entity The cosine similarity.

[0220] Let the conditional probability threshold be... , Take the maximum value of the conditional probability and make a judgment:

[0221] When the maximum conditional probability ≥Conditional probability threshold Select the standard entity name corresponding to the maximum conditional probability. As entities to be aligned The alignment result; Marked as a valid alignment pair.

[0222] when Alignment confidence is insufficient, alignment will not be performed.

[0223] All satisfying the conditional probability threshold The matching results between the entities to be aligned and the standard entities constitute a context-enhanced alignment set. .

[0224] Step S10: Extract auxiliary feature information, full name text information, and main name information of the entity from the unaligned entities to be aligned, forming a multi-view supplementary information set.

[0225] All misaligned entities to be aligned constitute an unknown set of unmatched entities. From the set of entities to be aligned Entities with insufficient alignment confidence are selected from the pool to form an unknown unmatched entity set. The set of unknown, unmatched entities must simultaneously satisfy two conditions:

[0226] Condition 1: Failed to meet the similarity threshold The filtering (i.e., all similarities in step S5) Less than the similarity threshold (The entity to be aligned).

[0227] Condition 2: In subsequent context-enhanced alignment, the maximum conditional probability is still less than the conditional probability threshold. (Right now ).

[0228] Therefore, the unknown set of unmatched entities The complete definition is:

[0229]

[0230] in, Indicates the first One entity to be aligned No valid standard entities were matched in the initial similarity screening; This indicates that the maximum match probability for this entity remains below the confidence threshold in subsequent context-enhanced alignment. Unknown set of unmatched entities. The entities in the model need to be supplemented with more multi-dimensional information to improve alignment accuracy.

[0231] Extracting multi-dimensional supplementary information of unknown and unmatched entities provides data support for multi-view similarity calculation.

[0232] Step S101: Extract auxiliary information from the structured metadata of unknown unmatched entities and encode it into auxiliary feature information using BERT.

[0233] From unknown unmatched entities Auxiliary information, such as achievement classification, award date, and awarding organization, is extracted from structured metadata and encoded using a pre-trained word vector model BERT to obtain auxiliary feature information, denoted as... ( ).in, Represents the set of unknown, unmatched entities. The first in Individual entities.

[0234] Step S102: Extract the full name text feature representation of the unknown unmatched entity to obtain the full name text information of the entity.

[0235] Extracting unknown unmatched entities The full name text feature representation yields the entity's full name text information, denoted as . ( ).

[0236] Step S103: The full name text information of the entity is encoded by BERT to obtain the entity name vector; the entity name vector is projected to obtain the entity name backbone information.

[0237] Entity full name text information The entity name vector is obtained by encoding the pre-trained word vector model BERT into high-dimensional semantic vectors. , For entity feature dimensions.

[0238] Perform principal component analysis (PCA) on the set of entity name vectors for all unknown and unmatched entities, calculate the covariance matrix, and solve for the eigenvalues ​​and eigenvectors. Select the entities with the largest variance contributions. Principal components ( ).

[0239] Entity Name Vector Project to front The subspace composed of principal components yields the dimensionality-reduced principal component vector, which represents the core information of the entity name, denoted as . .

[0240] The auxiliary feature information, full name text information, and core name information of the unknown unmatched entity constitute a multi-perspective supplementary information set for the unknown unmatched entity. .

[0241] Step S11: Calculate the auxiliary feature information similarity matrix, the entity full name semantic similarity matrix, and the entity name backbone semantic similarity matrix for the auxiliary feature information, and obtain a set of multi-view similarity matrices to achieve multi-view similarity fusion.

[0242] In this embodiment, as shown in Figure 5, a structural diagram of multi-view similarity fusion and row / column calibration optimization is constructed. Based on multi-dimensional supplementary information, a multi-view similarity matrix between unknown entities is built.

[0243] Step S111: Calculate the similarity of auxiliary feature information based on the intersection and union of auxiliary feature information to obtain the similarity matrix of auxiliary feature information.

[0244] Based on the intersection and union of auxiliary feature information of unknown and unmatched entities, the matching degree of auxiliary feature information is calculated, and an auxiliary feature information similarity matrix is ​​constructed. Auxiliary information similarity matrix elements The calculation method is as follows:

[0245]

[0246] in, The first one in the set of unknown unmatched entities , One entity, Indicates the number of elements in the set. Characterizing the first individual entities With the individual entities The degree of matching of auxiliary feature information.

[0247] Step S112: Encode the full name text information of the entity into a vector using BERT, calculate the cosine similarity, and obtain the semantic similarity matrix of the full name of the entity.

[0248] The full-name text information of entities is encoded into vectors using the pre-trained language model BERT, and cosine similarity is calculated to construct a semantic similarity matrix of full-name entities. Entity full name semantic similarity matrix elements The calculation method is as follows:

[0249]

[0250] in, , They represent the first , The full text information of an unknown, unmatched entity; This is a text encoding function that converts text into a vector. This represents the transpose of a vector. The L2 norm of a vector. Characterization and The full name is semantic matching degree.

[0251] Step S113: Encode the entity name backbone information into a vector using BERT, calculate the cosine similarity, and obtain the backbone name semantic similarity matrix.

[0252] The entity name backbone information is encoded into a vector using the pre-trained language model BERT, and cosine similarity is calculated to construct a semantic similarity matrix of backbone names. Semantic similarity matrix of main names elements The calculation method is as follows:

[0253]

[0254] in, , They represent the first , The backbone information of the entity names of the unknown, unmatched entities; text encoding function. The same pre-trained language model BERT is used as in step S112, but the input is entity name backbone information.

[0255] Auxiliary feature information similarity matrix Entity full name semantic similarity matrix and the semantic similarity matrix of the main names Construct a set of multi-view similarity matrices ;

[0256] Step S12: Normalize, enhance row confidence, and suppress column centers of the matrix in the multi-view similarity matrix set to obtain the final optimized similarity matrix.

[0257] An optimized unified similarity matrix is ​​obtained through adaptive weighting and noise reduction.

[0258] Step S121: Normalize the matrices in the multi-view similarity matrix set, and perform weighted fusion on the normalized multi-view similarity matrices to obtain the fused similarity matrix.

[0259] First, the multi-view similarity matrix set is normalized to... Interval: Auxiliary feature information similarity matrix Already No processing required; entity full name semantic similarity matrix Semantic similarity matrix of main names Original range pass:

[0260]

[0261]

[0262] Normalization to , This represents the normalized semantic similarity matrix of entity full names. This represents the normalized semantic similarity matrix of the main names.

[0263] Introducing learnable gating parameters ,satisfy The multi-view similarity matrix is ​​weighted and fused to obtain the fused similarity matrix. fusion similarity matrix elements The calculation method is as follows:

[0264]

[0265] in, The similarity value after fusion; This represents the normalized semantic similarity matrix of entity full names. Element; This represents the normalized semantic similarity matrix of the main branch names. Element.

[0266] Step S122: Perform row confidence enhancement on the fused similarity matrix to obtain the row-enhanced intermediate matrix.

[0267] For the fusion similarity matrix The line, corresponding to the first Unknown unmatched entities Iterate through all elements in the row. ( ), excluding its own matches, i.e. After selecting the elements, choose the element with the highest similarity within the row. The second largest similarity as :

[0268]

[0269] Calculate the maximum similarity Second largest similarity The difference The distinguishing factor between the best and second-best matches within this row is represented as follows:

[0270]

[0271] The difference is expressed by the Sigmoid function. Mapped to row weights of the (0,1) interval The formula is:

[0272]

[0273] in, The Sigmoid activation function is used to map the difference to smooth weights. , Learnable parameters (optimized via a loss function during training): Control the amplification factor of the difference. The baseline offset of the control weights is determined by mapping to (0,1). The purpose of mapping is to strengthen the weights of high-confidence rows and weaken the interference of low-confidence rows.

[0274] Use row weights For the We weight all elements of each row to obtain the row-enhanced intermediate matrix. intermediate matrix elements It is expressed as follows:

[0275]

[0276] Step S123: Perform column center suppression on the row-enhanced intermediate matrix to obtain the final optimized similarity matrix.

[0277] Column centrality is quantified based on the number of nearest neighbors for each column. Overly correlated columns in the row-enhanced intermediate matrix are suppressed by dynamic weights. The specific operation is as follows:

[0278] Preset nearest neighbor threshold The intermediate matrix after row enhancement The Column, corresponding to the first Unknown unmatched entities The similarity within a column is greater than the nearest neighbor threshold. The number of elements (excluding its own matches) ), as the number of nearest neighbors of that column. :

[0279]

[0280] in, Indicates the number of elements in the set. For the first The number of nearest neighbors of a column.

[0281] Based on the number of nearest neighbors Generate column suppression weights The formula is:

[0282]

[0283] in, The design is to avoid There are cases where the denominator is 0.

[0284] Use column weights For the We weight all the elements in the column to obtain the final optimized similarity matrix. The final optimized similarity matrix elements It is expressed as follows:

[0285]

[0286] Step S13: Filter the final optimized similarity matrix, construct a similarity graph, cluster the similarity graph, and obtain a set of entity clusters.

[0287] Step S131: Filter and retain the edges corresponding to the R neighbor nodes with the highest similarity in each row of the final optimized similarity matrix to obtain the Top-R adjacency edge set.

[0288] For each unknown unmatched entity This corresponds to the final optimized similarity matrix. The Okay, exclude its own matches. Then, retain the edges corresponding to the R neighboring nodes with the highest similarity to the current node, and keep only the final optimized similarity matrix. The edge relationships corresponding to the top R elements are used to obtain the Top-R adjacency edge set.

[0289] Step S132: Based on the edge similarity threshold, perform a secondary screening of the Top-R adjacent edge set to filter out weakly associated edges and obtain the filtered edge set.

[0290] Preset edge similarity threshold A secondary filtering is performed on the Top-R adjacency set: only those that meet the following conditions are retained. The edges are filtered to obtain the filtered edge set by filtering out weakly correlated edges with low confidence.

[0291] Step S133: Perform edge symmetry processing on the filtered edge set to construct a sparse adjacency matrix and a similarity graph.

[0292] Perform a reciprocity check on the filtered edge set: only retain bidirectional adjacent edges, i.e., if nodes are mutually adjacent... yes Top-R neighbors, and yes Top-R neighbors, then retain the edges Otherwise, remove the edge to ensure the reliability of the edge association.

[0293] Based on the edge set selected above, a sparse adjacency matrix is ​​constructed. :

[0294]

[0295] The sparse adjacency matrix corresponds to a similar graph. ,in For a set of nodes, This is the filtered edge set.

[0296] Step S134: Cluster the similar graphs to obtain a set of entity clusters.

[0297] Using a class graph clustering algorithm to cluster similar graphs Clustering is performed to obtain a set of entity clusters. ; Indicates the first A cluster of entities, containing unknown, unmatched entities with different representations but consistent semantics; Indicates the total number of entity clusters; within the same cluster Entities within the same category are identified as matching entities with different descriptions but consistent semantics.

[0298] Step S14: For each entity cluster in the entity cluster set, extract the common features of the entities within the cluster, generate standard entity information, form a new standard entity set, and write it into the standard entity library.

[0299] Standardized entity information is generated based on entity clusters and stored in the database to ensure consistency in entity alignment. The specific operations are as follows:

[0300] Step S141: For each entity cluster in the entity cluster group, extract the common features of the entities within the cluster to generate standard entity information.

[0301] The common characteristics of an entity include its standard name, aliases, and field prototypes.

[0302] Standard name determination: For each entity name within an entity cluster After being encoded into vectors using the pre-trained language model BERT, the average semantic similarity between the vector and all other entity names within the cluster is calculated:

[0303]

[0304] in, Indicates the first A cluster of entities, Indicates the first The number of entities in an entity cluster; Indicates the name of the entity currently to be evaluated within the cluster; Indicates cluster division Other entity names; For text encoding functions, Cosine similarity. Average semantic similarity is selected. The largest name is used as the standard name for this cluster.

[0305] Alias ​​organization: Mark semantically equivalent names within an entity cluster that are similar to the standard name, such as variants or abbreviations with different characters, as aliases of the standard name; when subsequent entities to be aligned match the aliases, they will be automatically associated with the corresponding standard name.

[0306] Field prototype construction: Extract common templates of structured information of entities within an entity cluster, such as achievement classification, award time, awarding unit, etc., as field prototypes for that type of entity; subsequent structured information of similar entities must follow this template.

[0307] Step S142: Associate the generated standard entity information as a new standard entity set and write it into the standard entity library.

[0308] The generated standard entity information, including standard names, aliases, and field prototypes, is associated with a new set of standard entities. Write it into the standard entity library; at the same time, write the entity cluster. All entities within the scope to be aligned are associated with this newly generated standard entity.

[0309] Step S15: Summarize the set of valid alignment pairs, the set of context-enhanced alignments, and the set of new standard entities for output alignment.

[0310] The aggregated output set represents all aligned entities, containing alignment results from the following three sources:

[0311] Source 1: Set of valid alignment pairs That is, meeting the similarity threshold Entities to be aligned Compared with the original standard entity Matching pairs ;

[0312] Source 2: Context-enhanced alignment sets That is, satisfying the conditional probability threshold Matching pairs of entities to be aligned with standard entities ;

[0313] Source 3: New Standard Entity Set This refers to the newly generated standard entity and the association relationship between the corresponding entities to be aligned within the cluster.

[0314] Output: The final set of entities This set contains all the contents of the previous three sets, that is... This covers all entities that have been aligned and newly generated.

[0315] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical entity alignment method for college students' academic activity data, characterized in that, The method includes the following steps: Step S1: Obtain the entity set to be aligned and load the standard entity library; Step S2: Perform semantic encoding on the entity set to be aligned and the standard entity library to obtain the entity feature set to be aligned and the standard entity library feature set; Step S3: Perform multi-level semantic enhancement on the entity feature set to be aligned to obtain multi-level semantic enhancement features; Step S4: Calculate the similarity between the multi-level semantic enhancement features and the standard entities to construct a similarity matrix; Step S5: Based on the similarity matrix, evaluate the alignment confidence by combining a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the standard entity with the highest posterior probability is the alignment result of the entity to be aligned, thus obtaining a set of effective alignment pairs. Conversely, if not aligned, proceed with subsequent steps S6~S9 for context-enhanced alignment; Step S6: Construct the context entity pair representation of the standard entity; Step S7: Enhance the context entity pair of the standard entity to obtain the context entity pair enhancement feature of the standard entity; Construct a query vector for the entity to be aligned based on the multi-level semantic enhancement features of the entity to be aligned; after updating the query vector of the entity to be aligned, obtain the enhanced features of the entity to be aligned; Step S8: linearly transform the enhanced features of the context entity pair of the standard entity to obtain the standard entity fusion features; Step S9: calculate the cosine similarity between the enhanced features of the entity to be aligned and the standard entity fusion features; calculate the conditional probability of the standard entity corresponding to the entity to be aligned based on the cosine similarity. If the maximum conditional probability is greater than or equal to the conditional threshold, the standard entity corresponding to the maximum conditional probability is the alignment result of the entity to be aligned, and obtain the context enhancement alignment set; Conversely, if not aligned, proceed to the alignment steps S10-S14. Step S10: Extract auxiliary feature information, full-name text information, and core name information from the unaligned entities to be aligned, forming a multi-view supplementary information set. Step S11: Calculate the auxiliary feature information similarity matrix, the full-name semantic similarity matrix, and the core name semantic similarity matrix for the entity full-name text information; obtain a multi-view similarity matrix set. Step S12: Normalize, enhance row confidence, and suppress column centers in the multi-view similarity matrix set to obtain the final optimized similarity matrix. Step S13: Filter the final optimized similarity matrix, construct a similarity graph, and cluster the similarity graph to obtain an entity cluster set. Step S14: Extract common features from each entity cluster in the entity cluster set, generate standard entity information, form a new standard entity set, and write it into the standard entity library. Step S15: Summarize and output the effective alignment pair set, and the upper and lower alignment pairs. The text enhancement alignment set and the new standard entity set; Step S10 includes the following steps: all unaligned entities to be aligned constitute an unknown unmatched entity set; Step S101: extract auxiliary information from the structured metadata of the unknown unmatched entities and encode it into auxiliary feature information through BERT; Step S102: extract the full name text feature representation of the unknown unmatched entities to obtain the full name text information of the entities; Step S103: encode the full name text information of the entities into entity name vectors through BERT; project the entity name vectors to obtain the entity name backbone information; Step S11 includes the following steps: Step S111: calculate the similarity of auxiliary feature information based on the intersection and union of auxiliary feature information to obtain the auxiliary feature information similarity matrix; Step S112: encode the full name text information of the entities into vectors through BERT, calculate the cosine similarity, and obtain the entity full name semantic similarity matrix; Step S113: encode the entity name backbone information into vectors through BERT, calculate the cosine similarity, and obtain the backbone name semantic similarity matrix; The auxiliary feature information similarity matrix, the entity full name semantic similarity matrix, and the backbone name semantic similarity matrix constitute a multi-view similarity matrix set.

2. The hierarchical entity alignment method for university students' academic activity data according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Encode the entity to be aligned using BERT to obtain the entity feature set; Step S22: Encode the standard entity using BERT to obtain the standard entity feature set.

3. The hierarchical entity alignment method for university students' academic activity data according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform a convolution operation on the feature set of entities to be aligned to obtain the local semantic features of the entities to be aligned under the sliding window; the local semantic features of all sliding windows constitute the local semantic features of the entities to be aligned; Step S32: Obtain the query matrix, key matrix, and value matrix from the local semantic features of the entities to be aligned through linear projection; Step S33: Calculate the semantic similarity between features by scaling the dot product of the query matrix and key matrix, and then normalize them using Softmax to obtain the attention weight matrix; Step S34: Use the attention weight matrix to weight the value matrix to obtain the semantic feature order after global association enhancement. Step S35: Perform row-level aggregation on the attention weight matrix to obtain the weight factor corresponding to each semantic feature; Step S36: Based on the weight factor, weight the local semantic features of the entity to be aligned under the sliding window to obtain the semantically enhanced local feature representation, and concatenate all semantically enhanced local feature representations to form a weighted feature sequence; Step S37: Input the weighted feature sequence into the feedforward neural network to refine the deep semantics, and unify the feature distribution through layer normalization to obtain the preliminary semantically enhanced features; Step S38: Perform max pooling and average pooling on the preliminary semantically enhanced features to concatenate them to obtain the final multi-level semantically enhanced features.

4. The hierarchical entity alignment method for university students' academic activity data according to claim 1, characterized in that, Step S6 includes the following steps: Step S61: Extract standard entity context information, encode the standard entity context information using BERT to obtain the standard entity context vector; Step S62: Concatenate the standard entity feature representation with the standard entity context vector to construct the standard entity context entity pair.

5. The hierarchical entity alignment method for university students' academic activity data according to claim 1, characterized in that, Step S7 includes the following steps: Step S71: Concatenate the multi-level semantic enhancement features of the entity to be aligned with the zero vector to obtain the query vector of the entity to be aligned; Step S72: Use the context entity pairs of the standard entities as both standard entity key vectors and standard entity value vectors; all standard entity key vectors and standard entity value vectors constitute the standard entity key matrix and standard entity value matrix, respectively; Step S73: Multiply the query vector of the entity to be aligned by the transpose of the standard entity key matrix to obtain the semantic similarity matrix between the entity to be aligned and all standard entity context entity pairs; Step S74: Based on the semantic similarity matrix, select the entity with the highest semantic matching degree. For each standard entity, extract the corresponding standard entity key matrix, standard entity value matrix, and semantic similarity submatrix; Step S75: Normalize the semantic similarity submatrix to obtain the entity pairs to be aligned. Pre-matching attention weights for the context entity pairs of the standard entity; Step S76: Perform matrix multiplication between the pre-matching attention weights and the standard entity value submatrix to obtain the enhanced features of the context entity pairs of the standard entity; S77: Perform matrix multiplication between the transposed pre-matching attention weights and the query vector of the entity to be aligned, resulting in the updated entity to be aligned; extract the first... The dimension obtains the updated enhanced features of the entity to be aligned.

6. The hierarchical entity alignment method for university students' academic activity data according to claim 1, characterized in that, Step S12 includes the following steps: Step S121: Normalize the matrices in the multi-view similarity matrix set, and perform weighted fusion on the normalized multi-view similarity matrices to obtain a fused similarity matrix; Step S122: Perform row confidence enhancement on the fused similarity matrix to obtain an intermediate matrix after row enhancement; Step S123: Perform column center suppression on the intermediate matrix after row enhancement to obtain the final optimized similarity matrix.

7. The hierarchical entity alignment method for college student academic activity data according to claim 1, characterized in that, Step S13 includes the following steps: Step S131: Filter and retain the edges corresponding to the R neighbor nodes with the highest similarity in each row of the final optimized similarity matrix to obtain the Top-R adjacency edge set; Step S132: Perform a second filtering on the Top-R adjacency edge set based on the edge similarity threshold to filter weakly related edges, obtaining the filtered edge set; Step S133: Perform edge symmetry processing on the filtered edge set to construct a sparse adjacency matrix and a similarity graph; Step S134: Cluster the similarity graph to obtain a set of entity clusters.

8. The hierarchical entity alignment method for college student academic activity data according to claim 1, characterized in that, Step S14 includes the following steps: Step S141: For each entity cluster in the entity cluster group, extract the common features of the entities within the cluster to generate standard entity information; Step S142: Associate the generated standard entity information as a new standard entity set and write it into the standard entity library.

Images