A method and system for archive retrieval based on historical semantic proxy graph convolutional networks
By constructing annotation-aware heterogeneous graphs and historical semantic proxies, the problems of semantic gap and information leakage in archival retrieval are solved, achieving deep semantic matching and robust relevance prediction, thereby improving the relevance and reliability of archival retrieval.
Patent Information
- Application Number
- CN202611427310.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing document retrieval methods struggle to capture the deep semantic relationship between document content and user needs, resulting in a semantic gap. Furthermore, their relevance prediction accuracy decreases when the target interaction does not allow for the viewing of annotation text, demonstrating insufficient robustness.
We construct a heterogeneous annotation perception graph, integrate node representations and edge features through a gating mechanism, learn user and archive graph representations, aggregate non-target historical annotation semantics of target users and archives, synthesize historical semantic proxies through a multilayer perceptron, predict relevance, and generate retrieval results.
It bridges the semantic gap of traditional literal matching, maintains consistency between training and online retrieval, avoids future information leakage, and provides stable relevance predictions and refined search results.
Smart Images

Figure CN123115391A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of archival informatization and intelligent retrieval technology, specifically to an archival retrieval method and system based on historical semantic proxy graph convolutional networks. Background Technology
[0002] With the deepening of e-government, digital libraries, and enterprise informatization, various institutions have accumulated massive amounts of electronic archives and historical documents. How to quickly and accurately retrieve the archives most relevant to user needs from these large-scale databases has become a key issue in the field of archival informatization. Traditional archival retrieval mainly relies on keyword matching, classification numbers, subject terms, and literal comparison of bibliographic metadata. This makes it difficult to depict the deep semantic relationship between archival content and user needs, resulting in a widespread "semantic gap" and limited relevance and utilization value of the retrieval results.
[0003] To improve the relevance of retrieval and recommendation, researchers have successively introduced collaborative filtering, matrix factorization, and neural collaborative filtering into information retrieval and personalized recommendation. Text-oriented deep models such as DeepCoNN, NARRE, DAML, and HFT further utilize historical review texts of users and items to model preferences, significantly improving the accuracy of relevance prediction. In recent years, graph neural networks have been widely used to characterize user-item interaction structures, with representative methods including NGCF, LightGCN, and the graph contrastive learning method RGCL that integrates reviews. Furthermore, TransNets, BERT4Rec, and SGL have also improved representation quality from the perspectives of text generation, sequence modeling, and supervision.
[0004] However, when directly applying the above methods to document retrieval scenarios, the following shortcomings exist: (1) Traditional retrieval based on keywords, classification numbers and bibliographic metadata is essentially a literal matching, which is difficult to deal with differences in expression and polysemy, and lacks a deep semantic understanding of the content of archives and user needs; (2) The augmentation method that integrates annotation / comment text relies on the annotation text of the target interaction pair during the training phase, but this text does not exist in the actual retrieval reasoning, causing future information leakage and a significant drop in relevance prediction accuracy after the model goes online; (3) Generative completion (such as TransNets-style text generation) directly fabricates target semantics when target information is missing, which is prone to semantic drift and noise. In scenarios with sparse data and cold start, it may amplify errors and lack robustness. (4) Due to the sparse nature of the archives and the constraint that the target annotations are invisible, existing methods are unable to provide stable and reliable relevance predictions. Summary of the Invention
[0005] The purpose of this invention is to propose a document retrieval method and system based on historical semantic proxy graph convolutional networks, which achieves deep semantic matching, training-inference consistency, robust proxy, and reliable relevance prediction.
[0006] According to a first aspect of the embodiments of this disclosure, a method for archive retrieval based on a historical semantic proxy graph convolutional network is provided, comprising the following steps: Construct a heterogeneous graph for annotation awareness, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features; On the annotation-aware heterogeneous graph, node representations and edge features are fused through a gating mechanism to learn user graph representations and archive graph representations; The non-target historical annotation semantics of the target user and the target file are aggregated and combined with the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron. Based on the user graph representation, the file graph representation, and the historical semantic agent, the relevance of the target user to the target file is predicted, and the candidate files are ranked according to the relevance to generate search results.
[0007] In one embodiment, the step of constructing the annotation-aware heterogeneous graph specifically includes: constructing a graph structure using elements in the user set and the archive set as nodes and historical usage records as edges; wherein each historical usage record is represented as a quadruple (u, a, r). ua ,t ua ), u represents the user, a represents the file, r ua For relevance level, t ua To utilize the annotated text; a pre-trained language model is used to process the annotated text. ua Encoded as edge semantic vector e ua and the edge semantic vector e ua As an edge feature, it is bound to the corresponding edge; according to the relevance level r ua The edges are distinguished to construct the annotation-aware heterogeneous graph.
[0008] In one embodiment, the user graph representation and the archive graph representation are obtained as follows: the message propagated from user u to archive a is defined as: m (r) u→a =(W·h u ⊙σ(w p ⊤ e ua )+W r ·e ua ⊙σ(w r ⊤ e ua ))⊙cu Among them, h u For the initial node representation of user u, e ua Let W and W' be edge semantic vectors. r For a learnable transformation matrix, w p w r Let σ(·) be the gated vector, σ(·) be the Sigmoid function, ⊙ denote the element-wise product, and c u The normalization coefficients are used; after stacking L layers of message passing and aggregation, the user graph representation and archive graph representation are obtained: g u ,g a =Encoder(G,ε) Where G is the annotation-aware heterogeneous graph, ε is its edge set, and Encoder(·) represents a multi-layer gated graph convolutional encoder.
[0009] In one embodiment, in the step of aggregating the non-target historical annotation semantics of the target user u and the target file a, the mean aggregation is performed on all non-target historical annotation semantic vectors of the target user u and the target file a respectively to obtain the user-side historical semantics e. u -a and archival historical semantics e a -u : e u -a =(1 / |R u -a |)·Σ j e uj e a -u =(1 / |R a -u |)·Σ v e av Among them, e uj For the semantic vector of the j-th historical annotation of user u, e av Let |R| be the semantic vector of the vth historical annotation in file a. u -a |、|R a -u | represents the number of historical annotations.
[0010] In one embodiment, a historical semantic proxy for the target interaction is synthesized using a multilayer perceptron. ua The method is as follows: ê ua =MLP([g u ||g a ||e u-a ||e a -u ‖(e u -a ⊙e a -u )‖|e u -a -e a -u |‖(g u ⊙g a )]) Among them, g u For user graph representation, g a For archival representation, e u -a For user-side historical semantics, e a -u For the historical semantics of the archives, (e u -a ⊙e a -u ) represents the common semantics between users and archive history, |e u -a -e a -u | represents the semantic differences between users and archive history, (g) u ⊙g a )Structured interaction items.
[0011] In one embodiment, the relevance of a target user to a target profile is predicted by first calculating the logits value z for each relevance level using a multilayer perceptron scoring function. ua,k : z ua,k =φ([g u ||g a ‖ê ua ‖ē u -a ‖ē a -u ||(g u ⊙g a )‖ê ua ⊙((g u +g a ) / 2)‖|ē u -a -ē a -u |]) k Where φ(·) is the multilayer perceptron scoring function, ê ua For historical semantic proxy, ē u -a , ē a -uFor the aggregated user-side historical semantics and archive-side historical semantics, g u For user graph representation, g a For archival representation, k represents the k-th relevance level; Then, the relevance rank distribution P(r) is obtained using the Softmax function. ua =k): P(r ua =k)=exp(z ua,k ) / Σ j∈R exp(z ua,j ) Finally, the expected decoding yields continuous correlation prediction values: r̂ ua =Σ k∈R k·P(r ua =k) Where R is the set of relevance levels.
[0012] In one embodiment, the step of sorting candidate files according to the relevance and generating search results specifically includes: during online retrieval, for a given target user, calculating the continuous relevance prediction value r̂ between the user and each candidate file. ua According to the aforementioned correlation prediction value r̂ ua Sort the candidate files in descending order from high to low, and output the sorted list of candidate files as the file search results for the target user.
[0013] According to a second aspect of the present disclosure, an archive retrieval system based on a historical semantic proxy graph convolutional network is provided, comprising: The annotation-aware heterogeneous graph construction module constructs an annotation-aware heterogeneous graph, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features. The gated graph encoding module learns user graph representation and archive graph representation by fusing node representation and edge features through a gating mechanism on the annotation-aware heterogeneous graph. The historical semantic agent synthesis module aggregates the non-target historical annotation semantics of the target user and the target file, and combines the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron. The relevance prediction and retrieval ranking module, based on the user graph representation, the file graph representation, and the historical semantic agent, predicts the relevance of the target user to the target file, and ranks the candidate files according to the relevance to generate retrieval results.
[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned document retrieval method based on a historical semantic proxy graph convolutional network.
[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned document retrieval method based on a historical semantic proxy graph convolutional network.
[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention utilizes annotation semantics and user-archive interaction structure for joint modeling within the same graph framework. By adaptively fusing node representations and edge semantics through a gating mechanism, it bridges the semantic gap of traditional literal matching and significantly improves the relevance of archive retrieval.
[0017] 2. When synthesizing historical semantic agents, this invention only aggregates the non-target historical annotation semantics of the target user and the target file, without using the annotation semantics of the target interaction pairs. This allows for the reconstruction of target semantics within the scope of information available during reasoning. Since the annotation semantics of the target interaction pairs are not used as reasoning input, data availability remains consistent between the training and online retrieval phases, fundamentally eliminating future information leakage and solving the core problem of target annotations being invisible during reasoning.
[0018] 3. The all-historical mean aggregation does not rely on target edge information and is not sensitive to the number of historical annotations. It is more stable than generative completion, effectively avoids semantic drift and noise amplification, and can still provide reliable relevance predictions in data sparsity and cold start scenarios.
[0019] 4. This invention uses classification-expectation decoding instead of direct regression, which preserves the correlation level distribution information, making the prediction results more detailed and the discrimination higher, and further improving the ability to characterize the utilization value of archives. Attached Figure Description
[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0021] Figure 1 This is a flowchart of an archive retrieval method based on a historical semantic proxy graph convolutional network. Detailed Implementation
[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0025] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.
[0026] Example 1: like Figure 1 As shown, this embodiment provides a document retrieval method based on a historical semantic proxy graph convolutional network, including the following steps: S1. Construct an annotation-aware heterogeneous graph, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features; Let the set of users be U, and the set of files be A = {a1, a2, ..., a...} N Each historical record is represented as a quadruple (u, a, r). ua ,t ua ), where u∈U is the user, a∈A is the file, and r ua ∈{1,2,3,4,5} represents the relevance (use value) level of user u to file a, t ua This corresponds to the annotation text.
[0027] A graph structure is constructed using elements from the user set U and the file set A as nodes and historical usage records as edges. For each historical usage record, a pre-trained language model (PLM) is used to process the usage annotation text t. ua Encoded as a fixed-dimensional edge semantic vector e ua And the semantic vector e of that edge ua As an edge feature, it is bound to the corresponding edge. Simultaneously, based on the relevance level r... ua Edges are differentiated, for example, by using relevance level as an edge type label or additional feature, so that the semantic vector of the edge between the same user-archive node pair carries the relevance level information of that use. This leads to the construction of an annotation-aware heterogeneous graph G=(U∪A,ε), where ε is its edge set.
[0028] In this embodiment, the pre-trained language model can use BERT to encode the annotation text into a 64-dimensional semantic vector; the dimension of the node and edge representations is set to 64.
[0029] S2. On the annotation-aware heterogeneous graph, node representations and edge features are fused through a gating mechanism to learn user graph representations and archive graph representations; A message propagated from user u along the annotated edge to file a is defined as: m (r) u→a =(W·h u ⊙σ(w p ⊤ e ua )+W r ·e ua ⊙σ(w r ⊤ e ua ))⊙c u Among them, h u For the initial node representation of user u, e ua Let W and W' be edge semantic vectors. r For a learnable transformation matrix, w p w r Let σ(·) be the gated vector, σ(·) be the Sigmoid function, ⊙ denote the element-wise product, and c u The normalization coefficient is used; the gating mechanism adaptively adjusts the fusion strength of structural information and textual information based on the semantic meaning of the annotations.
[0030] After stacking and aggregating L-layer message passing, we obtain the user graph representation and the archive graph representation: g u ,g a =Encoder(G,ε) Here, Encoder(·) represents a multi-layer gated graph convolutional encoder. In this embodiment, the graph convolution dropout is set to 0.7.
[0031] S3. Aggregate the non-target historical annotation semantics of the target user and the target file respectively, and combine the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron; During the actual retrieval and reasoning phase, the annotation text of the target interaction pair (target user - target file) does not exist, therefore the edge semantic vector e cannot be used. ua As input for reasoning, this embodiment designs a full-history semantic proxy branch to reconstruct the target semantics under the constraint that the target annotations are not visible during reasoning.
[0032] First, the mean aggregation is performed on all non-target historical annotation semantic vectors of the target user u and the target file a, respectively, to obtain the user-side historical semantic e. u -a and archival historical semantics e a -u : e u -a =(1 / |R u -a |)·Σ j e uj e a -u =(1 / |R a -u |)·Σ v e av Where e uj For the semantic vector of the j-th historical annotation of user u, e av Let |R| be the semantic vector of the vth historical annotation in file a. u -a |、|R a -u | represents the corresponding number of historical annotations. The non-target historical annotations refer to: historical annotations by the target user other than those related to the target file, and historical annotations by the target file other than those related to the target user. Mean aggregation is insensitive to the number of historical annotations and does not depend on any target-side information, naturally satisfying the availability constraint during inference.
[0033] Subsequently, the user graph representation of g will be used. u Archival diagram representation g a User-side historical semantics u -a Historical semantics of archives a -uIts commonalities and differences are jointly input into a multilayer perceptron to synthesize a historical semantic agent for target interaction. ua : ê ua =MLP([g u ||g a ||e u -a ||e a -u ‖(e u -a ⊙e a -u )‖|e u -a -e a -u |‖(g u ⊙g a )]) Among them, (e) u -a ⊙e a -u ) represents the common semantics between users and archive history, |e u -a -e a -u | represents the semantic differences between users and archive history, (g) u ⊙g a The structural interaction terms, MLP(·) is a multilayer perceptron. The resulting ê ua This is a semantic proxy that lacks target annotations, and does not use any target edge annotations throughout the process, fundamentally avoiding future information leakage and generative drift.
[0034] S4. Based on the user graph representation, file graph representation and historical semantic agent, predict the relevance of the target user to the target file, and sort the candidate files according to the relevance to generate search results.
[0035] To avoid the "mean collapse" problem caused by direct regression, this embodiment employs a classification-expectation decoding relevance prediction head. First, the logits value z for each relevance level is calculated using a multilayer perceptron scoring function. ua,k : z ua,k =φ([g u ||g a ‖ê ua ‖ē u -a ‖ē a -u ||(g u ⊙g a )‖ê ua ⊙((g u +g a) / 2)‖|ē u -a -ē a -u |]) k Where φ(·) is the multilayer perceptron scoring function, ê ua For historical semantic proxy, ē u -a , ē a -u For the aggregated user-side historical semantics and archive-side historical semantics, g u For user graph representation, g a For archival representation, k represents the k-th relevance level; Then, the relevance rank distribution P(r) is obtained using the Softmax function. ua =k): P(r ua =k)=exp(z ua,k ) / Σ j∈R exp(z ua,j ) Finally, the expected decoding yields continuous correlation prediction values: r̂ ua =Σ k∈R k·P(r ua =k) Where R={1,2,3,4,5} is the set of relevance levels.
[0036] It should be noted that, during online retrieval, for a given target user, a continuous relevance prediction value r̂ is calculated between them and each candidate file. ua According to the aforementioned correlation prediction value r̂ ua Sort the candidate files in descending order from high to low, and output the sorted list of candidate files as the file search results for the target user.
[0037] During training in this invention, the edge semantic vector e ua It serves only as a source of training supervision signals and never as inference input, thus strictly distinguishing between "training / test-obtainable" and "target-side specific" information in the data stream.
[0038] Example 2: This embodiment provides an archive retrieval system based on a historical semantic proxy graph convolutional network, including: The annotation-aware heterogeneous graph construction module constructs an annotation-aware heterogeneous graph, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features. The gated graph encoding module learns user graph representation and archive graph representation by fusing node representation and edge features through a gating mechanism on the annotation-aware heterogeneous graph. The historical semantic agent synthesis module aggregates the non-target historical annotation semantics of the target user and the target file, and combines the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron. The relevance prediction and retrieval ranking module, based on the user graph representation, the file graph representation, and the historical semantic agent, predicts the relevance of the target user to the target file, and ranks the candidate files according to the relevance to generate retrieval results.
[0039] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.
[0040] Example 3: An electronic device is provided for running the aforementioned "archive retrieval method based on historical semantic proxy graph convolutional network". The electronic device includes: a processor, a memory, and optional communication interface / display device / input device, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S4 of the method described in Embodiment 1, specifically including but not limited to: S1. Construct an annotation-aware heterogeneous graph, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features; S2. On the annotation-aware heterogeneous graph, node representations and edge features are fused through a gating mechanism to learn user graph representations and archive graph representations; S3. Aggregate the non-target historical annotation semantics of the target user and the target file respectively, and combine the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron; S4. Based on the user graph representation, file graph representation and historical semantic agent, predict the relevance of the target user to the target file, and sort the candidate files according to the relevance to generate search results.
[0041] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.
[0042] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to perform the method steps S1 to S4 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.
[0043] Experimental Validation: Since real-world archive annotation data often involves privacy concerns and publicly available samples are scarce, experiments were conducted on a publicly available benchmark dataset of comments, structurally isomorphic to the archive retrieval task, to objectively verify the effectiveness of this method. The publicly available dataset also consists of a "user-entry-text comment-rating" quadruple, strictly satisfying the key constraint that "target interactions are not visible to the text during inference," corresponding one-to-one with the "user-archive-annotation-relevance" setting in archive retrieval. In actual deployment, only the products in the publicly available data need to be replaced with archive entries, comments with annotations, and ratings with relevance levels; the model structure remains unchanged.
[0044] (1) Dataset The experiment used four publicly available benchmark datasets, the size and sparsity of which are shown in Table 1. Lower sparsity indicates sparser interactions and greater difficulty in modeling.
[0045] Table 1. Dataset Statistics (2) Evaluation indicators The root mean square error (RMSE) is used to evaluate the accuracy of correlation prediction, defined as: Where T is the test set, and r̂ ua With r ua These represent the predicted and actual relevance levels, respectively. A lower RMSE indicates a more accurate prediction and a more reliable search ranking.
[0046] (3) Experimental setup This method is implemented using PyTorch 2.4.0 and trained on a single NVIDIA Tesla V100 GPU. The Adam optimizer is used with a learning rate of 0.01, node and edge representation dimensions are both 64, and graph convolution dropout is 0.7. Historical semantic proxies employ full historical mean aggregation (without Top-K truncation), and the auxiliary contrast / distillation loss weights are reset to 0. The learning rate is halved if the validation metric does not improve for 20 consecutive epochs, and stopped early if it does not improve for 50 consecutive epochs. All methods are compared under the same data partitioning and evaluation protocol.
[0047] (4) Experimental results and analysis Table 2 shows the RMSE comparison of our method (HSP-GCN) with several representative baselines on four datasets (bold indicates the best for each column).
[0048] Table 2 Comparison of RMSE for different methods (lower is better) As shown in Table 2, our proposed method achieves the best RMSE on the Digital_Music, Toys_and_Games, and Clothing datasets, reducing the RMSE by approximately 2.1%, 3.6%, and 2.1% respectively compared to the strongest baseline RGCL, demonstrating a significant advantage in incorporating historical semantic proxies in low-to-medium density (relatively sparse) scenarios. On the densest CDs_and_Vinyl dataset, the traditional commenting method DAML achieves the best RMSE (0.9210). While our proposed method (0.9904) does not take first place, it is still significantly better than RGCL (1.0424) and LIRDRec, which are also graph contrastive learning methods. This indicates that text-based methods are more likely to benefit directly in scenarios with highly dense interactions and sufficient target text signals. The core value of our proposed method lies in robust relevance prediction in sparse scenarios where the target text is not visible.
[0049] To verify the contribution of each module, ablation experiments were conducted on Digital_Music and Toys_and_Games, and the results are shown in Table 3.
[0050] Table 3 Ablation test (RMSE, the smaller the better) It is evident that removing the full history proxy or removing the proxy branch both lead to a significant increase in RMSE (from 0.8632 to over 0.89 on Digital_Music), indicating that the full history semantic proxy is the key source of the accuracy improvement of this method, and the proxy branch and the graph encoder complement each other and are indispensable.
[0051] Graph convolution dropout has a significant impact on performance. Table 4 shows its sensitivity analysis on Digital Music.
[0052] Table 4. Graph Convolution Dropout Sensitivity (Digital_Music, RMSE) The optimal dropout value is 0.7; too low (0.3) will lead to overfitting, and too high (0.9) will lead to underfitting. Therefore, this method defaults to 0.7.
[0053] To examine the robustness of cold starts, cold / warm start samples were divided according to the threshold of the number of historical user interactions, and the results are shown in Table 5.
[0054] Table 5 Cold Start Robustness Analysis (Digital_Music, RMSE) The RMSE of cold-start samples (historical interaction count not exceeding the threshold) is generally higher than that of warm-start samples, but the gap between the two gradually narrows as the threshold is relaxed. Thanks to the fact that the full-history semantic agent can still provide effective target semantic reconstruction when the history is limited, this method still maintains reasonable prediction accuracy for cold-start users.
[0055] Further comparisons of different combinations of historical scope and agency strategies are shown in Table 6.
[0056] Table 6 Comparison of Historical Scope and Proxy Strategy (Digital_Music, RMSE) The combination of full-history proxy + mean aggregation + classification expectation decoding (i.e., this invention) achieves optimal results. It is worth noting that: using target-aware (TA) proxy is worse (0.8930) because target edge annotations are unavailable during inference, and forcibly using them will introduce future information leakage and drift; replacing the classification expectation head with the regression head slightly reduces accuracy (0.8670); using only the training set history instead of the entire history (0.8902) is also worse than the full-history scheme, which fully demonstrates the value of incorporating all non-target historical annotations for reconstructing target semantics.
[0057] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.
[0058] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0059] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A method for retrieving archives based on a historical semantic proxy graph convolutional network, characterized in that, Includes the following steps: Construct a heterogeneous graph for annotation awareness, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features; On the annotation-aware heterogeneous graph, node representations and edge features are fused through a gating mechanism to learn user graph representations and archive graph representations; The non-target historical annotation semantics of the target user and the target file are aggregated and combined with the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron. Based on the user graph representation, the file graph representation, and the historical semantic agent, the relevance of the target user to the target file is predicted, and the candidate files are ranked according to the relevance to generate search results.
2. The document retrieval method based on historical semantic proxy graph convolutional network according to claim 1, characterized in that, The specific steps for constructing the annotation-aware heterogeneous graph include: using elements from the user set and the archive set as nodes, and using historical usage records as edges to construct the graph structure; where each historical usage record is represented as a quadruple (u, a, r). ua ,t ua ), u represents the user, a represents the file, r ua For relevance level, t ua To utilize the annotated text; a pre-trained language model is used to process the annotated text. ua Encoded as edge semantic vector e ua and the edge semantic vector e ua As an edge feature, it is bound to the corresponding edge; according to the relevance level r ua The edges are distinguished to construct the annotation-aware heterogeneous graph.
3. The document retrieval method based on a historical semantic proxy graph convolutional network according to claim 1, characterized in that, The method for obtaining the user graph representation and the archive graph representation is as follows: the message propagated from user u to archive a is defined as: m (r) u→a =(W·h u ⊙σ(w p ⊤ have been ua )+W r ·have been ua ⊙σ(w r ⊤ have been ua ))⊙c u Among them, h u For the initial node representation of user u, e ua Let W and W' be edge semantic vectors. r For a learnable transformation matrix, w p w r Let σ(·) be the gated vector, σ(·) be the Sigmoid function, ⊙ denote the element-wise product, and c u The normalization coefficients are used; after stacking L layers of message passing and aggregation, the user graph representation and archive graph representation are obtained: g u ,g a =Encoder(G,ε) Where G is the annotation-aware heterogeneous graph, ε is its edge set, and Encoder(·) represents a multi-layer gated graph convolutional encoder.
4. The document retrieval method based on a historical semantic proxy graph convolutional network according to claim 1, characterized in that, In the step of aggregating the non-target historical annotation semantics of both the target user u and the target file a, the mean aggregation is performed on all non-target historical annotation semantic vectors of both the target user u and the target file a to obtain the user-side historical semantics e. u -a and archival historical semantics e a -u : and u -a =(1 / |R u -a |)·Σ j and uj and a -u =(1 / |R a -u |)·Σ v and av Among them, e uj For the semantic vector of the j-th historical annotation of user u, e av Let |R| be the semantic vector of the vth historical annotation in file a. u -a |、|R a -u | represents the number of historical annotations.
5. The document retrieval method based on a historical semantic proxy graph convolutional network according to claim 1, characterized in that, Synthesize historical semantic proxies for target interactions using a multilayer perceptron. ua The method is as follows: ê ua =MLP([g u ‖g a ‖e u -a ‖e a -u ‖(e u -a ⊙e a -u )‖|e u -a -e a -u |‖(g u ⊙g a )]) Among them, g u For user graph representation, g a For archival representation, e u -a For user-side historical semantics, e a -u For the historical semantics of the archives, (e u -a ⊙e a -u ) represents the common semantics between users and archive history, |e u -a -e a -u | represents the semantic differences between users and archive history, (g) u ⊙g a )Structured interaction items.
6. The document retrieval method based on a historical semantic proxy graph convolutional network according to claim 1, characterized in that, The method for predicting the relevance of a target user to a target profile is as follows: First, the logits value z for each relevance level is calculated using a multilayer perceptron scoring function. ua,k : z ua,k =φ([g u ‖g a ‖ê ua ‖ē u -a ‖ē a -u ‖(g u ⊙g a )‖ê ua ⊙((g u +g a ) / 2)‖|ē u -a -ē a -u |]) k Where φ(·) is the multilayer perceptron scoring function, ê ua For historical semantic proxy, ē u -a , ē a -u For the aggregated user-side historical semantics and archive-side historical semantics, g u For user graph representation, g a For archival representation, k represents the k-th relevance level; Then, the relevance rank distribution P(r) is obtained using the Softmax function. ua =k): P(r ua =k)=exp(z ua,k ) / Σ j∈R exp(z ua,j ) Finally, the expected decoding yields continuous correlation prediction values: r̂ ua =Σ k∈R k·P(r ua =k) Where R is the set of relevance levels.
7. The document retrieval method based on a historical semantic proxy graph convolutional network according to claim 1, characterized in that, The step of sorting candidate files according to the relevance and generating search results specifically involves: During online retrieval, for a given target user, calculating the continuous relevance prediction value r̂ between them and each candidate file. ua According to the aforementioned correlation prediction value r̂ ua Sort the candidate files in descending order from high to low, and output the sorted list of candidate files as the file search results for the target user.
8. An archive retrieval system based on a historical semantic proxy graph convolutional network, characterized in that, include: The annotation-aware heterogeneous graph construction module constructs an annotation-aware heterogeneous graph, with users and archives as nodes, usage relationships as edges, and the semantic vectors of user annotations as edge features. The gated graph encoding module learns user graph representation and archive graph representation by fusing node representation and edge features through a gating mechanism on the annotation-aware heterogeneous graph. The historical semantic agent synthesis module aggregates the non-target historical annotation semantics of the target user and the target file, and combines the user graph representation and the file graph representation to synthesize the historical semantic agent of the target interaction through a multilayer perceptron. The relevance prediction and retrieval ranking module, based on the user graph representation, the file graph representation, and the historical semantic agent, predicts the relevance of the target user to the target file, and ranks the candidate files according to the relevance to generate retrieval results.
9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the document retrieval method based on a historical semantic proxy graph convolutional network as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the archive retrieval method based on a historical semantic proxy graph convolutional network as described in any one of claims 1-7.
