A semantic retrieval method and system for an agricultural policy document

Abstract

The application discloses a kind of semantic retrieval method and system of agricultural policy document, method includes: obtaining multiple-user policy retrieval instruction, generates initial query semantic vector and is mapped to semantic index library and recalls initial candidate document set;Policy semantic evolution direction field is constructed using index library document semantic vector and promulgation time stamp, and semantic overlap degree is determined by bidirectional evolution consistency reconciliation score, and user intent topology is constructed;Select the highest node of connectivity in topological map as anchor point user, determine reference semantic anchor point;Calculate the semantic deviation of candidate document and the comprehensive score weighted by time factor;If not up to standard, then based on the mean vector of hit and deviated document, construct semantic calibration ray, and iteratively reconstruct query vector to up to standard in reverse disturbance;Standard document is output in descending order of score.The application introduces policy semantic evolution direction field, improves semantic overlap degree from static similarity to dynamic evolution consistency test, improves retrieval accuracy and timeliness.

Description

Technical Field

[0001] This invention belongs to the field of information retrieval technology, and in particular relates to a semantic retrieval method and system for agricultural policy documents. Background Technology

[0002] Agricultural policy documents serve as crucial information carriers guiding agricultural production, rural development, and increasing farmers' income. With the continuous issuance and updating of various agricultural policies by government departments at all levels, the number of policy documents has increased dramatically, covering multiple areas such as planting subsidies, livestock support, land transfer, agricultural insurance, and rural revitalization. How users can quickly and accurately retrieve policies relevant to their needs when faced with this massive amount of policy documents has become a pressing technical problem that needs to be solved.

[0003] Existing semantic retrieval methods typically encode user queries and documents into semantic vectors, and then perform matching and recall by calculating the cosine similarity between these vectors. However, agricultural policy documents have a significant time-sensitive characteristic—policies are revised, repealed, or replaced over time, and there is an inherent evolutionary logic between old and new policies. Existing semantic retrieval methods fail to fully utilize the evolutionary patterns between policy documents, nor do they consider the correlation between the intentions of different users, resulting in insufficient performance in terms of both timeliness and relevance of search results.

[0004] Furthermore, existing technologies typically measure the semantic relationships between document sets by calculating the similarity between each pair of documents and taking the arithmetic mean. This method can only reflect static semantic content overlap and cannot capture the deep connections between document sets in the dimension of policy evolution. Summary of the Invention

[0005] The purpose of this application is to provide a semantic retrieval method, system, electronic device, and storage medium for agricultural policy documents, in order to solve the problem that the prior art does not fully utilize the laws of policy evolution and the correlation of user intent.

[0006] In a first aspect, the present invention provides a semantic retrieval method for agricultural policy documents, comprising:

[0007] Obtain policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user;

[0008] Obtain a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor;

[0009] Each initial query semantic vector is mapped to the semantic index library for semantic retrieval, resulting in an initial candidate document set corresponding to each user. A user intent topology graph is then constructed based on the semantic overlap between the initial candidate document sets.

[0010] The node with the highest connectivity in the user intent topology graph is selected as the anchor user node, the initial candidate document set corresponding to the anchor user node is used as the anchor retrieval queue, and the semantic center of the anchor retrieval queue is determined as the reference semantic anchor.

[0011] Calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and adjust the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document.

[0012] Determine whether the comprehensive semantic matching score falls within the preset target score range;

[0013] If it does not fall into the target range, then based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, a reverse semantic perturbation is applied to the initial query semantic vector of the anchor user node to generate a reconstructed query semantic vector, and the semantic recall and comprehensive semantic matching score are recalculated based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range.

[0014] Candidate documents whose comprehensive semantic matching scores fall within the target score range in the anchor point retrieval queue, and the initial candidate document set corresponding to other nodes in the user intent topology graph, are sorted from high to low according to their comprehensive semantic matching scores to obtain the final retrieval results. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

[0015] Secondly, the present invention provides a semantic retrieval system for agricultural policy documents, comprising:

[0016] The first acquisition module is configured to acquire policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user.

[0017] The second acquisition module is configured to acquire a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor.

[0018] The construction module is configured to map each initial query semantic vector to the semantic index library for semantic retrieval, obtain an initial candidate document set corresponding to each user, and construct a user intent topology graph based on the semantic overlap between the initial candidate document sets.

[0019] The selection module is configured to select the node with the highest connectivity in the user intent topology graph as the anchor user node, use the initial candidate document set corresponding to the anchor user node as the anchor retrieval queue, and determine the semantic center of the anchor retrieval queue as the reference semantic anchor.

[0020] The correction module is configured to calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and to perform weighted correction on the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document.

[0021] The judgment module is configured to determine whether the comprehensive semantic matching score falls within a preset target score range;

[0022] The generation module is configured to, if not falling into the target range, apply reverse semantic perturbation to the initial query semantic vector of the anchor user node based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, generate a reconstructed query semantic vector, and re-execute the semantic recall and comprehensive semantic matching score calculation based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range.

[0023] The output module is configured to sort candidate documents in the anchor point retrieval queue whose comprehensive semantic matching score falls within the target score interval, as well as candidate documents in the initial candidate document set corresponding to other nodes in the user intent topology graph whose comprehensive semantic matching score falls within the target score interval, according to the comprehensive semantic matching score from high to low, to obtain the final retrieval result. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

[0024] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the semantic retrieval method for agricultural policy documents according to any embodiment of the present invention.

[0025] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the semantic retrieval method for agricultural policy documents according to any embodiment of the present invention.

[0026] The semantic retrieval method and system for agricultural policy documents disclosed in this application obtains policy retrieval instructions from multiple users and generates initial query semantic vectors. Each initial query semantic vector is mapped to a pre-built semantic index for semantic recall, resulting in an initial candidate document set. A policy semantic evolution direction field is constructed using the semantic vectors of all documents in the index and the promulgation timestamps. The semantic overlap is determined by calculating the bidirectional evolutionary consistency score and harmonic mean between any two document sets in the evolution direction field, and a user intent topology graph is constructed based on this. The node with the highest connectivity in the topology graph is selected as the anchor user, and its candidate document set is used as the anchor. The system retrieves a queue of reference semantic anchors; calculates the semantic deviation between each candidate document and the reference semantic anchor, and obtains a comprehensive semantic matching score after weighting and correction by a timeliness decay factor; if the score does not reach the preset target range, a semantic calibration ray is constructed based on the mean vector of the hit and deviation documents, and a reverse semantic perturbation is applied to the query vector of the anchor user, iteratively reconstructing until the score reaches the target; finally, the documents that meet the target for each user are output in sorted order by score; by introducing a policy semantic evolution direction field, the semantic overlap is upgraded from static content similarity to a consistency check of dynamic evolution logic, which significantly improves the accuracy and timeliness of agricultural policy retrieval. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0028] Figure 1 A flowchart illustrating a semantic retrieval method for agricultural policy documents according to an embodiment of the present invention;

[0029] Figure 2 This is a structural block diagram of a semantic retrieval system for agricultural policy documents provided in an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0032] Please see Figure 1 The diagram illustrates a flowchart of a semantic retrieval method for agricultural policy documents according to this application.

[0033] like Figure 1 As shown, the semantic retrieval method for agricultural policy documents specifically includes the following steps:

[0034] Step S101: Obtain policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user.

[0035] In this step, each policy retrieval instruction is sequentially input into the pre-trained syntactic parsing model, and the dependency syntax tree corresponding to each policy retrieval instruction is output. The dependency syntax tree contains word nodes and dependency relationship edges between word nodes.

[0036] In the dependency syntax tree, the core predicate verb is determined as the root node. Starting from the root node, the noun phrases in the subject branch are extracted as policy subject vector components, the noun phrases in the object branch are extracted as policy object vector components, and the time phrases and place phrases in the adverbial branch are extracted as time constraint vector components and space constraint vector components, respectively.

[0037] The policy subject vector component, policy object vector component, time constraint vector component, and spatial constraint vector component are weighted and concatenated according to preset dimension weights to generate the initial query semantic vector corresponding to each user.

[0038] In one specific embodiment, assume that the system simultaneously receives policy retrieval instructions submitted by three users, namely:

[0039] User A's search query: "Rice planting subsidy policy issued by the Department of Agriculture and Rural Affairs of a certain province in 2025";

[0040] User B's search query: "Minimum purchase price policy for wheat in a certain province from 2024 to 2026";

[0041] User C's search query: "Subsidy scheme for conservation tillage of black soil in a certain region during a certain period".

[0042] Sub-step S1011: Syntactic parsing and dependency syntax tree construction

[0043] The three policy retrieval instructions are sequentially input into a pre-trained syntactic parsing model. The syntactic parsing model can be a deep neural network dependency parser trained on a large-scale Chinese corpus, such as a model pre-trained on the BERT architecture and fine-tuned for Chinese dependency parsing tasks. The model performs word segmentation, part-of-speech tagging, and dependency relation prediction on the input text, outputting the corresponding dependency syntax tree.

[0044] Taking user A's search query "rice planting subsidy policy issued by the Department of Agriculture and Rural Affairs of a certain province in 2025" as an example, the output dependency syntax tree contains the following word nodes and dependency relation edges:

[0045] Root node: The core predicate verb "publish" (marked as ROOT);

[0046] Subject branch: Starting from "release", follow the subject-verb relationship and connect the noun phrase "a certain province's Department of Agriculture and Rural Affairs" (marked as the subject);

[0047] Object branch: Starting from "release", follow the verb-object relationship to connect the noun phrase "policy" (marked as the object), then follow the attributive relationship to connect the modifier "subsidy", and then follow the attributive relationship to connect "rice planting";

[0048] Adverbial branch: Starting from "released", follow the adverbial-head relationship and connect the time phrase "2025" (marked as a time adverbial).

[0049] Similarly, the search commands of users B and C can be parsed to obtain the corresponding dependency syntax tree structures.

[0050] Sub-step S1012: Semantic component extraction

[0051] In the dependency syntax tree, the core predicate verb is identified as the starting point of traversal (root node). If the core predicate verb is not present in the search command (e.g., the command is a pure noun phrase), then the core noun is identified as the root node. Starting from the root node, semantic components of four dimensions are extracted according to the preset dependency relationship path.

[0052] Example of User A:

[0053] Policy subject vector component: Extract the noun phrase "a certain province's Department of Agriculture and Rural Affairs" from the subject branch of the root node. Input this phrase into the pre-trained word vector model to obtain a 128-dimensional real-valued vector, which serves as the policy subject vector component.

[0054] Policy object vector component: Extract the noun phrase "rice planting subsidy policy" from the object branch of the root node. Encode this phrase as a 128-dimensional real-valued vector, which serves as the policy object vector component.

[0055] Time constraint vector component: Extract the time phrase "2025" along the adverbial branch of the root node. Encode this phrase as a 128-dimensional real-valued vector, which serves as the time constraint vector component.

[0056] Spatial constraint vector component: Extract the location phrase "a certain province" along the adverbial or attributive branch of the root node. Encode this phrase as a 128-dimensional real-valued vector, which serves as the spatial constraint vector component.

[0057] For missing dimensions in a search query, the corresponding vector components are filled with zero vectors. For example, in user B's search query "Minimum purchase price policy for wheat in a certain province from 2024 to 2026", the policy subject is not explicitly mentioned, so the policy subject vector component is filled with a 128-dimensional zero vector; in user C's search query "Subsidy scheme for protective tillage of black soil in a certain region during a certain period", the policy subject is not explicitly mentioned, and the time phrase "a certain period" needs to be mapped to a specific year interval through a preset time phrase parsing module, taking the starting year of the interval as the time constraint.

[0058] After extraction, each user's search command obtains four independent 128-dimensional semantic vector components.

[0059] Sub-step S1013: Dimension-weighted splicing

[0060] The four vector components are weighted and concatenated according to preset dimensional weights, which are set based on the distinguishability of different dimensions in agricultural policy retrieval tasks. In this embodiment, the dimensional weights are set as follows: the weight of the policy object vector component is 0.4, the weight of the policy subject vector component is 0.2, the weight of the time constraint vector component is 0.2, and the weight of the spatial constraint vector component is 0.2. The weights are set based on the fact that in agricultural policy retrieval, users are most concerned with the field and object (i.e., the object) involved in the policy, followed by the publishing subject, time, and location. The sum of all weights is 1.

[0061] The specific operation of weighted concatenation is as follows: multiply the policy subject vector component by 0.2, the policy object vector component by 0.4, the time constraint vector component by 0.2, and the spatial constraint vector component by 0.2. Then, concatenate the four weighted vectors along their respective dimensions to obtain a 512-dimensional composite vector. This 512-dimensional composite vector is the initial query semantic vector for user A. The initial query semantic vectors for users B and C can be obtained similarly.

[0062] The dependency syntax semantic vector construction method based on four-dimensional decomposition proposed in this step has the following advantages compared with existing techniques that encode the entire retrieval command into a single vector:

[0063] By using dependency parsing, agricultural policy retrieval commands are deconstructed into four semantic dimensions: policy subject, policy object, time constraint, and space constraint. These dimensions are encoded independently and then structurally fused together. This allows the generated initial query semantic vector to more accurately depict various aspects of the user's query intent in the semantic space. Compared to the problem of the four dimensions being mixed and difficult to distinguish in the whole sentence encoding method, this improves the expressive accuracy of the semantic vector from the source.

[0064] By introducing a dimension weighting mechanism, the retrieval system can flexibly adjust the contribution of different semantic dimensions according to the actual retrieval needs in the field of agricultural policy. For example, giving higher weight to the policy object dimension can enable the semantic recall process to prioritize matching the policy fields and objects that users are interested in, thereby improving the relevance ranking quality of the retrieval results.

[0065] For missing semantic dimensions in the retrieval command, zero vector padding is used to ensure that the generated initial query semantic vector does not introduce noise or bias due to missing information while maintaining a fixed dimension format, thus ensuring the stability and consistency of subsequent semantic recall calculations.

[0066] The four-dimensional decomposition vector structure provides traceable semantic granularity for subsequent steps such as user intent topology graph construction, semantic deviation calculation, and reverse semantic perturbation. When the search results need to be optimized, the system can trace back to the specific semantic dimension and analyze whether the expression of the subject, object, time, or space dimension caused the search deviation, thereby achieving more accurate query reconstruction.

[0067] Step S102: Obtain a pre-built hierarchical semantic index library for policy documents. Each policy document in the semantic index library is associated with and stored as a document semantic vector and a time-effect decay factor.

[0068] In this step, for each policy document collected, the following metadata fields are extracted: document title, full name of the issuing authority, level of the issuing authority (national, provincial, municipal, county), document type (law, administrative regulation, departmental rule, local regulation, local government rule, normative document, policy interpretation), promulgation timestamp (accurate to day), implementation timestamp, repeal timestamp (if repealed), and full text or summary text of the document.

[0069] For each policy document, the document title is concatenated with the full text or summary text, and then input into a pre-trained semantic encoder to generate a fixed-dimensional semantic vector for the document.

[0070] In this embodiment, the semantic encoder uses a pre-trained language model based on the BERT architecture, which is then fine-tuned using a large-scale Chinese policy corpus. The specific operation is as follows: the document title and full text are concatenated into an input sequence, the title and body are separated by a special delimiter, and the input is given to the fine-tuned BERT model; the hidden state vector corresponding to the label output by the last layer of the model is taken as the semantic representation of the document; this vector is mapped to a document semantic vector of a preset dimension (e.g., 512 dimensions) through a fully connected layer.

[0071] For lengthy policy documents, the entire text can be divided into segments, each segment encoded separately, and the average of the vectors from each segment taken as the final semantic vector of the document.

[0072] The characteristic of agricultural policy documents decaying over time needs to be quantified during the index building stage. For each policy document, a time-decrease factor is calculated based on the promulgation timestamp, and the time-decrease factor is associated with the document's semantic vector and stored in the semantic index library.

[0073] The calculated time-depreciation factor is associated with the corresponding document semantic vector and document metadata and stored in the index record of the semantic index library, forming a four-tuple storage structure of "document identifier - document semantic vector - time-depreciation factor - metadata field".

[0074] Step S103: Map each initial query semantic vector to the semantic index library for semantic retrieval to obtain an initial candidate document set corresponding to each user, and construct a user intent topology map based on the semantic overlap between the initial candidate document sets.

[0075] In this step, the vector cosine similarity between each initial query semantic vector and each document semantic vector in the semantic index is calculated. Policy documents with a vector cosine similarity exceeding a preset recall threshold are recalled and used as the initial candidate document set for each user.

[0076] Using the document semantic vectors of all policy documents in the semantic index library and the promulgation timestamps of all policy documents, a policy semantic evolution direction field is constructed;

[0077] For the first initial candidate document set corresponding to the first user and the second initial candidate document set corresponding to the second user, extract all cross-set document pairs that satisfy the first time sequence condition, wherein the first time sequence condition is: the first document belongs to the first initial candidate document set, the second document belongs to the second initial candidate document set, and the publication timestamp of the second document is later than the publication timestamp of the first document;

[0078] Extract all cross-set document pairs that satisfy the second time sequence condition, wherein the second time sequence condition is: the first document belongs to the first initial candidate document set, the second document belongs to the second initial candidate document set, and the publication timestamp of the second document is earlier than the publication timestamp of the first document;

[0079] For a certain cross-set document pair, calculate a certain actual evolution vector pointing from the document semantic vector of a certain first document to the document semantic vector of a certain second document, and perform path integration in the policy semantic evolution direction field according to the certain time difference between the promulgation timestamp of a certain first document and the promulgation timestamp of a certain second document, to obtain a certain expected evolution vector corresponding to the certain cross-set document pair.

[0080] Calculate a cosine similarity between a certain actual evolution vector and a certain expected evolution vector to obtain a certain evolutionary fit degree of a certain cross-set document pair, and calculate the arithmetic mean of the evolutionary fit degrees of all cross-set document pairs that satisfy the first temporal condition to obtain the first directional evolutionary consistency score from the first initial candidate document set to the second initial candidate document set.

[0081] Calculate the arithmetic mean of the evolutionary fit of all cross-set document pairs that satisfy the second temporal condition to obtain the second directional evolutionary consistency score from the second initial candidate document set to the first initial candidate document set;

[0082] Calculate the harmonic mean of the first directional evolutionary consistency score and the second directional evolutionary consistency score, and determine the harmonic mean as the semantic overlap between the first candidate document set and the second candidate document set;

[0083] Using each user as a node and the semantic overlap as the weight of the edges between nodes, an initial complete graph is constructed. Then, minimum spanning tree pruning is performed on the initial complete graph, retaining only edges with weights greater than a preset edge weight threshold, to obtain a sparse user intent topology graph.

[0084] In one specific embodiment, sub-step S1031: semantic recall, obtaining the initial candidate document set corresponding to each user.

[0085] Calculate the vector cosine similarity between the initial query semantic vectors of users A, B, and C and the semantic vectors of each document in the semantic index. Set the preset recall threshold to 0.65. After calculating the vector cosine similarity for each document in the semantic index, select policy documents with a vector cosine similarity exceeding 0.65 to form the initial candidate document sets for users A, B, and C, respectively.

[0086] Suppose the recall results are as follows: User A's initial candidate document set contains 5 documents, identified by document identifiers A1, A2, A3, A4, and A5, with promulgation timestamps ranging from 2019 to 2024; User B's initial candidate document set contains 6 documents, identified by document identifiers B1, B2, B3, B4, B5, and B6, with promulgation timestamps ranging from 2018 to 2025; User C's initial candidate document set contains 4 documents, identified by document identifiers C1, C2, C3, and C4, with promulgation timestamps ranging from 2020 to 2023. There is some overlap between the three sets, which means that some policy documents may be recalled by multiple users simultaneously. However, this direct intersection at the document level does not fully reflect the evolutionary relationship between user intentions.

[0087] Sub-step S1032: Constructing the policy semantic evolution direction field

[0088] By utilizing the document semantic vectors and promulgation timestamps of all N policy documents in the semantic index library, a policy semantic evolution direction field is constructed. This policy semantic evolution direction field is a vector field defined in a joint space spanned by the semantic space and the time axis, used to describe the directional patterns of policy semantic evolution over time.

[0089] The construction process consists of the following steps:

[0090] Sort all N policy documents in the semantic index from earliest to latest according to their promulgation timestamps to obtain a time-ordered document sequence. After sorting, the promulgation timestamp of the i-th policy document is: The document semantic vector is ,satisfy .

[0091] For each pair of temporally adjacent documents in the sorted document sequence, calculate the difference between the document semantic vectors to obtain the instantaneous semantic evolution vector:

[0092] ,

[0093] In the formula, Let i be the instantaneous semantic evolution vector between the i-th pair of temporally adjacent documents. Let i be the document semantic vector of the (i+1)th policy document. Let be the document semantic vector of the i-th policy document;

[0094] Simultaneously record the time difference between adjacent documents:

[0095] ,

[0096] In the formula, Let i be the time difference between the i-th pair of time-adjacent documents. This is the timestamp for the issuance of the (i+1)th policy document. This is the promulgation timestamp of the i-th policy document, sorted from earliest to latest by promulgation timestamp;

[0097] Perform local weighted regression on the time axis on all N-1 instantaneous semantic evolution vectors to obtain a smooth evolution direction vector field that can be continuously queried at any time point. Specifically, at any query time point t on the time axis, select the instantaneous semantic evolution vectors of all adjacent document pairs within a certain window width before and after t, and perform a weighted average using a Gaussian kernel function as the weight to obtain the smooth evolution direction vector at time t. The Gaussian kernel weight function is:

[0098] ,

[0099] ,

[0100] In the formula, The Gaussian kernel weights for querying time point t with respect to intermediate time points are defined, with values ​​ranging from (0,1]. Let be the midpoint of the i-th pair of time-adjacent documents. The bandwidth parameter of the Gaussian kernel. It is an exponential function with the natural constant e as its base;

[0101] After local weighted regression, the policy semantic evolution direction field is defined as a continuous queryable vector field. Given a starting time point... A starting semantic vector and a time difference The expected evolution vector is obtained by path integration along the evolution direction field. The numerical calculation method of the path integral is as follows: [The time interval is...] Divided into equal parts Each step is 0.1 years; from Beginning, in the Step, query Evolution direction vector at time step ,Will Move one step distance along the evolution direction vector and update ; After the step is completed, the extrapolated semantic vector is obtained. Then the expected evolution vector ;

[0102] Sub-step S1033: Extract cross-set document pairs that satisfy the first temporal condition.

[0103] For user A's initial candidate document set (containing documents A1 to A5, a total of 5 documents) and user B's initial candidate document set (containing documents B1 to B6, a total of 6 documents), extract all cross-set document pairs that satisfy the first temporal condition. The first temporal condition is: the first document belongs to user A's initial candidate document set, the second document belongs to user B's initial candidate document set, and the publication timestamp of the second document is later than the publication timestamp of the first document, i.e. .

[0104] To illustrate with a specific document example: User A's document A2 was published on March 15, 2020, and User B's document B3 was published on August 20, 2022, which satisfies... Therefore, the cross-set document pair (A2, B3) satisfies the first temporal condition and is included in the first direction of the cross-set document pair set. Conversely, user A's document A5 has an issue date of January 10, 2024, and user B's document B1 has an issue date of June 5, 2019, satisfying the condition. Therefore, (A5, B1) does not meet the first time sequence condition and is not included.

[0105] Iterate through all pairwise combinations of 5 documents from user A's set and 6 documents from user B's set, resulting in a total of 5 × 6 = 30 document pairs. Each pair is then checked to see if it meets the first temporal condition. In this embodiment, after filtering, a total of [number missing] documents are obtained. A cross-collection document pair that satisfies the first temporal condition.

[0106] Sub-step S1034: Extract cross-set document pairs that satisfy the second temporal condition.

[0107] The second timing condition is: the first document belongs to user A's initial candidate document set, the second document belongs to user B's initial candidate document set, and the publication timestamp of the second document is earlier than the publication timestamp of the first document, i.e. .

[0108] Symmetrical to sub-step S1033, all 30 document pairs are traversed, and document pairs that satisfy the second temporal condition are selected. In this embodiment, a total of [number] document pairs are obtained after selection. There are 3 cross-set document pairs that satisfy the second time sequence condition. The remaining 3 document pairs have two documents with the same publication timestamp, thus not satisfying any condition and are therefore excluded.

[0109] Sub-step S1035: Calculate the evolutionary fit of a cross-set document pair.

[0110] Taking the cross-set document pair (A2, B3) as an example, the calculation process of evolutionary fit is explained in detail:

[0111] Document A2's publication timestamp The release date is March 15, 2020, the release date of document B3. The date is August 20, 2022, with a time difference. Year;

[0112] The actual evolution vector is the vector difference between the document semantic vector of document A2 and the document semantic vector of document B3:

[0113] ,

[0114] In the formula, This is the actual evolution vector, representing the actual semantic change from document A2 to document B3. The document semantic vector of document A2. This is the document semantic vector for document B3;

[0115] The expected evolution vector is calculated as follows: As the initial semantic vector, The year is used as the time difference, and the integral is performed along the positive time direction in the policy semantic evolution direction field. The time interval [March 15, 2020, August 20, 2022] is divided into equal parts. Step (step size 0.1-0.1 years, rounded up); from To begin, at step k, query the time point. Evolution direction vector at the location ,renew After 25 steps, the extrapolated semantic vector is obtained. The expected evolution vector is:

[0116] ,

[0117] In the formula, This is the expected evolution vector for the cross-set document pair (A2, B3);

[0118] calculate and The cosine similarity between them serves as the evolutionary fit between cross-set document pairs (A2, B3);

[0119] For all For each cross-set document pair that satisfies the first temporal condition, the above calculation process is performed sequentially, resulting in 18 evolutionary fit values. Similarly, for all... Nine evolutionary fit values ​​were calculated for each cross-set document pair that met the second temporal condition.

[0120] First-direction evolutionary consistency score Defined as the arithmetic mean of the evolutionary fit of all cross-set document pairs that satisfy the first temporal condition:

[0121] ,

[0122] In the formula, Let be the actual evolution vector of the k-th cross-set document pair that satisfies the first temporal condition. Let be the expected evolution vector of the k-th cross-set document pair that satisfies the first temporal condition. The evolutionary fit of the k-th cross-set document pair;

[0123] Second-direction evolutionary consistency score Defined as the arithmetic mean of the evolutionary fit of all cross-set document pairs that satisfy the second temporal condition:

[0124] ,

[0125] In the formula, For the first The degree of evolutionary fit of cross-collection document pairs, For the first The actual evolution vector of a cross-set document pair that satisfies the second temporal condition. For the first The expected evolution vector of a cross-set document pair that satisfies the second temporal condition;

[0126] Sub-step S1038: Calculate semantic overlap

[0127] The semantic overlap between the initial candidate document set of user A and the initial candidate document set of user B is defined as the harmonic mean of the first-direction evolutionary consistency score and the second-direction evolutionary consistency score:

[0128] ,

[0129] In the formula, The semantic overlap between the initial candidate document set of user A and the initial candidate document set of user B;

[0130] In this embodiment, Significantly lower This indicates that when user B's intention is derived from user A's intention according to the policy evolution law, there is a certain degree of logical break in the evolution. If the arithmetic mean (approximately 0.732) is used, this break will be more pronounced. This approach partially masks the underlying issues; however, by using the harmonic mean (0.727), the lower score has a more significant negative impact on the result, allowing the final semantic overlap score to more honestly reflect the asymmetry of the bidirectional evolutionary logic between the two document sets. This characteristic aligns with the strict definition of semantic overlap in this scheme—a high level of semantic overlap can only be achieved when both document sets exhibit high evolutionary consistency in both bidirectional time dimensions, ensuring that semantic overlap reflects the essential policy logic connection rather than accidental one-way semantic similarity.

[0131] Following the same procedure, calculate the semantic overlap between user A and user C, and between user B and user C, respectively.

[0132] Using users A, B, and C as three nodes, and the semantic overlap calculated in sub-step S1038 as the weight of the edges between the nodes, an initial complete graph is constructed. Then, a minimum spanning tree pruning operation is performed on the initial complete graph to obtain the user intent topology graph.

[0133] Step S104: Select the node with the highest connectivity in the user intent topology graph as the anchor user node, use the initial candidate document set corresponding to the anchor user node as the anchor retrieval queue, and determine the semantic center of the anchor retrieval queue as the reference semantic anchor.

[0134] In this step, the mean of the document semantic vectors of all candidate documents in the anchor retrieval queue is calculated as the initial semantic center;

[0135] Using the initial semantic center as the initial cluster center, perform K-means clustering iteration on the document semantic vectors of all candidate documents in the anchor point retrieval queue until the cluster centers converge to obtain the target cluster center;

[0136] Select the document semantic vectors of a preset number of candidate documents that are closest to the target cluster center after clustering, calculate the weighted average of the preset number of document semantic vectors, and determine the weighted average as the reference semantic anchor point, wherein the weight of the weighted average is inversely proportional to the distance of each candidate document from the target cluster center.

[0137] Step S105: Calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and perform weighted correction on the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document.

[0138] In this step, the promulgation timestamps of candidate documents in the anchor retrieval queue and the time coordinates of the reference semantic anchors in the pre-constructed policy semantic evolution direction field are obtained;

[0139] The document semantic vector of the candidate document is input into the policy semantic evolution direction field, and semantic evolution extrapolation is performed along the positive time direction to generate the first discrete sampling point sequence of the candidate document within the preset extrapolation time interval;

[0140] The reference semantic anchor is input into the policy semantic evolution direction field, and semantic evolution extrapolation is performed along the positive time direction to generate a second discrete sampling point sequence of the reference semantic anchor within a preset extrapolation time interval;

[0141] For each sampling time, the corresponding first sampling vector and second sampling vector are taken from the first discrete sampling point sequence and the second discrete sampling point sequence, respectively, and the instantaneous difference vector pointing from the first sampling vector to the second sampling vector is calculated.

[0142] Obtain the evolution direction vector of the policy semantic evolution direction field at the sampling time, and decompose the instantaneous difference vector into a vertical component parallel to the evolution direction vector and a horizontal component perpendicular to the evolution direction vector;

[0143] Calculate the magnitude of the transverse component at all sampling times, and integrate the magnitude along the preset extrapolation time interval to obtain the evolution orthogonal cumulative deviation product;

[0144] Obtain the absolute value of the time difference between the publication timestamp of the candidate document and the time coordinate of the reference semantic anchor point, and normalize the absolute value of the time difference to obtain the time penalty factor.

[0145] Multiplying the evolutionary orthogonal cumulative deviation product by the time penalty factor yields the semantic deviation between the candidate document and the reference semantic anchor.

[0146] Further, the publication timestamp of each candidate document in the anchor retrieval queue is obtained, and the publication timestamp of the candidate document with the latest publication timestamp is selected as the reference time base in the anchor retrieval queue.

[0147] Calculate the time difference between the reference time base and the publication timestamp of each candidate document to obtain the document time difference of the candidate document, and normalize the document time difference to obtain the normalized time difference;

[0148] The time-decrease factor of the candidate document is calculated based on the normalized time difference, and the semantic deviation of the candidate document is multiplied by the reciprocal of the time-decrease factor to obtain the time-calibrated semantic deviation.

[0149] The reciprocal of the time-calibrated semantic deviation is determined as the comprehensive semantic matching score of the candidate document.

[0150] In one specific embodiment, sub-step S1051: Obtain the promulgation timestamp of the candidate document, and the time coordinates of the reference semantic anchor point in the pre-constructed policy semantic evolution direction field.

[0151] Get the publication timestamp of candidate document A2 Its value is March 15, 2020, which can be expressed as a decimal year. Year.

[0152] Reference semantic anchor This is a 512-dimensional semantic vector obtained in step S104 through K-means clustering and weighted averaging. It does not directly carry temporal information. To determine the temporal coordinates of the reference semantic anchor point in the policy semantic evolution direction field, the following registration operation is performed: The cosine similarity is calculated for each time point in the policy semantic evolution direction field corresponding to the expected semantic vector. The time coordinate corresponding to the time point with the highest cosine similarity is selected as the registration time coordinate of the reference semantic anchor point. In this embodiment, the registration calculation yields... Year (corresponding to around July 1, 2021).

[0153] Sub-step S1052: Generate the semantic evolution trajectory of candidate documents

[0154] The document semantic vector of candidate document A2 Input the policy semantic evolution direction field constructed in step S103, perform semantic evolution extrapolation along the positive time direction, and generate the first discrete sampling point sequence of candidate document A2 within the preset extrapolation time interval.

[0155] Preset extrapolation time interval length Year, sampling step size In that year, a total of There are 1 sampling point, and the sampling times are respectively . ,in , .

[0156] The extrapolation process is as follows: From Initially, in step j, the policy semantic evolution direction field is queried at time point. Evolution direction vector at the location ,renew The updated version As sampling time sampling vector Sampling time sampling vector All sampled vectors constitute the first discrete sample point sequence. .

[0157] Sub-step S1053: Generate the semantic evolution trajectory of the reference semantic anchor.

[0158] Reference semantic anchors From the registration time coordinate Initially, the same extrapolation duration as sub-step S1052 is used. Year and sampling step In 2010, semantic evolution was extrapolated along the positive direction of time.

[0159] The starting time point of extrapolation is Sampling time is ;from Initially, in step j, the policy semantic evolution direction field is queried at time point. Evolution direction vector at the location ,renew The updated version As sampling time sampling vector Sampling time sampling vector All sampled vectors constitute the second discrete sample point sequence. .

[0160] Sub-step S1054: Calculate the instantaneous difference vector at each sampling time.

[0161] Because the starting points of the first discrete sampling point sequence and the second discrete sampling point sequence are different, the sampling times of the two sequences are not aligned on the absolute time axis. In order to perform point-by-point comparison, the two sequences need to be unified to the same absolute time coordinate.

[0162] The time alignment strategy adopted is as follows: taking the sampling time of the first discrete sampling point sequence as the benchmark, for each sampling time, the semantic vector of the corresponding time is obtained by linear interpolation in the second discrete sampling point sequence.

[0163] For each aligned sampling time Take the first sampling vector from the first discrete sampling point sequence. Take the second sampling vector from the interpolated second sequence. Calculate the instantaneous difference vector ;

[0164] Sub-step S1055: Decompose the instantaneous difference vector into longitudinal and transverse components.

[0165] Obtaining the policy semantic evolution direction field at the sampling time Evolution direction vector , the instantaneous difference vector Decomposed into parallel The longitudinal component and perpendicular to Horizontal component, vertical component for exist Projection vector in the direction, lateral component The residual is the instantaneous difference vector minus the longitudinal component.

[0166] Sub-step S1056: Calculate the evolutionary orthogonal cumulative deviation product

[0167] Calculate the transverse component magnitude at all sampling times. The modulus is numerically integrated along a preset extrapolation time interval to obtain the evolution orthogonal cumulative deviation product.

[0168] Sub-step S1057: Calculate the time penalty factor

[0169] Get the publication timestamp of candidate document A2 Time coordinates of the reference semantic anchor point The absolute value of the time difference between them:

[0170] ,

[0171] Retrieve all anchor points in the search queue The timestamp of the release of the candidate documents and Find the absolute value of the time difference between them, and calculate the maximum value:

[0172] ,

[0173] in, The timestamp of the publication of the ii-th candidate document in the anchor retrieval queue;

[0174] Assuming the absolute values ​​of the time differences between the five documents are 1.29 / 1.29, 0.85 / 0.85, 2.10 / 2.10, 0.42 / 0.42, and 1.73 / 1.73 years respectively, then... Year.

[0175] Time penalty factor The ratio of the absolute value of the time difference for candidate document A2 to the absolute value of the maximum time difference:

[0176] ,

[0177] Sub-step S1058: Calculate semantic deviation

[0178] Multiplying the evolutionary orthogonal cumulative deviation product by the time penalty factor yields the semantic deviation between candidate document A2 and the reference semantic anchor.

[0179] Sub-step S1059: Determine the reference time base

[0180] Retrieve the publication timestamps of all M=5 candidate documents in the anchor retrieval queue. The publication timestamp of the candidate document with the latest publication timestamp is selected as the reference time base. ;

[0181] Assuming the publication timestamps of the five candidate documents, from smallest to largest, are June 2019 (2019.50-2019.50), March 2020 (2020.21-2020.21), January 2021 (2021.08-2021.08), September 2022 (2022.75-2022.75), and January 2024 (2024.03-2024.03), then... Year.

[0182] Sub-step S10510: Calculate the normalized time difference

[0183] For candidate document A2, calculate the reference time base. Timestamp of the release of candidate document A2 Document time difference between:

[0184] ,

[0185] Calculate the maximum document time difference among all candidate documents in the anchor retrieval queue:

[0186]

[0187] The document time differences of the five candidate documents are 4.53, 4.53, 3.82, 3.82, 2.95, 2.95, 1.28, 1.28, and 00 years, respectively. Year.

[0188] The normalized time difference is:

[0189] ,

[0190] In the formula, The document time difference for candidate document A2 is the difference between the reference time base and the publication timestamp of candidate document A2.

[0191] Sub-step S10511: Calculate semantic deviation

[0192] Calculate the time decay factor of candidate document A2 based on the normalized time difference:

[0193] ,

[0194] In the formula, The preset positive real attenuation coefficient, The time-decrease factor for candidate document A2;

[0195] The semantic deviation is obtained by multiplying the semantic deviation of candidate document A2 by the reciprocal of the time-decrease factor.

[0196] In summary, the method in this embodiment, by decomposing the instantaneous difference vector into vertical and horizontal components, achieves a fine distinction between "differences in evolutionary rhythm" and "divergence in evolutionary direction." The vertical component only reflects the difference in the order of evolution between candidate documents and reference semantic anchors, and does not represent a semantic deviation; the horizontal component reflects the true directional divergence. By accumulating and integrating only the magnitude of the horizontal component, the interference of rhythm differences caused by the order of document promulgation on the deviation assessment is effectively eliminated, enabling the deviation to accurately capture the substantial deviation of candidate documents in the policy evolution direction.

[0197] By introducing a time penalty factor, the time span information between candidate documents and reference semantic anchors is organically integrated into the calculation of semantic deviation. The larger the time span, the greater the difference between the policy and historical context of the candidate document and the historical context of the anchor user's core intent, and the heavier the corresponding deviation penalty. The time penalty factor adopts the normalization processing within the anchor retrieval queue, without relying on external system time, ensuring the consistency of evaluation results at different retrieval times;

[0198] By using a timeliness decay factor to weight and correct semantic deviation, the comprehensive semantic matching score can simultaneously take into account both semantic evolution consistency and policy timeliness. The reciprocal mechanism of the timeliness decay factor results in a greater deviation amplification penalty for documents issued earlier, ensuring that the final ranking results prioritize policy documents that both conform to the evolution direction of anchor user intent and have high timeliness and freshness, thereby improving the quality of search results in multi-user scenarios.

[0199] Step S106: Determine whether the comprehensive semantic matching score falls within the preset target score range.

[0200] In this step, if the comprehensive semantic matching score of each candidate document in the anchor retrieval queue falls within the target score range, it indicates that the current retrieval results have met the quality requirements and there is no need for iterative optimization. Therefore, skip step S107 and proceed directly to step S108.

[0201] Step S107: If the target is not found, then based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, apply a reverse semantic perturbation to the initial query semantic vector of the anchor user node to generate a reconstructed query semantic vector, and re-execute the semantic recall and comprehensive semantic matching score calculation based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range.

[0202] In this step, in the anchor point retrieval queue, all candidate documents whose comprehensive semantic matching score does not fall into the target score range are determined as the deviation document set, and all candidate documents whose comprehensive semantic matching score falls into the target score range are determined as the hit document set.

[0203] Obtain the document semantic vector of each hit document in the hit document set, calculate the first mean vector of the document semantic vectors of all hit documents, and obtain the document semantic vector of each deviation document in the deviation document set, calculate the second mean vector of the document semantic vectors of all deviation documents.

[0204] In the semantic space of the pre-constructed policy semantic evolution direction field, a semantic calibration ray is constructed with the first mean vector as the origin and the second mean vector as the far end. The semantic calibration ray starts from the first mean vector and points to the second mean vector.

[0205] Calculate the vertical projection point from the initial query semantic vector of the anchor user node to the semantic calibration ray, and determine the semantic vector corresponding to the vertical projection point as the semantic correction reference point;

[0206] Along the direction of the semantic calibration ray, the semantic correction reference point is translated by a preset step length in a direction away from the second mean vector, and the semantic vector obtained after translation is determined as the reconstructed query semantic vector.

[0207] In one specific embodiment, sub-step S1071 divides the off-target document set and the hit document set.

[0208] Among all 5 candidate documents in the anchor retrieval queue, the candidate documents whose comprehensive semantic matching score does not fall into the target score range are identified as the deviation document set, and the candidate documents whose comprehensive semantic matching score falls into the target score range are identified as the hit document set.

[0209] Based on the judgment result of step S106, candidate document A3 has a score of 0.438, which is less than 0.5, and is therefore classified into the off-target document set. Candidate documents A1, A2, A4, and A5 have scores of 0.615, 0.911, 0.799, and 2.439, respectively, all of which are greater than or equal to 0.5, and are therefore classified into the hit document set.

[0210] After the partitioning is completed, the off-target document set contains one document A3, and the hit document set contains four documents A1, A2, A4, and A5.

[0211] Sub-step S1072: Calculate the first mean vector and the second mean vector.

[0212] Obtain the document semantic vector for each hit document in the hit document set. The document semantic vectors for documents A1, A2, A4, and A5 are all 512-dimensional real-valued vectors pre-generated and stored in the semantic index library in step S102. Calculate the arithmetic mean of the document semantic vectors of these four hit documents to obtain the first mean vector. The first mean vector represents the cluster center of "documents meeting the requirements" in the semantic space within the anchor user's current search results, i.e., the semantic coordinates of the "correct search direction."

[0213] Similarly, obtain the document semantic vector of document A3 in the deviation document set. Since the deviation document set contains only one document, the second mean vector is the document semantic vector of document A3 itself. If the deviation document set contains multiple documents, calculate the arithmetic mean of the document semantic vectors of all deviation documents. The second mean vector represents the cluster center of "unsuitable documents" in the semantic space, i.e., the semantic coordinates of the "incorrect retrieval direction".

[0214] Sub-step S1073: Construct semantic calibration rays

[0215] In the semantic space spanned by the policy semantic evolution direction field constructed in step S103, a semantic calibration ray is constructed with the first mean vector as the origin and the second mean vector as the endpoint. The semantic calibration ray originates from the first mean vector and points to the second mean vector. The direction vector of the ray is obtained by subtracting the first mean vector from the second mean vector.

[0216] The physical meaning of this ray is that it defines the semantic space direction from the "correct result region" to the "incorrect result region". The points on the ray are arranged along the direction from hit to deviation, with the closer to the origin (first mean vector) the more correct, and the closer to the far end (second mean vector) the more incorrect.

[0217] Sub-step S1074: Calculate the semantic correction reference point

[0218] Calculate the vertical projection point from the initial query semantic vector of the anchor user node (user A) to the semantic calibration ray. The vertical projection point is the point on the semantic calibration ray that is closest to the initial query semantic vector.

[0219] The vertical projection point retains the effective information components of the initial query semantic vector along the semantic calibration ray direction, while eliminating components in the initial query semantic vector that are orthogonal to the ray direction and may be noise. The semantic vector corresponding to the vertical projection point is determined as the semantic correction reference point, serving as the starting point for subsequent translation operations.

[0220] Sub-step S1075: Generate reconstructed query semantic vector

[0221] Along the direction of the semantic calibration ray, the semantic correction reference point is shifted away from the second mean vector by a preset step size. The direction away from the second mean vector is the opposite direction of the semantic calibration ray, which is from the direction of deviation to hit. The semantic vector obtained after the shift is determined as the reconstructed query semantic vector.

[0222] The calibrated query points are actively moved in the opposite direction to the off-bound region, making the reconstructed query vector closer to the hit document region and further away from the off-bound region in the semantic space. The preset step size is controlled at a small value, for example, 0.1 times the length of the ray direction unit vector in this embodiment, to ensure that the adjustment range of each iteration is appropriate and to avoid over-correction that would cause oscillations in the recall results.

[0223] Sub-step S1076: Re-execute semantic recall and comprehensive semantic matching score calculation.

[0224] The initial query semantic vector of the anchor user node (User A) is replaced with the reconstructed query semantic vector, and the semantic recall operation in step S103 is re-executed. Since the reconstructed query vector has moved towards the hit area, its similarity to hit documents will increase during re-recall, while its similarity to off-target documents will decrease. For the newly recalled candidate document set, the comprehensive semantic matching score calculation in step S105 is re-executed. After the calculation is complete, the process returns to step S106 to determine again whether the score falls within the target score range.

[0225] If any candidate documents still fail to meet the score, the sub-steps of step S107 are repeated for the next iteration. During the iteration, the deviation document set and the hit document set are re-divided according to the score results of the current round, and the mean vector and semantic calibration ray are recalculated to ensure that each iteration is calibrated based on the latest retrieval feedback information.

[0226] When all candidate document scores fall within the target score range after a certain round of iteration, or when the number of iterations reaches the preset maximum number of iterations, the iteration is terminated, the current anchor point retrieval queue is taken as the final result, and the process proceeds to step S108.

[0227] Step S108: The candidate documents whose comprehensive semantic matching scores fall into the target score interval in the anchor point retrieval queue, and the candidate documents corresponding to other nodes in the user intent topology graph are sorted from high to low according to their comprehensive semantic matching scores to obtain the final retrieval results. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

[0228] In this step, sub-step S1081: Obtain the node composition of the user intent topology graph.

[0229] From the sparse user intent topology graph constructed in step S103, obtain a list of all user nodes. In this embodiment, the user intent topology graph contains three user nodes: user A, user B, and user C. User A is selected as the anchor user node in step S104. The remaining nodes are user B and user C, i.e., the nodes remaining in the user intent topology graph after removing the anchor user node.

[0230] Sub-step S1082: Obtain the final candidate document for the anchor user node.

[0231] After iterative optimization in step S107, the candidate documents in the anchor user node (User A)'s anchor retrieval queue have obtained a comprehensive semantic matching score that has been repeatedly calibrated. In the final judgment in step S106, the comprehensive semantic matching scores of all candidate documents in the anchor retrieval queue have fallen into the target score range.

[0232] From the anchor retrieval queue, all candidate documents whose comprehensive semantic matching scores fall within the target score range are selected. In this embodiment, the anchor retrieval queue contains 5 candidate documents A1 to A5, and all of them meet the criteria after step S106. Therefore, the final candidate documents for the anchor user are these 5 documents, each carrying a comprehensive semantic matching score after iterative optimization. These 5 documents and their scores are recorded as the anchor document result set.

[0233] Sub-step S1083: Obtain the final candidate documents for the remaining user nodes.

[0234] The remaining nodes in the user intent topology graph—User B and User C—are processed separately.

[0235] In step S103, User B obtains an initial candidate document set containing six documents, B1 to B6. In this step, for each candidate document in User B's initial candidate document set, a comprehensive semantic matching score is calculated one by one according to the comprehensive semantic matching score calculation method in step S105. The reference semantic anchor point used in the calculation process remains the same as the one determined in step S104, and the calculation of the time-related decay factor is also based on the reference time benchmark of the anchor point retrieval queue. After the calculation is completed, candidate documents whose comprehensive semantic matching scores fall within the target score range are selected from User B's initial candidate document set. Assuming that four of the six documents meet the target score (B1, B2, B4, and B6), these four documents and their scores are recorded as User B's document result set.

[0236] In step S103, User C obtains an initial candidate document set, containing four documents C1 to C4. Using the same processing flow as User B, a comprehensive semantic matching score is calculated for each document, and qualified documents are selected. Assuming that three of the four documents meet the criteria (C1, C3, and C4), these three documents and their scores are recorded as User C's document result set.

[0237] The remaining user nodes share the reference semantic anchor and reference time benchmark of the anchor user, which makes the scoring criteria consistent across the entire user group and ensures that the candidate documents of different users are comparable in the final ranking.

[0238] Sub-step S1084: Summarize and remove duplicates

[0239] The anchor document result set, the user B document result set, and the user C document result set are summarized and merged to obtain the summary document set.

[0240] During the aggregation process, deduplication is performed for cases where the same policy document may appear in multiple users' result sets. For example, document A2 might be retrieved by both user A and user B, existing in both result sets. The deduplication rule is as follows: for the same document, retain the record with the highest overall semantic matching score among all occurrences, and delete all other duplicate records. In the deduplicated aggregated document set, each policy document appears only once, carrying its highest overall semantic matching score.

[0241] In this embodiment, it is assumed that after summarizing and deduplicating, a total of 10 unique candidate documents are obtained.

[0242] Sub-step S1085: Sort output

[0243] All candidate documents in the deduplicated summary document set are sorted in descending order of their comprehensive semantic matching score. The document with the highest score is placed first, and the document with the lowest score is placed last.

[0244] After sorting, the sorted list of documents will be returned to each user as the final search results. Each record in the final search results list contains information such as document identifier, document title, promulgation timestamp, and comprehensive semantic matching score, allowing users to quickly locate the most relevant policy documents based on the sorting results.

[0245] In summary, the method of this application obtains policy retrieval instructions from multiple users and generates initial query semantic vectors; maps each initial query semantic vector to a pre-built semantic index for semantic retrieval, obtaining an initial candidate document set; constructs a policy semantic evolution direction field using the semantic vectors of all documents in the index and the promulgation timestamp; determines the semantic overlap by calculating the bidirectional evolutionary consistency score and harmonic mean between any two document sets in the evolution direction field, and constructs a user intent topology graph based on this; selects the node with the highest connectivity in the topology graph as the anchor user, and uses its candidate document set as the anchor retrieval queue. The process involves identifying reference semantic anchors; calculating the semantic deviation between each candidate document and the reference semantic anchor, and then weighting the result with a time-related decay factor to obtain a comprehensive semantic matching score. If the score does not reach the preset target range, a semantic calibration ray is constructed based on the mean vectors of the hit and deviation documents. This ray applies a reverse semantic perturbation to the query vector of the anchor user, and iteratively reconstructs the data until the score meets the target. Finally, the documents that meet the target for each user are sorted and output according to their scores. By introducing a policy semantic evolution direction field, the semantic overlap is transformed from static content similarity to a consistency check of dynamic evolution logic, significantly improving the accuracy and timeliness of agricultural policy retrieval.

[0246] Please see Figure 2 The diagram shows a structural block diagram of a semantic retrieval system for agricultural policy documents according to this application.

[0247] like Figure 2 As shown, the semantic retrieval system 200 includes a first acquisition module 210, a second acquisition module 220, a construction module 230, a selection module 240, a correction module 250, a judgment module 260, a generation module 270, and an output module 280.

[0248] The system comprises the following modules: a first acquisition module 210, configured to acquire policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction to obtain an initial query semantic vector for each user; a second acquisition module 220, configured to acquire a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with a document semantic vector and a time-related decay factor; a construction module 230, configured to map each initial query semantic vector to the semantic index library for semantic retrieval to obtain an initial candidate document set corresponding to each user, and construct a user intent topology graph based on the semantic overlap between the initial candidate document sets; a selection module 240, configured to select the node with the highest connectivity in the user intent topology graph as an anchor user node, use the initial candidate document set corresponding to the anchor user node as an anchor retrieval queue, and determine the semantic center of the anchor retrieval queue as a reference semantic anchor; and a correction module 250, configured to calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and correct the semantic deviation based on the initial query semantic vector of each user. A time-decrease factor is used to weight and correct the semantic deviation to obtain a comprehensive semantic matching score for each candidate document. A judgment module 260 is configured to determine whether the comprehensive semantic matching score falls within a preset target score range. A generation module 270 is configured to, if not, apply a reverse semantic perturbation to the initial query semantic vector of the anchor user node based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, generate a reconstructed query semantic vector, and re-execute semantic recall and comprehensive semantic matching score calculation based on the reconstructed query semantic vector until the comprehensive semantic matching score falls within the target score range. An output module 280 is configured to sort the candidate documents in the anchor retrieval queue whose comprehensive semantic matching scores fall within the target score range, and the candidate documents in the initial candidate document set corresponding to other nodes in the user intent topology graph whose comprehensive semantic matching scores fall within the target score range, according to their comprehensive semantic matching scores from high to low, to obtain the final retrieval result. The other nodes are the nodes in the user intent topology graph excluding the anchor user node.

[0249] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0250] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the semantic retrieval method for agricultural policy documents in any of the above method embodiments.

[0251] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0252] Obtain policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user;

[0253] Obtain a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor;

[0254] Each initial query semantic vector is mapped to the semantic index library for semantic retrieval, resulting in an initial candidate document set corresponding to each user. A user intent topology graph is then constructed based on the semantic overlap between the initial candidate document sets.

[0255] The node with the highest connectivity in the user intent topology graph is selected as the anchor user node, the initial candidate document set corresponding to the anchor user node is used as the anchor retrieval queue, and the semantic center of the anchor retrieval queue is determined as the reference semantic anchor.

[0256] Calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and adjust the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document.

[0257] Determine whether the comprehensive semantic matching score falls within the preset target score range;

[0258] If it does not fall into the target range, then based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, a reverse semantic perturbation is applied to the initial query semantic vector of the anchor user node to generate a reconstructed query semantic vector, and the semantic recall and comprehensive semantic matching score are recalculated based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range.

[0259] Candidate documents whose comprehensive semantic matching scores fall within the target score range in the anchor point retrieval queue, and the initial candidate document set corresponding to other nodes in the user intent topology graph, are sorted from high to low according to their comprehensive semantic matching scores to obtain the final retrieval results. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

[0260] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of a semantic retrieval system for agricultural policy documents, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely located relative to a processor, which can be connected to a semantic retrieval system for agricultural policy documents via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0261] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the semantic retrieval method for agricultural policy documents described in the above embodiment. The input device 330 can receive input numeric or character information and generate key signal inputs related to user settings and function control of the semantic retrieval system for agricultural policy documents. The output device 340 may include a display screen or other display device.

[0262] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0263] In one implementation, the above-described electronic device is applied to a semantic retrieval system for agricultural policy documents, serving as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0264] Obtain policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user;

[0265] Obtain a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor;

[0266] Each initial query semantic vector is mapped to the semantic index library for semantic retrieval, resulting in an initial candidate document set corresponding to each user. A user intent topology graph is then constructed based on the semantic overlap between the initial candidate document sets.

[0267] The node with the highest connectivity in the user intent topology graph is selected as the anchor user node, the initial candidate document set corresponding to the anchor user node is used as the anchor retrieval queue, and the semantic center of the anchor retrieval queue is determined as the reference semantic anchor.

[0268] Calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and adjust the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document.

[0269] Determine whether the comprehensive semantic matching score falls within the preset target score range;

[0270] If it does not fall into the target range, then based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, a reverse semantic perturbation is applied to the initial query semantic vector of the anchor user node to generate a reconstructed query semantic vector, and the semantic recall and comprehensive semantic matching score are recalculated based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range.

[0271] Candidate documents whose comprehensive semantic matching scores fall within the target score range in the anchor point retrieval queue, and the initial candidate document set corresponding to other nodes in the user intent topology graph, are sorted from high to low according to their comprehensive semantic matching scores to obtain the final retrieval results. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

[0272] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0273] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semantic retrieval method for agricultural policy documents, characterized in that, include: Obtain policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user; Obtain a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor; Each initial query semantic vector is mapped to the semantic index library for semantic retrieval, resulting in an initial candidate document set corresponding to each user. A user intent topology graph is then constructed based on the semantic overlap between the initial candidate document sets. The node with the highest connectivity in the user intent topology graph is selected as the anchor user node, the initial candidate document set corresponding to the anchor user node is used as the anchor retrieval queue, and the semantic center of the anchor retrieval queue is determined as the reference semantic anchor. Calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and adjust the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document. Determine whether the comprehensive semantic matching score falls within the preset target score range; If it does not fall into the target range, then based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, a reverse semantic perturbation is applied to the initial query semantic vector of the anchor user node to generate a reconstructed query semantic vector, and the semantic recall and comprehensive semantic matching score are recalculated based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range. Candidate documents whose comprehensive semantic matching scores fall within the target score range in the anchor point retrieval queue, and the initial candidate document set corresponding to other nodes in the user intent topology graph, are sorted from high to low according to their comprehensive semantic matching scores to obtain the final retrieval results. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

2. The semantic retrieval method for agricultural policy documents according to claim 1, characterized in that, The process of performing dependency parsing on each policy retrieval command to obtain the initial query semantic vector for each user includes: Each policy retrieval instruction is sequentially input into a pre-trained syntactic parsing model, which outputs a dependency syntax tree corresponding to each policy retrieval instruction. The dependency syntax tree contains word nodes and dependency relationship edges between word nodes. In the dependency syntax tree, the core predicate verb is determined as the root node. Starting from the root node, the noun phrases in the subject branch are extracted as policy subject vector components, the noun phrases in the object branch are extracted as policy object vector components, and the time phrases and place phrases in the adverbial branch are extracted as time constraint vector components and space constraint vector components, respectively. The policy subject vector component, policy object vector component, time constraint vector component, and spatial constraint vector component are weighted and concatenated according to preset dimension weights to generate the initial query semantic vector corresponding to each user.

3. The semantic retrieval method for agricultural policy documents according to claim 1, characterized in that, The step of mapping each initial query semantic vector to the semantic index for semantic retrieval to obtain an initial candidate document set corresponding to each user, and constructing a user intent topology graph based on the semantic overlap between the initial candidate document sets, includes: Calculate the vector cosine similarity between each initial query semantic vector and each document semantic vector in the semantic index, and recall policy documents whose vector cosine similarity exceeds a preset recall threshold as the initial candidate document set for each user. Using the document semantic vectors of all policy documents in the semantic index library and the promulgation timestamps of all policy documents, a policy semantic evolution direction field is constructed; For the first initial candidate document set corresponding to the first user and the second initial candidate document set corresponding to the second user, extract all cross-set document pairs that satisfy the first time sequence condition, wherein the first time sequence condition is: the first document belongs to the first initial candidate document set, the second document belongs to the second initial candidate document set, and the publication timestamp of the second document is later than the publication timestamp of the first document; Extract all cross-set document pairs that satisfy the second time sequence condition, wherein the second time sequence condition is: the first document belongs to the first initial candidate document set, the second document belongs to the second initial candidate document set, and the publication timestamp of the second document is earlier than the publication timestamp of the first document; For a certain cross-set document pair, calculate a certain actual evolution vector pointing from the document semantic vector of a certain first document to the document semantic vector of a certain second document, and perform path integration in the policy semantic evolution direction field according to the certain time difference between the promulgation timestamp of a certain first document and the promulgation timestamp of a certain second document, to obtain a certain expected evolution vector corresponding to the certain cross-set document pair. Calculate a cosine similarity between a certain actual evolution vector and a certain expected evolution vector to obtain a certain evolutionary fit degree of a certain cross-set document pair, and calculate the arithmetic mean of the evolutionary fit degrees of all cross-set document pairs that satisfy the first temporal condition to obtain the first directional evolutionary consistency score from the first initial candidate document set to the second initial candidate document set. Calculate the arithmetic mean of the evolutionary fit of all cross-set document pairs that satisfy the second temporal condition to obtain the second directional evolutionary consistency score from the second initial candidate document set to the first initial candidate document set; Calculate the harmonic mean of the first directional evolutionary consistency score and the second directional evolutionary consistency score, and determine the harmonic mean as the semantic overlap between the first candidate document set and the second candidate document set; Using each user as a node and the semantic overlap as the weight of the edges between nodes, an initial complete graph is constructed. Then, minimum spanning tree pruning is performed on the initial complete graph, retaining only edges with weights greater than a preset edge weight threshold, to obtain a sparse user intent topology graph.

4. The semantic retrieval method for agricultural policy documents according to claim 1, characterized in that, The step of using the initial candidate document set corresponding to the anchor user node as the anchor retrieval queue, and determining the semantic center of the anchor retrieval queue as the reference semantic anchor, includes: Calculate the mean of the document semantic vectors of all candidate documents in the anchor point retrieval queue, and use it as the initial semantic center; Using the initial semantic center as the initial cluster center, perform K-means clustering iteration on the document semantic vectors of all candidate documents in the anchor point retrieval queue until the cluster centers converge to obtain the target cluster center; Select the document semantic vectors of a preset number of candidate documents that are closest to the target cluster center after clustering, calculate the weighted average of the preset number of document semantic vectors, and determine the weighted average as the reference semantic anchor point, wherein the weight of the weighted average is inversely proportional to the distance of each candidate document from the target cluster center.

5. The semantic retrieval method for agricultural policy documents according to claim 3, characterized in that, The calculation of the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor includes: Obtain the promulgation timestamp of the candidate documents in the anchor retrieval queue, and the time coordinates of the reference semantic anchor in the pre-constructed policy semantic evolution direction field; The document semantic vector of the candidate document is input into the policy semantic evolution direction field, and semantic evolution extrapolation is performed along the positive time direction to generate the first discrete sampling point sequence of the candidate document within the preset extrapolation time interval; The reference semantic anchor is input into the policy semantic evolution direction field, and semantic evolution extrapolation is performed along the positive time direction to generate a second discrete sampling point sequence of the reference semantic anchor within a preset extrapolation time interval; For each sampling time, the corresponding first sampling vector and second sampling vector are taken from the first discrete sampling point sequence and the second discrete sampling point sequence, respectively, and the instantaneous difference vector pointing from the first sampling vector to the second sampling vector is calculated. Obtain the evolution direction vector of the policy semantic evolution direction field at the sampling time, and decompose the instantaneous difference vector into a vertical component parallel to the evolution direction vector and a horizontal component perpendicular to the evolution direction vector; Calculate the magnitude of the transverse component at all sampling times, and integrate the magnitude along the preset extrapolation time interval to obtain the evolution orthogonal cumulative deviation product; Obtain the absolute value of the time difference between the publication timestamp of the candidate document and the time coordinate of the reference semantic anchor point, and normalize the absolute value of the time difference to obtain the time penalty factor. Multiplying the evolutionary orthogonal cumulative deviation product by the time penalty factor yields the semantic deviation between the candidate document and the reference semantic anchor.

6. The semantic retrieval method for agricultural policy documents according to claim 1, characterized in that, The step of weighting and correcting the semantic deviation based on the time-decrease factor to obtain the comprehensive semantic matching score for each candidate document includes: Obtain the publication timestamp of each candidate document in the anchor point retrieval queue, and select the publication timestamp of the candidate document with the latest publication timestamp in the anchor point retrieval queue as the reference time base. Calculate the time difference between the reference time base and the publication timestamp of each candidate document to obtain the document time difference of the candidate document, and normalize the document time difference to obtain the normalized time difference; The time-decrease factor of the candidate document is calculated based on the normalized time difference, and the semantic deviation of the candidate document is multiplied by the reciprocal of the time-decrease factor to obtain the time-calibrated semantic deviation. The reciprocal of the time-calibrated semantic deviation is determined as the comprehensive semantic matching score of the candidate document.

7. The semantic retrieval method for agricultural policy documents according to claim 3, characterized in that, The step of applying a reverse semantic perturbation to the initial query semantic vector of the anchor user node based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, and generating a reconstructed query semantic vector, includes: In the anchor point retrieval queue, all candidate documents whose comprehensive semantic matching score does not fall into the target score range are determined as the deviation document set, and all candidate documents whose comprehensive semantic matching score falls into the target score range are determined as the hit document set. Obtain the document semantic vector of each hit document in the hit document set, calculate the first mean vector of the document semantic vectors of all hit documents, and obtain the document semantic vector of each deviation document in the deviation document set, calculate the second mean vector of the document semantic vectors of all deviation documents. In the semantic space of the pre-constructed policy semantic evolution direction field, a semantic calibration ray is constructed with the first mean vector as the origin and the second mean vector as the far end. The semantic calibration ray starts from the first mean vector and points to the second mean vector. Calculate the vertical projection point from the initial query semantic vector of the anchor user node to the semantic calibration ray, and determine the semantic vector corresponding to the vertical projection point as the semantic correction reference point; Along the direction of the semantic calibration ray, the semantic correction reference point is translated by a preset step length in a direction away from the second mean vector, and the semantic vector obtained after translation is determined as the reconstructed query semantic vector.

8. A semantic retrieval system for agricultural policy documents, characterized in that, include: The first acquisition module is configured to acquire policy retrieval instructions from at least one user, perform dependency parsing on each policy retrieval instruction, and obtain the initial query semantic vector corresponding to each user. The second acquisition module is configured to acquire a pre-built hierarchical semantic index library of policy documents, wherein each policy document in the semantic index library is associated with and stored with a document semantic vector and a time-effect decay factor. The module is configured to map each initial query semantic vector to the semantic index library for semantic retrieval, obtain an initial candidate document set corresponding to each user, and construct a user intent topology graph based on the semantic overlap between the initial candidate document sets. The selection module is configured to select the node with the highest connectivity in the user intent topology graph as the anchor user node, use the initial candidate document set corresponding to the anchor user node as the anchor retrieval queue, and determine the semantic center of the anchor retrieval queue as the reference semantic anchor. The correction module is configured to calculate the semantic deviation between the document semantic vector of each candidate document in the anchor retrieval queue and the reference semantic anchor, and to perform weighted correction on the semantic deviation according to the time decay factor to obtain the comprehensive semantic matching score of each candidate document. The judgment module is configured to determine whether the comprehensive semantic matching score falls within a preset target score range; The generation module is configured to, if not falling into the target range, apply reverse semantic perturbation to the initial query semantic vector of the anchor user node based on the semantic gradient information of each candidate document in the anchor retrieval queue relative to the reference semantic anchor, generate a reconstructed query semantic vector, and re-execute the semantic recall and comprehensive semantic matching score calculation based on the reconstructed query semantic vector until the comprehensive semantic matching score falls into the target score range. The output module is configured to sort candidate documents in the anchor point retrieval queue whose comprehensive semantic matching score falls within the target score interval, as well as candidate documents in the initial candidate document set corresponding to other nodes in the user intent topology graph whose comprehensive semantic matching score falls within the target score interval, according to the comprehensive semantic matching score from high to low, to obtain the final retrieval result. The other nodes are the nodes in the user intent topology graph excluding the anchor point user nodes.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method according to any one of claims 1 to 7.

Images