A hierarchical information recall and attribution method for expression two-stage execution

Abstract

The application discloses a layered information recall and attribution method for expression two-stage execution, and relates to the technical field of Internet information retrieval. The method parses a dynamic expression into a normalized logical tree and assigns a node number, simultaneously generates a tree abstract and a recursive evaluation rule, executes two-stage processing on the normalized logical tree, generates a plurality of query branches and corresponding branch execution credentials in an executable manner through a first processing stage, executes the query branches to obtain candidate information, and transfers the candidate information associated with the branch execution credentials to the second processing stage for triple verification; when the candidate information that passes the verification is de-duplicated, the branch execution credentials, node state vectors, node field hit matrices and verification results of each candidate information are merged to generate fusion evidence, and recall attribution and statistics are performed according to the fusion evidence. The application reduces the mismatch and false collection in the retrieval stage, and improves the completeness of the branch verification evidence and the explainability of the recall attribution.

Description

Technical Field

[0001] This invention relates to the field of Internet information retrieval technology, and in particular to a hierarchical information retrieval and attribution method based on two-stage expression execution. Background Technology

[0002] Existing topic information monitoring systems need to continuously acquire relevant information from internet search services, media databases, or site-specific search services, based on target subjects, business events, industry terms, and exclusion criteria, and generate result lists and periodic statistical data. In actual business operations, search criteria are usually not single keywords, but rather combined expressions containing AND, OR, parentheses grouping, and exclusion relationships. Different external search services have varying capabilities in supporting complex expressions, typically requiring the complex expressions to be broken down into simple queries and executed separately. However, splitting queries expands the candidate range, and the external results may not truly satisfy the original expression. There is a lack of verifiable consistency between candidate information and the original expression version and query branches, and logical evidence formed through consistency verification is easily lost after multiple queries and multi-channel deduplication. Summary of the Invention

[0003] In view of the above-mentioned prior art, the present invention provides a hierarchical information retrieval and attribution method for two-stage expression execution, which mainly solves the technical problems existing in the background art.

[0004] To achieve the above objectives, the technical solution of this invention is implemented as follows: This invention provides a hierarchical information retrieval and attribution method for two-stage expression execution, the method comprising the following steps: S1: Obtain dynamic expressions and rule templates, perform syntax analysis on the dynamic expressions, generate a normalized logic tree, and assign a node number to each leaf node of the normalized logic tree; S2: Generate a tree summary based on the normalized logic tree and the rule template, and construct the recursive evaluation rules of the normalized logic tree based on the tree summary; S3: Perform a two-stage processing on the normalized logic tree. The two-stage processing includes a first processing stage and a second processing stage. The first processing stage is: generate multiple externally executable query branches and corresponding branch execution credentials based on the normalized logic tree, execute the query branches to obtain candidate information, and then associate the candidate information with the branch execution credentials and pass them to the second processing stage. The second processing stage is as follows: according to the recursive evaluation rules, the candidate information is recursively evaluated to generate a node state vector and a node field hit matrix, and a branch consistency check is performed based on the branch execution certificate and the node state vector to determine whether the candidate information meets the constraints of the corresponding query branch. S4: Based on the results of the branch consistency verification, the candidate information that passes the verification is deduplicated. For multiple candidate information that are determined to be the same information, their respective branch execution credentials, node state vectors, node field hit matrices and branch consistency verification results are merged to generate fusion evidence. S5: Recall the candidate information and perform statistics based on the fused evidence.

[0005] As a preferred embodiment of the present invention, in step S1, the syntax analysis is performed according to preset syntax rules, including that expressions within parentheses are evaluated first, and the AND operator has higher precedence than the OR operator. The node types of the normalized logic tree include term nodes, AND nodes, OR nodes, and exclusion nodes. Each term node stores the term identifier, the original term, the normalized term, and the node path.

[0006] As a preferred embodiment of the present invention, the specific process of generating a tree digest based on the normalized logic tree and the rule template in step S2 is as follows: The normalized logic tree is deterministically serialized according to node number, node type, normalized terms, and node relationships. The serialization result is combined with the rule template and then hashed to obtain the tree digest.

[0007] As a preferred embodiment of the present invention, in step S2, the recursive evaluation rule defines the method of combining and evaluating different node types in the normalized logic tree layer by layer towards the root node, specifically including: The term node determines the matching result by judging whether the candidate information contains the term; The AND node determines the combined result by performing a logical AND operation on the matching results of each sub-expression; The OR node determines the combined result by performing a logical OR operation on the matching results of each sub-expression; The exclusion node determines the combined result by performing a logical NOT operation on the matching result of the subexpression.

[0008] As a preferred embodiment of the present invention, in step S3, the specific process of generating multiple externally executable query branches and corresponding branch execution credentials based on the normalized logic tree in the first processing stage includes: Generate corresponding query terms based on the term nodes, generate different query branches based on the OR nodes, and combine the query sets of each child node based on the AND nodes to obtain the multiple query branches; For each query branch, generate the required node vector and the excluded node vector for the corresponding query branch based on the node number; The tree summary, branch number, query text of the query branch, required node vector, and excluded node vector are combined and then hashed to obtain the branch execution credential.

[0009] As a preferred embodiment of the present invention, the specific process of associating the candidate information with the branch execution certificate and then transferring it to the second processing stage in step S3 includes: The candidate information is bound to the branch execution credential, tree summary, rule template, required node vector, and excluded node vector corresponding to the query branch that generated the candidate information to generate a candidate branch association object; for the candidate information of the query branch that is not associated with the dynamic expression, only the source query and recall channel identifier are added.

[0010] As a preferred embodiment of the present invention, the specific process of calculating the candidate information recursively and generating the node state vector and node field hit matrix according to the recursive evaluation rule in step S3 includes: The verification text in the candidate information is recursively evaluated layer by layer from the leaf node to the root node to obtain the hit state sequence corresponding to the leaf node of the normalized logic tree. Based on the hit state sequence, the node state vector is obtained. During the recursive evaluation process, for each hit leaf node, the specific field position corresponding to that leaf node in the verification text is recorded. All field hit records of the hit leaf nodes are combined to obtain the node field hit matrix.

[0011] As a preferred embodiment of the present invention, the specific process of performing branch consistency verification based on the branch execution certificate and the node state vector in step S3 includes: The branch consistency verification includes stage identity consistency verification, mandatory node verification, and excluded node verification. The stage identity consistency verification includes recalculating the branch execution credential based on the tree digest, branch number, query text, required node vector, and excluded node vector of the candidate information, comparing the recalculated branch execution credential with the branch execution credential bound to the candidate information, and if they match, the stage identity consistency verification passes. The mandatory node verification includes comparing the mandatory node vector in the branch execution certificate with the node status vector. If the candidate information matches all mandatory nodes in the mandatory node vector, the mandatory node verification passes; if there are any missing mandatory nodes, the missing mandatory nodes are recorded. The exclusion node verification includes comparing the exclusion node vector in the branch execution certificate with the node status vector. If the candidate information does not match any exclusion node in the exclusion node vector, the exclusion node verification passes; if there is a matched exclusion node, the matched exclusion node is recorded.

[0012] As a preferred embodiment of the present invention, the specific content of step S4 includes: A normalized URL is generated based on the URL of the candidate information and used as a deduplication key; when multiple candidate information has the same deduplication key, the multiple candidate information are determined to be the same information. For candidate information with the same deduplication key, the branch execution certificate, branch number, tree summary, rule template, required node vector, excluded node vector, node state vector, node field hit matrix and consistency check result of each candidate information are fused to obtain the fused evidence.

[0013] As a preferred embodiment of the present invention, the specific content of step S5 includes: Based on the node state vector and node field hit matrix retained in the fused evidence, the specific node and hit field position of the candidate information are determined, and interpretable attribution information is generated. Based on the set of recall channels retained in the fused evidence, candidate information is classified and statistically analyzed according to different recall channels.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention parses dynamic expressions into a normalized logic tree with uniform node numbers. In the first processing stage, it generates a branch execution credential for each query branch, containing a tree summary, a vector of required nodes, and a vector of excluded nodes, and binds candidates to this credential. In the second processing stage, it re-performs the normalized logic tree evaluation on the candidate information to generate node state vectors and node field hit matrices, and uses triple verification to determine whether the candidates satisfy the semantics of the original expression. During deduplication and fusion, the credential and node-level evidence are aggregated and merged instead of discarded. This reduces mismatches in the retrieval stage, reduces false positives caused by accidental term hits, improves the integrity of evidence after deduplication, and enhances the interpretability of recall causes. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the steps of a hierarchical information retrieval and attribution method that employs a two-stage expression execution approach. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0017] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0018] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0019] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0020] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0021] Please refer to the attached document. Figure 1 This application provides a hierarchical information retrieval and attribution method for two-stage expression execution, the method comprising the following steps: S1: Obtain the dynamic expression and rule template, perform syntax analysis on the dynamic expression, generate a normalized logic tree, and assign a node number to each leaf node of the normalized logic tree.

[0022] Specifically, in step S1, the syntax analysis is performed according to preset syntax rules, which include prioritizing the evaluation of expressions within parentheses and having higher precedence for the AND operator than the OR operator. Specifically, the node types of the normalized logic tree include term nodes, AND nodes, OR nodes, and exclusion nodes. Each term node stores a term identifier, the original term, the normalized term, and the node path. Specifically, a term node corresponds to a keyword or phrase in a dynamic expression, storing the term identifier, the original term, the normalized term, and the node path; an AND node corresponds to the "+" operator in a dynamic expression, indicating that all its sub-expressions must satisfy the condition; an OR node corresponds to the "|" operator in a dynamic expression, indicating that only one of its sub-expressions needs to satisfy the condition; and an exclusion node corresponds to an exclusion condition in a dynamic expression, indicating that its sub-expressions must not satisfy the condition.

[0023] Specifically, each leaf node is assigned a unique and stable node number within the current rule template according to a deterministic traversal order. This node number is used in subsequent steps to identify the position of the leaf node in the node state vector.

[0024] S2: Generate a tree summary based on the normalized logic tree and the rule template, and construct the recursive evaluation rules of the normalized logic tree based on the tree summary.

[0025] Specifically, the process of generating a tree digest based on the normalized logic tree and the rule template is as follows: The normalized logic tree is deterministically serialized according to node number, node type, normalized terms, and node relationships. The serialization result is combined with the rule template and then hashed to obtain the tree digest. The specific calculation formula is as follows:

[0026] In the formula, For tree summary; To standardize the logic tree; This is the rule template corresponding to the logic tree; This is the sequence obtained by deterministically serializing the normalized logic tree according to node number, node type, normalized terms, and node relationships; For summary functions; This is for field concatenation operations.

[0027] Furthermore, when mismatched parentheses, missing operands, or empty terms occur, the error location and reason are returned to avoid executing ambiguous queries.

[0028] Furthermore, in step S2, the recursive evaluation rule defines the method of combining and evaluating different node types in the normalized logic tree layer by layer towards the root node, specifically including: The term node determines the matching result by judging whether the candidate information contains the term; The AND node determines the combined result by performing a logical AND operation on the matching results of each sub-expression; The OR node determines the combined result by performing a logical OR operation on the matching results of each sub-expression; The exclusion node determines the combined result by performing a logical NOT operation on the matching result of the subexpression.

[0029] Specifically, let's set Represents dynamic expressions Candidate information The matching result is then evaluated using the following recursive rules:

[0030] In the formula, Its value belongs to the set ; The expression to be evaluated is a dynamic expression. Candidate information; For the normalized terms corresponding to term nodes; To obtain candidate information The review text is constructed from at least one of the following fields: title, abstract, body, source, and URL; For input text Perform text normalization functions such as case unification, whitespace compression, full-width / half-width conversion, and preset character normalization; The standardized candidate review text; This is an indicator function.

[0031] Specifically, the value is 1 if the condition inside the parentheses is true, and 0 otherwise. and These are dynamic expressions. The first subexpression and the second subexpression; the symbol " " is for logical AND operation; symbol " "For logical OR operation; symbol " "and Both represent logical NOT operations; the symbols "+" and "-" represent logical NOT operations. "" respectively represent the AND and OR relations in the input expression.

[0032] S3: Perform a two-stage processing on the normalized logic tree, the two-stage processing including a first processing stage and a second processing stage.

[0033] Furthermore, the first processing stage is as follows: generating multiple externally executable query branches and corresponding branch execution credentials based on the normalized logic tree, executing the query branches to obtain candidate information, and associating the candidate information with the branch execution credentials before passing them to the second processing stage.

[0034] Specifically, the process of generating multiple externally executable query branches based on the normalized logic tree is as follows: Based on the syntax capabilities supported by the target search service, the logic tree is deterministically expanded: term nodes generate corresponding query terms; OR nodes generate different query branches; and nodes combine the query sets of each child node; exclusion conditions that cannot be directly executed by external services are retained until the local review stage. For example, "(Subject A|Subject B)+(Ro-Ro|Ferry|Crossing the Sea)" can be expanded into six ordinary queries: "Subject A Ro-Ro", "Subject A Ferry", "Subject A Crossing the Sea", "Subject B Ro-Ro", "Subject B Ferry", and "Subject B Crossing the Sea".

[0035] Specifically, let's set For expression The deterministic expansion rule for the resulting ordinary query set can be expressed as:

[0036] In the formula, and These are the ordinary query sets obtained by expanding the first and second subexpressions, respectively; For from set A typical query selected from the options; For from set A typical query selected from the options; This is a set union operation; The join operation to connect two ordinary queries according to the syntax requirements of the target search service; the symbols "+" and "+" "" respectively represent the AND and OR relations in the input dynamic expression.

[0037] Furthermore, for a node, its theoretical number of expansions is the product of the number of expansions of its child nodes:

[0038] In the formula, and The first subexpression and the first subexpression Theoretical query count; symbol " "This is a multiplication operation. This formula can calculate the query combination size before actually calling the external search service."

[0039] Furthermore, the expanded query is deduplicated using a sequence-preserving method, and a maximum number of expanded queries is set. .in, The maximum number of query branches allowed to be sent to an external search service; To complete the sequential deduplication and select the actual set of queries to be executed according to a preset priority; This is the set of trimmed queries that were not executed because the maximum number of unfolded elements was exceeded. This represents the number of queries actually executed in the query set. When At that time, output the execution set and clipping collection ,satisfy The system also records the node path of each pruned query in the logical tree and the reason for pruning in the diagnostic log. This controls the number of external calls while preserving explainable evidence that queries were not executed.

[0040] Furthermore, the actual execution of the query set The branch records and their retrieval scope, under unified time range, site range, information source range and maximum return quantity constraints, send ordinary queries obtained by expanding the normalized logic tree to the target search service to obtain externally returned candidate information.

[0041] Furthermore, regarding the first Each query branch, based on its number... Leaf nodes generate required node vectors and Exclude Node Vectors When the number is When the leaf node is a positive node that the query branch must satisfy, When the leaf node is an excluded node that cannot be hit, The remaining positions are set to 0. The tree summary, branch number, query text of the query branch, required node vector, and excluded node vector are combined and then subjected to a hash operation. The specific formula is as follows:

[0042] In the formula, For the first Branch execution credentials for each query branch; To query the branch number; For the first The ordinary query text corresponding to each branch; For the first The required node vectors for each branch; For the first The vector of the excluded nodes of each branch; To execute the first The target search service identifier for each query branch; For the first The time range, site range, and source range of each query branch; This is for field concatenation operations.

[0043] Specifically, after generating the branch execution certificate, the branch record is saved. Based on the query text and search scope in the branch record, a regular query is sent to the target search service to obtain externally returned candidate information. During query execution, multiple query branches can be executed concurrently. This branch record can be used as branch constraint data for recalculation and verification in the second phase.

[0044] Furthermore, the second processing stage is as follows: according to the recursive evaluation rule, the candidate information is recursively evaluated to generate a node state vector and a node field hit matrix, and a branch consistency check is performed based on the branch execution certificate and the node state vector to determine whether the candidate information meets the constraints of the corresponding query branch. Specifically, the process of associating the candidate information with the branch execution credential and then transferring it to the second processing stage includes: The candidate information is bound to the branch execution credential, tree summary, rule template, required node vector, and excluded node vector corresponding to the query branch that generated the candidate information to generate a candidate branch association object; for the candidate information of the query branch that is not associated with the dynamic expression, only the source query and recall channel identifier are added.

[0045] Specifically, the data returned by different search services is transformed into a unified candidate structure, and a query branch number is attached to each candidate returned by the expression query. Tree Summary Rule template Required node vector Exclude node vectors Branch execution certificate The search service identifier, return location, recall channel, and recall time are specified. The binding rule template for writing candidate information is denoted as... When binding, make Candidate information and branch records are associated through credential identifiers to form candidate branch association objects. The associated object of the candidate branch serves as the input for node status verification, preventing expression mismatches caused by concurrent queries, rule updates, or cross-task merging.

[0046] Specifically, in the second processing stage, the process of recursively evaluating the candidate information according to the recursive evaluation rules to generate the node state vector and the node field hit matrix includes: The verification text in the candidate information is recursively evaluated layer by layer from the leaf node to the root node. During the recursive evaluation process, for each leaf node, it is determined whether the term corresponding to that leaf node appears in the verification text of the candidate information. If it appears, the leaf node is marked as a hit; otherwise, it is marked as a miss. A hit state sequence corresponding one-to-one with the leaf nodes of the normalized logic tree is obtained, and a node state vector is generated based on this hit state sequence. The node state vector is used to represent the hit status of the candidate information with each leaf node in the logic tree.

[0047] Simultaneously, during the recursive evaluation process, for each matched leaf node, the specific field position (such as title, abstract, body text, source, or URL) corresponding to that leaf node in the review text is recorded. The field match records of all matched leaf nodes are combined to obtain a node field match matrix. This node field match matrix is ​​used to record the specific field positions corresponding to the leaf nodes where candidate information matches.

[0048] Furthermore, the specific process of branch consistency verification based on branch execution credentials and node state vectors includes three stages: stage identity consistency verification, mandatory node verification, and excluded node verification. These three verification stages are independent of each other and are all necessary conditions for candidate success.

[0049] Specifically, the stage identity consistency verification includes recalculating the branch execution credential based on the tree digest, branch number, query text, required node vector, and excluded node vector of the candidate information, comparing the recalculated branch execution credential with the branch execution credential bound to the candidate information, and if they match, the stage identity consistency verification passes. The specific verification formula is as follows:

[0050] In the formula, For candidate information With the The stage identity consistency result of each query branch; Candidate information awaiting review; The query branch number bound to the candidate information; Branch execution credentials bound to candidate information.

[0051] Specifically, the value is 1 when the condition within the parentheses is true, and 0 otherwise; the symbol " "This is a multiplication of the results of two indicator functions. According to equation (C), a binding is performed only if the recalculated voucher is consistent with the candidate binding voucher, and the rule version of the candidate binding is consistent with the currently loaded logic tree rule version." ;otherwise, .

[0052] Specifically, the mandatory node verification includes comparing the mandatory node vector in the branch execution certificate with the node status vector. If the candidate information matches all mandatory nodes in the mandatory node vector, the mandatory node verification passes; if there are any missing mandatory nodes, the missing mandatory nodes are recorded.

[0053] Specifically, after completing the stage identity verification, the semantic completeness of the required nodes in the candidate pair branches is calculated:

[0054] In the formula, For candidate information Compared to the first The semantic completeness of the required nodes in each query branch; To determine the number of leaf nodes with uniform numbers in the normalized logic tree; The leaf node index number is given, and ; Summation operation performed by leaf node index; To take 1 and the numerical value The larger of the two is used to avoid a denominator of 0 when the number of required nodes is 0. When, it means that the candidate information hits all the required nodes in the query branch, when When this occurs, it means that at least one required node was not matched by the candidate information.

[0055] Specifically, the exclusion node verification includes comparing the exclusion node vector in the branch execution certificate with the node status vector. If the candidate information does not match any exclusion node in the exclusion node vector, the exclusion node verification passes; if there is a matched exclusion node, the matched exclusion node is recorded.

[0056] Furthermore, to separately identify the missing required nodes and the hit excluded nodes, the following two vectors are calculated:

[0057] In the formula, For candidate information Compared to the first Missing vectors of required nodes for each query branch; For candidate information Compared to the first The exclusion node hit vector for each query branch; For the first The vector of required nodes for each query branch; For the first The vector of excluded nodes for each query branch; It is a vector of length all 1s; This is an element-wise multiplication of two vectors of equal length.

[0058] Furthermore, candidate information is only accepted if it simultaneously satisfies the following constraints: consistency of stage identity, valid evaluation of the root node of the logic tree, and that all required nodes are hit while all excluded nodes are not hit. The constraints are as follows:

[0059] In the formula, For candidate information Compared to the first The acceptance result of each query branch, whose value belongs to the set. ; To standardize the logic tree root node pairs candidate information The result of the recursive evaluation; It represents the sum of the absolute values ​​of each component.

[0060] Specifically, when Candidate information is only accepted when the stage identity is inconsistent. If the stage identity is inconsistent, the corresponding normalized logic tree is reloaded or the query is re-executed; if a required node is missing, the record is... The node number corresponding to the non-zero component; when an excluded node is hit, record... The node number corresponding to the zero component in the non-zero component is selected and the candidate request is rejected.

[0061] S4: Based on the results of the branch consistency verification, the candidate information that passes the verification is deduplicated. For multiple candidate information that are determined to be the same information, their respective branch execution credentials, node state vectors, node field hit matrices and branch consistency verification results are merged to generate fused evidence.

[0062] Furthermore, the verified candidate information undergoes deduplication, specifically including: reading candidate information, source records, recall channels, and existing branch consistency verification results. The system normalizes the candidate URLs for protocol, domain name, path, and tracking parameters, prioritizing the use of normalized URLs as deduplication keys; when a URL is missing or unstable, a content fingerprint is generated based on the normalized title, source, publication date, or text summary as a backup deduplication key, thus obtaining the deduplication key corresponding to each candidate. The deduplication key for candidate information can be represented as:

[0063] In the formula, For candidate information Deduplication keys; The URL field for candidate information; For candidate information The title field; For candidate information The source field; For candidate information Summary field; This indicates that the field is empty; For URL normalization functions, For text normalization functions; For candidate information The standardized release date.

[0064] Preferably, when a URL exists and is stable, a normalized URL is used as the deduplication key; when a URL is missing, the title, source, publication date, and summary are used together to generate a content fingerprint. For the same deduplication key, a candidate set is defined. The evidence for its fusion is as follows:

[0065] In the formula, For the candidate set for deduplication Evidence of fusion; For a specific deduplication key value; For all satisfied A set of candidate information; For candidate information Source records; For candidate information A collection of recall channels; To match candidate information Establish a set of query branch numbers associated with vouchers; This is a set union operation.

[0066] Furthermore, when candidate information has been associated with one or more expression query branches, the branch execution credentials and node verification evidence corresponding to each query branch are further merged. When there is a field conflict, the master record is selected according to field integrity, search service trust level, publication time validity, or return order; aggregate evidence fields are merged using ordered deduplication to ensure that the same information is output only once without losing any query branch and its verification results.

[0067] S5: Based on the fused evidence, recall regression and statistics are performed on the candidate information, the specific contents of which include: First, based on the node state vectors and node field hit matrices preserved in the fused evidence, the specific nodes and hit field positions of the candidate information are determined, generating interpretable attribution information. For example, for a candidate piece of information, the output could be "This candidate was recalled through the branch 'Main Body A Passenger Roll-on', hitting the nodes 'Main Body A' (title) and 'Passenger Roll-on' (body text), with the missing node 'Crossing the Sea' (missed)." Attribution information enables business personnel to understand the specific reasons why each candidate piece of information was recalled.

[0068] Second, based on the set of recall channels retained in the fused evidence, candidate information is classified and statistically analyzed according to different recall channels. For example, the number of recall channels for dynamic expressions, the number of recall channels for fixed keywords, the number of recall channels for supplementary queries, and the number of multiple common recall channels are counted.

[0069] Output the number of unique information entries and the total count of the stable statistical matrix, respectively. Since the same piece of information may match multiple stable statistical keywords, the total count of the statistical matrix can be greater than the number of unique information entries. Define and display the two indicators separately.

[0070] As can be seen, this method parses dynamic expressions into normalized logic trees with unified node numbers and generates tree summaries as identifiers. In the first stage, the logic tree is expanded into multiple ordinary query branches, and a branch execution credential containing a tree summary, a vector of required nodes, and a vector of excluded nodes is generated for each branch. After the query is executed, the candidate information is bound to the credential that generated the candidate, so that each candidate can be traced back to a specific query branch. In the second stage, the normalized logic tree is recursively evaluated again for the candidate information to generate node state vectors and node field hit matrices. A triple verification is performed, namely, recalculating the identity consistency during the credential verification stage, comparing the vector of required nodes with the vector of node state to verify the integrity of required nodes, and comparing the vector of excluded nodes with the vector of node state to verify the exclusion constraints, to determine whether the candidate meets the constraints of the original expression. Finally, during deduplication and fusion, the branch execution credential, node state vector, and field hit evidence are collected and merged instead of discarded to form fused evidence for attribution analysis. Thus, throughout the entire process of dynamic expression degradation to ordinary query, asynchronous execution through heterogeneous services, and multi-channel deduplication and fusion, verifiable consistency between the candidate and the original dynamic expression semantics is maintained, and a complete node-level evidence chain is preserved.

[0071] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hierarchical information recall and attribution method for two-stage expression execution, characterized in that, The method includes the following steps: S1: Obtain the dynamic expression and rule template, perform syntax analysis on the dynamic expression, generate a normalized logic tree, and assign a node number to each leaf node of the normalized logic tree; S2: Generate a tree summary based on the normalized logic tree and the rule template, and construct the recursive evaluation rules of the normalized logic tree based on the tree summary; S3: Perform a two-stage processing on the normalized logic tree. The two-stage processing includes a first processing stage and a second processing stage. The first processing stage is: generate multiple externally executable query branches and corresponding branch execution credentials based on the normalized logic tree, execute the query branches to obtain candidate information, and then associate the candidate information with the branch execution credentials and pass them to the second processing stage. The second processing stage is as follows: according to the recursive evaluation rules, the candidate information is recursively evaluated to generate a node state vector and a node field hit matrix, and a branch consistency check is performed based on the branch execution certificate and the node state vector to determine whether the candidate information meets the constraints of the corresponding query branch. S4: Based on the results of the branch consistency verification, the candidate information that passes the verification is deduplicated. For multiple candidate information that are determined to be the same information, their respective branch execution credentials, node state vectors, node field hit matrices and branch consistency verification results are merged to generate fusion evidence. S5: Recall the candidate information and perform statistical analysis based on the fused evidence.

2. The hierarchical information recall and attribution method for two-stage expression execution according to claim 1, characterized in that, In step S1, the syntax analysis is performed according to preset syntax rules, which include prioritizing the evaluation of expressions within parentheses and having higher priority for the AND operator than the OR operator. The node types of the normalized logic tree include term nodes, AND nodes, OR nodes, and exclusion nodes. Each term node stores the term identifier, the original term, the normalized term, and the node path.

3. The hierarchical information recall and attribution method for two-stage expression execution according to claim 2, characterized in that, In step S2, the specific process of generating a tree summary based on the normalized logic tree and the rule template is as follows: The normalized logic tree is deterministically serialized according to node number, node type, normalized terms, and node relationships. The serialization result is combined with the rule template and then hashed to obtain the tree digest.

4. The hierarchical information recall and attribution method for two-stage expression execution according to claim 3, characterized in that, In step S2, the recursive evaluation rule defines the method of combining and evaluating different node types in the normalized logic tree layer by layer towards the root node, specifically including: The term node determines the matching result by judging whether the candidate information contains the term; The AND node determines the combined result by performing a logical AND operation on the matching results of each sub-expression; The OR node determines the combined result by performing a logical OR operation on the matching results of each sub-expression; The exclusion node determines the combined result by performing a logical NOT operation on the matching result of the subexpression.

5. The hierarchical information recall and attribution method for two-stage expression execution according to claim 4, characterized in that, In step S3, the specific process of generating multiple externally executable query branches and corresponding branch execution credentials based on the normalized logic tree in the first processing stage includes: Generate corresponding query terms based on the term nodes, generate different query branches based on the OR nodes, and combine the query sets of each child node based on the AND nodes to obtain the multiple query branches; For each query branch, generate the required node vector and the excluded node vector for the corresponding query branch based on the node number; The tree summary, branch number, query text of the query branch, required node vector, and excluded node vector are combined and then hashed to obtain the branch execution credential.

6. The hierarchical information recall and attribution method for two-stage expression execution according to claim 5, characterized in that, In step S3, the specific process of associating the candidate information with the branch execution credential and then transferring it to the second processing stage includes: The candidate information is bound to the branch execution credential, tree summary, rule template, required node vector, and excluded node vector corresponding to the query branch that generated the candidate information to generate a candidate branch association object; for the candidate information of the query branch that is not associated with the dynamic expression, only the source query and recall channel identifier are added.

7. The hierarchical information recall and attribution method for two-stage expression execution according to claim 6, characterized in that, In step S3, the specific process of calculating the candidate information recursively and generating the node state vector and node field hit matrix according to the recursive evaluation rule includes: The verification text in the candidate information is recursively evaluated layer by layer from the leaf node to the root node to obtain the hit state sequence corresponding to the leaf node of the normalized logic tree. Based on the hit state sequence, the node state vector is obtained. During the recursive evaluation process, for each hit leaf node, the specific field position corresponding to that leaf node in the verification text is recorded. All field hit records of the hit leaf nodes are combined to obtain the node field hit matrix.

8. The hierarchical information recall and attribution method for two-stage expression execution according to claim 7, characterized in that, In step S3, the specific process of performing branch consistency verification based on the branch execution certificate and the node state vector in the second processing stage includes: The branch consistency verification includes stage identity consistency verification, mandatory node verification, and excluded node verification. The stage identity consistency verification includes recalculating the branch execution credential based on the tree digest, branch number, query text, required node vector, and excluded node vector of the candidate information, comparing the recalculated branch execution credential with the branch execution credential bound to the candidate information, and if they match, the stage identity consistency verification passes. The mandatory node verification includes comparing the mandatory node vector in the branch execution certificate with the node status vector. If the candidate information matches all mandatory nodes in the mandatory node vector, the mandatory node verification passes; if there are any missing mandatory nodes, the missing mandatory nodes are recorded. The exclusion node verification includes comparing the exclusion node vector in the branch execution certificate with the node status vector. If the candidate information does not match any exclusion node in the exclusion node vector, the exclusion node verification passes; if there is a matched exclusion node, the matched exclusion node is recorded.

9. A hierarchical information recall and attribution method for two-stage expression execution according to claim 8, characterized in that, The specific content of step S4 includes: A normalized URL is generated based on the URL of the candidate information and used as a deduplication key; when multiple candidate information has the same deduplication key, the multiple candidate information are determined to be the same information. For candidate information with the same deduplication key, the branch execution certificate, branch number, tree summary, rule template, required node vector, excluded node vector, node state vector, node field hit matrix and consistency check result of each candidate information are fused to obtain the fused evidence.

10. The hierarchical information recall and attribution method for two-stage expression execution according to claim 9, characterized in that, The specific content of step S5 includes: Based on the node state vector and node field hit matrix retained in the fused evidence, the specific node and hit field position of the candidate information are determined, and interpretable attribution information is generated. Based on the set of recall channels retained in the fused evidence, candidate information is classified and statistically analyzed according to different recall channels.

Images