A multi-dimensional efficient recall method for complex time-sensitive documents

Abstract

The application discloses a kind of complex time-sensitive document multidimensional efficient recall method, it is related to document retrieval technical field, including the step of constructing document atlas: time-sensitive document is expressed as node in atlas, node attribute includes document content, effective date and state mark, node is connected by edge, edge connection is used to express semantic relationship type, including abolition, postponement, interpretation and update;Recall stage step: based on user query, execute the overall semantic vector of sentence with keyword vector, fusion generates new fusion vector, carries out the joint recall of keyword and text;Atlas recall step: after user submits query request, the state inspection of each document node in recall stage is carried out;Output result step: return valid document and query basis link;The method of the application significantly improves the timeliness, accuracy and explainability of document retrieval, effectively improves the retrieval experience and decision accuracy of user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document retrieval, more particularly, it relates to a multi-dimensional efficient recall method for complex time-sensitive documents. BACKGROUND

[0002] In the fields of taxation, corporate compliance, and finance, relevant time-sensitive documents exhibit characteristics such as large quantity, frequent updates, and strong timeliness. These time-sensitive documents are complex in content, covering multiple levels and fields, and have constructed a large and detailed document system. Specifically, the time-sensitive document system not only includes national-level macro documents, but also involves local-level specific implementation rules, industry regulatory documents, and corporate internal compliance requirements, etc. The documents at each level and in each field are intertwined and work together.

[0003] The relationship between old and new documents is complex, mainly manifested in the following types:

[0004] 1. Modification relationship: New documents may partially modify or supplement some provisions of old documents, which may involve document content, implementation scope, execution standards, etc. For example, with the development of technology or changes in the market, new energy vehicle subsidy documents may adjust the subsidy amount, applicable vehicle models, or application conditions, etc. to better adapt to new development needs.

[0005] 2. Related relationship: Some documents are related to each other, for example, one document may be an interpretation or supplementary explanation of another document. This relationship helps to better understand the intent and implementation details of the document, but also increases the complexity of document retrieval and understanding.

[0006] 3. Invalid relationship: With the change of situation or the introduction of new documents, old documents may be abolished or partially invalidated. For example, some outdated industry support-related documents may no longer be applicable due to industrial restructuring, but users may still encounter these invalid documents when searching.

[0007] 4. Postponement relationship: Some documents may be postponed for execution due to special circumstances, or their validity period may be extended. For example, the validity period of some tax preferential documents may be extended to continue to support enterprise development.

[0008] These complex relationships make it extremely difficult to manage and retrieve documents. When searching for documents, users not only need to find the current valid content, but also need to understand whether these documents are invalid and their relationship with other documents, in order to accurately grasp the evolution and impact of the documents.

[0009] However, traditional document retrieval methods mainly rely on vector retrieval and keyword retrieval, which have many limitations:

[0010] 1. Lack of precision: Ordinary vector recall + keyword recall this double path recall usually adopts fixed weight allocation, for example, the results of vector retrieval and keyword retrieval are combined in a fixed proportion. This fixed weight cannot be adjusted according to different query scenarios and data characteristics. In some scenarios, vector retrieval may be more suitable, while in other scenarios, keyword recall may be more advantageous. The fixed fusion method cannot dynamically adapt to these differences, resulting in suboptimal retrieval results.

[0011] 2. Unable to intelligently identify document invalidity and complex relationships: Traditional retrieval methods have obvious limitations when dealing with time-sensitive document information. They cannot intelligently identify invalid documents and cannot provide corresponding replacement or update suggestions. When users search for replacement documents for invalid documents, they often need to manually search and filter, which not only consumes time and effort, but also easily misses key information. In addition, for complex relationships between documents such as modification, correlation, invalidity, and postponement, traditional retrieval methods cannot effectively identify and display them. Users often need to spend a lot of time manually filtering and comparing to fully understand the overall picture of the documents and their evolution process, which greatly reduces retrieval efficiency and user experience.

[0012] 3. Time-sensitive issues: Documents have strong time sensitivity, but traditional retrieval methods cannot automatically filter out the latest valid documents, and users need to judge the validity of the documents themselves, which not only increases the difficulty of retrieval, but also may lead to incorrect decisions due to the use of outdated documents.

[0013] Therefore, traditional document retrieval methods have been unable to meet users' demands for accurate, efficient, and comprehensive retrieval, and there is an urgent need for more intelligent and efficient retrieval tools and methods to help users quickly and accurately obtain and understand the latest and most effective information. SUMMARY

[0014] The present application overcomes the shortcomings of the prior art and provides a multi-dimensional efficient recall method for complex time-sensitive documents that is accurate, efficient, and comprehensive.

[0015] The technical solution of the present application is as follows:

[0016] A multi-dimensional efficient recall method for complex time-sensitive documents, comprising the following steps:

[0017] (1) Building a document graph step: representing time-sensitive documents as nodes in a graph, node attributes include document content, effective date, and state markers, the state markers include valid and invalid; nodes are connected by edges, edge connections represent semantic relationship types, including abolition, postponement, interpretation, and update;

[0018] (2) Recall stage step: based on user query, perform the fusion of keyword vector and the overall semantic vector of the sentence to generate a new fusion vector, and perform joint recall of keywords and text;

[0019] (3) Graph recall step: after the user submits a query request, the state of each document node in the recall stage is checked: if the state is invalid, the alternative document node is traced through the semantic relationship edge connection in the graph; if the alternative document node state is delayed, the latest revised version is retrieved, and the valid document node is used to replace the invalid document node as the final recall result;

[0020] (4) Output result step: return the valid document and query basis link.

[0021] Further, the recall model in the step (2) is fine-tuned based on the BERT model, and the recall model specifically includes the following process:

[0022] (2.1) Positive and negative sample construction step: construct semantic samples and keyword samples; the positive sample of the semantic sample is to select the document paragraph related to the problem from the corpus, and the negative sample is to randomly select the document paragraph unrelated to the problem ; the positive sample of the keyword sample is to extract the core term from the problem by TF-IDF, and the negative sample is to randomly select from the non-core words of the problem ;

[0023] (2.2) Sample vectorization step: before vectorizing the positive and negative samples, vectorize the text and special identifier [CLS], and respectively splice to represent the overall semantic information of the text;

[0024] (2.3) Fusion vector step: fuse the keyword vector and the semantic vector, which fuses the keyword information and the semantic information through a length-aware dynamic gating mechanism; specifically, when the implicit semantic information cannot be captured by the length, when the query text is relatively short, it is determined that the query text is more related to the keyword, and when the query text is relatively long, it is determined that the query text is more related to the semantic, thereby setting a learnable length encoder.

[0025] Further, the splicing in the (2.2) sample vectorization step is as follows:

[0026]

[0027]

[0028]

[0029] wherein the semantic feature is the encoding of the BERT model for the [CLS] position, the keyword vector is the average of the encoding vectors for the positions other than the [CLS] position, represents a sample text to be vectorized, represents all position indexes in the input sequence except [CLS].

[0030] Further, the (2.3) fusion vector step maps the query length into a vector representation with semantic information as follows:

[0031]

[0032] wherein is a length vector constructed according to the text length of the question , with a dimension of , reduced to one dimension by a linear layer, and then normalized by an activation function to obtain the above query length gating weight α.

[0033] According to the query length gating weight α, the semantic vector and the keyword vector are fused, and the final information fused vector is as follows: .

[0034] Further, the user query and the document text are input into the recall model to generate the semantic vector and the keyword vector; the semantic vector and the keyword vector are fused through a dynamic gating mechanism to generate a fusion vector, and a cosine similarity is calculated based on the fusion vector to screen relevant documents.

[0035] Further, the dynamic gating mechanism includes generating a length vector according to the length of the user query text, normalizing the length vector into a weight parameter σ through a linear layer and an activation function; and dynamically adjusting the contribution degree of the keyword vector and the semantic vector using the weight parameter σ.

[0036] Further, the state marker and the semantic relationship edge are automatically updated based on a document timeliness rule to ensure the timeliness of the recall result.

[0037] Further, it also includes constructing a vector-level contrastive loss function:

[0038]

[0039] wherein represents the calculation of the cosine similarity between two texts, is the vector representation of the i-th sample in a training batch, is the vector representation of the positive sample in a training batch, is a temperature hyper-parameter to adjust the degree of attention of the model to difficult negative samples, B is all samples contained in a training batch.

[0040] A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the method as described above when executing the computer program.

[0041] A computer-readable storage medium storing a computer program, the program being executed to implement the method as described above.

[0042] The present application has the following advantages over the prior art:

[0043] The present application introduces the concept of "timeliness document graph", models the documents in a structured form, and annotates the semantic edge relationships of "modification", "correlation", "invalidation", and "delay" through the "valid / invalid" state in the graph structure, to realize the construction of the evolution path between documents. The multi-hop relationship tracking can be performed according to these semantic edges, and the current valid document version can be automatically inferred, to help users accurately obtain the target document and its historical change path, and significantly improve the timeliness, accuracy, and explainability of document retrieval.

[0044] The present application combines keyword recall, semantic recall, and established graph recall, to form a multi-strategy fusion mechanism with dynamically adjustable weights. First, keyword recall ensures the accurate matching ability for clear query intent, and can quickly lock the document node containing the core term; second, semantic recall captures the potential semantic association between the query and the document text through a deep vectorization model, to improve the adaptation ability to complex query methods such as ambiguity, synonymy, and rewriting; finally, graph recall uses the multi-hop chain relationship reasoning in the timeliness document graph to realize cross-node and cross-version document retrieval and automatic replacement.

[0045] In the fusion stage, a dynamic weight distribution and length-aware gating mechanism is introduced, which can automatically adjust the contribution of keyword recall results in the final sorting according to the length characteristics of the query text. For example, for short text and high keyword density queries, the system tends to increase the weight of keyword recall results; for long text or descriptive queries, the system tends to increase the weight of semantic recall. Through this mechanism, the recall strategy can be adaptively adjusted under different business scenarios and different query characteristics, to achieve a balance between high precision and high coverage. At the same time, with the help of the timeliness tracing and state verification function of the graph recall, it is ensured that the final output document result not only has high relevance, but also is the latest valid version, thereby effectively improving the user's search experience and decision-making accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a schematic diagram of the overall flow of the present application;

[0047] Figure 2 is a model training architecture diagram of the present application. DETAILED DESCRIPTION

[0048] Embodiments of the present application are described in detail below with reference to the attached drawing figures, wherein the same or like reference numerals and characters in the drawing figures throughout the drawings denote like features or elements having the same or similar functions. The embodiments described below are exemplary only, and are not to be construed as limiting the present application.

[0049] Those skilled in the art will appreciate that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0050] The reference numerals mentioned in the embodiments with respect to steps are only for the convenience of description, and there is no substantial sequence relationship. Different steps in different embodiments can be combined in different sequences to achieve the purpose of the present application.

[0051] The present application is further illustrated below with reference to the accompanying drawings and specific embodiments.

[0052] As shown in Figure 1 , 2 , a multi-dimensional efficient recall method for complex time-sensitive documents includes the following steps:

[0053] (1) Building a document graph step: representing time-sensitive documents as nodes in a graph, node attributes including document content, effective date and state marker, the state marker including valid, invalid and draft; nodes are connected by edges, edge connections represent semantic relationship types, including repeal, extension, interpretation and update.

[0054] That is, first, a structured graph system is constructed based on invalid documents, each document is regarded as a node in the graph, and the node attributes include document content, effective date and current state marker (such as "valid", "invalid", "draft", etc.). Different document nodes are connected to each other through semantic relationship edges, and the relationship types include but are not limited to "repeal", "extension", "interpretation", "update", which are used to describe the logical association and evolution path between documents. This graph structure provides a basis for subsequent multi-hop relationship reasoning and semantic recall.

[0055] (2) Recall stage step: based on user query, the key word vector is fused with the overall semantic vector of the sentence to generate a new fusion vector, and the contrast loss recall is carried out. It realizes the effective combination of the two by vectorizing method, which deeply fuses the key word information and semantic information into the same vector space, thereby realizing the effective combination of the two. Specifically, the BERT model after fine-tuning is used for vectorization processing, that is, the recall model. That is, the user query and the document text are input into the recall model to generate semantic vectors and key word vectors; the semantic vectors and key word vectors are fused through a dynamic gating mechanism to generate a fusion vector, and the cosine similarity is calculated based on the fusion vector to screen relevant documents.

[0056] The recall model can efficiently integrate the semantic features of the key words and the overall semantic features of the sentence to generate a more representative vector representation. Especially for the loss function, the vector distance between the user question and the relevant documents is narrowed, while the vector distance between the user question and the irrelevant documents is widened, thereby generating an enhanced vector representation that comprehensively considers semantics and key words. This design not only retains the explicitness of key words, but also fully captures the flexibility of semantics, significantly improving the accuracy and coverage of the search results.

[0057] The recall model is based on the BERT model and the specific process of the recall model is as follows:

[0058] (2.1) Positive and negative sample construction step: construct semantic samples and key word samples.

[0059] Positive sample of semantic sample Select the document paragraph related to the question from the corpus Negative sample Randomly select a document paragraph unrelated to the question

[0060] Positive sample of key word sample Extract core terms from the question Negative sample Randomly select non-core words from the question

[0061] Among them, the ratio of positive and negative samples is generally set to 1:1, and the corpus generally ensures that the positive sample is 2000, and the negative sample is 2000 or more.

[0062] (2.2) Sample vectorization step: before vectorizing the positive and negative samples, the text is vectorized with a special identifier [CLS], and is spliced respectively to represent the overall semantic information of the text. The spliced representation is as follows:

[0063] ​​

[0064]

[0065]

[0066] where semantic feature is the BERT model encoding for the [CLS] position, is a sample text to be vectorized, keyword vector is the average of the encoding vectors for the positions other than the [CLS] position, denotes all position indices in the input sequence except [CLS].

[0067] (2.3) Fusion vector step: fuse the keyword vector and the semantic vector, which is through a length-aware dynamic gating mechanism to fuse keyword information and semantic information. Specifically, when the implicit semantics cannot be represented by length, the recall model is used to capture it. When the query text is relatively short, it is determined that the query text is more related to the keyword, and when the query text is relatively long, it is determined that the query text is more related to the semantics, thereby setting a learnable length encoder.

[0068] Map the query length to a vector representation with semantic information as follows:

[0069]

[0070] where is a length vector constructed according to the text length of the question , with a dimension of , reduced to one dimension through a linear layer, and then normalized by activation function to obtain the above query length gating weight α. The model learns a specific vector representation for each possible text length during training to generate the query length gating weight α. During inference, the length vector is directly indexed based on the text length.

[0071] Vector fusion is performed on the semantic vector and the keyword vector according to the query length gating weight α, and the final information fused vector is as follows: .

[0072] As a preferred, a vector-level contrastive loss function is also constructed:

[0073]

[0074] where denotes the cosine similarity between two texts, is the vector representation of the i-th sample in a training batch, is the vector representation of positive samples in a training batch, is a temperature hyper-parameter that regulates the degree of attention the model pays to difficult negative samples, B is the set of all samples contained in a training batch.

[0075] (2.4) Document recall step:

[0076] For all documents in the knowledge base, their feature representations are obtained respectively through the above steps:

[0077]

[0078] where is the set of feature representations of all documents in the knowledge base, and n represents the number of all documents in the knowledge base.

[0079] By comparing the similarity between the original text vector and the knowledge base text vector, the TopN documents are selected as the recall results.

[0080]

[0081] where, is the final feature representation of the problem text, is a function that selects the TopN similarity.

[0082] (3) Graph recall step: After the user submits a query request, the state of each document node in the recall stage is checked: if the state is abolished, the alternative document node is traced through the semantic relationship edge connection in the graph; if the alternative document node is in the state of postponement, the latest revised version is retrieved, and the valid document node is used to replace the invalid document node as the final recall result.

[0083] The specific document graph query recall processing flow is as follows:

[0084] (3.1) State verification and link tracking step: After the user submits a document query request, the system performs state verification on the Top N document nodes P1 in the recall.

[0085] If P1 is in the state of "abolished", the alternative document tracing mechanism is automatically triggered, and the corresponding alternative document node P2 is traced along the document graph relationship chain.

[0086] (3.2) Alternative document state detection step: Detect the state of P2, if P2 is in the state of "postponed", the latest revised version needs to be found, and the document latest modified version retrieval sub-flow is started, if there is a latest revised version, then trace to the document node p3.

[0087] (3.3) Modified version confirmation step: based on the document time limit rule, within the extension period of P2, retrieve and extract its latest modified version P3. Verify the status of P3, if it is "valid" state, replace the initial recall result with the final version of the document text. Among them, the status label and semantic relationship edge are automatically updated based on the document time limit rule to ensure the timeliness of the recall result.

[0088] (4) Output result step: return valid documents and query basis link.

[0089] The specific implementation process is combined Figure 1 as follows:

[0090] Step 1: Start the process, assuming the input is "How to pay stamp duty?"

[0091] Step 2: User submits query, recalls relevant basis, and uses the basis recall model as shown in the model. The recall result is "Article 13. Taxpayers who are units shall report and pay stamp duty to the competent tax authorities at the location of their institutions; taxpayers who are individuals shall report and pay stamp duty to the competent tax authorities at the location of the tax certificate book or the residence of the taxpayer." Figure 2

[0092] Step 3: The system finds the corresponding basis node P1 through the recall result.

[0093] Step 4: Perform basis node P1 status verification:

[0094] If P1 is valid → Step 9.

[0095] If P1 is abolished → Step 5.

[0096] Step 5: Trigger alternative document retrieval, find P2.

[0097] Step 6: Perform P2 status verification:

[0098] If P2 is valid → Step 9.

[0099] If P2 is extended → Step 7.

[0100] Step 7: Retrieve the latest modified version P3 of P2.

[0101] Step 8: Perform P3 status verification:

[0102] If P3 is valid → Replace the initial query result with P3, and go to Step 9.

[0103] Step 9: Return valid basis and query basis link. ​

[0104] Step 10: End the process.

[0105] An instance of the recall model is combined as follows:

[0106] Step 2.1 Text encoding and feature extraction:

[0107] 1. Input the original text (example: "How to pay stamp duty?") to the BERT model.

[0108] 2. BERT encodes the text into a sequence of vectors ( to ) and a special vector . Among them as the semantic representation of the entire sentence (denoted as ).

[0109] 3. The vectors output by BERT ( to ) are processed through an average pooling layer to generate keyword representation ( ).

[0110] Step 2.2 Length information weight acquisition:

[0111] 1. Generate a text length vector according to the text length. During inference, only the numerical value of the text length is needed to obtain the length vector . Through a linear layer and an activation function, reduce to one dimension as σ, as the weight of the keyword vector.

[0112] Step 2.3 Feature fusion and loss calculation:

[0113] Weighted fusion of keyword representation ( ) and semantic representation ( ), the weight is controlled by σ, = (1 -σ) + σ , the output of the fused final feature is , and the loss is calculated based on to optimize the model parameters through backpropagation.

[0114] Step 2.4 Document recall:

[0115] For all documents in the knowledge base, their feature representations are obtained through the above steps, and the similarity between the original text vector and the knowledge base text is compared to select the top 5 documents as the recall result.

[0116] In one embodiment, the present application further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method steps provided by the above-mentioned embodiments when executing the computer program.

[0117] In one embodiment, the present application further provides a computer readable storage medium, having a computer program stored thereon, wherein the computer program is executable by a processor to implement the method steps provided by the above-mentioned embodiments.

[0118] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM) and the like.

[0119] Any combination of the technical features of the above-mentioned embodiments can be made, and in order to make the description simple, all possible combinations of the technical features in the above-mentioned embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0120] The application has several embodiments, which are described in more detail and in more detail, but should not be understood as limiting the scope of the patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, and these improvements and refinements should be considered as within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for multi-dimension efficient recall of complex time-sensitive documents, comprising: The method comprises the following steps: (1) constructing a document graph step: representing time-sensitive documents as nodes in a graph, node attributes including document content, effective date and state marker, the state marker including valid, abolished and delayed; Nodes are connected by edges, and edge connections represent semantic relationship types, including explanation and update; (2) recall stage step: based on user query, execute the keyword vector and the overall semantic vector of the sentence, generate a new fusion vector, and perform joint recall of keywords and text; (3) graph recall step: (3.1) state verification and link tracking step: after the user submits a document query request, the system performs state verification on the top N document nodes P1 recalled; if P1 is in the abolished state, the system automatically triggers the substitute document tracing mechanism and traces to the corresponding substitute document node P2 along the document graph relationship chain; (3.2) substitute document state detection step: detect P2 state, if P2 is in the delayed state, find the latest revised version, start the document latest modified version retrieval sub-process, if there is a latest revised version, trace to the document node P3; (3.3) modified version confirmation step: based on the document time limit rule, within the extension period of P2, retrieve and extract its latest modified version P3; verify the state of P3, if it is in the valid state, replace the initial recall result with the document text as the final version; (4) output result step: return the valid document and query basis link; The recall model in step (2) is fine-tuned based on the BERT model, and the recall model specifically includes the following process: (2.1) Positive and negative sample construction step: construct semantic sample and keyword sample; positive sample of semantic sample For selecting document paragraphs related to the question from the corpus, negative sample For randomly selecting document paragraphs unrelated to the question from the corpus; positive sample of keyword sample For extracting core terms from the question by TF-IDF, negative sample For randomly selecting from the non-core words of the question ​ (2.2) sample vectorization step: before vectorizing positive and negative samples, vectorize the text and special identifier [CLS], and respectively concatenate to represent the overall semantic information of the text; (2.3) fusion vector step: fuse keyword vector and semantic vector, which uses a length-aware dynamic gating mechanism to fuse keyword information and semantic information; Specifically, when the query text is relatively short, it is determined that the query text is more related to the keyword, and when the query text is relatively long, it is determined that the query text is more related to the semantic, thereby setting a learnable length encoder.

2. The method of claim 1, wherein, The concatenation in the (2.2) sample vectorization step is as follows: , , , wherein the semantic features is encoded as a [CLS] position BERT model encoding, keyword vector is the average of the encoding vectors for the other positions except the [CLS] position, represents a sample text to be vectorized, denotes all position indices in the input sequence except [CLS].

3. The method of claim 1, wherein, Input user query and document text into the recall model to generate semantic vector and keyword vector; generate a fusion vector by fusing the semantic vector and keyword vector through a dynamic gating mechanism, calculate the cosine similarity based on the fusion vector, and screen relevant documents.

4. The method of claim 3, wherein, The dynamic gating mechanism includes generating a length vector according to the length of the user query text, passing the length vector through a linear layer and an activation function to normalize the length vector into a weight parameter σ ; and dynamically adjusting the contribution degrees of the keyword vector and the semantic vector using the weight parameter σ.

5. The method of claim 1, wherein, The state marker and semantic relationship edge are automatically updated based on the document time limit rule to ensure the timeliness of the recall result.

6. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1-5.

7. A computer readable storage medium storing a computer program, characterized in that, The program is executed to realize the method of any one of claims 1-5.

Images