Bid project identification generation method, device and electronic equipment

Abstract

The application provides a bid project identifier generation method and device and electronic equipment, and belongs to the technical field of data information management. The method comprises the following steps: obtaining a target information text, extracting project core metadata entities and unstructured project descriptions; generating a feature vector according to the project core metadata entities and performing a hash operation to obtain a metadata hash value; converting the unstructured project descriptions into deep semantic fingerprint vectors by using a contrast learning and transformer architecture; fusing the metadata hash value and the deep semantic fingerprint vectors to obtain fused identifier data, and performing conflict detection to generate a globally unique bid project identifier. The application can improve the accuracy of cross-platform bid project identification and the stability of unique identifier generation.

Description

Technical Field

[0001] This invention relates to the field of data information management technology, specifically to a method, apparatus, and electronic device for generating bidding project identifiers. Background Technology

[0002] In the bidding market, the same bidding project is often repeatedly published on multiple government procurement websites, enterprise bidding platforms, and third-party information aggregation platforms, accompanied by subsequent texts such as correction notices, change instructions, and winning bid announcements. To effectively integrate this fragmented information, it is necessary to uniquely identify and associate bidding projects published across platforms, thereby supporting application scenarios such as project deduplication, lifecycle tracking, and data analysis. Currently, the industry generally uses string matching or rule engines based on explicit fields such as project title, bidding entity, and publication time to generate project identifiers. Some solutions also introduce simple hash functions to encode key fields in order to achieve unique differentiation of projects.

[0003] In the process of developing this invention, the inventors discovered that existing methods have at least the following drawbacks: First, different publishing platforms may use abbreviations, full names, parenthetical remarks, and misspellings to name the same project. Traditional string matching and hashing methods are highly sensitive to changes in text format; even minor textual alterations can lead to completely different identifiers, making it difficult to effectively associate substantially identical projects. Second, existing solutions largely rely on manually set entity recognition rules or keyword templates, which are insufficient to handle the diverse descriptive styles in massive amounts of data and lack the ability to understand the semantic level of text. They cannot distinguish between complex situations where "the descriptions are similar but the substance is different" and "the descriptions are different but the substance is the same," leading to identifier conflicts or missed detections. Third, bidding projects will issue various types of announcements (such as corrections, changes, and winning bids) during their lifecycle. Existing technologies only generate independent identifiers for each announcement, lacking a structured evolutionary association mechanism. This makes it impossible to dynamically track the complete process of a project from bidding to winning the bid, which is detrimental to subsequent data statistics and risk analysis.

[0004] Therefore, how to stably extract core project features from multi-source heterogeneous tender texts and generate globally unique identifiers that are resistant to textual perturbations in order to improve the accuracy of cross-platform project identification is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] In view of this, it is necessary to provide a bidding project identifier generation method, device and electronic device to solve the technical problems of low cross-platform project identification accuracy and high identifier conflict rate caused by the existing methods being sensitive to differences in text expression, lacking semantic understanding ability and lacking evolutionary association mechanism.

[0006] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for generating a bidding project identifier, comprising: Obtain the target tender text, and based on the target tender text, extract the core metadata entities and unstructured project descriptions; Based on the project's core metadata entity, a project core metadata feature vector is generated, and a hash operation is performed on the project core metadata feature vector to obtain the metadata hash value; By employing contrastive learning and the Transformer architecture, the unstructured project description is transformed into a deep semantic fingerprint vector in a high-dimensional semantic space. The metadata hash value is fused with the deep semantic fingerprint vector to obtain fused identifier data, and a conflict detection algorithm is executed on the fused identifier data to generate a globally unique bid project identifier.

[0007] In one possible implementation, the step of extracting the project's core metadata entities and unstructured project description based on the target tag text includes: The target tag text is encoded using a BiLSTM-CRF model to obtain the encoding layer output; Based on the output of the coding layer, the optimal prediction probability of the label sequence is determined using a conditional random field layer; Based on the tag sequence corresponding to the optimal prediction probability, key entities are extracted from the target tender text as the core metadata entities of the project. The key entities include at least one of the following: tenderer, project amount, and construction location.

[0008] In one possible implementation, the use of contrastive learning and a Transformer architecture to transform the unstructured project description into a deep semantic fingerprint vector includes: The unstructured item description is input into the Transformer encoder, and the same unstructured item description is forwarded twice using the Transformer encoder with different random deactivation masks to obtain two semantic representations as positive sample pairs. Obtain descriptions of other items in the same training batch as negative sample pairs, and select semantically similar but different identifier items as hard negative samples; Based on the positive sample pairs, the negative sample pairs, and the hard negative samples, the Transformer encoder is trained using a contrastive learning loss function, so that the semantic representations of similar items are close in distance in the semantic space, and the semantic representations of different items are far apart. The semantic embeddings output by the trained Transformer encoder are pooled to obtain the deep semantic fingerprint vector.

[0009] In one possible implementation, performing a hash operation on the core metadata feature vector of the project to obtain a metadata hash value includes: The core metadata feature vector of the project is encoded in hexadecimal to obtain the encoded string; The encoded string is hashed using a secure hash algorithm to obtain the metadata hash value.

[0010] In one possible implementation, fusing the metadata hash value with the deep semantic fingerprint vector includes: The deep semantic fingerprint vector is input to the encoder for encoding to obtain the encoded semantic features; The encoded semantic features are dimensionality-reduced by using a dimensionality-reducing mapping function, which is implemented by a fully connected neural network layer. The fused identifier data is obtained by performing an XOR operation between the metadata hash value and the dimensionality-reduced semantic features.

[0011] In one possible implementation, performing a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier includes: The fused identifier data will be used as the initial draft of the tender project identifier; Access the global UID index to check if the initial draft of the bidding project identifier already exists; A Bloom filter is configured at the front end of the global UID index library. The Bloom filter is used to perform an existence check on the initial draft of the bidding project identifier. If the Bloom filter determines that the initial draft of the bidding project identifier does not exist, it is passed directly and the initial draft of the bidding project identifier is output as the final identifier. If the Bloom filter determines that a conflict may exist or the database query confirms that a conflict has already existed, a spatiotemporal perturbation factor containing a nanosecond-level timestamp and a server node identifier is obtained. The spatiotemporal perturbation factor is then appended to the fused identifier data, and a new hash operation is performed to generate a new bidding project identifier. When fingerprint overlap occurs due to project changes or correction notices, an incrementing version identifier is appended to the end of the newly generated bid project identifier.

[0012] In one possible implementation, the method for generating the bid project identifier includes: Obtain the generated bid project identifier and use the bid project identifier as a node in the graph neural network; Obtain the legal and logical relationships and temporal evolution relationships between nodes, and construct the edges in the graph neural network based on the legal and logical relationships and temporal evolution relationships. The edges carry the announcement timestamp and the announcement type identification factor. The graph neural network is used to maintain the evolutionary relationship between projects based on the nodes and edges, and the evolutionary relationship includes the dynamic tracking of project changes and correction announcements.

[0013] In one possible implementation, employing the graph neural network to maintain the evolutionary relationship between items based on the nodes and edges includes: The embedding vector of the current node and the set of neighboring nodes of the current node are obtained by using graph sampling and aggregation algorithms; Aggregate the neighbor node embeddings in the neighbor node set to obtain the aggregation result; The embedding vector of the current node is concatenated with the aggregation result, and a nonlinear transformation is performed on the concatenated vector to generate the node embedding of the next layer. When a correction notice is generated, the gradient of the comparative association loss function with respect to the current node embedding vector is obtained, and the node embedding is updated according to the product of the gradient and the incremental learning rate to obtain the updated node embedding.

[0014] On the other hand, the present invention also provides a bidding project identifier generation device, comprising: The text information extraction module is used to obtain the target tender text and, based on the target tender text, extract the core metadata entities and unstructured project descriptions. The first vector extraction module is used to generate a project core metadata feature vector based on the project core metadata entity, and to perform a hash operation on the project core metadata feature vector to obtain a metadata hash value. The second vector extraction module is used to transform the unstructured project description into a deep semantic fingerprint vector in a high-dimensional semantic space by employing contrastive learning and the Transformer architecture. The fusion generation module is used to fuse the metadata hash value with the deep semantic fingerprint vector to obtain fused identifier data, and to perform a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier.

[0015] Thirdly, the present invention also provides an electronic device, including a communication interface, a memory, and a processor, wherein, The communication interface is used to acquire the target tag text; The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the bid project identifier generation method described in any of the above implementations.

[0016] The beneficial effects of this invention are as follows: The bidding project identifier generation method provided by this invention extracts key metadata entities and unstructured descriptions from the bidding announcement text, generates hash values ​​and deep semantic fingerprints respectively, and then fuses them for conflict detection. This method combines structured information and unstructured semantic features from the bidding announcement to form a global identifier resistant to textual perturbation. Specifically, contrastive learning and Transformer are used to generate semantic fingerprints, ensuring that projects with different descriptions but identical essence receive similar fingerprint vectors, thus avoiding identifier drift caused by format differences such as abbreviations. Furthermore, conflict detection is performed on the fusion result to further eliminate the risk of identifier duplication, effectively improving the accuracy of cross-platform bidding project identification and the stability of unique identifier generation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic flowchart of an embodiment of the bidding project identifier generation method provided by the present invention; Figure 2 For the present invention Figure 1 A schematic diagram of an embodiment of S101; Figure 3 For the present invention Figure 1 A schematic diagram of an embodiment of S102; Figure 4 For the present invention Figure 1 A schematic diagram of an embodiment of S103; Figure 5 For the present invention Figure 1 A schematic diagram of an embodiment of S104; Figure 6 For the present invention Figure 1 A schematic diagram of another embodiment of S104; Figure 7 A schematic flowchart of another embodiment of the bidding project identifier generation method provided by the present invention; Figure 8 A schematic diagram of an embodiment of the bidding project identifier generation device provided by the present invention; Figure 9 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0019] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0020] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] This invention provides a method, apparatus, and electronic device for generating bid project identifiers. The technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1 This is a schematic flowchart of an embodiment of the bidding project identifier generation method provided by the present invention, as shown below. Figure 1 As shown, the methods for generating the tender project identifier include: S101. Obtain the target tender text, and based on the target tender text, extract the core metadata entities and unstructured project descriptions. S102. Based on the project's core metadata entity, generate the project's core metadata feature vector, and perform a hash operation on the project's core metadata feature vector to obtain the metadata hash value. S103. Using contrastive learning and the Transformer architecture, unstructured project descriptions are transformed into deep semantic fingerprint vectors in a high-dimensional semantic space. S104. The metadata hash value and the deep semantic fingerprint vector are fused to obtain fused identifier data, and a conflict detection algorithm is executed on the fused identifier data to generate a globally unique bid project identifier.

[0024] Specifically, electronic devices obtain target tender information text from one or more tender information publishing platforms (such as government procurement websites or enterprise bidding systems) through communication interfaces.

[0025] The tender notice text may include the project name, the main body of the tender notice, the information of the tendering party, the budget amount, etc. The electronic device parses the text and extracts the core metadata entities of the project (such as the tendering party, project amount, construction location) as well as the remaining unstructured project descriptions (such as project overview, scope of tender, and other free text).

[0026] As an example, the metadata entities extracted from a certain tender notice text include "XX Transportation Bureau" as the tenderer, "50 million" as the project amount, and the unstructured description is "The first phase of the intelligent transportation project includes the installation of roadside sensing equipment and the development of the system platform".

[0027] After obtaining the metadata entity, this embodiment organizes it into a fixed-dimensional numerical vector, namely the project's core metadata feature vector. This vector can numerically characterize the project's key business attributes. A hash operation is then performed on this feature vector to obtain a fixed-length metadata hash value. The hash operation is irreversible and collision-resistant, ensuring that subtle changes in the metadata result in significant changes to the hash value, thus providing a unique basis for subsequent identifier generation.

[0028] For the unstructured project description portion, this embodiment uses a combination of contrastive learning and a Transformer architecture for processing.

[0029] Specifically, unstructured item descriptions are input into a pre-trained transformer encoder, which uses a self-attention mechanism to capture semantic relationships within the text and outputs a high-dimensional deep semantic fingerprint vector. This fingerprint vector represents the "semantic coordinates" of the item in the semantic space, causing items with similar descriptions to be close to each other in space, while items with different descriptions are far apart.

[0030] Next, the previously obtained metadata hash value and the deep semantic fingerprint vector are fused. The fusion method can employ bitwise XOR operations or concatenation followed by secondary hashing, resulting in a fused identifier. This fused identifier simultaneously carries structured metadata information and semantic-level deep features.

[0031] Finally, a conflict detection algorithm is executed on the fused identifier data to check whether the currently generated identifier is duplicated with an existing identifier. If no duplicate is found, it is directly output as a globally unique bid project identifier; if a duplicate is detected, a conflict handling mechanism (such as adding a spatiotemporal factor or version number) is triggered to regenerate until a globally unique identifier is obtained.

[0032] By combining structured metadata hashing with deep semantic fingerprinting, this embodiment can both use metadata to ensure the sensitivity of identifiers to key business fields and use semantic fingerprinting to resist disturbances from non-core textual expressions, thereby generating stable and unique cross-platform project identifiers, effectively improving the accuracy and anti-interference capability of bidding project identification.

[0033] In some embodiments of the present invention, such as Figure 2 As shown, step S101 extracts the core metadata entities and unstructured project description based on the target tender text, including: S201. The target tag text is encoded using the BiLSTM-CRF model to obtain the output of the encoding layer; S202. Based on the output of the coding layer, the optimal prediction probability of the label sequence is determined by the conditional random field layer. S203. Based on the tag sequence corresponding to the optimal prediction probability, extract key entities from the target tender text as core metadata entities for the project. The key entities include at least one of the following: tenderer, project amount, and construction location.

[0034] In this embodiment, the process of extracting the core metadata entity of the project in step S101 is implemented using a bidirectional long short-term memory network and a conditional random field (BiLSTM-CRF) model.

[0035] Specifically, the target target text is first segmented by character or word to obtain serialized input.

[0036] The input is then fed into a BiLSTM network, which uses LSTM layers in both the forward and backward directions to capture the contextual information at each position and outputs a sequence of encoding layer output vectors containing contextual semantics.

[0037] Next, a Conditional Random Field (CRF) layer is superimposed on the BiLSTM network. The role of the CRF layer is to use the label transition matrix to globally constrain the rationality of the label sequence.

[0038] For example, in the BIO labeling system, the "I-Org" label cannot appear before the "B-Org" label. This dependency between labels can be reflected by the probability values ​​in the transition matrix. The CRF layer receives the output of the coding layer as the emission matrix (i.e., the score of each position mapped to each label), combines it with the label transition matrix, calculates the normalized probability of all possible label sequences, and selects the label sequence with the highest probability as the optimal prediction result.

[0039] In one example, suppose the tender notice text contains a fragment such as "Tenderer: XX Transportation Bureau". After processing with BiLSTM-CRF, the model will label "XX Transportation Bureau" as "B-Org" and "I-Org", thus identifying it as the tenderer entity. Similarly, "50 million yuan" in "Investment amount: 50 million yuan" might be labeled as the "B-Money" entity.

[0040] The extracted key entities include the tenderer, project amount, construction location, etc., which constitute the core metadata entity set of the project.

[0041] Optionally, more entity types can be added based on business needs, such as "project number" and "contact person".

[0042] Furthermore, the label transition matrix A and emission matrix E involved in the BiLSTM-CRF model in this embodiment are constructed as follows: The label transition matrix has a dimension of (L+2)×(L+2), where L is the total number of preset label categories (e.g., "B-Org", "I-Org", "B-Money", "O", etc.), and the two additional dimensions are used to represent the start and end markers of the sequence. In the early stage of model training, the values ​​in the transition matrix can be randomly initialized. During training, through the backpropagation algorithm, each element A_{i,j} in the matrix is ​​continuously updated according to the actual statistical rules of the transition from label i to label j in the training corpus, automatically learning the legal order constraints between labels (e.g., "I-Org" must be preceded by "B-Org" or "I-Org", and cannot be "B-Money").

[0043] The emission matrix E is generated based on the output H of the BiLSTM encoding layer. The element E_{i,y_i} in the matrix represents the denormalized score of the feature vector at position i in the input sequence mapped to the label y_i. The learning objective of this matrix is ​​to capture the correlation strength between the current context information and a specific label. For example, when the context includes a prompt word like "bidder:", the emission matrix will increase the score for subsequent words labeled "B-Org". Through joint decoding with the CRF layer, the model can output the globally optimal label sequence, thereby accurately extracting key entities such as the bidder and project amount. This approach, combining the contextual modeling capabilities of BiLSTM and the sequence constraint capabilities of CRF, achieves higher extraction accuracy and generalization performance compared to independently labeled entities or rule-based methods.

[0044] This embodiment uses the BiLSTM-CRF model for sequence labeling, which can automatically learn entity boundaries and types in text without relying on manual rules, thereby improving the accuracy and generalization ability of metadata extraction and providing reliable input for the subsequent generation of stable item identifiers.

[0045] In some embodiments of the present invention, such as Figure 3 As shown, step S102 performs a hash operation on the core metadata feature vector of the project to obtain the metadata hash value, including: S301. Encode the core metadata feature vector of the project into hexadecimal format to obtain the encoded string; S302. Use a secure hash algorithm to perform a hash operation on the encoded string to obtain the metadata hash value.

[0046] In this embodiment, the specific implementation of hashing the feature vector of the project's core metadata in step S102 is as follows.

[0047] First, the generated project core metadata feature vector (usually an array of floating-point numbers or integers) is converted into a byte sequence suitable for hashing. A common method is to convert it to a hexadecimal string based on its numeric type. For example, assuming the feature vector is [128, 45, 200], the converted hexadecimal string can be a concatenated form of "80 2D C8". This string can completely preserve the numerical information of each component in the vector.

[0048] After obtaining the encoded string, a secure hash algorithm is used to perform a hash operation on it. A secure hash algorithm is a cryptographic hash function that can map an input of arbitrary length to an output of fixed length, and it has one-way transmission and strong collision resistance.

[0049] In this embodiment, the preferred algorithm is SHA-256 (a 256-bit variant of the Secure Hash Algorithm), which has an output length of 256 bits (32 bytes). Taking the hexadecimal string in the example as input, the SHA-256 operation yields a fixed-length hexadecimal digest, i.e., the metadata hash value.

[0050] It should be noted that the SHA-256 algorithm is only a preferred implementation. In alternative implementations, other secure hash algorithms such as SHA-3 and the Chinese national cryptographic standard SM3 can also be used, as long as they meet the requirements of collision resistance and irreversibility.

[0051] Through the processing in this embodiment, even if the metadata feature vector undergoes a slight numerical change (such as a change due to a conversion of monetary units), its hash result will produce a significant difference, thereby ensuring the identifier's sensitive detection capability for the authenticity of metadata. At the same time, the irreversible nature of hash operations also helps protect sensitive commercial information in the original identifier.

[0052] In some embodiments of the present invention, such as Figure 4As shown, step S103 employs contrastive learning and the Transformer architecture to transform unstructured project descriptions into deep semantic fingerprint vectors, including: S401. Input the unstructured item description into the Transformer encoder, and use the Transformer encoder with different random inactivation masks to perform two forward propagations on the same unstructured item description to obtain two semantic representations as positive sample pairs. S402. Obtain descriptions of other items in the same training batch as negative sample pairs, and select semantically similar but different identifier items as hard negative samples; S403. Based on positive sample pairs, negative sample pairs, and hard negative samples, a contrastive learning loss function is used to train the Transformer encoder, so that the semantic representations of similar items are close in the semantic space and the semantic representations of different items are far apart. S404. Pool the semantic embeddings output by the trained Transformer encoder to obtain a deep semantic fingerprint vector.

[0053] In this embodiment, the process of generating deep semantic fingerprint vectors in step S103 introduces a contrastive learning mechanism. The core objective of contrastive learning is to bring semantically similar text representations closer together while pushing apart semantically different text representations. This embodiment is based on SimCSE (Simple Contrastive Learning Sentence Embedding) and utilizes the random dropout mechanism within the transformer encoder to construct positive sample pairs.

[0054] Specifically, an unstructured project description (such as a text outlining the scope of a project tender) is input into the transformer encoder. This encoder contains multiple randomly deactivated layers. During each forward pass, these layers randomly discard a portion of neurons, resulting in slight differences in the output vectors obtained from two forward passes, even with the same input. This embodiment leverages this characteristic by performing two consecutive forward passes on the same descriptive text, yielding two semantic representation vectors. These two vectors are then used as positive sample pairs. Positive sample pairs represent different views of the same project and should theoretically have a very high degree of similarity.

[0055] Meanwhile, negative sample pairs are constructed using an in-batch negative sampling strategy. In a training batch, all other item descriptions within the batch, except for the current item itself and its positive samples, are considered negative samples of the current item.

[0056] In addition, to improve the model's discriminative ability, hard negative sample mining was introduced, which deliberately selects text pairs that are semantically similar but actually belong to different projects from the dataset (such as two "smart transportation" projects in different regions but with similar names), forcing the model to learn more subtle semantic differences.

[0057] During training, a contrastive learning loss function (such as InfoNCE loss) is used to optimize the parameters of the transformer encoder. This loss function encourages the vector representations of positive sample pairs to approach 1 in cosine distance, while forcing the distance of negative sample pairs to approach 0.

[0058] After sufficient training, the semantic embedding output by the transformer encoder has good semantic clustering properties. Finally, pooling is performed on the output—usually by taking the [CLS] bit vector from the encoder output, or by performing mean pooling on all word vectors—to obtain a fixed-dimensional deep semantic fingerprint vector.

[0059] It should be understood that the comparative learning training process can be pre-trained on publicly available bidding and tendering corpora, or fine-tuned on business data.

[0060] The deep semantic fingerprint vectors generated in this way can effectively resist non-substantial textual changes in project descriptions (such as typos, word order adjustments, and synonym replacements), making the fingerprint distance between different versions of the same project in the semantic space extremely close, and different projects can be effectively distinguished even if their titles are similar.

[0061] In some embodiments of the present invention, after generating a deep semantic fingerprint vector, this vector can be used to determine whether two tender texts from different sources belong to the same project. Specifically, the deep semantic fingerprint vector F1 of the first tender project and the deep semantic fingerprint vector F2 of the second tender project are obtained, and the cosine similarity between the two vectors is calculated. S The formula for calculating cosine similarity is the product of the dot product and the modulus. The value of cosine similarity ranges from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors in the semantic space, that is, the closer the descriptions of the two items are.

[0062] Then, obtain a pre-set threshold. τ (For example, 0.95). The calculated cosine similarity... S With threshold τ Comparison: If S Greater than τ If the first and second bidding projects are deemed to be substantially the same, the system assigns them the same globally unique bidding project identifier; if S Less than or equal to τ If they are not, they are classified as different items and assigned different identifiers. As a preferred method, a threshold... τ The threshold can be dynamically adjusted according to the business scenario and the required level of stringency. For example, a higher threshold (such as 0.98) can be set for highly sensitive projects, while a threshold of 0.92 can be set for general projects.

[0063] This embodiment enables item deduplication and association based on semantic similarity. Even if two item descriptions differ significantly in wording (such as "Intelligent Transportation Phase I" and "Intelligent Transportation Phase 1 Project"), as long as the cosine similarity of the semantic fingerprint exceeds the threshold, the system can automatically merge them into the same item, significantly improving the recall and accuracy of cross-platform item identification.

[0064] In some embodiments of the present invention, such as Figure 5 As shown, step S104 fuses the metadata hash value with the deep semantic fingerprint vector, including: S501. Input the deep semantic fingerprint vector into the encoder for encoding to obtain the encoded semantic features; S502. The encoded semantic features are reduced in dimension by a dimension reduction mapping function to obtain the reduced semantic features. The dimension reduction mapping function is implemented by a fully connected neural network layer. S503. Perform an XOR operation between the metadata hash value and the dimensionality-reduced semantic features to obtain the fused identifier data.

[0065] In this embodiment, the fusion process of metadata hash value and deep semantic fingerprint vector in step S104 is implemented in the following way.

[0066] First, the deep semantic fingerprint vector obtained in step S103 is input into an additional encoder for further encoding. This encoder can be a lightweight fully connected network or keep the original fingerprint unchanged (depending on the specific design). Its purpose is to perform appropriate nonlinear transformations on the semantic features to facilitate subsequent fusion operations.

[0067] Next, a dimensionality reduction mapping function is used to reduce the dimensionality of the encoded semantic features. Dimensionality reduction is necessary because deep semantic fingerprint vectors are typically high-dimensional (e.g., 768 or 1024 dimensions), while hash operations and identifier storage often require a shorter bit width. Simple physical truncation (e.g., taking the first N dimensions) will lose the semantic information of the latter half of the vector, potentially leading to different items producing the same vector fragment after truncation.

[0068] Therefore, the dimensionality reduction mapping function in this embodiment is implemented by a fully connected neural network layer, which maps the high-dimensional input to a low-dimensional dense vector space through a weight matrix. This fully connected layer can be simultaneously optimized during the contrastive learning training process, enabling it to learn to retain the most semantically discriminative feature components, thereby maximizing the retention of feature entropy.

[0069] As a preferred approach, the target dimension after dimensionality reduction can be matched with the number of bits in the metadata hash value, for example, both being 256 bits.

[0070] Subsequently, the dimensionality-reduced semantic features are XORed with the previously obtained metadata hash value. XOR is a bitwise logical operation; if two binary bits are the same, the result is 0, and if they are different, the result is 1.

[0071] Through XOR operations, the metadata hash value and semantic features are combined to form the final fused identifier data. This fused identifier data carries both structured business information and unstructured semantic information, and the extremely poor reversibility of XOR operations enhances the identifier's resistance to tampering.

[0072] It should be noted that the reason why the dimensionality reduction mapping function in this embodiment uses a fully connected neural network layer instead of a simple physical truncation of the first N dimensions is that simple truncation directly discards the semantic features of the latter half of the high-dimensional vector, which may cause items with different distinguishing qualities to generate the same low-dimensional vector fragments after truncation, thus causing identifier conflicts. The fully connected layer, by performing weighted linear combination and nonlinear transformation on all dimensions of the original vector, compresses global information into a low-dimensional space, so that each dimension after dimensionality reduction contains comprehensive semantic information from the original high-dimensional space. The parameters of this fully connected layer can be optimized synchronously during the contrastive learning training process, thereby learning the optimal dimensionality reduction projection direction and maximizing the preservation of feature entropy. Experiments show that using a fully connected layer for dimensionality reduction, compared to simple truncation, can reduce the probability of identifier conflicts by about an order of magnitude for the same number of bits.

[0073] Through the processing in this embodiment, the fused identifier data retains the unique identification capability of the metadata and embeds deep semantic features, providing information-rich intermediate results for subsequent conflict detection and identifier generation.

[0074] In some embodiments of the present invention, such as Figure 6 As shown, step S104 performs a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier, including: S601. Use the merged identifier data as the initial draft of the identifier for the bidding project; S602. Access the global UID index library to check if the initial draft of the bidding project identifier already exists; S603. Configure a Bloom filter in the front end of the global UID index library. Use the Bloom filter to perform an existence check on the initial draft of the bid project identifier. If the Bloom filter determines that it does not exist, it will pass directly and the initial draft of the bid project identifier will be output as the final identifier. S604. If the Bloom filter determines that a conflict may exist or the database query confirms that a conflict has already existed, then a spatiotemporal perturbation factor containing a nanosecond-level timestamp and a server node identifier is obtained. The spatiotemporal perturbation factor is appended to the fused identifier data, and a new hash operation is performed to generate a new bid project identifier. S605. When fingerprint overlap is caused by project changes or correction announcements, an incrementing version identifier shall be appended to the end of the newly generated bid project identifier.

[0075] In this embodiment, the conflict detection and identifier generation process in step S104 has the following specific implementation.

[0076] First, the fused identifier data obtained in the previous step is directly used as the initial draft of the tender project identifier. However, due to hash collisions or semantic fingerprint overlap in extreme cases, this initial draft may duplicate existing identifiers. Therefore, collision detection needs to be performed.

[0077] Electronic devices maintain a global UID index that stores all historically generated item identifiers. The device accesses this index via the network to check if a draft already exists. To speed up the query process, a Bloom filter is configured in front of the index. A Bloom filter is a highly space-efficient probabilistic data structure that can determine whether an element "absolutely does not exist" or "may exist." When the Bloom filter determines that the draft does not exist, it can be assured that the identifier is unique and directly outputs it as the final identifier without accessing the backend database. If the Bloom filter determines that it may exist, it further queries the actual index to confirm whether it is indeed a duplicate.

[0078] Once a conflict is confirmed, this embodiment will not simply add random numbers, but will instead employ a spatiotemporal perturbation factor strategy.

[0079] Specifically, the nanosecond-level timestamp of the current generation process and the server node identifier (Node ID) that generated the identifier are obtained. These two pieces of information are concatenated and appended to the end of the original merged identifier data, and then a new hash operation is performed. Because nanosecond-level timestamps have extremely high uniqueness, the rehashed identifier is almost impossible to conflict with again.

[0080] Furthermore, correction notices and change instructions frequently appear in bidding projects. These notices are semantically very similar to the original project, which may lead to fingerprint overlap, but legally they belong to different record entries. To handle this situation, this embodiment adopts an incremental version number method. When fingerprint overlap caused by project changes is detected, the system appends a version identifier, such as "-V1" or "-V2", to the end of the final output bidding project identifier. This version number increments with each change, thereby distinguishing different versions of the same project without changing the core identifier.

[0081] Alternatively, the version number can be automatically generated using the evolutionary relationships maintained by a graph neural network.

[0082] In this embodiment, the conflict detection algorithm ensures the global uniqueness of identifiers in the context of massive data, while the spatiotemporal perturbation factor and version number strategy provide a complete conflict resolution mechanism, which helps to improve the robustness of identifier generation and the clarity of project version management.

[0083] In some embodiments of the present invention, such as Figure 7 As shown, the method for generating the tender project identifier includes: S701. Obtain the generated bid project identifier and use the bid project identifier as a node in the graph neural network. S702. Obtain the legal and logical relationships and temporal evolution relationships between nodes, and construct the edges in the graph neural network based on the legal and logical relationships and temporal evolution relationships. Each edge carries an announcement timestamp and an announcement type identifier factor. S703. Employs a graph neural network to maintain the evolutionary relationships between projects based on nodes and edges. These relationships include dynamic tracking of project changes and correction announcements.

[0084] In this embodiment, the method further includes a step of dynamically tracking the entire lifecycle of the project.

[0085] Specifically, after generating each bid project identifier (UID), the UID is used as a node in a graph neural network (GNN). In addition to storing the UID itself, each node also encapsulates the corresponding metadata feature vector and deep semantic fingerprint vector, enabling the node to identify projects that are essentially the same but have slightly different descriptions.

[0086] Subsequently, edges are constructed between nodes based on the legal logic and chronological order of the projects. Edge types include, but are not limited to, "corrected from," "winning bid from," and "change originating from," each representing a different legal relationship. Each edge also carries the precise timestamp of the announcement's publication (e.g., "2026-05-28 14:30:00.123456") and an announcement type identifier factor (e.g., 0 for the original announcement, 1 for a correction announcement, and 2 for a winning bid announcement). The edge's timestamp and type factor together constitute the chronological thread of the project's evolution.

[0087] In one example, when a bidding project is first published, the system generates node Node_A. Subsequently, the project publishes its first correction notice. The system recognizes that the new notice has a very high semantic fingerprint similarity to Node_A, so it generates a new node Node_B and adds a directed edge between Node_B and Node_A, labeled "Correction," with the correction notice's timestamp and type factor 1. Later, the project is awarded, generating node Node_C, and adding an edge "Winning Bid" between Node_C and the latest node (Node_B), with the winning bid notice's timestamp and type factor 2.

[0088] By employing graph neural networks (GNNs), the features of all nodes along an evolutionary path can be progressively fused through message passing and node aggregation mechanisms. For example, from Node_A to Node_B and then to Node_C, GNNs can integrate the metadata and semantic information of the three nodes to form a final global view of the project. Even if the project undergoes multiple corrections or supplementary announcements, its complete evolutionary history is still structurally preserved in the graph. This embodiment uses GNNs to dynamically track project evolution relationships, which is beneficial for subsequent project risk analysis, historical backtracking, and correlation queries of multiple versions of announcements, thereby improving the information integrity of the entire lifecycle management of bidding projects.

[0089] In some embodiments of the present invention, step S703 employs a graph neural network to maintain the evolutionary relationship between projects based on nodes and edges, including: The embedding vector of the current node and the set of neighboring nodes of the current node are obtained by using graph sampling and aggregation algorithms; Aggregate the embeddings of neighbor nodes in the neighbor node set to obtain the aggregation result; The embedding vector of the current node is concatenated with the aggregation result, and a nonlinear transformation is performed on the concatenated vector to generate the node embedding of the next layer. When a correction notice is generated, the gradient of the comparative association loss function with respect to the current node embedding vector is obtained, and the node embedding is updated based on the product of the gradient and the incremental learning rate to obtain the updated node embedding.

[0090] In this embodiment, the specific training and update process of the graph neural network is implemented using the Graph Sampling and Aggregation Algorithm (GraphSAGE). GraphSAGE learns an aggregation function that allows it to collect information from a node's neighboring nodes and update the embedding representation of the current node.

[0091] Specifically, for any current node in the graph, we first obtain the node's current embedding vector and the set of embedding vectors of all its neighboring nodes. Neighboring nodes refer to other nodes directly connected to this node, such as the two nodes connected by the "correction" edge. Then, we perform an aggregation operation, inputting the embedding vectors of the neighboring nodes into an aggregation function (such as mean aggregation, max pooling aggregation, or LSTM aggregation) to obtain an aggregated result vector, which condenses the neighborhood information.

[0092] Next, the original embedding vector of the current node is concatenated with the aggregation result.

[0093] Finally, the concatenated vectors are fed into a fully connected layer with a non-linear activation function (such as ReLU) to output the node embedding representation for the next layer. Through multiple iterations, the embedding of each node gradually incorporates information from its multi-hop neighbors, thus reflecting the context of the entire project's evolution path.

[0094] When the system detects a new correction notice and generates a new node, the existing node embeddings need to be fine-tuned based on the new information to maintain the consistency of the graph logic.

[0095] This embodiment employs an incremental update mechanism. First, a contrastive association loss function is defined, which measures the semantic consistency and topological rationality between the new announcement node and the old nodes. Using the backpropagation algorithm, the gradient of the loss function with respect to the current node embedding vector is calculated. The gradient indicates the direction and magnitude in which the original node embedding should move in the semantic space to ensure that the new announcement accurately points to the same UID core node. Then, the gradient is multiplied by an incremental learning rate, and the product is added to the original node embedding vector to obtain the updated node embedding.

[0096] Incremental learning rate ρ It is a hyperparameter that is dynamically adjusted according to the actual situation, based on the following two factors.

[0097] The first factor is the semantic fingerprint consistency coefficient. This involves obtaining the deep semantic fingerprint vector corresponding to the correction announcement and the deep semantic fingerprint vector of the original announcement (or the previous version announcement), and calculating their cosine similarity. A higher similarity indicates that the two announcements are semantically closer, resulting in a smaller correction magnitude. In this case, the system will... ρ Set a larger value (e.g., above 0.8) to quickly integrate new node features into existing nodes, reducing the number of iterations. Conversely, if the similarity is low (e.g., below 0.7), it indicates that the correction involves a large amount of content, and a smaller value should be set. ρ (e.g., 0.2) to prevent excessive updates from causing map distortion.

[0098] The second factor is the legal time decay coefficient, which measures the time difference between the publication of the correction notice and the original notice. Considering the time-sensitive nature of bidding projects, a correction notice published more than three months later has little effect on correcting the initial project characteristics. Therefore, ρ The size is inversely proportional to the time difference between releases: the larger the time difference, the greater the impact of the time difference. ρ The smaller the value, the better. For example, corrections within a week can be implemented using... ρ =0.5, while corrections older than one month will automatically be... ρ Reduced to 0.1.

[0099] Through this two-factor dynamic adjustment mechanism, the incremental learning process can quickly respond to recent changes made by users while maintaining the stability and authenticity of long-term graphs, which is conducive to improving the accuracy of project evolution tracking.

[0100] Optionally, the settings can be based on the deep semantic fingerprint similarity between the two announcements and the time difference between their publication. If the semantic fingerprint similarity is extremely high, it indicates that the two announcements are essentially the same. ρ Set a larger value for faster fusion; if the similarity is low, then... ρ Smaller time differences prevent incorrect associations. The larger the time difference, the lower the weight of the change's impact on the initial node.

[0101] By employing this incremental graph learning approach, this embodiment can efficiently adapt to dynamic changes in the project while ensuring the stability of the graph, thereby improving the real-time performance and accuracy of evolution tracking.

[0102] To make the technical solution of this invention easier to understand, an actual bidding project is used as an example for explanation below. Suppose that a province has launched the "Smart Transportation Phase I" project, and the bidding announcement for this project is published on three different government procurement platforms, A, B, and C. The titles of the three announcements are slightly different: the title on platform A is "Public Bidding Announcement for Smart Transportation Phase I Project", the title on platform B is "Bidding for Smart Transportation Phase I (including roadside equipment) in a Certain Province", and the title on platform C is "Smart Transportation Phase I (Re-announcement)".

[0103] According to the implementation of this invention, the system obtains tender notice texts from three platforms respectively. The extracted metadata entities are all (Tenderer: XX Transportation Bureau, Amount: 50 million yuan). Although the wording of the unstructured descriptions differs, the core semantics are consistent. After Transformer encoding and comparative learning, the cosine similarity between each pair of deep semantic fingerprint vectors of the three announcements reaches 0.992 or higher, which is greater than the preset threshold of 0.95. The system determines that the three refer to the same project and assigns them the same globally unique identifier "TAX-2026-001". At the same time, this UID node is created in the graph, and the three source announcements are associated with this UID as three different "release" event nodes. Each association edge records the source of the announcement and the timestamp. Subsequently, the project releases a correction announcement and a winning bid announcement. The system continues to add new nodes and "correction" and "winning bid" edges, eventually forming a complete project lifecycle graph. This graph can be displayed using visualization tools (such as D3.js) for business personnel to trace the entire project.

[0104] As can be seen from the above examples, the present invention can effectively integrate the same project information scattered across multiple platforms, automatically deduplicate and generate a unified identifier, and support full-process tracking of project evolution, significantly improving the information integration efficiency of bidding data processing.

[0105] To better implement the bid project identifier generation method in this embodiment of the invention, based on the bid project identifier generation method, correspondingly, such as... Figure 8 As shown, this embodiment of the invention also provides a bidding project identifier generation device, the bidding project identifier generation device 800 comprising: The text information extraction module 801 is used to obtain the target tender text and, based on the target tender text, extract the core metadata entities of the project and the unstructured project description. The first vector extraction module 802 is used to generate a project core metadata feature vector based on the project core metadata entity, and to perform a hash operation on the project core metadata feature vector to obtain a metadata hash value. The second vector extraction module 803 is used to transform the unstructured project description into a deep semantic fingerprint vector in a high-dimensional semantic space by employing contrastive learning and the Transformer architecture. The fusion generation module 804 is used to fuse the metadata hash value with the deep semantic fingerprint vector to obtain fused identifier data, and to perform a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier.

[0106] The bidding project identifier generation device 800 provided in the above embodiments can realize the technical solutions described in the above bidding project identifier generation method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above bidding project identifier generation method embodiments, which will not be repeated here.

[0107] like Figure 9 As shown, the present invention also provides an electronic device 900. The electronic device 900 includes a processor 901, a memory 902, a display 903, and a communication interface 904. Figure 9 Only some components of the electronic device 900 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0108] In some embodiments, processor 901 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 902 or process data, such as the bidding project identifier generation method in this invention.

[0109] In some embodiments, processor 901 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 901 may be local or remote. In some embodiments, processor 901 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, etc., or any combination thereof.

[0110] In some embodiments, memory 902 may be an internal storage unit of electronic device 900, such as a hard disk or memory of electronic device 900. In other embodiments, memory 902 may also be an external storage device of electronic device 900, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 900.

[0111] Furthermore, the memory 902 may include both internal storage units of the electronic device 900 and external storage devices. The memory 902 is used to store application software and various types of data installed on the electronic device 900.

[0112] In some embodiments, display 903 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. Display 903 is used to display information from electronic device 900 and to display a visual user interface.

[0113] Communication interface 904 is used to acquire target tender information text. Specifically, communication interface 904 can be a wired communication interface or a wireless communication interface, used to receive raw tender information text data from external data sources, bidding platforms, government procurement websites, or database servers. The acquired target tender information text is transmitted to processor 901 to execute the steps in the tender project identifier generation method described in this invention. In some embodiments, communication interface 904 is also used to establish a communication connection with a globally unique identifier index library to complete a conflict detection query operation for tender project identifiers.

[0114] In one embodiment, when the processor 901 executes the bid item identifier generation program in the memory 902, the following steps can be implemented: Obtain the target tender text, and based on the target tender text, extract the core metadata entities and unstructured project description; Based on the project's core metadata entity, generate the project's core metadata feature vector, and perform a hash operation on the project's core metadata feature vector to obtain the metadata hash value; By employing contrastive learning and the Transformer architecture, unstructured project descriptions are transformed into deep semantic fingerprint vectors in a high-dimensional semantic space; The metadata hash value and the deep semantic fingerprint vector are fused to obtain fused identifier data. A conflict detection algorithm is then performed on the fused identifier data to generate a globally unique bid project identifier.

[0115] It should be understood that when the processor 901 executes the tender project identifier generation program in the memory 902, in addition to the functions mentioned above, it can also perform other functions, as detailed in the description of the corresponding method embodiments above.

[0116] Furthermore, the embodiments of the present invention do not specifically limit the type of the electronic device 900 mentioned. The electronic device 900 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 900 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0117] Accordingly, this application also provides a computer-readable storage medium for storing a computer-readable program or instruction. When the program or instruction is executed by a processor, it can implement the steps or functions of the bidding project identifier generation method provided in the above-described method embodiments.

[0118] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0119] The above provides a detailed description of the bidding project identifier generation method, apparatus, and electronic device provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for generating a bid project identifier, characterized in that, include: Obtain the target tender text, and based on the target tender text, extract the core metadata entities and unstructured project descriptions; Based on the project's core metadata entity, a project core metadata feature vector is generated, and a hash operation is performed on the project core metadata feature vector to obtain the metadata hash value; By employing contrastive learning and the Transformer architecture, the unstructured project description is transformed into a deep semantic fingerprint vector in a high-dimensional semantic space. The metadata hash value is fused with the deep semantic fingerprint vector to obtain fused identifier data, and a conflict detection algorithm is executed on the fused identifier data to generate a globally unique bid project identifier.

2. The method for generating a bid project identifier according to claim 1, characterized in that, The step of extracting core metadata entities and unstructured project descriptions based on the target tag text includes: The target tag text is encoded using a BiLSTM-CRF model to obtain the encoding layer output; Based on the output of the coding layer, the optimal prediction probability of the label sequence is determined using a conditional random field layer; Based on the tag sequence corresponding to the optimal prediction probability, key entities are extracted from the target tender text as the core metadata entities of the project. The key entities include at least one of the following: tenderer, project amount, and construction location.

3. The method for generating bid project identifiers according to claim 1, characterized in that, The method employs contrastive learning and the Transformer architecture to transform the unstructured project description into a deep semantic fingerprint vector, including: The unstructured item description is input into the Transformer encoder, and the same unstructured item description is forwarded twice using the Transformer encoder with different random deactivation masks to obtain two semantic representations as positive sample pairs. Obtain descriptions of other items in the same training batch as negative sample pairs, and select semantically similar but different identifier items as hard negative samples; Based on the positive sample pairs, the negative sample pairs, and the hard negative samples, the Transformer encoder is trained using a contrastive learning loss function, so that the semantic representations of similar items are close in distance in the semantic space, and the semantic representations of different items are far apart. The semantic embeddings output by the trained Transformer encoder are pooled to obtain the deep semantic fingerprint vector.

4. The method for generating a bid project identifier according to claim 1, characterized in that, The process of performing a hash operation on the core metadata feature vector of the project to obtain the metadata hash value includes: The core metadata feature vector of the project is encoded in hexadecimal to obtain the encoded string; The encoded string is hashed using a secure hash algorithm to obtain the metadata hash value.

5. The method for generating a bid project identifier according to claim 1, characterized in that, The step of fusing the metadata hash value with the deep semantic fingerprint vector includes: The deep semantic fingerprint vector is input to the encoder for encoding to obtain the encoded semantic features; The encoded semantic features are dimensionality-reduced by using a dimensionality-reducing mapping function, which is implemented by a fully connected neural network layer. The fused identifier data is obtained by performing an XOR operation between the metadata hash value and the dimensionality-reduced semantic features.

6. The method for generating a bid project identifier according to claim 1, characterized in that, The process of performing a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier includes: The fused identifier data will be used as the initial draft of the tender project identifier; Access the global UID index to check if the initial draft of the bidding project identifier already exists; A Bloom filter is configured at the front end of the global UID index library. The Bloom filter is used to perform an existence check on the initial draft of the bidding project identifier. If the Bloom filter determines that the initial draft of the bidding project identifier does not exist, it is passed directly and the initial draft of the bidding project identifier is output as the final identifier. If the Bloom filter determines that a conflict may exist or the database query confirms that a conflict has already existed, a spatiotemporal perturbation factor containing a nanosecond-level timestamp and a server node identifier is obtained. The spatiotemporal perturbation factor is then appended to the fused identifier data, and a new hash operation is performed to generate a new bidding project identifier. When fingerprint overlap occurs due to project changes or correction notices, an incrementing version identifier is appended to the end of the newly generated bid project identifier.

7. The method for generating a bid project identifier according to any one of claims 1 to 6, characterized in that, include: Obtain the generated bid project identifier and use the bid project identifier as a node in the graph neural network; Obtain the legal and logical relationships and temporal evolution relationships between nodes, and construct the edges in the graph neural network based on the legal and logical relationships and temporal evolution relationships. The edges carry the announcement timestamp and the announcement type identification factor. The graph neural network is used to maintain the evolutionary relationship between projects based on the nodes and edges, and the evolutionary relationship includes the dynamic tracking of project changes and correction announcements.

8. The method for generating a bid project identifier according to claim 7, characterized in that, The method of employing the graph neural network to maintain the evolutionary relationship between projects based on the nodes and edges includes: The embedding vector of the current node and the set of neighboring nodes of the current node are obtained by using graph sampling and aggregation algorithms; Aggregate the neighbor node embeddings in the neighbor node set to obtain the aggregation result; The embedding vector of the current node is concatenated with the aggregation result, and a nonlinear transformation is performed on the concatenated vector to generate the node embedding of the next layer. When a correction notice is generated, the gradient of the comparative association loss function with respect to the current node embedding vector is obtained, and the node embedding is updated according to the product of the gradient and the incremental learning rate to obtain the updated node embedding.

9. A device for generating a tender project identifier, characterized in that, include: The text information extraction module is used to obtain the target tender text and, based on the target tender text, extract the core metadata entities and unstructured project descriptions. The first vector extraction module is used to generate a project core metadata feature vector based on the project core metadata entity, and to perform a hash operation on the project core metadata feature vector to obtain a metadata hash value. The second vector extraction module is used to transform the unstructured project description into a deep semantic fingerprint vector in a high-dimensional semantic space by employing contrastive learning and the Transformer architecture. The fusion generation module is used to fuse the metadata hash value with the deep semantic fingerprint vector to obtain fused identifier data, and to perform a conflict detection algorithm on the fused identifier data to generate a globally unique bid project identifier.

10. An electronic device, characterized in that, Includes communication interfaces, memory, and processors, among which, The communication interface is used to acquire the target tag text; The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the bidding project identifier generation method according to any one of claims 1 to 8.

Images