A Method and System for Tender Document Detection Based on Large Language Model

Abstract

The application relates to the technical field of office automation, and in particular relates to a bidding document detection method and system based on a large language model, wherein a model is trained based on training data to generate a large language model; a target bidding document is acquired and parsed into structured data; the structured data is divided into a plurality of text segments, a mapping relationship between each legal compliance detection point and each text segment is established, all detection points are traversed, and the corresponding text segments are called as inputs to judge whether the text segments exist detection points, and a legal compliance detection result is output; clause content with quantification limitation characteristics in the bidding document is extracted, and a target enterprise knowledge graph is constructed and compared with an enterprise knowledge graph library; meanwhile, whether the bidding document has tendentious expression content is identified, and a fair competition detection result is output. The application can automatically and intelligently detect the bidding document, greatly reduces the cost and time consumption of manual examination, and simultaneously reduces the interference of human subjective factors as much as possible.

Description

Technical Field

[0001] This application relates to the field of office automation technology, and in particular to a method and system for detecting tender documents based on a large language model. Background Technology

[0002] As a core document in the bidding process, the tender document not only contains a detailed description of the technical requirements and commercial conditions, but its compliance and fairness are also crucial to the bidding process. Therefore, the review of tender documents typically involves two aspects: legality and compliance testing, and fair competition testing. The former refers to matching and analyzing the content of the tender document against relevant laws, administrative regulations, and policy documents to determine whether its provisions meet regulatory requirements; the latter focuses on whether the tender document contains factors that affect the equal opportunities for potential bidders, thus impacting the environment of fair competition in form or substance.

[0003] To achieve the aforementioned inspection of bidding documents, two main existing technologies exist. One is the traditional manual inspection method, which relies primarily on manual review. Inspectors often use office software to check formatting and content verification, combining this with their personal experience to determine whether the content of the bidding documents meets compliance and fairness requirements. The other method utilizes traditional localized bidding document inspection software. This type of software typically has a pre-set compliance clause rule library. Users in a local environment use a static matching mechanism to identify clauses and analyze the structure of the bidding documents, outputting the inspection results. Some software also includes features such as category suggestions and keyword annotation to assist reviewers in conducting item-by-item verification.

[0004] However, with the increasing complexity of tender documents and the continuous updating of relevant policies and regulations, human experience-based detection methods are often inefficient when handling large volumes of text or complex logical structures. They are also susceptible to interference from subjective human factors and are slow to respond to policy updates. Meanwhile, software tools based on static rule matching typically rely on fixed templates and keyword settings, making them prone to errors when detecting tender documents with flexible language or deeply nested logic. Therefore, how to intelligently detect tender documents to improve their efficiency and accuracy remains a significant challenge. Summary of the Invention

[0005] This application provides a method and system for detecting tender documents based on a large language model. This system enables intelligent detection of tender documents, significantly reducing the cost and time consumption of manual review, improving the efficiency and accuracy of tender document detection, and minimizing interference from subjective human factors. The technical solution provided in this application is as follows:

[0006] Firstly, this application provides a method for detecting tender documents based on a large language model, the method comprising:

[0007] The pre-prepared training data is used to train the preset model and generate a large language model suitable for the bidding and tendering field.

[0008] Obtain the target tender document to be tested, and parse the target tender document into structured data;

[0009] The system identifies legality and compliance checkpoints, calls the large language model, divides the structured data obtained from parsing the target tender document into several text fragments, establishes a mapping relationship between each checkpoint and each text fragment, iterates through all checkpoints and calls the corresponding text fragments as input, uses the large language model to determine whether there are checkpoints in the text fragments, and outputs the legality and compliance check results.

[0010] The large language model is invoked to extract the clauses with quantitative limiting features from the parsed tender documents and construct a target enterprise knowledge graph. The target enterprise knowledge graph is then compared with an enterprise knowledge graph database. Simultaneously, the large language model is invoked to identify whether the parsed tender documents contain biased expressions. The comparison results of the knowledge graphs and the identification results of the biased expressions are output as fair competition detection results.

[0011] In one specific implementation scheme, training a pre-prepared model based on pre-prepared training data to generate a large language model suitable for the bidding and tendering field includes:

[0012] The training data is cleaned and deduplicated. The training data includes tender document templates, samples, or historical cases from multiple common projects.

[0013] The key bidding elements involved in the training data are labeled and annotated to establish a structured dataset suitable for supervised language modeling.

[0014] Using the seqGPT architecture as the training platform, a pre-labeled structured dataset is input into a pre-defined general-purpose large language model through supervised learning to train a large language model suitable for the bidding and tendering field.

[0015] In a specific feasible implementation, the step of determining whether a text fragment contains a detection point using a large language model and outputting a legality and compliance detection result includes:

[0016] During the detection process, an independent judgment strategy is adopted, and the judgment of each text segment does not interfere with each other. If any text segment is identified by the large language model as containing illegal content that matches the detection point, it is determined that the detection point is hit in the current bidding document.

[0017] Record the detected points and their corresponding text fragments to form a structured detection result output as a legality and compliance detection result.

[0018] In one specific implementation, establishing the mapping relationship between each detection point and each text segment includes:

[0019] Let the set of detection points be... The set of text fragments in the target tender document is The similarity between each detection point and the segment is expressed by the following formula:

[0020] ;

[0021] in, and These represent the detection points. and text fragments semantic vectors, This indicates a semantic concatenation operation. for The square of the norm, This is a moderating coefficient for semantic similarity. For constant terms, This is the adjustment coefficient for the structural similarity index term. For syntactic structural similarity, Represents a position-based variable power factor. The calculation method is as follows:

[0022] ;

[0023] in, and These are parameters used to control the weight magnitude and decay rate, respectively. For text fragments In the normalization location of the tender documents, Indicates the detection point The corresponding focus location center;

[0024] After calculating the similarity, for each detection point, it iterates through its similarity values ​​with all text segments. When a certain similarity value exceeds the set judgment threshold, it is considered that there is a mapping relationship between the detection point and the corresponding text segment. In turn, it constructs a set of text segment indexes for each detection point and establishes a many-to-many mapping relationship between each legitimate and compliant detection point and each text segment.

[0025] In a specific feasible implementation, the step of calling the large language model to extract the clauses with quantitative limiting features from the parsed tender document and constructing a target enterprise knowledge graph includes:

[0026] The language big data model is used to perform semantic understanding and content segmentation on the structured data obtained from parsing the target tender document, and the clauses with quantitative limiting features in the tender document are automatically identified.

[0027] The language big model combines contextual semantics to cluster the extracted terms, and performs semantic classification and intent annotation to form semantic clusters.

[0028] Each semantic dimension in the semantic cluster is mapped to a first-level node in the target enterprise knowledge graph. The corresponding quantitative terms under each cluster are embedded in the graph structure as attribute entries of that node, thus constructing the target enterprise knowledge graph of the tender document.

[0029] In one specific implementation scheme, the step of calling the large language model to identify whether the parsed tender document contains biased content includes:

[0030] Establish a blacklist semantic set and a whitelist semantic set. The blacklist semantic set contains sensitive statement feature templates known to have a guiding or exclusive tendency, while the whitelist semantic set includes reasonable expressions that have a specific orientation but are acceptable due to legal authorization, policy exemption, etc.

[0031] The language big model is invoked to perform semantic modeling and context parsing on the structured data obtained from parsing the target tender document, and semantic vectorization is performed on each text segment to extract its semantic expression features;

[0032] The degree of matching between the text fragment and the sensitive template is determined by calculating the similarity between the semantic vector of the text fragment and the semantic vector of each sensitive sentence feature template;

[0033] If a text fragment is identified as having semantic expression that satisfies any sensitive semantic template in the blacklist and cannot be matched with a rationale in the whitelist set, then the text fragment is considered to have hit a bias detection point.

[0034] In a specific implementation scheme, determining the degree of matching between the text fragment and the sensitive template by calculating the similarity between the semantic vector of the text fragment and the semantic vector of each sensitive statement feature template includes:

[0035] Let the semantic vector of the text segment to be judged be... The semantic vector of the blacklist template set is The formula for calculating the matching degree is as follows:

[0036] ;

[0037] in, Indicates the text fragment and the first Cosine similarity of blacklist templates For the first The matching threshold for each blacklist template;

[0038] After the matching degree is calculated, a global threshold is set to determine whether the text fragment tends to hit the blacklist. If the matching degree is greater than or equal to the global threshold, the text fragment is considered to have a significant semantic match with the blacklist template and is judged to have sensitive semantic risks; if the matching degree is less than the global threshold, it is considered not to have hit the blacklist.

[0039] Secondly, this application provides a tender document detection system based on a large language model, which adopts the following technical solution:

[0040] A tender document detection system based on a large language model includes:

[0041] The large language model training module is used to train a pre-prepared model based on pre-prepared training data to generate a large language model suitable for the bidding and tendering field.

[0042] The tender document parsing module is used to obtain the target tender document to be tested and parse the target tender document into structured data;

[0043] The legality and compliance detection module is used to sort out legality and compliance detection points, call the large language model, divide the structured data obtained from parsing the target bidding document into several text fragments, establish a mapping relationship between each detection point and each text fragment, traverse all detection points and call the corresponding text fragments as input, use the large language model to determine whether there are detection points in the text fragments, and output the legality and compliance detection results.

[0044] The fair competition detection module is used to call the large language model to extract the clauses with quantitative limiting features in the parsed bidding documents and construct a target enterprise knowledge graph. The target enterprise knowledge graph is then compared with the enterprise knowledge graph library. At the same time, the large language model is called to identify whether the parsed bidding documents have biased expressions. The comparison results of the knowledge graph and the identification results of the biased expressions are output as the fair competition detection results.

[0045] Thirdly, this application provides an electronic device, the device including a processor and a memory; the memory stores a program, the program being loaded and executed by the processor to implement a tender document detection method based on a large language model as described in the first aspect.

[0046] Fourthly, this application provides a computer-readable storage medium storing a program that, when executed by a processor, is used to implement a tender document detection method based on a large language model as described in the first aspect.

[0047] In summary, the beneficial effects of this application include at least the following:

[0048] (1) By constructing a large language model specifically for the bidding and tendering field, and combining the structured parsing and semantic modeling capabilities of bidding texts, it is possible to accurately identify risk content related to laws, regulations, policy requirements, and negative lists from massive amounts of text without the need for preset templates and rules. Unlike traditional detection methods based on keyword matching or rule engines, the large language model can understand long-distance dependencies, implicit references in context, and complex logical nesting. Even when faced with bidding content that is diverse in expression, flexible in word choice, or even evasive in description, it can still effectively parse its actual semantics. By establishing a many-to-many semantic mapping relationship between detection points and text fragments, and combining it with the large language model to judge each fragment one by one, it avoids misjudgments and omissions caused by the lack of semantic adaptability of static rules, greatly improving the coverage and accuracy of legality and compliance risk identification, and providing regulatory authorities and industry practitioners with an intelligent and efficient compliance review tool.

[0049] (2) In fair competition detection, two collaborative paths are introduced: semantic clustering + knowledge graph comparison and sensitive expression identification + blacklist / whitelist correction. This enables the system to not only identify biased expressions at the linguistic level, but also to uncover hidden targeted behavior at the structural and data logic levels. The large language model first extracts all restrictive clauses with quantitative characteristics in the bidding documents and forms semantic clusters through contextual semantic clustering, further constructing a target enterprise knowledge graph. Based on this, by comparing with the enterprise sample knowledge graph library, it assesses whether the number of enterprises meeting the graph requirements is lower than a set threshold to determine whether there is a tailored problem. At the same time, the large language model can also perform vectorized analysis on sensitive semantics in the bidding text, perform deep matching with the blacklist template, and combine with the exempted expressions in the whitelist for correction, effectively avoiding misjudgment due to legitimate and reasonable expressions. This multi-dimensional collaborative identification mechanism can significantly improve the sensitivity and accuracy of fair competition risks, strengthen the intelligent judgment ability of improper restrictive conditions in bidding documents, and thus effectively ensure the openness and fairness of the bidding market.

[0050] To address the shortcomings of existing detection methods that rely on human experience or static rule configuration, this paper introduces a technical approach combining large language models and structured semantic analysis to achieve intelligent compliance and fair competition detection of bidding documents. Specifically, based on systematic training of a pre-set model on relevant training corpora in the bidding field, the large language model acquires the ability to identify policy and regulatory directions and differences in expression, thereby supporting the semantic parsing and judgment of bidding documents. In implementation, the bidding documents are first parsed into structured data. Then, the large language model participates in dividing the text into segments, establishing semantic mapping relationships between detection points and segments, and relying on deep semantic understanding to determine whether each segment contains illegal expressions. Simultaneously, in fair competition detection, the large language model can automatically extract clauses with quantitatively restrictive features from the bidding documents. By constructing a target enterprise knowledge graph and comparing it with an existing enterprise sample graph library, it identifies whether there are obvious targeted settings. Furthermore, the large model can also identify potential biased or suggestive expressions in the text and, combined with a blacklist / whitelist mechanism, avoid misjudgments, thus accurately outputting detection conclusions. This technical solution not only improves detection coverage but also breaks through the reliance of traditional rule-based methods on keywords and format templates. It has stronger semantic adaptability and policy evolution response capabilities, which can greatly reduce the cost and time consumption of manual review, improve the efficiency and accuracy of tender document detection, and effectively reduce interference from human subjective judgment, thus realizing intelligent review of tender documents.

[0051] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the tender document detection method based on a large language model in this application embodiment.

[0053] Figure 2 This is a flowchart illustrating the process of verifying the legality and compliance of bidding documents in this application embodiment.

[0054] Figure 3 This is a schematic diagram of the process for testing the fair competition of tender documents in this application embodiment.

[0055] Figure 4 This is a block diagram of an electronic device for detecting tender documents based on a large language model, as described in this application embodiment. Detailed Implementation

[0056] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0057] Optionally, this application uses the tender document detection method based on a large language model provided in various embodiments as an example for application in an electronic device. The electronic device is a terminal or a server. The terminal can be a computer, tablet computer, etc. This embodiment does not limit the type of electronic device.

[0058] Reference Figure 1 This is a flowchart illustrating a tender document detection method based on a large language model provided in one embodiment of this application. The method includes at least the following steps:

[0059] Step S101: Train the preset model based on the pre-prepared training data to generate a large language model suitable for the bidding field.

[0060] In step S101, to achieve a deep understanding of the structure, linguistic features, and expression style of bidding documents, a large language model suitable for bidding scenarios needs to be trained and constructed first. This large language model is based on existing general-purpose large language models based on the seqGPT architecture, and is trained in a targeted manner using high-quality corpora with industry characteristics, thereby improving its adaptability to the unique semantics, logical structure, and usage style of bidding documents.

[0061] Optionally, Deepseek is selected as an existing general-purpose large language model in this application. Other existing general-purpose large language models, such as chatgpt, can also be selected. This application does not impose any restrictions on the specific selection of existing general-purpose large language models.

[0062] The pre-prepared training data mainly includes tender document templates, examples, or historical cases covering multiple common projects such as engineering construction. The sources of the training data include, but are not limited to, publicly available projects or accessible websites related to those projects. It should be noted that the training data must be timely, typically selecting data published recently. After the training data is collected, it undergoes cleaning and deduplication. Cleaning includes removing invalid characters, standardizing encoding formats, and regularizing paragraph structures. Deduplication involves introducing text similarity calculations to remove semantically repetitive or highly consistent document content, ensuring the training data has sufficient breadth and information diversity. After data cleaning, structured labeling is performed on the training data. Specifically, key bidding elements involved in the training data are labeled and annotated, such as project name, bidder qualifications, performance terms, and scoring distribution, establishing a structured dataset suitable for supervised language modeling.

[0063] During the model training phase, a language generation architecture with contextual modeling capabilities is selected and fine-tuned using domain data. Optionally, the seqGPT architecture is used as the training platform. This architecture has the ability to model language in complex documents. Through task-driven training on a large-scale domain corpus, the model can generate sentences that conform to the language habits of bidding documents and identify and parse expression structures with specific review meanings. During training, a labeled structured dataset is input through supervised learning to build the model's ability to recognize lexical combinations, semantic cues, and inter-sentence logic.

[0064] After training, a model capability evaluation is performed, with evaluation metrics including field extraction accuracy, structural integrity, and semantic consistency score. Evaluation methods include automated test set scoring and manual verification of a subset of samples to ensure the model has stable output and generalization ability. Once the evaluation is passed, the model is deployed to a usable runtime environment to support the actual operation of language recognition and processing functions.

[0065] Step S102: Obtain the target tender document to be tested and parse the target tender document into structured data.

[0066] In step S102, after obtaining the target tender document to be detected, a preset parsing tool is used to extract text and convert the format of the target tender document, identify the chapter divisions, clause order, and paragraph boundaries, segment the original text into semantic units, and unify the encoding format. Through structural reconstruction and field classification operations, the target tender document is parsed into structured data to adapt to the input format of the large language model, thereby achieving standardized modeling preparation for the content.

[0067] Step S103: Identify the legality and compliance checkpoints, call the large language model, divide the structured data obtained from parsing the target tender document into several text fragments, establish a mapping relationship between each checkpoint and each text fragment, traverse all checkpoints and call the corresponding text fragments as input, use the large language model to determine whether there are checkpoints in the text fragments, and output the legality and compliance check results.

[0068] Reference Figure 2In step S103, firstly, laws, regulations, policy documents, negative lists of tenderers, and negative lists of government procurement are reviewed, and multiple legality and compliance checkpoints are manually extracted. Each checkpoint reflects expressions or structural settings that should be avoided in the tender documents, with clear judgment criteria and corresponding descriptive text. Simultaneously, a large language model is invoked to slice the structured data parsed from the target tender documents, dividing it into several content segments according to paragraphs, clauses, or semantic units. Then, for each checkpoint, a semantic vector is generated based on its descriptive text. For each slice of text in the target tender document, its semantic vector is generated. The semantic representation of each segment includes not only its main text content but may also selectively include its chapter title or clause number to enhance structural context understanding. Then, by calculating the semantic similarity between the checkpoint and each text segment, it is determined whether the segment semantically belongs to the expression category of interest to a certain checkpoint. Once the similarity exceeds a set threshold, a mapping relationship is established. Furthermore, a confidence threshold can be set, retaining only the top N similarity-ranked mappings. Many-to-many mappings are also supported, allowing one detection point to correspond to multiple text segments, and one text segment to correspond to multiple detection points, thus establishing a many-to-many mapping between each legal compliance detection point and each text segment. Finally, all legal compliance detection points are iterated through one by one. For each detection point, one or more text segments associated with it in the mapping relationship are called as the content to be judged. For each called text segment, input content containing detection point hints is constructed and input into a large language model for semantic analysis. The large language model determines whether the text segment contains relevant illegal expressions or inappropriate restrictions by understanding the correspondence between the policy requirements represented by the detection point and the actual meaning of the text in the tender document. During the detection process, an independent judgment strategy is adopted: the judgment of each segment does not interfere with each other. If any segment is identified by the model as containing illegal content consistent with the detection point, the detection point is determined to be hit in the current tender document. The hit detection points and their corresponding text segments are recorded, forming a structured detection result output, i.e., the legal compliance detection result.

[0069] Furthermore, as a preferred embodiment, in another feasible embodiment, the mapping relationship between each legality and compliance detection point and each text segment can also be established as follows:

[0070] First, let the set of detection points be... The set of text fragments in the target tender document is The similarity between each detection point and the segment is expressed by the following formula:

[0071] ;

[0072] in, and These represent the detection points. and text fragments semantic vectors, This indicates a semantic concatenation operation. for The square of the norm, This is a moderating coefficient for semantic similarity. For constant terms, This is the adjustment coefficient for the structural similarity index term. Syntactic structural similarity is calculated in this embodiment by constructing syntactic dependency trees for each detection point and segment, and then using the approximate similarity calculation method based on Tree Edit Distance. A higher similarity indicates greater structural similarity. This represents a position-based variable power factor used to dynamically adjust the weights between specific detection points and segments.

[0073] The calculation method is as follows:

[0074] ;

[0075] in, and These are parameters used to control the weight magnitude and decay rate, respectively. For text fragments In the normalization location of the tender documents, Indicates the detection point The corresponding center of attention is determined in advance through historical sample statistics.

[0076] The above formula integrates three levels of features: semantic concatenation, structural similarity weighting, and position power adjustment. This aims to achieve a more refined and comprehensive characterization of the correlation between tender document fragments and legality / compliance detection points. After concatenating the detection point vector and the fragment vector, its... The square of the norm introduces asymmetry and the ability to reconstruct the overall semantic space, helping to capture more complex semantic combinations. Many non-compliant expressions in tender documents are not semantically identical but are hidden through structural deformation and clause rearrangement. Therefore, introducing syntactic structural similarity obtained from TreeEdit Distance during the matching process allows the system to maintain high recognition ability even when faced with diverse expressions, avoiding misjudgments and omissions. Tender documents typically have strong format norms; for example, qualification requirements are often found in the "Instructions to Bidders" or "Qualification Review" sections. Therefore, introducing a regulation term based on document normalization position gives higher weight to segments closer to the center of the focus position during matching, improving the model's context adaptability and directional sensitivity. The combination of exponential decay and power regulation terms results in an exponential amplification of structurally similar and reasonably positioned matching pairs in the final score, while redundant matches without structural similarity are quickly suppressed, greatly improving the model's selectivity for high-quality mapping relationships.

[0077] After calculating the similarity, each detection point is iterated through with all text segments for its similarity value. When a certain similarity value exceeds a set judgment threshold, the detection point is considered to have a high semantic and structural correlation with the corresponding text segment, thus establishing a mapping relationship between the two. In this way, a corresponding text segment index set is constructed for each detection point, thereby establishing a many-to-many mapping relationship between each legitimate and compliant detection point and each text segment.

[0078] In another feasible embodiment, the mapping relationship between each legality and compliance detection point and each text segment can also be established as follows:

[0079] Let the set of detection points be... The set of text fragments in the target tender document is In this embodiment, each legality and compliance detection point is first checked. Extract its semantic embedding vector Then, a lightweight feedforward network is used to generate the gating vector for the detection point. For each text fragment Generate its semantic vector as The mapping score is defined as follows:

[0080] ;

[0081] in, It is a non-linear activation function that ensures the output is between 0 and 1, reflecting the non-linear attention intensity of the detection point to the text segment. It is an element-wise product. This indicates the transpose operation. The weight matrix is ​​a trainable asymmetric matrix, which makes semantic matching no longer symmetric and allows for directional bias between detection points and text fragments. It is a structured positional encoding vector of a text segment. Indicates legality and compliance testing points The corresponding semantic focus locations are all in the positional encoding space. The attenuation coefficient is greater than 0. The distance is Euclidean.

[0082] In the design of the above formula, the mapping scoring function integrates two core dimensions: semantic relevance and structural position matching. The first term introduces a detection point weight vector to weight the semantic representation of the text fragment, enabling the model to maintain sensitivity to key feature words at the semantic level and improve the discriminative ability of semantic matching. The second term introduces a soft constraint mechanism for structural position, scoring the distance between the fragment position and the position of interest of the detection point through a Gaussian kernel function, strengthening the model's ability to understand the structural context. This design is particularly suitable for document structures with regular formats and clear chapters, such as tender documents. Traditional matching methods often rely on pure semantic similarity or static keyword matching, ignoring the high correspondence between structural positions and detection points in tender documents, making it difficult to accurately locate the corresponding clauses, especially when there is semantic overlap or nested clauses, which can easily lead to mismatches. This formula nonlinearly couples and fuses structural information and semantic representation, guiding the semantic comparison process through position priors, improving the matching accuracy between detection points and target content, and exhibiting better robustness and generalization ability when dealing with long documents and multi-level structures.

[0083] After calculating the mapping score between each detection point and each text segment, the mapping score is compared with a preset threshold. If the score is greater than or equal to the preset threshold, it is considered that there is a mapping relationship between the detection point and the text segment.

[0084] Step S104: Use a large language model to extract the clauses with quantitative limiting features from the parsed tender documents and construct a target enterprise knowledge graph. Compare the target enterprise knowledge graph with the enterprise knowledge graph database. At the same time, use a large language model to identify whether the parsed tender documents have biased expressions. Output the comparison results of the knowledge graph and the identification results of the biased expressions as the fair competition detection results.

[0085] Reference Figure 3In the fair competition detection step S104, the first step is to identify qualification thresholds or exclusive requirements in the bidding documents that target specific companies. This requires combining knowledge graph construction and comparison methods to automatically extract quantitative clauses, standardize their expression, and determine the matching status of companies. Subsequently, for expressions in the bidding documents that may hinder fair competition, the focus is on identifying biased, suggestive, or targeted statements, such as clauses with exclusive characteristics like "preferential consideration for XX company," "must have specific engineering experience," or "restricted use of a certain brand of equipment."

[0086] Specifically, the first step involves using a large language model to perform semantic understanding and content segmentation on the structured data obtained from parsing the target tender document. This automatically identifies clauses in the tender document with quantitative limiting features, such as "the company's annual output value must not be less than XX," "the company must possess XX awards," and "the legal representative must have more than ten years of industry experience." After extracting clauses with quantitative limiting features, the large language model clusters these clauses based on contextual semantics, performing semantic classification and intent annotation. This results in semantic clusters with core semantic dimensions such as company size, qualification level, performance honors, and team experience. The purpose of introducing these semantic clusters after extracting the quantitative feature clauses is to perform semantic normalization and intent aggregation on these clauses, forming a clearly structured and dimensionally defined intermediate representation layer. Each semantic dimension in the semantic cluster (such as company size, project experience, industry honors, etc.) is mapped to a first-level node in the target company's knowledge graph. The corresponding quantitative clauses under each cluster (such as registered capital not less than 300 million yuan) are embedded in the graph structure as attribute entries for that node, ensuring that the graph nodes not only have semantic consistency but also retain quantitative information closely corresponding to the original text. Based on semantic clusters, a target enterprise knowledge graph is further constructed for the tender documents. After constructing the target enterprise knowledge graph, each semantic dimension node in the graph is mapped to one or more quantitative requirements, maintaining consistency with the actual enterprise data structure. Subsequently, the target graph is compared item by item with a pre-constructed enterprise knowledge graph library. The knowledge graph in the enterprise knowledge graph library comes from publicly available data resources and covers a large amount of structured information such as registered capital, years of operation, qualification categories, and award records of various types of enterprises. During the comparison process, each enterprise instance in the enterprise knowledge graph library is verified one by one according to the requirements clauses in the semantic clusters. It is determined whether the number of enterprises that meet all quantitative requirements is lower than a preset adaptation threshold, which is usually set to 2 to 3. If the number of enterprises that meet all quantitative requirements is lower than the preset adaptation threshold, it proves that the number of enterprises that meet the requirements is too small, and it is considered that the quantitative clause has obvious suspicion of targeted setting. The comparison result of the target enterprise knowledge graph, that is, the comparison result of the number of enterprises that meet all quantitative requirements and the preset threshold, is used as part of the subsequent output fair competition detection result. This type of comparison does not rely on manually specifying a particular list of companies, but rather judges the universal applicability of setting thresholds based on objective data, thus avoiding misjudgments caused by subjective judgment in traditional methods.

[0087] It's important to note that the core purpose of introducing semantic clustering is to improve the construction quality and semantic consistency of the target enterprise's knowledge graph, thereby enhancing the accuracy and generalization ability of subsequent graph comparison stages. Conventional knowledge graph construction methods typically treat each quantitative requirement in the tender document as an independent node and directly incorporate it into the graph structure. This approach easily leads to a chaotic graph dimension and blurred semantic boundaries, especially in texts with numerous variations in expression or nested logic, where unified categorization is often difficult. For example, while "the company's registered capital is over 300 million yuan" and "its annual turnover reaches 500 million yuan" have different wording, they both fall under the semantic dimension of enterprise size from the perspective of the tender's intent. Building separate graphs without differentiation results in dimension fragmentation, reducing matching efficiency and accuracy in subsequent comparison stages. In contrast, the introduction of semantic clustering utilizes a large language model to perform contextual understanding and intent recognition of the tender clauses, aggregating and categorizing multiple clauses with superficial differences but pointing to the same review purpose. This achieves semantic compression and intent alignment before graph construction, ensuring that graph nodes possess unified dimensional meaning and logical independence. This processing not only improves the clarity of the graph structure but also significantly reduces dimensional redundancy and semantic ambiguity in subsequent comparisons, making the quantification requirements under each dimension more representative. Simultaneously, semantic clustering can automatically avoid the problem of "inconsistent expression but identical substance" caused by different language expressions, improving the large model's tolerance for diverse clauses. Therefore, introducing a semantic clustering mechanism is not redundant but rather a key technical step in achieving structural uniformity, improving comparison efficiency, and enhancing semantic accessibility.

[0088] In implementation, while comparing the aforementioned enterprise knowledge graph, the fair competition detection also needs to identify whether there are biased or suggestive language expressions in the target bidding documents. Specifically, a blacklist semantic set and a whitelist semantic set are first established. The blacklist semantic set contains sensitive statement feature templates known to have a guiding or exclusive tendency, including direct wording (such as "must be from a certain organization's background") and implicit descriptions (such as "preferential selection of companies with local offices"). The whitelist semantic set includes reasonable expressions that, while having a specific target, are acceptable due to legal authorization, policy exemptions, etc. Subsequently, a large language model is used to perform semantic modeling and contextual analysis on the structured data obtained from parsing the target bidding documents. Semantic vectorization is performed on each text fragment, and its semantic expression features are extracted. Based on the degree of matching between the semantics of each text fragment and the blacklist set, the sensitivity risk of the fragment is initially determined. At the same time, a reverse matching mechanism with the whitelist set is introduced to avoid false judgments. During the judgment process, each segment undergoes independent risk analysis. If the large language model identifies that the semantic expression of the segment satisfies any sensitive semantic template in the blacklist and cannot be matched with a rationale in the whitelist set, the segment is considered to have hit a bias detection point and is judged to have a potential risk of hindering fair competition. All hit segments will be recorded and marked as risk output items, and will serve as another component of the fair competition detection result.

[0089] The reverse matching mechanism of the whitelist further judges text fragments identified by the large language model as hitting sensitive templates in the blacklist by combining contextual semantics. If the fragment contains a reasonable business explanation or legal basis (such as the requirement for local service capabilities due to special project needs), it matches the corresponding exemption template in the whitelist, thereby eliminating the risk mark of the fragment and avoiding misjudging reasonable and compliant expressions as biased content. This mechanism effectively improves the accuracy of detection, reduces the false judgment rate, and ensures that fair competition detection is more refined and meets actual business needs.

[0090] Optionally, the large language model determines the degree of matching between the text fragment semantic vector and the semantic vector of each sensitive sentence feature template by calculating the cosine similarity.

[0091] Furthermore, as a preferred method, the large language model can also determine the degree of matching between the text segment and the sensitive template by calculating the similarity between the semantic vector of the text segment and the semantic vector of each sensitive sentence feature template, where the semantic vector of the text segment to be judged is... The semantic vector of the blacklist template set is The formula for calculating the matching degree is as follows:

[0092] ;

[0093] in, Indicates the text fragment and the first Cosine similarity of blacklist templates For the first A matching threshold is set for each blacklist template. Similarity scores below this threshold are not included in the matching score to avoid interference from irrelevant templates. The formula first subtracts the threshold and uses a maximum function to filter similarity scores below the threshold to zero, thus achieving the purpose of screening and reducing the influence of irrelevant templates. The numerator accumulates the excess similarity scores of all templates exceeding the threshold, highlighting effective matches. The denominator uses a square root structure to summarize and penalize the mismatch scores of all templates, preventing an excessively large numerator due to too many templates and maintaining a reasonable normalization of the matching score. This design improves computational efficiency while enhancing sensitivity to key matching templates, avoiding the drawback of traditional average cosine similarity being easily diluted by a large number of low-relevance templates.

[0094] After calculating the matching degree, a global threshold is set to determine whether the text fragment tends to hit the blacklist. If the matching degree is greater than or equal to the global threshold, the text fragment is considered to have a significant semantic match with the blacklist template and is judged to have sensitive semantic risks. If the matching degree is less than the global threshold, it is considered not to have hit the blacklist.

[0095] In summary, addressing the shortcomings of existing detection methods that rely on manual experience or static rule configuration, this paper proposes a technical approach combining large language models and structured semantic analysis to achieve intelligent compliance and fair competition detection of bidding documents. Specifically, based on systematic training of a pre-defined model on relevant training corpora in the bidding field, the large language model acquires the ability to identify policy and regulatory implications and differences in expression, thereby supporting the semantic parsing and judgment of bidding documents. In implementation, the bidding documents are first parsed into structured data. The large language model then participates in dividing the text into segments, establishing semantic mapping relationships between detection points and segments, and relying on deep semantic understanding to determine whether each segment contains any illegal expressions. Simultaneously, in fair competition detection, the large language model can automatically extract clauses with quantitatively defined characteristics from the bidding documents. By constructing a target enterprise knowledge graph and comparing it with an existing enterprise sample graph library, it identifies whether there are obvious targeted settings. Furthermore, the large model can also identify potential biased or suggestive expressions in the text and, combined with a blacklist / whitelist mechanism, avoid misjudgments, thus accurately outputting detection conclusions. This technical solution not only improves detection coverage but also breaks through the reliance of traditional rule-based methods on keywords and format templates. It has stronger semantic adaptability and policy evolution response capabilities, which can greatly reduce the cost and time consumption of manual review, improve the efficiency and accuracy of tender document detection, and effectively reduce interference from human subjective judgment, thus realizing intelligent review of tender documents.

[0096] This application also provides a tender document detection system based on a large language model, which includes at least the following modules:

[0097] The large language model training module is used to train a pre-prepared model based on pre-prepared training data to generate a large language model suitable for the bidding and tendering field.

[0098] The tender document parsing module is used to obtain the target tender document to be tested and parse the target tender document into structured data;

[0099] The legality and compliance detection module is used to identify legality and compliance detection points. It calls a large language model to divide the structured data obtained from parsing the target tender document into several text fragments, establishes a mapping relationship between each detection point and each text fragment, traverses all detection points and calls the corresponding text fragments as input, uses the large language model to determine whether there are detection points in the text fragments, and outputs the legality and compliance detection results.

[0100] The fair competition detection module is used to extract clauses with quantitative limiting features from the parsed tender documents using a large language model and construct a target company knowledge graph. The target company knowledge graph is then compared with the company knowledge graph database. Simultaneously, the large language model is used to identify whether the parsed tender documents contain biased expressions. The comparison results of the knowledge graph and the identification results of the biased expressions are output as the fair competition detection results.

[0101] For relevant details, please refer to the above method implementation examples.

[0102] Figure 4 This is a block diagram of an electronic device provided in one embodiment of this application. The device includes at least a processor 401 and a memory 402.

[0103] Processor 401 may include one or more processing cores, such as a quad-core processor or an octa-core processor. Processor 401 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 401 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 401 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 401 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0104] The memory 402 may include one or more computer-readable storage media, which may be non-transitory. The memory 402 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 402 are used to store at least one instruction, which is executed by the processor 401 to implement the tender document detection method based on a large language model provided in the method embodiments of this application.

[0105] In some embodiments, the electronic device may also optionally include: a peripheral device interface and at least one peripheral device. The processor 401, memory 402, and peripheral device interface can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface via a bus, signal line, or circuit board. Indicatively, peripheral devices include, but are not limited to: radio frequency circuits, touch displays, audio circuits, and power supplies.

[0106] Of course, electronic devices may also include fewer or more components, and this embodiment does not limit this.

[0107] Optionally, this application also provides a computer-readable storage medium storing a program that is loaded and executed by a processor to implement the tender document detection method based on a large language model described in the above method embodiments.

[0108] Optionally, this application also provides a computer product including a computer-readable storage medium storing a program, which is loaded and executed by a processor to implement the tender document detection method based on a large language model described in the above method embodiments.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting a tender document based on a large language model, characterized in that, The method comprises: training a preset model based on pre-prepared training data to generate a large language model applicable to the bidding field; obtaining a target bidding document to be detected, and parsing the target bidding document into structured data; combing legal compliance detection points, calling the large language model, dividing the structured data obtained by parsing the target bidding document into a plurality of text segments, establishing a mapping relationship between each detection point and each text segment, traversing all detection points and calling the corresponding text segments as input, judging whether the text segment contains the detection point through the large language model, and outputting a legal compliance detection result; the establishment of the mapping relationship between each detection point and each text segment comprises: Let the set of detection points be , and the set of text segments of the target tender document be . The similarity between each detection point and segment is represented by the following formula: ； wherein, and denote semantic vectors of detection points and text segments , denotes a semantic concatenation operation, is the square of the norm, is a regulation coefficient of semantic similarity, is a constant term, is a regulation coefficient of the structural similarity index term, is the syntactic structure similarity, denotes a position-based variable power factor, is calculated as follows: ； wherein, and are parameters for controlling the weight magnitude and the decay rate, respectively, is a text segment a normalized position in the tender document, denotes a detection point a corresponding center of attention position; after calculating the similarity, traversing the similarity values of each detection point with all text segments, and when a similarity value exceeds a set determination threshold, considering that the detection point and the corresponding text segment have a mapping relationship; sequentially constructing the corresponding text segment index set for each detection point, and establishing a many-to-many mapping relationship between each legal compliance detection point and each text segment; Or, set the detection point set as , the text segment set of the target bidding document as , the first pair of each legal compliance detection point extract its semantic embedding vector , and then generate the gate vector of the detection point through a layer of lightweight feedforward network , and the semantic vector of each text segment is , and the mapping score is defined as follows: ； wherein, is a nonlinear activation function, is an element-wise product, denotes a transpose operation, is a trainable asymmetric weight matrix, is a structured positional encoding vector of the text segment, denotes a legal compliance detection point corresponding to a semantic focus center position, is a decay coefficient, is an Euclidean distance; after calculating the mapping score of each detection point and each text segment, the mapping score and a preset threshold are compared, and if it is greater than or equal to the preset threshold, it is considered that the detection point and the text segment have a mapping relationship. calling the large language model to extract the content of the clauses with quantitative limitation characteristics in the parsed bidding document and constructing a target enterprise knowledge graph, comparing the target enterprise knowledge graph with an enterprise knowledge graph library; and simultaneously calling the large language model to identify whether the parsed bidding document has a biased expression content, and outputting the comparison result of the knowledge graph and the identification result of the biased expression content as a fair competition detection result. 2.The tender document detection method based on a large language model according to claim 1, wherein, The training of the preset model based on the pre-prepared training data to generate a large language model applicable to the bidding field comprises: cleaning and deduplicating the training data, which includes bidding document templates, templates or historical cases of a plurality of common projects; labeling and annotating the key bidding elements involved in the training data, establishing a structured data set applicable to supervised language modeling; using a seqGPT architecture as a training platform, inputting the annotated structured data set into a preset general large language model through supervised learning, and training a large language model applicable to the bidding field. 3.The method of claim 1, wherein, The judgment of the text segment by the large language model whether the detection point exists, and the output of the legal compliance detection result comprise: in the detection process, an independent judgment strategy is adopted, and the judgment of each text segment does not interfere with each other, if any text segment is identified by the large language model as containing illegal content consistent with the detection point, it is determined that the detection point is hit in the current bidding document; recording the hit detection point and the corresponding text segment to form a structured detection result output as a legal compliance detection result. 4.The method of claim 1, wherein, The calling of the large language model to extract the content of the clauses with quantitative limitation characteristics in the parsed bidding document and the construction of the target enterprise knowledge graph comprise: calling the language large model to perform semantic understanding and content segmentation on the structured data obtained by parsing the target bidding document, and automatically identifying the content of the clauses with quantitative limitation characteristics in the bidding document; The language large model clusters the extracted clauses by combining context semantics, and performs semantic classification and intent labeling on the clauses, forming semantic clusters in the semantic dimension; Each semantic dimension in the semantic cluster set is mapped to a first-level node in the target enterprise knowledge graph, and the corresponding quantitative clauses under each cluster are embedded into the graph structure as attribute items of the node, thereby constructing the target enterprise knowledge graph of the tender document. 5.The method of claim 1, wherein, The calling of the large language model to identify and analyze the tender document includes: A blacklist semantic set and a whitelist semantic set are established, the blacklist semantic set contains sensitive sentence feature templates with known guiding or exclusive tendencies, and the whitelist semantic set includes acceptable reasonable expressions; The language large model is called to perform semantic modeling and context analysis on the structured data obtained by analyzing the target tender document, and each text segment is subjected to semantic vectorization processing and extraction of its semantic expression features; The similarity between the semantic vector of the text segment and the semantic vector of each sensitive sentence feature template is calculated to determine the matching degree of the text segment and the sensitive template; if the semantic expression of the text segment meets any sensitive semantic template in the blacklist and cannot match a reasonable explanation item in the whitelist set, it is considered that the text segment hits the tendency detection point. 6.The method of claim 5, wherein, The similarity between the semantic vector of the text segment and the semantic vector of each sensitive sentence feature template is calculated to determine the matching degree of the text segment and the sensitive template; if the semantic expression of the text segment meets any sensitive semantic template in the blacklist and cannot match a reasonable explanation item in the whitelist set, it is considered that the text segment hits the tendency detection point. Let the semantic vector of the text segment to be determined be , the semantic vector of the blacklist template set be , and the calculation formula of the matching degree be as follows: ; wherein, represents a cosine similarity of the text segment to the blacklist template, is a match threshold for the blacklist template; It includes: 7.A tender document detection system based on a large language model, characterized by A large language model training module is used to train a preset model based on pre-prepared training data to generate a large language model suitable for the bidding field; A tender document analysis module is used to obtain a target tender document to be detected and analyze the target tender document into structured data; A legal compliance detection module is used to sort legal compliance detection points, call the large language model, divide the structured data obtained by analyzing the target tender document into text segments, establish a mapping relationship between each detection point and each text segment, traverse all detection points and call the corresponding text segments as input, determine whether the text segment contains the detection point through the large language model, and output the legal compliance detection result; The mapping relationship between each detection point and each text segment includes: After calculating the similarity, the similarity values between each detection point and all text segments are traversed, and when a similarity value exceeds a set determination threshold, it is considered that the detection point and the corresponding text segment have a mapping relationship; the corresponding text segment index set is constructed for each detection point in turn, and a many-to-many mapping relationship between each legal compliance detection point and each text segment is established. Let the set of detection points be , and the set of text segments of the target bidding document be The similarity between each detection point and segment is represented by the following formula: ； wherein, and denote semantic vectors of detection points and text snippets respectively, denotes a semantic concatenation operation, is the square of the ℓ2 norm, is a tuning coefficient for semantic similarity, is a constant term, is a tuning coefficient for the structural similarity index term, is the syntactic structure similarity, denotes a position-based variable power factor, is calculated as follows: ； wherein, and are parameters for controlling the weight magnitude and the decay rate, respectively, is a text segment is a normalized position in the tender document, denotes a detection point corresponds to a center of attention position; ​ Or, set the detection point set as , the text segment set of the target bidding document as , the first pair of each legal compliance detection point extract its semantic embedding vector , and then generate the gate vector of the detection point through a layer of lightweight feedforward network , and the semantic vector of each text segment is , and the mapping score is defined as follows: ； wherein, is a nonlinear activation function, is an element-wise product, denotes a transpose operation, is a trainable asymmetric weight matrix, is a structured positional encoding vector of the text segment, denotes a legal compliance detection point corresponding to a semantic focus center position, is a decay coefficient, is an Euclidean distance; after calculating the mapping score of each detection point and each text segment, the mapping score is compared with a preset threshold value, if it is greater than or equal to the preset threshold value, it is considered that the detection point and the text segment have a mapping relationship; a fair competition detection module is used to call the large language model to extract the clause content with quantified limited characteristics in the parsed bidding document and construct a target enterprise knowledge graph, and compare the target enterprise knowledge graph with an enterprise knowledge graph library; at the same time, the large language model is called to identify whether the parsed bidding document has a biased expression content, and the comparison result of the knowledge graph and the identification result of the biased expression content are taken as a fair competition detection result output.

8. An electronic device, comprising: The device comprises a processor and a memory; the memory stores a program, which is loaded and executed by the processor to implement the method for detecting a tender document based on a large language model according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program, which is executed by the processor to implement the method for detecting a tender document based on a large language model according to any one of claims 1 to 6.

Images