A large model-based tender multi-dimensional intelligent review method

Abstract

This invention relates to the field of intelligent document review technology, and in particular to a multi-dimensional intelligent review method for tender documents based on a large model. The method receives tender documents and extracts the original paragraph sequence; uses a machine learning model to perform binary classification of each paragraph to determine the title, obtaining candidate titles, and performs matching and alignment based on a three-level title template library to generate three-level title format review data; based on a list of required clauses, it identifies missing clauses through keyword matching and TF-IDF similarity, generating completeness missing data; uses a large model to extract key information and identifies implicit risks based on this information; performs condition matching based on an industry risk rule library to identify explicit risks; calculates risk scores for each risk point, determines the level, and generates a risk list; and generates a multi-dimensional review report based on the format review data, completeness missing data, and risk list, achieving automated comprehensive review of tender document format, completeness, and risks.

Description

Technical Field

[0001] This invention relates to the field of intelligent document review technology, and in particular to a multi-dimensional intelligent review method for tender documents based on a large model. Background Technology

[0002] In the bidding process, the quality of the tender documents directly affects the success or failure of the bid. Traditional manual review methods are time-consuming and labor-intensive, and are prone to overlooking key clauses or risk points due to fatigue or lack of experience. With the development of artificial intelligence technology, automated tender review has gradually become a research hotspot.

[0003] In existing technologies, common tender document review methods mainly rely on rule engines or keyword matching, such as string matching of the list of required terms and threshold judgment of the price range. While these methods can quickly detect explicit compliance issues, they are difficult to identify implicit risks at the semantic level, such as contradictory promises, implicit brand-biased clauses, and ambiguous wording.

[0004] On the other hand, in recent years, large language models (LLM) have performed well in text understanding and can be directly used to determine whether there are risk points in the tender documents. However, there are two problems with relying entirely on LLM: First, large models may produce illusions and over-alarm for risks that do not exist; second, they lack a systematic understanding of document structure, especially their poor adaptability to format specifications such as three-level headings, which can easily lead to misjudging the body text as a heading or misidentifying the heading level.

[0005] In addition, existing methods typically handle format verification, integrity checks, and risk detection independently, failing to form an integrated, collaborative review process.

[0006] Therefore, there is an urgent need for a multi-dimensional intelligent review method for tender documents that integrates the advantages of rules and large models and can adaptively handle title formats, so as to improve the comprehensiveness, accuracy and automation of the review. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a multi-dimensional intelligent review method for tender documents based on a large model, which can automatically complete the verification of the title format specifications, the check of the completeness of the clauses, and the detection of explicit / implicit risks in the tender documents, and realize the efficient linkage between the rule engine and the large model, thereby improving the accuracy and interpretability of the review results.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-dimensional intelligent review method for tender documents based on a large model includes the following steps: S1. Receive the tender document, extract the content of the tender document, and obtain the original paragraph sequence; S2. Using a machine learning model, each original paragraph in the original paragraph sequence is classified sentence by sentence to determine whether it is a title, so as to obtain all candidate titles. Based on a preset three-level title template library, the candidate titles are matched and aligned to generate three-level title format review data. S3. Based on a pre-set list of essential terms, identify and monitor missing terms through keyword matching and TF-IDF vector similarity, and generate completeness missing data for missing terms. S4. Use a large model to extract key information from the original paragraph sequence, and use the large model based on the extracted key information to identify implicit risks such as inconsistencies, ambiguities, and biased clauses. S5. Based on a pre-defined industry risk rule base, explicit risks are identified through condition matching. S6. For each risk point identified from the explicit and implicit risks, calculate the risk score, determine the risk level, and generate a risk list. S7. Based on the three-level heading format review data, the data on missing integrity, and the risk list, generate a multi-dimensional review report.

[0009] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in the above-described method for multi-dimensional intelligent review of tender documents based on a large model.

[0010] The beneficial effects of this invention are as follows: This application provides a multi-dimensional intelligent review method for tender documents based on a large model, realizing an integrated review process from document parsing, adaptive verification of title format, and detection of clause completeness, to the identification of explicit / implicit risks through the linkage of rules and the large model; this invention uses a machine learning model instead of fixed regular expressions in title recognition, which can adapt to different numbering styles of tender documents; in risk identification, it combines a rule engine and a large model to make up for the defects of pure rule methods in omitting semantic risks and pure large model methods in producing illusions; the final generated structured review report can be used to guide the revision of tender documents, significantly improving the automation level and reliability of tender document review. Attached Figure Description

[0011] Figure 1 This is a flowchart of a multi-dimensional intelligent review method for tender documents based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0012] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0013] Before detailing the embodiments of this application, some related concepts will first be explained: Large Language Model (LLM): Refers to a deep learning model with a huge number of parameters and powerful natural language understanding and generation capabilities, such as the GPT series, LLaMA, and Wenxin Yiyan. In this invention, it is used for implicit risk identification and key information extraction.

[0014] Rule engine: A software module that performs logical judgments based on preset "condition-conclusion" pairs to quickly detect explicit compliance risks.

[0015] Dual-stream model: The neural network structure for title detection proposed in this invention processes text semantic features and layout style features separately, and finally fuses them for classification.

[0016] MechZeta activation function: This invention proposes a novel activation function that introduces damped sinusoidal oscillations in the negative region outside the positive linear region of ReLU, thereby enhancing the model's ability to discriminate boundary samples.

[0017] HSWCL loss function: Hierarchical semantic weighted contrastive loss, which combines cross-entropy, contrastive learning, and hierarchical constraints to train the title detection model.

[0018] In existing technologies, tender document review systems typically treat title detection as simple rule matching, such as detecting lines in the style of "Chapter X" or "XY". However, the formats of tender documents vary greatly among different bidders; some use "1. Project Overview" as a first-level title, while others use "I" or "Chapter 1". Existing regular expression methods struggle to cover all variations and cannot handle numbering errors caused by OCR or format conversion. Furthermore, risk detection modules often operate in isolation, and the rule engine relies on a manually maintained risk database, resulting in slow responses to new risks; large models, while flexible, are uncontrollable. This invention is proposed to solve the aforementioned technical problems.

[0019] To at least solve the above problems, please refer to Figure 1 Embodiment 1 of the present invention provides a multi-dimensional intelligent review method for tender documents based on a large model, including the following steps: S1. Receive the tender document, extract the content of the tender document, and obtain the original paragraph sequence.

[0020] In this embodiment, a tender document (such as a PDF or Word document) is received, and the text content is extracted using an existing document parsing library (such as Apache PDFBox or python-docx). The document is then segmented into paragraphs to obtain the original paragraph sequence. Simultaneously, the layout information of each paragraph (such as font size, bolding, indentation, and position within the page) is recorded for later use.

[0021] S2. Using a machine learning model, each original paragraph in the original paragraph sequence is classified sentence by sentence to determine whether it is a title, so as to obtain all candidate titles. Based on a preset three-level title template library, the candidate titles are matched and aligned to generate three-level title format review data. The machine learning model employs a two-stream model network structure, which includes: Input layer: Extract text semantic features, page layout features, and style format features from each of the original input paragraphs.

[0022] In this embodiment, three types of features are extracted for each paragraph: (1) Text semantic features: the paragraph text is segmented and BERT tokenized to obtain a token sequence; (2) Page layout features: paragraph line number, first line indentation, left indentation, paragraph front spacing, paragraph back spacing, etc., which are normalized and used as vectors; (3) Style format features: whether it is bold, whether it is underlined, whether it is centered, font name one-hot encoding, etc.

[0023] Dual-stream encoder: The semantic encoding stream feeds the text semantic features into the Transformer encoder to generate a semantic representation vector. Specifically, the semantic encoding stream uses a 6-layer Transformer encoder (pre-trained BERT-base-Chinese initialized) to output a semantic representation vector.

[0024] The layout features and style format features are concatenated by the layout style encoding stream and then fed into a two-layer fully connected network to output a structure representation vector. Feature fusion layer: The semantic representation vector and the structural representation vector are concatenated and fed into the attention fusion module (self-attention mechanism) to output a comprehensive feature vector.

[0025] A MechZeta activation function is introduced before the classification output layer for feature representation processing: ; Where max(0, x) represents taking the larger value between 0 and x. This represents the oscillation amplitude coefficient, which is a learnable parameter. This represents the oscillation frequency, which is a learnable parameter. Indicates the damping coefficient. This indicates that oscillations are introduced entirely within the negative value range.

[0026] This activation function produces oscillatory decay in the negative value region, which can preserve the differential information in the negative values, thereby improving the recall rate of boundary titles.

[0027] Classification Output Layer: A fully connected layer maps the combined features to binary classification probabilities, outputting the probability that the paragraph is a title. The features processed by MechZeta are then passed through another fully connected layer (input dimension d, output 2), and finally through Softmax to obtain the probability that the paragraph is a title.

[0028] The loss function used during the training of the machine learning model is expressed as follows: ; in, and To balance the hyperparameters, the initial value is... =0.3, =0.2. The cross-entropy loss is the standard binary cross-entropy loss, which guarantees the accuracy of the basic classification. The formula is as follows: ; in, N This represents the total number of paragraph samples in a batch. Indicates the true label, 1 indicates a title, 0 indicates no title. This represents the probability of the title output by the model; To contrast the learning loss, paragraph features of the same category (heading or non-heading) attract each other in the representation space, while features of different categories repel each other, effectively alleviating the problem of imbalanced positive and negative samples. The formula is expressed as: ; in, B This indicates the number of anchor samples used for comparison in the current batch. Indicates sample i The normalized eigenvectors, Indicates and Positive sample feature vectors of the same type This means that the denominator contains all the values ​​in the batch. One sample, including Positive samples and all negative samples This indicates the preset temperature over-parameter; To apply hierarchical constraint loss, this component utilizes the natural hierarchical structure of the tender document's three-level headings (e.g., Level 1 heading—Level 2 heading—Level 3 heading—body text) to impose document-level constraints on the model's output. The implementation is as follows: When labeling the data, additionally label the hierarchy for each title sample. ∈{1,2,3}; the hierarchical label for non-title samples is 0.

[0029] The model outputs the title probability. Simultaneously, an auxiliary classification head is added to output a 4-dimensional probability distribution. , respectively representing "non-heading" and "level 1 / 2 / 3 heading".

[0030] The hierarchical constraint loss constructs a 4-class auxiliary task using the natural hierarchy of headings (levels 1 / 2 / 3 and non-headings) in the tender document, and forces the hierarchical distribution of the model output to be consistent with the true hierarchy through KL divergence, reducing the error of misclassifying body text as headings. The formula is expressed as: ; in, This represents the true hierarchical distribution in one-hot form, and KL represents the KL divergence, which measures the difference between the predicted and true distributions. This indicates the preset hierarchical weights, which can be set. , , (Third-level headings provide more detailed information and carry greater weight.) (Non-title weight is slightly lower).

[0031] Matching and aligning the candidate titles based on a preset three-level heading template library includes the following steps: Calculate the semantic similarity between the candidate title and each title in the third-level title template library. The semantic similarity is determined based on a weighted sum of vector cosine similarity and edit distance. When the semantic similarity is greater than a preset threshold, the candidate title is aligned to the corresponding title in the template library and the mapping relationship is recorded; otherwise, it is marked as a format deviation.

[0032] The generation of the third-level heading format review data includes the following steps: Based on the mapping relationship and the format deviation, three levels of heading format review data are generated, including a list of missing headings, a list of redundant headings, and a list of incorrect headings.

[0033] In this embodiment, after training, the model is used to predict whether each paragraph is a title, resulting in a candidate title list. Then, matching and alignment are performed based on a pre-built three-level title template library (e.g., a specification for an industry tender document: Chapter 1 → 1.1 → 1.1.1). Specifically, the semantic similarity (cosine similarity + edit distance weighted sum) between the candidate titles and each title in the template is calculated. When the similarity is greater than a threshold (e.g., 0.85), a mapping relationship is established; those that cannot be matched are marked as format deviations. Finally, three-level title format review data is generated, including a list of missing titles, a list of redundant titles, and a list of misordered titles.

[0034] This step utilizes an adaptive dual-stream model and novel activation and loss functions to achieve highly accurate automatic detection and standardization suggestions for tender title formats, making it particularly suitable for tender scenarios with diverse formats.

[0035] S3. Based on a pre-set list of essential terms, identification and monitoring are performed through keyword matching and TF-IDF vector similarity to generate complete missing data for missing terms.

[0036] In this embodiment, based on a pre-set list of essential clauses (such as "letter of tender", "legal representative's identity certificate", "quotation details", "commitment letter against collusion in bidding", etc.), the existence of each clause is determined using keyword matching and TF-IDF vector similarity. For missing clauses, their names and locations are recorded to generate complete missing data.

[0037] S4. Use a large model to extract key information from the original paragraph sequence, and use the large model based on the extracted key information to identify implicit risks such as inconsistencies, ambiguities, and biased clauses. The key information includes the project name, bid price, construction period, warranty period, and qualification certificate number.

[0038] In this embodiment, firstly, key information is extracted using a large model (such as GPT-4) or a lightweight named entity recognition model: project name, bid price, construction period, warranty period, qualification certificate number, etc. Then, the entire tender document and the extracted key information are used to construct a prompt template, requiring the large model to detect three types of hidden risks: (1) inconsistencies (such as the construction period in the technical solution being inconsistent with the construction period in the commercial quotation); (2) ambiguous statements (such as "the equipment performance is not lower than the national standard, but the specific indicators are yet to be determined"); (3) biased clauses (such as exclusive statements such as "it is recommended to use a certain brand, or equivalent performance"). The large model outputs the description, location, and confidence level (0~1) of each risk point. At the same time, risk points not detected by the rule engine are recorded.

[0039] S5. Based on a pre-set industry risk rule base, explicit risks are identified through condition matching.

[0040] In this embodiment, the industry risk rule base is stored as condition-conclusion pairs, such as "quote exceeds budget limit by 20% → high risk", "missing 'No Collusion Bidding Commitment Letter' → high risk", "warranty period < 1 year → medium risk", etc. By matching the extracted key information with the rule conditions, an explicit risk list is generated, with each risk point containing a risk value and a rule number.

[0041] In this embodiment, a two-layer risk detection architecture is implemented based on steps S4 and S5: Underlying layer: Preset rule engine Built-in industry risk rule base, for example: Quote exceeding budget limit → High risk Warranty period < 1 year → Medium risk Lack of "No Collusion Bidding Commitment Letter" → High Risk The rules adopt a condition-conclusion format to quickly identify explicit risk points and output the risk level and trigger rule number.

[0042] Upper layer: Large model semantic risk analyzer The full text of the tender document and the key information extracted in step four are used to construct a prompt template, requiring the large model to detect three types of hidden risks: Inconsistent: For example, the construction period in the technical plan is inconsistent with the construction period in the quotation.

[0043] Ambiguous wording: such as "the equipment performance is no less than the national standard, but the specific indicators are yet to be determined".

[0044] Prejudicative clauses: such as "specifying certain brand parameters to exclude other competitors".

[0045] The large model outputs a description, location, and confidence level (0~1) for each risk point.

[0046] S6. For each risk point identified from the explicit and implicit risks, calculate the risk score, determine the risk level, and generate a risk list. The calculation of the risk score includes: For the same risk point obtained from the identification of both explicit and implicit risks, the risk score calculation formula is expressed as follows: Score=α×RuleScore+(1-α)×LLMConfidence; Wherein, RuleScore is the risk value corresponding to the explicit risk in the industry risk rule base, LLMConfidence is the confidence level corresponding to the implicit risk identified by the large model, and α is the weight coefficient. For the risk points obtained solely through the identification of the explicit risks, the risk score is the risk value corresponding to the explicit risk in the industry risk rule base; For the risk point obtained solely through the identification of the latent risk, determine whether the confidence level is greater than a preset filtering threshold. If not, filter the risk point; otherwise, the risk value calculation formula for the risk point is as follows: Score=(1-α)×LLMConfidence.

[0047] The weighting coefficient adopts a dynamic adjustment strategy, with an initial value of 0.7; For the same risk point obtained from the identification of the explicit risk and the implicit risk, when the confidence level is greater than 0.9, α is adjusted to 0.4; For the risk points obtained solely through the identification of the latent risks, if the confidence level is greater than the preset filtering threshold, then α is temporarily adjusted to 0.2.

[0048] In this embodiment, after calculating the scores of all risk points, the risk is divided into high / medium / low risk according to the threshold, a risk list is generated, and the source of the judgment is noted (rule / large model / combination of both).

[0049] Furthermore, in this embodiment, after reviewing every 10 tender documents, the hidden risk patterns frequently detected by the large model (such as "specifying more than three exclusive parameters") are summarized, manually reviewed, and transformed into new rule entries, which are then added to the preset rule base. This allows the rule engine to continuously learn the semantic patterns discovered by the large model.

[0050] This linkage mechanism can ensure the interpretability and determinism of the rule engine, discover semantic risks not covered by the rule base using the large model, and expand the feedback loop through rules. After processing a number of tender documents, the hidden risk patterns detected by the large model at high frequency are transformed into new rules and added to the rule base through manual review, so as to realize the continuous evolution of the system.

[0051] S7. Based on the three-level heading format review data, the data on missing integrity, and the risk list, generate a multi-dimensional review report.

[0052] In this embodiment, the comprehensive score of the tender document is calculated based on the number and severity of deviations in the format review data, the number and importance weight of missing clauses in the integrity data, and the level of each risk point in the risk list, using a preset scoring card (e.g., integrity weight 0.3, format specification weight 0.2, and risk deductions accumulated). A visual HTML / PDF report is ultimately output, containing the three-level heading format review data, the integrity data, and the risk list, and can further provide optimization and modification suggestions for the three-level heading format based on the large model for manual review.

[0053] Embodiment 2 of the present invention is as follows: A storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in the above-described method for multi-dimensional intelligent review of tender documents based on a large model.

[0054] This invention provides a multi-dimensional intelligent review method for tender documents based on a large model, which comprehensively realizes adaptive validation of title format, detection of clause integrity, linked identification of explicit and implicit risks, and continuous evolution of review rules, and has the following significant beneficial effects: First, it significantly improves the adaptability and accuracy of title format review. By fusing text semantics and layout style features through a dual-stream neural network model and introducing the MechZeta activation function and HSWCL loss function, it can adapt to the numbering styles and typesetting differences of different tender documents. Compared with traditional regular expression matching methods, it improves the F1 score of title detection and the accuracy of hierarchy recognition, effectively solving the problem of misjudgment caused by non-standard format.

[0055] Second, it comprehensively covers both explicit compliance risks and implicit semantic risks. The rule engine quickly hits hard red lines (such as exceeding price limits or missing clauses), and the large model deeply detects inconsistencies, ambiguities, and biased clauses. By dynamically weighting and fusing the data, it determines the risk score, which not only avoids false alarms caused by the illusion of the large model, but also makes up for the semantic blind spots that the rule base cannot cover.

[0056] Third, it achieves closed-loop self-evolution of review rules. New risk patterns detected with high confidence by the large model can be reverse-engineered into rule base entries, enabling the system to continuously improve during use and reducing manual maintenance costs. The entire review process, from document parsing to report generation, is completed in an integrated manner, significantly reducing the workload of manual review, shortening bid review time, and significantly improving bidding efficiency and bid quality.

[0057] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-dimensional intelligent review method for tender documents based on a large model, characterized in that, Including the following steps: S1. Receive the tender document, extract the content of the tender document, and obtain the original paragraph sequence; S2. Using a machine learning model, each original paragraph in the original paragraph sequence is classified sentence by sentence to determine whether it is a title, so as to obtain all candidate titles. Based on a preset three-level title template library, the candidate titles are matched and aligned to generate three-level title format review data. S3. Based on a pre-set list of essential terms, identify and monitor missing terms through keyword matching and TF-IDF vector similarity, and generate completeness missing data for missing terms. S4. Use a large model to extract key information from the original paragraph sequence, and use the large model based on the extracted key information to identify implicit risks such as inconsistencies, ambiguities, and biased clauses. S5. Based on a pre-defined industry risk rule base, explicit risks are identified through condition matching. S6. For each risk point identified from the explicit and implicit risks, calculate the risk score, determine the risk level, and generate a risk list. S7. Based on the three-level heading format review data, the data on missing integrity, and the risk list, generate a multi-dimensional review report.

2. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 1, characterized in that, Matching and aligning the candidate titles based on a preset three-level heading template library includes the following steps: Calculate the semantic similarity between the candidate title and each title in the third-level title template library. The semantic similarity is determined based on a weighted sum of vector cosine similarity and edit distance. When the semantic similarity is greater than a preset threshold, the candidate title is aligned to the corresponding title in the template library and the mapping relationship is recorded; otherwise, it is marked as a format deviation.

3. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 2, characterized in that, The generation of the third-level heading format review data includes the following steps: Based on the mapping relationship and the format deviation, three levels of heading format review data are generated, including a list of missing headings, a list of redundant headings, and a list of incorrect headings.

4. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 1, characterized in that, The key information includes the project name, bid price, construction period, warranty period, and qualification certificate number.

5. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 1, characterized in that, The calculation of the risk score includes: For the same risk point obtained from the identification of both explicit and implicit risks, the risk score calculation formula is expressed as follows: Score=α×RuleScore+(1-α)×LLMConfidence; Wherein, RuleScore is the risk value corresponding to the explicit risk in the industry risk rule base, LLMConfidence is the confidence level corresponding to the implicit risk identified by the large model, and α is the weight coefficient. For the risk points obtained solely through the identification of the explicit risks, the risk score is the risk value corresponding to the explicit risk in the industry risk rule base; For the risk point obtained solely through the identification of the latent risk, determine whether the confidence level is greater than a preset filtering threshold. If not, filter the risk point; otherwise, the risk value calculation formula for the risk point is as follows: Score=(1-α)×LLMConfidence.

6. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 5, characterized in that, The weighting coefficient adopts a dynamic adjustment strategy, with an initial value of 0.7; For the same risk point obtained from the identification of the explicit risk and the implicit risk, when the confidence level is greater than 0.9, α is adjusted to 0.4; For the risk points obtained solely through the identification of the latent risks, if the confidence level is greater than the preset filtering threshold, then α is temporarily adjusted to 0.

2.

7. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 1, characterized in that, The machine learning model employs a two-stream model network structure, which includes: Input layer: Extracts text semantic features, page layout features, and style format features from each of the original input paragraphs; Dual-stream encoder: The semantic encoding stream feeds the text semantic features into the Transformer encoder to generate a semantic representation vector; The layout features and style format features are concatenated by the layout style encoding stream and then fed into a two-layer fully connected network to output a structure representation vector. Feature fusion layer: concatenates the semantic representation vector with the structural representation vector, feeds it into the attention fusion module, and outputs a comprehensive feature vector; Classification output layer: The fully connected layer maps the comprehensive features into binary classification probabilities and outputs the probability value that the paragraph is the title.

8. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 7, characterized in that, A MechZeta activation function is introduced before the classification output layer for feature representation processing: ； Where max(0, x) represents taking the larger value between 0 and x. This represents the oscillation amplitude coefficient, which is a learnable parameter. This represents the oscillation frequency, which is a learnable parameter. Indicates the damping coefficient. This indicates that oscillations are introduced entirely within the negative value range.

9. The method for multi-dimensional intelligent review of tender documents based on a large model according to claim 7, characterized in that, The loss function used during the training of the machine learning model is expressed as follows: ； in, and To balance hyperparameters, The cross-entropy loss is expressed by the formula: ； in, N This represents the total number of paragraph samples in a batch. Indicates the true label, 1 indicates a title, 0 indicates no title. This represents the probability of the title output by the model; To compare the learning loss, the formula is expressed as: ； in, B This indicates the number of anchor samples used for comparison in the current batch. Indicates sample i The normalized eigenvectors, Indicates and Positive sample feature vectors of the same type This means that the denominator contains all values ​​in the batch. One sample, including Positive samples and all negative samples This indicates the preset temperature over-parameter; The hierarchical constraint loss is expressed by the formula: ； in, This represents the true hierarchical distribution in one-hot form, and KL represents the KL divergence, which measures the difference between the predicted and true distributions. This indicates the preset hierarchical weights.

10. A storage medium having a computer program stored thereon, characterized in that, When the processor executes the computer program, it implements the steps in the multi-dimensional intelligent review method for tender documents based on a large model as described in any one of claims 1-9.

Images