A large model-based bidding document labeling method, device and medium

Abstract

This application discloses a method, device, and medium for annotating bidding documents based on a large-scale model. The method includes: acquiring the bidding documents; preprocessing the bidding documents; deconstructing the preprocessed bidding documents to obtain multiple text data blocks; determining a pre-set annotation task; determining the natural language instructions corresponding to the text data blocks based on the annotation task; processing the text data blocks and corresponding natural language instructions through a large-scale language model to annotate the text data blocks and obtain structured annotation results; and associating the structured annotation results with the bidding documents to form bidding documents with semantic tags. This application significantly improves document processing efficiency, annotation accuracy, and intelligent retrieval capabilities by deconstructing bidding documents into structured text blocks and combining natural language instructions to drive large-scale model automatic annotation and generate semantic tags.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, device and medium for annotating bidding documents based on a large model. Background Technology

[0002] Tendering and bidding are crucial rules for allocating various public resources and fairly selecting partner companies. These transactions span the entire process, including bidding, tendering, opening, evaluation, awarding, on-site services, and supervision. With the continuous development of digital transaction platforms and the deepening of the "AI+" initiative, the tendering and bidding field is accelerating the application of artificial intelligence in key areas such as tender planning, document preparation and testing, bidding services, bid opening and evaluation assistance, awarding decisions, on-site management, and collaborative supervision. This aims to improve transaction efficiency, ensure transparent transactions, facilitate competitive selection, and enhance regulatory effectiveness. However, in practice, the reliance on paid general-purpose models often leads to high costs, significant data security risks, and unstable business adaptability. The industry tends to train smaller, customizable models that can be independently controlled and deployed locally. However, training these custom models heavily relies on high-quality labeled data. Tendering and bidding involve multiple business scenarios with significant differences in sample types, labeling systems, and judgment criteria, requiring a massive sample size. Furthermore, domain experts, such as bid evaluation experts and tendering managers, are unable to participate in large-scale manual labeling for extended periods due to their actual business responsibilities, resulting in a severe shortage of high-quality training data. This has become a core technological bottleneck restricting the development and large-scale application of custom models across multiple scenarios. Summary of the Invention

[0003] To address the aforementioned issues, this application proposes a method for annotating bidding documents based on a large-scale model, comprising: acquiring bidding documents; preprocessing the bidding documents; deconstructing the preprocessed bidding documents to obtain multiple text data blocks; determining a pre-set annotation task; determining the natural language instructions corresponding to the text data blocks based on the annotation task; processing the text data blocks and the corresponding natural language instructions using a large-scale language model to annotate the text data blocks and obtain structured annotation results; and associating the structured annotation results with the bidding documents to form bidding documents with semantic tags.

[0004] In one example, the preprocessed bidding document is deconstructed to obtain multiple text data blocks. Specifically, this includes: splitting the bidding document according to pre-set splitting criteria to divide the long document into multiple independently processable text segments, where the splitting criteria include chapters, clauses, scoring items, and business rules; and converting the text segments into structured text blocks, wherein the structured text blocks of the bidding document include block identifiers, chapters, page ranges, text content, and citation relationships, and the structured text blocks of the tender document include tender entity identifiers.

[0005] In one example, a pre-set annotation task is determined, and the natural language instructions corresponding to the text data block are determined based on the annotation task. Specifically, this includes: determining the application scenario corresponding to the bidding document; determining the corresponding scenario type and task type based on the application scenario, where the scenario type includes bidding planning scenario, bidding document detection scenario, intelligent assisted bid evaluation scenario, bid compliance self-inspection scenario, collusion identification scenario, and complaint handling scenario; determining the annotation task type corresponding to the text data block based on the task type, where the annotation task type includes question answering task, classification task, extraction task, and judgment task; and generating corresponding natural language instructions based on the scenario type, the annotation task type, and the semantic type of the text data block, where the natural language instructions include structured prompts, the text content to be analyzed, and domain-level rules.

[0006] In one example, the text data block is labeled based on the natural language instructions using a pre-defined large language model to obtain structured labeling results. Specifically, this includes: inputting the natural language instructions and the corresponding text data block into the large language model; performing semantic processing through the large language model to generate labeling results; the semantic processing steps include semantic understanding, feature extraction, rule matching, and sample generation; and performing format verification and structure conversion on the labeling results to obtain structured labeling results that conform to a preset data pattern. The structured labeling results include label tags, label content, confidence scores, original text citations used for labeling, and reasoning basis.

[0007] In one example, the method further includes: performing dataset clarification and governance on the annotation results in the semantically labeled bidding documents, wherein the dataset clarification and governance steps include deduplication, normalization, conflict resolution, scene attribution verification, label consistency verification, and low-confidence sample isolation; and classifying and storing the clarified annotation results according to scene type and task type to form a training dataset that can be called for training small models.

[0008] In one example, the method further includes: performing automated quality assessment on the classified and stored training dataset through a preset multi-dimensional quality assessment mechanism, wherein the multi-dimensional quality assessment mechanism includes any one or more of the following: rule verification, model consistency verification, sample difficulty stratification assessment, scene coverage analysis, and sampling index review; and filtering the samples in the training dataset according to the assessment results of the automated quality assessment to remove unqualified samples and retain high-value samples suitable for model training.

[0009] In one example, the method further includes: determining the risk level and scenario importance level of each sample based on the assessment results of the automated quality assessment; extracting samples from the training dataset based on the risk level and the scenario importance level to review the extracted samples; incorporating the approved extracted samples into the high-quality training dataset, and returning the unapproved extracted samples to the sample pool for subsequent re-labeling and retraining.

[0010] In one example, the method further includes: training a pre-defined small model based on the high-quality training dataset to enable the small model to possess vertical business capabilities, including question answering, extraction, classification, and judgment capabilities; transferring the domain reasoning and rule interpretation capabilities from the large language model to the small model through knowledge distillation; deploying the distilled small model to a target bidding business scenario and collecting model output results, user revision opinions, and business feedback to determine new sample return data; and periodically performing intelligent annotation, data updates, manual review, model retraining, and redistillation according to business changes, rule updates, and scenario expansion to achieve continuous evolution of the small model's capabilities.

[0011] On the other hand, this application also proposes a bidding document annotation device based on a large model, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the bidding document annotation device based on a large model to perform: the method described in any of the examples above.

[0012] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to be the method described in any of the examples above.

[0013] This application deconstructs bidding documents through structured deconstruction and generates natural language instructions adaptively based on the scenario. This enables large models to accurately perform various annotation tasks such as question answering, classification, extraction, and judgment, significantly improving the automation level of annotation and the accuracy of semantic understanding. It also significantly alleviates the pain points of traditional manual annotation, such as low efficiency, high cost, and inconsistent standards. The solution further introduces a dataset clarity governance and multi-dimensional automated quality assessment mechanism. Differentiated manual sampling is conducted based on risk level and scenario importance, ensuring the high quality and compliance of training data and effectively solving the problems of self-assessment distortion by single models and model contamination by low-quality samples. Building on this, knowledge distillation efficiently transfers the reasoning capabilities of large models to dedicated small models, reducing inference costs and deployment barriers while retaining the professional judgment capabilities of vertical domains. Finally, through post-deployment business feedback and periodic iterative training, continuous evolution and adaptive optimization of model capabilities are achieved. This provides a technical path for the intelligent application of the entire bidding process that is engineering-featured, long-term maintainable, cost-effective, and highly reliable. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a bidding document annotation method based on a large model, as described in an embodiment of this application. Figure 2 This is a schematic diagram of a bidding document annotation device based on a large model, as described in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0017] like Figure 1 As shown, in order to solve the above problems, this application provides a method for annotating bidding documents based on a large model, the method including: S101. Obtain the bidding documents, preprocess the bidding documents, and deconstruct the preprocessed bidding documents to obtain multiple text data blocks.

[0018] The acquired bidding documents include multi-source data such as tender documents, bid documents, evaluation records, regulations and policies, historical cases, complaint handling materials, and business Q&A records. The acquired multi-source data undergoes format conversion, text cleaning, anonymization, and segmentation. The preprocessed bidding documents are then deconstructed to obtain multiple text data blocks. Specifically, this includes: splitting the bidding documents according to pre-set splitting criteria to divide long documents into multiple independently processable text segments; the splitting criteria include chapters, clauses, scoring items, and business rules; and converting the text segments into structured text blocks. For tender documents, the structured text block includes at least: a block identifier, the chapter it belongs to, page number range, text content, and citation relationships; for bid documents, the structured text block also includes at least a bidder identifier.

[0019] Among them, the block identifier is used to uniquely identify the text block; the chapter to which it belongs is used to indicate the position of the text block in the original document; the page range is used to record the page number range corresponding to the text block; the text content is the substantive content of the text block; the reference relationship is used to indicate the relationship between the text block and other chapters, clauses or requirements; and the bidder identifier is used to identify the bidder to which the text block belongs.

[0020] Through the above deconstruction process, the bidding documents are transformed from unstructured long documents into a sequence of structured text blocks, which facilitates subsequent structured storage, accurate retrieval, and sample association [0†L92-L93].

[0021] S102. Determine the pre-set annotation task, and determine the natural language instructions corresponding to the text data block according to the annotation task.

[0022] The process involves determining a pre-set annotation task and, based on that task, identifying the corresponding natural language instructions for the text data block. Specifically, this includes: determining the application scenario corresponding to the bidding document; and, based on the application scenario, determining the corresponding scenario type and task type. The scenario types include bidding planning scenarios, bidding document inspection scenarios, intelligent assisted bidding scenarios, bidding compliance self-inspection scenarios, collusion detection scenarios, and complaint handling scenarios. The annotation task type corresponding to the text data block is determined based on the task type, which includes question-answering tasks, classification tasks, extraction tasks, and judgment tasks. Based on the scenario type, the annotation task type, and the semantic type of the text data block, corresponding natural language instructions are generated. These instructions include structured prompts, the text content to be analyzed, and domain-level rules. For example, in an intelligent assisted bidding scenario, if the text data block is a technical clause in a bidding document, natural language instructions for extracting technical requirements can be generated; if the text data block is a technical response in a bidding document, natural language instructions for judging the compliance of the response can be generated. By adaptively generating natural language instructions based on the scenario and task type, the large language model can accurately execute different types of annotation tasks.

[0023] S103. The text data block and the corresponding natural language instruction are processed by a large language model to annotate the text data block and obtain structured annotation results.

[0024] The text data blocks are labeled using a pre-defined large language model based on the natural language instructions to obtain structured labeling results. Specifically, this includes: inputting the natural language instructions and the corresponding text data blocks into the large language model; performing semantic processing through the large language model to generate labeling results. The semantic processing steps include semantic understanding, feature extraction, rule matching, and sample generation. Specifically, the large language model is used to perform semantic understanding, feature extraction, rule matching, and sample generation on preprocessed documents and business data, automatically constructing question-answering training sets, classification training sets, extraction training sets, and judgment training sets to replace or supplement expert manual labeling. The labeling results undergo format validation and structure conversion to obtain structured labeling results conforming to a preset data pattern. The structured labeling results include label tags, label content, confidence scores, original text citations used for labeling, and reasoning basis. The original text citations are used to constrain subsequent generation and judgment processes, reducing the risk of content deviating from the original text; the reasoning chain field records the reasoning process and basis from the original text evidence to the output conclusion.

[0025] In one embodiment, one model is used to generate training or problem samples, while another model independently evaluates the sample quality, thereby improving the objectivity and stability of sample generation and quality control. This approach significantly enhances the automation of annotation and the accuracy of semantic understanding, effectively alleviating the problems of low efficiency, high cost, and inconsistent standards associated with traditional manual annotation.

[0026] S104. Associate the structured annotation results with the bidding document to form a bidding document with semantic tags.

[0027] The structured annotation results are then associated with the corresponding text data blocks in the original bidding document to form a bidding document with semantic tags. Specifically, annotation labels, annotation content, and original text citations in the annotation results can be bound to the block identifiers of the text data blocks, so that each text block has a corresponding semantic tag. The resulting bidding document with semantic tags retains the complete content of the original document while adding structured semantic information, facilitating subsequent retrieval, analysis, and model training.

[0028] In one embodiment, the annotation results in the semantically labeled bidding documents undergo dataset de-duplication and processing. The dataset de-duplication and processing steps include deduplication, normalization, conflict resolution, scene attribution verification, label consistency verification, and low-confidence sample isolation. The de-duplicated annotation results are then categorized and stored according to scene type and task type to form a training dataset that can be used for small model training. Through dataset de-duplication and processing, a well-structured, standardized, and scenario-specific training dataset is created.

[0029] In one embodiment, a pre-defined multi-dimensional quality assessment mechanism is used to automatically assess the quality of the categorized and stored training dataset. This multi-dimensional quality assessment mechanism includes one or more of the following: rule validation, model consistency verification, sample difficulty stratification assessment, scene coverage analysis, and sampling indicator verification. Based on the results of the automated quality assessment, samples in the training dataset are filtered to remove unqualified samples and retain high-value samples suitable for model training. Specifically, the Data Juicer tool can be used to call an external evaluation model to automatically score and filter the training dataset across five dimensions: rule validation, model consistency verification, sample difficulty stratification, scene coverage analysis, and sampling indicator verification. This automatically removes low-quality unqualified samples and retains high-value samples suitable for model training. This multi-dimensional quality assessment mechanism effectively solves the problems of self-assessment distortion by a single model and model contamination by low-quality samples.

[0030] In one embodiment, the risk level and scenario importance level of each sample are determined based on the assessment results of the automated quality assessment; samples are then extracted from the training dataset according to the risk level and scenario importance level for review; the approved samples are included in the high-quality training dataset, and the unapproved samples are returned to the sample pool for subsequent re-labeling and retraining. Specifically, the assessed samples can be pushed to domain experts or business personnel for random inspection, revision, and confirmation according to risk level and scenario importance, forming a high-quality training dataset that has been manually reviewed. The sampling ratio can be 1% to 5% of the total, and should cover different score ranges, different requirement types, different judgment difficulties, and different scenario categories.

[0031] In one embodiment, a pre-defined small model is trained based on the high-quality training dataset to enable the small model to possess vertical business capabilities. These vertical business capabilities include question answering, extraction, classification, and judgment. Specifically, based on an approved dataset, a small model specifically designed for the bidding and tendering domain is trained according to scenario or task type, enabling the model to possess vertical business capabilities such as question answering, extraction, classification, and judgment.

[0032] Knowledge distillation is used to transfer the domain reasoning and rule interpretation capabilities from the large language model to the smaller model. Specifically, the trained smaller model is evaluated for performance and compared with the auxiliary capabilities of the large model. Through knowledge distillation, parameter compression, and inference optimization, the domain capabilities of the large model are transferred to the smaller model, reducing deployment costs and improving operational efficiency. The distilled smaller model is deployed to the target bidding and tendering business scenario, and model output results, user revision opinions, and business feedback are collected to determine new sample feedback data. Specifically, the distilled smaller model is deployed to specific application scenarios such as bidding planning, document detection, assisted bid evaluation, collusion identification, and complaint handling; model output results, user revision opinions, and business feedback are collected as new rounds of sample feedback data.

[0033] In one embodiment, based on business changes, rule updates, and scenario expansions, intelligent annotation, data updates, manual review, model retraining, and redistillation are periodically performed to achieve continuous evolution of the small model's capabilities. Specifically, in practical applications, the system records model output results, expert revision opinions, and newly added business samples, and feeds these back to the aforementioned steps, forming a continuous training closed loop of "sample generation, quality assessment, manual review, model training, scenario application, and feedback feedback." Through this continuous training mechanism, the model's capabilities are continuously optimized as business changes occur.

[0034] In one embodiment, a dedicated small model is trained based on the high-quality training dataset and continuously iterated. First, obtain the tender documents and bid documents, and break them down according to chapters, clauses, scoring items, or business rules. Convert the resulting text fragments into structured text blocks. Based on the structured text blocks of the tender documents, automatically extract the tender requirements using a large language model, rule engine, or a combination of both to form a list of tender requirements. Classify each requirement in the list of tender requirements according to requirement type, judgment method, business rules, scenario tags, and task type. For each requirement in the list of tender requirements, retrieve and identify the corresponding response content in the structured text blocks of the bid documents, and establish a "tender requirement - bid response" mapping relationship. Based on the mapping relationship, automatically generate structured training samples. Perform deduplication, integrity verification, conflict detection, and quality assessment on the training samples. Include the samples that pass the automatic cleaning and quality assessment into the available training dataset. Extract samples from the available training dataset according to a preset ratio, and have them manually reviewed by domain experts or business personnel. Use the high-quality training dataset obtained above to train a small model specifically for the tendering and bidding domain, and transfer the capabilities of the large model in complex reasoning and rule interpretation to the small model through knowledge distillation. Deploy the trained small model in intelligent assisted bidding or other related business scenarios. In practical applications, the system records the model output results, expert revision opinions, and newly added business samples, and then feeds them back to the aforementioned steps to form a continuous training loop.

[0035] In one embodiment, a nine-dimensional hierarchical sample labeling system is implemented for the annotation results. Specifically, this system covers standardized labeling paradigms across nine dimensions: sample credibility, difficulty, compliance risk, conclusion clarity, evidence completeness, regulatory relevance, data quality, regulatory type, and special sample attributes. Each dimension is further subdivided into multi-level quantitative grading standards, accurately distinguishing differentiated data types such as high-confidence high-quality samples, low-confidence samples requiring review, high-risk violation samples, and extremely difficult / adversarial / counterexample boundary samples. A closed-loop automatic sorting mechanism is established based on multi-dimensional labels. After unified sample labeling, the dataset is precisely split according to label combination rules: high-quality, usable samples are stratified and allocated according to difficulty to generate a training set; difficult, adversarial, and counterexample samples are grouped into a test set; medium-difficulty samples requiring review are assigned to a validation set; and low-confidence, high-risk, and samples requiring review are automatically pushed to a manual sampling queue. This multi-dimensional labeling system enables precise sampling review and automatic dataset classification and filtering, specifically improving the model's compliance judgment and boundary case robustness training effects.

[0036] In one embodiment, a dual-model architecture of an independent generation model and an evaluation model is used to achieve differentiated verification, avoiding the result bias caused by self-evaluation of a single model, and accurately identifying low-quality question-and-answer samples that fail to meet the quality, accuracy, and completeness standards. A closed-loop automatic repair mechanism is established for unqualified samples, in which defective samples are sent back to the LLM annotation generation model for regeneration, iterative correction, and then sent back to the independent evaluation model for re-examination. At the end of the process, a manual sampling review by maintenance personnel is added to focus on marginal cases and difficult samples, further ensuring the credibility of the dataset. The entire solution relies on dual-model cross-validation to achieve automated identification of question-and-answer data quality, and completes fully automatic repair of low-quality samples through closed-loop regeneration of defective samples.

[0037] In one embodiment, this application deconstructs bidding documents through structured deconstruction and adaptively generates natural language instructions based on the scenario. This enables the large model to accurately perform various annotation tasks such as question answering, classification, extraction, and judgment, significantly improving the automation level of annotation and the accuracy of semantic understanding. This significantly alleviates the pain points of traditional manual annotation, such as low efficiency, high cost, and inconsistent standards. The solution further introduces a dataset clarity governance and multi-dimensional automated quality assessment mechanism. Differentiated manual sampling is conducted based on risk level and scenario importance, ensuring the high quality and compliance of training data and effectively solving the problems of self-assessment distortion of single models and model contamination by low-quality samples. Based on this, knowledge distillation efficiently transfers the reasoning capabilities of the large model to dedicated small models, reducing reasoning costs and deployment barriers while retaining the professional discrimination capabilities of vertical domains. Finally, through post-deployment business feedback and periodic iterative training, continuous evolution and adaptive optimization of model capabilities are achieved, providing a technical path for the intelligent application of the entire bidding process that is engineering-featured, long-term maintainable, cost-effective, and highly reliable.

[0038] like Figure 2 As shown in the illustration, this application also provides a bidding document annotation device based on a large model, including: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable a large-model-based tender document annotation device to perform the method described in any of the embodiments above.

[0039] This application also provides a non-volatile computer storage medium storing computer-executable instructions, which are configured as described in any of the above embodiments.

[0040] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0041] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0042] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0043] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0044] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0045] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0046] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0051] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0052] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0053] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0054] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for annotating bidding documents based on a large model, characterized in that, include: Obtain the bidding documents, preprocess the bidding documents, and deconstruct the preprocessed bidding documents to obtain multiple text data blocks; Determine a pre-set annotation task, and determine the natural language instructions corresponding to the text data block based on the annotation task; The text data blocks are annotated based on the natural language instructions using a pre-determined large language model to obtain structured annotation results; The structured annotation results are associated with the bidding documents to form bidding documents with semantic tags.

2. The method according to claim 1, characterized in that, The preprocessed bidding documents are deconstructed to obtain multiple text data blocks, specifically including: The bidding document is split according to a pre-set splitting index to divide the long document into multiple independently processable text segments. The splitting index includes chapters, clauses, scoring items, and business rules. The text fragments are converted into structured text blocks, wherein the structured text blocks of the tender documents include block identifiers, chapters, page ranges, text content, and reference relationships, and the structured text blocks of the bid documents include bid subject identifiers.

3. The method according to claim 1, characterized in that, Determine a pre-set annotation task, and determine the natural language instructions corresponding to the text data block based on the annotation task, specifically including: Determine the application scenario corresponding to the bidding document, and determine the corresponding scenario type and task type based on the application scenario. The scenario type includes bidding planning scenario, bidding document detection scenario, intelligent assisted bid evaluation scenario, bid compliance self-inspection scenario, collusion and bid rigging identification scenario, and complaint handling scenario. The annotation task type corresponding to the text data block is determined according to the task type. The annotation task type includes question answering task, classification task, extraction task, and judgment task. Based on the scenario type, the annotation task type, and the semantic type of the text data block, corresponding natural language instructions are generated. The natural language instructions include structured prompt words, the text content to be analyzed, and domain classification rules.

4. The method according to claim 1, characterized in that, The text data blocks are annotated based on the natural language instructions using a pre-determined large language model to obtain structured annotation results, specifically including: The natural language instructions and the corresponding text data blocks are input into the large language model, and semantic processing is performed through the large language model to generate annotation results. The semantic processing steps include semantic understanding, feature extraction, rule matching, and sample generation. The annotation results are formatted and structured to obtain structured annotation results that conform to a preset data pattern. The structured annotation results include annotation labels, annotation content, confidence scores, original text citations used for annotation, and reasoning basis.

5. The method according to claim 1, characterized in that, The method further includes: The annotation results in the bidding documents with semantic tags are subjected to dataset clarification and governance. The steps of dataset clarification and governance include deduplication, normalization, conflict resolution, scene attribution verification, label consistency verification, and low-confidence sample isolation. The clarified annotation results are classified and stored according to scene type and task type to form a training dataset that can be used for training small models.

6. The method according to claim 5, characterized in that, The method further includes: The training dataset after classification and storage is automatically evaluated through a preset multi-dimensional quality evaluation mechanism. The multi-dimensional quality evaluation mechanism includes any one or more of the following: rule verification, model consistency verification, sample difficulty stratification evaluation, scene coverage analysis, and sampling index review. Based on the evaluation results of the automated quality assessment, the samples in the training dataset are screened to remove unqualified samples and retain high-value samples suitable for model training.

7. The method according to claim 6, characterized in that, The method further includes: The risk level and scenario importance level of each sample are determined based on the assessment results of the automated quality assessment. Samples are drawn from the training dataset based on the risk level and the importance level of the scenario, and the drawn samples are reviewed. The approved samples are included in the high-quality training dataset, while the unapproved samples are returned to the sample pool for subsequent re-annotation and retraining.

8. The method according to claim 7, characterized in that, The method further includes: The preset small model is trained based on the high-quality training dataset so that the small model has vertical business capabilities, including question answering, extraction, classification and judgment capabilities. Knowledge distillation is used to transfer the domain reasoning and rule interpretation capabilities from the large language model to the small model. The distilled small model is deployed to the target bidding business scenario, and the model output results, user revision opinions and business feedback are collected to determine the new sample feedback data; Based on business changes, rule updates, and scenario expansion, we periodically perform intelligent annotation, data updates, manual review, model retraining, and redistillation to achieve continuous evolution of the small model's capabilities.

9. A bidding document annotation device based on a large model, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the large-model-based tender document annotation device to perform the method described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to be the method as described in any one of claims 1-8.

Images