Bid invitation file key information extraction method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of information processing, and relates to a bid invitation file key information extraction method and system based on artificial intelligence, the system comprises a file input module, an AI extraction module, a self-learning module and an output module, and the method comprises file input, AI extraction, preliminary report generation, user decision, feedback / learning, re-extraction and new report until accuracy. Through user-driven self-learning circulation, the system not only can process standard bid invitation files, but also can adapt to personalized requirements (such as terms of a specific industry), and the extraction efficiency and accuracy are remarkably improved. In addition, the modular design facilitates expansion, supports cloud deployment, and is suitable for bid invitation scenes such as enterprises and governments.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method and system for extracting key information from tender documents based on artificial intelligence. The system utilizes a large language model to automatically extract key information from tender documents and has self-learning capabilities. Background Technology

[0002] Tender documents are common in government procurement and corporate bidding activities. They typically contain a large amount of textual information, such as deadlines (e.g., registration deadline, bid security submission deadline, bid closing date), scoring criteria, eligibility requirements, and contract terms. This crucial information is often scattered throughout lengthy documents, making manual extraction inefficient and prone to omission. Furthermore, users may overlook potentially important points, such as environmental requirements or breach of contract risks.

[0003] In existing technologies, some tools use optical character recognition (OCR) or simple keyword search to process documents. However, these methods cannot understand the semantic context of the text, leading to inaccurate extraction and a lack of dynamic learning capabilities, failing to automatically expand the extraction categories based on user feedback or new data. Therefore, an intelligent system is needed that can achieve accurate extraction using advanced Large Language Models (LLM) and discover more key information through a self-learning mechanism, thereby improving the systematicness and comprehensiveness of bidding analysis.

[0004] Currently, there are many inventions on the market related to information extraction from bidding documents based on large language models. For example, the invention patent (application number 202410091142.7, entitled "A Method for Extracting Information from Bidding Documents Based on Large Language Models") has the following technical features: 1. Comprehensive data preparation process: Data collection: The dataset broadly covers bidding documents from multiple fields, including construction, transportation, and information technology, and supports various formats (PDF, Word, Excel). This ensures the diversity and representativeness of the dataset, preventing the model from being biased towards a single scenario.

[0005] Preprocessing and Enhancement: Non-text files are converted using OCR (Optical Character Recognition), and text cleaning is performed (e.g., removing irrelevant content, correcting errors, and standardizing terminology). Data augmentation employs synonym replacement, back-translation (translating and then back-translating), and rule-based distortion to expand the dataset size and improve the model's generalization ability. This is the innovative aspect, capable of handling real-world data noise. 2. Refined data annotation and prompt word design: Labeling: Establish detailed guidelines, label key information (such as project number, registration date), and implement quality control. This provides high-quality labeled data for model training.

[0006] Prompt word design: By analyzing the file structure, prompt words (such as JSON format instructions) are designed and optimized, and tested using models such as GPT-4 to select the best prompt. This guides LLM to extract information more accurately, demonstrating the application of Prompt Engineering, a key innovation in the LLM era, and can significantly improve the performance of zero-shot or few-shot learning. 3. Efficient model training and fine-tuning: We select a pre-trained LLM (such as BERT, GPT, LLaMA) and fine-tune it using LoRA (Low-Rank Adaptation). LoRA adjusts model parameters by adding a low-rank matrix without requiring full training, saving computational resources (suitable for domain-specific adaptations in bidding documents). This allows the model to handle complex semantics (such as technical terms and implicit meanings) and optimizes at the self-attention and feedforward layers, adapting to the structured requirements of bidding documents. Compared to full parameter fine-tuning, LoRA is more efficient, demonstrating the advancements of Parameter Efficient Fine-Tune (PEFT).

[0007] However, this patent has the following problems: 1. High data dependency and high cost: The method heavily relies on extensive data collection, preprocessing, and annotation (e.g., nationwide, multi-channel data collection). This can lead to high costs (human resources, time) and data privacy risks (bidding documents are often sensitive). If the data is incomplete (e.g., lacking information on certain niche industries), the model's generalization ability will be limited.

[0008] 2. Insufficient self-learning and adaptability: Although there are fine-tuning and prompts for optimization, the lack of a true "real-time self-learning" mechanism (such as dynamically updating the model based on user feedback, rather than one-time training) may prevent it from adapting to new types of tender documents or personalized user needs.

[0009] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The purpose of this invention is to address the problems of inaccurate extraction of key information from tender documents and lack of dynamic optimization in existing technologies. To this end, an artificial intelligence-based method and system for extracting key information from tender documents is provided. This method achieves automated extraction through a Large Language Model (LLM) and introduces a self-learning mechanism to support user feedback-driven iterative optimization, ensuring the accuracy and adaptability of the extraction results.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for extracting key information from tender documents based on artificial intelligence includes the following steps: Step S1: The user uploads the tender documents through the file input module, and the system preprocesses them to generate parsable text; Step S2: The AI ​​extraction module uses an LLM model to parse the text and extract key information; Step S3: The output module generates a preliminary structured report based on the extraction results for users to view; Step S4: After reviewing the preliminary report, the user determines whether they are satisfied. If satisfied, the process ends; if not, the feedback stage begins. Step S5: Collect user feedback through the self-learning module, update the model parameters, and loop back to the AI ​​extraction module to reprocess the file and generate a new report; Step S6: Repeat steps S2-S5 until the user is satisfied with the report, at which point the system will stop working.

[0012] The following is a further technical solution for the method of this invention, with key information including time nodes, scoring criteria, qualification requirements, and contract terms.

[0013] The following are further technical solutions that define the method in this invention, and user feedback includes adding keywords and deleting incorrectly extracted information.

[0014] An AI-based key information extraction system for tender documents, used to implement the aforementioned AI-based key information extraction method for tender documents, includes: The file input module is used to receive and preprocess user-uploaded tender documents, which include files in various formats. The module uses optical character recognition technology to convert unstructured files into parsable text, ensuring the standardization of input data. The AI ​​extraction module uses a pre-trained LLM model to perform semantic analysis and extract key information from pre-processed text. The self-learning module is used to achieve dynamic optimization of the system by updating model parameters by collecting user feedback; The output module is responsible for converting the extracted results into a structured report.

[0015] The following is a further defined technical solution for the system in this invention: the AI ​​extraction module includes a semantic parsing unit and an information matching unit; The semantic parsing unit uses natural language processing technology for word segmentation, entity recognition, and semantic understanding to identify key information in the tender documents. The information matching unit maps the parsing results to structured categories based on preset templates or LLM generative matching.

[0016] The following is a further defined technical solution for the system in this invention: the self-learning module includes a feedback collection unit and a model update unit; The feedback collection unit records user modifications to the extraction results, including adding specific keywords or deleting incorrectly extracted information; The model update unit uses machine learning techniques to fine-tune LLM parameters, expand extraction categories, or optimize algorithms to adapt to new scenarios.

[0017] The following are further technical solutions for the system in this invention, which also include cloud servers; Cloud servers are used to provide computing resources to support system deployment, enabling remote access and data storage; The file input module transmits preprocessed text to the AI ​​extraction module; the AI ​​extraction module and the self-learning module are connected in a two-way feedback loop to achieve iterative learning; the final result is presented through the output module. The cloud server connects all modules and supports scalability.

[0018] Compared with the prior art, the present invention has the following technical effects: This invention utilizes a user-driven self-learning loop, enabling the system to not only process standard bidding documents but also adapt to personalized needs (such as industry-specific terminology), significantly improving extraction efficiency and accuracy. Furthermore, its modular design facilitates expansion, supports cloud deployment, and is suitable for bidding scenarios in enterprises, governments, and other sectors.

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a flowchart of the method operation of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] like Figure 1 As shown, an AI-based key information extraction system for tender documents is provided, mainly composed of a document input module 1, an AI extraction module 2, a self-learning module 3, an output module 4, and a cloud server 5. The document input module 1, AI extraction module 2, self-learning module 3, and output module 4 can optionally be deployed on the cloud server 5. All modules are connected via data transmission, forming a modular structure.

[0024] File Input Module 1: Used to receive and preprocess user-uploaded tender documents, supporting multiple formats such as PDF, Word, or Excel. It uses Optical Character Recognition (OCR) technology to convert unstructured documents into parsable text, ensuring the standardization of input data.

[0025] AI Extraction Module 2: The core processing unit, utilizing a pre-trained large language model (LLM, such as the GPT series or similar models) to perform semantic analysis and extract key information from the pre-processed text. Due to its reliance on LLM models, it may be affected by model bias or computational resources; therefore, robustness measures (such as multi-model fusion) are added to the implementation. This module includes: Semantic parsing unit 2a: uses natural language processing (NLP) technology for word segmentation, entity recognition and semantic understanding to identify core elements in the tender document (such as time nodes and scoring criteria).

[0026] Information matching unit 2b: Based on a preset template or LLM generative matching, the parsing results are mapped to structured categories.

[0027] AI Extraction Module 2 is more intelligent than traditional rule-based methods and can handle complex tender documents (such as time nodes, scoring criteria, and qualification requirements).

[0028] Self-learning module 3: Enables dynamic optimization of the system by updating model parameters through user feedback. This module includes: Feedback collection unit 3a: Records user modifications to the extraction results, such as adding specific keywords (e.g., "eligibility requirements") or deleting incorrectly extracted information (e.g., misidentified redundant data).

[0029] Model update unit 3b: Use machine learning techniques (such as LoRA low-order adaptation) to fine-tune LLM parameters, expand extraction categories or optimize algorithms to adapt to new scenarios.

[0030] Self-learning module 3 forms a closed-loop learning process, enhancing the system's adaptability (e.g., adapting to specific industry terminology).

[0031] Output module 4: Responsible for converting the extracted results into user-friendly structured reports, supporting export of tables, reports, or visualizations (such as timeline charts or rating matrices).

[0032] Cloud Server 5 (optional): Provides computing resources to support system deployment, enabling remote access and data storage. Due to the sensitive information involved in processing the tender documents, Cloud Server 5 requires encryption protection.

[0033] The data flow between modules is shown in the attached figure. Figure 1 As shown: File input module 1 transmits preprocessed text to AI extraction module 2; AI extraction module 2 and self-learning module 3 are connected in a bidirectional feedback loop to achieve iterative learning; the final result is presented through output module 4. Cloud server 5 can connect all modules via dotted lines, supporting scalability.

[0034] like Figure 2 As shown, this paper presents an artificial intelligence-based method for extracting key information from tender documents. An iterative loop design is employed to ensure extraction accuracy. The specific steps are as follows: Step 1, File Input: The user uploads the tender document through the file input module 1. The system performs preprocessing (such as OCR conversion) to generate parsable text.

[0035] Step 2, AI Extraction: The AI ​​extraction module 2 uses LLM to parse the text and extract key information (such as time points, scoring criteria, qualification requirements, etc.).

[0036] Step 3, Initial Output: Output module 4 generates an initial structured report (in table, report, or visualization format) based on the extraction results for users to view.

[0037] Step 4: User Decision: After reviewing the preliminary report, the user determines whether they are satisfied (i.e., whether the information is accurate enough). If satisfied, the process ends; if not, it proceeds to the feedback stage.

[0038] Step 5, Self-Learning: The self-learning module 3 collects user feedback, such as adding keywords (to expand the extraction scope) or deleting incorrectly extracted information (to improve accuracy). The system updates the model parameters and loops back to the AI ​​extraction module 2 to reprocess the file and generate a new report.

[0039] Iterative loop: Repeat steps 2-5 until the user is satisfied with the report, at which point the system terminates its work. This loop mechanism ensures that the system can adaptively optimize based on user feedback, gradually improving extraction accuracy.

[0040] This emphasizes user interaction and incremental improvement, unlike one-off extraction systems.

[0041] Dynamic optimization vs. static extraction: Existing bidding information extraction tools (such as those based on keyword matching or simple OCR) lack feedback mechanisms, and their accuracy is limited by the initial model. The solution of this invention introduces an iterative learning loop (user satisfaction decision point), allowing the system to "adapt" to new data and improve accuracy.

[0042] The integration of LLM with user interaction: Combining the generation capabilities of LLM with user feedback to form a "human-machine collaboration" model. This is relatively new in the bidding field, capable of handling semantically ambiguous or non-standard documents and solving the "long tail problem" (extraction errors in rare scenarios).

[0043] Visualization and output diversity: Output module 4 supports tables, reports, or visualizations (such as timeline charts), enhancing the user experience beyond basic text extraction.

[0044] The working process of this embodiment will be further described below: Simple scenario: Processing PDF format tender documents (initial extraction, no iteration).

[0045] In this scenario, the system processes a standard PDF tender document to demonstrate the basic extraction process.

[0046] Input data: The user uploads a PDF file, and the following is a fragment of its content (text after OCR preprocessing): Tender Announcement: Municipal Road Engineering Project. Bid Submission Deadline: November 1, 2025. Bid Opening Date: November 2, 2025. Evaluation Criteria: Technical proposal accounts for 60 points, price quotation accounts for 40 points. Qualification Requirements: Enterprise qualification level no lower than Level II.

[0047] Processing steps: The file input module 1 performs OCR conversion to generate parsable text. The AI ​​extraction module 2 uses LLM (e.g., GPT-4 model) for semantic parsing and information matching, with preset extraction categories including "time node" and "scoring criteria". Without user feedback, a preliminary report is generated directly.

[0048] Output: Module 4 generates a structured table report: Table 1:

[0049] Therefore, this demonstrates the efficiency of LLM in unstructured text, but iterative optimization may be necessary if the file is complex.

[0050] Complex scenario: Processing Word format bidding documents (with two iterations based on user feedback).

[0051] This scenario demonstrates a self-learning mechanism for processing Word documents containing ambiguous terms.

[0052] Input data: Word document fragment (preprocessed text): Project: Bridge construction. Response deadline: December 15, 2025; Bid security deadline: December 15, 2025. Evaluation method: Comprehensive evaluation; technical section: 100 points (50 points for innovation, 50 points for implementation plan); commercial section: 50 points. Additional requirements: Environmental compliance.

[0053] Processing Steps: The file input module 1 converts the text. The initial AI extraction module 2 outputs a preliminary report, which the user finds unsatisfactory (e.g., "environmental compliance" was not extracted as a qualification requirement). The process then moves to the self-learning module 3: Feedback collection unit 3a records the user input "add keyword: environmental compliance," and model update unit 3b uses LoRA fine-tuning (learning rate 0.001, 1 iteration) to expand the extraction categories. The process loops back to the AI ​​extraction module 2 to reprocess the output report. The user finds that the bid bond deadline and response deadline are the same but equally important, requiring model learning and updates, and is therefore unsatisfactory. The process then moves back to the self-learning module 3: Feedback collection unit 3a records the user input "add keyword: bid bond deadline," and model update unit 3b uses LoRA fine-tuning (learning rate 0.001, 1 iteration) to expand the extraction categories.

[0054] Output (after iteration): Output module 4 generates a structured table report.

[0055] Table 2:

[0056] Therefore, by emphasizing dynamic optimization driven by user feedback, the problem of insufficient adaptability of existing technologies is solved, and it can be applied to most government and state-owned enterprise bidding projects.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention using the methods and techniques disclosed above, or modify it into equivalent embodiments with equivalent changes, without departing from the scope of the present invention. Therefore, all equivalent changes made based on the shape, structure, and principle of the present invention without departing from the scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for extracting key information from tender documents based on artificial intelligence, characterized in that, Includes the following steps: Step S1: The user uploads the tender documents through the file input module, and the system preprocesses them to generate parsable text; Step S2: The AI ​​extraction module uses an LLM model to parse the text and extract key information; Step S3: The output module generates a preliminary structured report based on the extraction results for users to view; Step S4: After reviewing the preliminary report, the user determines whether they are satisfied. If satisfied, the process ends; if not, the feedback stage begins. Step S5: Collect user feedback through the self-learning module, update the model parameters, and loop back to the AI ​​extraction module to reprocess the file and generate a new report; Step S6: Repeat steps S2-S5 until the user is satisfied with the report, at which point the system will stop working.

2. The method for extracting key information from tender documents based on artificial intelligence as described in claim 1, characterized in that, Key information includes timelines, scoring criteria, eligibility requirements, and contract terms.

3. The method for extracting key information from tender documents based on artificial intelligence as described in claim 1, characterized in that, User feedback included adding keywords and removing incorrectly extracted information.

4. A key information extraction system for tender documents based on artificial intelligence, used to implement the key information extraction method for tender documents based on artificial intelligence as described in any one of claims 1-3, characterized in that, include: The file input module is used to receive and preprocess user-uploaded tender documents, which include files in various formats. The module uses optical character recognition technology to convert unstructured files into parsable text, ensuring the standardization of input data. The AI ​​extraction module uses a pre-trained LLM model to perform semantic analysis and extract key information from pre-processed text. The self-learning module is used to achieve dynamic optimization of the system by updating model parameters by collecting user feedback; The output module is responsible for converting the extracted results into a structured report.

5. The key information extraction system for tender documents based on artificial intelligence as described in claim 4, characterized in that, The AI ​​extraction module includes a semantic parsing unit and an information matching unit; The semantic parsing unit uses natural language processing technology for word segmentation, entity recognition, and semantic understanding to identify key information in the tender documents. The information matching unit maps the parsing results to structured categories based on preset templates or LLM generative matching.

6. The AI-based key information extraction system for tender documents as described in claim 4, characterized in that, The self-learning module includes a feedback collection unit and a model update unit; The feedback collection unit records user modifications to the extraction results, including adding specific keywords or deleting incorrectly extracted information; The model update unit uses machine learning techniques to fine-tune LLM parameters, expand extraction categories, or optimize algorithms to adapt to new scenarios.

7. The AI-based key information extraction system for tender documents as described in claim 4, characterized in that, It also includes cloud servers; Cloud servers are used to provide computing resources to support system deployment, enabling remote access and data storage; The file input module transmits preprocessed text to the AI ​​extraction module; the AI ​​extraction module and the self-learning module are connected in a two-way feedback loop to achieve iterative learning; the final result is presented through the output module. The cloud server connects all modules and supports scalability.

Images