A method for extracting information from bidding documents

Abstract

This application relates to the field of text processing, and more particularly to a method for extracting information from bidding documents. The method includes: segmenting the bidding document into pages, identifying the corresponding text for each page; generating supplementary text descriptions for images and tables within the pages and appending them to the end of the corresponding text to form an enhanced text block sequence; matching tags from the text block sequence according to a pre-built hierarchical tagging system, and generating corresponding prompt word templates based on the tags and a pre-built prompt word template library; inputting the prompt word templates, the enhanced text block sequence, and the contextual text summary as a combined input to a large language model to obtain a structured extraction result with hierarchical relationships; matching the extracted entity content with a local dictionary, and after successful matching, aggregating and organizing the results to output a structured data file. This method reduces the risk of illusions in the generated content without requiring model retraining.

Description

Technical Field

[0001] This application relates to the field of text processing, and in particular to a method for extracting information from bidding documents. Background Technology

[0002] In the bidding process for equipment procurement in engineering construction projects, a large amount of unstructured documents are involved, such as tender documents and bid documents. These documents are usually in PDF format and have complex content, including descriptive technical parameters, qualification requirements, and other textual information, as well as embedded technical drawings, test reports, complex tables, and other visual content. Accurately and efficiently extracting key entity information, such as equipment model, design pressure, and supplier qualification certificate number, from these multimodal documents and converting it into structured data that can be directly processed by computers is a crucial step in achieving intelligent management and decision-making in the bidding process.

[0003] In recent years, with the rise of large language model technology, technical solutions have emerged that directly utilize such models for end-to-end information extraction. While these solutions possess strong semantic understanding capabilities, they have significant limitations when applied to scenarios with extremely high accuracy requirements, such as bidding and tendering processes.

[0004] On the one hand, the inherent illusion phenomenon of large language models may cause them to generate information that does not exist in the original text, which is unacceptable for rigorous bidding activities.

[0005] On the other hand, if a large model is fine-tuned or retrained in order to improve the extraction accuracy in a specific domain, it requires huge computational resources and time costs, making it difficult to promote the solution in practical applications and greatly reducing its practicality. Summary of the Invention

[0006] To address the aforementioned problems, this application provides a method for extracting information from bidding documents, including:

[0007] The bidding documents to be processed are divided into pages, and OCR recognition is performed on each page to obtain the corresponding text;

[0008] For each image and table on the page, a visual language model is used to generate supplementary text descriptions, which are then appended to the end of the corresponding text on the page to form an enhanced text block sequence.

[0009] Tags are matched from the text block sequence according to the pre-built hierarchical tag system, and corresponding prompt word templates are generated according to the tags and the pre-built prompt word template library. The prompt word templates are designed according to the hierarchical tag system so that the large language model is constrained by the hierarchical tag system when generating content.

[0010] The prompt word template, the enhanced text block sequence, and the context text summary are combined and input into the large language model, wherein the context text summary is the text summary of the pages before and after the page corresponding to the prompt word template, and the large language model obtains a structured extraction result with hierarchical relationships.

[0011] The extracted entity content is matched with the local dictionary. After a successful match, the results are aggregated and organized to output a structured data file.

[0012] Specifically, after the step of forming the enhanced text block sequence, a convolutional neural network is used to extract the image feature vector of the page, and a natural language processing encoder is used to extract the text feature vector of the corresponding text of the page. The cosine similarity between the image feature vector and the text feature vector is calculated. When the cosine similarity is lower than a preset threshold, the page is marked as a low-confidence page for manual review.

[0013] Specifically, the pre-constructed hierarchical tagging system includes primary tags and secondary tags. The primary tags include technical parameter categories, qualification certificate categories, performance capability categories, and multimodal association categories.

[0014] The secondary labels under the technical parameters category include equipment model, design pressure, and volume;

[0015] The secondary labels under the qualification certificate category include business license number and ISO certification number.

[0016] The secondary labels under the category of performance capability include financial statements and type test reports for the past three years;

[0017] The secondary tags under the multimodal association class include drawing interface annotations and chart data values.

[0018] Specifically, the method for matching tags from the text block sequence based on a pre-constructed hierarchical tag system includes:

[0019] Calculate the semantic similarity between the text block content and the keywords of each level of tags, and match the tag with the highest similarity as the dominant tag of the current text block.

[0020] Specifically, the prompt word template includes a requirement instruction, which controls the large language model to output the identified entities in JSON format according to the hierarchical structure of the hierarchical tag system, and explicitly marks the parent-child hierarchical relationship between entities in the JSON.

[0021] Specifically, the prompt word template library is a dynamic prompt word library. During the execution of the method, the extraction results of the large language model for specific entities are continuously monitored. When an entity is not recognized three times in a row or its recognition confidence is consistently lower than 0.8, an update mechanism is automatically triggered to extract the context sentences in which the entity appears and use the large language model to generate candidate prompt words that emphasize different expressions of the entity. After manual review, these prompt words are included in the word library.

[0022] Specifically, the contextual text summary is generated by extracting key sentences from the text of the two pages before and the two pages after the current page, so as to provide cross-page contextual semantic background for the large language model.

[0023] Specifically, in the step of matching the extracted entity content with the local dictionary, if the matching fails, a notification mechanism is automatically triggered to remind business personnel to check and feed back the latest information to the system for correcting the extraction results or marking them as pending confirmation.

[0024] Specifically, the method for aggregating and organizing the results includes:

[0025] Standardize numerical values ​​to international standard units and convert dates to a unified format;

[0026] The Sentence-BERT model is used to calculate the semantic vector similarity of entity representations and to normalize synonyms with similarity higher than 0.9 into standard terms.

[0027] The system determines whether to perform deduplication based on the specific chapter or topic in which the entity appears in the document, removing duplicate entities only when the context and topic are consistent.

[0028] Specifically, the use of a visual language model to generate supplementary text descriptions is performed on pages containing technical drawings or complex tables. The generated descriptions are used to explain structural relationships in images or to supplement data relationships in tables that were not fully extracted by OCR.

[0029] This application has the following technical advantages:

[0030] This invention integrates multimodal information processing and context awareness mechanisms, and guides an existing large language model to extract information based on a constructed domain knowledge system. Without the need to retrain the model, it significantly reduces the risk of generated content illusion, ensures the high precision and high reliability required in bidding scenarios, and greatly reduces the cost of technology deployment and maintenance. It provides a practical and efficient solution for realizing intelligent bidding document processing. Attached Figure Description

[0031] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts.

[0032] Figure 1 This is a flowchart illustrating the steps of a method for extracting bidding document information in an embodiment of this application. Detailed Implementation

[0033] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0034] The tendency of large language models to generate illusory content stems from their generation mechanism, which is based on probabilistic models trained on massive amounts of data. Their goal is to generate semantically fluent and logical text, rather than strictly adhering to the given source text. While this characteristic is an advantage in open-domain dialogue, it becomes a fatal flaw in scenarios requiring absolute accuracy, such as bidding and tendering. Although fine-tuning all parameters by introducing large amounts of high-quality domain data can alleviate this problem to some extent, it introduces another equally intractable challenge—enormous resource consumption. Large language models possess hundreds of billions or even trillions of parameters; fine-tuning them requires massive computing clusters and training time of days or even weeks, the economic and time costs of which are prohibitive for most enterprises. More importantly, the norms, terminology, and entity relationships in the bidding and tendering domain are not static; changes can necessitate a new, expensive round of fine-tuning. This high adaptation cost severely restricts the large-scale practical application of the technology. Therefore, the core issue lies in how to efficiently and accurately "inject" domain knowledge into the model without large-scale modifications to the model itself, and effectively constrain its output behavior.

[0035] Based on the above background, this application proposes a method for extracting information from bidding documents, such as... Figure 1 As shown, it includes the following steps:

[0036] The bidding documents to be processed are divided into pages, and OCR recognition is performed on each page to obtain the corresponding text;

[0037] For each page, a visual language model is used to generate supplementary text descriptions for the images and tables, which are then appended to the end of the corresponding text on the page to form an enhanced text block sequence.

[0038] Tags are matched from the text block sequence based on a pre-built hierarchical tag system, and corresponding prompt word templates are generated based on the tags and a pre-built prompt word template library. The prompt word templates are designed according to the hierarchical tag system so that the large language model is constrained by the hierarchical tag system when generating content.

[0039] The prompt word template, the enhanced text block sequence, and the context text summary are used as combined inputs to the large language model. The context text summary is the text summary of the pages before and after the page corresponding to the prompt word template. The large language model obtains a structured extraction result with hierarchical relationships.

[0040] The extracted entity content is matched with the local dictionary. After a successful match, the results are aggregated and organized to output a structured data file.

[0041] In one embodiment of this application, the entire method begins with the tender document to be processed, typically in PDF format. First, data preprocessing is performed, using a document parsing library to segment the PDF file into independent page images. For each page, an optical character recognition engine is used to recognize the text and obtain the original text content of that page. Tender documents usually contain numerous technical drawings, product diagrams, complex tables, and scanned copies of company qualification certificates bearing official seals. These elements often contain key information, but pure OCR processing has limited effectiveness. Therefore, this embodiment introduces a visual language model to perform in-depth analysis of these non-text elements. For example, for a structural diagram of a pressure vessel, the visual language model can generate descriptive text such as, "This drawing shows a vertical pressure vessel with a safety valve interface A at the top, an inlet B and an outlet C marked on the side, and a support at the bottom." This supplementary description, generated by the model after understanding the image semantics, is appended to the OCR-recognized text on the same page, forming a more complete enhanced text block. The enhanced text blocks of all pages sequentially form the enhanced text block sequence of the entire document, laying a high-quality data foundation for subsequent analysis. In this embodiment, the use of a visual language model to generate supplementary text descriptions is performed on pages containing technical drawings or complex tables. The generated descriptions are used to explain structural relationships in images or to supplement data relationships in tables that were not fully extracted by OCR.

[0042] In this embodiment, a hierarchical tagging system deeply aligned with the bidding and tendering field is pre-constructed, including primary and secondary tags. Primary tags include technical parameters, qualification certificates, performance capability, and multimodal association. Secondary tags under the technical parameters category include equipment model, design pressure, and volume. Secondary tags under the qualification certificates category include business license number and ISO certification number. Secondary tags under the performance capability category include financial statements for the past three years and type test reports. Secondary tags under the multimodal association category include drawing interface annotations and chart data values. This system precisely defines the entity types to be extracted and their hierarchical relationships. For example, "Equipment A" is the parent entity, while "Model B" and "Design Pressure C" are its child entities. Based on this tagging system, an initial prompt word template library is designed. These templates contain explicit instructions requiring the large language model to strictly adhere to the above hierarchical structure, output recognition results in a standardized JSON format, and explicitly annotate the parent-child relationships between entities. In this embodiment, the method for matching tags from a text block sequence based on the pre-constructed hierarchical tagging system includes:

[0043] Calculate the semantic similarity between the text block content and the keywords of each level of tags, and match the tag with the highest similarity as the dominant tag of the current text block.

[0044] In the information extraction process, to improve accuracy, especially in addressing the issues of referential resolution and contextual dependency, this embodiment pairs each text block with its contextual text summary. The contextual text summary is generated by extracting key sentences from the text of the two pages preceding and following the current page, providing a cross-page contextual semantic background for the large language model. Specifically, the text content of the two pages preceding and following the page corresponding to the current text block is extracted, and a concise contextual summary is generated using a text summarization model or key sentence extraction algorithm. Subsequently, the prompt word template, the enhanced text block of the current page, and its contextual summary are combined to form a complete input, which is submitted to a large language model that supports long text and mixed text-image input for processing. Since the prompt words already embed structured domain knowledge constraints, the large language model, guided by this, can accurately locate key information from the text and output structured data with clear hierarchical relationships in the required JSON format. This application utilizes prompt word engineering to replace model parameter updates, thereby achieving effective control over model behavior with almost no additional computational cost and reducing the risk of the model generating illusions.

[0045] After the initial extraction is complete, the entity verification stage begins. The extracted entity information, such as supplier names and certificate numbers, is matched against a locally maintained, regularly updated qualification certificate dictionary database. If a match is successful, the result is considered highly reliable. If a match fails, for example, due to the discovery of a new supplier not yet included or an incorrect certificate number format, the system will automatically trigger a notification mechanism to remind business personnel to conduct manual verification and intervention, ensuring the accuracy of the final result. Finally, all verified extraction results are aggregated and organized using methods including:

[0046] Standardize numerical values ​​to international standard units and convert dates to a unified format;

[0047] The Sentence-BERT model is used to calculate the semantic vector similarity of entity representations and normalize synonyms with a similarity higher than 0.9 into standard terms. In practice, this model is based on calculating cosine similarity to obtain semantic vector similarity. The value range of cosine similarity is between -1 and 1, where -1 indicates complete irrelevance and 1 indicates complete semantic consistency. The closer the value is to 1, the more semantically similar the two entity representations are. In the bidding document processing scenario, the semantic similarity value is usually distributed in the range of 0 to 1. This embodiment uses 0.9 as the judgment threshold for synonym normalization. When the semantic vector similarity result of two entity contents is higher than this threshold, they are judged to have a highly consistent semantic connotation.

[0048] The system determines whether to perform deduplication based on the specific chapter or topic in which an entity appears in the document, removing duplicate entities only when the context and topic are consistent. Ultimately, the system outputs a complete, accurate, and structured data file that can be directly used for subsequent analysis, comparison, and decision support.

[0049] To further enhance the robustness and adaptability of the system, this embodiment also includes a dynamic prompt word library mechanism. The prompt word template library is a dynamic prompt word library that continuously monitors the extraction results of specific entities by the large language model during method execution. When an entity is not identified three times consecutively or its identification confidence score remains below 0.8, an update mechanism is automatically triggered. This mechanism extracts the contextual statements in which the entity appears and uses the large language model to generate candidate prompt words emphasizing different expressions of the entity. These candidate prompt words are then manually reviewed and added to the library. For example, if the specific entity "design pressure" is not identified three times consecutively, or its identification confidence score remains below the threshold of 0.8, the system determines that the current prompt words for that entity may not be effective enough, thus automatically triggering an update process. The update mechanism extracts the contextual statements in which the entity was not correctly identified and then uses the analytical capabilities of the large language model to generate new candidate prompt words. These new prompt words emphasize synonymous expressions of the entity. The generated candidate prompt words are then manually reviewed and confirmed by domain experts before being officially added to the dynamic prompt word library. The dynamic suggestion lexicon mechanism can proactively adapt to new terms, expressions, or local idioms that may appear in bidding documents, further enhancing its adaptability to domain changes without requiring model retraining.

[0050] To ensure the quality of input data, this embodiment also includes an image-text consistency verification step. After forming the enhanced text block sequence, a convolutional neural network is used to extract the image feature vector of the page, and a natural language processing encoder is used to extract the text feature vector of the corresponding text on the page. The cosine similarity between the image feature vector and the text feature vector is calculated. When the cosine similarity is lower than a preset threshold, the page is marked as a low-confidence page for manual review. Specifically, the value range of cosine similarity is between -1 and 1, with the value closer to 1 indicating greater similarity. A confidence threshold of 0.7 is set. When the calculated cosine similarity is lower than 0.7, it indicates a significant semantic deviation between the OCR-recognized text and the original image, such as missing large sections of text or misrecognizing key numbers. Such pages are marked as "low-confidence," their paths are recorded, and they are pushed to the manual review queue for intervention. This reduces the probability of subsequent extraction failures due to input data errors from the source, increasing the reliability of the entire process.

[0051] Obviously, the embodiments described above are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0052] It should be understood that when the terms "first," "second," etc., are used in the claims, description, and drawings of this application, they are only used to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" used in the description and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

Claims

1. A method for extracting information from a tender document, characterized by, The method comprises the following steps: The bidding document to be processed is divided into pages, and OCR recognition is performed on each page to obtain corresponding text; A visual language model is used to generate a supplementary text description for each image and table in the page and append it to the end of the corresponding text of the page to form an enhanced text block sequence; According to a pre-constructed hierarchical label system, a label is matched from the text block sequence, and a corresponding prompt word template is generated according to the label and a pre-constructed prompt word template library, wherein the prompt word template is designed according to the hierarchical label system so that a large language model is constrained by the hierarchical label system when generating content; The prompt word template, the enhanced text block sequence, and a context text summary are input into the large language model as a combination, wherein the context text summary is a text summary of the previous and subsequent pages of the page corresponding to the prompt word template, the large language model obtains a structured extraction result with a hierarchical relationship, and the context text summary is generated by extracting key sentences from the previous two pages and the subsequent two pages of the current page; The extracted entity content is matched with a local dictionary, and the result is aggregated and arranged after the matching to output a structured data file; The prompt word template library is a dynamic prompt word library, which continuously monitors the extraction result of a specific entity by the large language model during the execution of the method, and when a certain entity is not recognized for three times in succession or the recognition confidence is continuously lower than 0.8, an update mechanism is automatically triggered to extract the context sentences in which the entity appears and generate candidate prompt words that emphasize different expressions of the entity by using the large language model, and the candidate prompt words are included in the word library after being manually reviewed, so that the dynamic prompt word library mechanism actively adapts to new terms, new expressions, or local colloquialisms in the bidding document; Further comprising a step of verifying the consistency of images and texts, after the step of forming the enhanced text block sequence, an image feature vector of the page is extracted using a convolutional neural network, and a text feature vector of the corresponding text of the page is extracted using a natural language processing encoder, a cosine similarity between the image feature vector and the text feature vector is calculated, and when the cosine similarity is lower than a preset threshold, the page is marked as a low-confidence page for manual review; The prompt word template contains a requirement instruction, which controls the large language model to output the recognized entity in JSON format according to the hierarchical structure of the hierarchical label system, and explicitly labels the parent-child hierarchical relationship between entities in JSON.

2. The method of claim 1, wherein, The pre-constructed hierarchical label system comprises primary labels and secondary labels, the primary labels comprise technical parameter labels, qualification certificate labels, performance capability labels, and multi-modal association labels; The secondary labels under the technical parameter labels comprise equipment models, design pressures, and volumes; The secondary labels under the qualification certificate labels comprise business license numbers and ISO certification numbers; The secondary labels under the performance capability labels comprise financial reports in the past three years and type test reports; The secondary labels under the multi-modal association labels comprise drawing interface annotations and chart data values.

3. The method of claim 1, wherein, The method of matching a label from the text block sequence according to the pre-constructed hierarchical label system comprises: The semantic similarity between the content of the text block and the keywords of the labels at each level is calculated, and the label with the highest similarity is matched as the dominant label of the current text block.

4. The method of claim 1, wherein, In the step of matching the extracted entity content with the local dictionary, if the matching fails, a notification mechanism is automatically triggered to remind the business personnel to check and feed back the latest information to the system for correcting or marking the extraction result as a pending state.

5. The method of claim 1, wherein, The method for aggregating and sorting the results comprises: unifying the numerical values into international standard units and converting the dates into a unified format; calculating the semantic vector similarity of the entity expressions using the Sentence-BERT model and normalizing the synonyms with a similarity higher than 0.9 into standard terms; judging whether to remove duplicates in combination with the specific chapter theme of the entity appearing in the document, and only removing repeated entities when the context theme is consistent.

6. The method of claim 1, wherein, The generated description is used to explain the structural relationship in the image or supplement the data relationship in the table that is not completely extracted through OCR.

Images