Intelligent bidding and tendering information identification method

Abstract

The invention discloses an intelligent bidding and tendering information identification method, which relates to the technical field of engineering management, and comprises the following steps: standardizing and preprocessing a bidding and tendering information document format; constructing a natural language processing model special for the bidding and tendering field; training a multi-level information extraction model based on deep learning; carrying out cross-document information integration and correlation analysis; performing anomaly detection and verification based on rules; semantic understanding and sentiment analysis; automatic subject matter description and technical specification extraction; a real-time dynamic monitoring and information updating mechanism; the invention discloses a graph-based decision support and recommendation system. According to the intelligent bidding and tendering information identification method, efficient and intelligent identification of bidding and tendering information is realized by constructing a natural language processing model specially aiming at the field of bidding and tendering. The bidirectional coding capability of BERT and the sequence modeling capability of LTM are combined, and a hierarchical attention mechanism is introduced, so that the model can accurately identify various key information in bidding and tendering documents.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering management, and in particular to an intelligent bidding information recognition method. Background Art

[0002] Information management and analysis in the bidding process usually involves a large amount of documents and data, which often come from multiple sources such as different bidding platforms, contract documents, bids, technical and financial statements, etc. Since bidding documents usually have complex structures and contents, and contain a large amount of industry terms, legal terms, technical requirements and financial data, traditional manual processing and traditional information extraction methods often face great challenges. Existing technologies mostly rely on rule-based or template-based methods, which usually require manual definition of a large number of rules and are difficult to cope with the ever-changing content and format of bidding documents.

[0003] Although existing natural language processing (NLP) technologies, such as BERT, GPT and other pre-trained models, can process general text data, in the field of bidding, the terms and expressions involved in the documents are highly professional. It is difficult for general models to effectively identify industry-specific entities and relationships. Existing named entity recognition (NER) models are also mostly based on general models, which cannot fully understand the complex context in bidding, and can easily lead to inaccurate entity recognition or missed recognition of important information. In addition, traditional multi-document information integration and relationship analysis often rely on simple text comparisons and cannot handle deep associations and reasoning across document information. Therefore, how to improve the accuracy and efficiency of information recognition in bidding documents through deep learning technology combined with industry knowledge and contextual information has become an important issue facing current technology. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides an intelligent bidding information identification method to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: an intelligent bidding information identification method, comprising the following steps:

[0006] S1. Standardization and preprocessing of bidding information document formats;

[0007] S2. Build a natural language processing model dedicated to the bidding field;

[0008] S3, multi-level information extraction model training based on deep learning;

[0009] S4, cross-document information integration and correlation analysis;

[0010] S5. Rule-based anomaly detection and verification;

[0011] S6, semantic understanding and sentiment analysis;

[0012] S7, automated subject matter description and technical specification extraction;

[0013] S8, real-time dynamic monitoring and information update mechanism;

[0014] S9. Graph-based decision support and recommendation system.

[0015] To further optimize the technical solution, in step S1, the bidding documents include PDF, Word, and Excel file formats, and the bidding documents are converted into a unified format, which is a readable plain text format;

[0016] Extract text data from documents through OCR optical character recognition technology or API interface, identify and remove noise content, and ensure the integrity and accuracy of text content;

[0017] At the same time, the paragraphs, tables, and titles in the document are structured for information extraction in the subsequent recognition process.

[0018] To further optimize the technical solution, in step S2, before performing text analysis, the natural language processing model is trained through a large number of bidding documents to collect industry-specific terms, keywords, and common grammatical structures, increase the ability to recognize high-frequency words and industry terms in bidding documents, and enable the model to efficiently process content related to bidding.

[0019] Further optimizing the technical solution, the natural language processing model includes a word embedding layer, a multi-layer self-attention mechanism, a domain term enhancement module, a bidirectional encoding mechanism, and a domain-specific loss function;

[0020] The word embedding layer is based on a pre-trained embedding matrix, which is optimized specifically for the bidding field. It is fine-tuned using large-scale bidding document data to capture industry terms and legal terms in specific bidding fields.

[0021] Multi-layer self-attention mechanism, which helps the model to obtain vocabulary relationships from global and local contexts when processing long documents;

[0022] The domain term enhancement module adjusts the self-attention results by introducing a domain term library to enhance the model's sensitivity to specific domain terms;

[0023] A bidirectional encoding mechanism is used to ensure that the model simultaneously considers the dependencies between the contexts in the bidding documents, including the long-distance dependencies of complex contract terms and technical specifications;

[0024] The loss function is used to ensure that the model focuses on and accurately identifies special terms in bidding documents.

[0025] To further optimize this technical solution, the domain terminology enhancement module has a built-in domain terminology library Adjust the self-attention results:

[0026]

[0027] Among them, H i is the context representation obtained through the self-attention mechanism, λ is the adjustment parameter, and T t is the term embedding, sim(e i , t) is the similarity between word embedding and term embedding. This module is used to enhance the sensitivity of the model to domain-specific terms.

[0028] To further optimize the technical solution, the natural language processing model, when used, includes the following process:

[0029] Data preprocessing and word embedding generation: The text in the bidding documents is converted into a standardized input format through OCR and preprocessing. The domain-specific word embedding generator is used to convert the text into word vectors and input them into the word embedding layer of the model.

[0030] Self-attention mechanism and term enhancement module processing: After the word embedding is generated, it is input into the multi-head self-attention mechanism to extract contextual information in the text, focusing on long-distance dependencies. The domain term enhancement module uses the industry term library to fine-tune the word vector, so that the model pays more attention to key terms in the industry and their associated contexts;

[0031] Bidirectional encoding and context integration: Through the bidirectional encoding mechanism, the context of the bidding documents is comprehensively processed to ensure the effective decoding of complex information such as contract terms and technical specifications;

[0032] Key information extraction and classification: Using the trained model, key entities are automatically identified from the document, including project name, amount, and term, and classified. Through the domain-specific loss function, the model focuses on key information to ensure high precision and high recall of the extraction results.

[0033] To further optimize the technical solution, in step S3, a multi-level named entity recognition model is constructed to automatically identify key information such as project name, time node, and amount;

[0034] The named entity recognition model includes an input embedding layer, a bidirectional LSTM layer, a hierarchical attention mechanism, a named entity recognition layer, and a loss function;

[0035] Input embedding layer: the input text is embedded by BERT, which captures the contextual relationship between words.

[0036] Bidirectional LSTM layer, after the word embedding generated by BERT, it is input into the bidirectional LSTM network for sequence modeling;

[0037] Hierarchical attention mechanism, which is used by the model to pay weighted attention to different information at different levels;

[0038] The named entity recognition layer is used to convert the output of the model into a label for each word to complete the named entity recognition task;

[0039] The loss function is used to consider the conventional cross entropy loss and introduce constraints on the semantic relevance between entities.

[0040] To further optimize the technical solution, the named entity recognition layer is a linear layer added on the basis of LSTM and attention mechanism, which converts the output of the model into the label of each word to complete the named entity recognition task;

[0041] The labels of entities include: project name, bidder, amount, date; conditional random field CRF is used as the output layer to capture the dependencies between labels and ensure the consistency between adjacent entities; the model of the named entity recognition layer is as follows:

[0042] Y i = CRF(A i );

[0043] Among them, Y i Representation word x i The CRF layer considers the constraints and dependencies between entity tags to avoid label conflicts in recognition.

[0044] To further optimize the technical solution, in step S4, through graph analysis technology, key elements of entities, time, and amount in different documents are associated to form a multi-dimensional information network, combined with entity linking and relationship extraction technology to achieve automatic comparison and integration of information between different documents.

[0045] To further optimize the technical solution, in step S5, a set of rule templates are defined to verify whether the identified information complies with conventional logic or industry standards;

[0046] If the numerical range of the budget amount is consistent with the project scale, the bid deadline cannot be later than the date of publication of the tender notice.

[0047] Further optimizing the technical solution, in step S7, the description of the subject matter utilizes a combination of image recognition and natural language processing technology to automatically parse the subject matter image, technical drawings and text description in the document;

[0048] The technical specifications are analyzed through machine learning models on technical parameters, material specifications, and performance indicators to ensure that the extracted subject matter information is accurate and compared and verified with the descriptions in other relevant documents.

[0049] To further optimize the technical solution, in step S8, a real-time dynamic monitoring and information update mechanism is designed to obtain information changes on the bidding platform in real time through crawler technology and API interface, and update and adjust the existing information; when a bidding document is revised or postponed, the system automatically identifies and notifies the relevant parties to ensure that each party can obtain the latest bidding information in a timely manner.

[0050] Compared with the prior art, the present invention provides an intelligent bidding information identification method, which has the following beneficial effects:

[0051] 1. This intelligent bidding information recognition method achieves efficient and intelligent recognition of bidding information by building a natural language processing model specifically for the bidding field. By combining the bidirectional encoding capability of BERT with the sequence modeling capability of LSTM and introducing a hierarchical attention mechanism, the model can accurately identify various key information in bidding documents. This method can automatically adapt to the structure and language characteristics of bidding documents, reduce the need for manual intervention and rule setting, and significantly improve the accuracy and efficiency of information extraction.

[0052] 2. This intelligent bidding information recognition method can realize cross-document information chain association through multi-document information integration and association analysis, help users fully understand various important information in the bidding process, and improve the intelligent level of decision support. Therefore, this method can effectively solve the problems of insufficient accuracy, poor adaptability and difficulty in information integration in traditional bidding information recognition technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of a flow chart of an intelligent bidding information identification method proposed by the present invention;

[0054] Figure 2 The composition and flow chart of the natural language processing model in the intelligent bidding information recognition method proposed by the present invention;

[0055] Figure 3 This is a schematic diagram of the composition of a named entity recognition model in an intelligent bidding information recognition method proposed in the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Embodiment 1:

[0058] See also Figure 1-Figure 3 , an intelligent bidding information identification method, comprising the following steps:

[0059] S1. Standardization and preprocessing of bidding information document formats

[0060] First, the format of the bidding documents must be standardized and preprocessed. Bidding documents often exist in multiple formats, including PDF, Word, Excel and other file formats, which may differ in structure, fonts, tables, etc. In order to facilitate subsequent automatic recognition and analysis, these files must be converted into a unified format, such as a readable plain text format. The text data in the document is extracted through OCR (optical character recognition) technology or API interface, and the noise content is identified and removed to ensure the integrity and accuracy of the text content. At the same time, the paragraphs, tables, and titles in the document are structured to facilitate the extraction of information in the subsequent recognition process.

[0061] S2. Build a natural language processing model dedicated to the bidding field

[0062] In this embodiment, before performing text analysis, a natural language processing model specifically for the bidding field must be built. Although ordinary language models can process general texts, in bidding documents, due to the involvement of a large number of industry terms, technical specifications, legal terms, etc., general language models cannot accurately understand these professional vocabulary. Through training with a large number of bidding documents, the terms, keywords and common grammatical structures of specific industries are collected, so that the model can efficiently process content related to bidding. In this process, it is possible to consider using a pre-trained model for fine-tuning, using architectures such as BERT or GPT, to increase the recognition ability of high-frequency words and industry terms in bidding documents.

[0063] S3. Multi-level information extraction model training based on deep learning

[0064] In this embodiment, the bidding documents contain a large amount of information, such as project name, bidding unit, budget amount, contract terms, bidding conditions, etc., and this information is usually distributed in different levels and locations of the document. In order to accurately identify this information, a multi-level named entity recognition model must be built to automatically identify the key information of project name, time node, and amount;

[0065] The named entity recognition model includes an input embedding layer, a bidirectional LSTM layer, a hierarchical attention mechanism, a named entity recognition layer, and a loss function;

[0066] Input embedding layer: the input text is embedded by BERT, which captures the contextual relationship between words.

[0067] Bidirectional LSTM layer, after the word embedding generated by BERT, it is input into the bidirectional LSTM network for sequence modeling;

[0068] Hierarchical attention mechanism, which is used by the model to pay weighted attention to different information at different levels;

[0069] The named entity recognition layer is used to convert the output of the model into a label for each word to complete the named entity recognition task;

[0070] The loss function is used to consider the conventional cross entropy loss and introduce constraints on the semantic relevance between entities.

[0071] S4. Cross-document information integration and correlation analysis

[0072] In this embodiment, the bidding information recognition is not limited to the content of a single document, but also involves the information integration between different documents. For example, bidding notices, bids, contract documents, etc. often need to extract information from different documents and associate them to form a complete bidding information chain. Through graph analysis technology, the key elements of entities, time, and amount in different documents are associated to form a multi-dimensional information network, combined with entity linking and relationship extraction technology to achieve automatic comparison and integration of information between different documents.

[0073] S5. Rule-based anomaly detection and verification

[0074] In this embodiment, some abnormal situations may occur during the bidding process, such as non-compliance with contract terms, abnormal budget amounts, time scheduling conflicts, etc. Therefore, in the process of information identification, a rule-based anomaly detection mechanism is designed. By defining a set of rule templates, it is verified whether the identified information conforms to conventional logic or industry standards;

[0075] If the numerical range of the budget amount is consistent with the project scale, the bid deadline cannot be later than the publication date of the tender notice. For anomalies that cannot be detected by rules, deep learning models can be combined for further anomaly identification to improve the intelligence level of the system.

[0076] S6. Semantic Understanding and Sentiment Analysis

[0077] In this embodiment, the bidding documents not only contain a large amount of digital and clause information, but also involve implicit information such as the semantic understanding of the contract, the intentions of the bidders, and the needs of the tendering party. Therefore, in addition to basic structured data extraction, semantic understanding and sentiment analysis are also required. Based on sentiment analysis and contextual reasoning technology in natural language processing, the emotional tendency and potential meaning of the language in the document are analyzed. For example, the expression of the bidder in the technical proposal may imply a certain emotion or attitude. The system can identify the actual commitment of the bidder and its potential intentions through in-depth analysis of keywords and context, so as to help reviewers make a more comprehensive judgment.

[0078] S7. Automated object description and technical specification extraction

[0079] In this example, the subject matter description and technical specifications are the core parts of the bidding documents, and accurate understanding of these contents is crucial for both bidders and reviewers. The goal is to automatically extract and standardize the subject matter description and technical requirements. The subject matter description uses a combination of image recognition and natural language processing technology to automatically parse the subject matter images, technical drawings and text descriptions in the document;

[0080] The technical specifications are analyzed through machine learning models on technical parameters, material specifications, and performance indicators to ensure that the extracted subject matter information is accurate and compared and verified with the descriptions in other relevant documents.

[0081] S8. Real-time dynamic monitoring and information update mechanism

[0082] In this embodiment, during the bidding process, information may be frequently updated, such as changes in the bidding notice, adjustments to the bid evaluation results, etc. Therefore, a real-time dynamic monitoring and information update mechanism is designed to obtain information changes on the bidding platform in real time through crawler technology and API interfaces, and update and adjust existing information; when a bidding document is revised or postponed, the system automatically identifies and notifies the relevant parties to ensure that each party can obtain the latest bidding information in a timely manner.

[0083] S9. Graph-based decision support and recommendation system

[0084] In this embodiment, the identification of bidding information is not only the extraction and organization of data, but also requires the provision of intelligent decision support for bidders and tendering parties. Therefore, a decision support and recommendation system based on a graph can be designed. The system establishes a knowledge graph in the field of bidding, associates various types of information and entities (such as bidding units, bidders, project types, historical bidding situations, etc.), and combines graph analysis technology to help users make more intelligent decisions. For example, based on past bidding data, the system can predict the probability of winning a certain type of project, or recommend the most suitable bidding plan and team configuration, reduce manual intervention, and improve decision-making efficiency.

[0085] Embodiment 2:

[0086] See also Figure 2 and Figure 3 , based on the intelligent bidding information recognition method described in Example 1, in step S2, a natural language processing model is designed.

[0087] The natural language processing model includes a word embedding layer, a multi-layer self-attention mechanism, a domain term enhancement module, a bidirectional encoding mechanism, and a domain-specific loss function;

[0088] The word embedding layer is based on a pre-trained embedding matrix, which is optimized specifically for the bidding field. It is fine-tuned using large-scale bidding document data to capture industry terms and legal terms in specific bidding fields.

[0089] Multi-layer self-attention mechanism, which helps the model to obtain vocabulary relationships from global and local contexts when processing long documents;

[0090] The domain term enhancement module adjusts the self-attention results by introducing a domain term library to enhance the model's sensitivity to specific domain terms;

[0091] A bidirectional encoding mechanism is used to ensure that the model simultaneously considers the dependencies between the contexts in the bidding documents, including the long-distance dependencies of complex contract terms and technical specifications;

[0092] The loss function is used to ensure that the model focuses on and accurately identifies special terms in bidding documents.

[0093] The word embedding layer looks like this:

[0094] E = {e1, e2, ..., e n},e i =W e ·x i

[0095] Among them, x i is the input vocabulary, W eIt is a pre-trained embedding matrix optimized for the bidding field. Different from the embedding layer of the ordinary BERT, the embedding matrix W here is e It is fine-tuned through large-scale bidding document data and is able to capture industry terminology and legal terms in specific fields.

[0096] The multi-layer self-attention mechanism is as follows:

[0097]

[0098] Among them, Q, K, V are query, key and value vectors generated based on the input embedding. is the scaling factor. The multi-head attention mechanism is used to enhance the processing capability of complex bidding document structures.

[0099] Domain terminology enhancement module, by introducing a domain term library Adjust the self-attention results:

[0100]

[0101] Among them, H i is the context representation obtained through the self-attention mechanism, λ is the adjustment parameter, and T t is the term embedding, sim(e i , t) is the similarity between word embedding and term embedding. This module enhances the model’s sensitivity to domain-specific terms.

[0102] The bidirectional encoding mechanism encodes each input sequence forward and backward to capture richer semantic relationships:

[0103] H=Transformer(E),O=Bidirectional(H)

[0104] This bidirectional encoding ensures that the model can simultaneously consider the dependencies between contexts in bidding documents, especially the long-distance dependencies of complex contract terms and technical specifications.

[0105] In order to further optimize the performance of the model in bidding scenarios, a loss function combining domain characteristics is designed. This loss function adds additional penalty terms for important industry terms and key information on the basis of conventional cross entropy loss.

[0106]

[0107] Among them, term_importance(x i ) represents the current word x i domain importance (determined by the domain vocabulary), p term (y i) is the predicted probability after adjusting the importance of the term, and α is the adjustment parameter. This loss function ensures that the model focuses on and accurately identifies special terms in bidding documents.

[0108] Furthermore, when the natural language processing model is used, it includes the following process:

[0109] Data preprocessing and word embedding generation: The text in the bidding documents is converted into a standardized input format through OCR and preprocessing. The domain-specific word embedding generator is used to convert the text into word vectors and input them into the word embedding layer of the model.

[0110] Self-attention mechanism and term enhancement module processing: After the word embedding is generated, it is input into the multi-head self-attention mechanism to extract contextual information in the text, focusing on long-distance dependencies. The domain term enhancement module uses the industry term library to fine-tune the word vector, so that the model pays more attention to key terms in the industry and their associated contexts;

[0111] Bidirectional encoding and context integration: Through the bidirectional encoding mechanism, the context of the bidding documents is comprehensively processed to ensure the effective decoding of complex information such as contract terms and technical specifications;

[0112] Key information extraction and classification: Using the trained model, key entities are automatically identified from the document, including project name, amount, and term, and classified. Through the domain-specific loss function, the model focuses on key information to ensure high precision and high recall of the extraction results.

[0113] By introducing domain terminology enhancement modules, custom loss functions and bidirectional encoding mechanisms, the recognition capability of bidding and tendering texts is effectively improved. The model can automatically understand and process complex contract terms, technical specifications and other industry-specific content, providing an efficient and intelligent solution for bidding and tendering information recognition.

[0114] In step S3, the designed named entity recognition model is designed.

[0115] The named entity recognition model includes an input embedding layer, a bidirectional LSTM layer, a hierarchical attention mechanism, a named entity recognition layer, and a loss function;

[0116] Input embedding layer: the input text is embedded by BERT, which captures the contextual relationship between words.

[0117] Bidirectional LSTM layer, after the word embedding generated by BERT, it is input into the bidirectional LSTM network for sequence modeling;

[0118] Hierarchical attention mechanism, which is used by the model to pay weighted attention to different information at different levels;

[0119] The named entity recognition layer is used to convert the output of the model into a label for each word to complete the named entity recognition task;

[0120] The loss function is used to consider the conventional cross entropy loss and introduce constraints on the semantic relevance between entities.

[0121] The input embedding layer looks like this:

[0122] E input =BERT(X)

[0123] Where X = {x1, x2, ..., x n} is the vocabulary sequence of the input text, E input It is the word vector generated by the BERT model.

[0124] The bidirectional LSTM layer can capture contextual information in the text, especially processing long-distance dependencies and avoiding missing key information. The output of the bidirectional LSTM can be expressed by the following formula:

[0125] H i =Bi-LSTM(E input [i]) for i=1,2,...,n

[0126] Among them, H i Represents the bidirectional LSTM encoding results of each word, which will provide rich contextual information for subsequent entity recognition.

[0127] Hierarchical Attention Mechanism,Since bidding documents usually contain multiple levels of key information (such as project name, budget amount, contract terms, etc.), this information is sometimes scattered in the text and their semantic levels are inconsistent. The hierarchical attention mechanism is as follows:

[0128] A i =Attention(H i , H l , ..., H n )

[0129] Among them, A i Represents the attention weight of each word. The model automatically assigns higher weights to important entities by calculating the relevance of the current word with other words. The hierarchical attention mechanism can automatically adjust the attention to key information such as project name, amount, time node, etc. according to the structure of the bidding document.

[0130] Named Entity Recognition Layer,Finally, a linear layer is added on top of the LSTM and attention mechanism to convert the output of the model into labels for each word to complete the named entity recognition task.,The labels of entities can include: project name, bidder, amount, date, etc.,The conditional random field (CRF) is used as the output layer to capture the dependencies between labels and ensure the consistency between adjacent entities.

[0131] Y i = CRF(A i )

[0132] Among them, Y i Representation word x i The CRF layer considers the constraints and dependencies between entity tags and avoids the label conflict problem in recognition.

[0133] The loss function is as follows:

[0134]

[0135] Among them, y i represents the entity label predicted by the model, p(y i ) is the predicted probability of the label, ε critical Represents a collection of "key entities" in the domain (such as project name, bidder, budget amount, etc.), is an indicator function that determines whether the current label belongs to a "key entity". γ is a tuning parameter used to enhance the model's ability to identify key entities.

[0136] When using:

[0137] During the training process, the model optimizes the loss function through the back-propagation algorithm to gradually improve the accuracy of entity recognition. A large-scale bidding document dataset is used for training, and a stratified cross-validation technique is used to ensure the generalization ability of the model. In the early stages of training, the focus is mainly on the recognition of general entities; in the later stages, the focus is gradually strengthened on bidding-specific entities (such as project names, contract terms, etc.).

[0138] In actual use, when a bidding document is input, the document is first embedded using BERT, and then the embedding result is input into the bidirectional LSTM for context modeling. Next, the key entities in the document are accurately extracted through the hierarchical attention mechanism and entity recognition layer. Finally, the model outputs the entity label of each word and structures the extracted information into formats such as tables and database entries as required.

[0139] When the model is actually applied, the following scenarios are designed:

[0140] Suppose there is a bidding document with the following content:

[0141] "Project name: XX Expressway Construction Project, Bidding unit: XX Construction Company, Budget amount: 50 million yuan, Bidding deadline: December 31, 2024."

[0142] Through the model, we can extract the following information:

[0143] Project Name: XX Expressway Construction Project

[0144] Bidding unit:XX Construction Company

[0145] Budget amount: 50 million yuan

[0146] Bid submission deadline: December 31, 2024

[0147] By combining BERT's bidirectional encoding, LSTM's sequence modeling capabilities, and hierarchical attention mechanisms, it is possible to accurately extract multi-level, multi-category information from bidding documents. Through domain-specific loss functions, the model strengthens the ability to identify key entities, ensuring high accuracy and high recall of information extraction.

[0148] The beneficial effects of the present invention are:

[0149] 1. This intelligent bidding information recognition method achieves efficient and intelligent recognition of bidding information by building a natural language processing model specifically for the bidding field. By combining the bidirectional encoding capability of BERT with the sequence modeling capability of LSTM and introducing a hierarchical attention mechanism, the model can accurately identify various key information in bidding documents. This method can automatically adapt to the structure and language characteristics of bidding documents, reduce the need for manual intervention and rule setting, and significantly improve the accuracy and efficiency of information extraction.

[0150] 2. This intelligent bidding information recognition method can realize cross-document information chain association through multi-document information integration and association analysis, help users fully understand various important information in the bidding process, and improve the intelligent level of decision support. Therefore, this method can effectively solve the problems of insufficient accuracy, poor adaptability and difficulty in information integration in traditional bidding information recognition technology.

[0151] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0152] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent bidding information identification method, characterized in that: The following steps are involved: S1. Standardization and preprocessing of bidding information document formats; S2. Build a natural language processing model dedicated to the bidding field; S3, multi-level information extraction model training based on deep learning; S4, cross-document information integration and correlation analysis; S5. Rule-based anomaly detection and verification; S6, semantic understanding and sentiment analysis; S7, automated subject matter description and technical specification extraction; S8, real-time dynamic monitoring and information update mechanism; S9. Graph-based decision support and recommendation system.

2. According to claim 1, a smart bidding information identification method is characterized in that: In the step S1, the bidding documents include PDF, Word, and Excel file formats, and the bidding documents are converted into a unified format, which is a readable plain text format; Extract text data from documents through OCR optical character recognition technology or API interface, identify and remove noise content, and ensure the integrity and accuracy of text content; At the same time, the paragraphs, tables, and titles in the document are structured for information extraction in the subsequent recognition process.

3. The intelligent bidding information identification method according to claim 1 is characterized in that: In step S2, before performing text analysis, the natural language processing model is trained through a large number of bidding documents to collect industry-specific terms, keywords, and common grammatical structures, increase the ability to recognize high-frequency words and industry terms in bidding documents, and enable the model to efficiently process content related to bidding.

4. The intelligent bidding information identification method according to claim 3 is characterized in that: The natural language processing model includes a word embedding layer, a multi-layer self-attention mechanism, a domain term enhancement module, a bidirectional encoding mechanism, and a domain-specific loss function; The word embedding layer is based on a pre-trained embedding matrix, which is optimized specifically for the bidding field. It is fine-tuned using large-scale bidding document data to capture industry terms and legal terms in specific bidding fields. Multi-layer self-attention mechanism, which helps the model to obtain vocabulary relationships from global and local contexts when processing long documents; The domain term enhancement module adjusts the self-attention results by introducing a domain term library to enhance the model's sensitivity to specific domain terms; A bidirectional encoding mechanism is used to ensure that the model simultaneously considers the dependencies between the contexts in the bidding documents, including the long-distance dependencies of complex contract terms and technical specifications; The loss function is used to ensure that the model focuses on and accurately identifies special terms in bidding documents.

5. The intelligent bidding information identification method according to claim 4 is characterized in that: The domain term enhancement module has a built-in domain term library Adjust the self-attention results: Among them, H i is the context representation obtained through the self-attention mechanism, λ is the adjustment parameter, and T t is the term embedding, sim(e i , t) is the similarity between word embedding and term embedding. This module is used to enhance the sensitivity of the model to domain-specific terms.

6. The intelligent bidding information identification method according to claim 4 is characterized in that: When the natural language processing model is used, it includes the following process: Data preprocessing and word embedding generation: The text in the bidding documents is converted into a standardized input format through OCR and preprocessing. The domain-specific word embedding generator is used to convert the text into word vectors and input them into the word embedding layer of the model. Self-attention mechanism and term enhancement module processing: After the word embedding is generated, it is input into the multi-head self-attention mechanism to extract contextual information in the text, focusing on long-distance dependencies. The domain term enhancement module uses the industry term library to fine-tune the word vector, so that the model pays more attention to key terms in the industry and their associated contexts; Bidirectional encoding and context integration: Through the bidirectional encoding mechanism, the context of the bidding documents is comprehensively processed to ensure the effective decoding of complex information such as contract terms and technical specifications; Key information extraction and classification: Using the trained model, key entities are automatically identified from the document, including project name, amount, and term, and classified. Through the domain-specific loss function, the model focuses on key information to ensure high precision and high recall of the extraction results.

7. The intelligent bidding information identification method according to claim 1 is characterized in that: In step S3, a multi-level named entity recognition model is constructed to automatically identify key information such as project name, time node, and amount; The named entity recognition model includes an input embedding layer, a bidirectional LSTM layer, a hierarchical attention mechanism, a named entity recognition layer, and a loss function; Input embedding layer: the input text is embedded by BERT, which captures the contextual relationship between words. Bidirectional LSTM layer, after the word embedding generated by BERT, it is input into the bidirectional LSTM network for sequence modeling; Hierarchical attention mechanism, which allows the model to pay weighted attention to different information at different levels; The named entity recognition layer is used to convert the output of the model into a label for each word to complete the named entity recognition task; The loss function is used to consider the conventional cross entropy loss and introduce constraints on the semantic relevance between entities.

8. The intelligent bidding information identification method according to claim 7 is characterized in that: The named entity recognition layer is a linear layer added on the basis of LSTM and attention mechanism, which converts the output of the model into the label of each word to complete the named entity recognition task; The labels of entities include: project name, bidder, amount, date; conditional random field CRF is used as the output layer to capture the dependencies between labels and ensure the consistency between adjacent entities; the model of the named entity recognition layer is as follows: Y i =CRF(A i ); Among them, Y i Representation word x i The CRF layer considers the constraints and dependencies between entity tags to avoid label conflicts in recognition.

9. The intelligent bidding information identification method according to claim 1 is characterized in that: In step S4, through graph analysis technology, key elements of entities, time, and amount in different documents are associated to form a multi-dimensional information network, combined with entity linking and relationship extraction technology to achieve automatic comparison and integration of information between different documents.

10. The intelligent bidding information identification method according to claim 1, characterized in that: In step S5, a set of rule templates are defined to check whether the identified information complies with conventional logic or industry standards; If the numerical range of the budget amount is consistent with the project scale, the bid deadline cannot be later than the date of publication of the tender notice.

11. The intelligent bidding information identification method according to claim 1, characterized in that: In step S7, the description of the subject matter is automatically parsed from the subject matter pictures, technical drawings and text descriptions in the document by combining image recognition and natural language processing technology; The technical specifications are analyzed through machine learning models on technical parameters, material specifications, and performance indicators to ensure that the extracted subject matter information is accurate and compared and verified with the descriptions in other relevant documents.

12. The intelligent bidding information identification method according to claim 1, characterized in that: In step S8, a real-time dynamic monitoring and information updating mechanism is designed to obtain information changes on the bidding platform in real time through crawler technology and API interface, and to update and adjust the existing information; When a bidding document is revised or postponed, the system automatically identifies and notifies the relevant parties to ensure that each party can obtain the latest bidding information in a timely manner.

Images