Intelligent bidding document quality evaluation and auxiliary generation system based on multi-dimensional detection

Abstract

The invention relates to the technical field of bidding information processing and document auditing, and discloses an intelligent bidding document quality evaluation and auxiliary generation system based on multi-dimensional detection, and the system comprises an information extraction module which is used for analyzing a bidding document file and extracting various nodes; the knowledge graph construction module is used for receiving the nodes and constructing a response evidence chain type knowledge graph; the risk assessment module is used for identifying a response path in the knowledge graph, judging a path mode and calculating a quantitative risk score; and the report generation module and the auxiliary generation module are used for generating a bidding document quality evaluation report containing specific modification suggestions according to the risk evaluation result. According to the method, a traditional manual auditing process is automated, multi-dimensional detection means such as text analysis, image detection and external credit query are comprehensively applied, quantitative evaluation of bidding response risks, accurate positioning of high-risk items and automatic generation of modification suggestions are achieved, and therefore the efficiency, depth and accuracy of bidding document quality auditing are improved.

Description

Technical Field

[0001] This invention relates to the field of bidding information processing and document review technology, and in particular to an intelligent bidding document quality assessment and auxiliary generation system based on multi-dimensional detection. Background Technology

[0002] In current bidding activities, the quality of the tender documents is one of the key factors determining the success or failure of a bid. A high-quality tender document needs to accurately and completely respond to all the requirements of the tender document and provide true and valid supporting materials. Therefore, it is essential to conduct a rigorous quality review of the tender documents before the bid deadline.

[0003] However, the existing tender document quality review work mainly faces the following challenges: the quality assessment process of tender documents relies heavily on manual review. Reviewers need to spend a lot of energy reading and comparing the massive amount of content in the tender documents and bid documents word by word, and manually cross-referencing and analyzing the summary. The traditional method is not only extremely time-consuming and labor-intensive, but the review cycle is often calculated in hours or even days. Moreover, it is prone to errors due to human fatigue or negligence, resulting in low review efficiency and high labor costs.

[0004] Furthermore, the potential risks in tender documents are multi-dimensional, not only reflected in the consistency of text content, but also in the authenticity of image elements such as seals and the background and credit status of related companies. Existing review methods often lack integrated composite detection capabilities. Reviewers need to use different tools or conduct multiple independent external queries to verify text, images, and credit information separately. This fragmented process and inefficient sequential operation not only increases the workload but also makes it difficult to discover deep-seated related risks, affecting the comprehensiveness and depth of the review.

[0005] Finally, the output of traditional manual review is usually a qualitative and descriptive list of opinions, lacking standardized quantitative assessment. The modification suggestions given by reviewers are often subjective and fail to be sorted by risk level. As a result, the personnel modifying the tender documents need to read the entire review report to make sense of it, and it is difficult to quickly locate the most urgent and critical issues. Due to the lack of quantitative guidance on risks and direct location of specific defects, the decision-making cycle from problem discovery to effective solution is lengthened, which seriously affects the efficiency of iterative optimization of tender document quality. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection. This system addresses the shortcomings of existing technologies that rely on manual review or only automate partial steps in tender document quality assessment, resulting in time-consuming, labor-intensive, and inefficient processes. Furthermore, assessments often focus on superficial checks such as text plagiarism and typo detection, lacking in-depth automated analysis capabilities to verify the inherent logical consistency and evidentiary validity among the tender document requirements, bid responses, and supporting materials. This leads to inconsistent review efficiency and difficulty in effectively identifying implicit performance risks caused by broken evidence chains or weak probative value.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The information extraction module is configured to parse the tender documents and bid documents to be analyzed, and extract demand nodes, response nodes, entity nodes, evidence nodes and seal entity nodes from the tender documents and bid documents;

[0009] The knowledge graph construction module is configured to receive various types of nodes extracted by the information extraction module and construct a response evidence chain knowledge graph containing the various types of nodes according to a preset relationship. The preset relationship includes at least the response relationship between the response node and the demand node, and the proof relationship between the response node and the evidence node.

[0010] The risk assessment module is configured to identify one or more response paths originating from a specific demand node in a response-evidence chain knowledge graph constructed by the knowledge graph construction module, and calculate a risk score based on the path pattern of the identified response paths.

[0011] The report generation module is configured to receive the risk score calculated by the risk assessment module and generate a tender quality assessment report based on the risk score.

[0012] In one specific embodiment, the information extraction module includes:

[0013] A named entity recognition engine is configured to identify organizations or individuals from the text content of the tender documents and bid documents, and output the identified organizations or individuals as entity nodes.

[0014] A seal detection engine is configured to detect seal images in the tender documents and bid documents, and to identify text information within the seal images in order to generate a seal entity node.

[0015] The credit detection engine is configured to query credit information associated with entity nodes through an external interface and append the queried credit information as an attribute of the corresponding entity node.

[0016] Preferably, the demand node, response node, entity node, evidence node, and seal entity node are defined as follows:

[0017] A requirement node represents a specific clause in the tender document that the bidder is required to respond to.

[0018] A response node represents the descriptive or commitment content provided in the tender document in response to the requirement node.

[0019] Entity nodes represent objective entities such as organizations, personnel, or specific projects involved in bidding activities;

[0020] Evidence nodes represent qualification certificates, contracts, or other supporting documents used to support the content of the response nodes;

[0021] A seal entity node represents a seal object identified from the document that contains image and text information.

[0022] Preferably, the path patterns identified by the risk assessment module include at least the following four types:

[0023] A direct proof path, wherein the response path is composed of a demand node, a response node and an evidence node that directly proves the response node, connected in sequence, to represent a low-risk and formally complete response structure;

[0024] Subcontracting dependency path, the response path contains one or more entity nodes that are not the bidding entity itself, in order to identify the dependency relationship of the bidding proposal on third-party entities;

[0025] A weak proof path, wherein the endpoint of the response path is an evidence node with weak probative force, used to identify response items with insufficient supporting material;

[0026] A broken chain of evidence path is a response path that fails to connect to any valid evidence node after starting from the response node, used to identify invalid responses that are completely lacking in supporting evidence.

[0027] In one specific embodiment, the risk assessment module is configured to calculate the risk score, which is based on the following two quantitative indicators:

[0028] The path pattern base risk weight, determined by the path pattern of the response path, is used to quantify the structural risk of the response path.

[0029] And a node defect penalty factor, determined by the attributes of nodes on the response path, used to quantify the content defects of the response path.

[0030] Specifically, for a given demand node and its corresponding response path Its risk score The calculation is performed using the following formula:

[0031] ;

[0032] in, This represents a response path identified in the knowledge graph for a specific requirement node. Indicates the first indivual; Represents the response path The final risk score is a non-negative real number; Represents a path used to determine the response path. The function of the corresponding mode category outputs one of the four types mentioned above: direct proof path, sub-package dependency path, weak proof path, or evidence chain break path. This represents a function that maps path patterns to their underlying risk weights. The underlying risk weight for each path pattern is a pre-defined non-negative value used to quantify the inherent structural risk of different path patterns. The representative constitutes the response path A node; Summation operation, its subscript This is a conditional expression used to limit the summation range to the response path. Each node included ; Represents a node used for computation A function of the defect penalty factor. The node defect penalty factor is a non-negative numerical value used to quantify the node. Content defects inherent in its own attributes; Each item, through path pattern determination, establishes a basic risk value, which reflects the structural risk of the response path. For example, the basic risk weight corresponding to a broken chain of evidence path is higher than the basic risk weight corresponding to a direct proof path; The total number of content defects in the path is obtained by summing the defect penalty factors of all nodes on the path. Each item constitutes a risk amplification factor. When there are no node defects on the path, this factor is 1, and it increases as the number of node defects on the path increases. The entire formula multiplies the structural risk (basic risk weight) of the path by the content defects (risk amplification factor) to obtain a comprehensive risk score. .

[0033] Furthermore, the risk assessment module calculates the node defect penalty factor. This is achieved by evaluating the nodes on the response path. This is accomplished by the attributes of the node. The attributes of the node originate from the output of the information extraction module. In a specific embodiment, the node defect penalty factor... The calculation is performed using the following formula:

[0034] ;

[0035] in, This represents a set of preset non-negative weighting coefficients used to adjust the degree of influence of different types of defects; Represents a credit defect function, when node When the node is an entity and its credit information attributes contain negative records, this function outputs a positive value; otherwise, it outputs 0. This represents a content consistency defect function, when a node... If a node is an evidence node and the entity information it contains is inconsistent with the entity information promised by the associated response node, this function outputs a positive value; otherwise, it outputs 0. This represents a text error defect function, and its output is displayed in the node. The number of text errors detected in the text content is a non-negative integer; Represents a seal anomaly defect function, when node This function outputs a positive value when the evidence node is in an abnormal state and its associated seal information is in an abnormal state; otherwise, it outputs 0.

[0036] The formula quantifies the node's comprehensive content-related defects by weighted summation of multiple defects (credit, consistency, text, and seal) present in a node, i.e., the node defect penalty factor. .

[0037] Preferably, the system further includes:

[0038] The auxiliary generation module is configured to automatically generate targeted suggestions for modifying the tender document content when the risk assessment module identifies a specific high-risk path pattern.

[0039] In one specific embodiment, the modification suggestions in the auxiliary generation module specifically include:

[0040] When the risk assessment module identifies the response path as a broken chain of evidence, it generates a suggestion to supplement the corresponding supporting documents.

[0041] When the risk assessment module identifies the response path as a weak proof path, it generates a suggestion to replace the weak proof file with a strong proof file.

[0042] Preferably, the tender quality assessment report generated by the report generation module specifically includes the following contents:

[0043] A list of response items sorted by risk score;

[0044] The analysis also includes a detailed analysis of the response items that rank high in risk score in the response item list. The detailed analysis indicates the path pattern corresponding to each such response item and points out the specific node defects that lead to the increase in risk score.

[0045] Preferably, in addition to establishing response-to-relationships and proof-to-relationships, the knowledge graph construction module is also configured to establish affiliation relationships between entity nodes and stamp-to-relationships between seal entity nodes and evidence nodes, in order to support the risk assessment module in identifying more complex response paths.

[0046] In summary, the present invention has at least one of the following beneficial technical effects:

[0047] 1. This invention integrates multiple steps that previously required manual reading, comparison, analysis, and summarization into an automated data processing workflow through the coordinated operation of an information extraction module, a knowledge graph construction module, a risk assessment module, and a report generation module. This workflow begins with the parsing of the original tender documents and continues until a structured evaluation report is generated, all without human intervention. This reduces the manual review time of hours or even days to minutes, significantly decreasing labor costs and time consumption.

[0048] 2. This invention integrates multiple engines, including named entity recognition, seal detection, and external credit query, within its information extraction module. This enables simultaneous, automated detection of text content, image information, and related entity backgrounds in a single process. This design avoids the inefficient operation of switching between different tools or performing multiple independent queries required in traditional work, achieving a one-time, parallel scan of potential risk points across different dimensions, significantly improving the efficiency of problem detection.

[0049] 3. This invention automatically calculates the quantitative risk score for each response item through a risk assessment module. The report generation module then automatically generates a sorted list based on the scores, clearly indicating the path patterns and defect nodes corresponding to high-risk items. This approach eliminates the need for reviewers to read the entire document for secondary analysis, allowing them to quickly pinpoint the most critical issues and directly obtain specific, actionable modification guidelines. This shortens the decision-making time from problem discovery to resolution and improves the efficiency of bid quality optimization. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the system flow of the present invention;

[0051] Figure 2 This is a schematic diagram of the internal structure and data flow of the information extraction module of the present invention;

[0052] Figure 3 This is a schematic diagram illustrating the working principle of the knowledge graph construction module of the present invention;

[0053] Figure 4 This is a schematic diagram illustrating the functional division and workflow of the risk assessment module of the present invention;

[0054] Figure 5 This is a schematic diagram illustrating the collaborative operation of the report generation module and the auxiliary generation module of the present invention. Detailed Implementation

[0055] The following is in conjunction with the appendix Figure 1 -Appendix Figure 5 The present invention will be further described in detail below.

[0056] This invention provides an intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection.

[0057] See attached document Figure 1 This invention provides an intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection, including: an information extraction module, a knowledge graph construction module, a risk assessment module, and a report generation module, as well as an auxiliary generation module.

[0058] The operating environment of the system can be a computing device, which includes at least one processor and a memory. The memory stores computer program instructions, and the processor executes the computer program instructions to realize the functions of the system of the present invention. The computing device can be a physical server, a virtual machine, or a containerized instance deployed on a cloud platform. Its operating system can be Linux or Windows Server. It receives electronic documents of tender documents and bid documents to be processed through a file interface, such as PDF or DOCX format files. After processing, it outputs an evaluation report through a data interface.

[0059] In the workflow, the information extraction module first receives and parses the input tender documents and bid documents, extracts structured node data from them, and outputs the node data to the knowledge graph construction module. The knowledge graph construction module processes the received node data according to preset association rules, establishes the connection relationship between nodes, and thus constructs a response evidence chain knowledge graph.

[0060] The response evidence chain knowledge graph is then transmitted to the risk assessment module. The risk assessment module performs a graph traversal algorithm in the graph to identify the response path from demand to evidence, and performs pattern classification and risk quantification on the identified path. The risk assessment module outputs the calculated quantified risk results, along with the corresponding path information and node information, to the report generation module. The report generation module generates a structured tender quality assessment report based on the received data.

[0061] In one embodiment, the information extraction module is configured to extract five types of nodes: demand nodes, response nodes, entity nodes, evidence nodes, and seal entity nodes. The knowledge graph construction module then establishes various connections based on these nodes, including response relationships and proof relationships.

[0062] The risk assessment module calculates a specific response path using the following formula. Risk score :

[0063] ;

[0064] in, This represents a response path identified in the knowledge graph for a specific requirement node. Indicates the first indivual; Represents the response path The final risk score is a non-negative real number; Represents a path used to determine the response path. The function of the corresponding mode category outputs one of the four types mentioned above: direct proof path, sub-package dependency path, weak proof path, or evidence chain break path. This represents a function that maps path patterns to their underlying risk weights. The underlying risk weight for each path pattern is a pre-defined non-negative value used to quantify the inherent structural risk of different path patterns. The representative constitutes the response path A node; Summation operation, its subscript This is a conditional expression used to limit the summation range to the response path. Each node included ; Represents a node used for computation A function of the defect penalty factor, wherein the node defect penalty factor is a non-negative numerical value used to quantify the node. Content defects inherent in its own attributes; Each item, through path pattern determination, establishes a basic risk value, which reflects the structural risk of the response path. For example, the basic risk weight corresponding to a broken chain of evidence path is higher than the basic risk weight corresponding to a direct proof path; The total number of content defects in the path is obtained by summing the defect penalty factors of all nodes on the path. Each item constitutes a risk amplification factor. When there are no node defects on the path, this factor is 1, and it increases as the number of node defects on the path increases. The entire formula multiplies the structural risk (basic risk weight) of the path by the content defects (risk amplification factor) to obtain a comprehensive risk score. .

[0065] The node defect penalty factor The calculation is performed using the following formula:

[0066] ;

[0067] in, This represents a set of preset non-negative weighting coefficients used to adjust the degree of influence of different types of defects; Represents a credit defect function, when node When the node is an entity and its credit information attributes contain negative records, this function outputs a positive value; otherwise, it outputs 0. This represents a content consistency defect function, when a node... If a node is an evidence node and the entity information it contains is inconsistent with the entity information promised by the associated response node, this function outputs a positive value; otherwise, it outputs 0. This represents a text error defect function, and its output is displayed in the node. The number of text errors detected in the text content is a non-negative integer; Represents a seal anomaly defect function, when node This function outputs a positive value when the evidence node is in an abnormal state and its associated seal information is in an abnormal state; otherwise, it outputs 0.

[0068] See attached document Figure 2 , attached Figure 2 This is a schematic diagram of the internal structure and data flow of the information extraction module in one embodiment of the present invention.

[0069] The information extraction module is responsible for receiving electronic documents of tender documents and bid documents, and automatically extracting various nodes required to build a knowledge graph. After receiving a file in a format such as PDF or DOCX, the information extraction module first performs a preprocessing step, converting the file content into a sequence of plain text and a collection of image objects to facilitate subsequent processing by the engine. For the extraction of demand nodes and response nodes, natural language processing technology is used to analyze the preprocessed text sequence.

[0070] In one embodiment, a text classification model finely tuned to a corpus in the bidding and tendering domain is used to semantically classify paragraphs or sentences in a document, automatically identifying which belong to bidding requirements and which belong to bid responses.

[0071] In another embodiment, extraction is performed by matching a preset keyword template (e.g., sentences containing "the bidder must" or "requirements" are identified as requirements, and sentences containing "our commitment" or "specific responses are as follows" are identified as responses). Each identified requirement or response is created as an independent node, and its position in the original text is recorded.

[0072] For entity node extraction, the information extraction module integrates a named entity recognition engine. The named entity recognition engine processes all text content of the tender documents and bid documents. It can use a sequence labeling model based on bidirectional long short-term memory network and conditional random field (Bi-LSTM-CRF). The sequence labeling model is trained on a large amount of labeled data and can identify and extract predefined categories of entities from the text, such as organization name, personnel name, project name, date, amount, etc. Each identified entity is created as an entity node and is attached with its type and text content attributes.

[0073] For the creation of evidence nodes, the logical structure and content layout of the document are analyzed. When a response node is identified that mentions specific supporting materials (for example, the text says "See Appendix 3: XX Project Performance Contract"), the corresponding attachment (for example, the page image or content of Appendix 3) will be marked as an evidence node. The evidence node contains a reference to its original file location and establishes a preliminary relationship with the response node that mentions it.

[0074] For the extraction of seal entity nodes, the information extraction module also integrates a seal detection engine. First, all image objects in the document are processed using an object detection model (e.g., YOLOv5 model). This object detection model has been trained on a dataset containing various seal images and can identify the seal region in the image and output the bounding box coordinates of its position. Next, the system crops and preprocesses the seal image within the bounding box (e.g., image enhancement, affine transformation). Then, an optical character recognition (OCR) engine optimized for seal characters is applied to extract the text information in the seal image. Finally, the system generates a seal entity node, which contains the identified seal text and the physical location attribute of the seal in the document.

[0075] The information extraction module further includes a credit detection engine for obtaining credit information of entity nodes. The credit detection engine receives entity nodes of organizational type output by the named entity recognition engine. For each such entity node, the credit detection engine sends an HTTP query request to an external, public, or authorized enterprise credit information database through a pre-configured application programming interface (API). The data body of the HTTP query request contains the accurate name of the organization. The engine receives the structured data (e.g., JSON format) returned by the API, parses the credit information such as administrative penalties, legal proceedings, and records of dishonesty of the enterprise, and organizes this queried credit information into a set of attributes attached to the corresponding entity node.

[0076] After completing all the above extraction tasks, the information extraction module organizes all the generated nodes (requirement nodes, response nodes, entity nodes, evidence nodes, and seal entity nodes) and their respective attributes into a unified structured data format, such as a data stream containing multiple JSON objects. Each JSON object represents a node, and its type, content, source, attributes, and location information in the source file are recorded in detail. Finally, the information extraction module transmits this structured node data stream to the knowledge graph construction module for subsequent processing.

[0077] See attached document Figure 3 , attached Figure 3 This is a schematic diagram illustrating the working principle of the knowledge graph construction module in one embodiment of the present invention.

[0078] The knowledge graph construction module receives a structured data stream containing various types of nodes output by the information extraction module. The core function of the knowledge graph construction module is to establish semantic connections, i.e. edges, between these discrete nodes based on a set of preset association rules and algorithms, thereby forming a unified and structured response evidence chain knowledge graph. The response evidence chain knowledge graph can be stored in a graph database (e.g., Neo4j).

[0079] To establish a response-to-request relationship between response nodes and request nodes, the knowledge graph construction module can use a text semantic similarity calculation method. Specifically, for the text content of each request node and each response node, the knowledge graph construction module calculates its sentence vector representation using a pre-trained language model (e.g., Sentence-BERT). Then, by calculating the cosine similarity between the sentence vectors of different request-response node pairs, the response node with the highest similarity score (e.g., 0.85) is established as a response-to-request relationship with the corresponding request node.

[0080] The calculation formula is as follows:

[0081] ;

[0082] in, The sentence vector is obtained after the text content of the demand node is calculated by a pre-trained language model. The sentence vector is obtained after the text content of the response node is calculated by a pre-trained language model. Describe a function used to compute vectors. sum vector The similarity score between them, the output value range is The closer the value is to 1, the closer the two vectors are in direction, meaning the more similar the corresponding text semantics are. The dot product of vectors is calculated by multiplying corresponding elements of the two vectors one by one, and then summing all the products. The dot product measures how well two vectors are aligned in a direction; This represents the length of a vector in multidimensional space. It is calculated by squaring each element of the vector, summing the results, and then taking the square root. Norms are used to normalize the length of vectors to eliminate the influence of text length on similarity calculations; The summation symbol indicates that the expression following i=1 to n is summed.

[0083] To establish a proof relationship between a response node and an evidence node, the knowledge graph construction module executes an entity matching process. First, it extracts the entities mentioned in the text of the response node (e.g., project name, certificate number). At the same time, it also extracts entities from the text content or OCR results of the corresponding evidence node (e.g., scanned contract or certificate document). When the entity declared in the response node is consistent with the entity contained in the evidence node, the knowledge graph construction module establishes a proof relationship between the two nodes.

[0084] To establish the stamp-and-evidence relationship between the seal entity node and the evidence node, the knowledge graph construction module utilizes the node's positional attributes. The information extraction module has provided the bounding box coordinates of the seal entity node on the document page, as well as the page range corresponding to the evidence node. The knowledge graph construction module uses geometric calculations to determine whether the bounding box of the seal entity node falls within the page range of a certain evidence node. If the determination result is yes, a stamp-and-evidence relationship is established between the seal entity node and the corresponding evidence node.

[0085] To establish affiliation relationships between entity nodes, the knowledge graph construction module can apply a relation extraction model. The relation extraction model analyzes the context sentences in which entity nodes appear and identifies lexical patterns that indicate affiliation relationships (e.g., employee B of company A, project C is constructed by company D). When such patterns are identified, affiliation relationships are established between the corresponding entity nodes (e.g., entity B and entity A, entity C and entity D).

[0086] After all the relationships mentioned above have been established, a response evidence chain knowledge graph stored in the form of triples (head node, relation, tail node) is constructed. This response evidence chain knowledge graph, as a whole data structure, is transmitted to the risk assessment module for subsequent path analysis and risk calculation.

[0087] See attached document Figure 4 , attached Figure 4 This is a schematic diagram illustrating the functional division and workflow of the risk assessment module in one embodiment of the present invention.

[0088] The risk assessment module receives the response evidence chain knowledge graph transmitted by the knowledge graph construction module. The function of the risk assessment module is to analyze the graph structure, identify and evaluate the response path starting from each demand node, and finally output a quantitative risk score for each response item.

[0089] In one embodiment, the risk assessment module first traverses all the demand nodes in the knowledge graph. For each demand node, the risk assessment module starts from that demand node and uses graph traversal algorithms such as depth-first search (DFS) or breadth-first search (BFS) to search for paths along the edges formed by relationships such as response and proof, until it reaches a terminal node with no out-degree (such as an evidence node) or the path cannot be extended further. In this way, the risk assessment module can identify one or more complete response paths associated with a demand node.

[0090] After identifying the response path, the risk assessment module performs pattern determination on each path based on a set of deterministic logical rules. In a specific embodiment, these rules are as follows:

[0091] If a path is structured such that it starts from a demand node, reaches a response node, and then reaches an evidence node, and the preset proof strength level attribute of the evidence node is strong, then the path is determined to be a direct proof path.

[0092] If a path starts from a response node but fails to connect to any valid evidence node, the path is considered a broken evidence chain path.

[0093] If the endpoint of a path is an evidence node, but the proof strength level attribute of that node is preset to weak (e.g., only a letter of commitment rather than third-party proof), then the path is determined to be a weak proof path.

[0094] If a path contains one or more entity nodes that are not the bidding entity itself (determined by comparing the node entity name with the bidder name stored in the system), then the path is determined to be a subcontracting dependency path.

[0095] After determining the path pattern, the risk assessment module begins to quantify the risk score. Specifically, for a given response path... Its risk score The calculation is performed using the following formula:

[0096] ;

[0097] in, This represents a response path identified in the knowledge graph for a specific requirement node. Indicates the first indivual; Represents the response path The final risk score is a non-negative real number; Represents a path used to determine the response path. The function of the corresponding mode category outputs one of the four types mentioned above: direct proof path, sub-package dependency path, weak proof path, or evidence chain break path. This represents a function that maps path patterns to their underlying risk weights. The underlying risk weight for each path pattern is a pre-defined non-negative value used to quantify the inherent structural risk of different path patterns. The representative constitutes the response path A node; Summation operation, its subscript This is a conditional expression used to limit the summation range to the response path. Each node included ; Represents a node used for computation A function of the defect penalty factor. The node defect penalty factor is a non-negative numerical value used to quantify the node. Content defects inherent in its own attributes; Each item, through path pattern determination, establishes a basic risk value, which reflects the structural risk of the response path. For example, the basic risk weight corresponding to a broken chain of evidence path is higher than the basic risk weight corresponding to a direct proof path; The total number of content defects in the path is obtained by summing the defect penalty factors of all nodes on the path. Each item constitutes a risk amplification factor. When there are no node defects on the path, this factor is 1, and it increases as the number of node defects on the path increases. The entire formula multiplies the structural risk (basic risk weight) of the path by the content defects (risk amplification factor) to obtain a comprehensive risk score. .

[0098] in, It is a function that maps path patterns to their underlying risk weights. In the system, it is implemented as a pre-defined key-value lookup table. For example, the lookup table maps a broken chain of evidence path pattern to a higher value (e.g., 50), a weak proof path to a medium value (e.g., 20), and a direct proof path to a lower value (e.g., 5).

[0099] The node defect penalty factor The calculation is performed using the following formula:

[0100] ;

[0101] in, This represents a set of preset non-negative weighting coefficients used to adjust the degree of influence of different types of defects; Represents a credit defect function, when node When the node is an entity and its credit information attributes contain negative records, this function outputs a positive value; otherwise, it outputs 0. This represents a content consistency defect function, when a node... If a node is an evidence node and the entity information it contains is inconsistent with the entity information promised by the associated response node, this function outputs a positive value; otherwise, it outputs 0. This represents a text error defect function, and its output is displayed in the node. The number of text errors detected in the text content is a non-negative integer; Represents a seal anomaly defect function, when node If the node is an evidence node and its associated seal information is in an abnormal state, this function outputs a positive value (e.g., 1); otherwise, it outputs 0.

[0102] After calculating the final risk score for each response path, the risk assessment module packages the score, along with its corresponding path pattern, information on all nodes on the path, and specific node defects, into a structured data object and sends it to the report generation module.

[0103] See attached document Figure 5 , attached Figure 5 This is a schematic diagram illustrating the collaborative operation of the report generation module and the auxiliary generation module in one embodiment of the present invention.

[0104] The report generation module receives a set of structured data objects output by the risk assessment module. Each data object corresponds to a response path and its complete risk assessment result. The function of this report generation module is to generate a structured tender quality assessment report based on these data.

[0105] As attached Figure 5 As shown, in the specific implementation, the report generation module includes a data sorter and a report assembler. The data sorter first sorts all received response path data objects according to their risk scores. The values ​​are sorted in descending order. Then, the report assembler calls a preset report template (e.g., an HTML or PDF template file) and fills the sorted data into the specified area of ​​the template. The report template defines the layout and content structure of the output report. For example, the report includes an overall risk summary section and a detailed risk analysis section.

[0106] The overall risk summary section displays all evaluated response items in list form, each labeled with its corresponding final risk score. The detailed risk analysis section provides complete analytical information for response items with risk scores higher than a preset threshold, including:

[0107] The path pattern determined for this response item, the specific node defect type that contributes the most to the risk score, and the source location of the original text or evidence.

[0108] In one embodiment, an auxiliary generation module is also included, the input of which is connected to the output of the risk assessment module, and the output of which is connected to the report generation module.

[0109] As attached Figure 5 As shown, the auxiliary generation module contains a rule mapping library, which uses path pattern or node defect type as the query key and pre-written suggestion text as the value, and stores them as key-value pairs.

[0110] When the risk assessment module outputs a data object with a risk score higher than a preset threshold, the auxiliary generation module is triggered. The auxiliary generation module takes the path pattern or specific defect type of the high-risk data object as input and queries the rule mapping library.

[0111] For example: when the input path pattern is a broken chain of evidence, retrieve and output the corresponding suggested text from the mapping library: suggest supplementing the evidence documents to support this response.

[0112] When a node's defect type is "Seal Abnormality Defect," the output suggestion text is: "It is recommended to verify and use a valid seal consistent with the bidding entity."

[0113] These suggested texts generated by the auxiliary generation module are sent to the report assembler of the report generation module. The report assembler inserts the received suggested texts into the tender quality assessment report and displays them side by side with the detailed analysis of the corresponding high-risk items, thereby completing the generation of the final report.

[0114] To further clarify the technical solution and workflow of the system described in this invention, the following will be explained in conjunction with two specific application scenarios.

[0115] Scenario 1: Assessment of Low-Risk Response Items

[0116] In a specific application scenario, a clause in the tender document states: "The bidder must provide a valid ISO9001 quality management system certification." The corresponding response in the tender document is: "We have passed ISO9001 quality management system certification. The certificate is detailed in Appendix 5, which includes a valid scanned copy of the ISO9001 certificate belonging to the bidding entity."

[0117] The system receives and processes the above documents. The information extraction module extracts the text "Bidders must provide a valid ISO9001 quality management system certification" from the bidding documents and creates a requirement node. Subsequently, it extracts the response text "We have passed ISO9001 quality management system certification. The certificate is detailed in Appendix 5" from the bidding documents and creates a response node. At the same time, it processes the scanned copy of the certificate in Appendix 5 into an evidence node.

[0118] The knowledge graph construction module receives these three nodes, establishes a response relationship between the response node and the demand node through semantic similarity calculation, and establishes a proof relationship between the response node and the evidence node through entity matching. Thus, a response path is constructed that starts from the demand node, passes through the response nodes in sequence, and finally reaches the evidence node.

[0119] The risk assessment module receives and identifies this response path. Based on logical rules, the path structure conforms to the pattern of starting from the demand node, reaching the response node, and then reaching the evidence node, and the preset probative value level of the evidence node is strong. Therefore, this path pattern is determined to be a direct proof path.

[0120] When calculating the risk score, the basic risk weight corresponding to the direct proof path is obtained from a pre-defined lookup table, for example... After inspecting the attributes of all nodes along the path, no text errors, credit defects, or seal anomalies were found; therefore, the defect penalty factor for all nodes was set to zero. All values ​​are 0, and the final calculated risk score is: The report generation module will then mark this response item as low risk in the report.

[0121] Scenario 2: Assessment of High-Risk Response Items

[0122] In another specific application scenario, a clause in the tender document stated that bidders must provide at least one similar project performance contract with a contract amount of no less than RMB 5 million within the past three years. The corresponding response in the tender document stated that the bidder completed a smart city management platform project in 2023 with a contract amount of RMB 6 million, which met the requirement. However, the contract document for this project was not provided in any of the attachments to the tender document.

[0123] The system receives and processes the above files. The information extraction module creates a requirement node containing the requirements of the tender document and a response node containing the response text of the tender document. Since no supporting materials corresponding to the response item were found in the file, the system does not create a related evidence node.

[0124] The knowledge graph construction module receives these two nodes and establishes a response relationship between them. The response path thus constructed starts from the demand node, terminates after reaching the response node, and fails to connect to any evidence node.

[0125] The risk assessment module identified this response path and determined through logical rules that the path failed to connect to any evidence node after starting from the response node, thus classifying the path pattern as a broken evidence chain path.

[0126] When calculating the risk score, the high base risk weights corresponding to the broken evidence chain paths are retrieved from a pre-defined lookup table, for example... Assuming that there are no other content defects on the path, and the sum of the defect penalty factors is 0, the final calculated risk score is: .

[0127] The report generation module receives the high-risk score and the pattern determination result of the broken evidence chain path. At the same time, the auxiliary generation module is triggered. Based on the pattern of the broken evidence chain path, it retrieves the corresponding suggested text from the rule mapping library. Finally, in the output report, the report generation module marks this response item as high-risk and attaches the modification suggestions provided by the auxiliary generation module: it is recommended to supplement the supporting supporting documents for this response.

Claims

1. A smart tender document quality assessment and auxiliary generation system based on multi-dimensional detection, characterized in that, include: The information extraction module is configured to parse the tender documents and bid documents to be analyzed, and extract the requirement nodes, response nodes, entity nodes and evidence nodes from the tender documents and bid documents; The knowledge graph construction module is configured to receive various types of nodes extracted by the information extraction module and construct a response evidence chain knowledge graph according to a preset relationship. The preset relationship includes at least the response relationship between the response node and the demand node, and the proof relationship between the response node and the evidence node. The risk assessment module is configured to identify one or more response paths originating from a specific demand node in a response evidence chain knowledge graph constructed by the knowledge graph construction module, and calculate a risk score based on the path pattern of the identified response paths. The report generation module is configured to receive the risk score calculated by the risk assessment module and generate a tender quality assessment report based on the risk score.

2. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 1, characterized in that, The information extraction module includes: A named entity recognition engine is configured to identify organizations or individuals from the text content of the tender documents and bid documents, and output the identified organizations or individuals as entity nodes. A seal detection engine is configured to detect seal images in the tender documents and bid documents, identify text information within the seal images, and then generate a seal entity node. The credit detection engine is configured to query credit information associated with entity nodes through an external interface and append the queried credit information as an attribute of the corresponding entity node.

3. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 1, characterized in that, The demand node, response node, entity node, evidence node, and seal entity node are defined as follows: A requirement node represents a specific clause in the tender document that the bidder is required to respond to. A response node represents the descriptive or commitment content provided in the tender document in response to the requirement node. Entity nodes represent objective entities such as organizations, personnel, or specific projects involved in bidding activities; Evidence nodes represent qualification certificates, contracts, or other supporting documents used to support the content of the response nodes; A seal entity node represents a seal object identified from the document that contains image and text information.

4. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 1, characterized in that, The path patterns identified by the risk assessment module include at least the following four types: A direct proof path, wherein the response path is composed of a demand node, a response node and an evidence node that directly proves the response node, connected in sequence, to represent a low-risk and formally complete response structure; Subcontracting dependency path, the response path contains one or more entity nodes that are not the bidding entity itself, in order to identify the dependency relationship of the bidding proposal on third-party entities; A weak proof path, wherein the endpoint of the response path is an evidence node with weak probative force, used to identify response items with insufficient supporting material; A broken chain of evidence path is a response path that fails to connect to any valid evidence node after starting from the response node, used to identify invalid responses that are completely lacking in supporting evidence.

5. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 4, characterized in that, The risk assessment module is configured to calculate the risk score, which is based on the following two quantitative indicators: The path pattern base risk weight, determined by the path pattern of the response path, is used to quantify the structural risk of the response path. And a node defect penalty factor, determined by the attributes of nodes on the response path, used to quantify the content defects of the response path.

6. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 5, characterized in that, The risk assessment module calculates the node defect penalty factor by evaluating the attributes of the nodes on the response path. The attributes of the node are derived from the output of the information extraction module, and the calculation of the node defect penalty factor is based on at least one or more of the following attributes: Credit information attributes of entity nodes; The consistency between the entity information within the evidence node and the entity information promised by the response node; Text errors exist in the text content of response nodes or entity nodes; Abnormal status of seal information associated with evidence nodes.

7. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 4, characterized in that, Also includes: The auxiliary generation module is used to automatically generate targeted suggestions for modifying the tender document content when the risk assessment module identifies a specific high-risk path pattern.

8. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 7, characterized in that, The modification suggestions in the auxiliary generation module specifically include: When the risk assessment module identifies the response path as a broken chain of evidence, it generates a suggestion to supplement the corresponding supporting documents. When the risk assessment module identifies the response path as a weak proof path, it generates a suggestion to replace the weak proof file with a strong proof file.

9. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 1, characterized in that, The tender quality assessment report generated by the report generation module specifically includes: A list of response items sorted by risk score; The analysis also includes a detailed analysis of the response items that rank high in risk score in the response item list. The detailed analysis indicates the path pattern corresponding to each such response item and points out the specific node defects that lead to the increase in risk score.

10. The intelligent tender document quality assessment and auxiliary generation system based on multi-dimensional detection according to claim 1, characterized in that, In addition to establishing response and proof relationships, the knowledge graph construction module is also configured to establish affiliation relationships between entity nodes and stamp relationships between seal entity nodes and evidence nodes, in order to support the risk assessment module in identifying more complex response paths.

Images