Auxiliary bidding method based on knowledge graph

Abstract

The invention discloses an auxiliary bidding method based on a knowledge graph, and belongs to the technical field of auxiliary bidding, and the method comprises the following steps: S1, capturing a bidding text through a crawler, and preprocessing the bidding text; s2, sequence labeling: constructing a BERT-BILSTM-CRF fusion model in which an attention mechanism is introduced, and extracting a named entity of the bid inviting text by using the constructed BERT-BILSTM-CRF fusion model in which the attention mechanism is introduced; s3, relation classification: constructing a BiGRU-attention relation classification model, and performing relation prediction on the named entities by utilizing the constructed BiGRU-attention relation classification model; and S4, according to a relationship classification result, combining the named entities and the relationships to obtain effective entities, and storing the effective entities in a database. By adopting the auxiliary bidding method based on the knowledge graph, important information in the bidding document can be captured, and subsequent quotation and bidding document writing are facilitated.

Description

Technical Field

[0001] The present invention relates to the field of assisted bidding technology, and in particular to an assisted bidding method based on knowledge graph. Background Art

[0002] Bidding refers to the process in which suppliers, contractors or service providers respond to procurement requirements issued by the tendering party (usually a government agency, enterprise or other organization) and submit quotations and technical solutions.

[0003] In the bidding process, correctly interpreting the bidding documents has the following advantages: 1. Avoid bid rejection: The bidding documents usually list various requirements and regulations in detail, such as format, content, submission time, etc. If the bidding documents do not meet these requirements, they may be directly deemed invalid (i.e. "bid rejection"). Therefore, by accurately understanding each requirement in the bidding documents, you can ensure the integrity and compliance of the bidding documents, thereby avoiding the risk of bid rejection due to minor errors; 2. Correct response scoring criteria: The bidding documents usually clearly indicate the evaluation criteria and weights. Only by fully understanding these criteria can you prepare the bidding documents in a targeted manner, ensure that the technical solutions and commercial terms can meet the requirements of the evaluation committee and improve the score.

[0004] 3. Effectively customize bidding strategies: (1) Understand project requirements: By reading the bidding documents in depth, you can fully understand the background, objectives, scope and technical requirements of the project, which will help you formulate a technical solution and quotation strategy that meets the needs of the tenderer. (2) Identify key points: The bidding documents may contain some key terms or special requirements, which are often the focus of bid evaluation. Through careful interpretation, you can identify these key points and respond to them in the bidding documents.

[0005] 4. Increase the probability of winning the bid: (1) Accurate positioning: Through in-depth analysis of the bidding documents, you can more accurately identify your strengths and weaknesses, and thus make the best choice in terms of technical solutions, quotations and services. (2) Competitive advantage: Understanding the implicit information and potential needs in the bidding documents can make the bidding documents more in line with the actual needs of the tenderer, thus standing out from many competitors.

[0006] 5. Facilitate subsequent tender writing and quotation: After accurately understanding the bidding documents, you can communicate and coordinate with team members more effectively to ensure that all preparations are carried out as planned, which not only improves work efficiency but also reduces rework and delays caused by misunderstandings.

[0007] In order to interpret the bidding documents, it is generally necessary for multiple professionals to interpret the bidding documents multiple times, and then conduct internal reviews. If necessary, experts need to be consulted, so there are still some problems: 1. High degree of subjectivity: The bidding documents are interpreted by multiple professionals multiple times. Different people may have different understandings and interpretations, resulting in the final bidding documents being not unified and consistent.

[0008] 2. High labor costs: Multiple interpretations and internal reviews require a lot of human resources and time investment, which increases the cost of the bidding process.

[0009] 3. Incomplete understanding: Even with the participation of multiple professionals, it is difficult to ensure that everyone can fully understand all the details of the bidding documents, especially for complex technical or legal terms, which may lead to misunderstandings. Summary of the invention

[0010] The purpose of this invention is to provide an auxiliary bidding method based on knowledge graph to solve the above technical problems.

[0011] To achieve the above object, the present invention provides an auxiliary bidding method based on knowledge graph, comprising the following steps: S1. Capture the bidding text through a crawler and pre-process the bidding text; S2, sequence labeling: construct a BERT-BILSTM-CRF fusion model with an attention mechanism, and use the constructed BERT-BILSTM-CRF fusion model with an attention mechanism to extract named entities from the bidding text preprocessed in step S1; S3, Relation classification: construct a BiGRU-attention relationship classification model, and use the constructed BiGRU-attention relationship classification model to predict the relationship of the named entities extracted in step S2; S4. According to the relationship classification result obtained in step S3, the named entities and relationships are combined to obtain valid entities, and stored in the database.

[0012] Preferably, step S2 specifically includes the following steps: S21. Use the attention mechanism to process the feature sequence in the bidding text and generate a global feature representation. The expression of the attention mechanism is: (1); In the formula, Indicates The global feature representation of features; Indicates the length of the feature sequence; Indicates The feature pair The level of attention paid to each feature; Indicates The initial representation of features; in, (2); In the formula, Indicates the calculation for Features and A function of feature similarity; Indicates the calculation for Features and A function of feature similarity; S22, training the BERT model using the MLM training method, and using the trained BERT model to generate a bidirectional semantic representation based on the feature sequence processed in step S21; S23. Use BILSTM to further process the feature sequence and capture context information: Use BILSTM to input the context information of the bidding text into a forward LSTM and a reverse LSTM respectively, and output two vectors at each time node, and then connect the two vectors output at the same time node to form the final output of BILSTM, which will then be sent to the attention unit; S24. Consider the context information and use the classification labels of the CRF model for classification labeling.

[0013] Preferably, in step S23, the LSTM includes a forget gate for determining discarded information, an input gate for determining retained information, and an output gate for outputting final information; Among them, the forget gate is Always hide status and the current input , and after passing through the Sigmoid function, it outputs a value between 0 and 1, and the output of the forget gate The expression is as follows: (3); In the formula, Represents the Sigmoid function; Indicates the Always hide status For the current moment The weight matrix of the impact degree; Indicates that it is used to adjust the current input For the current moment The weight matrix of the impact degree; Represents the forget gate bias term; In the input gate, the hidden state and the current input Determine the input of the input gate, and its output expression is as follows: (4); (5); In the formula, represents the output of the input gate; Represents adjusting the hidden state from the previous time step to the current input gate The weight matrix of the impact degree; Indicates that it is used to adjust the current input The weight matrix of the influence on the input gate state; represents the input gate bias term; represents the activation function; Represents the weight matrix used to adjust the influence of the hidden layer on the candidate memory unit; Indicates the process used to process the current input The weight matrix of Represents the bias term of the candidate memory unit; Update cell status: (6); In the formula, express Cell status at each moment; express The cell state at a given moment; In the output gate, through the hidden state and the current input Determine the output gate value, and the output gate outputs The expression is as follows: (7); (8); In the formula, Indicates the hidden state to be adjusted The weight matrix for the output gate value; Indicates adjustment of the current input The weight matrix affecting the state of the output gate; represents the output gate bias term; express Always hide the status; The prediction output expression is as follows: (9); In the formula, express The predicted output at the moment; Indicates the hidden state to be adjusted A weight matrix of the influence on the predicted output; Represents the bias term used to adjust the output baseline value.

[0014] Preferably, step S3 specifically includes the following steps: S31. Build BiGRU-attention relationship classification model: (10); In the formula, Indicates that the forward GRU is The hidden state of the moment; Indicates that the forward GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; , , and Both represent weight matrices; Represents the hidden state vector after nonlinear transformation; represents the bias term; represents the attention weight; represents the transformation vector; Represents the final output; S32, extracting named entity-relationship-named entity triples using the constructed BiGRU-attention relation classification model; S33, knowledge fusion: calculate the similarity of the extracted named entity-relationship-named entity triples to identify named entities with the same meaning, and merge the named entities with the same meaning; S34, Entity disambiguation: Use clustering methods to perform entity disambiguation based on contextual information; S35. Store in the knowledge base.

[0015] Preferably, in step S33, the Word2vec model is used to calculate the spatial distance of the words that make up the named entity, and the spatial distance is used to represent the semantic similarity between the words, and the calculated semantic similarity is compared with a set threshold. If the calculated semantic similarity is greater than the set threshold, they are merged, otherwise they are retained.

[0016] Preferably, after step S4, an indicator evaluation is also included, and the selected evaluation indicators include precision P, recall R and accuracy F1.

[0017] Therefore, the present invention adopts the above-mentioned auxiliary bidding method based on knowledge graph, which has the following beneficial effects: 1. Automation and efficiency: By using crawler technology to automatically crawl the bidding text, a lot of time for manual information collection can be saved, which makes the whole process more efficient and reduces the possibility of human error; 2. Accurate information extraction: Using the BERT-BILSTM-CRF fusion model for sequence labeling can more accurately extract named entities from the preprocessed bidding text. The BERT model performs well in understanding context, while the BILSTM-CRF model is good at capturing dependencies in sequences. The combination of the two can improve the accuracy of named entity recognition. At the same time, the introduction of the attention mechanism can make the model pay more attention to key parts when processing long texts, thereby further improving the accuracy of entity recognition. 3. Accurate relationship classification: Using the BiGRU-attention relationship classification model to predict the relationship between the extracted named entities can better understand and analyze the complex relationships between entities. The BiGRU model can effectively capture sequence information, and the attention mechanism can help the model focus on the most important features for relationship classification, thereby improving the accuracy of relationship classification. 4. Structured data storage and query: Named entities and relationships are combined into valid entities and stored in the database, which can transform unstructured text information into structured knowledge graphs. This not only facilitates subsequent data management and query, but also supports complex association analysis and reasoning, providing strong support for bidding decisions; 5. Intelligent decision support: The knowledge graph-based method can provide richer contextual information and multi-dimensional correlation analysis, helping users to more comprehensively understand the content and requirements of the bidding text, and help make more informed decisions during the bidding process, thereby increasing the winning rate. 6. Scalability and adaptability: This method has a flexible framework and can improve performance by continuously optimizing model parameters or introducing new algorithms. It has good scalability and adaptability. In summary, the present invention provides users with efficient, accurate and intelligent bidding support through automated and precise information extraction and relationship classification, as well as structured data storage.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of an auxiliary bidding method based on knowledge graph of the present invention. DETAILED DESCRIPTION

[0020] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, or the positions or positional relationships in which the invented product is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In the description of the present invention, it should also be noted that, unless otherwise clearly specified and limited, the terms "setting", "installation", and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0021] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, a knowledge graph-based auxiliary bidding method includes the following steps: S1. Capture the bidding text through a crawler and pre-process the bidding text; in this embodiment, the bidding text is pre-processed by data cleaning, abnormal value removal, etc.; S2, sequence labeling: construct a BERT-BILSTM-CRF fusion model with an attention mechanism, and use the constructed BERT-BILSTM-CRF fusion model with an attention mechanism to extract named entities from the bidding text preprocessed in step S1; Step S2 specifically includes the following steps: S21. Use the attention mechanism to process the feature sequence in the bidding text and generate a global feature representation. The expression of the attention mechanism is: (1); In the formula, Indicates The global feature representation of features; Indicates the length of the feature sequence; Indicates The feature pair The level of attention paid to each feature; Indicates The initial representation of features; in, (2); In the formula, Indicates the calculation for Features and A function of feature similarity; Indicates the calculation for Features and A function of feature similarity; S22, training the BERT model using the MLM training method, and using the trained BERT model to generate a bidirectional semantic representation based on the feature sequence processed in step S21; S23. Use BILSTM to further process the feature sequence and capture context information: Use BILSTM to input the context information of the bidding text into a forward LSTM and a reverse LSTM respectively, and output two vectors at each time node, and then connect the two vectors output at the same time node to form the final output of BILSTM, which will then be sent to the attention unit; In step S23, the LSTM includes a forget gate for determining discarded information, an input gate for determining retained information, and an output gate for outputting final information; Among them, the forget gate is Always hide status and the current input , and after passing through the Sigmoid function, it outputs a value between 0 and 1, and the output of the forget gate The expression is as follows: (3); In the formula, Represents the Sigmoid function; Indicates the Always hide status For the current moment The weight matrix of the impact degree; Indicates that it is used to adjust the current input For the current moment The weight matrix of the impact degree; Represents the forget gate bias term; In the input gate, the hidden state and the current input Determine the input of the input gate, and its output expression is as follows: (4); (5); In the formula, represents the output of the input gate; Represents adjusting the hidden state from the previous time step to the current input gate The weight matrix of the impact degree; Indicates that it is used to adjust the current input The weight matrix of the influence on the input gate state; represents the input gate bias term; represents the activation function; Represents the weight matrix used to adjust the influence of the hidden layer on the candidate memory unit; Indicates the process used to process the current input The weight matrix of Represents the bias term of the candidate memory unit; Update cell status: (6); In the formula, express Cell status at each moment; express The cell state at a given moment; In the output gate, through the hidden state and the current input Determine the output gate value, and the output gate outputs The expression is as follows: (7); (8); In the formula, Indicates the hidden state to be adjusted The weight matrix for the output gate value; Indicates adjustment of the current input The weight matrix affecting the state of the output gate; represents the output gate bias term; express Always hide the status; The prediction output expression is as follows: (9); In the formula, express The predicted output at the moment; Indicates the hidden state to be adjusted A weight matrix of the influence on the predicted output; Represents the bias term used to adjust the output baseline value.

[0023] S24. Consider the context information and use the classification labels of the CRF model for classification labeling.

[0024] S3, Relation classification: construct a BiGRU-attention relationship classification model, and use the constructed BiGRU-attention relationship classification model to predict the relationship of the named entities extracted in step S2; Step S3 specifically includes the following steps: S31. Build BiGRU-attention relationship classification model: (10); In the formula, Indicates that the forward GRU is The hidden state of the moment; Indicates that the forward GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; , , and Both represent weight matrices; Represents the hidden state vector after nonlinear transformation; represents the bias term; represents the attention weight; represents the transformation vector; Represents the final output; S32, extracting named entity-relationship-named entity triples using the constructed BiGRU-attention relation classification model; S33, knowledge fusion: calculate the similarity of the extracted named entity-relationship-named entity triples to identify named entities with the same meaning, and merge the named entities with the same meaning; In step S33, the Word2vec model is used to calculate the spatial distance of the words that make up the named entity, and the spatial distance is used to represent the semantic similarity between the words. The calculated semantic similarity is compared with the set threshold. If the calculated semantic similarity is greater than the set threshold, they are merged, otherwise they are retained.

[0025] S34, Entity disambiguation: Use clustering methods to perform entity disambiguation based on contextual information; S35. Store in the knowledge base.

[0026] S4. According to the relationship classification result obtained in step S3, the named entities and relationships are combined to obtain valid entities, and stored in the database.

[0027] After step S4, an indicator evaluation is also included, and the selected evaluation indicators include precision P, recall R and accuracy F1.

[0028] To demonstrate the effectiveness of the present invention, the fusion model of the present invention is compared with the existing model under the same data set.

[0029] Table 1 Evaluation results ;

[0030] Thereby the effectiveness of the present invention is proved.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A knowledge graph-based bidding assistance method, characterized in that: The following steps are involved: S1. Capture the bidding text through a crawler and pre-process the bidding text; S2, sequence labeling: construct a BERT-BILSTM-CRF fusion model with an attention mechanism, and use the constructed BERT-BILSTM-CRF fusion model with an attention mechanism to extract named entities from the bidding text preprocessed in step S1; S3, Relation classification: construct a BiGRU-attention relationship classification model, and use the constructed BiGRU-attention relationship classification model to predict the relationship of the named entities extracted in step S2; S4. According to the relationship classification result obtained in step S3, the named entities and relationships are combined to obtain valid entities, and stored in the database.

2. The knowledge graph-based bidding assistance method according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Use the attention mechanism to process the feature sequence in the bidding text and generate a global feature representation. The expression of the attention mechanism is: （1）； In the formula, Indicates The global feature representation of features; Indicates the length of the feature sequence; Indicates The feature pair The level of attention paid to each feature; Indicates The initial representation of features; in, （2）； In the formula, Indicates the calculation for Features and A function of feature similarity; Indicates the calculation for Features and A function of feature similarity; S22, training the BERT model using the MLM training method, and using the trained BERT model to generate a bidirectional semantic representation based on the feature sequence processed in step S21; S23. Use BILSTM to further process the feature sequence and capture context information: Use BILSTM to input the context information of the bidding text into a forward LSTM and a reverse LSTM respectively, and output two vectors at each time node, and then connect the two vectors output at the same time node to form the final output of BILSTM, which will then be sent to the attention unit; S24. Consider the context information and use the classification labels of the CRF model for classification labeling.

3. The knowledge graph-based bidding assistance method according to claim 2 is characterized in that: In step S23, the LSTM includes a forget gate for determining discarded information, an input gate for determining retained information, and an output gate for outputting final information; Among them, the forget gate is Always hide status and the current input , and after passing through the Sigmoid function, it outputs a value between 0 and 1, and the output of the forget gate The expression is as follows: （3）； In the formula, Represents the Sigmoid function; Indicates the Always hide status For the current moment The weight matrix of the impact degree; Indicates that it is used to adjust the current input For the current moment The weight matrix of the impact degree; Represents the forget gate bias term; In the input gate, the hidden state and the current input Determine the input of the input gate, and its output expression is as follows: （4）； （5）； In the formula, represents the output of the input gate; Represents adjusting the hidden state from the previous time step to the current input gate The weight matrix of the impact degree; Indicates that it is used to adjust the current input The weight matrix of the influence on the input gate state; represents the input gate bias term; represents the activation function; Represents the weight matrix used to adjust the influence of the hidden layer on the candidate memory unit; Indicates the process used to process the current input The weight matrix of Represents the bias term of the candidate memory unit; Update cell status: （6）； In the formula, express Cell status at each moment; express The cell state at a given moment; In the output gate, through the hidden state and the current input Determine the output gate value, and the output gate outputs The expression is as follows: （7）； （8）； In the formula, Indicates the hidden state to be adjusted The weight matrix for the output gate value; Indicates adjustment of the current input The weight matrix affecting the state of the output gate; represents the output gate bias term; express Always hide the status; The prediction output expression is as follows: （9）； In the formula, express The predicted output at the moment; Indicates the hidden state to be adjusted A weight matrix of the influence on the predicted output; Represents the bias term used to adjust the output baseline value.

4. The knowledge graph-based bidding assistance method according to claim 3 is characterized in that: Step S3 specifically includes the following steps: S31. Build BiGRU-attention relationship classification model: （10）； In the formula, Indicates that the forward GRU is The hidden state of the moment; Indicates that the forward GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; Indicates that the reverse GRU is The hidden state of the moment; , , and Both represent weight matrices; Represents the hidden state vector after nonlinear transformation; represents the bias term; represents the attention weight; represents the transformation vector; Represents the final output; S32, extracting named entity-relationship-named entity triples using the constructed BiGRU-attention relation classification model; S33, knowledge fusion: calculate the similarity of the extracted named entity-relationship-named entity triples to identify named entities with the same meaning, and merge the named entities with the same meaning; S34, Entity disambiguation: Use clustering methods to perform entity disambiguation based on contextual information; S35. Store in the knowledge base.

5. The knowledge graph-based bidding assistance method according to claim 4 is characterized in that: In step S33, the Word2vec model is used to calculate the spatial distance of the words that make up the named entity, and the spatial distance is used to represent the semantic similarity between the words. The calculated semantic similarity is compared with the set threshold. If the calculated semantic similarity is greater than the set threshold, they are merged, otherwise they are retained.

6. The knowledge graph-based bidding assistance method according to claim 5 is characterized in that: After step S4, an indicator evaluation is also included, and the selected evaluation indicators include precision P, recall R and accuracy F1.

Images