A method and system for recommending tendered products

Abstract

This application provides a method for recommending tendered products, comprising: determining the users involved in the tender documents through a user preference agent, and obtaining the recent product matching preference logic of the users; obtaining historical similar product matching corresponding to the tender documents through a parameter learning agent, and obtaining a second candidate product to be recalled and its corresponding product parameter matching logic; obtaining a third candidate product to be recalled by matching a new product list and a main product list based on the product name and product parameters in the tender documents; obtaining a high-dimensional semantic score by performing forward reasoning through a shallow scoring network based on the preference score, product parameter score, fusion score of each product in the recalled product list, as well as the one-way coverage, reverse coverage, and two-way coverage of each product's product parameters; and reordering the products in the recalled product list based on the high-dimensional semantic score to obtain a recommendation list.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for recommending tendered products. Background Technology

[0002] During the bidding process, product requirements can usually only be obtained through the company's bidding announcement. Ordinary users cannot obtain detailed information about the bidding process. Therefore, it is often more difficult to match products more accurately and provide users with more precise product matching than to recommend products on ordinary social media platforms.

[0003] Therefore, given the limited sources of information and the scarcity of bidding details, improving the accuracy of product recommendations for bidding has become an urgent technical problem to be solved. Summary of the Invention

[0004] This application provides a method and system for recommending tendered products, addressing the technical problem of low accuracy in recommending tendered products when information sources are limited and tender details are scarce.

[0005] Firstly, this application provides a method for recommending tendered products, the method comprising:

[0006] Obtain the product names and parameters from the tender documents;

[0007] The user preference agent determines the order allocation user involved in the tender document, obtains the order allocation preference logic of the order allocation user for recent order allocation products, and uses the recent order allocation products as the first candidate products to be recalled.

[0008] The system obtains historical similar product lists corresponding to the tender documents through a parameter learning agent. Based on the product name score and bidirectional coverage of product parameters for each product in the historical similar product lists, it obtains the second candidate product to be recalled and the corresponding product parameter matching logic. Based on the product name and product parameters in the tender documents, it matches the new product list and the main product list to obtain the third candidate product to be recalled.

[0009] The list of the first candidate product, the second candidate product, and the third candidate product shall be the recall product list;

[0010] The user preference agent scores the preferences of the products in the recall product list.

[0011] The parameter learning agent scores the product parameters of the products in the recall product list.

[0012] By learning the order matching preference logic and the product parameter order matching logic through the fusion agent, a fusion score is generated using the product name and product parameters in the tender document and the product name and product parameters in the recall product list as prompt words.

[0013] A shallow scoring network performs forward reasoning based on the preference score, product parameter score, fusion score, and unidirectional, inverse, and bidirectional coverage of each product's parameters in the recalled product list to obtain a high-dimensional semantic score. The products in the recalled product list are then rearranged based on the high-dimensional semantic score to obtain a recommendation list.

[0014] Optionally, the matching users involved in the tender documents are determined through a user preference agent, and the matching preference logic of the matching users for recent matching products is obtained, including:

[0015] The user preference agent obtains the historical product demand and historical order allocation corresponding to N recent order allocation products for each order allocation user.

[0016] The user preference agent determines the corresponding order matching logic based on the matching results between the product requirements in the tender document and the historical product requirements, where N is a positive integer.

[0017] Optionally, the logic for obtaining the second candidate product to be recalled and its corresponding product parameters based on the bidirectional coverage of the product name score and product parameters of each product in historical similar order matching includes:

[0018] Obtain the product descriptions corresponding to historical similar orders in the historical order database of the tender documents;

[0019] For each product parameter in the product parameter list of the tender document, the effective count of the product parameter in the historical allocation database is calculated according to the corresponding historical products. Based on the effective count, the one-way coverage of each product parameter from the tender document to the historical allocation database and the reverse coverage from the historical allocation database to the tender document are calculated.

[0020] The bidirectional coverage of each product parameter is obtained by multiplying the reverse coverage and the unidirectional coverage.

[0021] Based on the product name and the product name in the tender document, obtain the product name score;

[0022] By combining the product name score and the bidirectional coverage of product parameters for each product in historical similar order data, the top K products are selected as the second candidate products to be recalled.

[0023] Based on the product parameters of the top K products and the product parameters in the tender document, the corresponding parameter matching logic is determined, where K is a positive integer.

[0024] Optionally, by combining the product name score and the bidirectional coverage of product parameters for each product in historical similar order data, the top K products are selected as the second candidate products to be recalled, including:

[0025] The first vector representation corresponding to the product name in the tender document is obtained based on the Embedding model.

[0026] The second vector representation of each product name in the historical order database is obtained based on the Embedding model;

[0027] The similarity between the first vector representation and the second vector representation is used as the product name score for the corresponding historical product.

[0028] The product parameter score corresponding to the historical product is obtained by weighting the product name score and the bidirectional coverage of the corresponding product parameters.

[0029] The top K historical products with the highest product parameter scores will be selected as the second candidate products for recall.

[0030] Optionally, based on the matching results between the product requirements in the tender document and the historical product requirements, the corresponding order allocation preference logic is determined, including:

[0031] The user preference agent, based on a large language model, uses historical product requirements and historical order allocation as prompt words, and takes the product requirements in the tender document as input to generate corresponding order allocation preference logic.

[0032] Optionally, the learning agent, based on the product names and parameters in the tender documents, matches the new product list and the main product list to obtain a third candidate product to be recalled, including:

[0033] The learning agent uses the parameters to match the product name score and product parameter bidirectional coverage of each product in the new product list based on the product name and product parameters in the tender document, and obtains the new product recall products.

[0034] The learning agent uses the parameters to determine popular recall products based on the popularity of each product in the main product list.

[0035] Based on the newly recalled products and the popular recalled products, a third candidate product to be recalled is obtained.

[0036] Optionally, obtain the product names and parameters from the tender documents, including:

[0037] The structured product requirement table in the tender document is obtained through OCR (Optical Character Recognition) and / or entity recognition technology. The structured product requirement table includes product name and product parameters.

[0038] Optionally, the shallow scoring network comprises two fully connected layers, and the method further includes:

[0039] The two-layer fully connected network maps the preference score, product parameter score, fusion score of each product, as well as the one-way coverage, reverse coverage and two-way coverage of each product's product parameters to a high-dimensional semantic space, and performs forward reasoning to output the corresponding high-dimensional semantic score.

[0040] On the other hand, embodiments of this application provide a multi-agent-based bidding product recommendation system, including:

[0041] The information extraction module is used to obtain the product names and product parameters from the tender documents;

[0042] The user preference agent identifies the order allocation user involved in the tender document, obtains the order allocation preference logic of the order allocation user for recent order allocation products, and uses the recent order allocation products as the first candidate products to be recalled.

[0043] A parameter learning agent is used to obtain historical similar product lists corresponding to the tender documents. Based on the product name score and bidirectional coverage of product parameters for each product in the historical similar product lists, it obtains the second candidate product to be recalled and the corresponding product parameter matching logic. Based on the product name and product parameters in the tender documents, it matches the new product list and the main product list to obtain the third candidate product to be recalled.

[0044] The recall module is used to generate a recall product list from the list of the first candidate product, the second candidate product, and the third candidate product.

[0045] The user preference agent is also used to score preferences for products in the recall product list;

[0046] The parameter learning agent is also used to score the product parameters of the products in the recalled product list.

[0047] The fusion agent is used to learn the order matching preference logic and the product parameter order matching logic, and generates a fusion score using the product names and parameters in the tender document and the product names and parameters in the recalled product list as prompt words;

[0048] A shallow scoring network is used to perform forward reasoning based on the preference score, product parameter score, fusion score, and unidirectional coverage, reverse coverage, and bidirectional coverage of each product's product parameters in the recalled product list to obtain a high-dimensional semantic score. Based on the high-dimensional semantic score, the products in the recalled product list are rearranged to obtain a recommendation list.

[0049] On the other hand, an electronic device is provided, comprising:

[0050] processor;

[0051] The memory stores computer-readable instructions, which, when executed by the processor, implement the methods described above.

[0052] The beneficial effects of the technical solutions provided in this application include at least the following:

[0053] In this embodiment, user preference logic and preference scores are obtained from historical order matching data. The system learns from the order matching experience of different users and expands the scope of product searches by using a hybrid search method combining product name and product parameters. This avoids the limitations of searching with a single field, especially addressing discrepancies caused by inconsistencies in product name descriptions from different bidding entities (e.g., display screen vs. monitor), thus improving search accuracy. Furthermore, by using a parameter learning agent to obtain bidirectional coverage of product parameters from the bidding documents, the system can better learn the importance of different product parameters. By fusing the agent to learn user preference logic and product parameter order matching logic, a more accurate fusion score is obtained. Since forward reasoning is achieved across six dimensions based on user preference scores, product parameter scores, and fusion scores, as well as one-way, reverse, and bidirectional coverage, the final recommended products better cover user needs, improving user satisfaction. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of a method for recommending tendered products provided in an embodiment of this application;

[0056] Figure 2 This is a schematic diagram of a structured product requirement table provided in an embodiment of this application;

[0057] Figure 3 This is a schematic diagram of the bidding product recommendation system in the embodiments of this application;

[0058] Figure 4 This is a schematic diagram of a bidding product recommendation process provided in an embodiment of this application. Detailed Implementation

[0059] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0060] The bidding product recommendation method provided in this application combines a large language model and an intelligent agent to process bidding products. When making intelligent product recommendations, it comprehensively considers user preferences and the importance of product parameters, taking into account the different order matching experience and different understandings of products of different users.

[0061] See Figure 1 As shown, the method for recommending products in this tender includes the following steps:

[0062] S101: Obtain the product name and product parameters from the tender documents;

[0063] S102: Determine the order allocation user involved in the tender document through the user preference intelligent agent, obtain the order allocation preference logic of the order allocation user's recent order allocation products, and use the recent order allocation products as the first candidate products to be recalled;

[0064] S103: Obtain historical similar product lists corresponding to the tender documents through a parameter learning agent; based on the product name score and bidirectional coverage of product parameters for each product in the historical similar product lists, obtain the second candidate product to be recalled and the corresponding product parameter matching logic; based on the product name and product parameters in the tender documents, match the new product list and the main product list to obtain the third candidate product to be recalled.

[0065] S104: The list of the first candidate product, the second candidate product, and the third candidate product shall be the recall product list;

[0066] S105: Use a user preference agent to score the preferences of products in the recall product list;

[0067] S106: The parameter learning agent scores the product parameters of the products in the recall product list.

[0068] S107: By learning the order matching preference logic and the product parameter order matching logic through the fusion agent, a fusion score is generated using the product name and product parameters in the tender document as prompt words, and the product name and product parameters in the tender document and the product name and product parameters in the recall product list as prompt words.

[0069] S108: A high-dimensional semantic score is obtained by forward reasoning based on the preference score, product parameter score, fusion score, and one-way coverage, reverse coverage, and two-way coverage of each product's product parameters in the recalled product list using a shallow scoring network. The products in the recalled product list are then rearranged based on the high-dimensional semantic score to obtain a recommendation list.

[0070] In this application, "large language model" refers to natural language processing models that have been trained on a large amount of data and have hundreds of millions or even trillions of parameters. These models are able to understand and generate natural language text through deep learning and machine learning techniques, such as llama2 and Qwen7B.

[0071] An intelligent agent refers to an AI assistant that can think and act autonomously. It can understand complex instructions from large language models (LLMs) and call on tools to complete tasks.

[0072] In this embodiment, user preference logic and preference scores are obtained from historical order matching data. The system learns from the order matching experience of different users and expands the scope of product searches by using a hybrid search method combining product name and product parameters. This avoids the limitations of searching with a single field, especially addressing discrepancies caused by inconsistencies in product name descriptions from different bidding entities (e.g., display screen vs. monitor), thus improving search accuracy. Furthermore, by using a parameter learning agent to obtain bidirectional coverage of product parameters from the bidding documents, the system can better learn the importance of different product parameters. By fusing the agent to learn user preference logic and product parameter order matching logic, a more accurate fusion score is obtained. Since forward reasoning is achieved across six dimensions based on user preference scores, product parameter scores, and fusion scores, as well as one-way, reverse, and bidirectional coverage, the final recommended products better cover user needs, improving user satisfaction.

[0073] The above S101 can be implemented in the following way:

[0074] Obtain the structured product requirement form from the tender document for Optical Character Recognition (OCR) and / or entity recognition technologies. The structured product requirement form includes product name and product parameters.

[0075] In this embodiment, the original tender document (which can be in PDF, DOC, or other formats) can be parsed using technologies such as OCR and entity recognition to extract a structured product requirement table from the original tender document PDF. A single tender document may include multiple tender products, and each product has corresponding product requirements.

[0076] The above step S102 can be implemented, but is not limited to, in the following ways:

[0077] The user preference agent obtains the historical product demand and historical order allocation corresponding to N recent order allocation products for each order allocation user.

[0078] The user preference agent determines the corresponding order matching logic based on the matching results between the product requirements in the tender document and the historical product requirements, where N is a positive integer.

[0079] The recently matched products can be a specified number of matched products that are closest to the current time, or matched products that are within a preset time period from the current time, such as matched products from the past year.

[0080] Optionally, a user preference agent, based on a large language model, generates corresponding allocation preference logic by using historical product requirements and historical allocations as prompt words and the product requirements in the tender document as input.

[0081] Optionally, the user preference agent processes the product names and product parameters corresponding to the multiple first recalled products based on the Transformer model, Long Short-Term Memory Network (LSTM), Graph Inference, or Natural Language Processing (NLP), and generates multiple order matching preference logics corresponding to the first recalled products by combining the bidirectional coverage of the product parameters corresponding to each first recalled product.

[0082] In step S103 above, the parameter learning agent can, but is not limited to, obtain the bidirectional coverage of each product parameter in the tender document through the following methods:

[0083] Obtain the product descriptions corresponding to historical similar orders in the historical order database of the tender documents;

[0084] For each product parameter in the product parameter list of the tender document, the effective count of the product parameter in the historical allocation database is calculated according to the corresponding historical products. Based on the effective count, the one-way coverage of each product parameter from the tender document to the historical allocation database and the reverse coverage from the historical allocation database to the tender document are calculated.

[0085] The bidirectional coverage of each product parameter is obtained by multiplying the reverse coverage and the unidirectional coverage.

[0086] Based on the product name and the product name in the tender document, obtain the product name score;

[0087] By combining the product name score and the bidirectional coverage of product parameters for each product in historical similar order data, the top K products are selected as the second candidate products to be recalled.

[0088] Based on the product parameters of the top K products and the product parameters in the tender document, the corresponding parameter matching logic is determined, where K is a positive integer.

[0089] The second candidate product to be recalled may be obtained, but is not limited to, through the following methods:

[0090] The first vector representation corresponding to the product name in the tender document is obtained based on the Embedding model.

[0091] The second vector representation of each product name in the historical order database is obtained based on the Embedding model;

[0092] The similarity between the first vector representation and the second vector representation is used as the product name score for the corresponding historical product.

[0093] The product parameter score corresponding to the historical product is obtained by weighting the product name score and the bidirectional coverage of the corresponding product parameters.

[0094] The top K historical products with the highest product parameter scores will be selected as the second candidate products for recall.

[0095] For example, Figure 2 The equipment list details are a structured product requirement table parsed from a certain tender document. This table contains 4 product names and 4 corresponding product parameters. The product name for item number 1 is "LED display unit", the corresponding product parameter is "1. Fully enclosed...", and the final recommended product model is "DS-D4...".

[0096] For example, when the product name is "camera", the product parameters may include: camera parameters, exposure parameters, lens specifications, camera size, etc.

[0097] Assume the product name in the structured product requirements table is A. query The set of product parameters is

[0098] B query= [b 1 query b 2query ... b n query ].

[0099] In product recommendation methods, a multi-path recall approach can be used to filter out a small number of potentially relevant products from the entire candidate item set and include them in the subsequent product recommendation process. This can reduce the number of products that need to be processed in the later stages and improve the overall system response rate. On the other hand, it can also quickly filter out irrelevant items and avoid recommending invalid products.

[0100] The first method of recall can be to recall products from the historical order database:

[0101] The historical order library can include x orders A. doc1 A doc2 ...A docx The corresponding parameter set can include B doc1 B doc2 ...B docx .

[0102] The second recall method can be to recall products from the new product list and the main product list.

[0103] This multi-channel recall process can be implemented, but is not limited to, through the following methods:

[0104] Example 1:

[0105] Several new products whose release time is closest to the current time will be designated as new product recall products;

[0106] Several key products designated in the list of key products will be included as popular recall products.

[0107] Based on the newly recalled products and the popular recalled products, a third candidate product to be recalled is obtained.

[0108] Example 2: The learning agent uses the parameters to match the bidirectional coverage of the product name score and product parameters of each product in the new product list based on the product name and product parameters in the tender document, and obtains the new product recall products.

[0109] The learning agent uses the parameters to determine popular recall products based on the popularity of each product in the main product list.

[0110] Based on the newly recalled products and the popular recalled products, a third candidate product to be recalled is obtained.

[0111] Optionally, the fusion agent learns the order matching logic determined by the user preference agent and the order matching logic determined by the parameter learning agent, and generates a fusion score based on the product parameter score and preference score of each product, using the product name and product parameters in the tender document as prompts.

[0112] For example, the parameter learning agent can generate corresponding product matching prompts based on the parameter importance, one-way coverage, reverse coverage, and two-way coverage of the product parameters corresponding to each product in the tender document, as the product parameter matching logic.

[0113] In optional embodiments of this application, the importance of product parameters can be calculated, but is not limited to, based on the following process:

[0114] (1) Obtain the first vector representation E corresponding to the product name in the tender document. query =Embedding(A query ).

[0115] (2) Calculate the second vector representation E of the product name of each candidate product in the historical order matching database of the order matching user. doc =Embedding(A doc ).

[0116] (3) Calculate the cosine similarity between the first vector representation and the second vector representation, and use it as the product name score S for candidate products. name ：

[0117] (4) Obtain the product description corresponding to each candidate product. For each product parameter in the product parameter list of the tender document, calculate the effective count n of the product parameter in the historical order matching database according to the corresponding candidate product. b .

[0118] (5) Calculate the parameter importance of each product parameter based on the valid count of each product parameter.

[0119] Alternatively, the importance of each product parameter in the historical order matching database can be obtained using the following formula:

[0120]

[0121] Where w(b) represents the importance of the product parameter, b represents the product parameter, and n b This represents the valid count of product parameters, and N represents the total number of all parameters in the historical order database.

[0122] (6) For each product parameter, calculate the single coverage from the tender document to the historical order database and the reverse coverage from the historical order database to the tender document according to the importance of the parameter.

[0123]

[0124] in, This indicates the coverage of a single item from the tender documents to the historical allocation database. This represents the sum of parameter importance from the intersection of the tender document and the historical allocation database. This represents the sum of the importance of the parameters in the tender documents; This indicates the intersection between the tender documents and the historical allocation database. This indicates the reverse coverage from the historical order database to the tender documents. This represents the sum of parameter importance from the intersection of the tender document and the historical allocation database. This represents the sum of parameter importance in the historical allocation database.

[0125] (7) For each product parameter, calculate the product of the maximum values ​​of the corresponding one-way coverage and reverse coverage to obtain the corresponding two-way coverage, which is used as the product parameter score for that product parameter:

[0126] Among them, S params Characterizes the two-way coverage of product parameters. For example, the weight hyperparameter. Setting it to 0.8 can prevent issues caused by a large number of parameters in the doc file. Minimal coverage, multiplied by reverse coverage, can prevent bias.

[0127] (8) Obtain the list of recalled products based on the product name score and product parameter score of the product name in the tender document.

[0128] For example, the list of recalled products can be determined in the following manner:

[0129]

[0130] Among them, S total This represents the list of recalled products, and λ1 and λ2 represent the recall parameters, where λ1 can be set to 0.9 and λ2 can be set to 0.05.

[0131] Assuming that in the first phase, λ1=0.9 is used to recall the top 100 products by product name score, and in the second phase, λ2=0.05 is used to recall the top 10 products by product parameter score, there will be 110 products in the recall product list.

[0132] In one alternative embodiment, several historical order lists that are most similar to the structured product requirement list can also be retrieved from the historical order list database based on similarity. Assuming there are 10 historical order lists, these 10 order lists can be used... express.

[0133] See Figure 3 As shown in the embodiments of this application, a bidding product recommendation system can also be provided, including:

[0134] The information extraction module is used to obtain the product names and product parameters from the tender documents;

[0135] The user preference agent identifies the order allocation user involved in the tender document, obtains the order allocation preference logic of the order allocation user for recent order allocation products, and uses the recent order allocation products as the first candidate products to be recalled.

[0136] A parameter learning agent is used to obtain historical similar product lists corresponding to the tender documents. Based on the product name score and bidirectional coverage of product parameters for each product in the historical similar product lists, it obtains the second candidate product to be recalled and the corresponding product parameter matching logic. Based on the product name and product parameters in the tender documents, it matches the new product list and the main product list to obtain the third candidate product to be recalled.

[0137] The recall module is used to generate a recall product list from the list of the first candidate product, the second candidate product, and the third candidate product.

[0138] The user preference agent is also used to score preferences for products in the recall product list;

[0139] The parameter learning agent is also used to score the product parameters of the products in the recall product list.

[0140] The fusion agent is used to learn the order matching preference logic and the product parameter order matching logic, and uses the product name and product parameters in the tender document as prompt words to generate a fusion score based on the product parameter score and preference score of each product.

[0141] A shallow scoring network is used to perform forward reasoning based on the preference score, product parameter score, fusion score, and one-way coverage, reverse coverage, and two-way coverage of each product's product parameters in the recalled product list to obtain a high-dimensional semantic score. Based on the high-dimensional semantic score, the products in the recalled product list are rearranged to obtain the recalled list.

[0142] See Figure 4 As shown, from the historical order database, based on similarity, several historical orders that are most similar to the structured product requirement list are retrieved. , with C sim All products in the list are used as a recall candidate product list I. I is then input into the bidding product recommendation system. The bidding product recommendation system may also include an evaluation layer, which is used to evaluate whether to invoke a user preference agent or a parameter learning agent.

[0143] If the user preference agent is invoked, the user preference agent can obtain the preference score M of the user who is assigned the order. user And the reasons for the rating r user The reason for this rating is used as the logic for user order matching preferences;

[0144] If the parameter learning agent is invoked, the parameter learning agent can obtain the product parameter score M for the user who is assigning the order. param And the reasons for the rating r param This rating reason will be used as the logic for configuring product parameters.

[0145] For example, in a shallow scoring network, user order matching logic and product parameter order matching logic can be added to the sorting prompts of the fusion agent by using placeholders in fixed slots ("##User Preference Supplement" and "##Parameter Knowledge Supplement"), thus obtaining the fusion score M. sort .

[0146] The shallow scoring network generates a re-ranked list of recommended products based on two single-item coverage and two-way coverage, combined with fusion scoring.

[0147] For example, a shallow scoring network includes two fully connected layers. The process of generating a re-ranked list of recommended products based on two single-item coverages and two-way coverages, combined with a fused score, may include:

[0148] The two-layer fully connected network maps each product's preference score, product parameter score, fusion score, and the one-way coverage, reverse coverage, and two-way coverage of each product's product parameters to a high-dimensional semantic space, performs forward reasoning, and outputs the corresponding high-dimensional semantic score.

[0149] Example 1: Suppose there are 20 products in the recall product list, and each product can get 6 dimensions of rating (preference rating, product parameter rating, fusion rating, one-way coverage, reverse coverage and two-way coverage). The size of the input matrix is ​​X = (20, 6). Normalization is performed to ensure that each rating ∈ [0, 1].

[0150] Example 2: Assume the number of neurons in two fully connected networks , This is the output layer. Among them, The weight matrix represents the weights of different layers of the model, which are obtained from offline training (during the offline training sample feature construction process, the label of the matching product is 1, the label of the non-matching product is 0, and the loss function is cross-entropy loss).

[0151] In the forward inference process based on the above six dimensions of scoring:

[0152] The first reasoning step, with input X, involves... After matrix multiplication, the first-level high-dimensional representation of X1 = (20, 128) is obtained;

[0153] The second reasoning, with input X1, is performed using... After matrix multiplication, we obtain the second-level high-dimensional representation of X2 = (20, 128);

[0154] The third inference, with input X2, passes through the output layer. The matrix multiplication operation yields the weighted representation of X3 = (20, 1);

[0155] After normalizing the X3 output from the output layer, map it to ∈[0, 1] to obtain the final rating probability.

[0156] Assuming the number of candidate products to be recalled is 20, after multiple agents calculate the preference score, product parameter score, and fusion score for each product, as well as the one-way coverage, reverse coverage, and two-way coverage of each product's product parameters, a 6*20 input is obtained. This input enters the shallow scoring network, and after forward propagation, it is mapped to a high-dimensional semantic space of 128*20. Then, after forward propagation, the dimensionality is reduced to a 1*20 vector. The scores (such as preference score, product parameter score, and fusion score) are normalized, compressing all scores into a 0-1 space.

[0157] The above normalization process can be achieved by inputting elements of different dimensions into the following formula to obtain the normalized result.

[0158]

[0159] Where x represents an element in the original input vector. represents the minimum value in the original input vector, max is the maximum value in the original input vector, and x' represents the normalized result.

[0160] Because the output is normalized, the differences in the dimensions of different scores and coverages can be avoided. For example, the scores generated by the agent (e.g., preference scores, product parameter scores, and fusion scores) ∈ (0, 100), while the coverages (e.g., one-way coverage, reverse coverage, or two-way coverage) ∈ (0, 1). Since the scores and coverages have different dimensions, most of the weights may be dominated by the scores. Therefore, normalizing the scores can compress them to the same spatial dimension as the coverages, avoiding the inaccuracies caused by the difference in dimensions, thereby enhancing the stability of the network's forward inference and accelerating the inference speed.

[0161] In an optional embodiment of this application, the user preference agent can be used for:

[0162] For each order placement user, obtain the historical product demand and historical order placement corresponding to the N recent order placement products of that order placement user;

[0163] Based on the matching results between the product requirements in the tender documents and the historical product requirements, the corresponding order allocation preference logic is determined, where N is a positive integer.

[0164] In an optional embodiment of this application, the user preference agent can be used for:

[0165] The user preference agent, based on a large language model, uses historical product requirements and historical order allocation as prompt words, and takes the product requirements in the tender document as input to generate corresponding order allocation preference logic.

[0166] In an optional embodiment of this application, the parameter learning agent can be used for:

[0167] Obtain the product descriptions corresponding to historical similar orders in the historical order database of the tender documents;

[0168] For each product parameter in the product parameter list of the tender document, the effective count of the product parameter in the historical allocation database is calculated according to the corresponding historical products. Based on the effective count, the one-way coverage of each product parameter from the tender document to the historical allocation database and the reverse coverage from the historical allocation database to the tender document are calculated.

[0169] The bidirectional coverage of each product parameter is obtained by multiplying the reverse coverage and the unidirectional coverage.

[0170] Based on the product name and the product name in the tender document, obtain the product name score;

[0171] By combining the product name score and the bidirectional coverage of product parameters for each product in historical similar order data, the top K products are selected as the second candidate products to be recalled.

[0172] Based on the product parameters of the top K products and the product parameters in the tender document, the corresponding parameter matching logic is determined, where K is a positive integer.

[0173] In an optional embodiment of this application, the parameter learning agent can be used for:

[0174] The first vector representation corresponding to the product name in the tender document is obtained based on the Embedding model.

[0175] The second vector representation of each product name in the historical order database is obtained based on the Embedding model;

[0176] The similarity between the first vector representation and the second vector representation is used as the product name score for the corresponding historical product.

[0177] The product parameter score corresponding to the historical product is obtained by weighting the product name score and the bidirectional coverage of the corresponding product parameters.

[0178] The top K historical products with the highest product parameter scores will be selected as the second candidate products for recall.

[0179] In an optional embodiment of this application, the parameter learning agent can be used for:

[0180] Based on the product names and parameters in the tender documents, the product name score and the bidirectional coverage of the product parameters for each product in the new product list are matched to obtain the new product recall products.

[0181] Based on the popularity of each product in the main product list, identify the most popular recall products;

[0182] Based on the newly recalled products and the popular recalled products, a third candidate product to be recalled is obtained.

[0183] In an optional embodiment of this application, the information extraction module can be used for:

[0184] The structured product requirement table in the tender document is obtained by using optical character recognition (OCR) and / or entity recognition technologies. The structured product requirement table includes product name and product parameters.

[0185] In this embodiment, user preference logic and preference scores are obtained from historical order matching data. The system learns from the order matching experience of different users and expands the scope of product searches by using a hybrid search method combining product name and product parameters. This avoids the limitations of searching with a single field, especially addressing discrepancies caused by inconsistencies in product name descriptions from different bidding entities (e.g., display screen vs. monitor), thus improving search accuracy. Furthermore, by using a parameter learning agent to obtain bidirectional coverage of product parameters from the bidding documents, the system can better learn the importance of different product parameters. By fusing the agent to learn user preference logic and product parameter order matching logic, a more accurate fusion score is obtained. Since forward reasoning is achieved across six dimensions based on user preference scores, product parameter scores, and fusion scores, as well as one-way, reverse, and bidirectional coverage, the final recommended products better cover user needs, improving user satisfaction.

[0186] In the embodiments of this application, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0187] It should also be understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0188] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via infrared, microwave, or other means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), solid-state drives, etc.

[0189] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0190] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0191] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0192] The communication interface is used for communication between the aforementioned electronic device and other devices. The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The aforementioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0193] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0194] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0195] The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method of recommending a product for bidding, characterized by, The method comprises: obtaining product name and product parameter in the bidding document; determining the product ordering user involved in the bidding document through the user preference agent, obtaining the product ordering preference logic of the recent product ordering of the product ordering user, and taking the recent product ordering as the first candidate product to be recalled; obtaining the historical similar product order corresponding to the bidding document through the parameter learning agent, obtaining the second candidate product to be recalled and the corresponding product parameter product ordering logic based on the product name score and the bidirectional coverage of the product parameter of each product in the historical similar product order, and matching the new product list and the main product list based on the product name and the product parameter in the bidding document to obtain the third candidate product to be recalled; taking the list of the first candidate product, the second candidate product and the third candidate product as the recall product list; scoring the preference of the product in the recall product list through the user preference agent; scoring the product parameter of the product in the recall product list through the parameter learning agent; learning the product ordering preference logic and the product parameter product ordering logic through the fusion agent, taking the product name and the product parameter in the bidding document and the product name and the product parameter of the product in the recall product list as the prompt, and generating the fusion score; obtaining the high-dimensional semantic score through forward reasoning based on the preference score, the product parameter score, the fusion score of each product in the recall product list, and the one-way coverage, the reverse coverage and the bidirectional coverage of the product parameter of each product through the shallow scoring network, and rearranging the products in the recall product list based on the high-dimensional semantic score to obtain the recommendation list.

2. The method of claim 1, wherein, determining the product ordering user involved in the bidding document through the user preference agent, obtaining the product ordering preference logic of the recent product ordering of the product ordering user, comprising: obtaining the historical product demand and historical product ordering corresponding to N recent product ordering products of each product ordering user through the user preference agent; determining the corresponding product ordering preference logic based on the matching result of the product demand in the bidding document and the historical product demand through the user preference agent, wherein N is a positive integer.

3. The method of claim 1, wherein, The method comprises: obtaining the product description corresponding to the historical product of the historical similar product order of the bidding document in the historical product order database; for each product parameter in the product parameter list of the bidding document, calculating the effective count of the product parameter appearing in the historical product order database according to the corresponding historical product, calculating the one-way coverage from the bidding document to the historical product order database and the reverse coverage from the historical product order database to the bidding document of each product parameter according to the effective count; obtaining the bidirectional coverage of each product parameter based on the product of the reverse coverage and the one-way coverage; obtaining the product name score based on the product name of the product and the product name in the bidding document; In combination with the product name score of each product in the historical similar bid list and the bidirectional coverage of the product parameters, topK products are obtained as the second candidate products to be recalled. Based on the product parameters of the topK products and the product parameters in the bid document, the corresponding parameter bid list logic is determined, where K is a positive integer.

4. The method of claim 3, wherein, In combination with the product name score of each product in the historical similar bid list and the bidirectional coverage of the product parameters, topK products are obtained as the second candidate products to be recalled, including: Based on the Embedding model, a first vector representation corresponding to the product name in the bid document is obtained, Based on the Embedding model, a second vector representation of each product name in the historical bid list library is obtained; The similarity between the first vector representation and the second vector representation is taken as the product name score of the corresponding historical product; The weighted result of the product name score of the historical product and the bidirectional coverage of the corresponding product parameters is taken as the product parameter score of the historical product; The topK historical products with the product parameter score are taken as the second candidate products to be recalled.

5. The method of claim 2, wherein, Based on the matching result of the product requirements in the bid document and the historical product requirements, the corresponding bid list preference logic is determined, including: Through the user preference agent, based on the large language model, the historical product requirements and the historical bid list are taken as prompt words, and the product requirements in the bid document are taken as input to generate the corresponding bid list preference logic.

6. The method of claim 2, wherein, Through the parameter learning agent, based on the product name and product parameters in the bid document, the new product list and the main product list are matched to obtain the third candidate product to be recalled, including: Through the parameter learning agent, based on the product name score of each product in the new product list and the bidirectional coverage of the product parameters, a new product recall product is obtained; Through the parameter learning agent, based on the heat of each product in the main product list, a hot recall product is obtained; Based on the new product recall product and the hot recall product, the third candidate product to be recalled is obtained.

7. The method of claim 1, wherein, The product name and product parameter in the bid document are obtained, including: Through optical character recognition (OCR) and / or entity recognition technology, a structured product requirement table in the bid document is obtained, including the product name and product parameter.

8. The method of claim 1, wherein, The shallow scoring network includes a two-layer fully connected network, and the method further includes: The two-layer fully connected network maps the preference score, product parameter score, and fusion score of each product, as well as the unidirectional coverage, reverse coverage, and bidirectional coverage of the product parameters of each product, to a high-dimensional semantic space for forward reasoning to output the corresponding high-dimensional semantic score. 9.A multi-agent based tender product recommendation system, characterized by, including: An information extraction module is configured to obtain the product name and product parameter in the bid document; A user preference agent is configured to determine a bid list user involved in the bid document, obtain a bid list preference logic of a recent bid list product of the bid list user, and take the recent bid list product as the first candidate product to be recalled; The parameter learning agent is configured to obtain a historical similar order corresponding to the tender document, and obtain a second candidate product to be recalled and a corresponding product parameter order logic based on a product name score and a bidirectional coverage of product parameters of each product in the historical similar order. The product parameter learning agent is configured to match a new product list and a main product list based on the product name and the product parameter in the tender document to obtain a third candidate product to be recalled. The recall module is configured to take a list of the first candidate product, the second candidate product, and the third candidate product as a recall product list. The user preference agent is further configured to score a product in the recall product list based on a preference. The parameter learning agent is further configured to score a product parameter of a product in the recall product list based on a product parameter. The fusion agent is configured to learn the order preference logic and the product parameter order logic, take the product name and the product parameter in the tender document and the product name and the product parameter of a product in the recall product list as a prompt, and generate a fusion score. The shallow scoring network is configured to perform forward reasoning based on the preference score, the product parameter score, and the fusion score of each product in the recall product list, and a unidirectional coverage, a reverse coverage, and a bidirectional coverage of the product parameter of each product to obtain a high-dimensional semantic score, and rearrange the products in the recall product list based on the high-dimensional semantic score to obtain a recommendation list.

10. An electronic device, comprising: The method comprises: a processor; a memory having computer readable instructions stored thereon, the computer readable instructions being executed by the processor to implement the method of any one of claims 1 to 8.

Images