Policy information recommendation method

Abstract

The invention discloses a policy information recommendation method, which solves the problem of low matching degree between policy recommendation and enterprise requirements in the prior art by combining the information extraction capability of a closed source large language model and the semantic understanding capability of an open source large language model, and achieves the technical effects of improving the policy recommendation accuracy and the user experience. Complex relations between enterprises and policies can be deeply understood and matched, the correlation and practicability of policy recommendation are improved, and an efficient and intelligent decision support tool is provided for science and technology enterprises and policy formulators.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information recommendation, and in particular to a policy information recommendation method. BACKGROUND

[0002] With the development of cloud computing, big data, artificial intelligence and other technologies, more and more digital service platforms and government official websites have emerged, aiming to provide customized policy information services for enterprises. These platforms use algorithmic technologies such as machine learning and natural language processing to automatically filter, classify and recommend from a large number of policy documents, improving the efficiency and accuracy of policy pushing.

[0003] Currently, the practice of recommending policies to enterprises is mostly based on traditional information dissemination channels and basic informationization means. These methods have promoted the popularization of policies to some extent, but also have some limitations. On the one hand, the policy recommendation process often relies on traditional channels such as official announcements, email notifications or industry conferences, and the information transmission efficiency and coverage of these channels are limited. Enterprises need to actively search and filter information suitable for themselves, which is time-consuming and laborious and easy to miss important policies. In addition, the push of information generally lacks personalized consideration, usually adopting a "one-size-fits-all" approach without fully considering the actual needs of different enterprises, development stages and the particularity of the industry they are in. This leads to a low matching degree between policy resources and enterprise needs, reducing the effectiveness of policy implementation and the participation of enterprises. There is still significant room for improvement in improving personalized matching, enhancing information filtering efficiency and keeping up with market and policy environment. SUMMARY

[0004] Therefore, it is necessary to propose a policy information recommendation method to solve the above problems.

[0005] A policy information recommendation method, the method comprising:

[0006] obtaining an enterprise introduction text and a policy file content; extracting enterprise information corresponding to the enterprise from the enterprise introduction text and policy information from the policy file content through a Prompt prompt template;

[0007] determining a plurality of enterprise field information corresponding to the enterprise information and a plurality of policy field information corresponding to the policy information based on a closed-source large language model;

[0008] splicing the plurality of enterprise field information to obtain an enterprise information sequence and splicing the plurality of policy field information to obtain a policy information sequence;

[0009] processing the enterprise information sequence and the policy information sequence through an open-source large language model to obtain a target enterprise feature representation and a target policy feature representation, respectively;

[0010] fusing the target enterprise feature representation through an attention mechanism to obtain an enterprise continuous feature representation, and fusing the target policy feature representation to obtain a policy continuous feature representation;

[0011] determining a corresponding enterprise information target feature representation based on the enterprise continuous feature representation, and determining a corresponding policy information target feature representation based on the policy continuous feature representation;

[0012] determining a similarity between the enterprise information and the policy information according to the enterprise information target feature representation and the policy information target feature representation; and when the similarity is greater than a preset threshold, the policy information is target detection information recommended to an enterprise corresponding to the enterprise information.

[0013] In one embodiment, the processing of the enterprise information sequence and the policy information sequence through the open-source large language model respectively to obtain the target enterprise feature representation and the target policy feature representation comprises:

[0014] encoding and processing the enterprise information sequence through an embedding layer of the open-source large language model to obtain an enterprise information semantic feature vector, and encoding and processing the policy information sequence to obtain a policy information semantic feature vector;

[0015] processing the enterprise information semantic feature vector through multiple self-attention layers of the open-source large language model in sequence to obtain the target enterprise feature representation, and processing the policy information semantic feature vector to obtain the target policy feature representation.

[0016] In one embodiment, the determining of the corresponding enterprise information target feature representation based on the enterprise continuous feature representation, and the determining of the corresponding policy information target feature representation based on the policy continuous feature representation comprises:

[0017] determining a first query matrix, a first key matrix and a first value matrix corresponding to the enterprise continuous feature representation based on the enterprise continuous feature representation; obtaining a first multi-head self-attention of the enterprise information through a multi-head self-attention mechanism in combination with the first query matrix, the first key matrix and the first value matrix; and performing a multi-head self-attention operation on the first multi-head self-attention to obtain the enterprise information target feature representation;

[0018] determining a second query matrix, a second key matrix and a second value matrix corresponding to the policy continuous feature representation based on the policy continuous feature representation; obtaining a second multi-head self-attention of the policy information through a multi-head self-attention mechanism in combination with the second query matrix, the second key matrix and the second value matrix; and performing a multi-head self-attention operation on the second multi-head self-attention to obtain the policy information target feature representation.

[0019] In one embodiment, the enterprise information, the policy information, the enterprise field information, the policy field information, the enterprise information sequence, and the policy information sequence are respectively implemented by the following expressions:

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] in, For enterprise information; For policy information; Text for introducing a company; The content of the policy document; This is a Prompt template; For enterprise field information; This refers to policy field information; It is a closed-source large language model; For enterprise information sequence; A sequence of policy information; This represents a splicing operation.

[0027] In one embodiment, the processing of the enterprise information sequence and the policy information sequence using an open-source large language model to obtain the target enterprise feature representation and the target policy feature representation is achieved through the following expression:

[0028]

[0029]

[0030] in, Representing the enterprise characteristics for each layer; Represent the policy characteristics for each layer; A self-attention layer for open-source large language models; for Length; It is the hidden state dimension of the open-source large language model; For the number of self-attention layers, there are Representing the characteristics of the target enterprise. This represents the characteristics of the target policy; when = 1, there is the following expression:

[0031]

[0032]

[0033] wherein, is the enterprise feature representation input to the first layer of the self-attention layer; is the policy feature representation input to the first layer of the self-attention layer.

[0034] In one embodiment, the fusion of the target enterprise feature representation through the attention mechanism to obtain an enterprise continuous feature representation and the fusion of the target policy feature representation to obtain a policy continuous feature representation are implemented through the following expression:

[0035]

[0036]

[0037] wherein, is the target enterprise feature representation, is the target policy feature representation; is the enterprise continuous feature representation; is the policy continuous feature representation; is the size of the low-dimensional space after mapping; , , and are all learnable parameters.

[0038] In one embodiment, the first multi-head self-attention and the second multi-head self-attention are implemented through the following expression:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] wherein, is a first query matrix; is a first key matrix; is a first value matrix; is a second query matrix; is a second key matrix; is a second value matrix; is an enterprise continuous feature representation; is a policy continuous feature representation; , , , , , are all attention head weights; is a matrix dimension; is a first initial multi-head attention; is a second initial multi-head attention; is a first multi-head self-attention; is a second multi-head self-attention.

[0050] In one embodiment, the enterprise information target feature representation, the policy information target feature representation, and the similarity are implemented by the following expressions:

[0051]

[0052]

[0053]

[0054] wherein, is an enterprise information target feature representation; is a policy information target feature representation; is a first query matrix; is a first key matrix; is a first value matrix; is a second query matrix; is a second key matrix; is a second value matrix; is an enterprise continuous feature representation; is a total weight of all attention heads for enterprise information; is a total weight of all attention heads for policy information; is a similarity.

[0055] In one embodiment, the closed-source large language model is ChatGPT-4 or Deepseek, and the open-source large language model is Qwen-7b or Llama-7b.

[0056] The present application combines the information extraction capabilities of closed-source large language models and the semantic understanding capabilities of open-source large language models, solving the problem of low matching degree between policy recommendations and enterprise needs in the prior art, achieving the technical effects of improving policy recommendation accuracy and user experience. It can better understand and match the complex relationship between enterprises and policies, improving the relevance and practicality of policy recommendations, providing an efficient and intelligent decision support tool for technology enterprises and policymakers. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0058] wherein:

[0059] Figure 1 An application environment diagram of the policy information recommendation method in one embodiment;

[0060] Figure 2 A flowchart of the policy information recommendation method in one embodiment;

[0061] Figure 3 A structural block diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0063] With the development of cloud computing, big data, artificial intelligence and other technologies, more and more digital service platforms and government official websites have emerged, dedicated to providing customized policy information services for enterprises. These platforms use algorithmic technologies such as machine learning and natural language processing to automatically filter, classify and recommend from a large number of policy documents, improving the efficiency and accuracy of policy pushing.

[0064] Currently, most methods for recommending policies to enterprises are based on traditional information dissemination channels and basic information technology. While these methods have promoted policy dissemination to some extent, they also exhibit limitations. On the one hand, the policy recommendation process often relies on traditional channels such as official announcements, email notifications, or industry conferences, which have limited information transmission efficiency and coverage. Enterprises need to actively search for and filter information applicable to themselves, which is time-consuming, labor-intensive, and prone to overlooking important policies. Furthermore, information delivery generally lacks personalized consideration, often adopting a "one-size-fits-all" approach without fully taking into account the actual needs, development stages, and industry specificities of different enterprises. This results in a low degree of matching between policy resources and enterprise needs, reducing the effectiveness of policy implementation and enterprise participation. Significant room for improvement remains in enhancing personalized matching, improving information filtering efficiency, and keeping pace with the market and policy environment. To address the aforementioned technical problems, this application provides a policy information recommendation method.

[0065] Figure 1 This is a diagram illustrating the application environment of the policy information recommendation method in one embodiment. (Refer to...) Figure 1 This policy information recommendation method is applied to a policy information recommendation system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. Terminal 110 is used to acquire enterprise introduction text and policy document content; extract enterprise information corresponding to the enterprise from the enterprise introduction text and policy information from the policy document content using a Prompt template; server 120 is used to determine multiple enterprise field information corresponding to the enterprise information and multiple policy field information corresponding to the policy information based on a closed-source large language model; concatenate multiple enterprise field information to obtain an enterprise information sequence and concatenate multiple policy field information to obtain a policy information sequence; process the enterprise information sequence and the policy information sequence respectively using an open-source large language model to obtain target enterprise feature representation and target policy feature representation; fuse the target enterprise feature representation to obtain a continuous enterprise feature representation and fuse the target policy feature representation to obtain a continuous policy feature representation; determine the corresponding enterprise information target feature representation based on the continuous enterprise feature representation; determine the corresponding policy information target feature representation based on the continuous policy feature representation; determine the similarity between the enterprise information and the policy information based on the enterprise information target feature representation and the policy information target feature representation; if the similarity is greater than a preset threshold, the policy information is target detection information recommended to the enterprise corresponding to the enterprise information.

[0066] As Figure 2 shown, in one embodiment, a policy information recommendation method is provided. The method can be applied to a terminal and a server, and the embodiment is exemplified by application to a terminal. The policy information recommendation method specifically includes the following steps:

[0067] S10: Obtain enterprise introduction text and policy file content; extract enterprise information corresponding to the enterprise from the enterprise introduction text through a Prompt prompt template and extract policy information from the policy file content ; ;

[0068] S20: Determine a plurality of enterprise field information corresponding to the enterprise information and a plurality of policy field information corresponding to the policy information based on a closed-source large language model ; ;

[0069] S30: Concatenate the plurality of enterprise field information to obtain an enterprise information sequence and concatenate the plurality of policy field information to obtain a policy information sequence ;

[0070] S40: Process the enterprise information sequence and the policy information sequence respectively through an open-source large language model to obtain a target enterprise feature representation and a target policy feature representation ; ;

[0071] S50: Fuse the target enterprise feature representation through an attention mechanism to obtain an enterprise continuous feature representation and fuse the target policy feature representation to obtain a policy continuous feature representation ;

[0072] S60: Determine the corresponding enterprise information target feature representation based on the enterprise continuous feature representation and determine the corresponding policy information target feature representation based on the policy continuous feature representation ; ;

[0073] S70: Determine a target policy information recommendation result according to the enterprise information target feature representation​​​​​​​​​​ and the policy information target feature representation determining the enterprise information and the policy information similarity ; the similarity is greater than a preset threshold, the policy information is the target detection information recommended by the enterprise corresponding to the enterprise information .

[0074] In one embodiment, for the processing of the enterprise information sequence and the policy information sequence in step S40 by the open source large language model, the target enterprise feature representation and the target policy feature representation are obtained, including:

[0075] S401: encode the enterprise information sequence by the embedding layer of the open source large language model to obtain an enterprise information semantic feature vector , and encode the policy information sequence to obtain a policy information semantic feature vector ;

[0076] S402: sequentially process the enterprise information semantic feature vector by multiple self-attention layers of the open source large language model to obtain the target enterprise feature representation , and process the policy information semantic feature vector to obtain the target policy feature representation .

[0077] In one embodiment, for the determination of the corresponding enterprise information target feature representation based on the enterprise continuous feature representation in step S60, the determination of the corresponding policy information target feature representation based on the policy continuous feature representation includes:

[0078] S601: determine the corresponding first query matrix , first key matrix and first value matrix based on the enterprise continuous feature representation ; combine the first query matrix , first key matrix and the first value matrix obtain the enterprise information the first multi-head self-attention ; and perform multi-head self-attention operation on the first multi-head self-attention to obtain the enterprise information target feature representation .

[0079] S602: determine its corresponding second query matrix , second key matrix , and second value matrix ; obtain the policy information second multi-head self-attention through multi-head self-attention mechanism combining the second query matrix , second key matrix , and second value matrix ; and perform multi-head self-attention operation on the second multi-head self-attention to obtain the policy information target feature representation .

[0080] In one embodiment, the enterprise information , the policy information , the enterprise field information , the policy field information , the enterprise information sequence , and the policy information sequence are respectively implemented by the following expressions:

[0081] (1)

[0082] (2)

[0083] (3)

[0084] (4)

[0085] (5)

[0086] (6)

[0087] wherein, is the enterprise information; is the policy information; is the enterprise introduction text; is the policy document content; ​This is a Prompt template; For enterprise field information; This refers to policy field information; It is a closed-source large language model; For enterprise information sequence; A sequence of policy information; This represents a splicing operation.

[0088] In one embodiment, the enterprise information sequence is processed using an open-source large language model. and the policy information sequence The corresponding processing yields the target enterprise feature representation. and target policy characteristics This can be achieved using the following expression:

[0089] (7)

[0090] (8)

[0091] in, Representing the enterprise characteristics for each layer; Represent the policy characteristics for each layer; A self-attention layer for open-source large language models; for Length; It is the hidden state dimension of the open-source large language model; For the number of self-attention layers, there are Representing the characteristics of the target enterprise. This represents the characteristics of the target policy; when When =1, we have the following expression:

[0092] (9)

[0093] (10)

[0094] in, The enterprise feature representation is input to the first layer of the self-attention layer; The policy feature representation is input to the first layer of the self-attention layer.

[0095] In one embodiment, the representation of the target enterprise features through an attention mechanism The fusion is performed to obtain the continuous feature representation of the enterprise. The target policy features are represented. The fusion yields a continuous feature representation of the policy. is achieved by the following expression:

[0096] (11)

[0097] (12)

[0098] wherein, is a target enterprise feature representation, is a target policy feature representation; is an enterprise continuous feature representation; is a policy continuous feature representation; is the size of the mapped low-dimensional space; , , and are all learnable parameters.

[0099] In one embodiment, the first multi-head self-attention and the second multi-head self-attention is achieved by the following expression:

[0100] (13)

[0101] (14)

[0102] (15)

[0103] (16)

[0104] (17)

[0105] (18)

[0106] (19)

[0107] (20)

[0108] (21)

[0109] (22)

[0110] wherein, is a first query matrix; is a first key matrix; is a first value matrix; is a second query matrix; is a second key matrix; is a second value matrix; is an enterprise continuous feature representation; is a policy continuous feature representation; , , , , , are attention head weights; is a matrix dimension; is a first initial multi-head attention; is a second initial multi-head attention; is a first multi-head self-attention; is a second multi-head self-attention.

[0111] In an embodiment, the enterprise information target feature representation , the policy information target feature representation , and the similarity are implemented by the following expressions:

[0112] (23)

[0113] (24)

[0114] (25)

[0115] wherein, is an enterprise information target feature representation; is a policy information target feature representation; is a first query matrix; is a first key matrix; is a first value matrix; is a second query matrix; is a second key matrix; is a second value matrix; is an enterprise continuous feature representation; is a total weight of all attention heads for enterprise information; is a total weight of all attention heads for policy information; is a similarity.

[0116] In an embodiment, the closed-source large language model is ChatGPT-4 or Deepseek; and the open-source large language model is Qwen-7b or Llama-7b.

[0117] In the information extraction phase, the Prompt template is used to guide the closed-source large language model to accurately extract key information about enterprises and policies. The Prompt template is customized based on the type and structure of the information to ensure the accuracy and relevance of the model's output. For example, for enterprise information, the Prompt template may include fields such as enterprise size, industry sector, and technology needs; for policy information, it may cover elements such as policy type, applicable scope, and support measures. Through the powerful text understanding capabilities of the closed-source large language model, combined with the guidance of the Prompt template, valuable information can be accurately extracted from a large amount of text.

[0118] In the policy recommendation phase, the open-source large language model is used for similarity calculation. First, the extracted enterprise information and policy information are converted into feature vectors. Then, based on the pre-training knowledge of the open-source large language model, the similarity between these feature vectors is calculated. The similarity calculation principle is based on the vector space model, which measures the distance or angle between feature vectors to assess their similarity. Specifically, methods such as cosine similarity can be used to calculate the cosine value between the feature vectors of enterprise information and policy information, resulting in a similarity score between 0 and 1. The higher the score, the higher the matching degree between enterprise information and policy information.

[0119] According to the similarity calculation results, the system generates a personalized policy recommendation list. The recommendation list is sorted in descending order of similarity scores to ensure that users first see the policy information that best matches their needs. In addition, other factors such as the timeliness and regional relevance of the policy can be combined to further optimize the recommendation results.

[0120] The method of the present invention combines the information extraction capabilities of the closed-source large language model and the similarity calculation advantages of the open-source large language model to achieve accurate policy information service recommendations. The present invention can better understand and match the complex relationship between enterprises and policies, improving the relevance and practicality of policy recommendations, and providing a high-efficiency, intelligent decision support tool for technology enterprises and policy makers.

[0121] The model optimization process for the present application is as follows:

[0122] Step 1: For multiple enterprise information and policy information, calculate their cosine similarity, and construct a similarity matrix. The rows of the similarity matrix represent enterprise information, and the columns represent policy information. Each element in the matrix represents the similarity between the corresponding enterprise information and policy information. Through the similarity matrix, the matching relationship between enterprise information and policy information can be visually displayed.

[0123] Step 2: Using only cosine similarity for policy recommendation makes it difficult to capture implicit conditions in policy texts (e.g., "high-tech enterprises" must meet multiple conditions such as R&D investment and patent quantity). Meanwhile, matching based on a single semantic space can easily confuse policies that are similar on the surface but unrelated in nature (e.g., "tax incentives" and "tax regulation"). To address this issue, we use a contrastive learning approach to enhance the distinction between similar policy texts. Specifically, for a given batch of policy samples , where is the size of the batch. We obtain its semantic enhanced representation through a semantic enhancement function, which is specified as follows:

[0124] (26)

[0125] where is a learnable enhancement matrix, is Gaussian noise, where , is a domain bias term. By applying the semantic enhancement function to each , we obtain . First, we set the original sample and its enhanced sample as positive samples, and the original sample and other enhanced samples in the same batch as negative samples through the InfoNCE loss function, and learn their enhanced representations, which are specified as follows:

[0126] (27)

[0127] where represents the similarity calculation operation, is the temperature coefficient of the InfoNCE loss function.

[0128] Meanwhile, for difficult samples in the same batch, i.e., similar policy texts, we innovatively designed a hard example weighting function to increase the distinction between difficult samples. Given difficult samples and , the specific formula is as follows:

[0129] (28)

[0130] where is the hard example temperature coefficient, which should be much smaller than the temperature coefficient in the InfoNCE function, mainly used to amplify the weight difference of difficult examples.

[0131] (29)

[0132] where For marginal parameters, the difficulty example similarity is forced to be at least lower than the positive example , Only punish the case of violating the marginal constraint.

[0133] Step 3: Optimize and adjust the calculated similarity to improve the accuracy and reliability of similarity calculation. Some optimization strategies can be used, such as normalization of feature vectors, setting of similarity threshold, weighted calculation of similarity, etc. For example, use Softmax function for normalization, Focal Loss loss function for similarity optimization, the specific formula is as follows:

[0134] (30)

[0135] (31)

[0136] where, is the positive sample in the th sample, is the th negative sample in the th sample, represents the exponential operation. represents the Focal Loss loss function, where and are weight factors for balancing the importance between positive and negative samples, is the probability.

[0137] Step 4: Finally, for model training, we balance the influence of different loss functions on model optimization through multi-task learning, the specific formula is as follows:

[0138] (32)

[0139] where, is a hyperparameter that controls the contribution of the basic contrast learning module, is a hyperparameter that controls the contribution of the difficult example weighted contrast learning module.

[0140] Step 5: According to the similarity calculation results, develop reasonable policy recommendation strategies. The recommended policy should consider the characteristics and needs of the enterprise, the scope of application and support of the policy, etc. to ensure the actual value and matching degree of the recommended policy to the enterprise. A variety of recommendation strategies can be used, such as similarity ranking-based recommendation, personalized recommendation based on enterprise needs, policy priority-based recommendation, etc.

[0141] Step 6: Based on the established recommendation strategy, select policy information that meets the criteria from the similarity matrix to generate a policy recommendation list. The policy recommendation list should include key information such as the policy name, release date, scope of application, and support measures, so that enterprises can quickly understand the main content and value of the policy. Display the generated policy recommendation list to enterprise users through a user interface. The user interface should be simple, clear, and easy to operate, supporting users to view, filter, and sort the recommendation results.

[0142] Figure 3 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program that, when executed by the processor, enables the processor to implement a policy information recommendation method. The internal memory may also store a computer program that, when executed by the processor, enables the processor to implement the policy information recommendation method. Those skilled in the art will understand that… Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0143] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0144] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0145] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A policy information recommendation method, characterized in that, The method includes: Obtain the company introduction text and policy document content; extract the company information from the company introduction text and the policy information from the policy document content using the Prompt prompt template; Based on a closed-source large language model, determine multiple enterprise field information corresponding to the enterprise information and multiple policy field information corresponding to the policy information; Multiple enterprise field information is concatenated to obtain an enterprise information sequence, and multiple policy field information is concatenated to obtain a policy information sequence; The target enterprise feature representation and the target policy feature representation are obtained by processing the enterprise information sequence and the policy information sequence respectively using an open-source large language model. The target enterprise feature representation is fused using an attention mechanism to obtain a continuous enterprise feature representation, and the target policy feature representation is fused to obtain a continuous policy feature representation; Based on the continuous characteristic representation of the enterprise, determine its corresponding enterprise information target characteristic representation; based on the continuous characteristic representation of the policy, determine its corresponding policy information target characteristic representation; The similarity between the enterprise information and the policy information is determined based on the target feature representation of the enterprise information and the target feature representation of the policy information; if the similarity is greater than a preset threshold, then the policy information is target detection information recommended to the enterprise corresponding to the enterprise information.

2. The policy information recommendation method according to claim 1, characterized in that, The process of processing the enterprise information sequence and the policy information sequence using an open-source large language model to obtain the target enterprise feature representation and the target policy feature representation includes: The enterprise information sequence is encoded using the embedding layer of an open-source large language model to obtain an enterprise information semantic feature vector, and the policy information sequence is encoded to obtain a policy information semantic feature vector. The target enterprise feature representation is obtained by sequentially processing the semantic feature vector of the enterprise information through multiple self-attention layers of the open-source large language model, and the target policy feature representation is obtained by processing the semantic feature vector of the policy information.

3. The policy information recommendation method according to claim 1, characterized in that, The enterprise information target feature representation is determined based on the enterprise continuous feature representation. Determining the corresponding policy information target feature representation based on the aforementioned policy continuity feature representation includes: Based on the continuous feature representation of the enterprise, determine its corresponding first query matrix, first key matrix, and first value matrix; combine the first query matrix, first key matrix, and first value matrix through a multi-head self-attention mechanism to obtain the first multi-head self-attention of the enterprise information; and perform multi-head self-attention operation on the first multi-head self-attention to obtain the target feature representation of the enterprise information. Based on the policy continuous feature representation, determine its corresponding second query matrix, second key matrix, and second value matrix; combine the second query matrix, second key matrix, and second value matrix through a multi-head self-attention mechanism to obtain the second multi-head self-attention of the policy information; and perform multi-head self-attention operation on the second multi-head self-attention to obtain the policy information target feature representation.

4. The policy information recommendation method according to claim 1, characterized in that, The enterprise information, the policy information, the enterprise field information, the policy field information, the enterprise information sequence, and the policy information sequence are respectively implemented by the following expressions: in, For enterprise information; For policy information; Introducing text for businesses; The content of the policy document; This is a Prompt template; For enterprise field information; This refers to policy field information; It is a closed-source large language model; For enterprise information sequence; A sequence of policy information; This represents a splicing operation.

5. The policy information recommendation method according to claim 2, characterized in that, The process of processing the enterprise information sequence and the policy information sequence using an open-source large language model to obtain the target enterprise feature representation and the target policy feature representation is achieved through the following expressions: in, Representing the enterprise characteristics for each layer; Represent the policy characteristics for each layer; A self-attention layer for open-source large language models; for Length; It is the hidden state dimension of the open-source large language model; For the number of self-attention layers, there are Representing the characteristics of the target enterprise. This represents the characteristics of the target policy; when When =1, the following expression holds: in, The enterprise feature representation is input to the first layer of the self-attention layer; The policy feature representation is input to the first layer of the self-attention layer.

6. The policy information recommendation method according to claim 1, characterized in that, The process of fusing the target enterprise feature representations using an attention mechanism to obtain continuous enterprise feature representations, and fusing the target policy feature representations to obtain continuous policy feature representations, is achieved through the following expression: in, Representing the characteristics of the target enterprise. This represents the characteristics of the target policy; Representing continuous characteristics of enterprises; Represented as a continuous feature of policy; It is the size of the low-dimensional space after mapping; , , and All of these are learnable parameters.

7. The policy information recommendation method according to claim 3, characterized in that, The first multi-head self-attention and the second multi-head self-attention are implemented by the following expressions: in, This is the first query matrix; This is the first bond matrix; This is the first-value matrix; This is the second query matrix; This is the second bond matrix; This is the second-value matrix; Representing continuous characteristics of enterprises; Represented as a continuous feature of policy; , , , , , All are attention head weights; Dimensions are matrices; For the first initial multi-head attention; For the second initial multi-head attention; For the first multi-head self-attention; This is the second multi-head self-attention.

8. The policy information recommendation method according to claim 7, characterized in that, The enterprise information target feature representation, the policy information target feature representation, and the similarity are implemented through the following expressions: in, Representation of enterprise information target characteristics; This represents the target characteristics of policy information; This is the first query matrix; This is the first bond matrix; This is the first-value matrix; This is the second query matrix; This is the second bond matrix; This is the second-value matrix; Representing continuous characteristics of enterprises; The total weight of all attention on enterprise information; The total weight of all attention on policy information; For similarity.

9. The policy information recommendation method according to claim 1, characterized in that, The closed-source large language model is ChatGPT-4 or Deepseek; the open-source large language model is Qwen-7b or Llama-7b.

Images