Fair competition review method, device, equipment and medium based on large language model

Abstract

The present disclosure relates to a fair competition review method, device, equipment and medium based on a large language model. Among them, the fair competition review method based on a large language model includes: obtaining a document to be reviewed; determining a reference case based on the document to be reviewed, and obtaining a target review regulation; processing the document to be reviewed, the reference case and the target review regulation according to a preset instruction template to generate a target model instruction; reviewing and processing the target model instruction to obtain a corresponding review conclusion. According to the embodiments of the present disclosure, the judgment accuracy and reliability of the processed documents can be improved, providing a more comprehensive and profound review result.

Description

Technical Field

[0001] The present disclosure relates to the application field of fair competition review technology, and in particular to a fair competition review method, device, equipment and medium based on a large language model. Background Art

[0002] Fair competition review is an important part of market supervision. It aims to maintain the fairness and openness of the market by systematically evaluating and analyzing policy documents, rules and regulations to ensure that their contents do not contain inappropriate clauses that exclude or restrict competition.

[0003] In recent years, with the development of technologies such as artificial intelligence, new opportunities have emerged in the field of fair competition review. In order to improve the efficiency and accuracy of review, many researchers have begun to explore the use of modern information technology to assist or replace manual review. The existing technical solutions mainly include distributed review systems based on blockchain, methods for text classification and violation identification using deep learning models, and semantic analysis of policy documents through pre-trained language models.

[0004] However, most solutions do not fully utilize historical case data that has been manually annotated and reviewed, which results in the accuracy and reliability of the model being limited when processing and analyzing some marginal or long-tail case data. In addition, the widely used discriminant models such as BERT (Bidirectional Encoder Representations from Transformers) mainly rely on text classification and feature extraction methods. These models can usually only identify possible problems in the text, but cannot provide detailed judgment analysis and matching of specific regulations. Summary of the invention

[0005] In order to solve the above technical problems, the present disclosure provides a fair competition review method, device, equipment and medium based on a large language model.

[0006] In a first aspect, the present disclosure provides a fair competition review method based on a large language model, comprising:

[0007] Get documents for review;

[0008] Determine reference cases based on the documents to be reviewed and obtain target review regulations;

[0009] According to a preset instruction template, the to-be-examined document, the reference case and the target review regulations are processed to generate a target model instruction;

[0010] The target model instruction is reviewed and processed to obtain a corresponding review conclusion.

[0011] In a second aspect, the present disclosure provides a fair competition review device based on a large language model, comprising:

[0012] Document acquisition module, used to obtain documents to be reviewed;

[0013] A first processing module, configured to determine a reference case based on the document to be reviewed and obtain a target review regulation;

[0014] A second processing module is used to process the to-be-examined document, the reference case and the target review regulations according to a preset instruction template to generate a target model instruction;

[0015] The third processing module is used to review and process the target model instruction to obtain a corresponding review conclusion.

[0016] In a third aspect, the present disclosure provides a fair competition review device based on a large language model, comprising:

[0017] processor;

[0018] A memory for storing executable instructions;

[0019] The processor is used to read executable instructions from the memory and execute the executable instructions to implement the fair competition review method based on a large language model of the first aspect.

[0020] In a fourth aspect, the present disclosure provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor implements the fair competition review method based on a large language model of the first aspect.

[0021] Compared with the prior art, the technical solution provided by the embodiments of the present disclosure has the following advantages:

[0022] The fair competition review method, device, equipment and medium based on the large language model of the disclosed embodiment can obtain the document to be reviewed, then determine the reference case based on the document to be reviewed, and obtain the target review regulations, and then process the document to be reviewed, the reference case and the target review regulations according to the preset instruction template, generate the target model instruction, and finally review and process the target model instruction to obtain the corresponding review conclusion. Thus, according to the document to be reviewed, the reference case can be determined and the target review regulations can be obtained, and the target model instruction can be generated in combination with the preset instruction template, and finally the review and processing can be performed to obtain the review conclusion, thereby improving the judgment accuracy and reliability of the processed document and providing a more comprehensive and profound review result. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0024] Figure 1 A flowchart of a fair competition review method based on a large language model provided in an embodiment of the present disclosure;

[0025] Figure 2 A schematic diagram of the structure of a preset instruction template provided in an embodiment of the present disclosure;

[0026] Figure 3 A flowchart of another fair competition review method based on a large language model provided in an embodiment of the present disclosure;

[0027] Figure 4 A schematic diagram of the structure of a fair competition review device based on a large language model provided in an embodiment of the present disclosure;

[0028] Figure 5 A schematic diagram of the structure of a fair competition review device based on a large language model provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0030] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0031] The term "including" and its variations used herein are open inclusions, i.e., "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the following description.

[0032] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0033] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0034] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0035] In order to solve the above problems, the embodiments of the present disclosure provide a fair competition review method, device, equipment and medium based on a large language model. Figures 1 to 3 The fair competition review method based on a large language model provided in an embodiment of the present disclosure is described in detail.

[0036] Figure 1 A flowchart of a fair competition review method based on a large language model provided in an embodiment of the present disclosure is shown.

[0037] In the embodiment of the present disclosure, the fair competition review method based on the large language model can be executed by an electronic device, wherein the electronic device may include but is not limited to devices such as computer devices, cloud servers or cloud server clusters.

[0038] like Figure 1 As shown, the fair competition review method based on a large language model may include the following steps.

[0039] S110, obtaining documents to be reviewed.

[0040] In the embodiment of the present disclosure, the electronic device can obtain the document to be reviewed.

[0041] Optionally, the document to be reviewed may be a document that requires a fair competition review.

[0042] Specifically, the electronic device can obtain a document to be reviewed provided by a user that requires a fair competition review.

[0043] S120. Determine a reference case based on the document to be reviewed, and obtain target review regulations.

[0044] In an embodiment of the present disclosure, the electronic device may determine a reference case based on the document to be reviewed and obtain target review regulations.

[0045] Optionally, the reference case may be a case related to the document to be reviewed and used for supplementary reference.

[0046] Optionally, the target review regulations may be the full text of a pre-set fair competition review regulations.

[0047] Specifically, after obtaining the document to be reviewed, the electronic device can determine the relevant reference cases based on the document to be reviewed, and obtain the full text of the pre-set fair competition review regulations, that is, obtain the target review regulations.

[0048] S130. According to a preset instruction template, the document to be reviewed, the reference case and the target review regulations are processed to generate a target model instruction.

[0049] In the disclosed embodiment, the electronic device may process the document to be reviewed, the reference case and the target review regulations according to a preset instruction template to generate a target model instruction.

[0050] Optionally, the preset instruction template may be a preset template for generating model instructions.

[0051] Optionally, the target model instruction may be an instruction to instruct the model to perform review processing.

[0052] Specifically, the electronic device can input the document to be reviewed, the reference case and the target review regulations into the preset instruction template according to the preset instruction template to generate a corresponding target model instruction.

[0053] S140: review and process the target model instruction to obtain a corresponding review conclusion.

[0054] In the disclosed embodiment, the electronic device may review and process the target model instruction to obtain a corresponding review conclusion.

[0055] Optionally, the review conclusion may be the relevant review results obtained after conducting the review.

[0056] Specifically, the electronic device can review and process the target model instructions to obtain relevant review conclusions after the review.

[0057] Thus, in the disclosed embodiment, it is possible to obtain a document to be reviewed, and then determine a reference case based on the document to be reviewed, and obtain the target review regulations, and then process the document to be reviewed, the reference case, and the target review regulations according to the preset instruction template to generate a target model instruction, and finally review and process the target model instruction to obtain the corresponding review conclusion. Thus, according to the document to be reviewed, it is possible to determine a reference case and obtain the target review regulations, and generate the target model instruction in combination with the preset instruction template, and finally review and process to obtain the review conclusion, thereby improving the judgment accuracy and reliability of the processed documents and providing a more comprehensive and profound review result.

[0058] Optionally, S120 may specifically include: based on the document to be reviewed, performing vector retrieval and recall processing on a plurality of historical cases in a preset case database to obtain the reference case.

[0059] In an embodiment of the present disclosure, the electronic device may perform vector retrieval and recall processing on a plurality of historical cases in a preset case database based on the document to be reviewed to obtain the reference case.

[0060] Optionally, the preset case database can be a pre-set database including multiple historical cases. Among them, an initial seed data set is formed by preprocessing and labeling data from various open source databases, government websites, and fair competition case analysis websites. These seed data (historical cases) are strictly screened to ensure that they cover violations of various fair competition review regulations. Then, using vectorization technology, these documents are converted into high-dimensional vector representations and stored in a specially built vector database.

[0061] Specifically, the electronic device can perform vector retrieval and recall processing on multiple historical cases in a preset case database based on the document to be reviewed, that is, based on the document to be reviewed, perform data retrieval and data recall on multiple historical cases in the preset case database to obtain corresponding reference cases.

[0062] Optionally, S120 may also specifically include: generating a first vector representation corresponding to the document to be reviewed, and generating a second vector representation corresponding to each historical case in the preset case database; performing similarity calculation based on the first vector representation and the second vector representation to obtain a corresponding similarity score; and determining the historical case with the highest similarity score as the reference case.

[0063] In an embodiment of the present disclosure, the electronic device may generate a first vector representation corresponding to the document to be reviewed, and generate a second vector representation corresponding to each historical case in the preset case database.

[0064] Optionally, the first vector representation may be a vector representing the document to be reviewed.

[0065] Optionally, the second vector representation may be a vector representing a historical case.

[0066] Specifically, the electronic device may generate a first vector representation v corresponding to the document to be reviewed. D . For example, through the self-attention mechanism, the relationship between each word and other words in the document to be reviewed is calculated to capture the complex dependencies within the sentence. In addition, each layer also includes a feed-forward network to further perform nonlinear transformations on the embedding. This multi-level processing generates a series of hidden states (hidden_states), which carry rich semantic information and fully characterize the content of the document to be reviewed. In order to convert these hidden states into vector representations of fixed dimension D, the pooling technology is introduced, using the hidden state H = {h1,h2,...,h n}Perform average pooling to achieve dimension reduction and global vectorization of documents. The specific calculation process is as follows:

[0067] H = LLM (Embed (D)).

[0068]

[0069] Furthermore, the electronic device may also generate a second vector representation corresponding to each historical case in the preset case database: For example, for each historical case C in the preset case database j , using the first vector representation v D The same method generates the second vector representation

[0070] Furthermore, the electronic device may perform similarity calculation based on the first vector representation and the second vector representation to obtain a corresponding similarity score.

[0071] Optionally, the similarity score may be used to indicate the degree to which two vectors are similar.

[0072] Specifically, the electronic device can measure the relationship between the to-be-reviewed document D and each case document C based on the co-occurrence combined with the cosine method. j The similarity between them is calculated by Rouge-L index and cosine similarity respectively, and the final similarity score is obtained by weight combination. Specifically, Rouge-L mainly measures the co-occurrence relationship between two texts based on the length of the longest common subsequence (LCS), which is defined as:

[0073]

[0074] in, is the recall rate, is the precision rate, β is usually set to 1 to balance the recall rate and precision rate. The document vector to be reviewed (the first vector representation) v D With each historical case document vector (second vector representation) The cosine similarity is calculated as follows:

[0075]

[0076] Among them, "·" represents the vector dot product, and "∥·∥" represents the Euclidean norm of the vector. Finally, the similarity between two documents is obtained by weighted combination of Rouge-L similarity and cosine similarity:

[0077]

[0078] Among them, α is a weight parameter, which is used to adjust the contribution ratio of co-occurrence information and semantic information in the final similarity. According to the specific task requirements, α can be optimized through experiments or cross-validation.

[0079] Furthermore, the electronic device may determine the historical case with the highest similarity score as the reference case.

[0080] Thus, by combining the co-occurrence features and deep semantic features of the text, a multi-dimensional characterization of the similarity between documents is achieved. The Rouge-L indicator captures explicit text matching information, while the cosine similarity provides implicit semantic associations, thereby effectively improving the matching accuracy and robustness. In short, based on the documents to be reviewed provided by the user, historical cases related to the fragment are accurately retrieved and recalled from the preset case database, providing the model with rich contextual information as a supplementary reference.

[0081] Optionally, S130 may specifically include: determining the system role of the preset instruction template based on the reference case, the target review regulations and preset generation requirements; determining the user role of the preset instruction template based on the document to be reviewed; and generating the target model instruction according to the system role and the user role.

[0082] In an embodiment of the present disclosure, the electronic device may determine the system role of the preset instruction template based on the reference case, the target review regulations and the preset generation requirements.

[0083] Optionally, the preset instruction template may include a system role.

[0084] Specifically, the electronic device may use the reference case, the target review regulations and the preset generation requirements as the system role of the preset instruction template.

[0085] Furthermore, the electronic device may determine the user role of the preset instruction template based on the document to be reviewed.

[0086] Optionally, the preset instruction template may include a user role.

[0087] Specifically, the electronic device may use the document to be reviewed as a user role of a preset instruction template.

[0088] Furthermore, the electronic device may generate the target model instruction according to the system role and the user role.

[0089] Figure 2 A schematic diagram of the structure of a preset instruction template provided by an embodiment of the present disclosure is shown.

[0090] like Figure 2 As shown, the electronic device can use the reference case, the target review regulations and the preset generation requirements as the system role of the preset instruction template, and use the document to be reviewed as the user role of the preset instruction template.

[0091] Optionally, S140 may specifically include: inputting the target model instruction into a pre-trained large language model, so that the large language model reviews and processes the target model instruction to obtain the review conclusion.

[0092] In the disclosed embodiment, after obtaining the target model instruction, the electronic device can input the target model instruction into a pre-trained large language model (LLM). The pre-trained large language model can review and process the target model instruction to obtain the review conclusion.

[0093] Among them, the review conclusions include the questionable original text, the basis for judgment and the violation of regulations.

[0094] For example, let T represent the full text of the Fair Competition Review Ordinance (target review ordinance), C = {C1, C2, ..., C K} represents the Top K reference cases retrieved based on the document to be reviewed D. The instructions of the system role can be expressed as:

[0095] P System =f(T,C).

[0096] P User =f(D).

[0097] P fina1 =g(P System ,P User).

[0098] O=h(P final ).

[0099] Among them, f is used to integrate the regulation text and reference cases into an input format that can be processed by the model and is used for format preprocessing of the document to be reviewed, and g is used to convert the system role instruction P System With user role instruction P User (i.e., documents to be reviewed) to form a complete model instruction input P finak O contains the result of the identification of the questionable fragment (the original text in question), the conclusion of the judgment analysis (the basis for judgment), and the specific content of the regulations suspected of violation (violation of regulations). h represents the reasoning and generation process of the large language model, which outputs the required review conclusion by modeling the context and generating the thinking chain.

[0100] Optionally, the fair competition review method based on a large language model may also include: based on a sequence parallel training strategy, performing multi-stage extended training on the large language model to be trained to obtain the large language model.

[0101] In an embodiment of the present disclosure, the electronic device may perform multi-stage extended training on the large language model to be trained based on a sequence parallel training strategy to obtain the large language model.

[0102] Specifically, the model's initial context window is set to 8k, which means that the model can simultaneously process input text with a length of no more than 8192 tokens. Using the rotational position encoding (RoPE) extrapolation technique, the model's context window is expanded from 8k to 32k during the inference phase, thereby constructing a dataset containing 32k contexts, and directly training the model's 32k context capabilities through sequential parallelism.

[0103] RoPE(x 8k )→x 32k .

[0104] Among them, x nk Represents context input data with nk length. Based on sequence parallel technology, when the model processes text beyond the limit of its original context window, it splits the input sequence into several subsequences of moderate length. These subsequences can be processed in parallel on different computing nodes, significantly improving computational efficiency. Each subsequence retains its logical association with other subsequences, maintaining overall semantic consistency through a shared global state or memory mechanism. After completing the 32k context window training, continue to apply RoPE extrapolation to further expand the context window to 128k. Similarly, the model is retrained by constructing a dataset suitable for 128k context.

[0105] RoPE(x32k )→x 128k .

[0106] Through the above multi-stage expansion and training, the model has a stronger ability to capture global information in a larger context window. This enables it to handle the vast majority of the length of documents to be reviewed during the review process and effectively allocate attention to a wider range of information. By gradually expanding the context window, this method ensures the consistency of training and reasoning, thereby improving the analysis ability and accuracy of complex texts to be reviewed.

[0107] To ensure the accuracy and reliability of the generated data, all outputs are subject to strict manual quality inspection and review. The quality inspection process not only verifies the consistency and correctness of the data, but also continuously optimizes the model generation strategy through a feedback mechanism. Finally, the quality-inspected data is included in the preset case database to provide richer and more diverse reference information for future review tasks.

[0108] Figure 3 A flowchart of another fair competition review method based on a large language model provided in an embodiment of the present disclosure is shown.

[0109] like Figure 3 As shown, the electronic equipment department can obtain the document to be reviewed, and perform vector retrieval and recall processing on multiple historical cases in the preset case database, that is, based on the document to be reviewed, data retrieval and data recall are performed on multiple historical cases in the preset case database, so as to obtain the corresponding reference case. Then, the reference case, the target review regulations and the preset generation requirements are used as the system role (system role) of the preset instruction template, and the document to be reviewed is used as the user role (user role) of the preset instruction template, and the target model instruction is generated according to the system role and the user role. The target model instruction is input into the pre-trained large language model LLM (for example, Llama and YAYI), and the pre-trained large language model can review the target model instruction to obtain the review conclusion, which may include the questionable original text, the basis for judgment and the violation of regulations. The electronic device can also perform multi-stage extended training on the large language model to be trained based on the sequence parallel training strategy to obtain the large language model.

[0110] Figure 4 A schematic structural diagram of a fair competition review device based on a large language model provided by an embodiment of the present disclosure is shown.

[0111] like Figure 4 As shown, the fair competition review device 400 based on a large language model may include a document acquisition module 410 , a first processing module 420 , a second processing module 430 and a third processing module 440 .

[0112] The document acquisition module 410 can be used to acquire documents to be reviewed.

[0113] The first processing module 420 can be used to determine a reference case based on the document to be reviewed and obtain target review regulations.

[0114] The second processing module 430 can be used to process the document to be reviewed, the reference case and the target review regulations according to a preset instruction template to generate a target model instruction.

[0115] The third processing module 440 can be used to review and process the target model instruction to obtain a corresponding review conclusion.

[0116] Thus, in the disclosed embodiment, it is possible to obtain a document to be reviewed, and then determine a reference case based on the document to be reviewed, and obtain the target review regulations, and then process the document to be reviewed, the reference case, and the target review regulations according to the preset instruction template to generate a target model instruction, and finally review and process the target model instruction to obtain the corresponding review conclusion. Thus, according to the document to be reviewed, it is possible to determine a reference case and obtain the target review regulations, and generate the target model instruction in combination with the preset instruction template, and finally review and process to obtain the review conclusion, thereby improving the judgment accuracy and reliability of the processed documents and providing a more comprehensive and profound review result.

[0117] In some embodiments of the present disclosure, the first processing module 420 may specifically include a first processing unit.

[0118] The first processing unit may be configured to perform vector retrieval and recall processing on a plurality of historical cases in a preset case database based on the document to be reviewed, so as to obtain the reference case.

[0119] In some embodiments of the present disclosure, the first processing module 420 may further specifically include a first generating unit, a vector calculating unit and a first determining unit.

[0120] The first generating unit may be used to generate a first vector representation corresponding to the document to be reviewed, and to generate a second vector representation corresponding to each historical case in the preset case database.

[0121] The vector calculation unit may be configured to perform similarity calculation based on the first vector representation and the second vector representation to obtain a corresponding similarity score.

[0122] The first determining unit may be configured to determine the historical case with the highest similarity score as the reference case.

[0123] In some embodiments of the present disclosure, the second processing module 430 may specifically include a second generating unit, a second determining unit and a third generating unit.

[0124] The second generating unit can be used to determine the system role of the preset instruction template based on the reference case, the target review regulations and the preset generation requirements.

[0125] The second determining unit may be configured to determine the user role of the preset instruction template based on the document to be reviewed.

[0126] The third generating unit may be configured to generate the target model instruction according to the system role and the user role.

[0127] In some embodiments of the present disclosure, the third processing module 440 may specifically include a second processing unit.

[0128] The second processing unit can be used to input the target model instruction into a pre-trained large language model, so that the large language model reviews and processes the target model instruction to obtain the review conclusion.

[0129] In some embodiments of the present disclosure, the fair competition review device 400 based on a large language model may further include a model training module.

[0130] The model training module can be used to perform multi-stage extended training on the large language model to be trained based on a sequential parallel training strategy to obtain the large language model.

[0131] In some embodiments of the present disclosure, the review conclusion may include the questionable original text, the basis for the determination, and the violation regulations.

[0132] It should be noted that Figure 4 The fair competition review device 400 based on the large language model shown can execute Figures 1 to 3 The various steps in the method embodiment shown in the figure are implemented Figures 1 to 3 The various processes and effects in the method embodiment shown are not described in detail here.

[0133] Figure 5 A structural schematic diagram of a fair competition review device based on a large language model provided by an embodiment of the present disclosure is shown.

[0134] In some embodiments of the present disclosure, Figure 5 The fair competition review device based on a large language model shown may be an electronic device. Specifically, the electronic device may include but is not limited to devices such as computer devices, cloud servers or cloud server clusters.

[0135] like Figure 5As shown, the fair competition review device based on a large language model may include a processor 501 and a memory 502 storing computer program instructions.

[0136] Specifically, the processor 501 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0137] The memory 502 may include a large capacity memory for information or instructions. By way of example and not limitation, the memory 502 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a tape, or a universal serial bus (USB) drive or a combination of two or more of these. Where appropriate, the memory 502 may include a removable or non-removable (or fixed) medium. Where appropriate, the memory 502 may be inside or outside the integrated gateway device. In a particular embodiment, the memory 502 is a non-volatile solid-state memory. In a particular embodiment, the memory 502 includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (Electrically Erasable Programmable ROM, EEPROM), an electrically rewritable ROM (EAROM) or a flash memory, or a combination of two or more of these.

[0138] The processor 501 reads and executes the computer program instructions stored in the memory 502 to perform the steps of the fair competition review method based on the large language model provided in the embodiment of the present disclosure.

[0139] In one example, the fair competition review device based on a large language model may further include a transceiver 503 and a bus 504. Figure 5 As shown, the processor 501, the memory 502 and the transceiver 503 are connected via a bus 504 and communicate with each other.

[0140] The bus 504 includes hardware, software, or both. For example, but not limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses or a combination of two or more of these. Where appropriate, the bus 504 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, the present application contemplates any suitable bus or interconnect.

[0141] The embodiments of the present disclosure also provide a computer-readable storage medium, which may store a computer program. When the computer program is executed by a processor, the processor implements the fair competition review method based on a large language model provided by the embodiments of the present disclosure.

[0142] The above-mentioned storage medium may, for example, include a memory 502 of computer program instructions, and the above-mentioned instructions may be executed by the processor 501 of the fair competition review device based on the large language model to complete the fair competition review method based on the large language model provided in the embodiment of the present disclosure. Optionally, the storage medium may be a non-temporary computer-readable storage medium, for example, the non-temporary computer-readable storage medium may be a ROM, a random access memory (Random Access Memory, RAM), a compact disc read-only memory (Compact Disc ROM, CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0143] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising" is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0144] The above description is only a specific embodiment of the present disclosure, so that those skilled in the art can understand or implement the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fair competition review method based on a large language model, characterized in that: include: Get documents for review; Determine reference cases based on the documents to be reviewed and obtain target review regulations; According to a preset instruction template, the to-be-examined document, the reference case and the target review regulations are processed to generate a target model instruction; The target model instruction is reviewed and processed to obtain a corresponding review conclusion.

2. The method according to claim 1, characterized in that The determining of a reference case based on the document to be reviewed includes: Based on the document to be reviewed, vector retrieval and recall processing is performed on multiple historical cases in a preset case database to obtain the reference case.

3. The method according to claim 2, characterized in that The step of performing vector retrieval and recall processing on a plurality of historical cases in a preset case database based on the document to be reviewed to obtain the reference case includes: Generate a first vector representation corresponding to the document to be reviewed, and generate a second vector representation corresponding to each historical case in the preset case database; Performing similarity calculation based on the first vector representation and the second vector representation to obtain a corresponding similarity score; The historical case with the highest similarity score is determined as the reference case.

4. The method according to claim 1, characterized in that: The step of processing the document to be reviewed, the reference case and the target review regulations according to the preset instruction template to generate a target model instruction includes: Determining the system role of the preset instruction template based on the reference case, the target review regulations and the preset generation requirements; Determining a user role of the preset instruction template based on the document to be reviewed; The target model instruction is generated according to the system role and the user role.

5. The method according to claim 1, characterized in that The reviewing and processing the target model instruction to obtain a corresponding review conclusion includes: The target model instruction is input into a pre-trained large language model so that the large language model examines and processes the target model instruction to obtain the examination conclusion.

6. The method according to claim 5, characterized in that The method further comprises: Based on the sequential parallel training strategy, multi-stage extended training is performed on the large language model to be trained to obtain the large language model.

7. The method according to claim 1, characterized in that The review conclusion includes the questionable original text, the basis for judgment and the violation of regulations.

8. A fair competition review device based on a large language model, characterized in that: include: Document acquisition module, used to obtain documents to be reviewed; A first processing module, configured to determine a reference case based on the document to be reviewed and obtain a target review regulation; A second processing module is used to process the to-be-examined document, the reference case and the target review regulations according to a preset instruction template to generate a target model instruction; The third processing module is used to review and process the target model instruction to obtain a corresponding review conclusion.

9. A fair competition review device based on a large language model, characterized in that: include: processor; A memory for storing executable instructions; The processor is used to read the executable instructions from the memory and execute the executable instructions to implement the fair competition review method based on a large language model as described in any one of claims 1 to 7.

10. A non-volatile computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the processor implements the fair competition review method based on a large language model as described in any one of claims 1 to 7.

Images