Search point generation method, system and device based on size model cooperation

Abstract

This invention discloses a method, system, and device for generating novelty search points based on a collaboration between large and small models. First, technical background text and requirement prompts are input for keyword extraction. The requirement prompts include the category or features of the novelty search keywords to be extracted. Then, based on the requirement prompts, a large model extracts novelty search keywords from the technical background text. These keywords are then filtered, and a small model generates novelty search point text from the retained keywords. Next, an evaluation classifier is used to classify and evaluate the generated novelty search point text. Novelty search point text that does not meet the requirements is rewritten by the large model based on the prompt information. Finally, novelty search point text that meets the requirements is generated. This invention achieves controllable automatic generation of novelty search point text under human-computer interaction. The generation process employs a collaborative approach between large and small models, reducing manual costs for novelty searches, improving the quality and completeness of generated novelty search points, and overcoming problems such as illusion errors that may arise from using a large model.

Description

Technical Field

[0001] This invention belongs to the field of computer technology and relates to a method, system and device for generating novelty points in information retrieval, specifically a method, system and device for generating novelty points in information retrieval based on big-small model collaboration. Background Technology

[0002] Novelty search has always been a crucial step in the iterative development of scientific research and technology, and a vital component of science and technology management and evaluation in my country. Based on literature review, it involves intelligence analysis to examine the novelty of the target project and produce a detailed summary report. Therefore, novelty search is a highly tedious and complex task, involving numerous technical points. Insufficient search completeness or a lack of systematic retrieval of the technical aspects and levels involved in the target project can easily lead to biased novelty search reports, impacting the innovation and development of the target project and ultimately wasting research resources. Therefore, in the context of human-computer interaction, designing a method to automatically, controllably, completely, and accurately generate novelty search points based on user needs has significant application value and broad development prospects.

[0003] Existing novelty search methods mainly consist of two steps. The main search approaches include subject search and category search. Subject search is the primary method for determining search terms; it uses words or phrases that reflect the content and characteristics of the novelty search target to search for relevant documents in the database. Category search, on the other hand, continues the search in the database based on the categories of things within the novelty search target, helping to categorize and organize the search objects. After determining the search terms through certain methods, a search expression needs to be constructed according to the novelty search objective. A typical search expression mainly consists of search terms combined with corresponding positional and logical operators, used to define the scope of the novelty search and control the accuracy of the detection. Therefore, novelty search personnel often need to continuously select search terms and construct search expressions to ensure that the novelty search results are as accurate and complete as possible. However, search expressions composed of positional operators, logical operators, and search terms lack the ability to express the novelty search objective and are difficult to reflect more complex relationships. Novelty points provide a semantic basis for the construction of search expressions, but currently, the construction of novelty points (search expressions) still relies to some extent on manual labor, increasing the manual cost of novelty searches. Therefore, based on the comprehensive existing research results and the analysis of relevant novelty point generation technologies, constructing a controllable novelty point generation architecture based on human-computer interaction and exploring the feasibility of the generation method in the automatic generation of novelty points has important theoretical significance and practical value. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, system, and device for generating novelty points for retrieval based on size model collaboration.

[0005] The first aspect of this invention provides a method for generating novelty points for retrieval based on size model collaboration, comprising the following steps:

[0006] Step 1: Input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or characteristics of the novelty search keywords to be extracted.

[0007] Step 2: Based on the requirements, use a large model to extract novelty search keywords from the technical background text;

[0008] Step 3: Filter the search keywords and use a small model to generate search point text for the retained keywords;

[0009] Step 4: Use an evaluation classifier to classify and evaluate the generated novelty search text; rewrite the novelty search text that does not meet the requirements by the large model based on the prompt information;

[0010] Step 5: Generate a search point text that meets the requirements.

[0011] Preferably, the requirement prompts mentioned in step 1 are used to guide the large model to generate novelty search point central words. The content includes allowing the large model to recommend corresponding novelty search point central words based on the categories or features retrieved from similar cases.

[0012] As a preferred embodiment, the large model described in step 2 consists of an encoding layer, several transformer layers, and an output layer.

[0013] The encoding layer is a linear neural network used to convert vocabulary unit indices into high-dimensional continuous vector representations;

[0014] The plurality of transformer layers are composed of preset transformer modules; each transformer module is composed of a mask-based multi-head self-attention mechanism module, a first normalization module, a forward propagation module, and a second normalization module connected in series; wherein, the output of the mask-based multi-head self-attention mechanism module and the original input are subjected to a ⊕ operation and then input to the first normalization module, and the output of the forward propagation module and the output of the first normalization module are subjected to a ⊕ operation and then input to the second normalization module; the ⊕ operation is an element-wise addition operation;

[0015] The output layer is a fully connected neural network.

[0016] Preferably, the small model described in step 3 consists of an encoding layer, several GLMBlock layers, and an output layer.

[0017] The encoding layer is a linear neural network used to convert vocabulary unit indices into high-dimensional continuous vector representations;

[0018] The plurality of GLMBlock layers are composed of stacked preset GLMBlock modules; each GLMBlock module consists of a first RMS normalization module, an attention module, a second RMS normalization module, and an MLP module; the first RMS normalization module pre-normalizes the input vector of the GLMBlock module, and its output is input to the attention module; the attention module performs context attention operation on the input vector, and the operation result is performed with the input vector that has not been pre-normalized for residual concatenation operation; the second RMS normalization module post-normalizes the residual concatenation operation result vector of the attention module, and then inputs it to the MLP module; the output vector of the MLP module and the residual concatenation operation result vector of the attention module are further subjected to residual concatenation operation before being output;

[0019] The output layer is a fully connected neural network.

[0020] Preferably, the training corpus of the small model in step 3 consists of input text and output text; the input text comes from the central words and technical background text extracted by the large model based on existing novelty points; the output text is the novelty point.

[0021] Preferably, the evaluation classifier in step 4 adopts a BERT-CNN architecture, consisting of a BERT layer and a convolutional neural network (CNN) layer connected in series. The input is the technical background text and the novelty search text generated from it. The BERT layer passes the CLS vector of the BERT layer classification head to the classifier classifier to obtain a classification of whether the novelty search information is complete. The CNN layer connects only the novelty search text part through a mask to evaluate the quality of the generated novelty search text. The output is a multi-label vector, with each dimension representing a different evaluation aspect.

[0022] Preferably, the evaluation classifier mentioned in step 4 is a trained evaluation classifier. During training, pseudo-corpus data is first constructed to train the evaluation classifier. The pseudo-corpus data is based on the abstract data of papers in the field, with complete abstracts as positive sample data and incomplete or modified abstracts as negative sample data.

[0023] A second aspect of the present invention provides a novelty point generation system for retrieval based on size model collaboration, comprising the following modules:

[0024] The input module is used to input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or characteristics of the novelty search keywords to be extracted.

[0025] The novelty search core keyword extraction module is used to extract novelty search core keywords from technical background text based on requirement prompts and a large model.

[0026] The novelty search point text generation module is used to filter novelty search keywords and generate novelty search point text from the retained keywords using a small model;

[0027] The classification and evaluation module is used to classify and evaluate the generated novelty search texts using an evaluation classifier; for novelty search texts that do not meet the requirements, the large model rewrites them based on the prompt information.

[0028] The retrieval module is used to generate search results text that meets the requirements.

[0029] A third aspect of the present invention provides a novelty point generation device for retrieval based on size model collaboration, comprising:

[0030] One or more processors;

[0031] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the novelty point generation method for retrieval based on size model collaboration.

[0032] A fourth aspect of the present invention provides a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the step of generating novelty points for retrieval based on size model collaboration.

[0033] Compared with the prior art, the beneficial effects of the present invention include:

[0034] (1) The present invention proposes a novelty point generation technology based on large and small model collaboration. It adopts a method of large model extraction combined with small model generation to overcome the defects such as illusion errors that may be caused by large models, realizes controllable text generation under human-computer interaction, and completes the automatic generation of novelty points for the target project.

[0035] (2) This invention combines the extraction and rewriting capabilities of large models with the subdivided domain generation capabilities of small models to achieve controllable novelty search point text generation under human-computer interaction, better understand and adapt to users' novelty search needs, reduce the cost of manually constructing novelty search points, and improve the completeness of novelty search results.

[0036] (3) This invention utilizes the implicit knowledge graph within the large model, which can extend knowledge based on the user's novelty search needs, help to mine multi-level related novelty search information, expand the scope of novelty search, and improve the work of innovation assessment. Attached Figure Description

[0037] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0038] Figure 1 This is a schematic diagram of the method according to an embodiment of the present invention;

[0039] Figure 2 This is a flowchart of a method according to an embodiment of the present invention;

[0040] Figure 3 This is a large model architecture diagram of an embodiment of the present invention;

[0041] Figure 4 This is a diagram of the transformer module architecture within a large model according to an embodiment of the present invention;

[0042] Figure 5 This is a small model architecture diagram of an embodiment of the present invention;

[0043] Figure 6 This is a diagram of the evaluation classifier architecture according to an embodiment of the present invention. Detailed Implementation

[0044] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] This embodiment analyzes the inherent defects in the novelty search field and the characteristics of large models in the deep learning field. Combining human-computer interaction and controllable text generation technology, it formulates a novelty point generation technology based on collaboration between large and small models. Under the guidance of user prompts, the large model extracts novelty search keywords from the background text of the target technology. After being filtered by the user, the model, fine-tuned with a specialized corpus, generates novelty points according to the search keywords selected by the user. In addition, this scheme also builds a novelty point quality assessment classifier to rewrite novelty points that do not meet the quality standards.

[0046] Please see Figure 1 and Figure 2 This embodiment provides a method for generating novelty points for retrieval based on size model collaboration, including the following steps:

[0047] Step 1: Input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or features of the novelty search core words to be extracted, so that the large model can search and recommend based on similar cases.

[0048] In one implementation, the demand suggestion is used to guide the large model to generate novelty search point keywords, including allowing the large model to recommend corresponding novelty search point keywords based on the categories or features retrieved from similar cases.

[0049] Step 2: Based on the requirements, use a large model to extract novelty search keywords from the technical background text by facets;

[0050] Please see Figure 3 and Figure 4 In one implementation, the large model consists of an encoding layer, several transformer layers, and an output layer.

[0051] The encoding layer is a linear neural network used to convert vocabulary unit indices into high-dimensional continuous vector representations;

[0052] The plurality of transformer layers are composed of preset transformer modules; each transformer module is composed of a mask-based multi-head self-attention mechanism module, a first normalization module, a forward propagation module, and a second normalization module connected in series; wherein, the output of the mask-based multi-head self-attention mechanism module and the original input are subjected to a ⊕ operation and then input to the first normalization module, and the output of the forward propagation module and the output of the first normalization module are subjected to a ⊕ operation and then input to the second normalization module; the ⊕ operation is an element-wise addition operation;

[0053] The output layer is a fully connected neural network.

[0054] For example, a power company wants to improve its material procurement and supply level and needs to understand the existing technologies in related fields, thus requiring a technology novelty search. Based on this method, the first step is to extract core keywords. The power company combines the technical background text with demand prompts to input into the large model. The input text for the large model is: "To respond to the State Grid's requirements for intensive material management and improve the level of material management, a power company needs to further improve the management level of centralized procurement. Expanding the scope of centralized procurement on the secondary platform and further sorting out procurement and supply strategies under various material demand characteristics will help further promote the company's intensive material management work and achieve the company's material management strategic goals." This project studies the company's existing material demand categories based on the framework of intensive material management, establishes a matching model between influencing factors and strategies, and formulates a research plan and implementation method for unified procurement and supply strategies for different material demand categories. It conducts in-depth research on the demand characteristics of different materials and establishes corresponding unified standard material supply strategies, creating a set of strategies covering material demand, planning, procurement methods, contracts, warehousing, and distribution. Standardized solutions, such as on-site services, are applied to the establishment of internal service standards and evaluation systems within the enterprise. This forms a relationship model between power materials and procurement and supply strategies (demand planning, procurement methods, warehousing management, material distribution, etc.), clarifying different supply models to unify and standardize the company's material supply standards and processes. The material procurement and supply strategy matching model, procurement and supply strategies for typical materials, and information system requirements designed in this project have been accepted by the business department, delivered to the system development department, and research work has been completed. In the next stage, in conjunction with subsequent system upgrades, the project's relevant results will be integrated into the ERP and e-commerce platforms. Please extract 10 key terms from the above text from the perspective of improving the level of material procurement and supply. The extracted key terms are as follows: State Grid, centralized material management, a power company, centralized procurement, secondary platform, procurement strategy, supply strategy, material demand characteristics, improved management level, supply model.

[0055] Step 3: Filter the search center words and use the fine-tuned GLM2 mini-model to generate search point text for the retained center words;

[0056] Please see Figure 5 In one implementation, the small model consists of an encoding layer, several GLMBlock layers, and an output layer;

[0057] The encoding layer is a linear neural network, where the weight matrix is ​​a trained word embedding matrix;

[0058] The plurality of GLMBlock layers are composed of stacked preset GLMBlock modules; each GLMBlock module consists of a first RMS normalization module, an attention module, a second RMS normalization module, and an MLP module; the first RMS normalization module pre-normalizes the GLMBlock input vector, and its output is input to the attention module; the attention module performs context attention operation on the input vector, and the operation result is residually concatenated with the input vector that has not been pre-normalized; the second RMS normalization module post-normalizes the residual concatenation result vector of the attention module, and inputs it to the MLP module; the output vector of the MLP module and the output vector of the residual concatenation operation of the attention module are further residually concatenated before being output;

[0059] The output layer is a fully connected neural network.

[0060] In one implementation, the training corpus of the small model consists of input text and output text;

[0061] The input text comes from the core words and technical background text extracted using a large model based on existing novelty search points; the output text is the novelty search point.

[0062] For example, after obtaining the core keywords for novelty searches, the power company selects three core keywords (centralized procurement, material demand characteristics, and supply model) and submits them to the small model ChatGLM3 to generate corresponding novelty search points. The prompt information entered by the user is: "Technical background text: In response to the State Grid's requirements for intensive material management and to improve the level of material management, a certain power company needs to further improve the management level of centralized procurement. Expanding the scope of centralized procurement on the secondary platform and further sorting out the procurement and supply strategies under various material demand characteristics will help further promote the company's intensive material management work and achieve the company's strategic goals for material management. This project studies the company's existing material demand categories based on the framework of intensive material management, establishes a matching model between influencing factors and strategies, and formulates a research plan and implementation method for unified procurement and supply strategies for different material demand category standards. It conducts in-depth research on the current demand characteristics of different materials and builds..." Establish a unified standard for material supply strategies, creating a set of standardized solutions covering material demand, planning, procurement methods, contracts, warehousing, distribution, and on-site services. These solutions will be applied to the establishment of internal service standards and evaluation systems, forming a relationship model between power materials and procurement and supply strategies (demand planning, procurement methods, warehousing management, material distribution, etc.). Different supply models will be clearly defined to standardize the company's material supply standards and processes. The material procurement and supply strategy matching model, procurement and supply strategies for typical materials, and information system requirements designed for this project have been accepted by the business department, delivered to the system development department, and research work has been completed. In the next phase, in conjunction with subsequent system upgrades, the project's results will be integrated into the ERP and e-commerce platforms. Keyword 1: Centralized procurement; Keyword 2: Material demand characteristics; Keyword 3: Supply model. Please generate a novelty search sentence for each key word based on the technical background text. The resulting novelty search points are:

[0063] 1. The management level of centralized procurement needs to be further improved;

[0064] 2. Develop research plans and implementation methods for unified procurement and supply strategies based on standards for different material demand categories;

[0065] 3. Establish a set of standard solutions covering material requirements, planning, procurement methods, contracts, warehousing, distribution, and on-site services, and apply them to the establishment of internal service standards and evaluation systems within the enterprise.

[0066] In one implementation, the small model can be continuously learned and updated to understand and handle emerging technological directions and their details. The performance of the current small model is periodically evaluated, and it is instructed to attempt to describe newer patented technologies to understand its limitations. The small model is then trained and updated using the latest patent texts, and its performance is tested before being deployed to the overall system.

[0067] Step 4: Construct pseudo-corpus data to train an evaluation classifier, and use the evaluation classifier to classify and evaluate the generated novelty search texts; for novelty search texts that do not meet the requirements, the large model rewrites them based on the prompt information;

[0068] Please see Figure 6 In one implementation, the evaluation classifier adopts a BERT-CNN architecture, consisting of a BERT layer and a convolutional neural network (CNN) layer connected in series. The input is technical background text and novelty search text generated therefrom. The BERT layer passes the CLS vector of the BERT layer classification head to the classifier classifier to obtain a classification of whether the novelty search information is complete. The CNN layer connects only the novelty search text part through a mask to evaluate the quality of the generated novelty search text. The output is a multi-label vector, with each dimension representing a different evaluation aspect.

[0069] In one implementation, to evaluate the quality of generated novelty search points, key sentences are extracted from abstract data of papers in the field using the GPT-4 model. Complete key sentences are used as positive sample data, while modified key sentences (added, deleted, or altered) are used as negative sample data to construct a pseudo-corpus. A fine-tuned BERT-CNN model is then trained on this data using a classification task to serve as an evaluation classifier for the novelty search point text. Simultaneously, this evaluation classifier is periodically assessed for its review capabilities and updated with new training to ensure the quality of its evaluation and review process.

[0070] In one implementation, a BERT-CNN-based evaluation classifier is used to classify and evaluate the generated novelty points. Unqualified novelty point texts are returned to the user, who then adjusts the prompts and inputs them into a large model to rewrite the novelty points.

[0071] For example, if the evaluation classifier deems the first novelty search point generated above unqualified, the system returns the novelty search point to the power company. The power company then adjusts the prompt based on this novelty search point, changing it to: "Technical Background Text: In response to the State Grid's requirements for intensive material management and to improve material management, a power company needs to further enhance its centralized procurement management. Expanding the scope of centralized procurement on the secondary platform and further streamlining procurement and supply strategies under various material demand characteristics will help further advance the company's intensive material management work and achieve its material management strategic goals." This project studies the company's existing material demand categories within the framework of intensive material management, establishes a matching model between influencing factors and strategies, and develops a research plan and implementation method for unified procurement and supply strategies for different material demand categories. It conducts in-depth research on the current demand characteristics of different materials and establishes corresponding unified standard material supply strategies. It establishes a set of standard solutions covering material demand, planning, procurement methods, contracts, warehousing, distribution, and on-site services, applicable to enterprises. The establishment of internal service standards and evaluation systems, and the formation of a relationship model between power materials and procurement and supply strategies (demand planning, procurement methods, warehousing management, material distribution, etc.), clarify different supply models to unify and standardize the company's material supply standards and processes. The material procurement and supply strategy matching model, procurement and supply strategies for typical materials, and information system requirements designed for this project have been accepted by the business department, delivered to the system development department, and research work has been completed. In the next stage, in conjunction with subsequent system upgrades, the project's relevant results will be integrated into the ERP and e-commerce platforms. Keyword: Centralized Procurement. The original novelty search text was "The management level of centralized procurement needs to be further improved." This novelty search did not reflect the details of "centralized procurement." Please rewrite the original novelty search sentence (within 20 characters) based on the technical background text and key words. The rewritten novelty search is: "Improving centralized procurement, responding to the State Grid's intensive management, optimizing procurement strategies, promoting material centralization, achieving management goals, and ensuring standardized supply." This novelty search text passed the evaluation classifier check.

[0072] Step 5: Generate a search point text that meets the requirements.

[0073] This embodiment also provides a novelty point generation system for retrieval based on size model collaboration, including the following modules:

[0074] The input module is used to input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or characteristics of the novelty search keywords to be extracted.

[0075] The novelty search core keyword extraction module is used to extract novelty search core keywords from technical background text based on requirement prompts and a large model.

[0076] The novelty search point text generation module is used to filter novelty search keywords and generate novelty search point text from the retained keywords using a small model;

[0077] The classification and evaluation module is used to classify and evaluate the generated novelty search texts using an evaluation classifier; for novelty search texts that do not meet the requirements, the large model rewrites them based on the prompt information.

[0078] The retrieval module is used to generate search results text that meets the requirements.

[0079] This embodiment also provides a novelty point generation device for retrieval based on size model collaboration, including:

[0080] One or more processors;

[0081] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the novelty point generation method for retrieval based on size model collaboration.

[0082] This embodiment also provides a readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the step of generating novelty points for retrieval based on size model collaboration.

[0083] This invention combines the extraction and rewriting capabilities of a large model with the domain-specific generation capabilities of a small model to achieve controllable novelty search point text generation under human-computer interaction. This better understands and adapts to users' novelty search needs, reduces the cost of manually constructing novelty search points, and improves the completeness of novelty search results. This method is executed by an electronic device, which can be a server or a computer. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0084] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0085] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A novelty point generation method for retrieval based on size model collaboration, characterized in that, Includes the following steps: Step 1: Input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or characteristics of the novelty search keywords to be extracted. Step 2: Based on the requirements, use a large model to extract novelty search keywords from the technical background text; The large model consists of an encoding layer, several transformer layers, and an output layer. The encoding layer is a linear neural network used to convert vocabulary unit indices into high-dimensional continuous vector representations; The plurality of transformer layers are composed of preset transformer modules; each transformer module is composed of a mask-based multi-head self-attention mechanism module, a first normalization module, a forward propagation module, and a second normalization module connected in series; wherein, the output of the mask-based multi-head self-attention mechanism module and the original input are subjected to a ⊕ operation and then input to the first normalization module, and the output of the forward propagation module and the output of the first normalization module are subjected to a ⊕ operation and then input to the second normalization module; the ⊕ operation is an element-wise addition operation; The output layer is a fully connected neural network; Step 3: Filter the search keywords and use a small model to generate search point text for the retained keywords; The small model consists of an encoding layer, several GLMBlock layers, and an output layer. The encoding layer is a linear neural network, where the weight matrix is ​​a trained word embedding matrix. The plurality of GLMBlock layers are composed of stacked preset GLMBlock modules; each GLMBlock module consists of a first RMS normalization module, an attention module, a second RMS normalization module, and an MLP module; the first RMS normalization module pre-normalizes the input vector of the GLMBlock module, and its output is input to the attention module; the attention module performs context attention operation on the input vector, and performs residual concatenation operation on the result with the input vector that has not been pre-normalized; the second RMS normalization module post-normalizes the residual concatenation operation result vector of the attention module, and inputs it to the MLP module; the output of the MLP module and the output of the residual concatenation operation of the attention module are further subjected to residual operation before being output; The output layer is a fully connected neural network; Step 4: Use an evaluation classifier to classify and evaluate the generated novelty search text; rewrite the novelty search text that does not meet the requirements by the large model based on the prompt information; Step 5: Generate a search point text that meets the requirements.

2. The novelty point generation method for retrieval based on size model collaboration according to claim 1, characterized in that: The requirement prompts mentioned in step 1 are used to guide the large model to generate novelty search point central words. The content includes allowing the large model to recommend corresponding novelty search point central words based on the categories or features retrieved from similar cases.

3. The novelty point generation method for retrieval based on size model collaboration according to claim 1, characterized in that: The small model described in step 3 has a training corpus consisting of input text and output text; The input text comes from the core words and technical background text extracted using a large model based on existing novelty search points; the output text is the novelty search point.

4. The novelty point generation method for retrieval based on size model collaboration according to claim 1, characterized in that: The evaluation classifier described in step 4 adopts a BERT-CNN architecture, consisting of a BERT layer and a convolutional neural network (CNN) layer connected in series. The input is the technical background text and the novelty search text generated based on it. The BERT layer passes the CLS vector of the BERT layer classification head to the classifier classifier to obtain a classification of whether the novelty search information is complete. The CNN layer connects only the novelty search text part through a mask to evaluate the quality of the generated novelty search text. The output is a multi-label vector, with each dimension representing a different evaluation aspect.

5. The method for generating novelty points for retrieval based on size model collaboration according to any one of claims 1-4, characterized in that: The evaluation classifier mentioned in step 4 is a trained evaluation classifier. During training, pseudo-corpus data is first constructed to train the evaluation classifier. The pseudo-corpus data is based on the abstract data of papers in the field. Complete abstracts are used as positive sample data, while incomplete or modified abstracts are used as negative sample data.

6. A novelty point generation system for retrieval based on size model collaboration, characterized in that, Includes the following modules: The input module is used to input the technical background text and requirement prompts for extracting keywords. The requirement prompts include the category or characteristics of the novelty search keywords to be extracted. The novelty search core keyword extraction module is used to extract novelty search core keywords from technical background text based on requirement prompts and a large model. The large model consists of an encoding layer, several transformer layers, and an output layer. The encoding layer is a linear neural network used to convert vocabulary unit indices into high-dimensional continuous vector representations; The plurality of transformer layers are composed of preset transformer modules; each transformer module is composed of a mask-based multi-head self-attention mechanism module, a first normalization module, a forward propagation module, and a second normalization module connected in series; wherein, the output of the mask-based multi-head self-attention mechanism module and the original input are subjected to a ⊕ operation and then input to the first normalization module, and the output of the forward propagation module and the output of the first normalization module are subjected to a ⊕ operation and then input to the second normalization module; the ⊕ operation is an element-wise addition operation; The output layer is a fully connected neural network; The novelty search point text generation module is used to filter novelty search keywords and generate novelty search point text from the retained keywords using a small model; The small model consists of an encoding layer, several GLMBlock layers, and an output layer. The encoding layer is a linear neural network, where the weight matrix is ​​a trained word embedding matrix. The plurality of GLMBlock layers are composed of stacked preset GLMBlock modules; each GLMBlock module consists of a first RMS normalization module, an attention module, a second RMS normalization module, and an MLP module; the first RMS normalization module pre-normalizes the input vector of the GLMBlock module, and its output is input to the attention module; the attention module performs context attention operation on the input vector, and performs residual concatenation operation on the result with the input vector that has not been pre-normalized; the second RMS normalization module post-normalizes the residual concatenation operation result vector of the attention module, and inputs it to the MLP module; the output of the MLP module and the output of the residual concatenation operation of the attention module are further subjected to residual operation before being output; The output layer is a fully connected neural network; The classification and evaluation module is used to classify and evaluate the generated novelty search texts using an evaluation classifier; for novelty search texts that do not meet the requirements, the large model rewrites them based on the prompt information. The retrieval module is used to generate search results text that meets the requirements.

7. A novelty point generation device for retrieval based on size model collaboration, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the novelty point generation method for retrieval based on size model collaboration as described in any one of claims 1 to 5.

8. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the step of generating novelty points for retrieval based on size model collaboration as described in any one of claims 1 to 5.

Images