Scientific and technical literature enhanced retrieval method and system

Abstract

The invention relates to a scientific and technical literature enhanced retrieval method and system, and belongs to the technical field of natural language processing. Comprising the following steps: acquiring a query request, and rewriting the query request into a plurality of sub-query requests; vectorizing the plurality of sub-query requests to obtain a query request vector; according to the vector, performing vector retrieval in a vector database to obtain a matched fragment; the vector database obtaining method comprises the steps that scientific and technical literatures in a query request are received, and text content of the scientific and technical literatures is extracted; performing enhancement processing on the text content to obtain an enhanced text; performing block processing on the enhanced text to obtain a plurality of block contents; carrying out vectorization processing on the plurality of block contents and then storing the block contents into a vector database; according to the matched fragments, the query request and the plurality of sub-queries, obtaining comprehensive prompt information; and inputting the comprehensive prompt information into the large language model to generate a retrieval result. According to the method, accurate and complete answer results can be provided, and the accuracy and the retrieval efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a method and system for enhanced retrieval of scientific and technological documents. Background Art

[0002] With the rapid development of science and technology, the number of scientific papers continues to increase, containing a wealth of knowledge and information. To efficiently extract relevant knowledge from these papers, RAG (Retrieval-Augmented Generation) technology has been introduced to the field of scientific paper question answering. However, in practical applications, directly applying traditional RAG methods to scientific paper question answering often encounters numerous problems, resulting in ambiguous or missing results, or even inability to obtain appropriate answers.

[0003] First, scientific papers have a clear and fixed writing framework, with different sections focusing on different emphases. For example, the introduction mainly introduces the research background and purpose, the methods section details the experimental design and operational steps, the results section presents the experimental data and phenomena, and the discussion section analyzes and summarizes the research results. However, the same statements may appear in different sections, which makes it easy to recall a large amount of noisy data when relying solely on language similarity retrieval. In addition, the entry points of different users' questions are also different, further exacerbating this problem, which seriously affects the accuracy and relevance of the generated results.

[0004] Secondly, scientific papers contain large amounts of data, much of which is presented in the form of images. These papers typically only discuss information that is strongly relevant to the topic, while potentially overlooking other underlying information or phenomena. When a user's question relates to one of these underlying issues, existing RAG technology struggles to capture valid information due to the lack of corresponding data, thus failing to provide more accurate and comprehensive results.

[0005] Finally, a single figure in a scientific paper is often composed of multiple sub-figures, each of which may depict different experimental conditions, data comparisons, or phenomenon analyses. In such cases, simply extracting information from the image is challenging. Existing RAG technology often fails to effectively extract key information relevant to the question when processing such complex images, resulting in a failure to provide users with accurate answers.

[0006] To sum up, when traditional technologies are applied to scientific paper question and answer, due to the writing characteristics of scientific papers, the way data is presented, and the existence of complex images, it is impossible to efficiently obtain relevant knowledge and it is difficult to meet users' needs for accurate and comprehensive answers. Summary of the Invention

[0007] To this end, the technical problem to be solved by the present invention is to overcome the problems of the existing technology in processing scientific paper questions and answers, such as vague answer results, missing information, and low accuracy and retrieval efficiency.

[0008] In a first aspect, to solve the above technical problems, the present invention provides a method for enhanced retrieval of scientific and technological literature, comprising:

[0009] Obtaining a query request, and classifying and rewriting the query request into multiple sub-query requests;

[0010] Vectorizing the multiple sub-query requests to obtain a query request vector;

[0011] According to the query request vector, a vector search is performed in a vector database to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database;

[0012] Obtaining comprehensive prompt information according to the matching fragments, the query request, and the multiple sub-queries;

[0013] The comprehensive prompt information is input into a large language model to generate a search result.

[0014] In one embodiment of the present invention, the steps of enhancing the text content to obtain enhanced text are:

[0015] Determine and determine the core content of the text;

[0016] According to the chapter titles of the scientific literature, the core content is segmented for the first time to obtain multiple segmentation blocks;

[0017] Acquire image information in the plurality of segmented blocks, and enhance the image information to obtain a text description corresponding to the image information;

[0018] Add the text description to the end of the corresponding image information reference paragraph to obtain a new segmentation block;

[0019] Each of the new segmented blocks is segmented for the second time to obtain enhanced text.

[0020] In one embodiment of the present invention, the steps of obtaining image information in the plurality of segmented blocks, enhancing the image information, and obtaining a text description corresponding to the image information are as follows:

[0021] Extracting a picture of the picture information and a caption corresponding to the picture;

[0022] Determine whether the image is a single image based on the caption corresponding to the image; if the image is not a single image, segment the image into multiple independent single images, and repeat the determination for each single image until all segmented images are confirmed to be single images; if the image is a single image, obtain relevant information about the image;

[0023] Inputting the relevant information into the input multimodal large model to obtain enhanced information of the image;

[0024] A text description corresponding to the image information is obtained according to the enhanced information.

[0025] In one embodiment of the present invention, enhancing the image information further includes obtaining potential information points of the image information, and the method for obtaining the potential information points is:

[0026] The image information and related information are input into a multimodal model to obtain the viewpoint expressed by the image and potential information points not explicitly stated in the text; or, a combination of multiple models is used to first input the image information and its contextual information into a multimodal model to obtain the viewpoint expressed by the image, and then the obtained information is given to other models to obtain potential information points.

[0027] In one embodiment of the present invention, the relevant information of the picture includes the label of the current segment, the paragraph citing the picture, and the title of the paragraph.

[0028] In one embodiment of the present invention, before classifying and rewriting the obtained query request, the process further includes: determining whether the query request is a single question; and when the query request is not a single question, splitting the query request into multiple single questions.

[0029] In one embodiment of the present invention, according to the query request vector, a vector search is performed in a vector database to obtain matching fragments, and the process further includes filtering out content that is not associated with the query request vector.

[0030] In one embodiment of the present invention, obtaining the comprehensive prompt information also includes obtaining a prompt template.

[0031] In a second aspect, to solve the above technical problems, the present invention provides a system for enhancing retrieval of scientific and technological literature, comprising:

[0032] A query processing module, configured to obtain a query request and classify and rewrite the query request into multiple sub-query requests;

[0033] a query vectorization module, configured to vectorize the multiple sub-query requests to obtain a query request vector;

[0034] A retrieval module is configured to perform a vector search in a vector database based on the query request vector to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database;

[0035] The result output module is used to obtain comprehensive prompt information according to the matching fragments, the query request and the multiple sub-queries; and input the comprehensive prompt information into the large language model to generate a search result.

[0036] On the third aspect, in order to solve the above technical problems, the present invention provides an APP, including the above-mentioned enhanced retrieval system for scientific and technological literature.

[0037] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0038] (1) The method and system for enhanced retrieval of scientific and technological documents described in the present invention significantly improve the accuracy of information retrieval by subdividing the query request submitted by the user into several sub-queries and performing vectorization processing on these sub-queries. This method not only enhances the flexibility of the retrieval process, enabling it to adapt to various complex retrieval requirements, but also optimizes the storage and retrieval efficiency of data by constructing a structured vector database. The construction of the vector database involves extracting, enhancing, segmenting and vectorizing the text content of scientific and technological documents, and these steps provide a solid foundation for achieving fast and accurate information retrieval. In addition, the application of vectorization processing and vector database significantly speeds up the retrieval speed, enables users to quickly obtain the required information, and enhances the system's ability to handle complex queries. The present invention not only improves the retrieval accuracy and efficiency, but also enhances the flexibility and robustness of the system, providing users with an efficient and reliable scientific and technological document retrieval solution.

[0039] (2) This invention uses a multi-path recall strategy to effectively collect accurate information from the literature modules relevant to the query while avoiding the interference of a large amount of irrelevant noise data. This method not only improves the accuracy of information retrieval, but also uses the selected relevant data to assist the model in generating more accurate and comprehensive answers.

[0040] (3) This invention effectively enhances the depth and breadth of information retrieval by feeding images and their context into a large multimodal model. This approach not only captures information explicitly expressed in the image but also uncovers potential information not explicitly mentioned in the text, thereby improving the accuracy and relevance of the information. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein

[0042] Figure 1 This is a flow chart of a method for enhancing retrieval of scientific and technological literature in a preferred embodiment of the present invention;

[0043] Figure 2 A flowchart of query request classification and rewriting in a preferred embodiment of the present invention;

[0044] Figure 3 A schematic diagram of query request classification and rewriting in a preferred embodiment of the present invention;

[0045] Figure 4 This is a flow chart of obtaining enhanced text from a vector database in a preferred embodiment of the present invention;

[0046] Figure 5 A flowchart of obtaining text description from a vector database in a preferred embodiment of the present invention;

[0047] Figure 6 A schematic diagram of the location of adding text descriptions in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0049] Example 1

[0050] Reference Figure 1 As shown, the embodiment of the present invention provides a method for enhanced retrieval of scientific and technological literature, including but not limited to the following steps:

[0051] S1. Obtain a query request and classify and rewrite the query request into multiple sub-query requests;

[0052] S2. Vectorize multiple sub-query requests to obtain a query request vector.

[0053] S3. Perform a vector search in a vector database based on the query request vector to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database;

[0054] S4. Obtaining comprehensive prompt information based on the matching fragments, the query request, and the multiple sub-queries;

[0055] S5. Input the comprehensive prompt information into the large language model to generate retrieval results.

[0056] An embodiment of the present invention provides a method for enhanced retrieval of scientific and technological literature, which significantly improves the accuracy of information retrieval by decomposing the user's query request into multiple sub-queries and performing vectorized processing on them. This method not only enhances the flexibility of the retrieval process, enabling it to adapt to diverse retrieval needs, but also optimizes data storage and retrieval efficiency by constructing a structured vector database. The process of constructing a vector database includes text content extraction, enhancement processing, block processing, and vectorized storage of scientific and technological literature, which lays the foundation for fast and accurate information extraction. In addition, the use of vectorized processing and vector databases also greatly improves the retrieval speed, enabling users to quickly obtain the required information and support the processing capabilities of complex queries. The embodiment of the present invention not only improves the accuracy and efficiency of retrieval, but also enhances the flexibility and robustness of the system.

[0057] Specifically, refer to Figure 2 As shown, in step S1, the query request initiated by the user is obtained, recorded as query, and is classified and rewritten to be decomposed into multiple sub-query requests. This process includes a preliminary determination of whether the query request query is a single question; if the query request is not a single question, it is split into multiple single questions. The specific steps are:

[0058] S110: Receive a query request input by a user.

[0059] S120: Input the query request into a large language model (LLM), and analyze the model to obtain a label list for the current query request. LLMs are deep learning models with billions to trillions of parameters that can understand and generate natural language and are widely used in tasks such as question answering, text generation, and translation.

[0060] S130. Determine the nature of the query request based on the tag list to determine whether it is a single question; if the query request is a single question, directly proceed to step S2 for further processing; if the query request includes multiple questions, execute step S140.

[0061] S140, reference Figure 3 As shown in FIG, the original query request is classified and rewritten by LLM, and is decomposed into multiple sub-query requests of single questions.

[0062] By executing steps S110 to S140, the embodiment of the present invention can effectively decompose a complex query request into multiple sub-query requests that are easier to manage and process. This process not only enhances the accuracy of the query, but also significantly improves the overall processing efficiency.

[0063] Specifically, in step S2, for the selection of vectorization processing methods, the most appropriate method can be selected according to the specific application scenario and requirements, or multiple methods can be used in combination to improve the vectorization effect and model performance.

[0064] For example, when processing queries containing complex lexical relationships, each subquery is initially vectorized using TF-IDF or a bag-of-words model. PCA is then used for dimensionality reduction to reduce the data's dimensionality while retaining the most important features. Furthermore, embedding techniques can be incorporated into the vectorization process to further enrich semantic information and improve query accuracy.

[0065] Specifically, in step S3, the specific steps of obtaining the vector database are:

[0066] S310: Receive scientific and technological documents in the query request, perform document parsing on the documents, and extract text content of the scientific and technological documents.

[0067] S320: Enhance the text content to obtain enhanced text. Figure 4 As shown, the following steps are included:

[0068] S321. Identify and determine the core content of the text, while filtering out non-core parts such as the table of contents, references, and acknowledgments.

[0069] S322. Perform a first segmentation of the core content of the scientific literature based on the chapter titles to obtain multiple segments, and assign a label to each segment and record it. The chapter titles include abstract, introduction, methods, results, etc. The label format can be M_F_i_dj, where M represents the segment; F represents the first segmentation; S represents the second segmentation; i represents the label (such as abstract, introduction, conclusion, etc.); dj represents the description of the jth image in the corresponding restriction, and d is the abbreviation of description. In this embodiment, the segmentation segment obtained by the first segmentation is labeled M_F_i.

[0070] S323, obtain the image information in the multiple segmented blocks, and enhance the image information to generate the corresponding text description M_F_i_dj. Figure 5 As shown, the specific steps are:

[0071] Step 1: Extract the image information and the corresponding caption of the image.

[0072] Step 2: To determine whether the image is a single image, first check whether its caption contains labels indicating multiple images, such as "1), 2)" or "a), b)". If the image is not a single image, use image processing techniques, such as OpenCV or self-trained models, to segment the image into multiple independent single images. Repeat this determination for each single image until all segmented images are confirmed to be single images. If the image is a single image, obtain relevant information about the image. This information includes the label of the current segment, the paragraph that references the image, and the title of the paragraph.

[0073] Step 3: Input relevant information into the multimodal model to obtain enhanced information of the image. The enhanced information includes details, main information and purpose, potential information points, etc.

[0074] Furthermore, methods for obtaining potential information points can include: using a single model to input image information and related information into a multimodal model (such as a visual language model, or VLM model) to simultaneously obtain the viewpoint expressed in the image and potential information points not explicitly stated in the text; or using a combination of multiple models to first input the image information and its context into a multimodal model (such as a VLM model) to obtain the viewpoint expressed in the image, and then pass the obtained information to other models to obtain potential information points. Among these other models, include the DeepSeek-Reasoner series and GPT-4.

[0075] Step 4: Based on the enhanced information, obtain the text description corresponding to the image information.

[0076] S324, add the text description M_F_i_dj to the end of the corresponding picture information reference paragraph, such as Figure 6 As shown, a new segmentation block M_F_i_new is obtained. When dealing with multiple subgraphs, these text descriptions M_F_i_dj are added one by one according to the order in which the subgraphs appear in the original text, ensuring that the organization and presentation of the information are consistent with the logic and structure of the original text. This approach ensures the coherence and organization of the enhanced text, facilitating subsequent information retrieval and analysis.

[0077] S325: Perform a second segmentation on each new segmented block to obtain a second segmented block M_S_i_new, which is the enhanced text. Then, the second segmented block M_S_i_new is vectorized and stored in a vector database along with metadata information for subsequent retrieval.

[0078] S330: Divide the enhanced text into blocks to obtain multiple block contents.

[0079] S340: Perform vectorization processing on the contents of the multiple blocks, and store the results in a vector database.

[0080] Through steps S310 to S340, the in-depth utilization of image information is enhanced, the information retrieval process is optimized, and the efficiency of automated processing is improved. This series of steps brings obvious advantages to the processing and utilization of scientific and technological literature. In step S320, by enhancing paper information and adopting an innovative text segmentation method, key information is effectively aggregated, making the content of the document richer and more organized. This method not only improves the accuracy and efficiency of information retrieval, but also provides convenience for users to quickly access and understand scientific and technological literature. In step S323, by inputting the image and its context information into a large multimodal model, the depth and breadth of information retrieval are effectively enhanced. This method can not only obtain the information clearly stated in the image, but also dig out potential information points that are not explicitly mentioned in the text, thereby improving the accuracy and relevance of the information. By combining multiple models, the quality of the retrieval results can be further improved to ensure that rich and accurate information is provided to users, thereby significantly improving the user experience.

[0081] Furthermore, before performing the vector search in the vector database, unnecessary chapter information is filtered out using the tag list in step S120 to obtain the chapter for vector search. A text similarity search is performed on the chapter to obtain corresponding matching segments, where the matching segments include matching documents or information segments.

[0082] Furthermore, vector databases can use metadata format for data storage. Metadata is a type of data that describes data, providing detailed information about the dataset, including its attributes, structure, and content. In practical applications, the vector database storage format can be flexibly configured to optimize data management and retrieval efficiency based on specific needs and scenarios.

[0083] Specifically, before obtaining the comprehensive prompt information (abbreviated as prompt) in step S4, a predefined prompt template is also obtained, and a suitable prompt template can be selected from a predefined template library. The following exemplary description of the prompt template is given.

[0084] For example, if you are a world-class, high-level scientific researcher, please first extract the information from the figure and possible trends as detailed and comprehensive as possible. Then, based on the information in the figure, the caption, and the paragraph where the figure is cited, analyze and summarize the main information points of the figure and the main purpose of the paper using the figure. Finally, please analyze the potential information points in the figure that are not described in the context from multiple angles as much as possible, presenting them in a scientific, rigorous, and concise manner. The returned format is JSON, including the detailed information of the figure (detail), the main information and purpose (summary), and the potential information points (potential info), such as {"detail":"","summary":"","potential info":""}; the caption is represented by {caption}; the information of the paragraph where the figure is cited is represented by {content}; the title of the paper is represented by {title}; and the heading of the section containing the paragraph that cited the figure is represented by {section headings}.

[0085] For example, if the user's query request is not a single question, the prompt template can be described as follows: Imagine you are a world-class scientific researcher who is very familiar with the structure of a paper and the key points of the corresponding text under each chapter label. Please analyze whether the user's question is a single question. If it is multiple questions, please split it into an appropriate number of single questions. Then analyze the focus of each question and infer which chapter or chapters the corresponding answer is generally located. The chapter labels are one or more of [abstract, introduction, method and result, conclusion]. Finally, the result is returned in JSON format, such as [{"query1":"","label":[]},{"query2":"","label":[]}]. User question: "What are some micro-nanostructures that can achieve asymmetric optical transmission?"

[0086] Furthermore, comprehensive prompt information is obtained based on the prompt template, matching fragments, query requests and multiple sub-queries. The purpose of this comprehensive prompt information is to provide the model with a more detailed contextual background, thereby enhancing the relevance and accuracy of the generated answers. In this embodiment, a multi-channel recall strategy is adopted to collect valid information in the corresponding document module, which significantly improves the efficiency of information recall and avoids the interference of a large amount of irrelevant noise data. In addition, by utilizing this filtered information, the model can generate more accurate and comprehensive responses. This strategy not only optimizes the information retrieval process, but also enhances the quality and reliability of the answers, providing users with more accurate retrieval results.

[0087] Specifically, in step S5, the comprehensive prompt information is input into the LLM to generate accurate search results. The generated search results are further processed and optimized before being returned to the user, ensuring that the information provided is both accurate and relevant to meet the user's needs.

[0088] The present invention proposes an optimized method for scientific and technological literature retrieval. First, the documents are preprocessed, and they are labeled and divided into blocks according to the chapter titles. Next, the pictures in each chapter are split and analyzed, and the parts that are not clearly described in the pictures are reasonably inferred and described. Then, these enhanced picture information are reasonably integrated into the corresponding positions, for example, placed after the paragraph that quotes the picture, to adapt to the characteristics of scientific and technological literature question and answer. In addition, the embodiment of the present invention also classifies and rewrites the user's query request, establishes a mapping relationship between the query and the document chapter, and realizes refined retrieval. This process not only improves the accuracy of the retrieval, but also enhances the accuracy of the large language model when generating answers. Through the above steps, the embodiment of the present invention improves the efficiency and accuracy of scientific and technological literature retrieval, and provides users with more accurate and relevant information.

[0089] Furthermore, compared to the prior art, embodiments of the present invention demonstrate significant advantages in the field of scientific paper question-answering. For example, the technical approach disclosed in patent CN119046446A involves performing multimodal branching retrieval and recall on user queries, followed by rearrangement, and generating corresponding responses using a large language model. However, this technical approach differs significantly from the method employed by the present invention and cannot achieve the high accuracy achieved by the present invention. Another prior art, patent CN119169151A, primarily achieves mixed text and image arrangement by calculating the information similarity between the generated answer and the image description. This technical implementation differs from the method of the present patent and similarly cannot achieve the high accuracy achieved by the present invention. Although patent CN119003788A also targets the field of scientific papers, its technology primarily focuses on flowcharts, and its construction scheme is completely different from that of the present invention. Therefore, its scope of application is limited and it is not suitable for the application scenarios of the present invention. Furthermore, prior art CN119441523A constructs image descriptions through a document hierarchy to achieve document-based text and image question-answering. However, the present invention surpasses this technology in terms of the completeness and comprehensiveness of the generated results. The present invention marks the paper by chapter title in advance, and then performs a refined split and analysis of the pictures in each chapter title paragraph, especially adding a reasonable reasoning description of the potential undescribed points in the picture content, and appending the picture information to the reasonable paragraphs and then dividing it into blocks to better adapt to the characteristics of question and answer in the field of scientific and technological papers. In addition, the present invention also classifies and rewrites queries for user texts, builds a mapping relationship between queries and chapters, and significantly improves the accuracy of answers generated by large language models through refined retrieval. Therefore, compared with the existing technology, the present invention can provide accurate and complete answer results when processing scientific and technological paper questions and answers, and significantly improves the accuracy and retrieval efficiency, thereby showing higher performance and effects in practical applications.

[0090] Example 2

[0091] Based on the same inventive concept, this embodiment provides an enhanced retrieval system for scientific and technological literature. The principle of solving the problem is similar to the enhanced retrieval method for scientific and technological literature provided in Example 1, and the repeated parts will not be repeated.

[0092] This embodiment provides a system for enhancing retrieval of scientific and technological literature, including:

[0093] A query processing module is used to obtain a query request and classify and rewrite the query request into multiple sub-query requests;

[0094] The query vectorization module is used to vectorize multiple sub-query requests to obtain query request vectors;

[0095] The retrieval module is used to perform vector retrieval in a vector database based on a query request vector to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database;

[0096] The result output module is used to obtain comprehensive prompt information based on the matching fragments, query requests and multiple sub-queries; and input the comprehensive prompt information into the large language model to generate retrieval results.

[0097] Example 3

[0098] This embodiment provides an APP, including the enhanced retrieval system for scientific and technological literature provided in Example 2.

[0099] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0100] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0103] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for enhancing retrieval of scientific and technological literature, characterized in that: include: Obtaining a query request, and classifying and rewriting the query request into multiple sub-query requests; Vectorizing the multiple sub-query requests to obtain a query request vector; According to the query request vector, a vector search is performed in a vector database to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database; Obtaining comprehensive prompt information according to the matching fragments, the query request, and the multiple sub-queries; The comprehensive prompt information is input into a large language model to generate a search result.

2. A method for enhanced retrieval of scientific and technological literature according to claim 1, characterized in that: The steps of enhancing the text content to obtain enhanced text are: Determine and determine the core content of the text; According to the chapter titles of the scientific literature, the core content is segmented for the first time to obtain multiple segmentation blocks; Acquire image information in the plurality of segmented blocks, and enhance the image information to obtain a text description corresponding to the image information; Add the text description to the end of the corresponding image information reference paragraph to obtain a new segmentation block; Each of the new segmented blocks is segmented for the second time to obtain enhanced text.

3. A method for enhanced retrieval of scientific and technological literature according to claim 2, characterized in that: The steps of obtaining the image information in the plurality of segmented blocks, enhancing the image information, and obtaining a text description corresponding to the image information are as follows: Extracting a picture of the picture information and a caption corresponding to the picture; Determine whether the image is a single image based on the caption corresponding to the image; if the image is not a single image, segment the image into multiple independent single images, and repeat the determination for each single image until all segmented images are confirmed to be single images; if the image is a single image, obtain relevant information about the image; Inputting the relevant information into the input multimodal large model to obtain enhanced information of the image; A text description corresponding to the image information is obtained according to the enhanced information.

4. A method for enhanced retrieval of scientific and technological literature according to claim 2, characterized in that: Enhancing the image information further includes obtaining potential information points of the image information, and the method for obtaining the potential information points is: The image information and related information are input into a multimodal model to obtain the viewpoint expressed by the image and potential information points not explicitly stated in the text; or, a combination of multiple models is used to first input the image information and its contextual information into a multimodal model to obtain the viewpoint expressed by the image, and then the obtained information is given to other models to obtain potential information points.

5. A method for enhanced retrieval of scientific and technological literature according to claim 3, characterized in that: The relevant information of the image includes the label of the current segment, the paragraph that references the image, and the title of the paragraph.

6. A method for enhanced retrieval of scientific and technological literature according to claim 1, characterized in that: Before classifying and rewriting the obtained query request, the method further includes: determining whether the query request is a single question; and when the query request is not a single question, splitting the query request into multiple single questions.

7. A method for enhanced retrieval of scientific and technological literature according to claim 1, characterized in that: According to the query request vector, a vector search is performed in a vector database to obtain matching fragments, and the process also includes filtering out content that is not related to the query request vector.

8. A method for enhanced retrieval of scientific and technological literature according to claim 1, characterized in that: Before obtaining the comprehensive prompt information, it also includes obtaining the prompt template.

9. A system for enhancing the retrieval of scientific and technological literature, characterized in that: include: A query processing module, configured to obtain a query request and classify and rewrite the query request into multiple sub-query requests; a query vectorization module, configured to vectorize the multiple sub-query requests to obtain a query request vector; A retrieval module is configured to perform a vector search in a vector database based on the query request vector to obtain matching fragments; wherein the vector database is obtained by: receiving scientific documents in the query request and extracting text content of the scientific documents; enhancing the text content to obtain enhanced text; dividing the enhanced text into blocks to obtain multiple block contents; vectorizing the multiple block contents and storing them in the vector database; The result output module is used to obtain comprehensive prompt information according to the matching fragments, the query request and the multiple sub-queries; and input the comprehensive prompt information into the large language model to generate a search result.

10. An APP, characterized in that: Including the enhanced retrieval system for scientific and technological literature as described in claim 9.

Images