A scientific research information retrieval matching method and device based on a large model

Abstract

The application discloses a scientific research information retrieval matching method and device based on a large model, and the method comprises the following steps: collecting to-be-retrieved scientific research text information; the to-be-retrieved scientific research text information comprises a plurality of search terms; the to-be-retrieved scientific research text information is processed to obtain a search graph; and the search graph is subjected to search processing by using a preset scientific research knowledge graph to obtain search matching result information.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent processing of text data and the field of equipment system evaluation, and in particular to a scientific research information retrieval matching method and device based on a large model. BACKGROUND

[0002] In the current field of scientific research information retrieval, the traditional retrieval method faces the following technical problems:

[0003] Insufficient semantic understanding: existing technologies rely on keyword matching or simple vector representation, making it difficult to capture complex semantic associations in scientific research texts (such as term synonymy and domain context dependence), resulting in semantic deviations between retrieval results and user needs. For example, although BERT-based vectorization can handle some semantics, it lacks modeling of structured relationships between search terms.

[0004] Data noise interference: Non-relevant terms, format errors, or domain-specific words (such as general words mixed into professional literature) often exist in scientific research texts, and traditional preprocessing methods (such as simple word frequency filtering) cannot accurately remove noise, resulting in the introduction of redundant information during retrieval graph construction, affecting matching accuracy.

[0005] Low efficiency of multi-modal retrieval fusion: Existing retrieval systems mostly use a single search strategy (such as relying only on semantic search or vector matching), which cannot balance retrieval speed and accuracy. For example, embedded vector search is efficient but lacks semantic reasoning ability, and RAG method can combine knowledge base but has slow response speed.

[0006] Rough relationship modeling: In traditional knowledge graph construction, relationships between entities are mostly based on predefined rules or simple similarity calculations, making it difficult to dynamically capture deep associations implied in scientific research texts (such as the decay of cross-domain term relevance and the context differentiation of polysemous words), resulting in insufficient structural expression of the retrieval graph.

[0007] The above technical problems result in low retrieval accuracy and efficiency of existing scientific research information retrieval methods. SUMMARY

[0008] The present application mainly solves the problems of insufficient semantic understanding, data noise interference, low efficiency of multi-modal retrieval fusion, and rough relationship modeling in existing scientific research information retrieval methods. The present application discloses a scientific research information retrieval matching method and device based on a large model.

[0009] In a first aspect of an embodiment of the present application, a scientific research information retrieval matching method based on a large model is disclosed, comprising:

[0010] S1, collecting scientific research text information to be retrieved; the scientific research text information to be retrieved includes a plurality of search terms;

[0011] S2, processing the to-be-retrieved scientific research text information to obtain a retrieval graph;

[0012] S3, using a preset scientific research knowledge graph to perform retrieval processing on the retrieval graph to obtain retrieval matching result information.

[0013] The processing of the to-be-retrieved scientific research text information to obtain a retrieval graph comprises:

[0014] S21, preprocessing the to-be-retrieved scientific research text information to obtain preprocessed embedding vector information;

[0015] S22, constructing a knowledge graph based on the preprocessed embedding vector information to obtain a retrieval graph.

[0016] The preprocessing of the to-be-retrieved scientific research text information to obtain preprocessed embedding vector information comprises:

[0017] S211, performing text vectorization representation on the to-be-retrieved scientific research text information to obtain a text embedding vector;

[0018] S212, performing data cleaning processing on the text embedding vector to obtain first text information;

[0019] S213, performing category checking processing on the first text information to obtain preprocessed embedding vector information.

[0020] The constructing of a knowledge graph based on the preprocessed embedding vector information to obtain a retrieval graph comprises:

[0021] S221, using a preset retrieval word grouping rule to group all retrieval words of the preprocessed embedding vector information to obtain a plurality of grouped word sets; the grouped word set comprises a plurality of retrieval words;

[0022] S222, performing relevance calculation on each grouped word set and a corresponding word library to obtain a relevance calculation result;

[0023] S223, taking a grouped word set with a relevance calculation result greater than a preset relevance threshold as entity information of a retrieval graph;

[0024] S224, calculating a relationship value between each entity information;

[0025] S225, constructing a relationship matrix using the relationship value between each entity information; an element of the i-th row and the j-th column of the relationship matrix is a relationship value between an i-th retrieval word in a grouped word set corresponding to a first entity information and a j-th retrieval word in a grouped word set corresponding to a second entity information, which is calculated;

[0026] S226, edge information calculation is performed on the relation matrix to obtain edge information between entity information of the search graph;

[0027] S227, a search graph is constructed by using all the entity information and edge information;

[0028] The expression of the edge information calculation is:

[0029]

[0030] Wherein, W N1,M1 (x) is a Whitakker function, x is an input value of the Whitakker function, N1 and M1 are respectively a row dimension and a column dimension of the relation matrix, G ij is an element of the i-th row and the j-th column of the relation matrix, σ ij is an element of the i-th singular vector obtained by singular value decomposition of the relation matrix, σ i0 is a mean value of the i-th singular vector, τ i is an i-th singular value obtained by singular value decomposition of the relation matrix, τ0 is a mean value of all singular values, and b is the edge information.

[0031] The search graph is searched in the preset scientific research knowledge graph by using a knowledge graph semantic search method to obtain first search result information.

[0032] S31, the search graph is searched in the preset scientific research knowledge graph by using a knowledge graph semantic search method to obtain first search result information.

[0033] S32, the search graph is searched in the preset scientific research knowledge graph by using an embedding vector search method to obtain second search result information.

[0034] S33, the search graph is searched in the preset scientific research knowledge graph by using an RAG search method to obtain third search result information.

[0035] S34, all the search result information is fused and matched to obtain search matching result information.

[0036] The search matching result information is obtained by fusing and matching all the search result information, including:

[0037] S341, all the search result information is text-aligned to obtain aligned search result information.

[0038] S342, an important coefficient value is calculated for each aligned search result information.

[0039] S343, using the important coefficient value, all the aligned search result information is fused and calculated to obtain the retrieval matching result information.

[0040] The expression of the important coefficient calculation is:

[0041]

[0042] Wherein, zy is the important coefficient value, beta i is the i-th entity information of the retrieval graph, lambda i is the average value of all edge information of the i-th entity information of the retrieval graph, M2 is the total number of entity information of the retrieval graph, |alpha| represents the modulus value of the aligned search result information alpha, and cos(alpha, beta i ) represents the cosine similarity calculation result.

[0043] The expression of the fusion calculation is:

[0044]

[0045] Wherein, epsilon j is the j-th element of the retrieval matching result information, alpha ij is the j-th element of the i-th search result information after alignment.

[0046] In a second aspect of the embodiment of the present application, a scientific research information retrieval matching device based on a large model is disclosed, and the device comprises:

[0047] A memory in which an executable program code is stored;

[0048] A processor coupled with the memory;

[0049] The processor calls the executable program code stored in the memory to execute the scientific research information retrieval matching method based on a large model.

[0050] In a third aspect of the embodiment of the present application, a computer storage medium is disclosed, and the computer storage medium stores computer instructions, which are called by a computer to execute the scientific research information retrieval matching method based on a large model.

[0051] In a fourth aspect of the embodiment of the present application, an information data processing terminal is disclosed, which is used to implement the scientific research information retrieval matching method based on a large model.

[0052] The present application has the following beneficial effects:

[0053] The application improves the purity and semantic representation ability of input data by pre-training a language model (such as BERT) for text vectorization, combining data cleaning (smooth noise, deleting outliers) and category checking (filtering irrelevant words based on a scientific text library), and laying a high-quality foundation for subsequent graph construction.

[0054] The application dynamically filters high-correlation entities based on correlation calculation (integrating Jaccard similarity and Jaro-Winkler distance) of the grouping word set and the word library, and accurately depicts the semantic association between entities through relationship matrix and edge information calculation (introducing Whitaker function to process singular value decomposition results), so that the search graph can be closer to the knowledge structure of the scientific field.

[0055] The application combines knowledge graph semantic search, embedded vector search (supporting brute force / tree structure index) and RAG search, takes into account semantic understanding, search efficiency and knowledge base reasoning ability, and improves the matching accuracy and recall rate in complex scientific scenarios by weighted fusion of multi-source results through important coefficients.

[0056] The application introduces decision factors and constant factors in relationship value calculation, dynamically adjusts the relationship weight between entities according to the maximum correlation, adapts to the semantic distribution characteristics of scientific texts in different fields, and enhances the model generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The application method is an implementation flowchart. DETAILED DESCRIPTION

[0058] In order to better understand the content of the application, an embodiment is given here.

[0059] Figure 1 The application method is an implementation flowchart.

[0060] In the first aspect of the embodiment of the application, a scientific information retrieval matching method based on a large model is disclosed, comprising:

[0061] S1, collecting to obtain the scientific text information to be searched;The scientific text information to be searched includes a plurality of search words;

[0062] S2, processing the scientific text information to be searched to obtain a search graph;

[0063] S3, using a preset scientific knowledge graph to process the search graph to obtain a search matching result information;

[0064] The scientific text information to be searched is processed to obtain a search graph, comprising:

[0065] S21, pre-processing the to-be-retrieved scientific research text information to obtain pre-processed embedded vector information;

[0066] S22, constructing a knowledge graph for the pre-processed embedded vector information to obtain a retrieval graph;

[0067] The pre-processing of the to-be-retrieved scientific research text information to obtain pre-processed embedded vector information comprises:

[0068] S211, text vectorization representation of the to-be-retrieved scientific research text information to obtain text embedded vector;

[0069] S212, data cleaning processing of the text embedded vector to obtain first text information;

[0070] S213, category checking processing of the first text information to obtain pre-processed embedded vector information.

[0071] The text vectorization representation can be realized by using a pre-trained language model (such as BERT, RoBERTa, ALBERT, etc.).

[0072] The data cleaning processing comprises smoothing noise data, smoothing or deleting outliers;

[0073] The category checking processing is to delete the text in the first text information that is not in the scientific research text library from the first text information by using a preset scientific research text library to obtain pre-processed embedded vector information.

[0074] The constructing of the knowledge graph for the pre-processed embedded vector information to obtain a retrieval graph comprises:

[0075] S221, grouping all retrieval words of the pre-processed embedded vector information by using a preset retrieval word grouping rule to obtain a plurality of grouped word sets; the grouped word set comprises a plurality of retrieval words;

[0076] S222, calculating the correlation of each grouped word set with the corresponding word library to obtain a correlation calculation result;

[0077] S223, taking the grouped word set with a correlation calculation result greater than a preset correlation threshold as entity information of the retrieval graph; one correlation calculation result greater than a preset correlation threshold is taken as one entity information;

[0078] S224, calculating the relationship value between each entity information;

[0079] S225, constructing a relationship matrix using the relationship values between each entity information; an element in the ith row and jth column of the relationship matrix is a relationship value between an ith search term in a group word set corresponding to a first entity information and a jth search term in a group word set corresponding to a second entity information;

[0080] The first entity information and the second entity information are two entity information for which the relationship values are calculated.

[0081] S226, performing edge information calculation on the relationship matrix to obtain edge information between entity information of a search graph;

[0082] S227, constructing a search graph using all the entity information and the edge information.

[0083] The expression for calculating the relevance of the group word set and the corresponding word library is:

[0084]

[0085] where xg is the relevance, co1(f i ,ck j ) represents the first relevance calculation, co2(f i ,ck j ) represents the second relevance calculation, f i represents an ith search term in the group word set, ck j represents a jth word in the word library corresponding to the group word set, and N and M represent the total number of search terms included in the group word set and the total number of words included in the word library corresponding to the group word set, respectively.

[0086] The first relevance calculation can use Jaccard similarity, and the second relevance calculation can use Jaro-Winkler distance.

[0087] The relevance calculation of the group word set and the corresponding word library combines Jaccard similarity (co1, which measures the degree of set overlap) and Jaro-Winkler distance (co2, which measures the string edit distance), simultaneously capturing the semantic inclusion relationship (such as "machine learning" and "deep learning") and spelling similarity (such as "algorithm" and "algorithms") of the terms, avoiding the limitations of a single indicator. By performing a nonlinear transformation on the relevance value using a tan function and normalizing the maximum relevance within the group, the contribution of high-relevance terms is highlighted and the interference of low-relevance terms is suppressed, making the entity selection of the group word set more accurate.

[0088] The expression for calculating the relationship value between each entity information is:

[0089]

[0090] where G is the relationship value, Co3 ij represents the relevance value of the i-th search term of the first entity information and the j-th search term of the second entity information for calculating the relationship value, max{Co3 ij} represents the maximum value of the relevance values of all detected terms between the first entity information and the second entity information for calculating the relationship value, δ is a preset constant factor, and ψ is a preset decision factor. The relevance value can adopt Jaccard similarity.

[0091] When the maximum relevance max{Co3 ij} between entities exceeds the decision factor ψ, a linear decay model is adopted to weaken the influence of extremely high correlation values and avoid that a single strong correlation hides the overall semantic relationship; otherwise, a global weighted average is adopted to ensure that the relationship of weakly related entities is still effectively modeled. The constant factor δ is used to adjust the decay rate of the relationship value to adapt to the distribution characteristics of term relevance in different fields (such as more concentrated term relevance in the biomedical field and more dispersed term relevance in the information technology field), thereby improving the adaptability of the model to cross-field data.

[0092] The expression for calculating the edge information is:

[0093]

[0094] where W N1,M1 (x) is a Whitaker function, x is an input value of the Whitaker function, N1 and M1 are the row dimension and column dimension of the relationship matrix respectively, G ij is an element in the i-th row and j-th column of the relationship matrix, σ ij is the j-th element of the i-th singular vector obtained by singular value decomposition of the relationship matrix, σ i0 is the mean value of the i-th singular vector, τ i is the i-th singular value obtained by singular value decomposition of the relationship matrix, τ0 is the mean value of all singular values, and b is the edge information.

[0095] The edge information calculation combines the singular values (reflecting the overall matrix energy) and the singular vectors (reflecting the matrix feature distribution) of the relationship matrix to capture the global structure and local details of entity relationships, so that the edge information can reflect both the relationship strength and the graph topology characteristics. The Whitaker function is used to perform nonlinear mapping on the weighted distance and to convert the numerical characteristics into edge weights suitable for graph construction, thereby enhancing the expression ability of complex semantic relationships (such as indirect association of cross-field terms and transitivity of multi-hop relationships). i -G ij and σ i0 -σ ij , capture the global structure and local details of entity relationships, so that the edge information can reflect both the relationship strength and the graph topology characteristics. The Whitaker function is used to perform nonlinear mapping on the weighted distance and to convert the numerical characteristics into edge weights suitable for graph construction, thereby enhancing the expression ability of complex semantic relationships (such as indirect association of cross-field terms and transitivity of multi-hop relationships).

[0096] The preset retrieval word grouping rule includes a plurality of word libraries; each word library includes a plurality of entity words; each grouping word set has a corresponding word library;

[0097] The retrieval graph is searched in the preset scientific research knowledge graph by using a knowledge graph semantic search method to obtain first search result information;

[0098] S31, using a knowledge graph semantic search method, searching the retrieval graph in the preset scientific research knowledge graph to obtain first search result information;

[0099] S32, using an embedding vector search method, searching the retrieval graph in the preset scientific research knowledge graph to obtain second search result information;

[0100] S33, using a RAG search method, searching the retrieval graph in the preset scientific research knowledge graph to obtain third search result information;

[0101] S34, performing fusion matching calculation processing on all obtained search result information to obtain retrieval matching result information.

[0102] The embedding vector search method can use brute force search or tree structure index search.

[0103] The fusion matching calculation processing on all obtained search result information to obtain retrieval matching result information includes:

[0104] S341, performing text alignment processing on all search result information to obtain aligned search result information;

[0105] S342, calculating an important coefficient value for each aligned search result information;

[0106] S343, using the important coefficient value to perform fusion calculation on all aligned search result information to obtain retrieval matching result information.

[0107] The expression of the important coefficient calculation is:

[0108]

[0109] Wherein, zy is the important coefficient value, β i is the i-th entity information of the retrieval graph, λ i is the average value of all edge information of the i-th entity information of the retrieval graph, M2 is the total number of entity information of the retrieval graph, |α| represents the modulus value of the aligned search result information α, and cos(α,β i) represents the cosine similarity calculation result;

[0110] The expression of the fusion calculation is:

[0111]

[0112] wherein, ε j is the jth element of the retrieval matching result information, and α ij is the jth element of the ith search result information after alignment.

[0113] The RAG search method combines retrieval and generation, provides accurate answers, and enhances understanding of context. After the user presents a query, the problem and the text in the knowledge base are first converted into high-dimensional vector representations through semantic vectorization technology. The most relevant text passages are retrieved from the knowledge base based on similarity calculation. The retrieval results are input into the generation model as context, and a generative model (such as T5, BART, or GPT) is used for natural language answer generation. This architecture of RAG ensures seamless integration of retrieval results and generation process, making the answer more accurate and contextually relevant.

[0114] The knowledge graph semantic search method can use a knowledge graph semantic search method based on representation learning or a knowledge graph semantic search method based on a section;

[0115] The text vectorization representation of the to-be-retrieved scientific research text information obtains a text embedding vector. A pre-trained language model (such as BERT, RoBERTa, ALBERT, etc.) can be used to perform semantic representation on the text data and convert it into a high-dimensional vector. These vectors can capture the semantic features of the text, facilitating subsequent retrieval and calculation. In specific scenarios, the pre-trained model can be fine-tuned to further improve the semantic accuracy of the embedding, ensuring a deeper understanding of domain-specific vocabulary or context. In addition, a single-tower or double-tower structure model can be selected according to specific task requirements to optimize the distribution of semantic vectors in high-dimensional space, thereby providing a more accurate basis for retrieval. High-performance vector retrieval tools are used to index and store the generated vectors. These tools use algorithms such as HNSW and IVF to quickly find the most similar vector set to the query vector. Retrieval processes can also incorporate inverted indexes, approximate nearest neighbor search, and other techniques to improve efficiency. To improve the accuracy of retrieval results, a reordering model can be used to semantically reorder the preliminary retrieval results, ensuring that the returned information is highly relevant to the user query.

[0116] The retrieved relevant vector information is input as a condition to a generation model (such as GPT, T5, BLOOM, etc.) to inject context information into the generation process through prompt learning or explicit knowledge insertion. This not only improves the accuracy of the generated content, but also enhances the performance of the model in specific fields.

[0117] In a second aspect, the application discloses a large model-based scientific research information retrieval matching device, which comprises:

[0118] A memory storing executable program code;

[0119] A processor coupled to the memory;

[0120] The processor calls the executable program code stored in the memory to execute the large model-based scientific research information retrieval matching method.

[0121] In a third aspect, the application discloses a computer storage medium storing computer instructions, which, when called by a computer, is used to execute the large model-based scientific research information retrieval matching method.

[0122] In a fourth aspect, the application discloses an information data processing terminal for implementing the large model-based scientific research information retrieval matching method.

[0123] The above is only an embodiment of the application and is not intended to limit the application. The application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the scope of the claims of the application.

Claims

1. A large model-based scientific research information retrieval matching method, characterized in that, The application relates to a scientific research text information retrieval method and device. S1, collecting to-be-retrieved scientific research text information; The to-be-retrieved scientific research text information comprises a plurality of search terms; S2, processing the to-be-retrieved scientific research text information to obtain a search graph, comprising: S21, preprocessing the to-be-retrieved scientific research text information to obtain preprocessed embedding vector information; S22, constructing a knowledge graph based on the preprocessed embedding vector information to obtain a search graph, comprising: S221, grouping all search terms of the preprocessed embedding vector information by using a preset search term grouping rule to obtain a plurality of grouped term sets; the grouped term set comprises a plurality of search terms; S222, calculating the relevance of each grouped term set and a corresponding term library to obtain a relevance calculation result; S223, taking a grouped term set with a relevance calculation result greater than a preset relevance threshold as entity information of the search graph; S224, calculating the relationship value between each entity information; S225, constructing a relationship matrix by using the relationship value between each entity information; the element of the i-th row and the j-th column of the relationship matrix is the relationship value between the i-th search term in the grouped term set corresponding to the first entity information and the j-th search term in the grouped term set corresponding to the second entity information; S226, calculating edge information between the entity information of the search graph by using the relationship matrix; S227, constructing the search graph by using all the entity information and the edge information; The expression of the edge information calculation is: wherein W N1,M1 (x) is a Whitakker function, x is an input value of the Whitakker function, N1 and M1 are respectively a row dimension and a column dimension of a relation matrix, G ij is an element of the i-th row and the j-th column of the relation matrix, σ ij is an element of the i-th singular vector obtained by singular value decomposition of the relation matrix, σ i0 is a mean value of the i-th singular vector, τ i is an i-th singular value obtained by singular value decomposition of the relation matrix, τ0 is a mean value of all singular values, and b is side information. S3, performing retrieval processing on the search graph by using a preset scientific research knowledge graph to obtain search matching result information. 2.The large model-based scientific research information retrieval matching method according to claim 1, wherein, The preprocessing of the to-be-retrieved scientific research text information to obtain preprocessed embedding vector information comprises: S211, performing text vectorization representation on the to-be-retrieved scientific research text information to obtain text embedding vector information; S212, performing data cleaning processing on the text embedding vector to obtain first text information; S213, performing category inspection processing on the first text information to obtain preprocessed embedding vector information. 3.The large model-based scientific research information retrieval matching method of claim 1, wherein, The retrieval processing of the search graph by using a preset scientific research knowledge graph to obtain search matching result information comprises: S31, searching the search graph in the preset scientific research knowledge graph by using a knowledge graph semantic search method to obtain first search result information; S32, searching the search graph in the preset scientific research knowledge graph by using an embedding vector search method to obtain second search result information; S33, searching the search graph in the preset scientific research knowledge graph by using an RAG search method to obtain third search result information; S34, performing fusion matching calculation processing on all the obtained search result information to obtain search matching result information. 4.The large model-based scientific research information retrieval matching method of claim 3, wherein, The fusion matching calculation processing on all the obtained search result information to obtain search matching result information comprises: S341, performing text alignment processing on all the search result information to obtain aligned search result information; S342, important coefficient calculation is performed on each aligned search result information to obtain an important coefficient value; S343, fusion calculation is performed on all aligned search result information by using the important coefficient value to obtain search matching result information. 5.The large model-based scientific research information retrieval matching method according to claim 4, wherein, An expression of the important coefficient calculation is: wherein zy is an important coefficient value, β i is the i-th entity information of the search graph, λ i is the average value of all edge information of the i-th entity information of the search graph, M2 is the total number of entity information of the search graph, |α| represents the modulus value of the aligned search result information α, and cos(α, β i ) represents the cosine similarity calculation result. An expression of the fusion calculation is: Wherein, ε j is the jth element of the search result information retrieved ij is the jth element of the ith search result information after alignment. 6.A large model-based scientific research information retrieval matching device, characterized in that, The device comprises: a memory storing executable program codes; a processor coupled with the memory; The processor invokes the executable program codes stored in the memory to execute the scientific research information retrieval matching method based on a large model according to any one of claims 1 to 5.

7. A computer storable medium, characterized by The computer storage medium stores computer instructions, which are invoked by a computer to execute the scientific research information retrieval matching method based on a large model according to any one of claims 1 to 5.

8. An information data processing terminal, characterized by The information data processing terminal is used to implement the scientific research information retrieval matching method based on a large model according to any one of claims 1 to 5.

Images