An AI intelligence-based literature information retrieval analysis system and method

Abstract

The application relates to the technical field of document information retrieval, and specifically discloses a document information retrieval analysis system and method based on AI intelligence, wherein the system comprises a representation vector output module, a preliminary retrieval result output module, an updated retrieval result module, a dynamic knowledge graph construction module, an enhanced document representation output module, a recommendation list output module and a knowledge graph feedback updating module.The representation vector output module extracts document text features, the preliminary retrieval result output module performs semantic enhancement retrieval and query expansion, the updated retrieval result module optimizes the preliminary result, the dynamic knowledge graph construction module constructs a graph based on the retrieval result, the enhanced document representation output module outputs enhanced document representation, the recommendation list output module provides personalized recommendation, and the knowledge graph feedback updating module updates the knowledge graph according to user interaction flow data.The application comprehensively improves the accuracy, intelligence and dynamic adaptability of document retrieval and can deeply meet user needs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document information retrieval, in particular to a document information retrieval analysis system and method based on AI intelligence. BACKGROUND

[0002] With the progress of science and technology and the popularization of the Internet, the amount of global scientific research data is growing exponentially, and the number of documents in various databases is increasing dramatically. Traditional keyword matching-based retrieval methods are difficult to cope with massive data, and cannot fully and accurately reflect user needs, resulting in low information retrieval efficiency and difficulty in filtering valuable information, and users are easily overwhelmed by irrelevant or low-quality content.

[0003] Current mainstream technologies include:

[0004] Keyword matching retrieval: based on Boolean logic and inverted index (such as Elasticsearch), but cannot capture semantic associations;

[0005] Citation network analysis: mining document influence through PageRank and other algorithms, but ignoring content semantics and cross-disciplinary associations;

[0006] Machine learning models: using LDA topic models or Word2Vec word vectors, but limited by shallow feature representation capabilities.

[0007] Therefore, the existing document retrieval method not only has low efficiency, but also easily misses important information, requires manual intervention in determining retrieval strategies and adjusting results, which may lead to human errors, and lacks intelligence, and cannot make personalized recommendations based on user search history and behavior patterns. SUMMARY

[0008] To overcome the shortcomings of the prior art, the present application provides a document information retrieval analysis system and method based on AI intelligence, which improves the accuracy, intelligence and dynamic adaptability of document retrieval, and deeply meets user needs.

[0009] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0010] A document information retrieval analysis system based on AI intelligence, comprising:

[0011] A representation vector output module for extracting document text features, building a multi-modal document representation model, and outputting document representation vectors;

[0012] A preliminary retrieval result output module for performing semantic enhanced retrieval, expanding user queries, and outputting a preliminary retrieval result set and a corresponding similarity score set;

[0013] The updating retrieval result module is configured to optimize the preliminary retrieval result set based on the similarity scores to obtain an updated retrieval result set.

[0014] The dynamic knowledge graph construction module is configured to construct a dynamic knowledge graph based on the updated retrieval result set.

[0015] The enhanced literature representation output module is configured to perform cross-disciplinary correlation analysis based on the dynamic knowledge graph and output enhanced literature representations in combination with literature representation vectors.

[0016] The recommendation list output module is configured to obtain user portrait data and output personalized recommendation lists in combination with the enhanced literature representations and the dynamic knowledge graph.

[0017] The knowledge graph feedback updating module is configured to collect user interaction flow data and output updated knowledge graphs in combination with the personalized recommendation lists.

[0018] Preferably, in the representation vector output module, the process of outputting the literature representation vector is as follows:

[0019] The BERT model is used to extract the [CLS] mark vector of the literature title and abstract , which is represented as:

[0020] ;

[0021] In the formula, is the literature title and abstract text content input into the BERT model, is the original output vector after the BERT model extracts the [CLS] mark vector of the literature title and abstract, which is denoted as the BERT output vector;

[0022] The BERT output vector is whitened, and the steps include:

[0023] Step 1: Calculate the mean value

[0024] ;

[0025] In the formula, i is the sample number, i = 1, 2, 3,..., N, N is the sample quantity, is the i-th element in , and is the mean value of the calculated vector;

[0026] Step 2: Centering processing

[0027] ;

[0028] In the formula, is the whitened the vector after the centralization processing;

[0029] Step three, eigenvalue decomposition of the covariance matrix:

[0030] ;

[0031] wherein, is the covariance matrix, is the eigenvector matrix, is the diagonal eigenvalue matrix, denotes the transpose of the vector;

[0032] Step four, dimensionality reduction processing:

[0033] ;

[0034] wherein, is the text feature vector after dimensionality reduction;

[0035] extract visual features, encode the figures and formulas in the literature using the CLIP model to obtain the visual feature vector : ;

[0036] ;

[0037] Integrate metadata features, extract metadata features M, including author index H, journal impact factor I, and citation network weight C;

[0038] ;

[0039] By weighted fusion of text features, visual features and metadata features, output the literature representation vector :

[0040] ;

[0041] wherein, is the ELU activation function, , , is a trainable weight matrix, which is learned from the training dataset to minimize the loss function of the retrieval task.

[0042] Preferably, in the preliminary retrieval result output module, the process of outputting the preliminary retrieval result set and the corresponding similarity score set is:

[0043] Based on the term correlation matrix S, combined with the TF-IDF value, the cosine similarity of BERT embedding is calculated, represented as:

[0044] ​ ;

[0045] wherein, is an element in the term correlation matrix S, representing the correlation between the mth query term and the nth query term , is the cosine similarity between the embedding vectors obtained after processing the query term and by the BERT model, is the term frequency-inverse document frequency value of the query term in the document set D, is the term frequency-inverse document frequency value of the query term in the document set D;

[0046] Based on the term correlation matrix S, the top 5 terms with the highest correlation are selected to expand the original query;

[0047] An abnormal term filtering mechanism is introduced to calculate the frequency deviation of the query term q :

[0048] ;

[0049] When > , the query term q is removed from the expanded query, is the standard deviation;

[0050] For each query term q, find the document d with the maximum dot product, and use the ColBERT architecture to calculate the similarity score of the query statement Q and the document set D :

[0051] ;

[0052] In the formula, Q is the query statement, which is the retrieval content input by the user, D is the document set, i.e. the document library to be retrieved, and are embedding functions for converting query terms q and documents d into vector representations, represents the transpose of the vector;

[0053] According to the output similarity score, the corresponding document is taken as the preliminary retrieval result;

[0054] The preliminary retrieval result set and the corresponding similarity score set are output.

[0055] Preferably, in the update retrieval result module, the process of obtaining the updated retrieval result set is:

[0056] A credibility evaluation matrix is constructed:

[0057] ;

[0058] wherein, is a document credibility index, is the number of citations of the document, is an author ranking based on author H-index level, is the impact factor of the journal, is a set weight factor of the document credibility index, is a set weight factor of the author ranking, is a set weight factor of the impact factor of the journal, and e is a natural constant;

[0059] Based on the document credibility index, the similarity score is dynamically corrected:

[0060] ;

[0061] wherein, is the corrected similarity score of the sentence Q and the document set D as a search score, is an activation function;

[0062] When <0.6 and from a non-core journal, remove from the results;

[0063] The updated search result set and the corresponding search score are obtained.

[0064] Preferably, in the dynamic knowledge graph construction module, the process of constructing the dynamic knowledge graph is:

[0065] Establishing entity nodes, including subject entity nodes and object entity nodes;

[0066] Based on the updated search result set, using an improved OpenNRE model, combining BiLSTM and context attention features, calculating entity relationship probability :

[0067] ;

[0068] wherein, s is the subject entity corresponding to the subject entity node, o is the object entity corresponding to the object entity node, r is the relationship type between s and o, and c is the context attention feature, is a weight matrix related to the relationship r;

[0069] If the entity relationship probability is greater than a set probability threshold, a graph edge is established for s and o, and an initial weight is assigned to the graph edge;

[0070] If the probability of entity relationship If the probability is not greater than the set probability threshold, then no graph edges are created for s and o;

[0071] Apply time-series decay to the initial weights of the graph edges:

[0072] ;

[0073] In the formula, Let be the weight of the graph edge at time t. The initial weights of the graph edges are given by time. Weight of time, is the attenuation coefficient with a value of 0.05, and e is the natural constant; output a dynamic knowledge graph.

[0074] Preferably, the process of obtaining the context attention feature c is as follows:

[0075] Obtaining the BiLSTM hidden state sequence a = 1, 2, 3, ..., z, where z is the number of elements in the hidden state sequence. Let be the hidden state vector of the a-th element in the hidden state sequence;

[0076] Calculate the attention weight of the a-th element in the hidden state sequence. :

[0077] ;

[0078] In the formula, For query vector, Let be the dimension of the vector;

[0079] Generate context attention features:

[0080] .

[0081] Preferably, in the enhanced document representation output module, the process of outputting the enhanced document representation is as follows:

[0082] Input a dynamic knowledge graph and extract the subject tag distribution matrix:

[0083] ;

[0084] In the formula, These are elements in the subject label distribution matrix, used to measure the subject. and disciplines The relationship between them For the discipline and disciplines Co-occurrence frequency, For the discipline information entropy of the discipline g and h are both discipline numbers;

[0085] calculating the discipline intersection degree :

[0086] ;

[0087] wherein, is the PageRank value of the entity in the literature d, which is used as the weight in the calculation of the discipline intersection degree;

[0088] generating the enhanced literature representation:

[0089] ;

[0090] wherein, is the enhanced literature representation, is the original literature representation, is the cross-discipline projection matrix, is an adjustable coefficient.

[0091] Preferably, in the recommendation list output module, the process of outputting the personalized recommendation list is:

[0092] constructing a user feature vector U, including search behavior, reading record, citation pattern, cooperation network, field preference and freshness demand:

[0093] ;

[0094] wherein, is the weight factor of the set search behavior, is the weight factor of the set reading record, is the weight factor of the set citation pattern, is the weight factor of the set cooperation network, is the weight factor of the set field preference, is the weight factor of the set freshness demand;

[0095] combining the similarity of the user feature vector and the literature representation vector, and the similarity of the similar user set, to obtain a recommendation score:

[0096] ;

[0097] wherein, is the recommendation score of the literature d for the user u, is the feature vector of the user u, is the enhanced literature representation of the literature d, is the similarity between the user feature vector and the enhanced literature representation of the literature d , is a set of similar users, v is one of the similar users, is a feature vector of similar user v is an enhanced document representation of document d is the similarity between is a set of similar users is the cardinality of the set, i.e., the number of users in the set;

[0098] According to the calculated Sort all documents;

[0099] Select the top x documents with the highest scores as the personalized recommendation list;

[0100] Output the personalized recommendation list.

[0101] Preferably, in the knowledge graph feedback updating module, the process of outputting the updated knowledge graph is:

[0102] Real-time capture user interaction behavior, build user interaction behavior set F:

[0103] User interaction behavior set F contains click, dwell time, download, citation, represented as:

[0104] ;

[0105] Calculate behavior weight :

[0106] ;

[0107] In the formula, is an indicator function;

[0108] Analyze the update entity influence weight :

[0109] ;

[0110] In the formula, is the old value of the entity influence weight;

[0111] When the node weight change rate is higher than 15%, reconstruct the local graph topology, after the local graph topology reconstruction is completed, integrate into the global knowledge graph, output the updated knowledge graph.

[0112] An AI intelligent-based literature information retrieval analysis method based on the above system, comprising the following steps:

[0113] S1, extract the literature text features to be searched, construct a multi-modal literature representation model, and output a literature representation vector;

[0114] S2, perform semantic enhanced retrieval, extend the user query, output a preliminary retrieval result set and a corresponding similarity score set;

[0115] S3, based on the similarity score, optimize the preliminary retrieval result set to obtain an updated retrieval result set;

[0116] S4, based on the updated retrieval result set, construct a dynamic knowledge graph;

[0117] S5, based on the dynamic knowledge graph, perform cross-disciplinary correlation analysis, and combine the literature representation vector to output an enhanced literature representation;

[0118] S6, obtain user portrait data, and combine the enhanced literature representation and the dynamic knowledge graph to output a personalized recommendation list;

[0119] S7, collect user interaction flow data, and combine the personalized recommendation list to output an updated knowledge graph.

[0120] The present application has the following beneficial effects:

[0121] The present application provides a literature information retrieval analysis system and method based on AI intelligence, the representation vector output module extracts the literature text features, and the multi-modal literature representation model can comprehensively depict the literature; the preliminary retrieval result output module performs semantic enhanced retrieval and query expansion, improves the retrieval accuracy, and enables the user to obtain more relevant literature. The knowledge graph feedback update module updates the knowledge graph according to the user interaction flow data, so that the system continuously learns and optimizes, adapts to new literature and user demand changes, and maintains the accuracy and effectiveness of the retrieval analysis.

[0122] The present application optimizes the preliminary result through the updated retrieval result module, eliminates low-quality or irrelevant literature, reduces the user's screening time and effort, and quickly focuses on valuable content. The dynamic knowledge graph construction module constructs the graph based on the retrieval result, displays the entities and relationships in the literature, performs cross-disciplinary correlation analysis to mine cross-disciplinary knowledge connections, and provides new ideas. The recommendation list output module combines the enhanced literature representation, the knowledge graph and the user portrait to provide personalized recommendations, meets different user demand preferences, and discovers potential useful literature. BRIEF DESCRIPTION OF DRAWINGS

[0123] Figure 1 It is a schematic diagram of the system module connection of the present application;

[0124] Figure 2 It is a schematic diagram of the method step flow of the present application;

[0125] Figure 3A logic block diagram for the system of the present application. DETAILED DESCRIPTION

[0126] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

[0127] Embodiment 1: As shown in the figure, an AI intelligent-based literature information retrieval analysis system includes: Figure 1

[0128] A representation vector output module is configured to extract literature text features to be retrieved, build a multi-modal literature representation model, and output literature representation vectors.

[0129] A preliminary retrieval result output module is configured to perform semantic enhancement retrieval, expand user queries, and output a preliminary retrieval result set and a corresponding similarity score set.

[0130] An updated retrieval result module is configured to optimize the preliminary retrieval result set based on the similarity score to obtain an updated retrieval result set.

[0131] A dynamic knowledge graph construction module is configured to construct a dynamic knowledge graph based on the updated retrieval result set.

[0132] An enhanced literature representation output module is configured to perform cross-disciplinary correlation analysis based on the dynamic knowledge graph, combine the literature representation vectors, and output enhanced literature representations.

[0133] A recommendation list output module is configured to obtain user portrait data, combine the enhanced literature representations and the dynamic knowledge graph, and output a personalized recommendation list.

[0134] A knowledge graph feedback update module is configured to collect user interaction flow data, combine the personalized recommendation list, and output an updated knowledge graph.

[0135] In the representation vector output module, the process of outputting the literature representation vectors is as follows:

[0136] The [CLS] token vector of the literature title and abstract is extracted using the BERT model, denoted as:

[0137]

[0138] In the formula, the input BERT model literature title and abstract text content is denoted as, the [CLS] token vector of the literature title and abstract extracted using the BERT model is denoted as the original output vector after the extraction, and is denoted as the BERT output vector; The BERT output vector is denoted as

[0139] ​​​​The whitening process is performed, and the steps include:

[0140] Step one, calculate the mean value:

[0141] ;

[0142] In the formula, i is the sample number, i = 1, 2, 3, …, N, N is the sample quantity, is the i-th element in , is the mean value of the calculated vector;

[0143] Step two, centering processing:

[0144] ;

[0145] In the formula, is the vector after centering processing ;

[0146] Step three, eigenvalue decomposition of the covariance matrix:

[0147] ;

[0148] In the formula, is the covariance matrix, is the eigenvector matrix, is the diagonal eigenvalue matrix, represents the transpose of the vector;

[0149] Step four, dimensionality reduction processing:

[0150] ;

[0151] In the formula, is the text feature vector after dimensionality reduction;

[0152] Extract visual features, use the CLIP model to encode the charts and formulas in the literature to obtain the visual feature vector :

[0153] ;

[0154] Integrate metadata features, extract metadata features M, including author index H, journal impact factor I, and citation network weight C;

[0155] ;

[0156] By weighted fusion of text features, visual features and metadata features, the literature representation vector is output :

[0157] ;

[0158] where, is an ELU activation function, , , is a trainable weight matrix learned from the training dataset to minimize the loss function of the retrieval task.

[0159] Combining text, visual and metadata features, the multi-modal way comprehensively depicts the literature. The text feature extraction uses the BERT model to capture semantic information, the visual feature encodes the chart formula through the CLIP model, and the metadata feature covers key information such as author and journal. The output literature representation vector is more complete and accurate.

[0160] The BERT output vector is whitened (mean, centering, eigenvalue decomposition and dimension reduction), which reduces the data dimension, removes redundancy and correlation, improves the calculation efficiency, highlights the key text features, and enhances the accuracy of subsequent retrieval analysis. It should be noted that a cross-modal alignment layer (such as CCA or cross-modal attention) needs to be introduced to normalize and project the features to a unified semantic space.

[0161] When different features are weighted and fused, the trainable weight matrix is learned from the training dataset to minimize the loss function according to the retrieval task, which flexibly adjusts the weight of each feature, adapts to different retrieval analysis requirements, and improves the system performance.

[0162] All parameters in the embodiment are de-dimensioned, and the weight factors in the embodiment are obtained based on the data stored in the database.

[0163] In the preliminary retrieval result output module, the process of outputting the preliminary retrieval result set and the corresponding similarity score set is:

[0164] Based on the term correlation matrix S, combined with the TF-IDF value, the cosine similarity of the BERT embedding is calculated, which is represented as:

[0165] ;

[0166] where, is an element in the term correlation matrix S, which represents the correlation between the mth query word and the nth query word , is the cosine similarity between the embedding vectors obtained after processing the query words and by the BERT model, is the cosine similarity between the embedding vectors obtained after processing the query words The term frequency-inverse document frequency value in the document set D, is the query word The term frequency-inverse document frequency value in the document set D;

[0167] Based on the term correlation matrix S, the top 5 terms with the highest correlation are selected to expand the original query;

[0168] An abnormal term filtering mechanism is introduced to calculate the frequency deviation of the query word q :

[0169] ;

[0170] When > The query word q is removed from the expanded query, is the standard deviation;

[0171] For each query word q, find the document d with the maximum dot product, and use the ColBERT architecture to calculate the similarity score of the query sentence Q and the document set D :

[0172] ;

[0173] In the formula, Q is the query sentence, which is the retrieval content input by the user, D is the document set, i.e. the document library to be searched, and are embedding functions used to convert query words q and documents d into vector representations, denotes the transpose of the vector;

[0174] According to the output similarity score, the corresponding document is taken as the preliminary retrieval result;

[0175] Output the preliminary retrieval result set and the corresponding similarity score set.

[0176] Based on the term correlation matrix, combined with the cosine similarity of BERT embedding and TF-IDF value, the correlation between query words can be more accurately measured, the top 5 related terms are selected to expand the original query, the user query content is supplemented and refined, the comprehensiveness and accuracy of the query are improved, and the user can find more suitable documents.

[0177] An abnormal term filtering mechanism is introduced to calculate the frequency deviation of the query word q, remove abnormal terms, avoid irrelevant or low-frequency abnormal terms from interfering with the retrieval results, purify the query content, and make the retrieval results more relevant and reliable.

[0178] The similarity score of the query statement and the document set is calculated using the ColBERT architecture, which considers the vector relationship of the query words and the words in the document, and can measure the matching degree of the two from a fine-grained level. Compared with the simple matching method, it can more accurately find the documents highly related to the query, and improve the precision and quality of the retrieval.

[0179] In the updating retrieval result module, the process of obtaining the updated retrieval result set is:

[0180] A credibility evaluation matrix is constructed:

[0181] ;

[0182] In the formula, is a document credibility index, is the number of citations of the document, is the author ranking based on the H-index level of the author, is the impact factor of the journal, is the set weight factor, is the set weight factor, is the set weight factor, and e is a natural constant;

[0183] Based on the document credibility index, the similarity score is dynamically corrected:

[0184] ;

[0185] In the formula, is the modified similarity score of the statement Q and the document set D, which is the retrieval score, is an activation function;

[0186] When <0.6 and from non-core journals, remove from the results;

[0187] The updated retrieval result set and the corresponding retrieval score are obtained.

[0188] By constructing the credibility evaluation matrix, the number of citations of the document, the author ranking and the impact factor of the journal and other multi-dimensional factors are considered, which can more comprehensively and scientifically evaluate the credibility of each document and provide reliable basis for subsequent processing.

[0189] Based on the document credibility index, the retrieval score is dynamically corrected, and the quality factors of the document itself are included in the retrieval scoring system. Compared with the score based on similarity alone, it can more reasonably reflect the matching degree of the document and the user's demand, and improve the quality of the retrieval result.

[0190] By setting clear screening conditions, i.e. removing the literature from the results when the literature credibility index is low and from non-core journals, possible low-quality or unreliable literature is effectively filtered out, making the final update search result set more accurate and valuable, saving users time and effort in screening literature.

[0191] In the dynamic knowledge graph construction module, the process of constructing a dynamic knowledge graph is:

[0192] Establishing entity nodes, including subject entity nodes and object entity nodes;

[0193] Based on the update search result set, using the improved OpenNRE model, combining BiLSTM and context attention features, calculating entity relationship probability :

[0194] ;

[0195] In the formula, s is the subject entity corresponding to the subject entity node, o is the object entity corresponding to the object entity node, r is the relationship type between s and o, and c is the context attention feature, is the weight matrix related to the relationship r;

[0196] If the entity relationship probability is greater than the set probability threshold, a graph edge is established for s and o, and an initial weight is assigned to the graph edge;

[0197] If the entity relationship probability is not greater than the set probability threshold, no graph edge is established for s and o;

[0198] The initial weight of the graph edge is subjected to time decay:

[0199] ;

[0200] In the formula, is the weight of the graph edge at time t, is the initial weight of the graph edge, i.e. the weight at time , and is the decay coefficient, with a value of 0.05, and e is the natural constant; output the dynamic knowledge graph.

[0201] The process of obtaining the context attention feature c is:

[0202] Obtain the BiLSTM hidden state sequence , a=1,2,3,...,z, z is the number of elements in the hidden state sequence, is the hidden state vector of the a-th element in the hidden state sequence;

[0203] The attention weight of the a-th element in the hidden state sequence is calculated :

[0204] ;

[0205] wherein, is a query vector, is the dimension of the vector;

[0206] Generate context attention features:

[0207] .

[0208] With the improved OpenNRE model, the BiLSTM and context attention features are fused to calculate the entity relationship probability, which can deeply mine the semantic information in the text, more accurately judge the relationship type between the subject entity and the object entity in the knowledge graph, and improve the construction quality of the knowledge graph.

[0209] The time decay is applied to the graph edge weight, and the edge weight is automatically adjusted over time, reflecting the timeliness of knowledge. It can effectively reduce the influence of outdated information, make the knowledge graph better reflect the dynamic changes of knowledge, and ensure that the content of the knowledge graph is consistent with the current actual situation.

[0210] By obtaining the BiLSTM hidden state sequence and calculating the attention weight to generate the context attention feature, the context information of the text is fully considered, and the understanding of the text semantics is enhanced. It helps to more accurately capture the semantic association between entities and construct a knowledge graph with more reasonable structure and richer semantics.

[0211] In the enhanced literature representation output module, the process of outputting the enhanced literature representation is:

[0212] Input the dynamic knowledge graph, extract the discipline label distribution matrix:

[0213] ;

[0214] wherein, is an element in the discipline label distribution matrix, used to measure the relationship between the discipline and the discipline , is the co-occurrence frequency of the discipline and the discipline , is the information entropy of the discipline , and g and h are discipline numbers;

[0215] Calculate the discipline intersection degree :

[0216] ;

[0217] wherein, is the PageRank value of the entity in document d, which is used as the weight when calculating the discipline intersection degree;

[0218] Generating enhanced document representation:

[0219] ;

[0220] wherein, is the enhanced document representation, is the original document representation, is the cross-discipline projection matrix, is the adjustable coefficient.

[0221] Based on the dynamic knowledge graph, the cross-discipline association analysis is performed by extracting the discipline label distribution matrix and calculating the discipline intersection degree, which can mine the potential relationship between different disciplines, find the fusion point of cross-discipline knowledge, and help researchers to carry out cross-discipline research and expand the research vision.

[0222] By integrating the discipline intersection degree information into the document representation and combining the original document representation vector to generate the enhanced document representation, the representation of the document is more comprehensive and rich, not only containing the content features of the document itself, but also reflecting its association information with other disciplines, which provides more valuable information for subsequent document retrieval, recommendation and other tasks.

[0223] The adjustable coefficient makes the module flexible to adjust the influence of the discipline intersection degree information on the enhanced document representation according to different application scenarios and requirements, enhancing the adaptability and versatility of the module.

[0224] In the recommendation list output module, the process of outputting the personalized recommendation list is as follows:

[0225] Constructing the user feature vector U, including search behavior, reading record, citation pattern, cooperation network, field preference and freshness demand:

[0226] ;

[0227] wherein, is the set weight factor of search behavior, is the set weight factor of reading record, is the set weight factor of citation pattern, is the set weight factor of cooperation network, is the set weight factor of field preference, is the set weight factor of freshness demand;

[0228] Among them, the search behavior is the number of search keywords used by the user in a period of time, the reading record is the total number of documents read by the user, the citation mode is the number of documents cited by the user in the creation of content, the cooperation network is the number of different fields involved in cooperation, the field preference is the maximum number of search times of the user in each field, and the freshness demand is the proportion of new documents read by the user;

[0229] The recommendation score is obtained by combining the similarity of the user feature vector and the document representation vector, and the similarity of the similar user set:

[0230] ;

[0231] In the formula, is the recommendation score of the user u to the document d, is the feature vector of the user u, is the enhanced document representation of the document d, is the similarity between the user feature vector and the enhanced document representation of the document d, is the similar user set, and v is one of the similar users, is the feature vector of the similar user v and the enhanced document representation of the document d, is the similarity between the feature vector of the similar user v and the enhanced document representation of the document d, is the cardinality of the similar user set , that is, the number of users in the set;

[0232] According to the calculated , all documents are sorted;

[0233] The top x documents with the highest scores are selected as the personalized recommendation list;

[0234] The personalized recommendation list is output.

[0235] The user feature vector is constructed to include search behavior, reading record and other multi-dimensional features, and each dimension is provided with a weight factor, which can comprehensively and accurately depict the behavior mode, interest preference and demand characteristics of the user, and lay a solid foundation for personalized recommendation.

[0236] The recommendation score is calculated by combining the similarity of the user feature vector and the document enhanced representation vector, and the similarity of the similar user set and the document, which not only considers the matching degree between the user and the document, but also learns from the preferences of similar users, so that the recommendation result is more in line with the actual needs of the user. It should be noted that when U is less than the threshold, the document feature and the query similarity are directly matched, that is, the recommendation is based on content.

[0237] In the knowledge graph feedback updating module, the process of outputting the updated knowledge graph is:

[0238] Real-time capture of user interaction behavior, build user interaction behavior set F:

[0239] User interaction behavior set F contains click, dwell time, download, citation, represented as:

[0240] ;

[0241] Calculate behavior weight :

[0242] ;

[0243] In the formula, is an indicator function;

[0244] Analysis of updating entity influence weight :

[0245] ;

[0246] In the formula, is the old value of entity influence weight;

[0247] When the node weight change rate is higher than 15%, the local graph topology is reconstructed, and after the local graph topology reconstruction is completed, it is integrated into the global knowledge graph, and the updated knowledge graph is output.

[0248] Real-time capture of user interaction behavior in personalized recommendation list, according to click, dwell time and other behaviors to calculate weight, and update entity influence weight. It can closely combine user actual use feedback, make the knowledge graph more close to user demand and interest, and improve the practicability of the knowledge graph to the user.

[0249] When the node weight change rate is higher, the local graph topology is reconstructed and integrated into the global knowledge graph, which can reflect the dynamic changes of knowledge and user interest in time. Continuously optimize the structure and content of the knowledge graph, so that it maintains timeliness and accuracy, and adapts to the developing information environment.

[0250] Based on user interaction behavior, adjust entity influence weight, which helps to more accurately reflect the association and importance between knowledge. Make the knowledge association in the knowledge graph more consistent with the actual situation, and provide more reliable knowledge basis for subsequent literature retrieval, recommendation and other tasks.

[0251] In combination with the above description, the system operation logic block diagram of the present application is shown as Figure 3 .

[0252] Compared with the traditional retrieval analysis system, the advantages of the present application are summarized as shown in Table 1.

[0253] Table 1 Summary of system advantages

[0254]

[0255] Example 2: Based on example 1, when the personalized recommendation list is based on and the user interaction flow data is collected, the updated knowledge graph is output, the updated knowledge graph is fed back to the enhanced literature representation output module, the cross-disciplinary correlation analysis is performed again based on the updated knowledge graph, the enhanced literature representation is output combined with the output literature representation vector, the recommendation list output module and the knowledge graph feedback update module are continued to be executed, until the knowledge graph feedback update module reaches the maximum number of iterations, the last iteration updated knowledge graph is output, and the last iteration updated knowledge graph is stored.

[0256] Example 3: As shown in the following, a literature information retrieval analysis method based on AI intelligence, based on the system of example 1, comprising the following steps: Figure 2

[0257] S1, extract the text features of the literature to be retrieved, construct a multi-modal literature representation model, and output a literature representation vector;

[0258] S2, perform semantic enhancement retrieval, expand user query, and output a preliminary retrieval result set and a corresponding similarity score set;

[0259] S3, based on the similarity score, optimize the preliminary retrieval result set to obtain an updated retrieval result set;

[0260] S4, based on the updated retrieval result set, construct a dynamic knowledge graph;

[0261] S5, perform cross-disciplinary correlation analysis based on the dynamic knowledge graph, and output enhanced literature representation combined with the literature representation vector;

[0262] S6, obtain user portrait data, and output a personalized recommendation list combined with the enhanced literature representation and the dynamic knowledge graph;

[0263] S7, collect user interaction flow data, and output an updated knowledge graph combined with the personalized recommendation list.​

Claims

1. An AI intelligence-based literature information retrieval analysis system, characterized by, Comprise: A representation vector output module for extracting text features of documents to be retrieved, constructing a multi-modal document representation model, and outputting document representation vectors; A preliminary retrieval result output module for performing semantic enhanced retrieval, expanding user queries, and outputting a preliminary retrieval result set and a corresponding similarity score set, the process being: Based on the term correlation matrix S, the cosine similarity of BERT embedding is calculated combined with the TF-IDF value, denoted as: ； wherein, is an element in the term correlation matrix S, representing the correlation between the mth query term and the nth query term , is the cosine similarity between the query term and the embedding vector obtained after processing by the BERT model, is the term frequency-inverse document frequency value of the query term in the document set D, is the term frequency-inverse document frequency value of the query term in the document set D; Based on the term correlation matrix S, the top 5 terms with the highest correlation are selected to expand the original query; An exception term filtering mechanism is introduced to calculate the frequency deviation degree of the query word q : ； When > the query word q is removed from the extended query, is the standard deviation; For each query term q, find the document d with the largest dot product, compute the similarity scores of the query sentence Q to the set of documents D using the ColBERT architecture : ； In the formula, Q is a query statement, which is the retrieval content input by a user, D is a document set, i.e., a document library to be retrieved, and are embedding functions for converting the query word q and the document d into vector representations, denotes the transpose of the vector. According to the output similarity score, the corresponding document is taken as the preliminary retrieval result; Output the preliminary retrieval result set and the corresponding similarity score set; An updated retrieval result module for optimizing the preliminary retrieval result set based on the similarity score to obtain an updated retrieval result set; A dynamic knowledge graph construction module for constructing a dynamic knowledge graph based on the updated retrieval result set; An enhanced document representation output module for cross-disciplinary correlation analysis based on the dynamic knowledge graph, combined with the document representation vector, and outputting enhanced document representation; A recommendation list output module for obtaining user portrait data, and outputting a personalized recommendation list combined with the enhanced document representation and the dynamic knowledge graph; A knowledge graph feedback update module for collecting user interaction flow data and outputting an updated knowledge graph combined with the personalized recommendation list.

2. The AI intelligence-based document information retrieval and analysis system according to claim 1, wherein, In the representation vector output module, the process of outputting the document representation vector is: Using a BERT model to extract [CLS] token vectors for document titles and abstracts is represented as: ； In the formula, is the title and abstract text content of the input BERT model, is the original output vector after the [CLS] label vector of the title and abstract extracted using the BERT model, which is denoted as the BERT output vector. bert output vectors whitening, the steps comprising: Step one, calculate the mean: ； where i is the sample number, i = 1, 2, 3,..., N, N is the number of samples, is the i-th element in the vector, is the mean of the calculated vector; Step two, centering processing: ； In the formula, is a vector after centralization processing on ; Step three, eigenvalue decomposition of the covariance matrix: ； wherein is a covariance matrix, is a matrix of eigenvectors, is a diagonal matrix of eigenvalues, denotes the transpose of a vector; Step four, dimensionality reduction processing: ； In the formula, is the reduced dimension text feature vector; Extract visual features, use CLIP model on figures in literature and equations Encode, get visual feature vector : ； Integrate metadata features and extract metadata features M, including author index H, journal impact factor I, and citation network weight C; ； The text features, visual features and metadata features are fused by weighting to output a document representation vector : ； wherein, is an ELU activation function, , , is a trainable weight matrix learned from a training dataset to minimize a loss function of the retrieval task. 3.The AI intelligent-based literature information retrieval and analysis system according to claim 1, wherein, In the updated retrieval result module, the process of obtaining the updated retrieval result set is: Construct a credibility evaluation matrix: ； wherein, is the document credibility index, is the number of citations of the document, is the author ranking based on the author H-index level, is the impact factor of the journal, is the set weight factor for the, is the set weight factor for the, is the set weight factor for the, e is the natural constant; Based on the document credibility index, dynamically correct the similarity score: ； wherein is the modified similarity score of the statement Q to the document collection D as a retrieval score, is the activation function; When <0.6 and from non-core journals, removed from results; Obtain the updated retrieval result set and the corresponding retrieval score.

4. The AI intelligence-based document information retrieval and analysis system according to claim 1, wherein, In the dynamic knowledge graph construction module, the process of constructing the dynamic knowledge graph is: Establish entity nodes, including subject entity nodes and object entity nodes; Based on the updated search result set, an improved OpenNRE model is used to calculate entity relationship probabilities, using BiLSTM and context attention features : ； In the formula, s is a subject entity corresponding to a subject entity node, o is an object entity corresponding to an object entity node, r is a relationship type between s and o, c is a context attention feature, is a weight matrix related to the relationship r; If the probability of entity relationship If the probability exceeds the set probability threshold, then a graph edge is created between s and o, and an initial weight is assigned to the graph edge. if the entity relationship probability is not greater than a set probability threshold, then no graph edge is established for s and o; Apply time decay to the initial weight of the graph edge: ； In the formula, is the weight of the edge of the graph at time t, is the initial weight of the edge of the graph, that is, the weight at time is the initial weight of the edge of the graph, that is, the weight at time is the decay coefficient, the value is 0.05, e is a natural constant; and the output dynamic knowledge graph.

5. The AI intelligence-based document information retrieval and analysis system according to claim 4, wherein, The process of obtaining the context attention feature c is: obtaining a BiLSTM hidden state sequence , a = 1, 2, 3, …, z, z is the number of elements in the hidden state sequence, is the hidden state vector of the a-th element in the hidden state sequence; computing an attention weight for the a-th element in the sequence of hidden states : ； wherein is a query vector, is the dimension of the vector; Generate context attention features: 。 6.The AI intelligent-based literature information retrieval and analysis system according to claim 1, wherein, In the enhanced document representation output module, the process of outputting the enhanced document representation is: Input the dynamic knowledge graph and extract the discipline label distribution matrix: ； wherein, is an element in the subject tag distribution matrix, which measures the relationship between subjects and subjects , is the co-occurrence frequency of subjects and subjects , is the information entropy of subjects , and g and h are subject numbers; Computational science interdisciplinarity : ； wherein, is the PageRank value of the entity in document d, which is used as a weight when calculating the degree of interdisciplinary. Generate enhanced document representation: ； wherein to enhance the literature characterization, to the original literature characterization, to the interdisciplinary projection matrix, to the adjustable coefficient. 7.The AI intelligent-based literature information retrieval and analysis system according to claim 1, wherein, In the recommendation list output module, the process of outputting the personalized recommendation list is: Construct a user feature vector U, including search behavior, reading record, citation pattern, cooperation network, field preference, and freshness demand: ； wherein, is a weight factor for a set search behavior, is a weight factor for a set reading record, is a weight factor for a set citation pattern, is a weight factor for a set collaboration network, is a weight factor for a set field preference, is a weight factor for a set freshness requirement; Combine the similarity of the user feature vector and the document representation vector, and the similarity of the similar user set to obtain a recommendation score: ； wherein, is the recommendation score for a document d by a user u, is the feature vector of a user u, is the enhanced document representation of a document d, is the feature vector of a user is the similarity between the enhanced document representation of a document d and the feature vector of a user is the set of similar users, v is one of the similar users, is the feature vector of a similar user v is the similarity between the enhanced document representation of a document d and the feature vector of a similar user is the cardinality of the set of similar users , that is, the number of users in the set.​​ According to the calculation All documents are sorted; Select the top x documents with the highest score as the personalized recommendation list; Output the personalized recommendation list. 8.The AI intelligent-based literature information retrieval and analysis system according to claim 1, wherein, In the knowledge graph feedback update module, the process of outputting the updated knowledge graph is: Real-time capture of user interaction behavior, construct a user interaction behavior set F: The user interaction behavior set F includes click, dwell time, download, citation, and is represented as: ； Computing behavior weights : ； wherein is an indicator function; Analyzing update entity influence weight : ； In the formula, is the old value of the entity influence weight; When the node weight change rate is higher than 15%, the local graph topology is reconstructed, and after the local graph topology reconstruction is completed, the updated knowledge graph is output by integrating into the global knowledge graph.

9. An AI intelligence-based literature information retrieval analysis method based on the system of any one of claims 1-8, characterized in that, Comprise the following steps: S1, extracting the text features of the literature to be searched, constructing a multi-modal literature representation model, and outputting a literature representation vector; S2, performing semantic enhanced retrieval, expanding user query, and outputting a preliminary retrieval result set and a corresponding similarity score set; S3, based on the similarity score, optimizing the preliminary retrieval result set to obtain an updated retrieval result set; S4, based on the updated retrieval result set, constructing a dynamic knowledge graph; S5, based on the dynamic knowledge graph, performing cross-disciplinary correlation analysis, combining the literature representation vector, and outputting an enhanced literature representation; S6, obtaining user portrait data, and combining the enhanced literature representation and the dynamic knowledge graph, outputting a personalized recommendation list; S7, collecting user interaction flow data, and combining the personalized recommendation list, outputting an updated knowledge graph.

Images