Intelligent popular science content retrieval interaction system and method combined with knowledge graph

Abstract

The invention discloses a science popularization content intelligent retrieval interaction system and method combined with a knowledge graph. The system comprises a knowledge graph construction unit, a multi-layer multi-head attention mechanism collaborative Transform coding unit, a knowledge graph embedding and Transform feature fusion unit, a retrieval result sorting unit based on fusion features, an interactive intention capture unit and a result optimization output unit. The method comprises the following steps: constructing a knowledge graph for multi-source science popularization data, extracting retrieval keyword features by applying Transform coding, fusing the knowledge graph and the retrieval keyword features, and calculating the similarity with science popularization content to finish initial sorting; and meanwhile, capturing user interaction behavior analysis intention change in real time, and readjusting and outputting an initial result in combination with knowledge graph semantic association. The method corresponds to operation steps of all units of the system, and intelligent retrieval interaction of popular science content is achieved. According to the system and the method, the problems of insufficient semantic understanding, poor interaction experience and the like of traditional retrieval are effectively solved, and the accuracy and the interaction intelligence of popular science content retrieval are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent retrieval, and in particular to an interactive system and method for intelligent retrieval of popular science content combined with a knowledge graph. Background Art

[0002] In the digital age, the means of disseminating popular science knowledge are becoming increasingly diverse. Multi-source, heterogeneous data, including text, images, and videos, constitutes a vast repository of popular science content. At the same time, public demand for popular science knowledge is becoming more diverse and personalized. Traditional search methods are unable to meet the demand for accurate and efficient access to popular science information. Building an intelligent and interactive system for popular science content retrieval is crucial for promoting the widespread dissemination of popular science knowledge and improving public scientific literacy.

[0003] Existing popular science content retrieval systems generally suffer from insufficient semantic understanding. Most systems rely on simple keyword matching techniques, which allow them to search only superficially within text and fail to deeply explore the semantic connections between entities within popular science content. For example, when a user searches for "the development history of artificial intelligence," the system may only return content containing the keywords "artificial intelligence" and "development history," ignoring semantically related information such as the evolution of related technologies and the contributions of key figures. This results in incomplete and inappropriate search results.

[0004] Furthermore, existing technologies have significant shortcomings in optimizing the user interaction experience. Traditional search systems typically present search results in a static manner and lack a dynamic analysis mechanism for user interaction behavior. During the user search process, behavioral data such as user browsing time, click preferences, and changes in secondary search keywords all contain information on adjustments to the user's search intent. However, existing systems are unable to capture and analyze this behavioral data in real time, making it difficult to dynamically optimize search results based on changes in user intent. This results in a lack of intelligence and personalization in the search interaction process, reducing the efficiency and experience of users in acquiring popular science knowledge. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides an interactive system and method for intelligent retrieval of popular science content combined with a knowledge graph.

[0006] The technical solution adopted by the present invention is an intelligent retrieval and interactive system for popular science content combined with a knowledge graph, including:

[0007] The knowledge graph construction unit is used to extract entities and identify relationships from multi-source popular science data, and to construct a knowledge graph containing entities, attributes, and relationships through semantic similarity calculation;

[0008] The Transformer encoding unit, which works with a multi-layer multi-head attention mechanism, converts the input popular science search keywords into word vectors and extracts features through a multi-layer stacked Transformer structure with different attention heads working together.

[0009] The knowledge graph embedding and Transformer feature fusion unit fuses the vector representations of entities and relationships in the knowledge graph with the feature vectors output by the Transformer encoding unit through tensor splicing and weighted summation;

[0010] The retrieval result ranking unit based on the fusion feature uses the fused feature vector to calculate the cosine similarity with each content in the popular science content library to generate the initial retrieval result ranking;

[0011] The interactive intention dynamic capture and feedback adjustment unit monitors user interactive behavior data in real time, extracts interactive behavior features, and analyzes changes in user retrieval intentions through features;

[0012] The retrieval result dynamic optimization output unit dynamically captures the intention changes analyzed by the feedback adjustment unit, and combines the semantic association information of the knowledge graph to readjust and output the initial retrieval result ranking.

[0013] The above units are connected in sequence. The knowledge graph construction unit outputs the knowledge graph to the knowledge graph embedding and Transformer feature fusion unit. The Transformer encoding unit with multi-layer multi-head attention mechanism outputs the feature vector to the knowledge graph embedding and Transformer feature fusion unit. The knowledge graph embedding and Transformer feature fusion unit transmits the fused vector to the retrieval result ranking unit based on the fused features. The retrieval result ranking unit based on the fused features outputs the initial retrieval result ranking to the retrieval result dynamic optimization output unit. The interactive intention dynamic capture and feedback adjustment unit transmits the analyzed intention change information to the retrieval result dynamic optimization output unit.

[0014] Furthermore, in the knowledge graph construction unit, the entity extraction formula used when constructing the knowledge graph is: Among them, E represents the extracted entity set, σ is the activation function, n is the number of features, and w i is the weight vector of the i-th feature, f i (D) is the i-th feature extracted from the multi-source popular science data D, b is the bias vector; the relationship recognition formula is: Among them, R represents the identified relationship set, is the set of all possible relations, and P(r|e1, e2, K) is the probability that there is a relation r between entity e1 and entity e2 in the knowledge graph K.

[0015] Furthermore, in the Transformer encoding unit with the multi-layer multi-head attention mechanism, the multi-head attention calculation adopts the formula: MultiHead(Q, K, V) = Concat(head1, ..., head h )W O ,in, Q is the query vector, K is the key vector, and V is the value vector. and W O is the trained weight matrix, h is the number of attention heads; the output feature vector calculation formula of the Transformer encoding unit is: H = LayerNorm(X + MHA(X)), where H is the output feature vector, X is the input word vector, MHA(X) is the calculation result of the multi-head attention mechanism on X, and LayerNorm is the layer normalization operation.

[0016] Furthermore, in the knowledge graph embedding and Transformer feature fusion unit, the knowledge graph embedding vector generation formula is: g = ∑ (s，p，o)∈G α (s，p，o) ·ψ(s, p, o), where g is the knowledge graph embedding vector, G is the set of triples in the knowledge graph, (s, p, o) represents the triple consisting of subject, predicate, and object, and α (s，p，o) is the weight of the triplet (s, p, o), ψ(s, p, o) is the vector representation of the triplet (s, p, o); the feature fusion formula is: F = β1·T+β2·g, where F is the fused feature vector, T is the feature vector output by the Transformer encoding unit, and β1 and β2 are fusion weight parameters.

[0017] Furthermore, in the retrieval result ranking unit based on fusion features, the cosine similarity is calculated using the formula: Among them, sim(F, C i ) represents the fusion feature vector F and the popular science content C i The cosine similarity of C i is the vector representation of the i-th popular science content in the popular science content library, ||F|| and ||C i || are vectors F and C respectively i The norm of the initial search results is sorted by the formula: rank init =sort([sim(F,C1),…,sim(F,C m )], desc = True), where rank initRanking the initial search results, m is the number of science popularization contents in the science popularization content library.

[0018] Further, in the interactive intention dynamic capture and feedback adjustment unit, the interactive behavior feature extraction formula is B = φ (A), wherein B is the extracted interactive behavior feature vector, A is the user interactive behavior data, and φ is the feature extraction function; the intention change analysis formula is ΔI = γ · (B - B prev ), wherein ΔI represents the user search intention change amount, B prev is the interactive behavior feature vector at the previous moment, and γ is the weight coefficient.

[0019] Further, in the search result dynamic optimization output unit, the weight adjustment formula based on the semantic association of the knowledge graph is: wherein w adj is the weight adjustment value of the search result, θ is the adjustment coefficient, Path (e q , e c ) is the path set of the entity e q corresponding to the query keyword in the knowledge graph and the entity e c in the science popularization content, (e i , r, e j ) is a triple in the path, and ρ (r) is the weight of the relationship r; the re-adjusted search result ranking formula is: rank final = sort ([sim (F, C1) + w adj1 , …, sim (F, C m ) + w adjm ], desc = True), wherein rank final is the re-adjusted search result ranking, and w adji is the weight adjustment value corresponding to the i-th science popularization content.

[0020] Further, between the knowledge graph construction unit and the Transformer encoding unit cooperating with the multi-layer multi-head attention mechanism, a data format adaptation conversion module is arranged, and the data conversion formula is: D trans = τ (D kg , D tf ), wherein D trans is the converted data, D kg is the data output by the knowledge graph construction unit, D tf is the data required by the Transformer encoding unit in the required format, and τ is the data conversion function.

[0021] Further, between the interactive intention dynamic capture and feedback adjustment unit and the search result dynamic optimization output unit, there is an intention change and search result association analysis module, and the association analysis formula is: Arel =μ(ΔI, rank init ), where A rel is the correlation between intention change and retrieval results, μ is the correlation analysis function, ΔI is the change in user retrieval intention, rank init Sort the initial search results.

[0022] The interactive method for intelligent retrieval of popular science content combined with knowledge graph includes the following steps:

[0023] Step S1: Start a knowledge graph construction program based on semantic association mining of heterogeneous data, perform entity extraction and relationship recognition operations on multi-source popular science data, and construct a knowledge graph data structure containing entities, attributes and relationships according to semantic similarity calculation rules;

[0024] Step S2: Run the Transformer encoding program with a multi-layer multi-head attention mechanism to convert the input popular science search keywords into word vectors, and perform feature extraction operations through a multi-layer stacked Transformer architecture with different attention heads working together to obtain keyword feature vectors;

[0025] Step S3: Execute the knowledge graph embedding and Transformer feature fusion program to fuse the vector representations of entities and relationships in the knowledge graph with the keyword feature vectors output by the Transformer encoding program according to the operation rules of tensor splicing and weighted summation to obtain a fused feature vector;

[0026] Step S4: Start the retrieval result sorting program based on the fusion feature, use the fusion feature vector and the vectors of each content in the popular science content library to generate an initial retrieval result sorting sequence through cosine similarity calculation;

[0027] Step S5: activating the interactive intention dynamic capture and feedback adjustment program, monitoring the user interactive behavior data in real time, using a preset feature extraction algorithm to extract interactive behavior features, and analyzing and judging changes in user search intentions based on the features;

[0028] Step S6: Start the dynamic optimization output program for the retrieval results, dynamically capture the interactive intention, and adjust the intention change information obtained by the feedback analysis program. Combined with the semantic association information of the knowledge graph, the initial retrieval result sorting sequence is readjusted, and the adjusted retrieval results are output and displayed.

[0029] Beneficial effects: The present invention proposes an interactive system and method for intelligent retrieval of popular science content combined with a knowledge graph. The present invention uses a knowledge graph construction unit to conduct in-depth analysis of multi-source popular science data, extract entities, identify relationships and construct a knowledge graph, and present the semantic associations in the popular science content in a structured manner; at the same time, the Transformer encoding unit coordinated by a multi-layer multi-head attention mechanism performs in-depth feature extraction on the retrieval keywords, and then through the knowledge graph embedding and Transformer feature fusion unit, the semantic information of the knowledge graph is organically combined with the keyword features, thereby achieving a deep semantic understanding of the retrieval needs and popular science content, so that the retrieval results can more comprehensively and accurately meet user needs.

[0030] In order to address the shortcomings of traditional technologies in optimizing user interaction experience, the present invention uses an interaction intention dynamic capture and feedback adjustment unit to monitor users' browsing time, click behavior and other interaction data in real time, extract interaction behavior features and analyze changes in user search intentions; the search result dynamic optimization output unit combines the semantic association information of the knowledge graph to readjust the initial search result ranking according to changes in user intentions, thereby realizing dynamic optimization of the search results, providing users with a personalized and intelligent search interaction experience, significantly improving the efficiency and satisfaction of users in acquiring popular science knowledge, and promoting the efficient dissemination of popular science knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a diagram of the system unit composition of the present invention;

[0032] Figure 2 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0033] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown in the figure, the intelligent retrieval and interactive system for popular science content combined with the knowledge graph includes:

[0035] The knowledge graph construction unit is used to extract entities and identify relationships from multi-source popular science data, and to construct a knowledge graph containing entities, attributes, and relationships through semantic similarity calculation;

[0036] Specifically, the core function of this unit is to deeply process multi-source popular science data and construct a structured knowledge graph. Popular science data comes from a wide range of sources, including text, images, videos, and other forms. These heterogeneous data contain rich popular science knowledge, but lack orderly organization. The unit uses entity extraction technology to identify key entities from massive data. These entities can be popular science concepts, scientific terms, people, events, etc. Then, using relationship recognition technology, it analyzes the associations between entities, such as causal relationships, subordinate relationships, and parallel relationships. Finally, through semantic similarity calculation, the extracted entities and identified relationships are integrated to construct a knowledge graph containing entities, attributes, and relationships. This process transforms the originally scattered and disordered popular science knowledge into a structured and semantic knowledge network, laying the foundation for subsequent retrieval and analysis.

[0037] During the implementation process, the unit first pre-processes multi-source popular science data and unifies the data format to facilitate subsequent processing. Then, algorithms and models in fields such as natural language processing and computer vision are used to extract entities and identify relationships. For example, when processing text data, named entity recognition algorithms can be used to find the entities therein, and the relationships between entities can be determined through syntactic analysis and semantic understanding algorithms. For image and video data, it is necessary to use technologies such as image recognition and video content analysis to extract key information and convert it into entities and relationships that can be used to construct knowledge graphs. By continuously optimizing algorithms and models, the accuracy and completeness of entity extraction and relationship identification can be improved, ensuring that the constructed knowledge graph can truly and comprehensively reflect the knowledge system in the field of popular science.

[0038] The Transformer encoding unit, which works with a multi-layer multi-head attention mechanism, converts the input popular science search keywords into word vectors and extracts features through a multi-layer stacked Transformer structure with different attention heads working together.

[0039] Specifically, this unit is mainly responsible for deep feature extraction of the input popular science search keywords. In the popular science search scenario, the keywords entered by users are often relatively concise, but they may contain complex semantic information. The unit first converts the keywords into word vectors, mapping each word into a vector in a high-dimensional space, so that the semantic relationship between words can be reflected through vector operations. Then, the word vectors are processed using a Transformer structure with multiple layers stacked and different attention heads working together. Each attention head can capture the semantic associations between keywords from different angles, and the multi-layer structure can gradually abstract and refine these semantic associations, and finally output a vector representation containing rich semantic features. In this way, the unit can deeply understand the user's search intentions and provide more representative features for subsequent retrieval and matching.

[0040] In implementation, the multi-layer structure in the Transformer encoding unit processes the word vectors in a certain order. Each layer contains multiple attention heads that work in parallel to calculate different attention weights, capturing information at different semantic levels. During the calculation, the model focuses on the keywords related to the user's search intent by adjusting the attention weights, while ignoring irrelevant information. Meanwhile, the unit also uses techniques such as normalization and residual connection to optimize the model's training and performance, preventing gradient vanishing or explosion problems and improving the model's stability and generalization ability. After the collaborative processing of multiple layers and multiple attention heads, the final feature vector accurately represents the user's input keywords, providing strong support for precise science popularization content retrieval.

[0041] The knowledge graph embedding and Transformer feature fusion unit fuses the vector representation of entities and relationships in the knowledge graph with the feature vector output by the Transformer encoding unit through tensor concatenation and weighted summation.

[0042] Specifically, the main task of this unit is to organically fuse the semantic information contained in the knowledge graph with the keyword features extracted by the Transformer encoding unit. The knowledge graph stores rich knowledge in the field of science popularization in a structured form, and the vector representation of entities and relationships contains a large amount of semantic association information. The feature vector output by the Transformer encoding unit accurately reflects the user's search intent. Through tensor concatenation and weighted summation, this unit fuses vectors from two different sources, making the fused vector contain both the global semantic information of the knowledge graph and the local features of the user's search keywords, forming a more comprehensive and expressive feature representation. This fused feature can better describe the relationship between the user's search requirements and the science popularization content, providing a more effective basis for the sorting and screening of search results.

[0043] During the implementation process, it is first necessary to convert the entities and relationships in the knowledge graph into vector representations, which can be achieved through knowledge graph embedding technology, such as TransE, RotatE and other algorithms. At the same time, obtain the keyword feature vector output by the Transformer encoding unit. Then, the knowledge graph embedding vector and the keyword feature vector are connected according to certain rules using tensor splicing to form a new vector. Then, through the weighted summation operation, according to the importance of different data sources, corresponding weights are assigned to different parts of the spliced ​​vector to obtain the final fused feature vector. When determining the weights, model training and optimization algorithms can be used to adjust according to the actual retrieval effect, so that the fused features can maximize the accuracy and relevance of the retrieval, thereby improving the performance of the entire retrieval system.

[0044] The retrieval result ranking unit based on the fusion feature uses the fused feature vector to calculate the cosine similarity with each content in the popular science content library to generate the initial retrieval result ranking;

[0045] Specifically, this unit calculates and sorts the similarity of all the contents in the popular science content library based on the fused feature vector output by the knowledge graph embedding and the Transformer feature fusion unit. The fused feature vector integrates the semantic information of the user's search intention and popular science knowledge, and can accurately describe the degree of match between user needs and popular science content. The unit measures the similarity between them by calculating the cosine similarity between the fused feature vector and each content vector in the popular science content library. The higher the value of the cosine similarity, the higher the degree of match between the popular science content and the user's search needs. According to the calculated similarity value, the content in the popular science content library is sorted to generate the initial search result sorting, and the popular science content most relevant to the user's needs is placed in front, providing the user with a search result list that better meets their expectations.

[0046] In specific implementation, the unit first converts each popular science content in the popular science content library into a vector representation, which can be achieved through technologies similar to knowledge graph embedding or Transformer encoding, ensuring that the vector representation of the popular science content is in the same semantic space as the fused feature vector. Then, for each popular science content vector, the cosine similarity is calculated with the fused feature vector. In order to improve computational efficiency, some optimization algorithms and data structures, such as KD trees, hash tables, etc., can be used to reduce the amount of calculation and query time. After completing the similarity calculation of all popular science content, the popular science content is sorted from high to low according to the similarity value to generate the initial search results. In addition, the unit can also further adjust and optimize the sorting results based on actual needs and in combination with other factors, such as the update time and popularity of the popular science content, so that the final search results can better meet the diverse needs of users.

[0047] The interactive intention dynamic capture and feedback adjustment unit monitors user interactive behavior data in real time, extracts interactive behavior features, and analyzes changes in user retrieval intentions through features;

[0048] Specifically, this unit focuses on real-time monitoring of user behavioral data during the retrieval interaction process, extracting key features from it, and analyzing the dynamic changes in user retrieval intentions. When users use the popular science retrieval interactive system, their browsing behavior, clicking behavior, input behavior, etc. will generate a large amount of data. These data contain implicit information about the user's satisfaction with the retrieval results and the adjustment of retrieval intentions. The unit collects and analyzes these behavioral data in real time to extract features that can reflect changes in user intentions, such as the user's dwell time on a certain retrieval result, the frequency of clicking on different types of popular science content, and the modification of search keywords. Through in-depth analysis of these features, it is possible to determine whether the user's retrieval intention has changed, as well as the direction and degree of change, providing a basis for the dynamic optimization of retrieval results.

[0049] In terms of implementation, the unit first establishes a user behavior data collection mechanism, and collects various user behavior data in real time by setting data collection points in the system interface and interaction process. Then, data mining and machine learning algorithms are used to preprocess and extract features from the collected data. For example, user browsing time data is processed by time series analysis methods, and user click behavior data is analyzed by classification algorithms. Next, an intention change analysis model is constructed, and the extracted behavioral features are used to predict changes in user retrieval intentions. The model can be based on deep learning architectures, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), to capture time series information and long-term dependencies in user behavior data. By continuously training and optimizing the model, the accuracy and timeliness of the analysis of changes in user retrieval intentions can be improved, ensuring that subtle changes in user intentions can be captured in a timely manner, and providing accurate guidance for the adjustment of retrieval results.

[0050] The retrieval result dynamic optimization output unit dynamically captures the intention changes analyzed by the feedback adjustment unit, combines the semantic association information of the knowledge graph, and readjusts and outputs the initial retrieval result ranking.

[0051] Specifically, this unit dynamically captures the interactive intention and obtains the user search intention change information obtained through analysis by the feedback adjustment unit, and combines it with the semantic association information of the knowledge graph to readjust and output the initial search result ranking. When a change in the user's search intention is detected, the original initial search results may no longer meet the user's needs. At this time, the unit uses the rich semantic associations in the knowledge graph to mine popular science content related to the user's new intention and adjust the ranking weights of these contents. By considering the relationship, attributes, and semantic similarity between entities in the knowledge graph, the degree of match between each popular science content and the user's new intention is re-evaluated, thereby dynamically optimizing the initial search results so that the adjusted search results can better meet the user's current search needs and provide users with more accurate popular science content that is more in line with their intentions.

[0052] During the specific implementation process, the unit first receives the user search intention change information transmitted by the interactive intention dynamic capture and feedback adjustment unit, and clarifies the direction and focus of the user's intention change. Then, based on the semantic association information of the knowledge graph, each popular science content in the initial search results is re-evaluated. For example, based on the path relationship between entities in the knowledge graph, the degree of indirect association between popular science content and the user's new intention is judged; through the attribute information of the entity, popular science content that better meets the user's specific needs is screened out. During the evaluation process, a specific weight calculation method is used to comprehensively consider multiple factors to reallocate the ranking weight for each popular science content. Finally, the popular science content is re-sorted according to the new weight, and the adjusted search results are output and displayed to the user. At the same time, the unit can also optimize the display format of the output results according to the system performance and user experience requirements, such as pagination display, highlighting key content, etc., to improve the convenience and satisfaction of users in obtaining popular science knowledge.

[0053] The above units are connected in sequence. The knowledge graph construction unit outputs the knowledge graph to the knowledge graph embedding and Transformer feature fusion unit. The Transformer encoding unit with multi-layer multi-head attention mechanism outputs the feature vector to the knowledge graph embedding and Transformer feature fusion unit. The knowledge graph embedding and Transformer feature fusion unit transmits the fused vector to the retrieval result ranking unit based on the fused features. The retrieval result ranking unit based on the fused features outputs the initial retrieval result ranking to the retrieval result dynamic optimization output unit. The interactive intention dynamic capture and feedback adjustment unit transmits the analyzed intention change information to the retrieval result dynamic optimization output unit.

[0054] Preferably, in the knowledge graph construction unit, the entity extraction formula used when constructing the knowledge graph is: Among them, E represents the extracted entity set, σ is the activation function, n is the number of features, and wi is the weight vector of the i-th feature, f i (D) is the i-th feature extracted from the multi-source popular science data D, b is the bias vector; the relationship recognition formula is: Among them, R represents the identified relationship set, is the set of all possible relations, and P(r|e1, e2, K) is the probability that there is a relation r between entity e1 and entity e2 in the knowledge graph K.

[0055] Specifically, in the entity extraction stage, by performing feature analysis on multi-source popular science data, each feature is assigned a specific weight, and then processed by the activation function to achieve accurate screening of key entities. The weight and bias parameters can be iteratively optimized through model training to ensure the accuracy of entity extraction. In terms of relationship recognition, based on the constructed knowledge graph structure, the system calculates the probability values ​​of various types of relationships between entity pairs, and selects the relationship type with the highest probability as the recognition result. This process makes full use of the global semantic information of the knowledge graph and enhances the reliability and logic of relationship recognition. In actual implementation, the data must be preprocessed first, and the above algorithm must be applied after the format is unified. At the same time, the parameters and algorithm models must be continuously optimized to improve the quality and efficiency of knowledge graph construction.

[0056] Preferably, in the Transformer encoding unit coordinated by the multi-layer multi-head attention mechanism, the multi-head attention calculation adopts the formula: MultiHead(Q, K, V) = Concat(head1, ..., head h )W O ,in, Q is the query vector, K is the key vector, and V is the value vector. and W O is the trained weight matrix, h is the number of attention heads; the output feature vector calculation formula of the Transformer encoding unit is: H = LayerNorm(X + MHA(X)), where H is the output feature vector, X is the input word vector, MHA(X) is the calculation result of the multi-head attention mechanism on X, and LayerNorm is the layer normalization operation.

[0057] Specifically, the multi-head attention mechanism processes the query, key and value vectors in parallel through multiple independent attention heads, each of which captures the semantic association of the input sequence from a different perspective, and then splices the output results of each head and converts them through a weight matrix to realize multi-dimensional extraction of semantic information. In the output feature vector calculation stage, the data distribution is stabilized through layer normalization operation, the gradient vanishing is avoided through residual connection, and the original word vector and the multi-head attention calculation result are fused to finally output a vector containing deep semantic features. In the implementation process, the number of attention heads, weight matrix and other parameters need to be reasonably set, and the model needs to be trained and optimized through a large amount of data, so that the encoding unit can accurately extract the core semantic features of the popular science retrieval keywords and provide strong support for subsequent retrieval.

[0058] Preferably, the knowledge graph embedding is in the Transformer feature fusion unit, and the knowledge graph embedding vector generation formula is: g = ∑ (s，p，o)∈G α (s，p，o) · ψ(s, p, o), wherein g is the knowledge graph embedding vector, G is the set of triples in the knowledge graph, (s, p, o) represents a triple composed of a subject, a predicate and an object, α (s，p，o) is the weight of the triple (s, p, o), and ψ(s, p, o) is the vector representation of the triple (s, p, o); the feature fusion formula is: F = β1·T + β2·g, wherein F is the fused feature vector, T is the feature vector output by the Transformer encoding unit, and β1 and β2 are fusion weight parameters.

[0059] Specifically, when generating the knowledge graph embedding vector, each triple in the graph is assigned a weight, and the weighted sum of the vector representations is obtained, so that the semantic importance of different triples can be reflected in the embedding vector. In the feature fusion stage, the knowledge graph embedding vector and the feature vector output by the Transformer encoding unit are combined in proportion by setting the fusion weight parameters, and the weight parameters can be dynamically adjusted according to the actual retrieval effect to balance the contribution of the two to the fused feature vector. In the implementation, the graph structure needs to be first converted into a vector representation using the knowledge graph embedding algorithm, and then fused and calculated with the Transformer feature, and the weight parameters are continuously optimized so that the fused feature can reflect both the global semantics of the knowledge graph and the local features of the user's retrieval intention.

[0060] Preferably, in the retrieval result sorting unit based on the fused feature, the cosine similarity calculation uses the formula: wherein sim(F, C i ) represents the cosine similarity between the fused feature vector F and the popular science content C i , C i is the vector representation of the i-th popular science content in the popular science content library, and ||F|| and ||C i|| are vectors F and C respectively i The norm of the initial search results is sorted by the formula: rank init =sort([sim(F,C1),…,sim(F,C m )], desc = True), where rank init is the sorting of the initial search results, and m is the number of popular science contents in the popular science content library.

[0061] Specifically, the system quantifies the degree of matching between the fused feature vector and the popular science content vector by calculating the cosine similarity in the semantic space. The higher the similarity value, the stronger the content relevance. Based on the similarity calculation results, all the contents in the popular science content library are sorted in descending order to generate the initial search results. During implementation, each content in the popular science content library must first be converted into a vector form to ensure that it is in the same semantic space as the fused feature vector, and then similarity calculation and sorting operations are performed. In order to improve computing efficiency, optimized data structures and algorithms can be used. At the same time, the sorting results can be adjusted again in combination with additional information such as the update time and popularity of the content to improve the practicality of the search results.

[0062] Preferably, in the interactive intention dynamic capture and feedback adjustment unit, the interactive behavior feature extraction formula is: B = φ (A), where B is the extracted interactive behavior feature vector, A is the user interactive behavior data, and φ is the feature extraction function; the intention change analysis formula is: ΔI = γ·(BB prev ), where ΔI represents the change in user search intention, B prev is the interaction behavior feature vector of the previous moment, and γ is the weight coefficient.

[0063] Specifically, interactive behavior feature extraction processes user behavior data such as browsing time and click operations through specific functions and converts them into feature vectors that can represent user behavior patterns. Intent change analysis quantifies the degree of change in user search intent by comparing the difference in behavioral feature vectors between the current and previous moments and multiplying them by a weight coefficient. The weight coefficient is set based on the difference in the impact of different behavior types on intent changes. During implementation, user behavior data needs to be collected in real time, and feature extraction must be performed using data mining algorithms. At the same time, a reasonable intent change analysis model must be constructed. By continuously training and optimizing model parameters, it is ensured that the dynamic changes in user search intent can be captured in a timely and accurate manner, providing a basis for optimizing search results.

[0064] Preferably, in the retrieval result dynamic optimization output unit, the weight adjustment formula based on the semantic association of the knowledge graph is: Among them, w adj is the weight adjustment value of the search result, θ is the adjustment coefficient, Path(e q , ec ) is the entity e corresponding to the query keyword q and entities in popular science content c The path set in the knowledge graph, (e i ,r,e j ) is the triple in the path, ρ(r) is the weight of relation r; the ranking formula of the readjusted retrieval results is: rank final =sort([sim(F,C1)+w adj1 ,…,sim(F,C m )+w adjm ], desc = True), where rank final is the sorting of the readjusted search results, w adji is the weight adjustment value corresponding to the i-th popular science content.

[0065] Specifically, the system assigns weights to the relationships on the path based on the entity path relationship between the query keywords and the popular science content in the knowledge graph, sums up all the path relationship weights and multiplies them by the adjustment coefficient to obtain the weight adjustment value of each popular science content. This value reflects the degree of association between the content and the user query at the semantic level of the knowledge graph. Based on the initial search results, the cosine similarity score is added to the weight adjustment value, and the results are re-sorted in descending order to generate the final search results. During implementation, it is necessary to first build an efficient knowledge graph path search algorithm, accurately calculate the path relationship weights, and then adjust the results based on the initial similarity score to ensure that the optimized search results are more in line with the user's dynamically changing search intentions.

[0066] Preferably, a data format adaptation conversion module is provided between the knowledge graph construction unit and the Transformer encoding unit coordinated with the multi-layer multi-head attention mechanism, and its data conversion formula is: D trans =τ(D kg , D tf ), where D trans is the converted data, D kg The data output by the knowledge graph construction unit, D tf is the data in the format required by the Transformer encoding unit, and τ is the data conversion function.

[0067] Specifically, due to the differences in data formats and structures processed by the two units, the data format adaptation and conversion module uses specific data conversion functions to map the structured data output by the knowledge graph construction unit into the format required by the Transformer encoding unit, ensuring smooth data transmission and processing between the two units. During implementation, it is necessary to deeply analyze the data characteristics of the two units and design targeted data conversion rules and functions to preserve the key semantic information of the data during the conversion process, avoiding data loss or semantic deviation due to format conversion, and laying the foundation for subsequent feature extraction and fusion.

[0068] Preferably, there is an intention change and retrieval result association analysis module between the interaction intention dynamic capture and feedback adjustment unit and the retrieval result dynamic optimization output unit, and its association analysis formula is: rel =μ(ΔI, rank init ), where A rel is the correlation between intention change and retrieval results, μ is the correlation analysis function, ΔI is the change in user retrieval intention, rank init Sort the initial search results.

[0069] Specifically, the modules analyze the association between interaction intent and search results, and the association between intent changes and search results. These modules utilize association analysis functions to comprehensively analyze changes in user search intent and the initial search result rankings, exploring the inherent connections and influence patterns between the two and clarifying the degree to which changes in intent affect the relevance of each search result. During implementation, a suitable association analysis model must be constructed. Through analysis and training of large amounts of user interaction data and search results, the model parameters are optimized to enable the model to accurately identify the association between intent changes and search results. This provides a precise basis for adjusting the dynamic optimization output unit for search results, thereby enhancing the interactive intelligence and user satisfaction of the search system.

[0070] like Figure 2 As shown in FIG, the interactive method for intelligent retrieval of popular science content combined with knowledge graph includes the following steps:

[0071] Step S1: Start a knowledge graph construction program based on semantic association mining of heterogeneous data, perform entity extraction and relationship recognition operations on multi-source popular science data, and construct a knowledge graph data structure containing entities, attributes and relationships according to semantic similarity calculation rules;

[0072] Step S2: Run the Transformer encoding program with a multi-layer multi-head attention mechanism to convert the input popular science search keywords into word vectors, and perform feature extraction operations through a multi-layer stacked Transformer architecture with different attention heads working together to obtain keyword feature vectors;

[0073] Step S3: Execute the knowledge graph embedding and Transformer feature fusion program to fuse the vector representations of entities and relationships in the knowledge graph with the keyword feature vectors output by the Transformer encoding program according to the operation rules of tensor splicing and weighted summation to obtain a fused feature vector;

[0074] Step S4: Start the retrieval result sorting program based on the fusion feature, use the fusion feature vector and the vectors of each content in the popular science content library to generate an initial retrieval result sorting sequence through cosine similarity calculation;

[0075] Step S5: activating the interactive intention dynamic capture and feedback adjustment program, monitoring the user interactive behavior data in real time, using a preset feature extraction algorithm to extract interactive behavior features, and analyzing and judging changes in user search intentions based on the features;

[0076] Step S6: Start the dynamic optimization output program for the retrieval results, dynamically capture the interactive intention, and adjust the intention change information obtained by the feedback analysis program. Combined with the semantic association information of the knowledge graph, the initial retrieval result sorting sequence is readjusted, and the adjusted retrieval results are output and displayed.

[0077] This application has constructed a multi-level semantic parsing system. The knowledge graph construction unit can accurately identify entities from multi-source popular science data such as text and images, deeply explore the logical relationships between entities, integrate fragmented popular science knowledge into a structured and semantic knowledge network, and realize deep semantic modeling of popular science content. The Transformer encoding unit, which works in conjunction with a multi-layer multi-head attention mechanism, performs refined feature extraction on the search keywords input by the user to capture the semantic associations behind the keywords. The two are organically combined through the knowledge graph embedding and the Transformer feature fusion unit to deeply fuse the structured semantic information of the knowledge graph with the search keyword features, so that the system can accurately understand the user's search intention and the semantic connotation of popular science content, and completely change the shallow search mode of traditional keyword matching. The search results not only include surface keywords, but also cover rich semantic extension content related to them, which greatly improves the comprehensiveness and accuracy of the search results.

[0078] In terms of optimizing the user interaction experience, the system has established a dynamic interactive response mechanism. The interactive intention dynamic capture and feedback adjustment unit monitors the user's browsing time, click behavior, secondary search keywords and other interactive data during the retrieval process in real time, accurately extracts user behavior characteristics and analyzes the dynamic change trend of their retrieval intentions. The retrieval result dynamic optimization output unit combines the semantic association information of the knowledge graph, intelligently rearranges the initial retrieval results in real time according to the changes in user intentions, and dynamically adjusts the display priority of the retrieval results. Compared with the traditional static presentation of retrieval results, the system realizes the real-time adaptation of retrieval services and user needs, providing users with personalized and dynamic intelligent retrieval interactive experience, effectively solving the problems of untimely interactive feedback in traditional systems and difficulty in meeting the real-time needs of users in retrieval results, and significantly improving the efficiency and satisfaction of users in acquiring popular science knowledge.

[0079] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0080] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent retrieval and interactive system for popular science content combined with knowledge graphs, characterized by: include: The knowledge graph construction unit is used to extract entities and identify relationships from multi-source popular science data, and to construct a knowledge graph containing entities, attributes, and relationships through semantic similarity calculation; The Transformer encoding unit, which works with a multi-layer multi-head attention mechanism, converts the input popular science search keywords into word vectors and extracts features through a multi-layer stacked Transformer structure with different attention heads working together. The knowledge graph embedding and Transformer feature fusion unit fuses the vector representations of entities and relationships in the knowledge graph with the feature vectors output by the Transformer encoding unit through tensor splicing and weighted summation; The retrieval result ranking unit based on the fusion feature uses the fused feature vector to calculate the cosine similarity with each content in the popular science content library to generate the initial retrieval result ranking; The interactive intention dynamic capture and feedback adjustment unit monitors user interactive behavior data in real time, extracts interactive behavior features, and analyzes changes in user retrieval intentions through features; The retrieval result dynamic optimization output unit dynamically captures the intention changes analyzed by the feedback adjustment unit, and combines the semantic association information of the knowledge graph to readjust and output the initial retrieval result ranking.

2. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the knowledge graph construction unit, the entity extraction formula used when constructing the knowledge graph is: Among them, E represents the extracted entity set, σ is the activation function, n is the number of features, and w i is the weight vector of the i-th feature, f i (D) is the i-th feature extracted from the multi-source popular science data D, b is the bias vector; the relationship recognition formula is: Among them, R represents the identified relationship set, is the set of all possible relations, and P(r|e1, e2, K) is the probability that there is a relation r between entity e1 and entity e2 in the knowledge graph K.

3. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the Transformer encoding unit of the multi-layer multi-head attention mechanism, the multi-head attention calculation adopts the formula: MultiHead(Q, K, V) = Concat(head1, ..., head h )W O ,in, Q is the query vector, K is the key vector, and V is the value vector. and W O is the trained weight matrix, h is the number of attention heads; the output feature vector calculation formula of the Transformer encoding unit is: H = LayerNorm(X + MHA(X)), where H is the output feature vector, X is the input word vector, MHA(X) is the calculation result of the multi-head attention mechanism on X, and LayerNorm is the layer normalization operation.

4. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the knowledge graph embedding and Transformer feature fusion unit, the knowledge graph embedding vector generation formula is: g = ∑ (s，p，o)∈G α (s，p，o) ·ψ(s, p, o), where g is the knowledge graph embedding vector, G is the set of triples in the knowledge graph, (s, p, o) represents the triple consisting of subject, predicate, and object, and α (s，p，o) is the weight of the triplet (s, p, o), ψ(s, p, o) is the vector representation of the triplet (s, p, o); the feature fusion formula is: F = β1·T+β2·g, where F is the fused feature vector, T is the feature vector output by the Transformer encoding unit, and β1 and β2 are fusion weight parameters.

5. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the retrieval result ranking unit based on fusion features, the cosine similarity is calculated using the formula: Among them, sim(F, C i ) represents the fusion feature vector F and the popular science content C i The cosine similarity of C i is the vector representation of the i-th popular science content in the popular science content library, ||F|| and ||C i || are vectors F and C respectively i The norm of the initial search results is sorted by the formula: rank init =sort([sim(F,C1),…,sim(F,C m )], desc = True), where rank init is the sorting of the initial search results, and m is the number of popular science contents in the popular science content library.

6. The intelligent retrieval and interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the interactive intention dynamic capture and feedback adjustment unit, the interactive behavior feature extraction formula is: B = φ (A), where B is the extracted interactive behavior feature vector, A is the user interactive behavior data, and φ is the feature extraction function; the intention change analysis formula is: ΔI = γ·(BB prev ), where ΔI represents the change in user search intention, B prev is the interaction behavior feature vector of the previous moment, and γ is the weight coefficient.

7. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: In the retrieval result dynamic optimization output unit, the weight adjustment formula based on the semantic association of the knowledge graph is: Among them, w adj is the weight adjustment value of the search result, θ is the adjustment coefficient, Path(e q , e c ) is the entity e corresponding to the query keyword q and entities in popular science content c The path set in the knowledge graph, (e i ,r,e j ) is the triple in the path, ρ(r) is the weight of relation r; the ranking formula of the readjusted retrieval results is: rank final =sort([sim(F,C1)+w adj1 ,…,sim(F,C m )+w adjm ], desc = True), where rank final is the sorting of the readjusted search results, w adji is the weight adjustment value corresponding to the i-th popular science content.

8. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1 is characterized in that: A data format adaptation conversion module is provided between the knowledge graph construction unit and the Transformer encoding unit coordinated with the multi-layer multi-head attention mechanism, and its data conversion formula is: D trans =τ(D kg , D tf ), where D trans is the converted data, D kg The data output by the knowledge graph construction unit, D tf is the data in the format required by the Transformer encoding unit, and τ is the data conversion function.

9. The intelligent retrieval interactive system for popular science content combined with knowledge graph according to claim 1, characterized in that: Between the interactive intention dynamic capture and feedback adjustment unit and the retrieval result dynamic optimization output unit, there is an intention change and retrieval result association analysis module, and its association analysis formula is: ref =μ(ΔI, rank init ), where A rel is the correlation between intention change and retrieval results, μ is the correlation analysis function, ΔI is the change in user retrieval intention, rank init Sort the initial search results.

10. An interactive method for intelligent retrieval of popular science content combined with knowledge graphs, characterized in that: The following steps are involved: Step S1: Start a knowledge graph construction program based on semantic association mining of heterogeneous data, perform entity extraction and relationship recognition operations on multi-source popular science data, and construct a knowledge graph data structure containing entities, attributes and relationships according to semantic similarity calculation rules; Step S2: Run the Transformer encoding program with a multi-layer multi-head attention mechanism to convert the input popular science search keywords into word vectors, and perform feature extraction operations through a multi-layer stacked Transformer architecture with different attention heads working together to obtain keyword feature vectors; Step S3: Execute the knowledge graph embedding and Transformer feature fusion program to fuse the vector representations of entities and relationships in the knowledge graph with the keyword feature vectors output by the Transformer encoding program according to the operation rules of tensor splicing and weighted summation to obtain a fused feature vector; Step S4: Start the retrieval result sorting program based on the fusion feature, and use the fusion feature vector and the vectors of each content in the popular science content library to generate an initial retrieval result sorting sequence through cosine similarity calculation; Step S5: activating the interactive intention dynamic capture and feedback adjustment program, monitoring the user interactive behavior data in real time, using a preset feature extraction algorithm to extract interactive behavior features, and analyzing and judging changes in user search intentions based on the features; Step S6: Start the dynamic optimization output program for the retrieval results, dynamically capture the interactive intention, and adjust the intention change information obtained by the feedback analysis program. Combined with the semantic association information of the knowledge graph, the initial retrieval result sorting sequence is readjusted, and the adjusted retrieval results are output and displayed.

Images