Cultural computing-based curation method

Abstract

The invention discloses a cultural calculation-based curation method, and the specific implementation scheme is as follows: inputting multi-modal data, theme data and an auxiliary knowledge graph of candidate exhibits, and outputting a theme correlation pool; converting the exhibit and user data of the theme-related pool into a core exhibit list, and performing descending sorting on the core exhibit list according to the final sorting score and the grouping sorting; and integrating the core exhibit lists after descending sorting and carrying out exhibition physical constraint to obtain an exhibition scheme of the core exhibit lists. According to the method, the problems that the relevance between the exhibit and the exhibition theme of the traditional policy exhibition method often depends on subjective judgment, the exhibition scheme is difficult to accurately match the interests and preferences of different audience groups, and the narrative logic of the exhibition is scattered are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cultural computing, and in particular to a curation method based on cultural computing. BACKGROUND

[0002] With the development of information technology, social computing, big data and artificial intelligence have injected new vitality into humanistic research, and have given birth to the interdisciplinary research field of "cultural computing". Cultural computing aims to model, analyze and interpret cultural phenomena through computing technology, to realize the deep mining, effective dissemination and innovative development of cultural content, and to promote digital humanities research and promote the prosperity of culture.

[0003] Curation, as an important means of cultural heritage and knowledge dissemination, has long relied on the professional knowledge, artistic intuition and personal experience of curators. This "experience-driven" mode centered on qualitative analysis has unique value in the excavation of humanistic connotations, but faces many challenges in terms of scale, precision and scientificity. There are several significant problems with traditional curation methods: first, the relevance between exhibits and exhibition themes often depends on subjective judgment, resulting in "fuzzy theme relevance" and a lack of objective quantitative basis; second, the curation plan is difficult to accurately match the interest preferences of different audience groups, leading to "insufficient user appeal"; finally, the narrative logic of the exhibition is mostly presented in a loose linear or parallel structure, leading to "scattered narrative logic" and making it difficult to form a story line with strong appeal and logical consistency. SUMMARY

[0004] Therefore, the present application provides a curation method based on cultural computing to solve the problems of traditional curation methods, such as the relevance between exhibits and exhibition themes often depending on subjective judgment, the difficulty of accurately matching the interest preferences of different audience groups, and the scattered narrative logic of the exhibition.

[0005] The present application provides a curation method based on cultural computing, comprising:

[0006] inputting multi-modal data of candidate exhibits, theme data and auxiliary knowledge graph, and outputting a theme-related pool;

[0007] transforming the exhibits and user data of the theme-related pool into a core exhibit list, and arranging the core exhibit list in descending order according to the final sorting score and group sorting;

[0008] integrating the core exhibit list after descending sorting and performing exhibition physical constraints to obtain a curation plan for the core exhibit list.

[0009] The output of the theme-related pool includes,

[0010] The surface word association between the exhibits and the theme is calculated based on the TF-IDF word vector, and obviously irrelevant items are filtered to realize the word-level feature matching.

[0011] The theme keyword vector is set as T=[t1, t2, … t n ], and the exhibit text word vector is E=[e1, e2, … e n ], wherein n represents the total number of elements in the sequence, t n represents the keyword weight, and e n represents the keyword frequency in the exhibit.

[0012] The cosine similarity S 相似度 is solved based on the theme keyword vector and the exhibit text word vector.

[0013]

[0014] When S 相似度 is greater than or equal to the similarity threshold value, the exhibit enters the semantic-level multi-modal fusion matching; based on the fusion of text semantics and visual features, the exhibits with different words but related semantics are mined, and the semantic-level multi-modal fusion score S 融合 is calculated, and if S 融合 is greater than or equal to the fusion threshold value, the exhibit enters the knowledge graph association verification.

[0015] The knowledge graph association verification mines the indirect association between the exhibit and the theme through the knowledge graph, obtains the association strength S 主题相关度 , and if S 主题相关度 is greater than or equal to the theme relevance threshold value, the exhibit is included in the theme-related pool.

[0016] The operation of the semantic-level multi-modal fusion matching specifically includes,

[0017] The text semantics and the visual features are fused to mine the exhibits with different words but related semantics; the BERT model is used to convert the theme description and the exhibit text into 768-dimensional semantic vectors V T , V E , the text semantic cosine similarity S text is calculated, the ResNet is used to extract the feature vectors I T , I E of the visual symbols associated with the exhibit and the theme, the visual feature similarity S image is calculated, and the semantic-level multi-modal fusion score S 融合 is solved based on the text semantic cosine similarity S text and the visual feature similarity S image .

[0018] S 融合 =α×S text +β×S image .

[0019] α+β=1;

[0020] Where α represents the weight of text semantics and β represents the weight of visual features;

[0021] Calculate the semantic-level multimodal fusion score S 融合 If S 融合 When the value is greater than or equal to the fusion threshold, the exhibit enters the knowledge graph association verification.

[0022] The specific steps for knowledge graph association verification include:

[0023] Knowledge graph association verification uses knowledge graphs to mine indirect associations between exhibits and themes, obtaining the association strength S. 主题相关度 If S 主题相关度 If the exhibit is greater than or equal to the theme relevance threshold, it will be included in the theme relevance pool.

[0024] The correlation strength S 主题相关度 The calculations include,

[0025] Let the shortest path from the exhibit entity to the core theme entity be P = [r1, r2, ... r]. k ], where k represents the total number of elements in the sequence, r k To represent the relation weight, then

[0026] The process of converting exhibit and user data from the theme-related pool into a core exhibit list, and then sorting the core exhibit list in descending order according to the final ranking score and cluster ranking, includes the following steps:

[0027] The user data includes user behavior data, emotional preferences, and group characteristics;

[0028] Based on the exhibits in the theme-related pool, a precise matching model of user interests and exhibit features is constructed through quantitative analysis of user behavior data, emotional preferences, and group characteristics. Finally, an exhibit ranking list that takes into account both theme relevance and user appeal is output.

[0029] User attraction is scored based on three dimensions: behavioral interaction, emotional inclination, and group matching. These scores are then weighted and combined to form a total user interest score.

[0030] The calculation of the behavioral interaction score includes,

[0031] Behavioral interaction score = (standardized value of play count × 0.1) + (like rate × 0.2) + (share rate × 0.3) + (deep dwell time × 0.4);

[0032] Sentiment score = (sentiment polarity score / 5 × 50) + (interest keyword matching degree × 50);

[0033] Group matching score = Σ(group weight × matching degree between exhibit and group);

[0034] S 用户兴趣 = Behavioral interaction score × Behavioral interaction weight + Sentimental tendency score × Sentimental tendency weight + Group matching score × Group matching weight;

[0035] The final ranking score S is obtained by weighting and fusing the correlation strength and user interest from two dimensions. 最终排序 S 最终排序 =γ×S 主题相关度 +(1-γ)×S 用户兴趣 Where γ represents the association strength weight;

[0036] The optimal value of γ is obtained by dynamically optimizing the association strength weight γ using a regression model or reinforcement learning.

[0037] Group sorting assigns weights based on user needs type, obtaining a group-specific weight value.

[0038] The process of integrating the core exhibit list after sorting it in descending order and applying exhibition physical constraints to obtain the curatorial scheme for the core exhibit list includes the following steps:

[0039] The core exhibits list, sorted in descending order, is filtered to refine the narrative framework parameters, spatial constraint data, and user experience parameters of the exhibits. A curatorial plan is then developed, which includes a narrative framework diagram, a list of key exhibits, a space allocation table, a sequence diagram, and a list of equipment requirements.

[0040] The narrative framework diagram assigns exhibits to corresponding narrative nodes according to preset narrative types, generates core tags for each narrative node, calculates the matching degree between exhibit theme tags and narrative node tags using cosine similarity, assigns exhibits to the node with the highest matching degree, and outputs the subset of exhibits bound to each narrative node.

[0041] Obtaining the list of key exhibits includes,

[0042] The exhibits were categorized according to their priority level to obtain a list of key exhibits;

[0043] The importance level is determined based on a weighted average of the final ranking score and the importance of the node. The comprehensive score of the exhibits within each narrative node is normalized to a normalized score within the node ranging from 0 to 100, and a weight is assigned to each narrative node. The importance score = (normalized score within the node × 0.7) + (node ​​weight × 30).

[0044] Exhibits with a key score of ≥80 are classified as core exhibits, exhibits with a key score of 60-79 are classified as important exhibits, and exhibits with a key score of <60 are classified as auxiliary exhibits.

[0045] It also includes,

[0046] Based on the priority level, physical attributes of the exhibits, and user interaction needs, the display space for each exhibit is allocated. The basic area for core exhibits is 2-5㎡, the basic area for important exhibits is 1-2㎡, and the basic area for auxiliary exhibits is 0.5-1㎡.

[0047] The required area of ​​exhibits is adjusted based on the weight of user interaction, the weight of exhibit physical size, and the weight of narrative nodes.

[0048] If the exhibits are highly interactive, the required area is 1.5 times the base area;

[0049] If the exhibit is an extra-large exhibit, the required area is 1.2 times the base area;

[0050] If the exhibit is a high-importance node exhibit, the required area for the exhibit is the exhibit space allocation coefficient within the high-importance node × the baseline area;

[0051] The total area allocated to all exhibits shall be less than or equal to the available area of ​​the exhibition hall. If it exceeds this area, the area shall be reduced in order of priority from low to high.

[0052] Beneficial effects: This application uses algorithms such as multi-level matching and two-dimensional fusion to transform abstract concepts of theme relevance and user interests into calculable scores, thus solving the subjectivity problem of traditional curation;

[0053] Deeply integrate multimodal data to break through the limitations of a single information source and uncover explicit and implicit relationships;

[0054] By adjusting weights and sorting by group, the system adapts to different exhibition positioning and user groups, achieving a personalized narrative experience for each individual.

[0055] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0056] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein:

[0057] Figure 1 This is based on the flowchart provided in this application. Detailed Implementation

[0058] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0059] like Figure 1 As shown, this application provides a curatorial method based on cultural computing, including:

[0060] S1: Input multimodal data, topic data, and auxiliary knowledge graph of candidate exhibits; output a topic-related pool. It should be noted that:

[0061] The output of the topic-related pool includes,

[0062] TF-IDF word vectors are used to calculate the surface-level lexical associations between exhibits and topics, filtering out obviously irrelevant terms and achieving lexical-level feature matching.

[0063] Set the topic keyword vector to T = [t1, t2, ... t]. n The word vectors of the exhibit text are E = [e1, e2, ..., e]. n ], where n represents the total number of elements in the sequence, t n Indicates keyword weight, e n This indicates the frequency of keywords appearing in the exhibits;

[0064] The cosine similarity S is calculated based on the topic keyword vector and the exhibit text word vector. 相似度 ,

[0065]

[0066] When S 相似度 When the similarity threshold is ≥, the exhibits enter the semantic-level multimodal fusion matching; based on the fusion of text semantics and visual features, exhibits with different words but semantically related are identified, and the semantic-level multimodal fusion score S is calculated. 融合 If S 融合 When the threshold is reached or greater than the fusion threshold, the exhibit enters the knowledge graph association verification.

[0067] Knowledge graph association verification uses knowledge graphs to mine indirect associations between exhibits and themes, obtaining the association strength S. 主题相关度 If S 主题相关度 If the exhibit is greater than or equal to the theme relevance threshold, it will be included in the theme relevance pool.

[0068] The semantic-level multimodal fusion matching operation specifically includes:

[0069] By integrating textual semantics and visual features, exhibits with different vocabulary but semantically related characteristics are identified. The BERT model is used to transform the topic descriptions and exhibit texts into 768-dimensional semantic vectors VT and VE, and the textual semantic cosine similarity S is calculated. text We used ResNet to extract the feature vectors IT and IE of visual symbols associated with exhibits and the theme, and calculated the visual feature similarity S. image Based on textual semantic cosine similarity S text Similarity to visual features S image Solve for the semantic-level multimodal fusion score S 融合 ,

[0070] S 融合 =α×S text +β×S image ;

[0071] α+β=1;

[0072] Where α represents the weight of text semantics and β represents the weight of visual features;

[0073] Calculate the semantic-level multimodal fusion score S 融合 If S 融合 When the value is greater than or equal to the fusion threshold, the exhibit enters the knowledge graph association verification.

[0074] The specific steps for knowledge graph association verification include:

[0075] Knowledge graph association verification uses knowledge graphs to mine indirect associations between exhibits and themes, obtaining the association strength S. 主题相关度 If S 主题相关度 If the exhibit is greater than or equal to the theme relevance threshold, it will be included in the theme relevance pool.

[0076] The correlation strength S 主题相关度 The calculations include,

[0077] Let the shortest path from the exhibit entity to the core theme entity be P = [r1, r2, ... r]. k ], where k represents the total number of elements in the sequence, r k To represent the relation weight, then

[0078] Multimodal data includes metadata, text descriptions, and visual features;

[0079] Each candidate exhibit is given a "multimodal feature label" that covers textual descriptions, visual features, semantic associations, etc., which forms the basis for similarity calculation.

[0080] Table 1: Multimodal Data Table

[0081]

[0082] Thematic data includes thematic keyword vectors and thematic semantic chains;

[0083] Extracting the "core feature system of the theme" transforms the abstract exhibition theme into a "computable feature vector";

[0084] Table 2: Thematic Data

[0085]

[0086]

[0087] Auxiliary knowledge graphs include domain entity association networks;

[0088] Table 3: Auxiliary Related Data

[0089]

[0090] S2: Transform the exhibit and user data from the theme-related pool into a core exhibit list, and sort the core exhibit list in descending order according to the final ranking score and cluster ranking. It should be noted that:

[0091] The process of converting exhibit and user data from the theme-related pool into a core exhibit list, and then sorting the core exhibit list in descending order according to the final ranking score and cluster ranking, includes the following steps:

[0092] The user data includes user behavior interactions (play count, share rate, etc.), sentiment tendencies (comment polarity, keywords), and group characteristics (age, interest tags);

[0093] Based on the exhibits in the theme-related pool, a precise matching model of user interests and exhibit features is constructed through quantitative analysis of user behavior data, emotional preferences, and group characteristics. Finally, an exhibit ranking list that takes into account both theme relevance and user appeal is output.

[0094] The quantification of user interests relies on multi-dimensional, high-quality data input, which needs to cover three layers of data: "behavior-emotion-group". Noise (such as inflated traffic data and invalid comments) should be eliminated through preprocessing to form a standardized "user interest feature library".

[0095] Table 4: User Feature Database

[0096]

[0097]

[0098] Transforming preprocessed user data into "user interest scores" requires multi-dimensional weighting and machine learning optimization to ensure that the scores accurately reflect the exhibits' appeal to users.

[0099] User attraction is scored based on three dimensions: behavioral interaction, emotional inclination, and group matching. These scores are then weighted and combined to form a total user interest score.

[0100] The calculation of the behavioral interaction score includes,

[0101] Behavioral interaction score = (standardized value of play count × 0.1) + (like rate × 0.2) + (share rate × 0.3) + (deep dwell time × 0.4);

[0102] Sentiment score = (sentiment polarity score / 5 × 50) + (interest keyword matching degree × 50);

[0103] Group matching score = Σ(group weight × matching degree between exhibit and group);

[0104] S 用户兴趣 = Behavioral interaction score × Behavioral interaction weight + Sentimental tendency score × Sentimental tendency weight + Group matching score × Group matching weight;

[0105] The final ranking score S is obtained by weighting and fusing the correlation strength and user interest from two dimensions. 最终排序 S 最终排序 =γ×S 主题相关度 +(1-γ)×S 用户兴趣 Where γ represents the association strength weight;

[0106] The optimal value of γ is obtained by dynamically optimizing the association strength weight γ using a regression model or reinforcement learning.

[0107] By dynamically optimizing the association strength weight γ using regression models or reinforcement learning, the subjective bias in setting the association strength weight γ value can be avoided.

[0108] In the initial stages of curating, we suggested using a linear regression model to obtain the γ value, derived from S. 最终排序 =γ×S 主题相关度 +(1-γ)×S 用户兴趣 The following linear regression formula is obtained: y 预测 =b0+b1×(S 主题相关度 -S 用户兴趣 );

[0109] Among them, y 预测 b0 and b1 represent the model's predicted value for the final performance of an exhibit, and are two coefficients obtained through training of the representational regression model.

[0110] By collecting a large amount of data and feeding it into a linear regression model for training, the model automatically learns and determines the optimal values ​​of b0 and b1 by minimizing the error between the predicted and actual values. Once the model is trained, γ can be extracted from the learned coefficients; that is, the optimal value of γ = b1; the average user score = b0.

[0111] After collecting a large amount of data through curation, it is recommended to use reinforcement learning to obtain the γ value;

[0112] The preliminary regression model reward function is given as: R = β1×M + β2×T + β3×U;

[0113] Where β1 + β2 + β3 = 1, β1, β2, and β3 represent weighting coefficients, R is the reward function, M is the actual effect indicator, T is the topic relevance, and U is the user interest. Different weighting variations can adapt to subsequent iterations. This reward function can automatically find a dynamic γ value that maximizes the combined performance of "exhibition reach" and "topic rating" through continuous trial and error and the receipt of rewards.

[0114] Group sorting assigns weights based on user needs type, obtaining a group-specific weight value.

[0115] S3: Integrate the core exhibit list after sorting in descending order and apply exhibition physical constraints to obtain the curatorial scheme for the core exhibit list. It should be noted that:

[0116] The process of integrating the core exhibit list after sorting it in descending order and applying exhibition physical constraints to obtain the curatorial scheme for the core exhibit list includes the following steps:

[0117] The core exhibits list, sorted in descending order, is filtered to refine the narrative framework parameters, spatial constraint data, and user experience parameters of the exhibits. A curatorial plan is then developed, which includes a narrative framework diagram, a list of key exhibits, a space allocation table, a sequence diagram, and a list of equipment requirements.

[0118] The narrative framework diagram assigns exhibits to corresponding narrative nodes according to preset narrative types, generates core tags for each narrative node, calculates the matching degree between exhibit theme tags and narrative node tags using cosine similarity, assigns exhibits to the node with the highest matching degree, and outputs the subset of exhibits bound to each narrative node.

[0119] Obtaining the list of key exhibits includes,

[0120] The exhibits were categorized according to their priority level to obtain a list of key exhibits;

[0121] The importance level is determined based on a weighted average of the final ranking score and the importance of the node. The comprehensive score of the exhibits within each narrative node is normalized to a normalized score within the node ranging from 0 to 100, and a weight is assigned to each narrative node. The importance score = (normalized score within the node × 0.7) + (node ​​weight × 30).

[0122] Exhibits with a key score of ≥80 are classified as core exhibits, exhibits with a key score of 60-79 are classified as important exhibits, and exhibits with a key score of <60 are classified as auxiliary exhibits.

[0123] It also includes,

[0124] Based on the priority level, physical attributes of the exhibits, and user interaction needs, the display space for each exhibit is allocated. The basic area for core exhibits is 2-5㎡, the basic area for important exhibits is 1-2㎡, and the basic area for auxiliary exhibits is 0.5-1㎡.

[0125] The required area of ​​exhibits is adjusted based on the weight of user interaction, the weight of exhibit physical size, and the weight of narrative nodes.

[0126] If the exhibits are highly interactive, the required area is 1.5 times the base area;

[0127] If the exhibit is an extra-large exhibit, the required area is 1.2 times the base area;

[0128] If the exhibit is a high-importance node exhibit, the required area for the exhibit is the exhibit space allocation coefficient within the high-importance node × the baseline area;

[0129] The total area allocated to all exhibits shall be less than or equal to the available area of ​​the exhibition hall. If it exceeds this area, the area shall be reduced in order of priority from low to high.

[0130] It also includes,

[0131] Based on the narrative type, the exhibits within each node are ordered to ensure that users can understand the narrative logic when viewing them in sequence.

[0132] 1) Timeline narrative:

[0133] Extract the time attributes of the exhibits (such as "Tang Dynasty Zhenguan era" and "Song Dynasty Xining era") and convert them into timestamps (such as 640 AD and 1070 AD).

[0134] Arranged in ascending order by timestamp (from ancient times to the present), with "transition notes" inserted at key time points (occupying 5% of the space).

[0135] 2) Thematic branching narrative:

[0136] Construct a "logical relationship diagram" of exhibits within a node (such as the causal chain of "tea picking tools → tea making utensils → tea drinking utensils" in the "tea utensils" node);

[0137] The order is determined by using a topological sorting algorithm to ensure that the preceding exhibits are the "prerequisites" for the following exhibits (e.g., show the "tea picking tools" first, and then show the "tea cake made with the tools").

[0138] 3) Interactive narrative:

[0139] Clustering based on user interest tags (such as "parent-child path" and "historical deep path");

[0140] Generate a sequence of "core exhibits + optional exhibits" for each path (core exhibits are mandatory, and optional exhibits are selected by the user in real time).

[0141] Match the optimal display format based on the type of exhibits, user interaction needs, and priority level to improve information delivery efficiency;

[0142] 1) Basic rule base:

[0143] Core exhibits: The default format is "physical objects + multimedia" (e.g., physical objects accompanied by 3D projections to recreate usage scenarios);

[0144] Key exhibits: "Physical objects + illustrated explanations" (high-resolution illustrations account for 60%);

[0145] Supporting exhibits: "Replicas + scrolling electronic screen display" (space-saving).

[0146] 2) Adjustment of user requirements:

[0147] For exhibits tagged with "high interaction demand" (such as those with more than 30% of user comments saying "want to touch it"), force matching with "touchable replica" or "AR interaction";

[0148] For exhibits that are "visually dependent" (such as paintings and clothing), priority will be given to matching them with "high-definition flat display + focused lighting".

[0149] 3) Machine learning optimization: Train a classification model (such as random forest) using historical data, input exhibit features (type, size, user tags), and output the optimal display format (accuracy target ≥85%).

[0150] When conflicts arise in terms of space, order, or form (such as insufficient space due to the size of the core exhibit), the following rules shall be followed to resolve them:

[0151] Space conflict: Prioritize reserving space for core exhibits and combine auxiliary exhibits for display (e.g., 3 auxiliary exhibits share one booth);

[0152] Order conflict: If the timeline contradicts the logical order (e.g., an exhibit is displayed earlier but logically should be displayed later), a "parallel viewing area" is set up (users can choose the order themselves);

[0153] Format conflict: If exhibits with high interactive requirements conflict with the weight-bearing capacity of the exhibition hall (such as large AR equipment being too heavy), replace them with "light interactive formats" (such as scanning QR codes to watch animations).

[0154] Taking the exhibition themed "The Refined Life of Song Dynasty Literati" as an example:

[0155] Step 1: Topic Relevance Screening

[0156] Input: Candidate exhibits (such as Jian kiln rabbit fur teacups, Su Shi's "Cold Food Festival Scroll", Song Dynasty guqin, etc.), theme keyword vector (T = [\text{Song Dynasty}(0.2),\text{literati}(0.3),\text{refined life}(0.3),\text{tea affairs}(0.1),\text{calligraphy and painting}(0.1)]).

[0157] Calculation process:

[0158] The text word vector of Jian ware rabbit fur tea bowl (E=[\text{Song Dynasty}(0.4),\text{literati}(0.3),\text{tea affairs}(0.5),...]), S 相似度 =0.72 (≥similarity threshold 0.3);

[0159] BERT semantic vector matching (S_{\text{text}} = 0.8), visual feature (rabbit fur pattern) matching (S_{\text{image}} = 0.7), S 融合 =0.8×0.6+0.7×0.4=0.76 (≥ fusion threshold 0.6);

[0160] Knowledge graph path: Jian kiln tea bowl → Tea competition → Song Dynasty literati (path length 2, sum of relation weights = 3), S 主题相关度 =3 / 2=1.5 (≥ topic relevance threshold 1.5).

[0161] Output: The theme-related pool contains 20 exhibits, and all three levels of Jian kiln hare's fur teacups have achieved the required scores.

[0162] Step 2: Optimizing User Interests

[0163] Input: Theme-related exhibits, user data (standardized value of Jian kiln tea bowl playback volume 0.8, sharing rate 0.35, standardized value of dwell time 2.0; sentiment polarity +4.5, keyword matching degree 0.9; target group is "women aged 25-35", matching degree 0.8).

[0164] Calculation process:

[0165] Behavioral interaction score = 0.8 × 0.1 + 0.35 × 0.3 + 2.0 × 0.6 = 1.385 → 85 points;

[0166] Emotional tendency score = (4.5 + 5) / 10 × 50 + 0.9 × 50 = 45 + 45 = 90 points;

[0167] Group matching score = 0.6 × 80 (female group weight 0.6) + 0.4 × 60 = 72 points;

[0168] Total user interest score = 85 × 0.4 + 90 × 0.3 + 72 × 0.3 = 82.6 points;

[0169] Overall ranking score (γ=0.5, topic relevance 80 points)=80×0.5+82.6×0.5=81.3 points.

[0170] Output: List of core exhibits (10 items), with Jian kiln hare's fur teacups ranking first.

[0171] Step 3: Narrative Scheme Generation

[0172] Input: Core exhibit, narrative branch "Tea Gathering" (weight 0.4), exhibition area 100㎡.

[0173] Calculation process:

[0174] Narrative embedding: Jian kiln tea bowls have a 0.9 match with the "Tea Gathering" tag, and are assigned to this branch;

[0175] Key score = (normalized score within the branch 0.9 × 70) + (0.4 × 30) = 63 + 12 = 75 points → Important exhibits (upgraded to core due to user interaction needs after adjustment);

[0176] Space allocation: Basic area 2㎡ × 1.5 (interaction weight) = 3㎡, total branch space 95㎡ (≤100㎡);

[0177] The order is as follows: "tea utensils → tea competition → tea poetry" is the first one, and the format is "physical objects + AR interaction".

[0178] Output: "Tea Gathering" branch plan (with Jian kiln teacups as the core, 3㎡, AR interaction).

[0179] As the case studies demonstrate, "the method transforms abstract curatorial logic into actionable steps through explicit algorithmic formulas, providing a technical path for the precise and personalized planning of cultural exhibitions."

[0180] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A curatorial method based on cultural computing, characterized in that, include: Input multimodal data, topic data, and auxiliary knowledge graph of candidate exhibits, and output a topic-related pool; The exhibits and user data from the theme-related pools are transformed into a core exhibit list, and the core exhibit list is sorted in descending order according to the final ranking score and the group ranking. The core exhibit list, sorted in descending order, is integrated and subjected to physical constraints to obtain a curatorial scheme for the core exhibit list.

2. The curatorial method based on cultural computing according to claim 1, characterized in that: The output of the topic-related pool includes, TF-IDF word vectors are used to calculate the surface-level lexical associations between exhibits and topics, filtering out obviously irrelevant terms and achieving lexical-level feature matching. Set the topic keyword vector to T = [t1, t2, ... t]. n The word vectors of the exhibit text are E = [e1, e2, ..., e]. n ], where n represents the total number of elements in the sequence, t n Indicates keyword weight, e n This indicates the frequency of keywords appearing in the exhibits; The cosine similarity S is calculated based on the topic keyword vector and the exhibit text word vector. 相似度 , When S 相似度 When the similarity threshold is ≥, the exhibits enter the semantic-level multimodal fusion matching; based on the fusion of text semantics and visual features, exhibits with different words but semantically related are identified, and the semantic-level multimodal fusion score S is calculated. 融合 If S 融合 When the threshold is reached or greater than the fusion threshold, the exhibit enters the knowledge graph association verification. Knowledge graph association verification uses knowledge graphs to mine indirect associations between exhibits and themes, obtaining the association strength S. 主题相关度 If S 主题相关度 If the exhibit is greater than or equal to the theme relevance threshold, it will be included in the theme relevance pool.

3. The curatorial method based on cultural computing according to claim 2, characterized in that: The semantic-level multimodal fusion matching operation specifically includes: By integrating textual semantics and visual features, exhibits with different words but related semantics are identified; the BERT model is used to transform topic descriptions and exhibit texts into 768-dimensional semantic vectors V. T V E Calculate the semantic cosine similarity S of the text. text We used ResNet to extract feature vectors I of visual symbols that associate exhibits with the theme. T I E Calculate visual feature similarity S image Based on textual semantic cosine similarity S text Similarity to visual features S image Solve for the semantic-level multimodal fusion score S 融合 , S 融合 =α×S text +β×S image ; α+β＝1； Where α represents the weight of text semantics and β represents the weight of visual features; Calculate the semantic-level multimodal fusion score S 融合 If S 融合 When the value is greater than or equal to the fusion threshold, the exhibit enters the knowledge graph association verification.

4. The curatorial method based on cultural computing according to claim 3, characterized in that: The specific steps for knowledge graph association verification include: Knowledge graph association verification uses knowledge graphs to mine indirect associations between exhibits and themes, obtaining the association strength S. 主题相关度 If S 主题相关度 If the exhibit is greater than or equal to the theme relevance threshold, it will be included in the theme relevance pool. The correlation strength S 主题相关度 The calculations include, Let the shortest path from the exhibit entity to the core theme entity be P = [r1, r2, ... r]. k ], where k represents the total number of elements in the sequence, r k To represent the relation weight, then 5. The curatorial method based on cultural computing according to claim 4, characterized in that: The process of converting exhibit and user data from the theme-related pool into a core exhibit list, and then sorting the core exhibit list in descending order according to the final ranking score and cluster ranking, includes the following steps: The user data includes user behavior data, emotional preferences, and group characteristics; Based on the exhibits in the theme-related pool, a precise matching model of user interests and exhibit features is constructed through quantitative analysis of user behavior data, emotional preferences, and group characteristics. Finally, an exhibit ranking list that takes into account both theme relevance and user appeal is output. User attraction is scored based on three dimensions: behavioral interaction, emotional inclination, and group matching. These scores are then weighted and combined to form a total user interest score. The calculation of the behavioral interaction score includes, Behavioral interaction score = (standardized value of play count × 0.1) + (like rate × 0.2) + (share rate × 0.3) + (deep dwell time × 0.4); Sentiment score = (sentiment polarity score / 5 × 50) + (interest keyword matching degree × 50); Group matching score = Σ(group weight × matching degree between exhibit and group); S 用户兴趣 = Behavioral interaction score × Behavioral interaction weight + Sentimental tendency score × Sentimental tendency weight + Group matching score × Group matching weight; The final ranking score S is obtained by weighting and fusing the correlation strength and user interest from two dimensions. 最终排序 S 最终排序 =γ×S 主题相关度 +(1-γ)×S 用户兴趣 Where γ represents the association strength weight; The optimal value of γ is obtained by dynamically optimizing the association strength weight γ using a regression model or reinforcement learning. Group sorting assigns weights based on user needs type, obtaining a group-specific weight value.

6. A curatorial method based on cultural computing according to claim 1 or 5, characterized in that: The process of integrating the core exhibit list after sorting it in descending order and applying exhibition physical constraints to obtain the curatorial scheme for the core exhibit list includes the following steps: The core exhibits list, sorted in descending order, is filtered to refine the narrative framework parameters, spatial constraint data, and user experience parameters of the exhibits. A curatorial plan is then developed, which includes a narrative framework diagram, a list of key exhibits, a space allocation table, a sequence diagram, and a list of equipment requirements.

7. A curatorial method based on cultural computing according to claim 6, characterized in that: The narrative framework diagram assigns exhibits to corresponding narrative nodes according to preset narrative types, generates core tags for each narrative node, calculates the matching degree between exhibit theme tags and narrative node tags using cosine similarity, assigns exhibits to the node with the highest matching degree, and outputs the subset of exhibits bound to each narrative node.

8. The curatorial method based on cultural computing according to claim 7, characterized in that: Obtaining the list of key exhibits includes, The exhibits were categorized according to their priority level to obtain a list of key exhibits; The importance level is determined based on a weighted average of the final ranking score and the importance of the node. The comprehensive score of the exhibits within each narrative node is normalized to a normalized score within the node ranging from 0 to 100, and a weight is assigned to each narrative node. The importance score = (normalized score within the node × 0.7) + (node ​​weight × 30). Exhibits with a key score of ≥80 are classified as core exhibits, exhibits with a key score of 60-79 are classified as important exhibits, and exhibits with a key score of <60 are classified as auxiliary exhibits.

9. A curatorial method based on cultural computing according to claim 8, characterized in that: It also includes, Based on the priority level, physical attributes of the exhibits, and user interaction needs, the display space for each exhibit is allocated. The basic area for core exhibits is 2-5㎡, the basic area for important exhibits is 1-2㎡, and the basic area for auxiliary exhibits is 0.5-1㎡. The required area of ​​exhibits is adjusted based on the weight of user interaction, the weight of exhibit physical size, and the weight of narrative nodes. If the exhibits are highly interactive, the required area is 1.5 times the base area; If the exhibit is an extra-large exhibit, the required area is 1.2 times the base area; If the exhibit is a high-importance node exhibit, the required area for the exhibit is the exhibit space allocation coefficient within the high-importance node × the baseline area; The total area allocated to all exhibits shall be less than or equal to the available area of ​​the exhibition hall. If it exceeds this area, the area shall be reduced in order of priority from low to high.

Images