Cultural relic knowledge extraction and visualization method

Abstract

The invention relates to the technical field of cultural relic analysis, and discloses a cultural relic knowledge extraction and visualization method, which comprises the following steps: processing an original sentence to obtain a feature vector and relation information of the sentence, inputting the relation information and the feature vector into an attention unit, and distributing different weights to context information of a current relation; fusing the feature vectors of the sentences output by the attention unit and the relation information to realize cultural relic knowledge extraction; constructing a visual view based on the identified data; the problems of particularity of knowledge content and text word formation in the cultural relic field, structure looseness and knowledge overlapping are solved; the extracted cultural relics are only explored and visualized, information of the cultural relics is displayed in a visual interaction mode, a user is helped to understand characteristics and evolution trends in the cultural relic development process under the historical culture background, repeated work of the user on cultural relic knowledge exploration is reduced, and the user experience is improved. And the analysis efficiency and exploration depth of the cultural relic knowledge by the user are improved.

Description

Technical Field

[0001] This invention relates to the field of cultural relic analysis technology, specifically to a method for extracting and visualizing cultural relic knowledge. Background Technology

[0002] Cultural relics are important carriers of human history and culture, and their protection and inheritance are of profound significance. Studying the evolution of cultural relics helps to understand the trajectory of ancient cultural exchanges and the transmission of technology between different civilizations. With the rapid development of information technology, museums, online encyclopedia platforms, books, documents, and archaeological data have accumulated a vast amount of information resources on cultural relics. These resources are rich in content, but the lengthy text often makes it difficult for researchers to quickly locate the basic information of the relics, affecting the efficient use of cultural relic data. Therefore, extracting meaningful knowledge from the massive amount of data has become crucial for cultural relic research. Existing technologies in the field of knowledge extraction and visualization analysis of cultural relics are insufficient in terms of domain adaptability, lack of dynamic analysis capabilities, and limited interactive functions, making it difficult to meet the needs of cultural relic research for efficient knowledge extraction, multi-dimensional correlation analysis, and intuitive display. Summary of the Invention

[0003] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a method for extracting and visualizing cultural relic knowledge.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for extracting and visualizing knowledge about cultural relics, comprising the following steps:

[0006] The original sentence is processed to obtain its feature vector.

[0007] The latent relations in the original sentence are classified to obtain relation information.

[0008] The relation information and feature vectors are input into the attention unit, and different weights are assigned to the context information of the current relation.

[0009] The feature vectors and relational information of the sentences output by the attention unit are fused, and cultural relic knowledge extraction is achieved using BiLSTM and CRF units;

[0010] The extracted cultural relic knowledge is identified based on attribute dimensions, cultural relic dimensions, and global dimensions.

[0011] A visualization view is built based on the identified data.

[0012] In this invention, preferably, the process of processing the original sentence to obtain the sentence's feature vector specifically includes the following steps:

[0013] Convert the sentence into a character sequence U = {u0, u1, u2, u3, ..., un}, where u v Let v represent the v-th character in sentence U, n be the length of the sequence, and u0 = [CLS] represent the start marker of the sentence. n =[SEP] indicates a sentence separator;

[0014] After introducing position embeddings and sentence embeddings into the character sequence, they are input into the BERT model (representing the position of a word in the sentence and the global semantics of the entire sentence, respectively). After BERT encoding, the semantic feature vector B = {b0, b1, b2, ..., b} of the sentence can be obtained. n}。 );

[0015] The BERT model uses complete cultural relic domain entities for masking, masks cultural relic phrases in sentences and predicts the entire phrase, captures the deep semantic connections between cultural relic words, and obtains the feature vector representation of the sentence.

[0016] In this invention, preferably, the feature vector of the sentence is also enhanced:

[0017] The Stanford CoreNLP dependency parser is used to extract dependency relations from the original input sentence and construct the corresponding dependency relation matrix.

[0018] A graph attention neural network (GAT) model is used to construct a graph with n nodes based on the dependency matrix and the original sentence, where the extracted words are nodes and the dependencies are edges connecting the nodes;

[0019] The feature vector representation of the obtained sentence is input into the GAT model. By concatenating the semantic feature vector and the syntactic feature vector of the text sequence, the final output feature vector is obtained.

[0020] In this invention, preferably, obtaining the relationship information specifically includes:

[0021] Average pooling and fully connected layers are used to predict embedding relations for a given sentence embedding C:

[0022] c pool =Avgpool(C),

[0023] P r =σ(W r c pool +b r ),

[0024] In the formula, Avgpool is the average pooling operation, Wr is the trainable weight matrix, σ is the sigmoid activation function, and br is the relevant parameter. This is modeled as a multi-label binary classification task. If the classification probability exceeds a threshold, the model assigns label 1 to the corresponding relation; otherwise, it assigns label 0. In subsequent work, this paper only needs to focus on these potential relations.

[0025] In this invention, preferably, the relational information and feature vectors are input into the attention unit, and the specific processing includes:

[0026] Calculate the attention score for each feature vector;

[0027] By using attention parameters to perform weighted summation of the words in the sentence, a sentence representation under the current relation is obtained;

[0028] Relationship information is preserved based on gating strategies;

[0029] The embedded relational information is fused with the text encoding to obtain the embedded vector representation for entity recognition.

[0030] In this invention, preferably, the BiLSTM unit calculates the forward contextual latent vector of each word from the beginning of the sentence and the backward contextual latent vector of each word from the end of the sentence. Concatenating the forward and backward latent vectors yields bidirectional context-dependent word vectors.

[0031] At the same time, the CRF unit identifies the dependencies between contexts;

[0032] Cultural relic knowledge is extracted based on bidirectional context-dependent word vectors and dependencies. Preferably, in this invention, the attribute dimensions include time attributes, spatial location attributes, and other enumerable attributes.

[0033] In this invention, preferably, the cultural relic dimension includes the importance and similarity of the cultural relics.

[0034] In this invention, preferably, the global dimensions include the number of cultural relics, the diversity of cultural relics, the importance of cultural relics, the diversity of attributes, and the innovativeness of attributes.

[0035] In this invention, preferably, the visualization view includes a basic information and control view, a timeline view, a cultural relic information view, an attribute evolution view, a map view, and a cultural relic relationship view, wherein the attribute evolution view includes a single-attribute module view and a multi-attribute module view.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] The method of this invention efficiently extracts knowledge in the field of cultural relics, solving the problems of the special nature of knowledge content and text word formation, loose structure, and knowledge overlap in the field of cultural relics; and it further explores and visualizes the extracted cultural relics knowledge, presenting the information of cultural relics in an intuitive and interactive way, helping users understand the characteristics and evolution trends of cultural relics in the historical and cultural context, thereby reducing repetitive work for users in exploring cultural relics knowledge, and improving the efficiency and depth of users' analysis of cultural relics knowledge. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a method for extracting and visualizing cultural relic knowledge as described in this invention.

[0039] Figure 2 This is a schematic diagram of the process of extracting cultural relic knowledge in the method for extracting and visualizing cultural relic knowledge described in this invention.

[0040] Figure 3 This is a schematic diagram of the structure of the visualization view described in this invention.

[0041] Figure 4 This is a schematic diagram illustrating the property evolution described in this invention.

[0042] Figure 5 This is a schematic diagram of the cultural relic relationship view and its interaction method as described in this invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0045] Please see Figure 1This invention, in a preferred embodiment, provides a method for extracting and visualizing cultural relic knowledge, addressing issues such as the unique characteristics of knowledge content and textual word formation in the field of cultural relics, the loose structure, and the overlap of knowledge. Based on the extracted cultural relic knowledge, it conducts cultural relic knowledge exploration and visualization analysis, displaying information about cultural relics in an intuitive and interactive manner to help users understand the characteristics and evolutionary trends of cultural relics in the historical and cultural context. The specific steps include:

[0046] The original sentence is processed to obtain its feature vector.

[0047] The latent relations in the original sentence are classified to obtain relation information.

[0048] The relation information and feature vectors are input into the attention unit, and different weights are assigned to the context information of the current relation.

[0049] The feature vectors and relational information of the sentences output by the attention unit are fused, and cultural relic knowledge extraction is achieved using BiLSTM and CRF units;

[0050] The extracted cultural relic knowledge is identified based on attribute dimensions, cultural relic dimensions, and global dimensions.

[0051] A visualization view is built based on the identified data.

[0052] Please see Figure 2 In this embodiment, the process of processing the original sentence to obtain the sentence's feature vector specifically includes the following steps:

[0053] Convert the sentence into a character sequence U = {u0, u1, u2, u3, ..., u n}, where u v Let v represent the v-th character in sentence U, n be the length of the sequence, and u0 = [CLS] represent the start marker of the sentence. n =[SEP] indicates a sentence separator;

[0054] After introducing positional embeddings and sentence embeddings into the character sequence and inputting them into the BERT model, the positional embeddings and sentence embeddings represent the position of a word in the sentence and the global semantics of the entire sentence, respectively. After BERT encoding, the semantic feature vector B = {b0, b1, b2, ..., b...} is obtained. n}

[0055] The BERT model uses complete cultural relic entity in the cultural relic domain for masking, obscures through cultural relic phrases in the sentence and predicts the whole phrase, captures the deep semantic connections between cultural relic vocabulary, and obtains the feature vector representation of the sentence. Among them, the BERT model is pre-trained with text data in the cultural relic domain to fully adapt to the word formation rules and semantic features in the cultural relic domain. First, text data of cultural relics are collected from the museum online platform, and these data are cleaned and preprocessed to remove noise data and special characters unrelated to cultural relic research. During the training process, BERT adopts the random masking method. The original method is to randomly select 15% of the vocabulary from the input samples for masking. However, for Chinese, each word is a single character, and the above masking method does not consider the information of phrases. For example, for the sentence "The blue-and-white dragon-patterned porcelain bottle of the Ming Dynasty is painted with tumbling dragon patterns in cobalt blue throughout", random masking may result in "The blue-and-white [Mask] dragon-patterned porcelain bottle of the Ming Dynasty is painted with [Mask] blue throughout and tumbling [Mask] dragon patterns". Therefore, complete cultural relic entities in the cultural relic domain are used for masking, and by obscuring some cultural relic phrases in the sentence and predicting the whole phrase, the model is allowed to capture the deep semantic connections between cultural relic vocabulary. The entire training process is carried out under a distributed training architecture and is optimized through forward propagation, loss calculation, and parameter update. The AdamW optimizer is used for training. This optimizer combines a weight decay mechanism and can effectively prevent overfitting. During the training process, the perplexity and accuracy of the model are monitored through the validation set to evaluate the model's text generation ability and ensure that the model can converge stably at each training stage. After pre-training, the BERT model can generate more accurate word vector representations in subsequent tasks, thus providing a solid foundation for text semantic understanding.

[0056] When using the pre-trained BERT to encode a sentence, first the sentence is converted into a character sequence U = {u0, u1, u2, u3,..., un}, where uv represents the v-th character in the sentence U, n is the length of the sequence, u0 = [CLS] represents the start token of the sentence, and un = [SEP] represents the separator of the sentence. This step converts the text sequence into a vector of fixed length to capture its semantic information. When inputting into the model, position embeddings and sentence embeddings are also introduced, which represent the position of the word in the sentence and the global semantics of the whole sentence respectively. After encoding by BERT, the semantic feature vector B = {b0, b1, b2,..., bn} of the sentence can be obtained.

[0057] In this embodiment, after the BERT model generates the semantic feature vector encoding, it further includes enhancing the feature vector of the sentence:

[0058] The StanfordCoreNLP dependency parser is adopted to extract the dependency relationships according to the input original sentence and construct the corresponding dependency relationship matrix;

[0059] Since different dependency relations and dependency words have different effects on sentence grammatical features, a graph attention neural network (GAT) model is adopted. Based on the dependency relation matrix and the original sentence, a graph with n nodes is constructed, where the extracted words are nodes and the dependency relations are edges connecting the nodes.

[0060] The feature vector representation of the obtained sentence is input into the GAT model. By concatenating the semantic feature vector and the syntactic feature vector of the text sequence, the final output feature vector is obtained.

[0061] Specifically, for example, in the text "The Jiahu bone flute, as an ancient wind instrument, is mainly made of natural materials such as animal bones and horn," "Jiahu bone flute" is the subject of the sentence, forming a subject-predicate relationship with "as," "material" and "selected" forming a verb-object relationship, and "animal bones" and "horn" are parallel components. Based on the recognized dependency relationships, a matrix Mn×n is constructed. According to the input text and the relation matrix, a graph with n nodes is constructed, and GAT is defined as layer L. The calculation formula for node vector updates is as follows:

[0062]

[0063] In the formula, and Let i, j, and k represent the vector representations of nodes i, j, and k in the l-th layer of the graph neural network, respectively. M represents the correlation coefficient between nodes i and j. ij W represents the dependency relationship between nodes i and j. (l) This represents the trainable weight matrix of the l-th layer. Let represent the parameters of the l-th layer. LeakyReLU is an improved linear activation function.

[0064] The semantic feature vector B and the dependency relation matrix Mn×n are input into the GAT. After L layers of processing, the syntactic feature vector S = {s1, s2, s3, ..., sn} is obtained. By concatenating the semantic feature vector and the syntactic feature vector of the text sequence, the final output feature vector is obtained, as shown in the following formula:

[0065] c i =[b i s i ],

[0066] In the formula, bi is the semantic feature vector of the i-th word, and si is the syntactic feature vector of the i-th word. In the embedding layer, the text sequence feature vector is C = {c1, c2, c3, ..., cn}.

[0067] In this embodiment, during knowledge extraction, since the number of relations in a sentence is far less than the total number of predefined relations, the distribution of relations in the dataset often suffers from class imbalance. Therefore, this application first identifies the possible relations in the input text, effectively reducing the computation of redundant relations in subsequent operations. The model only needs to perform entity recognition on these potential relations. To predict potential relations, average pooling and fully connected layers are used. For a given sentence embedding C:

[0068] c pool =Avgpool(C),

[0069] P r =σ(W r c pool +b r ),

[0070] In the formula, Avgpool is the average pooling operation, Wr is the trainable weight matrix, σ is the sigmoid activation function, and br is the relevant parameter. It is modeled as a multi-label binary classification task. If the classification probability exceeds a threshold, the model assigns label 1 to the corresponding relation; otherwise, it assigns label 0.

[0071] In this implementation, words in a sentence play different roles under different relations. To measure the importance of different parts, relation embeddings and sentence embeddings are input into the attention unit, assigning different weights to the contextual information of the current relation. Relation information and feature vectors are input into the attention unit, let r... j The embedding vector for relation j is used for specific processing, including:

[0072] Calculate the attention score of the feature vector:

[0073] S(c i r j ) = v T tanh(W c c i +W p r j ),

[0074]

[0075] Among them, v and W c and W p These are trainable parameters;

[0076] By using attention parameters to perform a weighted summation of the words in the sentence, we obtain the sentence representation under the current relation:

[0077]

[0078] After obtaining the relation-enhanced sentence representation, in order to obtain a more accurate input representation containing the current relation information and to ensure that the relation-enhanced sentence representation is meaningful for subsequent entity recognition, a gating strategy is used to preserve the relation information:

[0079]

[0080] u j =f⊙tanh(W g3 s j +b3),

[0081] Where σ represents the sigmoid activation function, W g1 b1, W g2 b2, W g3 b3 and b3 are training parameters. First, the usefulness of the original embedding representation and the relation-enhanced embedding representation for entity extraction is determined. Then, relation information is preserved based on the calculation results. Finally, the embedded relation information is fused with the text encoding to obtain the entity recognition embedding vector representation hi,j, as shown in the following formula, where i represents the encoding of the i-th word, and j represents the j-th relation:

[0082] h i，j =[c i u j ].

[0083] In this implementation, the dependencies between entities and relationships are fully considered. The BiLSTM and CRF modules work together to capture long-distance dependencies while focusing on important information, thus completing the entity recognition process. The BiLSTM unit calculates the forward contextual latent vector for each word from the beginning of the sentence and the backward contextual latent vector for each word from the end of the sentence. Concatenating the forward and backward latent vectors yields bidirectional context-dependent word vectors, as shown in the following formula:

[0084]

[0085] In entity recognition tasks within the field of cultural relics, the contextual information of a sentence often contains many annotation constraints, meaning that the label of the current word usually has a certain relationship with the labels of its neighboring words. Therefore, when annotating each character individually, the results of sequence labeling often suffer from inaccuracies. For example, logically, the internal label of a cultural relic's name entity cannot appear after the initial label of an excavation site entity. Therefore, the sequential dependencies between labels must be considered for accurate entity recognition. Thus, the CRF unit simultaneously identifies the dependencies between contextual relationships. Within the CRF unit, the dependencies between labels can be captured through model training, thereby effectively reflecting the connections between labels during prediction.

[0086] To address the common triple overlap problem in relation extraction, the entity recognition module performs sequence labeling on the head and tail entities separately. This strategy can handle issues such as entity pair overlap (EPO) and single entity overlap (SEO). Taking the head entity as an example, the calculation method is as follows:

[0087] P i，j =CRF(h′) i，j ).

[0088] This model, which combines BiLSTM and CRF, can improve prediction results for entities with complex contextual dependencies in cultural relic entity recognition by using reasonable label dependencies, and better adapt to the labeling needs of different entity types.

[0089] In this embodiment, the attribute dimensions include time attributes, spatial location attributes, and other enumerable attributes. The time attribute is a linear sequence, represented as... Where N is the length of the time series, t i Represents the i-th time and satisfying t i <t i+1 , t ni Indicates t i The corresponding dynasty, for example (618, Tang Dynasty). The location attribute is a spatial sequence, which is represented in this paper as... Where M is the length of the spatial sequence, [lxi, lyi] represents the latitude and longitude of the location, and lni represents the location name corresponding to [lxi, lyi], for example ([94.66, 40.14], Dunhuang). Other attributes do not have linear or spatial relationships, therefore they only contain a set of enumerable attribute sequences. For example, the cultural relic type attribute can be represented as... Each attribute value is considered a feature of the cultural relic. This paper calculates the importance of features from the perspectives of single features and combined features.

[0090] Specifically, the dimensions of cultural relics include their importance and similarity. For a single cultural relic, its importance is comprehensively considered from the perspectives of attribute importance, cultural influence, and scale. The attribute importance of a cultural relic can be calculated using the same method as the aforementioned feature importance calculation, assessing its relative importance within the relic based on the representativeness, innovativeness, and cultural value of each attribute within a specific spatiotemporal context. For each feature f appearing in a cultural relic... j The importance scores of each feature can be weighted and summed. i How to calculate the importance of attributes:

[0091]

[0092] Regarding cultural influence, it's important to consider whether the historical context of a cultural relic is related to significant historical events. Historical events often have a profound impact on the culture of specific regions and periods; therefore, cultural relics closely associated with these events typically have higher cultural influence. For example, Zhang Qian's mission to the Western Regions is a crucial event in Chinese history. Cultural relics appearing at times and locations related to Zhang Qian's mission typically receive higher importance scores. Considering different events (ej), the calculation method for the cultural influence importance of cultural relic (ri) is shown in the following formula.

[0093]

[0094] The scale of an artifact is measured by considering the number of components, its size and weight, and the quantity of decorations, craftsmanship, and materials used. Larger artifacts are more likely to have been used in public ceremonies and celebrations of significant historical events, representing a richer level of historical and cultural significance. Take the chime bells from the Tomb of Marquis Yi of Zeng as an example. This set of bells is extremely large; it is not merely a single musical instrument but a component of a complex ritual and musical system. The calculation method for the importance of the scale of artifact *ri* is shown in the following formula, where *j* attributes need to be calculated in terms of quantity, such as the number of components and decorations, and *n*... j Indicates the number of attributes, s represents the size of the artifact, w represents the weight of the artifact, and N represents the weight of the artifact. j s max and w max It represents the maximum value of each factor, used for normalization.

[0095]

[0096] Therefore, the importance of artifact ri is a weighted sum of the three importance factors mentioned above, as shown in the following formula:

[0097]

[0098] Where wk represents the weight of the three factors mentioned above.

[0099] For a given artifact ri and rj, their similarity is calculated using the following formula:

[0100]

[0101] Where v ik and v jk It is a cultural relic. i With r j The k-th eigenvalue in the matrix.

[0102] For a given group of cultural relics, the affiliation between the relics and the group is calculated by weighted summation of their features, as shown in the following formula:

[0103]

[0104] Where w = [w1, w2, ..., w k [v1, v2, ..., v] represents the fusion weight vector, r = [v1, v2, ..., v] k ] represents the feature vector of the cultural relic, where each feature v i The value represents the artifact's performance on this feature, with a weight w. i This reflects the importance of this feature in group ratings.

[0105] To determine the affiliation between cultural relics and their groups, a linear regression model was used as the classifier. Linear regression offers good interpretability, providing continuous output values ​​for each cultural relic, making it suitable for ranking and evaluating whether a relic belongs to a group. During training, a stochastic gradient descent optimizer was used to optimize the model and obtain the optimal fusion weights. Initially, cultural relics within a group were used as positive training samples. The model parameters were updated using backpropagation, progressively optimizing the weights wi of each feature to accurately reflect the similarity between each relic and the group. The group score CS(v) of a cultural relic was used to assess its affiliation within the group; cultural relics with a group score exceeding 1.0 were directly assigned to the group.

[0106] In this embodiment, the global dimensions include the number of cultural relics, the diversity of cultural relics, the importance of cultural relics, the diversity of attributes, and the innovativeness of attributes.

[0107] Specifically, a higher quantity of a certain type of cultural relic in a given period usually indicates a wider range of production, use, and dissemination of that type of relic at that time, reflecting its popularity. An increase in the quantity of relic usually signifies a more prosperous development of that type of relic during that historical period. Therefore, the score for the quantity factor of type R cultural relics... Where n represents the number of Class R artifacts in this period, and N represents the maximum number of Class R artifacts across all periods, used for normalization.

[0108] Specifically, the diversity of cultural relics reflects the degree of difference in the forms of expression of a certain type of cultural relic. Within a given period, the diversity of cultural relic types reflects the diverse needs of society for that type of relic; for example, musical relics can have different forms of expression such as musical instruments and figurines. Therefore, the score of the cultural relic diversity factor for category R cultural relics... Where n kind N represents the number of types of artifacts of category R during this period. kind The maximum number of R-type artifacts representing all periods is used for normalization.

[0109] Specifically, some cultural relics may bear witness to significant historical events, and these relics have a higher importance score. Assessing the importance of all cultural relics within a certain category during a period reflects the developmental level of that category. Using the cultural relic importance assessment method, therefore, the score of the cultural relic importance factor for category R cultural relics... r i For cultural relics belonging to category R, ​​I(r) i ) for cultural relics i Importance score.

[0110] Specifically, the diverse attributes of cultural relics within a given period indicate that these relics have undergone complex cultural integration and technological evolution, developing towards diversification, reflecting the greater prosperity of this type of cultural relic during that historical period. Assuming that type R cultural relics have m attributes, then the score for the diversity factor of attributes of type R cultural relics... Where n k N represents the number of categories of the k-th attribute of artifacts of category R during this period. k This represents the maximum number of the k-th attribute type of R-class cultural relics across all periods, used for normalization.

[0111] Specifically, the emergence of developmental and evolving attributes in a given period compared to an earlier period signifies attribute innovation during that time, reflecting cultural changes. For example, musical artifacts may have exhibited a new musical form, "playing music," during a particular period. Using the attribute innovation assessment method, the score for the attribute innovation factor of category R artifacts is thus determined. Where n represents the number of cultural relic attributes, m represents the number of attribute values ​​among the cultural relic attributes, and I2(f ij ) represents the importance of the j-th attribute value of the i-th cultural relic attribute.

[0112] Specifically, scores for each of the five factors are calculated separately. The overall score for a certain type of cultural relic is obtained by weighted summation of the five scores, as shown in the following formula:

[0113]

[0114] Where w and k represent the weights of the five factors to be considered.

[0115] In this embodiment, such as Figure 3 As shown, the visualization view includes a basic information and control view, a timeline view, a cultural relic information view, an attribute evolution view, a map view, and a cultural relic relationship view. The attribute evolution view includes single-attribute module views and multi-attribute module views. The various views are linked through diverse interactive methods, providing researchers with a complete exploration process.

[0116] Specifically, such as Figure 3(a) Displays basic information and control views. This view serves as the entry point and basic interactive interface for the cultural relic visualization and exploration system. Users enter a keyword in the input box, and the system automatically returns the types of cultural relics related to that keyword and their associated attributes. Clicking on a specific cultural relic type will recommend important information about that category, helping users quickly understand its value and characteristics. Clicking on an attribute type allows users to view an evolution graph of the attribute, showing its trend over time, region, or other factors, helping users understand the attribute's changes from various perspectives. Furthermore, users can control the weight of attributes, allowing for flexible exploration and more accurate information based on their personal interests and research needs, enhancing the interactivity and depth of cultural relic research and learning.

[0117] like Figure 3 As shown in (b), the artifact timeline view uses the visual metaphor of a mountain range of civilization as its core, integrating the mining of important knowledge along the timeline and multi-granularity interaction to construct an analytical space for exploring the development level and key knowledge of artifacts. This view is used to display the overall development level and important knowledge of artifacts along the timeline. The view consists of a mountain range of civilization representing the development level of artifacts and a scatter plot representing important knowledge. The mountain range of civilization is composed of multiple layers of mountain views, stacked in a streamlined manner. At each time point, the vertical height encodes the comprehensive development level of the artifact. This paper provides different perspectives to observe the development level of artifacts, including the number of artifacts, the diversity of artifacts, the importance of artifacts, the diversity of various attributes, and the perspective of innovation. Each perspective provides a layer of mountain views, and the multiple layers of mountain views are stacked to form the mountain range of civilization. This flow-map-based display method can intuitively show the evolution process of artifacts and help users understand the development level of artifacts in different historical stages.

[0118] like Figure 3 As shown in (c), the Cultural Relics Information View displays a list of important cultural relics and detailed information about specific relics. The Cultural Relics Information View shows the system's overall important cultural relics. When a user selects a specific time, location, or attribute, the Cultural Relics Information View automatically updates, displaying the specific important cultural relics related to those conditions. When a user clicks on a specific cultural relic, the Cultural Relics Information View displays detailed information about that relic, helping the user to fully understand all aspects of that relic.

[0119] like Figure 3 As shown in (d), the attribute evolution view is a view used in the cultural relic visualization and exploration system to display attribute changes, helping users to gain a deeper understanding of the evolution process of cultural relic attributes and the relationships between different attributes. This view supports both single-attribute and multi-attribute display modes, and supports different chart formats, providing users with a comprehensive analytical perspective. In single-attribute display mode, it also provides statistical information for that attribute (such as...). Figure 4(a) The system helps users analyze the distribution and trends of this attribute in cultural relics through different chart formats. It provides a variety of visualization charts for users to choose from, such as line charts, bar charts, pie charts, and rose charts. In addition, it can also display information on combinations of attribute values ​​(such as...). Figure 4 (b) Frequent combinations of attribute values ​​are presented in the form of heatmaps. The intensity of the heatmap colors indicates the frequency of occurrence of the combinations, helping users discover the correlations and patterns between different attribute values. In multiple attribute display modes, statistical information on multiple attributes ( Figure 4 (c) The system provides users with options for nested pie charts, overlaid bar charts, and single-attribute statistical charts. Furthermore, the system also supports displaying combined information of multiple attributes. Figure 4 (d) The system provides a confusion chart to display the frequency of two attribute values ​​appearing together in cultural relics. The confusion chart displays the combinations of two attribute values ​​in a matrix format, helping users identify common attribute combinations and their relationships. Users can freely choose the desired chart format and flexibly switch the display of attribute content. Users can also filter by specific time period, region, or attribute value to focus on a specific dimension for in-depth analysis. Through the interactive functions in the charts, users can view detailed information for each data point and intuitively observe more about attribute changes and their combination patterns.

[0120] like Figure 3 As shown in (e), the artifact relationship view is used to display the relationships between multiple artifacts across different attributes. Figure 5 This document presents a view of cultural relic relationships and its interactive features. The view visually illustrates the connections between cultural relics and their attributes through nodes and lines, helping users gain a deeper understanding of the representative characteristics and inherent relationships within a group of cultural relics. In the view, different nodes represent different types of attributes or cultural relics, and the size of each node reflects the importance of that attribute or relic, which is categorized into three different levels. Larger nodes indicate higher importance of the attribute or relic, or stronger representativeness within the group of relics. Users can easily identify key characteristics within a group of cultural relics by observing the size of the nodes.

[0122] In some other preferred embodiments of the present invention, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in the above embodiments.

[0123] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for extracting and visualizing cultural relic knowledge, characterized in that, Including the following steps: The original sentence is processed to obtain its feature vector. The latent relations in the original sentence are classified to obtain relation information. The relation information and feature vectors are input into the attention unit, and different weights are assigned to the context information of the current relation. The feature vectors and relational information of the sentences output by the attention unit are fused, and cultural relic knowledge extraction is achieved using BiLSTM and CRF units; The extracted cultural relic knowledge is identified based on attribute dimensions, cultural relic dimensions, and global dimensions. A visualization view is built based on the identified data.

2. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The specific steps involved in processing the original sentence to obtain its feature vector are as follows: Convert the sentence into a character sequence U = {u0, u1, u2, u3, ..., u n }, where u v Let v represent the v-th character in sentence U, n be the length of the sequence, and u0 = [CLS] represent the start marker of the sentence. n =[SEP] indicates a sentence separator; After incorporating positional embeddings and sentence embeddings into the character sequence, the data is input into the BERT model; The BERT model uses complete cultural relic domain entities for masking, masks cultural relic phrases in sentences and predicts the entire phrase, captures the deep semantic connections between cultural relic words, and obtains the feature vector representation of the sentence.

3. The method for extracting and visualizing cultural relic knowledge according to claim 2, characterized in that, It also includes feature vector enhancement for sentences: Extract dependency relations from the input original sentence and construct the corresponding dependency relation matrix; Based on the dependency matrix and the original sentence, construct a graph with n nodes, where the extracted words are nodes and the dependencies are edges connecting the nodes; The feature vector representation of the obtained sentence is input into the GAT model. By concatenating the semantic feature vector and the syntactic feature vector of the text sequence, the final output feature vector is obtained.

4. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The obtained relationship information specifically includes: Average pooling and fully connected layers are used to predict embedding relations for a given sentence embedding C: c pool =Avgpool(C), P r =σ(W r c pool +b r ), In the formula, Avgpool is the average pooling operation, Wr is the trainable weight matrix, σ is the sigmoid activation function, and br is the relevant parameter.

5. The method for extracting and visualizing cultural relic knowledge according to claim 4, characterized in that, The relational information and feature vectors are input into the attention unit, and the specific processing includes: Calculate the attention score for each feature vector; By using attention parameters to perform weighted summation of the words in the sentence, a sentence representation under the current relation is obtained; Relationship information is preserved based on gating strategies; The embedded relational information is fused with the text encoding to obtain the embedded vector representation for entity recognition.

6. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The BiLSTM unit calculates the forward contextual latent vector of each word from the beginning of the sentence and the backward contextual latent vector of each word from the end of the sentence. By concatenating the forward and backward latent vectors, bidirectional context-related word vectors can be obtained. At the same time, the CRF unit identifies the dependencies between contexts; Cultural relic knowledge is extracted based on bidirectional context-dependent word vectors and dependencies.

7. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The attribute dimensions include time attributes, spatial location attributes, and other enumerable attributes.

8. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The cultural relic dimension includes the importance and similarity of the cultural relics.

9. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The global dimensions include the number of cultural relics, the diversity of cultural relics, the importance of cultural relics, the diversity of attributes, and the innovativeness of attributes.

10. The method for extracting and visualizing cultural relic knowledge according to claim 1, characterized in that, The visualization views include basic information and control views, timeline views, cultural relic information views, attribute evolution views, map views, and cultural relic relationship views, wherein the attribute evolution views include single-attribute module views and multi-attribute module views.

Images