CN111930893A - Knowledge graph construction method for scenic spot abnormal events - Google Patents

Abstract

The invention discloses a knowledge map construction method for abnormal events in scenic spots, which includes the following major steps: collecting relevant Internet document data, examining the quality of the document data to decide further sorting work; extracting abnormality in the collected data The entities and relationships of events are connected with each other by a specific matching method; then the extracted entities are clustered into multiple entity clusters, so as to construct the pattern layer of the knowledge map of abnormal events; Each entity is divided into separate sub-graphs; at the end, the existing sub-graphs are merged to form a complete knowledge graph based on abnormal events. The invention combines the knowledge graph with the abnormal events of the scenic spot, and can sort out the relevant information of the abnormal events in the scenic spot in detail and quickly, so as to better select the countermeasures for the abnormal events, and the present invention can well deal with the complex entity relationship in the knowledge graph. to improve the data acquisition efficiency and construction quality of knowledge graphs.

Description

A knowledge graph construction method for abnormal events in scenic spots

technical field

The invention relates to the technical field of natural language processing, in particular to a knowledge map construction method for abnormal events in scenic spots.

Background technique

With the continuous development of the global economy and the continuous increase of the world population, the number of tourists in the scenic spot is increasing, so the scenic spot has also become a highly densely populated place, making it prone to various abnormal events that affect the order of the scenic spot and even cause serious public property damage loss. At present, the research on abnormal events in scenic spots mainly uses monitoring methods to prevent them. However, various abnormal events are complex and variable. Real-time monitoring can only be prepared after the event occurs, and cannot be prevented from the root cause, and the detection system cannot detect the event. Helps in decision-making for abnormal events in subsequent processing. How to quickly respond to abnormal events in scenic spots has become an urgent problem to be solved in related fields.

In recent years, knowledge graph has gradually become a popular scientific research field. The knowledge graph combines the theories and methods of applied mathematics, graphics, information visualization technology, information science and other disciplines with methods such as metrology citation analysis and co-occurrence analysis, and uses the visual graph to visualize the core structure and development history of the discipline. , cutting-edge fields, and the modern theory of the overall knowledge structure to achieve the purpose of multi-disciplinary integration. It is a series of different graphics showing the development process and structural relationship of knowledge. Visual technology is used to describe knowledge resources and their carriers, and to mine, analyze, construct, and draw. and show knowledge and the interconnections between them.

With its powerful semantic representation ability, knowledge spectrum graph is very suitable as a representation carrier of complex relationships involved in abnormal events. However, there are very few researches on the construction of knowledge graphs in the field of abnormal events in scenic spots. The existing knowledge graph construction technologies only involve the construction of conventional knowledge graphs, and the methods are less portable and less applicable to specific fields. The extraction and construction of knowledge in the field of abnormal events are based on traditional methods. The efficiency of acquisition and analysis is low, and it is impossible to quickly query and associate relevant information, and it cannot effectively help the formulation of response plans for abnormal events in scenic spots. . Therefore, designing an efficient construction method based on the knowledge graph of abnormal events in scenic spots is a very important issue at present.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, to provide a lightweight and efficient method for constructing a knowledge map of specific abnormal events in scenic spots, to realize rapid query of abnormal event information, and to provide a certain prevention and response plan for the occurrence of abnormal events. effective formulation.

The technical scheme that realizes the object of the present invention is:

A knowledge graph construction method for abnormal events in scenic spots, comprising the following steps:

S1: Use the constructed web crawler to collect data and documents related to abnormal events in scenic spots from the Internet, and then use the TF-IDF text similarity calculation method to judge the data relevance of the documents, and filter out the most suitable data documents to keep;

S2: Establish a semantic word segmentation dataset in the relevant information of the abnormal events in the scenic spot, mark the abnormal event entities and the part of speech of the relationship between the entities in the word segmentation dataset as nouns and verbs, and extract them from the crawled data documents according to the marked nouns and verbs Entity and entity relationship of abnormal events in scenic spots, and related entities and relationships through original data documents;

S3: Cluster the extracted entities to form multiple different entity clusters, each entity cluster contains several entity types, refer to the description structure of the professional document to construct the knowledge graph mode layer corresponding to the abnormal event, so as to Ontology is constructed;

S4: Find the document data corresponding to the entity in the entity cluster according to the entity relationship of the abnormal event extracted in step 2, find the description words that are similar to the entity in the document corresponding to the corresponding entity through cosine similarity calculation, and associate the entity with the entity. The relationship of similar description words is matched to obtain the corresponding sub-graph;

S5: Merge all sub-graphs: The sub-graphs are connected by relationship, and knowledge is merged to obtain a knowledge graph for abnormal events in scenic spots.

Further, the step of judging the data relevance of the document by the TF-IDF text similarity calculation method in the step S1 is:

S1-1: Establish a corresponding data quality judgment model, as shown in formula (1):

In the formula, S _i,j' represents the similarity between the jth document of the i-th layer and all the currently crawled documents, and S _i-1,j represents the j-th document of the i-1th layer and all the currently crawled documents. The similarity of documents, similarly, S _i,j',k represents the similarity between the jth document of the i-th layer and the kth document of the layer, where the link of the j'th document is between the jth document. In, Wi and T _i represent the weight values set by the _i -th layer;

S1-2: Set the threshold α, when the similarity S _i,j of the document to be judged and other documents is less than the threshold α, the current document is judged to be unqualified, and the number of documents that do not meet the conditions in this layer is counted as X _i , the number of documents whose content of the layer is duplicated with the determined document content is counted as Y _i ;

S1-3: Count the number Ni of all documents in this layer, calculate the failure rate of documents in this layer (Xi+Yi)/Ni, and set a new threshold β to judge whether to continue crawling the documents in this layer, if not qualified If the rate is greater than the threshold β, the crawling of the next layer of documents is stopped.

Further, the word segmentation data set in the step S2 is to mark the part of speech of the words in the word segmentation thesaurus in the document when crawling the document, and delete the stop words that appear in the document.

Further, the operation method of clustering in the step S3 is:

S3-1: Use the compound neural network model to train the word vectors for the extracted text entities to obtain distributed expressions containing grammatical and semantic feature words;

S3-2: After the distributed expression of words is completed, the KMeans unsupervised algorithm is used to cluster the entities.

Further, entity mapping is also adopted in the method for constructing the sub-atlas in the step S4, and the method is: first find the document corresponding to a certain entity, and then find out the related paper document similar to the document through the document link, And establish the mapping relationship between the entity and these documents.

Further, the method for merging graphs in the step S5 is: if the two subgraphs contain the same entity a, then the entity b connected to the entity a in one of the subgraphs is connected to the entity a in the other subgraph, Complete the merging of sub-graphs; if a sub-graph does not have the same entity as other sub-graphs, calculate the entity similarity of each sub-graph and then merge.

Further, the specific method for calculating the similarity between the sub-atlas is:

S5-1: Corresponding the crawled document data with the links of the upper-level documents, and divide them into documents linked at the same level and documents linked at different levels;

S5-2: A sub-graph corresponds to document A in the total knowledge graph, another sub-graph corresponds to document B in the knowledge graph, and there is no identical entity in the two sub-graphs, then calculate the entity in document A and the entity in document B and the first The similarity of all entities in other documents in layer i, the candidate sub-atlas with the highest similarity with the sub-atlas to be merged is screened out. The specific formula is as follows:

where G _k represents the candidate sub-map with the highest similarity to the sub-map to be merged, |G ₁ | and |Gt| represent the number of entities in the 1st and t-th candidate sub-maps, respectively, and a represents the sub-map to be merged The entities in , n represents the number of entities in the sub-graph to be merged corresponding to document G, b represents the entities in the candidate sub-graph, and s(a, b) represents the similarity between entities a and b;

where E represents the entity in the corresponding sub-graph to be merged in document A with the greatest similarity to all entities in the candidate sub-graph, |G _k | represents the number of entities in the candidate sub-graph with the greatest similarity to the sub-graph to be merged ;

E'=arg max{s(b,E)} (b=1,2,...,|G _k |) (4)

Among them, E' represents the entity with the largest similarity between the entity and the entity E in the candidate sub-graph. The entity E in the sub-graph to be merged corresponding to the document A is connected with the entity E' in the candidate sub-graph G _k , and the description is marked. is calculated by the model;

S5-3: Merge the candidate sub-graph corresponding to the document with the highest similarity with the sub-graph to be merged corresponding to document A.

The invention provides a method for extracting abnormal events in scenic spots based on a composite neural network. The method can reduce the dependence of event extraction on trigger words, avoid the influence of ambiguity of trigger words on category judgment, and can be effective in irregular sentence patterns. Extract events.

Description of drawings

Fig. 1 is the flow chart of the inventive method;

Fig. 2 is a flowchart of crawling document data and judging data quality;

Fig. 3 is the schematic diagram of the knowledge graph mode layer of the abnormal event of the scenic spot;

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific implementations. This implementation example is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation manner and a specific operation process, but the protection scope of the present invention is not limited to the following implementation example.

Example:

As shown in the flow chart of the method in FIG. 1 , the present embodiment takes a fire event as an example to provide a method for constructing a knowledge graph for abnormal events in scenic spots, and the method includes the following steps:

Step 1: Use the constructed web crawler to collect data and documents related to abnormal events in scenic spots from the Internet, and then use the TF-IDF text similarity calculation method to judge the data relevance of the documents, and filter out the most suitable data documents to retain. The specific process as shown in picture 2;

Step 1-1: Establish a corresponding data quality judgment model, as shown in formula (1):

In the formula, S _i,j' represents the similarity between the jth document of the i-th layer and all the currently crawled documents, and S _i-1,j represents the j-th document of the i-1th layer and all the currently crawled documents. The similarity of documents, similarly S _i,j',k represents the similarity between the jth document of the i-th layer and the k-th document of the layer, where the link of the j'th document is in the jth document , Wi and T _i represent the weight values set by the _i -th layer;

Step 1-2: Set the threshold α, when the similarity S _i,j of the document to be judged and other documents is less than the threshold α, the current document is determined to be unqualified, and the number of unqualified documents in this layer is counted as X _i , the number of documents whose content of the layer is duplicated with the determined document content is counted as Y _i ;

Step 1-3: Count the number Ni of all documents in this layer, calculate the failure rate (Xi+Yi)/Ni of documents in this layer, and set a new threshold β to judge whether to continue crawling documents in this layer, if not. If the pass rate is greater than the threshold β, the crawling of the next layer of documents is stopped.

In the process of document data crawling, judging the data quality of the obtained document while crawling, and deciding whether to carry out deeper data crawling, can improve the quality and efficiency of the obtained data, and reduce the later text cleaning and analysis work. difficulty. The data mining method in the present invention is also applicable in the process of constructing knowledge graphs in other fields;

Step 2: Establish a semantic word segmentation data set in the relevant information of the abnormal events in the scenic spot, mark the abnormal event entities in the word segmentation data set and the part-of-speech of the relationship between the entities and mark them as nouns and verbs, according to the marked nouns and verbs in the crawled data document Extract entities and entity relationships of abnormal events in scenic spots, and associate entities and relationships through original data documents. The word segmentation thesaurus is used to mark the part of speech of the words in the thesaurus in the document when crawling the content of the document, and the words belonging to the stop word library appearing in the document are directly deleted.

Establishing a text thesaurus for the field of fire events can make the extraction results more accurate. Part-of-speech tagging on the results can improve the accuracy and recall rate of specific event information extraction, and the part-of-speech information extraction is oriented to each sub-document to ensure that the entity alignment to reduce complexity.

Step 3: Cluster the extracted entities to form a number of different entity clusters, each entity cluster contains several entity types, refer to the description structure of the professional document to construct the knowledge graph mode layer corresponding to the abnormal event, as shown in the figure 3 shown;

The specific method of clustering operation is:

Step 3-1: Use the compound neural network model to train the word vectors for the extracted text entities, and obtain distributed expressions containing grammatical and semantic feature words;

Step 3-2: After completing the distributed expression of words, use KMeans unsupervised algorithm to cluster entities.

Step 4: Find the document data corresponding to the entity in the entity cluster according to the entity relationship of the abnormal event extracted in Step 2, and find the description words that are similar to the entity in the document corresponding to the corresponding entity through cosine similarity calculation. Match the relationship with the similar description words to obtain the corresponding sub-map;

In the method of constructing the sub-graph, entity mapping is also adopted. The specific method is: first find the document corresponding to an entity, and then find the related paper documents similar to the document through the document link, and establish the entity and these documents. Mapping relationship between documents;

Step 5: Merge all sub-graphs: Connect the sub-graphs by relationship, and merge knowledge to obtain a knowledge graph for abnormal events in scenic spots; the method of calculating the similarity between the candidate sub-graphs and the sub-graphs to be merged is:

Step 5-1: Corresponding the crawled document data with the links of the upper-level documents, and divide them into documents linked at the same level and documents linked at different levels;

Step 5-2: A sub-graph corresponds to document A in the general knowledge graph, and another sub-graph corresponds to document B in the knowledge graph. If there are no identical entities in the two sub-graphs, calculate the sum of the entity in document A and the entity in document B. The similarity of all entities in other documents of the i-th layer, and the candidate sub-map with the highest similarity to the sub-map to be merged is screened out. The specific formula is as follows:

where G _k represents the candidate sub-map with the highest similarity to the sub-map to be merged, |G ₁ | and |Gt| represent the number of entities in the 1st and t-th candidate sub-maps, respectively, and a represents the sub-map to be merged The entities in , n represents the number of entities in the sub-graph to be merged corresponding to document G, b represents the entities in the candidate sub-graph, and s(a, b) represents the similarity between entities a and b;

其中E表示文档A中对应的待合并子图谱中的实体与候选子图谱中所有实体相似度最大的实体，|G_k|表示与待合并子图谱相似度最大的候选子图谱中实体的个数；

where E represents the entity in the corresponding sub-graph to be merged in document A with the greatest similarity to all entities in the candidate sub-graph, |G _k | represents the number of entities in the candidate sub-graph with the greatest similarity to the sub-graph to be merged ;

E'=arg max{s(b,E)} (b=1,2,...,|Gk|) (4)

Among them, E' represents the entity with the largest similarity between the entity and the entity E in the candidate sub-graph. The entity E in the sub-graph to be merged corresponding to the document A is connected with the entity E' in the candidate sub-graph G _k , and the description is marked. is calculated by the model;

Step 5-3: Merge the candidate sub-graph corresponding to the document with the highest similarity with the sub-graph to be merged corresponding to document A.

The present invention constructs sub-graphs of related fields first, and then merges the knowledge graphs according to the sub-graphs, which can improve the construction efficiency of the knowledge graphs and the quality of the domain knowledge graphs.

The above is only a preferred embodiment of the present invention. For those skilled in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements should also be regarded as the present invention. the scope of protection of the invention.

Claims (7)

1. a knowledge map construction method for abnormal events in scenic spots, is characterized in that: comprise the following steps: S1: Use the constructed web crawler to collect data and documents related to abnormal events in scenic spots from the Internet, and then use the TF-IDF text similarity calculation method to judge the data relevance of the documents, and filter out the most suitable data documents to keep; S2: Establish a semantic word segmentation dataset in the relevant information of the abnormal events in the scenic spot, mark the abnormal event entities and the part of speech of the relationship between the entities in the word segmentation dataset as nouns and verbs, and extract them from the crawled data documents according to the marked nouns and verbs Entity and entity relationship of abnormal events in scenic spots, and related entities and relationships through original data documents; S3: Cluster the extracted entities to form multiple different entity clusters, each entity cluster contains several entity types, refer to the description structure of the professional document to construct the knowledge graph mode layer corresponding to the abnormal event, so as to Ontology is constructed; S4: Find the document data corresponding to the entity in the entity cluster according to the entity relationship of the abnormal event extracted in step 2, find the description words that are similar to the entity in the document corresponding to the corresponding entity through cosine similarity calculation, and associate the entity with the entity. The relationship of similar description words is matched to obtain the corresponding sub-graph; S5: Merge all sub-graphs: The sub-graphs are connected by relationship, and knowledge is merged to obtain a knowledge graph for abnormal events in scenic spots. 2. knowledge graph construction method according to claim 1 is characterized in that: in described step S1, the step of judging the data correlation degree of document by TF-IDF text similarity calculation method is: S1-1: Establish a corresponding data quality judgment model, as shown in formula (1):

In the formula, S _i,j' represents the similarity between the jth document of the i-th layer and all the currently crawled documents, and S _i-1,j represents the j-th document of the i-1th layer and all the currently crawled documents. The similarity of documents, similarly, S _i,j',k represents the similarity between the jth document of the i-th layer and the kth document of the layer, where the link of the j'th document is between the jth document. In, Wi and T _i represent the weight values set by the _i -th layer; S1-2: Set the threshold α, when the similarity S _i,j of the document to be judged and other documents is less than the threshold α, the current document is judged to be unqualified, and the number of documents that do not meet the conditions in this layer is counted as X _i , the number of documents whose content of the layer is duplicated with the determined document content is counted as Y _i ; S1-3: Count the number Ni of all documents in this layer, calculate the failure rate of documents in this layer (Xi+Yi)/Ni, and set a new threshold β to judge whether to continue crawling the documents in this layer, if not qualified If the rate is greater than the threshold β, the crawling of the next layer of documents is stopped. 3. The knowledge graph construction method according to claim 1, wherein the word segmentation data set in the step S2 is to mark the part of speech of the word in the word segmentation thesaurus in the marked document when crawling the document, and delete its occurrence Stop words in the document. 4. knowledge graph construction method according to claim 1 is characterized in that: the operation method of clustering in described step S3 is: S3-1: Use the compound neural network model to train the word vectors for the extracted text entities to obtain distributed expressions containing grammatical and semantic feature words; S3-2: After the distributed expression of words is completed, the KMeans unsupervised algorithm is used to cluster the entities. 5. The knowledge graph construction method according to claim 1 is characterized in that: entity mapping is also adopted in the method for constructing sub-graphs in the step S4, and its method is: first find the document corresponding to a certain entity, Then find the related paper documents similar to the document through the document link, and establish the mapping relationship between the entity and these documents. 6. The knowledge graph construction method according to claim 1 is characterized in that: the method for merging graphs in the step S5 is: if both subgraphs contain the same entity a, then one of the subgraphs is neutralized with the entity a The connected entity b is connected to the entity a in another sub-graph to complete the merging of the sub-graphs; if there is no entity in a sub-graph that is the same as other sub-graphs, the entity similarity of each sub-graph is calculated and then merged. . 7. knowledge graph construction method according to claim 6 is characterized in that: the method for the similarity between the described calculation subgraphs is: S5-1: Corresponding the crawled document data with the links of the upper-level documents, and divide them into documents linked at the same level and documents linked at different levels; S5-2: A sub-graph corresponds to document A in the total knowledge graph, another sub-graph corresponds to document B in the knowledge graph, and there is no identical entity in the two sub-graphs, then calculate the entity in document A and the entity in document B and the first The similarity of all entities in other documents in layer i, the candidate sub-atlas with the highest similarity with the sub-atlas to be merged is screened out. The specific formula is as follows:

where G _k represents the candidate sub-map with the highest similarity to the sub-map to be merged, |G ₁ | and |Gt| represent the number of entities in the 1st and t-th candidate sub-maps, respectively, and a represents the sub-map to be merged The entities in , n represents the number of entities in the sub-graph to be merged corresponding to document G, b represents the entities in the candidate sub-graph, and s(a, b) represents the similarity between entities a and b;

where E represents the entity in the corresponding sub-graph to be merged in document A with the greatest similarity to all entities in the candidate sub-graph, |G _k | represents the number of entities in the candidate sub-graph with the greatest similarity to the sub-graph to be merged ; E'=arg max{s(b,E)}(b=1,2,...,|G _k |) (4) Among them, E' represents the entity with the largest similarity between the entity and the entity E in the candidate sub-graph. The entity E in the sub-graph to be merged corresponding to the document A is connected with the entity E' in the candidate sub-graph G _k , and the description is marked. is calculated by the model; S5-3: Merge the candidate sub-graph corresponding to the document with the highest similarity with the sub-graph to be merged corresponding to document A.

Images