Real-time industrial dynamic event clustering method and information system based on industrial knowledge graph

Abstract

The application discloses a kind of real-time industrial dynamic event clustering method and intelligence system based on industrial knowledge graph.The method comprises the following steps: continuously obtaining near real-time updated industrial information documents from field industrial information data source;For each document, use a general large language model to identify event type and event feature argument, form a unique event feature description;Entity linking is carried out in the industrial knowledge graph, and the event subject name is linked to the enterprise standard name, and is reconstructed into a feature-enhanced event description text and vectorization;Calculate its time decay semantic similarity with the existing events in the dynamic event library, determine whether it is a new event by a judgment model, and update the event library accordingly.The present application extracts fine-grained event elements and disambiguates knowledge graph entities to generate consistent and unique event expressions, effectively improving the accuracy and robustness of event clustering, can process incremental information in near real time, realize real-time monitoring and clustering of dynamic events, and has the advantages of high timeliness and strong interpretability.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graphs, specifically relating to a real-time industry dynamic event clustering method and intelligence system based on industry knowledge graphs. Background Technology

[0002] With the development of information technology, information on various websites has exploded, and enterprises and research institutions urgently need to quickly extract valuable industry dynamics from massive amounts of information. Existing information clustering technologies mainly suffer from the following problems: 1. Poor industry adaptability: Existing clustering methods are mostly designed for general news information, failing to consider the relationships between industry entities (such as former company names, parent and subsidiary companies, company brand names, etc.), leading to the same industry event being misclassified into multiple events due to differences in description; 2. Insufficient real-time performance: Traditional offline clustering algorithms require recalculation of the entire dataset, failing to meet the need for immediate viewing of industry dynamics. Full-scale clustering calculations not only cause information lag but also lead to changes in cluster categories due to the addition of new nodes, making it impossible to guarantee consistency between previous and subsequent categories; 3. Information clustering lacks fine-grained event judgment and analysis, failing to support event-level clustering: Existing information clustering technologies typically use keyword matching clustering, which cannot handle the semantic relationships between industry entities; or they only cluster based on the overall text fingerprint of the information, lacking an understanding of key industry dynamics, failing to effectively handle the relationships between different fine-grained events, and even less able to effectively cluster and merge industry dynamics. 4. Inflexible clustering methods: Traditional clustering techniques usually require pre-setting the number of clusters or centers, and cannot dynamically and adaptively generate new events based on actual industry events.

[0003] Therefore, there is an urgent need for a fine-grained clustering and adaptive generation processing solution for information and dynamic industry events that is tailored to the characteristics of the industry and takes into account both real-time performance and accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a real-time industry dynamic event clustering method and intelligence system based on industry knowledge graphs, so as to solve the problems of domain adaptability, dynamic clustering flexibility, and timeliness and continuity in industry information event clustering.

[0005] The specific technical solution adopted in this invention is as follows: In a first aspect, the present invention provides a real-time industry dynamic event clustering method based on industry knowledge graphs, which includes: S1. Continuously acquire near real-time updated industry information documents from industry information data sources pre-built for the target industry sector; S2. For each newly acquired industry information document, use a general large language model to identify the event type of the industry event it describes and the event feature arguments that can ensure the uniqueness of the event. The event feature arguments include at least the event subject name and are composed of event type labels and all event feature arguments to form a unique feature description of the industry event described in each industry information document. S3. For each unique feature description obtained, the event subject name in the description is searched for similarity in the enterprise entity database of the industry knowledge graph. The candidate enterprise entities are judged for uniqueness based on the entity profile using a pre-trained large language model. The event subject name in the unique feature description is replaced with the standard enterprise name of the uniquely determined candidate enterprise entity. The description is then reconstructed into feature-enhanced event description text according to natural word order. Event feature vectors are obtained through text vectorization. S4. For each newly acquired industry information document, calculate the time decay semantic similarity between the industry event it describes and each existing event in the dynamic event database, and input it into the pre-trained judgment model to determine whether the industry event belongs to a new event, and update it to the dynamic event database according to the judgment result and the database entry rules.

[0006] As a preferred embodiment of the first aspect above, the specific implementation method of S1 is as follows: further subdivide the target industry sector into a series of sub-industries; select industry information websites that meet the requirements for update frequency and content quality as data sources, then configure a collection script program for each data source and schedule it on a regular basis, thereby periodically obtaining the latest industry information documents from the data source and storing them in an unstructured document database; use an industry sector classification model to identify the sub-industry of each obtained industry information document and mark it as at least one sub-industry.

[0007] As a preferred embodiment of the first aspect above, in step S2, before extracting the event type and event feature arguments of the industry event, it is necessary to pre-construct an event uniqueness argument identification template for each event type. The template defines the minimum related argument that can identify the uniqueness of the event corresponding to the event type through placeholders. The general large language model needs to first preliminarily identify the event type corresponding to the industry information document, and then determine the minimum related argument to be identified based on the event uniqueness argument identification template corresponding to the identified event type and replace it with the event feature argument in their respective placeholders to form a unique feature description.

[0008] As a preferred embodiment of the first aspect mentioned above, the industry knowledge graph is a knowledge graph constructed from industry domain knowledge, wherein the nodes include enterprises, talents, patents, industry links and administrative divisions, and the relationships between nodes are determined by prior domain knowledge; all enterprise nodes in the industry knowledge graph constitute the enterprise entity library.

[0009] As a preferred embodiment of the first aspect above, the method for determining the uniqueness of the obtained candidate enterprise entities in step S3 is as follows: For each candidate enterprise entity obtained through similarity retrieval in the enterprise entity database, its standard enterprise name and node attribute information are extracted from the industry knowledge graph and used as its profile features. Then, the profile features of all candidate enterprise entities and the unique feature description are input into a pre-trained large language model. The large language model infers and judges to determine the unique candidate enterprise entity that actually corresponds to the event subject name in the unique feature description.

[0010] As a preferred embodiment of the first aspect above, when performing natural word order reconstruction in S3, it is necessary to reconstruct the event type and event feature arguments of the industry event in the unique feature description into a natural language description statement of the industry event in accordance with the natural language expression order through template mapping or large language model generation.

[0011] As a preferred embodiment of the first aspect above, the specific implementation method of S4 is as follows: S41. Classify the historical industry information documents in the dynamic event database according to event type to form different types of event subsets; in each event subset, sort all industry information documents belonging to the same event by publication timestamp, take the first publication time as the event start time, and count the number of industry information documents published for the event every day. Take the time distance from the moment when the number of industry information documents on the day decays to 50% of the peak to the event start time as the half-life of the event; take the weighted median value of the half-life of all events in the event subset of each event type as the propagation half-life of the event type. S42. For each newly acquired industry information document, calculate the industry events it describes. Each existing event in the dynamic event library The time decay semantic similarity between the events is calculated, and the existing events with the highest similarity are selected as the most similar events; the formula for calculating the time decay semantic similarity is: In the formula: Indicates industry events Corresponding event feature vector and existing events Corresponding event feature vector Cosine similarity between them; and These represent industry events. and existing events Each has its own corresponding publication timestamp; is an adjustable hyperparameter; T is the propagation half-life of the event type to which the newly acquired industry information document belongs; S43, Industry Events temporal decay semantic similarity with the closest event Release time difference after compression and their respective event feature vectors and In a binary classification model pre-trained with fused feature vectors as input, the model determines the industry events described in newly acquired industry information documents. If the event is determined to be in the same industry as the most similar event, the event will be merged, and the newly acquired industry information document will be added as the new information source for the most similar event in the dynamic event database. If the event is determined not to be in the same industry, it will be recorded as a new industry event in the dynamic event database.

[0012] As a preferred embodiment of the first aspect mentioned above, in step S41, after the propagation half-life of each event type is calculated during initialization, it needs to be updated periodically through an online event closed-loop feedback adaptive update mechanism. The update method is as follows: within each update cycle, industry events with an additional number of information sources exceeding the threshold through event merging are marked as high-profile events and written into the cache. At the end of the update cycle, the propagation half-life of all high-profile events is calculated. Then, for each event type, the median value of the propagation half-life of high-profile events belonging to that event type is calculated. Finally, the calculation result is updated to the current propagation half-life of that event type using a Bayesian update method.

[0013] Secondly, the present invention provides a real-time industry dynamic event intelligence system based on industry knowledge graph, which is used to implement the real-time industry dynamic event clustering method based on industry knowledge graph as described in any of the first aspects above.

[0014] Thirdly, the present invention provides a computer electronic device, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the real-time industry dynamic event clustering method based on industry knowledge graph as described in any of the first aspects above.

[0015] Compared with the prior art, the present invention has the following advantages: Compared to existing technologies, the real-time dynamic event clustering method based on industry knowledge graphs provided by this invention monitors and clusters industry information at the fine-grained event level. This overcomes the limitations of previous methods that relied solely on deduplication of titles or content, which failed to address semantic diversity, effectively handle relationships between different fine-grained events, or effectively cluster and merge dynamic industry events. This invention eliminates ambiguity by decomposing fine-grained event element tuples and introducing entity links from the industry knowledge graph, generating unique event feature descriptions that are more focused on the event subject than news titles, more concise and clear than news text, and produce a more consistent event description structure, thus improving the accuracy and robustness of event clustering. Furthermore, the dynamic event monitoring system provided by this invention has the ability to process incremental industry information events in real time without requiring full clustering, offering advantages such as high timeliness, high generation accuracy, and traceable and interpretable results. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of a real-time industry dynamic event clustering method based on industry knowledge graphs. Figure 2 A diagram illustrating the storage of different industry information documents in an industry information knowledge base; Figure 3 This is a diagram of the event details page of the industry intelligence monitoring system. Figure 4 A schematic diagram of a company's profile in an industry knowledge graph; Figure 5 This is an exemplary architecture diagram of an industry intelligence monitoring system; Figure 6 This is a schematic diagram of the structure of a computer electronic device. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0018] In a preferred embodiment of the present invention, a real-time industry dynamic event clustering method based on industry knowledge graphs is provided. This method can simultaneously achieve fine-grained event clustering and adaptive generation of industry dynamic events, taking into account the characteristics of the industry field while ensuring real-time performance and accuracy. Figure 1As shown, the specific implementation process of this method includes S1 to S4. The specific implementation process of each step is described in detail below.

[0019] S1. Continuously acquire near real-time updated industry information documents from industry information data sources pre-built for the target industry sector.

[0020] It should be noted that the industry information data sources mentioned above refer to the data sources used to obtain industry information. In theory, these data sources can be any means, such as online websites and media news, but the information needs to be updated in a timely manner, that is, have "near real-time" characteristics. "Near real-time" means real-time or close to real-time (e.g., every few minutes).

[0021] In an embodiment of the present invention, the specific implementation method of S1 is as follows: S11. Further subdivide the target industry sector into a series of sub-sectors.

[0022] As a preferred implementation of this invention, a two-tiered tagging system of "top-level industry - sub-industry" can be adopted for each industry sector to form a multi-level industry classification framework, supporting the aggregation and analysis of information by industry sector. For example, if the top-level industry is artificial intelligence, then the sub-industries can be further divided into large-scale models, next-generation information technology, robotics, etc.

[0023] S12. Select industry information websites that meet the requirements for update frequency and content quality as data sources. Then, configure a collection script program for each data source and schedule it on a regular basis to retrieve the latest industry information documents from the data source and store them in an unstructured document database. Use an industry classification model to identify the sub-industry of each acquired industry information document and mark it as at least one sub-industry.

[0024] To ensure the reliability and timeliness of data sources, the aforementioned data sources should be high-quality industry information sources. High-quality industry information sources require comprehensive evaluation based on both update frequency and content quality to select the best data sources, ensuring comprehensive coverage and configurable granularity. Each sub-industry needs to select websites that provide industry information relevant to that sub-industry as high-quality industry information sources. However, since some websites can provide industry information for different sub-industries, the data sources for different sub-industries may overlap.

[0025] As a preferred implementation of this invention, a general-purpose web scraping program can be developed for batch websites. By configuring the scraping modules for each website, scraping scripts for batch data source websites can be obtained. Then, a scheduling tool is used to schedule the scraping program on a regular basis to obtain the latest original content of industry information. Next, an industry classification model is used to further subdivide each obtained industry information document into one or more subdivided industries, and finally, it is stored in an unstructured document database to form a knowledge base for industry information.

[0026] like Figure 2 As shown, this example demonstrates a series of different industry information documents stored in an industry information knowledge base, which are labeled with different sub-industries.

[0027] S2. For each newly acquired industry information document, use a general large language model to identify the event type of the industry event it describes and the event feature arguments that can ensure the uniqueness of the event. The event feature arguments include at least the event subject name, and the unique feature description of the industry event described in each industry information document is composed of the event type label and all event feature arguments.

[0028] It should be noted that event feature arguments that ensure the uniqueness of an event refer to related arguments that can uniquely describe a particular event. To ensure conciseness, the fewer such arguments, the better. Therefore, event feature arguments can be the minimum related arguments that can identify the uniqueness of an event. Since different types of industry events may have different characteristics, the event feature arguments used to uniquely describe the event may also be different. Therefore, in the embodiments of this invention, to ensure that the large language model can accurately extract event feature arguments that ensure the uniqueness of events, before extracting the event type and event feature arguments of industry events, it is necessary to adhere to the objective description of event uniqueness and define fine-grained event extraction arguments in the form of templates. Specifically, an event uniqueness argument identification template can be pre-constructed for each event type, and the template defines the minimum related arguments that can identify the uniqueness of the event type using placeholders. After an industry information document is input into the general large language model, the general large language model first needs to initially identify the event type corresponding to the industry information document. Then, based on the event unique argument identification template corresponding to the identified event type, it determines the minimum related argument to be identified and replaces it with the event feature argument in their respective placeholders to form a unique feature description.

[0029] In embodiments of the present invention, for example, for an industry-specific enterprise collaboration event, the event type label is "enterprise collaboration," and the minimum relevant arguments ensuring the uniqueness of the event are the participating "enterprise name" and the "collaboration details." Therefore, the event uniqueness argument identification template for an industry-specific enterprise collaboration event is defined as "[enterprise name] Enterprise collaboration [collaboration details]." As another example, for an industry-specific product launch event, the event type label is "product launch," and the minimum relevant arguments ensuring the uniqueness of the event are the launching product's "enterprise name" and the "new product name." Therefore, the event uniqueness argument identification template for a product launch event is "[enterprise name] Product launch [new product name]."

[0030] S3. For each unique feature description obtained, the event subject name in the description is searched for similarity in the enterprise entity database of the industry knowledge graph. The candidate enterprise entities are judged for uniqueness based on the entity profile using a pre-trained large language model. The event subject name in the unique feature description is replaced with the standard enterprise name of the uniquely determined candidate enterprise entity. The description is then reconstructed into feature-enhanced event description text according to natural word order. Event feature vectors are obtained through text vectorization.

[0031] It should be noted that the aforementioned industry knowledge graph is a knowledge graph constructed from industry domain knowledge. Nodes include enterprises, talent, patents, industry segments, and administrative divisions. Relationships between nodes are determined by prior domain knowledge. For example, the relationship between enterprises and patents, and the relationship between talent and patents, can be pre-defined and identified from the patent's bibliographic information. The relationship between enterprises and administrative divisions can be identified from the enterprise's registered address information. Relationships between nodes need to be determined based on different node types, including "[Enterprise] - <belongs to> - [Industry segment]", "[Enterprise] - <application> - [patent]", "[Enterprise] - <located to> - [Administrative division]", etc. Thus, the above nodes and relationships constitute a graph network structure of industry domain knowledge, i.e., the industry knowledge graph. All enterprise nodes in this industry knowledge graph constitute the aforementioned enterprise entity database.

[0032] It should be noted that the term "enterprise" in this invention is used to refer generally to all units and organizations, such as companies, universities, research institutes, non-profit organizations, etc.

[0033] In embodiments of the present invention, when performing a similarity search on the enterprise entity database of the industry knowledge graph using the event subject name in the uniqueness feature description, multiple candidate enterprise entities with similar names may be retrieved. When only one candidate enterprise entity exists, a uniqueness check can be performed without needing to establish an association between the event subject name and this candidate enterprise entity. However, when multiple candidate enterprise entities exist, a uniqueness check is required on the obtained candidate enterprise entities to ensure that the event subject name is associated with exactly one candidate enterprise entity. As a preferred implementation of the present invention, the specific method for uniqueness check is as follows: First, for each candidate enterprise entity (corresponding to a node in the industry knowledge graph) obtained through similarity retrieval in the enterprise entity database, its standard enterprise name and node attribute information are extracted from the industry knowledge graph and used as its profile features. The node attribute information that can be used as profile features can be filtered based on the actual recorded enterprise attribute fields. As a specific example, considering the specific circumstances of the company's enterprise knowledge base, the node attribute information used as profile features in this embodiment includes: the enterprise's unified social credit code, enterprise profile, business scope, geographical location, the industry classification field to which the enterprise belongs, and the bibliographic information and abstracts of the enterprise's 10 most recently applied invention patents.

[0034] Then, the profile features of all candidate enterprise entities, along with the unique feature description, are input into a pre-trained large language model (referred to as the entity link reasoning large language model for easy distinction). The large language model infers and determines the unique candidate enterprise entity that actually corresponds to the event subject name in the unique feature description, thereby obtaining the unique link between the event subject enterprise and the candidate enterprises in the knowledge base. Then, the enterprise entities extracted from the information are renamed and replaced using the standard enterprise names in the knowledge base, that is, the event subject name in the unique feature description is replaced with the standard enterprise name of the candidate enterprise entity. The unique feature description after name replacement is a templated text. For example, a unique feature description after name replacement might be "Alibaba (China) Co., Ltd. Product Release Qwen-Image-Layered Model". This text contains inconsistencies with natural language expression, therefore, natural language order reconstruction needs to be performed again. In embodiments of the present invention, when performing natural language order reconstruction, the event type and event feature arguments of the industry event in the unique feature description can be reconstructed into a natural language description statement of the industry event according to the natural language expression order through template mapping or large language model generation. If we denote the original uniqueness feature description as a text sentence consisting of an industry event type label and event feature arguments: ) ) Among them, label is the industry event type label. These are the characteristic arguments of the event.

[0035] The process of template mapping or large language model generation mentioned above is essentially a reordering and synonym substitution of these labels or arguments, with the actual meaning remaining unchanged.

[0036] When using template mapping, a mapping method can be defined for each event type. For example, for the product release event type mentioned above, it can be mapped from the original template "[Company Name] Product Release [New Product Name Released]" to a new natural language order template "[Company Name] Released New Product: [New Product Name Released]". However, when using a large language model for generation, the large language model can understand the meaning of the original unique feature description and re-express it as a synonymous natural language description. This natural language description is the feature-enhanced event description text. .

[0037] It is important to note that the aforementioned large language model for entity linking reasoning requires pre-training and fine-tuning. The fine-tuning method is as follows: First, use the profile features and unique features of the candidate enterprise entities generated in the previous steps to describe the event features. Then, manually annotate the correct entity link results as supervision labels and form a training corpus dataset consisting of text, instructions, and responses according to the pre-training data format. Finally, use the model pre-training platform and the corpus to fine-tune the local parameters of the general large language model to obtain a large language model that is better suited for entity linking reasoning.

[0038] Additionally, it should be noted that feature-enhanced event description text... By using word embedding tools to vectorize the text, event feature vectors of the industry's dynamic events are obtained. The specific form of the word embedding tool is not limited; it can be implemented using Bert, word2vec, etc.

[0039] Since industry information documents are constantly being updated and added to the knowledge base, steps S2 and S3 above need to be performed continuously for each updated industry information document. First, through steps S2 and S3, the first industry dynamic event is obtained. Feature-enhanced event description text and event feature vector The data is stored in a vector database. Then, for each updated industry information document, steps S2-S3 are repeated to obtain the industry dynamic events in the new information document. Feature-enhanced event description text and event feature vector For these acquired industry information documents, this invention requires clustering and storage, that is, performing dynamic information processing and incremental event clustering and merging based on whether they belong to the same enterprise and the same event. The process of dynamic information processing and incremental event clustering and merging is described in detail below.

[0040] S4. For each newly acquired industry information document, calculate the time decay semantic similarity between the industry event it describes and each existing event in the dynamic event database, and input it into the pre-trained judgment model to determine whether the industry event belongs to a new event, and update it to the dynamic event database according to the judgment result and the database entry rules.

[0041] In an embodiment of the present invention, the specific implementation method of step S4 above is as follows: S41. Classify the historical industry information documents in the dynamic event database according to event type to form different types of event subsets; in each event subset, sort all industry information documents belonging to the same event by publication timestamp, take the first publication time as the event start time, and count the number of industry information documents published for the event every day. Take the time distance from the moment when the number of industry information documents on the day decays to 50% of the peak to the event start time as the half-life of the event; take the weighted median value of the half-life of all events in the event subset of each event type as the propagation half-life of the event type.

[0042] The S41 step described above requires data mining based on historical industry information documents in the dynamic event database. Furthermore, the propagation half-life needs to be calculated separately for each event type, and the calculation timeframe should avoid including events that are too far in the past. For example, for the k-th type of industry event, industry information documents from the past two months should be included in the event subset. Because each industry information document is determined to be a new event during data entry, and if it is not a new event, it is associated with an existing event, thus creating a subset of events. The industry information documents within can be further clustered according to the events they pertain to, with each industry event... A series of events were linked to this incident. Related industry information documents. For a subset of events. Each industry event within the industry, for each individual industry event Related industry information documents are sorted by their publication timestamp, based on industry events. The first release time The earliest published industry news document's timestamp is used as the start time of the event, and then daily statistics are compiled for each industry event. The number of industry information documents published, including those related to industry events on day t. The number of industry information documents published is denoted as N(t), and the peak value of N(t) is determined as Nt. max Then determine how N(t) decays to the peak value N. max The number of days t corresponding to 50% 50% ,Will As an event in this industry propagation half-life Since the propagation half-life is obtained statistically based on event type, it ultimately applies to all industry events of the same type. propagation half-life By calculating the weighted median, the prior propagation half-life of the k-th type of industry event is obtained. The calculation formula is as follows: in This refers to all industry events of the same type. propagation half-life Take the median; It is an empirical compression factor to prevent overestimation of the long tail.

[0043] It should be noted that the above calculation of the propagation half-life relies on sufficient existing events and industry information documents. However, during the cold start phase of initially building a dynamic event library, different types of event subsets can be manually collected and then manually analyzed to determine the propagation half-life.

[0044] Furthermore, as a preferred embodiment of the present invention, when initializing the calculation of the prior propagation half-life for each event type... Subsequently, it should also be updated periodically through an online event closed-loop feedback adaptive update mechanism. The update method is as follows: Within each update cycle, industry events where the number of new information sources added via event merging exceeds a threshold are marked as high-profile events and written to the cache. At the end of the update cycle, the propagation half-life of all high-profile events is calculated. Then, for each event type, the median propagation half-life of high-profile events belonging to that event type is calculated. The calculation result is then updated to the current propagation half-life of that event type using a Bayesian update. The Bayesian update formula is as follows: = in For smoothing coefficients; The propagation half-life before updating the k-th event type (the initial value during the first Bayesian update is...) ); This represents the median propagation half-life of a high-temperature event belonging to the kth event type. The propagation half-life after the update for the k-th event type.

[0045] For example, using a monthly update cycle, each new industry event added to the database within that month... If an event is merged ≥ 3 times within the following 72 hours, it is marked as a high-temperature event; for high-temperature events, their propagation half-life is recalculated. And write it to the high-hot event cache pool set. Then, a Bayesian update based on the event half-life is triggered monthly, with the update formula as follows: = smoothness coefficient ∈ [0.6,0.8], ensuring a smooth change between historical priors and the latest observations.

[0046] After each Bayesian update, the propagation half-life obtained from the update needs to be calculated. Data control is performed, specifically as follows: If If the rate of change exceeds 15%, it is considered a trend drift and requires manual review. Data updates will only proceed after the review is approved. If the manual review fails, the propagation half-life obtained from this update will be discarded, and the original propagation half-life will be used instead. .

[0047] It should be noted that after each Bayesian update, the high-hot event cache pool needs to be cleared again. Then, it will enter the next update cycle.

[0048] S42. For each newly acquired industry information document, calculate the industry events it describes. Each existing event in the dynamic event library The time decay semantic similarity (TDSS similarity) between the events is calculated, and the existing events with the highest similarity are selected as the most similar events. The formula for calculating the time decay semantic similarity is as follows: In the formula: Indicates industry events Corresponding event feature vector and existing events Corresponding event feature vector Cosine similarity between them; and These represent industry events. and existing events Each corresponds to a publication timestamp, in hours; is an adjustable hyperparameter; T is the propagation half-life of the event type to which the newly acquired industry information document belongs.

[0049] The TDSS similarity calculation formula above contains two terms, the first of which is semantic cosine similarity. The second term is the time decay factor. , = T represents the propagation half-life of the event type to which the newly acquired industry information document belongs. To demonstrate industry events and existing events The time difference between the release of information, in hours. TDSS similarity calculation constrains the similarity results to between 0 and 1 by introducing a time decay coefficient and function transformation. The closer the similarity is to 1, the more likely the two events are the same dynamic event.

[0050] However, it should be noted that, due to existing events... Release timestamp Event feature vector These definitions are actually for industry information documents, not for industry events. Defined by, therefore for any existing event You can keep only the existing event. The publication timestamp of an industry information document Event feature vector As this existing event Release timestamp Event feature vector When a new industry information document is added to the database, it needs to be determined whether it belongs to an existing event in the database based on its corresponding event feature vector. If it does, the industry information document is merged into the same event for addition to the database, and a set of publication timestamps is retained for the same event. Event feature vector .

[0051] S43, Industry Events temporal decay semantic similarity with the closest event Release time difference after compression and their respective event feature vectors and In a binary classification model pre-trained with fused feature vectors as input, the model determines the industry events described in newly acquired industry information documents. If the event is determined to be in the same industry as the closest event, the event will be merged, and the newly acquired industry information document will be added as a new information source for the closest event in the dynamic event database. If the event is determined not to be in the same industry, it will be recorded as a new industry event in the dynamic event database, that is, a new industry event will be created, and the newly acquired industry information document will be used as an information source for the new event.

[0052] It should be noted that the fusion feature vector in the above input binary classification decision model is... , here These all refer to industry events described in newly acquired industry information documents within the dynamic event library. The most similar and closest events, i.e., those in the dynamic event database that are most closely related to industry events. The existing event with the highest TDSS similarity.

[0053] like Figure 3 As shown, an example is displayed on the event details page of an industry intelligence monitoring system. This page centrally displays related information sources belonging to the same event, and users can click to view other related information. Furthermore, the event and information clustering results of this invention can be further linked to relevant enterprise nodes in the industry knowledge graph. For example... Figure 4 As shown, an example is presented of a company profile in an industry knowledge graph, where the company's recent news and dynamics are displayed in the Recent Dynamics module by adding links to industry information.

[0054] It should be noted that the method steps shown in S1 to S4 above can essentially be implemented in the form of computer programs or software functional modules.

[0055] Therefore, based on the same inventive concept, the present invention also provides a real-time industry dynamic event intelligence system based on industry knowledge graph, corresponding to the real-time industry dynamic event clustering method based on industry knowledge graph provided in the above embodiments. It implements the real-time industry dynamic event clustering method based on industry knowledge graph provided in the above embodiments by combining functional modules.

[0056] like Figure 5 As shown, as one implementation of the aforementioned real-time industry dynamic event intelligence system based on industry knowledge graphs, it can be implemented by combining functional modules such as an information timed collector, an industry information filter, an industry event identification and fine-grained event argument parser, an industry event subject entity linker, and an industry dynamic event generator to achieve the corresponding method flow.

[0057] The aforementioned timed information collector includes a high-quality information data source knowledge base, an information collector, and a timed trigger. The high-quality information data source knowledge base is used to define the data source matrix for information collection, filtering high-quality data source sites with high update frequency and quality for collection; the information collector contains the collection code for each data source section, used to execute the collection program; the timed trigger is used to control the triggering of the collection program at regular intervals to obtain the latest updated information documents.

[0058] The aforementioned industry information filter is used to initially clean information documents, retain effective information, and use the trained industry sector classification model to filter industry-related information documents.

[0059] The aforementioned industry event identification and fine-grained event argument parser includes an industry event argument recognition template library and an event argument recognition model. The industry event argument recognition template library defines and stores event tags and unique event argument extraction templates for all industry events that need to be identified. The event argument recognition model utilizes the parsing capabilities of a large language model to extract and parse fine-grained arguments for various types of industry events.

[0060] The aforementioned industry event entity linker includes a pre-built industry knowledge graph, an event description generator, and an entity linking reasoning model.

[0061] The aforementioned industry knowledge graph comprises various entity nodes and their relationships within an industry sector, providing an objective profile of the event's subject. The industry knowledge graph consists of nodes such as enterprises, industry segments, patents, and administrative divisions. Edge relationships include "[Enterprise] - <belongs to> - [Industry segment]", "[Enterprise] - <applies to> - [Patent]", and "[Enterprise] - <located to> - [Administrative division]", forming a graph network structure.

[0062] The event description generator described above is used to generate event description statements that can uniquely represent an event. It generates event description statements according to the event tags and arguments and follows the word order rules.

[0063] The aforementioned entity linking reasoning model is used to infer and determine the link relationships between entities such as enterprises in information events and entity nodes in the industry knowledge graph. The pre-training corpus utilizes the background of event statements, candidate enterprise entities, and manually annotated entity link results to form an entity linking corpus. The entity linking reasoning model is obtained through fine-tuning training based on a general large language model. The standardized name of an entity can be obtained by associating it with entities in the industry knowledge graph.

[0064] The aforementioned industry dynamic event generator includes an event vector converter, a time-decay event similarity calculator, a same event determiner, an industry dynamic event incremental updater, and an industry event half-life dynamic adaptation module. The event vector converter uses the standardized names of the subjects obtained from the event subject linker to standardize and transform event descriptions, forming feature-enhanced unique event descriptions, and encodes the events into feature vectors through a vectorization model. The time-decay event similarity calculator calculates the similarity between a new event and existing events in the database, based on the dynamic event half-life, to obtain the most similar industry event. The same event determiner contains a trained determination model used to determine whether the new event and the most similar event in the database belong to the same industry dynamic event, obtaining the determination result. The event half-life dynamic adaptation module periodically performs adaptive dynamic calculations of the event half-life based on high-profile events and stores the results in a cache for use by the event similarity calculation module. The industry dynamic event incremental updater is responsible for performing subsequent data processing on the incremental event judgment results. For example, if it is the same event, the new information event will be merged into the existing event record in the database and the information source of the event will be updated; if it does not belong to the same event, a new event record will be added to the industry dynamic event database.

[0065] Furthermore, based on the same inventive concept, such as Figure 6 As shown, the present invention also provides a computer electronic device corresponding to the real-time industry dynamic event clustering method based on industry knowledge graph provided in the above embodiments, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the real-time industry dynamic event clustering method based on industry knowledge graph as described above when executing the computer program.

[0066] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, 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 the present invention.

[0067] Therefore, based on the same inventive concept, this invention provides a computer-readable storage medium corresponding to the real-time industry dynamic event clustering method based on industry knowledge graph. The storage medium stores a computer program, which, when executed by a processor, can realize the real-time industry dynamic event clustering method based on industry knowledge graph as described above.

[0068] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the real-time industry dynamic event clustering method based on industry knowledge graph as described above.

[0069] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S4.

[0070] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0071] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0072] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0073] The embodiments described above are merely some preferred implementations of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A real-time industry dynamic event clustering method based on industry knowledge graph, characterized in that, include: S1. Continuously acquire near real-time updated industry information documents from industry information data sources pre-built for the target industry sector; S2. For each newly acquired industry information document, use a general large language model to identify the event type of the industry event it describes and the event feature arguments that can ensure the uniqueness of the event. The event feature arguments include at least the event subject name and are composed of event type labels and all event feature arguments to form a unique feature description of the industry event described in each industry information document. S3. For each unique feature description obtained, the event subject name in the description is searched for similarity in the enterprise entity database of the industry knowledge graph. The candidate enterprise entities are judged for uniqueness based on the entity profile using a pre-trained large language model. The event subject name in the unique feature description is replaced with the standard enterprise name of the uniquely determined candidate enterprise entity. The description is then reconstructed into feature-enhanced event description text according to natural word order. Event feature vectors are obtained through text vectorization. S4. For each newly acquired industry information document, calculate the time decay semantic similarity between the industry event it describes and each existing event in the dynamic event database, and input it into the pre-trained judgment model to determine whether the industry event belongs to a new event, and update it to the dynamic event database according to the judgment result and the database entry rules.

2. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, The specific implementation method of S1 is as follows: further divide the target industry into a series of sub-industries; select industry information websites that meet the requirements of update frequency and content quality as data sources, then configure a collection script program for each data source and schedule it on a regular basis, and then periodically obtain the latest industry information documents from the data source and store them in an unstructured document database; use the industry classification model to identify the sub-industry of each obtained industry information document and mark it as at least one sub-industry.

3. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, In step S2, before extracting the event type and event feature arguments of industry events, it is necessary to pre-construct an event uniqueness argument identification template for each event type. The template defines the minimum related argument that can identify the uniqueness of the event corresponding to the event type through placeholders. The general large language model needs to first preliminarily identify the event type corresponding to the industry information document, and then determine the minimum related argument to be identified based on the event uniqueness argument identification template corresponding to the identified event type and replace it with the event feature argument in their respective placeholders to form a unique feature description.

4. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, The industry knowledge graph is a knowledge graph constructed from industry domain knowledge, where nodes include enterprises, talents, patents, industry links and administrative divisions, and the relationships between nodes are determined by prior domain knowledge; all enterprise nodes in the industry knowledge graph constitute the enterprise entity library.

5. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, The method for determining the uniqueness of the obtained candidate enterprise entities in S3 is as follows: For each candidate enterprise entity obtained through similarity retrieval in the enterprise entity database, its standard enterprise name and node attribute information are extracted from the industry knowledge graph and used as its profile features. Then, the profile features of all candidate enterprise entities and the unique feature description are input into a pre-trained large language model, and the large language model infers and judges to determine the unique candidate enterprise entity that actually corresponds to the event subject name in the unique feature description.

6. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, When performing natural word order reconstruction in S3, it is necessary to reconstruct the event type and event feature arguments of the industry event in the unique feature description into a natural language description statement of the industry event in accordance with the natural language expression order through template mapping or large language model generation.

7. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, The specific implementation method of S4 is as follows: S41. Classify the historical industry information documents in the dynamic event database according to event type to form different types of event subsets; in each event subset, sort all industry information documents belonging to the same event according to the publication timestamp, take the first publication time as the event start time, and count the number of industry information documents published for the event every day. The time distance from the moment when the number of industry information documents on the day decays to 50% of the peak value to the event start time is taken as the half-life of the event. The weighted median of the half-lives of all events in the event subset of each event type is taken as the propagation half-life of that event type. S42. For each newly acquired industry information document, calculate the industry events it describes. Each existing event in the dynamic event library The time decay semantic similarity between them is used to select the existing events with the highest similarity as the most similar events; The formula for calculating the time decay semantic similarity is: In the formula: Indicates industry events Corresponding event feature vector and existing events Corresponding event feature vector Cosine similarity between them; and These represent industry events. and existing events Each has its own corresponding publication timestamp; These are adjustable hyperparameters; T represents the propagation half-life of the event type to which the newly acquired industry information document belongs; S43, Industry Events temporal decay semantic similarity with the closest event Release time difference after compression and their respective event feature vectors and In a binary classification model pre-trained with fused feature vectors as input, the model determines the industry events described in newly acquired industry information documents. If the event is determined to be in the same industry as the most similar event, the event will be merged, and the newly acquired industry information document will be added as the new information source for the most similar event in the dynamic event database. If the event is determined not to be in the same industry, it will be recorded as a new industry event in the dynamic event database.

8. The real-time industry dynamic event clustering method based on industry knowledge graph as described in claim 1, characterized in that, In step S41, after the propagation half-life of each event type is calculated during initialization, it needs to be updated periodically through an online event closed-loop feedback adaptive update mechanism. The update method is as follows: within each update cycle, industry events with more than the threshold number of new information sources added by event merging are marked as high-heat events and written into the cache. At the end of the update cycle, the propagation half-life of all high-heat events is calculated. Then, for each event type, the median value of the propagation half-life of high-heat events belonging to that event type is calculated. The calculation result is then updated to the current propagation half-life of that event type using a Bayesian update method.

9. A real-time industry dynamic event intelligence system based on industry knowledge graphs, characterized in that, This method is used to implement the real-time industry dynamic event clustering method based on industry knowledge graph as described in any one of claims 1 to 8.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the real-time industry dynamic event clustering method based on industry knowledge graph as described in any one of claims 1 to 8.

Images