Early warning event detection method and electronic device based on large model summary vector

Abstract

The application provides a big model summary vector-based early warning event detection method, and belongs to the technical field of big models, and comprises the following steps: S1, using a big model to generate a semantic vector for each appeal event; S2, using semantic similarity for clustering; S3, using a big model to process all events in each pre-defined category obtained in step S2 to obtain a summary vector of the pre-defined category; S4, calculating the cosine similarity between the semantic vector of each appeal event obtained in step S1 and each summary vector obtained in step S3, and reclassifying according to the cosine similarity to obtain actual categories; S5, counting the number of appeal events contained in each actual category obtained in step S4, and the actual category containing appeal events satisfying a specific requirement is a leading category, and all appeal events contained in the leading category are leading events; the method has generalization capability, and does not need to pre-classify events, and can efficiently and autonomously detect leading events.

Description

Technical Field

[0001] This invention belongs to the field of large model technology, specifically relating to a method and electronic device for detecting emerging events based on large model summary vectors. Background Technology

[0002] Events are a key focus of online information, while precursory events are foreshadowing events of a potential major event (which may involve social and livelihood issues, politically sensitive issues, or sudden incidents). They can reveal signs that a major event is about to occur. Identifying and analyzing precursory events can help us to promptly grasp the occurrence and development trend of major events.

[0003] In the realm of public complaints, a work order is defined as an event. Unlike online events, public complaint events do not have quantifiable metrics such as page views, clicks, reads, comments, or shares. This significantly impacts algorithm technology and system design.

[0004] In network incidents, existing technologies generally use business rule matching, text clustering, and conditional causal relationship extraction algorithms to achieve the desired results.

[0005] (1) Business rule matching is generally judged by using a certain number of criteria, such as key business categories, whether it contains keywords or multiple keywords (synonyms), time range limitation, etc.

[0006] (2) Text clustering. Text clustering algorithms generally involve the following steps:

[0007] Data preprocessing: Preprocessing the raw text, such as removing stop words and punctuation marks, to obtain meaningful features.

[0008] Feature extraction: Extract text features using methods such as bag-of-words model, TF-IDF algorithm, and word embedding, and convert the text into a numerical vector.

[0009] TF-IDF (term frequency–inverse document frequency) here refers to a commonly used weighting technique for information retrieval and data mining. It is often used to mine keywords in articles, and the algorithm is simple and efficient, often used in industry for initial text data cleaning.

[0010] Clustering: Use K-Means, DBSCAN, hierarchical clustering algorithms, etc. to cluster text vectors.

[0011] K-Means here refers to a clustering algorithm based on Euclidean distance, which considers two objects to be more similar the closer they are.

[0012] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based spatial clustering algorithm that divides regions with sufficient density into clusters and discovers clusters of arbitrary shapes in a noisy spatial database. It defines a cluster as the largest set of density-connected points.

[0013] (3) Conditional Causal Relationship Extraction Algorithm

[0014] This method is essentially a rule-fusion supervised classification algorithm. The specific process is shown in CN110705597A, and its main content is:

[0015] The causal events in a causal event pair are designated as causal events and stored in a causal event sample library. Using the data from this library as the training set, a first causal event classifier based on machine learning is trained. The causal relationships between causal event pairs are used as causal event judgment rules and stored in a causal event judgment rule library. A second rule-based causal event classifier is then constructed using this rule library. Events are extracted from a specified network platform to obtain multiple structured events. Structured events referring to the same event within these multiple structured events are unified into a single common-reference event. This common-reference event is then generalized to obtain abstract events of the network platform. The first and second causal event classifiers are used to process these abstract events respectively, and the combined results are used as the detection result for causal events on the network platform.

[0016] The problems with the above technical solution are as follows:

[0017] (1) Business rules can solve some known problems, but cannot exhaustively list new things that are not covered by the rules. This solution needs to address the problem of poor generalization ability of business rules.

[0018] (2) Clustering algorithms are essentially clustering known samples. It is impossible to predict in advance how many categories they can be divided into, nor can it determine how many initiating events there will be. This solution needs to address the problem of pre-selecting and setting the number of cluster categories.

[0019] (3) The conditional causal relationship extraction algorithm involves too much manual annotation, but annotation and supervised training lead to decreased efficiency, and the generalization ability will also be greatly affected over time. This solution needs to use unsupervised methods to identify early events while ensuring generalization ability, so as to improve the response capability of early events. Summary of the Invention

[0020] This invention addresses the shortcomings of existing technologies by providing a method and electronic device for detecting emerging events based on large model summary vectors. It has generalization capabilities and can efficiently and autonomously detect emerging events without prior event classification, manual annotation, or supervised training.

[0021] To achieve the above objectives, the present invention adopts the following technical solution:

[0022] The seed event detection method based on large model summary vectors includes the following steps:

[0023] S1. Use a large model to generate semantic vectors for each appeal event; reduce the dimensionality of the semantic vectors to obtain a dimensionality-reduced vector; extract statistical features from the semantic vectors to obtain a statistical feature vector; and concatenate the dimensionality-reduced vector and the statistical feature vector to obtain a concatenated vector.

[0024] S2. Clustering using semantic similarity, which includes the following steps:

[0025] S21. Select any one appeal event from all appeal events, calculate the cosine similarity between the concatenated vectors of the appeal event and all other appeal events. Two appeal events with a cosine similarity of not less than a set value are classified into the same predefined category. If no appeal event has a cosine similarity of not less than the set value with the appeal event, then the appeal event is classified into a separate predefined category.

[0026] S22. Remove the categorized events from all the claims, and then repeat step S21.

[0027] S23. Repeat step S22 until all the claims have been categorized.

[0028] S3. Process all events in each predefined category obtained in step S2 through a large model to obtain a summary vector for that predefined category; the dimension of the summary vector is the same as the dimension of the semantic vector;

[0029] S4. Calculate the cosine similarity between the semantic vector of each appeal event obtained in step S1 and the summary vector obtained in step S3. If the cosine similarity between the semantic vector of a certain appeal event and multiple summary vectors is not less than a set value, then classify the appeal event into the predefined category corresponding to the summary vector with the largest cosine similarity, and classify the predefined category into an independent actual category. If the cosine similarity between the semantic vector of a certain appeal event and any summary vector is less than a set value, then classify the appeal event unit into an independent actual category.

[0030] S5. Count the number of appeal events contained in each actual category obtained in step S4. The actual category containing the number of appeal events that meets the specific requirements is the initiation category. All appeal events contained in the initiation category are initiation events.

[0031] Preferably, in step S1, the semantic vector of the i-th appeal event generated using the large model is V. i o , where i = 1, 2, 3, ..., N, and N is the total number of claims to be analyzed;

[0032] semantic vector V through UMAP i o Dimensionality reduction is performed to obtain the dimensionality-reduced vector V. i p ;

[0033] The statistical feature vector of the i-th claim is V. i s The statistical feature vector V i s Including semantic vector V i o The arithmetic mean and several quantiles;

[0034] The splicing vector is V i V i The dimension is a reduced-dimensional vector V i p With statistical eigenvector V i s The sum of dimensions.

[0035] Preferably, the cosine similarity of the concatenated vectors in step S21 is calculated using the following formula.

[0036]

[0037] Where V k Let be the concatenated vector of the k-th appeal event, where k ≠ i; m is the number of dimensions of the concatenated vector. and These represent the concatenated vector V. i and V k The value taken in the t-th dimension;

[0038] Preferably, the setting value in step S21 is 0.9, and the predefined category obtained in step S2 is C. j , where j = 1, 2, ..., n, and n is the total number of predefined categories.

[0039] Preferably, step S3 processes the predefined category C using a large model. j The resulting summary vector is Sj , where j = 1, 2, ..., n.

[0040] Preferably, in step S4, the semantic vector V of the appeal event is calculated using the following formula. i o With the summary vector S j cosine similarity

[0041]

[0042] Among them, (V) i o ) t and Semantic vector V i o With the summary vector S j The value at the t-th dimension; w is the semantic vector V i o With the summary vector S j Dimensions.

[0043] Preferably, the actual category obtained in step S4 is CE. x x = 1, 2, ..., u, where u is the total number of actual categories.

[0044] Preferably, in step S5, the xth actual category CE x The number of claims included is satisfy Actual Category CE x For seedling categories, α and β are calculated using the following formula:

[0045] α=max(2,Ceil(N×0.5%)), β=max(10,Ceil(N×5%))

[0046] The Ceil() function represents rounding up to the nearest integer.

[0047] A computer-readable storage medium storing a computer program that causes a computer to execute any of the above-described methods for detecting emerging events based on large model summary vectors.

[0048] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the nascent event detection method based on large model summary vectors as described above.

[0049] The beneficial effects of this invention are:

[0050] (1) Utilize the generative capabilities of the large model to generate summaries in a predefined category space, and then generate summary vectors. Calculate the similarity between the semantic vector of each (people's livelihood) appeal event and all summary vectors, and logically determine the actual category to which the (people's livelihood) appeal event belongs or form a new actual category based on the similarity value.

[0051] By forming a predefined category space through an unsupervised method, and using the summary vector of the predefined category space as a ubiquitous reference frame, the similarity is calculated one by one and a threshold judgment is made. This method solves the problem that the number of cluster spaces (i.e., cluster categories) is unknown in advance.

[0052] The summary vector generated by using a large model has a stronger generalization ability than using the semantics of the original event.

[0053] (2) The semantic vector (i.e., the original vector) of the public demand event is reduced in dimension and then concatenated with the statistical feature vector (i.e. the feature vector) of the original vector to form a new semantic vector (i.e., the concatenated vector).

[0054] The concatenated vector has a stronger generalization ability than the original vector, and it helps to form a predefined category space with strong generalization ability when calculating semantic similarity.

[0055] (3) Select any concatenated vector of a public demand event and calculate its similarity value with the concatenated vectors of other public demand events. Apply a threshold judgment to this similarity value to determine whether the events should be grouped into one category. Repeat this process for the remaining public demand events.

[0056] This embodies the unsupervised thinking. Since this method does not compare any appeal events within the category for the appeal events that have already been categorized (even if two appeal events have not been compared before, they will not be compared again), compared with the pairwise comparison method for all public welfare appeal events (pairwise comparison is required regardless of whether they have been categorized, as long as two objects have not been compared), it requires less computation and has a simpler judgment logic. Moreover, since the category aggregation is performed by concatenating vectors, this helps to form a predefined category space with better generalization ability.

[0057] (4) By using the method of the present invention to determine the actual category classification, and using the business limit upper and lower limit numerical fusion method to identify the total number of public demand events, the emerging events can be identified. This is more robust and faster than the causal method, and also enables more emerging events to be recalled. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0060] Before explaining the method provided by this invention, it should be noted that the large model used in this invention is an LLM (Large Language Model): A large language model is an artificial intelligence model designed to understand and generate human language, and has approximately tens of billions of parameters. Large language models are pre-trained on a large amount of text data, and then, through supervised fine-tuning and reinforcement learning based on human feedback, they acquire capabilities such as text generation, text understanding, and interactive dialogue.

[0061] like Figure 1 As shown, this invention provides a method for detecting emerging events based on large model summary vectors, including the following steps (note that this embodiment focuses on events related to people's livelihood demands; some parts omit "people's livelihood," but all refer to events related to people's livelihood demands):

[0062] S1. Generate semantic vectors for each public demand event based on a large model. Where i represents the i-th public demand event, i = 1, 2, 3, ..., N, and N represents the total number of events in the scope of analysis; here defined... The number of dimensions is 128.

[0063] semantic vector V through UMAP i o Dimensionality reduction is performed to obtain the dimensionality-reduced vector V. i p , defined here The dimension is 32. It can also have other dimensions, but they must be less than 32. Dimensions.

[0064] UMAP (Uniform Manifold Approximation and Projection) here refers to a data dimensionality reduction technique that can be used for general nonlinear dimensionality reduction. This technique significantly improves computation speed and better preserves the global structure of the data.

[0065] Statistical feature vector V is obtained by performing statistical feature extraction on the semantic vector. i s The statistical feature vector V i s include The vector has 10 dimensions: arithmetic mean, 10th percentile, 20th percentile, 30th percentile, 40th percentile, 50th percentile, 60th percentile, 70th percentile, 80th percentile, and 90th percentile.

[0066] The dimensionality reduction vector and the statistical feature vector are concatenated to obtain the concatenated vector V. i The spliced ​​vector V i There are a total of 42 dimensions.

[0067] splicing vector V i Compared to semantic vector V i o It has a stronger generalization ability.

[0068] S2. Clustering using semantic similarity, which includes the following steps:

[0069] S21. Select any one of the all claim events (e.g., the kth claim event) and calculate the concatenation vector V of that claim event. k The concatenation vector V with all other claim events i The Euclidean distance similarity (i.e., cosine similarity) for (i≠k) is calculated using the following formula.

[0070]

[0071] Where V k Let i be the concatenated vector of the k-th claim event, i ≠ k; m is the dimension of the concatenated vector, which is m = 42 in this embodiment (it can be other values ​​depending on the situation); and These represent the concatenated vector V. i and V k The value taken in the t-th dimension.

[0072] sim(V i V k )≥0.90 as V k A predefined category C with semantic similarity k and V i and V k Classified as this predefined category C j ; All satisfying sim(V i V k V ≥ 0.90 i The corresponding appeal events are all classified into the predefined category C along with the k-th appeal event. k .

[0073] Here, 0.90 is used as a threshold (set value). The higher the threshold is adjusted, the stricter the requirements for semantic similarity. The threshold size can be adjusted according to the actual situation.

[0074] If the cosine similarity between the concatenated vectors of all claim events and the concatenated vector of the k-th claim event is less than 0.9, then the k-th claim event is assigned to a predefined category C. k (C here) k With C above k They are the same concept, both representing the number value of a predefined category to which the k-th claim event belongs.

[0075] S22. Remove the categorized claims from all claims to be analyzed, and then repeat step S21.

[0076] S23. Repeat step S22 until all claims have been categorized, ultimately forming n predefined categories C. j , where j = 1, 2, ..., n; n is the total number of predefined categories.

[0077] S3. Process all events in each predefined category obtained in step S2 using a large model to obtain a summary vector for that predefined category; specifically:

[0078] For each predefined category C j All the requests within the scope are used to generate predefined categories C using a large model. j The corresponding summary, i.e., the predefined category C j The main content of public demands incidents; generating C using a large model. j The semantic vector S of class summary j , j = 1, 2, ..., n.

[0079] S4. Calculate the semantic vector V for each claim event obtained in step S1. i o With each summary vector S obtained in step S3 j The cosine similarity is calculated using the following formula to determine the semantic vector V of the appeal event. i o With the summary vector S j cosine similarity

[0080]

[0081] Among them, (V) i o ) t and Semantic vector V i o With the summary vector S jThe value at the t-th dimension; w is the semantic vector V i o With the summary vector S j The dimension is w = 128 in this embodiment (w can be other values ​​depending on the situation).

[0082] If multiple similarity values ​​exist (Here, 0.85 is also a set value and can be changed according to actual needs), where j = 1, 2, ..., n, then... The corresponding appeal events are assigned to the predefined category with the highest similarity, and this predefined category is then classified into an independent actual category, thus generating a semantic vector. and summary vector S j If the cosine similarity is not less than 0.85, then the abstract vector S... j The corresponding predefined category is renamed (classified) into a new actual category, which is equivalent to simply changing the name of the predefined category. The reason for this change is that the predefined category is more generalized than the actual category, while the actual category is a specific classification made for all the claims to be analyzed. The two need to be distinguished.

[0083] If all Where j = 1, 2, ..., n, the event of the claim is classified into a separate actual category.

[0084] After calculating each semantic vector V i o With each summary vector S j The cosine similarity is used to classify the vectors, and the final actual vector category CE is obtained after classification based on the cosine similarity. x x = 1, 2, ..., u; u is the total number of actual categories. Note that u and n can be the same or different, depending on the actual situation.

[0085] S5. Iterate through the number of appeal events in all actual categories, specifically the x-th actual category CE. x The number of claims included is satisfy Actual Category CE x For seedling categories, α and β are calculated using the following formula:

[0086] α=max(2,Ceil(N×0.5%)), β=max(10,Ceil(N×5%))

[0087] The Ceil() function rounds up to the nearest integer; max(x,y) represents the larger of x and y; N represents the total number of events to be analyzed; and all claims included in the tipping point category are tipping point events.

[0088] The present invention also provides a computer-readable storage medium storing a computer program that causes a computer to execute the above-described method for detecting emerging events based on large model summary vectors.

[0089] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the above-described method for detecting emerging events based on large model summary vectors.

[0090] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting emerging events based on large model summary vectors, characterized in that, Includes the following steps: S1. Use a large model to generate semantic vectors for each appeal event; reduce the dimensionality of the semantic vectors to obtain a dimensionality-reduced vector; extract statistical features from the semantic vectors to obtain a statistical feature vector; and concatenate the dimensionality-reduced vector and the statistical feature vector to obtain a concatenated vector. S2. Clustering using semantic similarity, which includes the following steps: S21. Select any one appeal event from all appeal events, calculate the cosine similarity between the concatenated vector of the appeal event and all other appeal events, and classify two appeal events with a cosine similarity of not less than a first set value into the same predefined category. If no appeal event has a cosine similarity of not less than the first set value with the appeal event, then the appeal event is classified into a separate predefined category. S22. Remove the categorized events from all the claims, and then repeat step S21. S23. Repeat step S22 until all the claims have been categorized. S3. Process all events in each predefined category obtained in step S2 through a large model to obtain a summary vector for that predefined category; the dimension of the summary vector is the same as the dimension of the semantic vector; S4. Calculate the cosine similarity between the semantic vector of each appeal event obtained in step S1 and the summary vector obtained in step S3; if the cosine similarity between the semantic vector of a certain appeal event and multiple summary vectors is not less than the second set value, then classify the appeal event into the predefined category corresponding to the summary vector with the largest cosine similarity, and classify the predefined category into an independent actual category; if the cosine similarity between the semantic vector of a certain appeal event and any summary vector is less than the second set value, then classify the appeal event into an independent actual category. S5. Count the number of appeal events contained in each actual category obtained in step S4. The actual category containing the number of appeal events that meets the specific requirements is the initiation category. All appeal events contained in the initiation category are initiation events.

2. The seed event detection method based on large model summary vectors according to claim 1, characterized in that: The semantic vector generated using the large model in step S1 is , where i = 1, 2, 3, ..., N, and N is the total number of claims to be analyzed; semantic vectors via UMAP Dimensionality reduction is performed to obtain the dimensionality-reduced vector. ; The statistical feature vector of the i-th claim event is The statistical feature vector Including semantic vectors The arithmetic mean and several quantiles; The splicing vector is , The dimension is a reduced-dimensional vector With statistical eigenvectors The sum of dimensions.

3. The seed event detection method based on large model summary vectors according to claim 2, characterized in that: The cosine similarity of the concatenated vectors in step S21 is calculated using the following formula. . in Let be the concatenated vector of the k-th appeal event, where k ≠ i; m is the number of dimensions of the concatenated vector. and These represent the concatenated vectors. and In the The values ​​can be taken in each dimension.

4. The seed event detection method based on large model summary vectors according to claim 3, characterized in that: In step S21, the first set value is 0.9, and the predefined category obtained in step S2 is... ,in , This represents the total number of predefined categories.

5. The seed event detection method based on large model summary vectors according to claim 4, characterized in that: Step S3 involves processing predefined categories using a large model. The resulting summary vector is .

6. The seed event detection method based on large model summary vectors according to claim 5, characterized in that: In step S4, the semantic vector of the appeal event is calculated using the following formula. With summary vector cosine similarity in, and Semantic vectors With summary vector The value at the t-th dimension; w is the semantic vector. With summary vector Dimensions.

7. The seed event detection method based on large model summary vectors according to claim 6, characterized in that: The actual category obtained in step S4 is , , where u is the total number of actual categories.

8. The seed event detection method based on large model summary vectors according to claim 7, characterized in that: The xth actual category in step S5 The number of claims included is ,satisfy Actual category The category is the initial signs, among which and Calculated using the following formula in The function represents rounding up to the nearest integer.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the nascent event detection method based on large model summary vectors as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the nascent event detection method based on large model summary vectors as described in any one of claims 1-8.

Images