A hotspot event mining method, system, device and storage medium

Abstract

The application discloses a hotspot event mining method, system, device and storage medium, wherein the method comprises the following steps: obtaining a government work order set of an event to be mined; determining a first work order abstract of each sub government work order according to the government work order set; determining a first semantic similarity according to the first work order abstract and a sentence coding model; the first semantic similarity is used to represent the semantic similarity of any two sub government work orders corresponding to the first work order abstract; determining a first entity similarity according to the government work order set; the first entity similarity is used to represent the entity similarity of government entities corresponding to any two sub government work orders; determining a target similarity according to the first semantic similarity and the first entity similarity; performing clustering processing on all sub government work orders in the government work order set according to the target similarity to determine a hotspot event. The method can improve the accuracy of hotspot event mining. The application can be widely applied in the technical field of data processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a hot event mining method, system, device and storage medium. BACKGROUND

[0002] In recent years, with the work of government affairs services moving online, a large number of government work orders are formed in government databases and saved in the form of natural language texts. These government work orders contain a large amount of information related to hot events, so analyzing and mining them can discover various events that are currently hot.

[0003] In related technologies, the mainstream method of hot event mining is based on a clustering algorithm, that is, a pre-trained language model and the semantic similarity of government work order texts are used for clustering processing, and hot events are screened from the clustering results. However, the pre-trained language model of this method is for open fields, and it cannot effectively represent some professional concepts and vocabularies in the government field and is difficult to annotate government work orders, which leads to the inability to use downstream tasks to optimize the pre-trained language model, ultimately reducing the accuracy of hot event mining, and making this method unable to well adapt to the needs of the government professional field. Therefore, there are still technical problems to be solved in related technologies. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems existing in the prior art.

[0005] To this end, one purpose of an embodiment of the present application is to provide a hot event mining method, system, device and storage medium, which can improve the accuracy of hot event mining.

[0006] In order to achieve the above technical purpose, the technical solution adopted by the embodiments of the present application comprises: obtaining a set of government work orders of events to be mined, the set of government work orders comprising a plurality of sub-government work orders; determining a first work order abstract of each of the sub-government work orders according to the set of government work orders; determining a first semantic similarity according to the first work order abstract and a sentence encoding model; the first semantic similarity being used to represent the semantic similarity of any two sub-government work orders corresponding to the first work order abstract; determining a first entity similarity according to the set of government work orders; the first entity similarity being used to represent the entity similarity of government entities corresponding to any two sub-government work orders; determining a target similarity according to the first semantic similarity and the first entity similarity; and performing clustering processing on all of the sub-government work orders in the set of government work orders according to the target similarity to determine a hot event.

[0007] In addition, the hot event mining method according to the above-mentioned embodiments of the present application can have the following additional technical features:

[0008] Further, in the embodiments of the present application, the step of determining the first ticket summary of each of the sub-government service tickets according to the set of government service tickets specifically comprises: constructing a sentence correlation graph of the sub-government service ticket; determining the first node weight of each node corresponding to a sentence according to the sentence correlation graph; and determining the sentences corresponding to the nodes with the first node weight in the top sequence as the first ticket summary.

[0009] Further, in the embodiments of the present application, the step of determining the first semantic similarity according to the first ticket summary and the sentence encoding model specifically comprises: determining a plurality of sentence vectors according to the first ticket summary and the sentence encoding model; and determining the first semantic similarity according to any two of the plurality of sentence vectors.

[0010] Further, in the embodiments of the present application, the step of determining the first entity similarity according to the set of government service tickets specifically comprises: extracting government entities contained in all sentences of each of the sub-government service tickets of the set of government service tickets, determining a government entity list corresponding to each of the sub-government service tickets, and determining the entity similarity of any two of the sub-government service tickets according to any two of the government entity lists and a similarity calculation formula; wherein the similarity calculation formula comprises:

[0011]

[0012] wherein EL i and EL j are the government entity lists corresponding to any two of the sub-government service tickets; ∩ is an intersection operation, ∪ is a union operation, and ESim(i,j) is the first entity similarity.

[0013] Further, in the embodiments of the present application, the step of determining the target similarity according to the first semantic similarity and the first entity similarity specifically comprises: determining the target similarity according to the first semantic similarity, the first entity similarity, and a fusion formula; wherein the fusion formula comprises:

[0014] FSim(i,j) = λ × SSim(i,j) + (1-λ) × ESim(i,j)

[0015] wherein SSim(i,j) is the first semantic similarity, ESim(i,j) is the first entity similarity, FSim(i,j) is the target similarity, and 0 < λ < 1.

[0016] Further, in the embodiment of the present application, the step of constructing the sentence correlation graph of the government work order specifically includes: splitting the government work order by sentence and filtering according to the number of words to determine a sentence list; determining sentence vectors of any two sentences according to the sentence encoding model and the sentence list, and taking the cosine similarity of the two sentence vectors as the first correlation degree of the two sentences; determining the number of government entities according to the sentence list; and determining the sentence correlation graph of the government work order according to the first correlation degree and the number of government entities.

[0017] Further, in the embodiment of the present application, the step of determining the first semantic similarity according to any two sentence vectors in the plurality of sentence vectors specifically includes: extracting the sentence vectors of any two sentences as a first vector and a second vector, respectively; determining the first semantic similarity according to the first vector, the second vector, and a similarity calculation formula, the similarity calculation formula including:

[0018]

[0019] where SSim(i,j) is the first semantic similarity, h ia is the first vector, h jb is the second vector, sim(,) is a similarity operator, is a maximum value operation.

[0020] On the other hand, the embodiment of the present application also provides a hot event mining system, which includes: an acquisition unit configured to acquire a government work order set of a to-be-mined event, the government work order set including a plurality of sub-government work orders; a first processing unit configured to determine a first work order abstract of each of the sub-government work orders according to the government work order set; a second processing unit configured to determine a first semantic similarity according to the first work order abstract and a sentence encoding model, the first semantic similarity being used to represent semantic similarity of any two sub-government work orders corresponding to the first work order abstract; a third processing unit configured to determine a first entity similarity according to the government work order set, the first entity similarity being used to represent entity similarity of government entities corresponding to any two sub-government work orders; a fourth processing unit configured to determine a target similarity according to the first semantic similarity and the first entity similarity; and a fifth processing unit configured to perform clustering processing on all the sub-government work orders in the government work order set according to the target similarity to determine a hot event.

[0021] On the other hand, the present application also provides a hot event mining device, which includes:

[0022] at least one processor;

[0023] at least one memory configured to store at least one program;

[0024] When the at least one program is executed by the at least one processor, the at least one processor implements the hotspot event mining method according to any one of the summary.

[0025] In addition, the application also provides a storage medium, wherein the storage medium stores processor-executable instructions, and the processor-executable instructions are used for executing the hotspot event mining method according to any one of the above when executed by a processor.

[0026] The advantages and beneficial effects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be understood by the practice of the present application:

[0027] The application can determine the government entity similarity of any two government work orders and the semantic similarity of any two sub-government work orders in the work order abstract of the government work order according to the government work order set, and determine the hotspot event according to the semantic similarity and the entity similarity. This method combines semantic similarity and government entity similarity, enhances the clustering ability of the abstract of the government work order through government entities, can improve the adaptation ability of the model to the government field, and can improve the accuracy of hotspot event mining. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The step schematic diagram of the hotspot event mining method in one specific embodiment in the application;

[0029] Figure 2 The step schematic diagram of the step of determining the first work order abstract of each sub-government work order according to the government work order set in one specific embodiment in the application;

[0030] Figure 3 The step schematic diagram of the step of determining the first semantic similarity according to the first work order abstract and the sentence encoding model in one specific embodiment in the application;

[0031] Figure 4 The step schematic diagram of the step of determining the first entity similarity according to the government work order set in one specific embodiment in the application;

[0032] Figure 5 The step schematic diagram of the step of constructing the sentence correlation graph of the sub-government work order in one specific embodiment in the application;

[0033] Figure 6 The example diagram of the positive and negative related sentence sample pair annotation based on government entities in one specific embodiment in the application;

[0034] Figure 7 The flowchart of the hotspot event mining method in one specific embodiment in the application;

[0035] Figure 8 A flow chart of a sentence encoding model training process in one specific embodiment of the present application;

[0036] Figure 9 A structural diagram of a hot event mining system in one specific embodiment of the present application;

[0037] Figure 10 A structural diagram of a hot event mining device in one specific embodiment of the present application. DETAILED DESCRIPTION

[0038] The principles and processes of the hot event mining method, system, device and storage medium in the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0039] Referring to Figure 1 The hot event mining method of the present application comprises the following steps:

[0040] S1, obtaining a set of government work orders of events to be mined.

[0041] In this step, the set of government work orders can include a plurality of sub-government work orders, which can include a large amount of information related to hot events. In a specific application, the set of government work orders can be stored in a database, and a data processor can establish a connection with a device provided with the database, which can realize data transmission. The connection can be a wired connection or a wireless connection. The wired connection mode can include the connection between the mobile device and the host computer, and can also include the connection between the host computer and the host computer, as well as other now known or future developed devices and the host computer. The wireless connection mode can include, but is not limited to, 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (Ultra Wide Band) connection and other now known or future developed wireless connection modes.

[0042] S2, determining a first work order abstract of each of the sub-government work orders according to the set of government work orders.

[0043] In this step, the first work order abstract can include an abstract composed of a number of sentences in the front of the sentence ordering in the sub-government work order. For the sentence ordering, the processor can order according to the weight of the sentence node of the sub-government work order, or can be ordered in other ways, such as ordering according to the frequency of the keyword of the sentence of the sub-government work order or ordering the number of sentences of the sub-government work order. For example, the embodiment orders all sub-government work orders in the government work order set according to the weight of the sentence node of the sub-government work order. The processor can generate a sentence association graph corresponding to each sub-government work order, and process each sentence association graph using a specific algorithm to obtain the maximum weight of the corresponding node of each sentence in the sentence association graph. It should be noted that in some embodiments of the present application, the first work order abstract can include the top 8 sentences in a sub-government work order, or the top 5 sentences in a sub-government work order, or the top 3 sentences. The number of sentences is not limited here.

[0044] S3, determining a first semantic similarity according to the first work order abstract and a sentence coding model.

[0045] In this step, the first semantic similarity can represent the semantic similarity of any two sub-government work orders corresponding to the first work order abstract, and the sentence coding model is an artificial intelligence model that has been trained. The processor can input the first work order abstract into the trained sentence coding model to obtain the sentence vector corresponding to the work order abstract of any two sub-government work orders, and then determine the semantic similarity corresponding to any two sub-government work orders according to the two sentence vectors.

[0046] S4, determining a first entity similarity according to the government work order set.

[0047] In this step, the first entity similarity can be used to represent the entity similarity of the government entities corresponding to any two sub-government work orders. The processor extracts the government entities contained in all sentences in each government work order from the government work order set, can determine the government entity list corresponding to each government work order, and calculates the entity similarity of any two government work orders according to the government entity list corresponding to any two government work orders.

[0048] S5, determining a target similarity according to the first semantic similarity and the first entity similarity. In this step, the target similarity can be the similarity after the first semantic similarity and the first entity similarity are fused. The processor can fuse the semantic similarity and the entity similarity of any two government work orders by a formula to obtain the final target similarity.

[0049] S6, cluster all the sub-government work orders in the set of government work orders according to the target similarity, and determine a hot event.

[0050] In this step, all the sub-government work orders can be clustered according to the final similarity calculated by all the sub-government work orders, and the cluster containing the number of sub-government work orders exceeding the set threshold is taken as a hot event. It should be noted that in the embodiments, the preset threshold can be 10, 100 or other values, and the specific number is not limited herein. The clustering algorithm can use K-Means, DBSCAN or other existing clustering algorithms.

[0051] Further, with reference to Figure 2 In some embodiments of the present application, the step of determining the first work order abstract of each sub-government work order according to the set of government work orders can specifically include:

[0052] S201, constructing a sentence association graph of a sub-government work order;

[0053] S202, determining a first node weight of each node corresponding to a sentence according to the sentence association graph;

[0054] S203, determining the sentences corresponding to the nodes with the top several first node weights as the first work order abstract.

[0055] In this embodiment, a sentence association graph can be generated for each sub-government work order. The TextRank algorithm can be used to process each sentence association graph to obtain the node weights of all nodes in each sentence association graph. The sentence association graph can be a graph in which each sentence is a node, and an edge is created between two nodes if the association degree of the sentences is greater than a specified threshold. For any government work order, the node weights of all nodes in the sentence association graph of the government work order are sorted in descending order, and the sentences represented by the nodes with the top several weights are obtained as the abstract of the current government work order. The several weights are a positive integer number of weights.

[0056] Further, with reference to Figure 3 In some embodiments of the present application, the step of determining the first semantic similarity according to the first work order abstract and the sentence encoding model can specifically include:

[0057] S301, determining a plurality of sentence vectors according to the first work order abstract and the sentence encoding model;

[0058] S302, determining the first semantic similarity according to any two sentence vectors in the plurality of sentence vectors.

[0059] In the embodiment, the first work order summary can be input into the trained sentence encoding model to obtain the same number of sentence vectors as the number of sentences of the first work order summary. Any two sentence vectors extracted from the same number of sentence vectors as the number of sentences of the first work order summary can obtain the semantic similarity of any two sub-government work orders corresponding to the first work order summary. As for the training of the sentence encoding model, the embodiment can first construct a government text corpus and a government entity knowledge base; and train the constructed sentence encoding model using the government text corpus to obtain the trained sentence encoding model; and then optimize the trained sentence encoding model using the government entity knowledge base and the government text corpus to obtain the optimized sentence encoding model, wherein the optimized sentence encoding model is the trained sentence encoding model.

[0060] Further, as for the training of the sentence encoding model, the embodiment can select a target sentence sample from the government text corpus, and randomly select another several sentence samples from the government text corpus to form a sample batch with the target sentence sample; send the target sentence sample into a sentence encoding model with a random dropout mechanism twice to obtain two different encoding vectors; send the other sentence samples in the sample batch into the sentence encoding model with the random dropout mechanism to obtain the encoding vector corresponding to each sentence sample; calculate the loss function according to the two different encoding vectors and the encoding vector corresponding to each sentence sample; the calculation formula is as follows:

[0061]

[0062] wherein, is the loss function, h k and h k(*) are different encoding vectors, sim() is a cosine similarity function, τ is a hyperparameter, x a is the a-th sentence in the sample batch, h a is the encoding vector corresponding to x a ; the above loss function is used for back propagation of the sentence encoding model to update the parameters in the sentence encoding model; the above steps of selecting a target sentence sample from the government text corpus and randomly selecting another several sentence samples from the government text corpus to form a sample batch with the target sentence sample are repeated until each sentence sample in the government text corpus is processed once, and finally the optimized sentence encoding model is obtained. It should be noted that the sentence encoding model of the embodiment can use the BERT model.

[0063] Further, the construction of the government entity knowledge base can include the following steps: comprehensively referring to the existing government term library, department list, current hot events and / or policy list, and constructing the government entity knowledge base through manual review; wherein each government entity is represented as a list of alias keywords.

[0064] Further, for the training of the sentence encoding model, the embodiment can also use alias matching to extract the list of government entities contained in each sentence sample in the government text corpus, and label the sentence samples containing t or more same government entities as positive correlation; t is a positive integer; select a target sentence sample from the government text corpus, first randomly select at most N / 2-1 sentence samples positively correlated with the target sentence sample from the government text corpus, denoted as the positive sample set; then randomly select N-|P(k)| sentence samples non-positively correlated with the target sentence sample from the government text corpus, and combine the selected non-positive sentence samples, the target sentence sample and the positive sample set to form a sample batch; send each sentence sample in the sample batch to the trained sentence encoding model to obtain the sentence vector corresponding to each sentence; then calculate the training loss, and the calculation formula is:

[0065]

[0066] wherein, is the training loss value, x p is the pth sentence sample in the positive sample set P(k); h p is the sentence vector corresponding to x p ; h k is the sentence vector corresponding to the target sentence sample; x a is the ath sentence sample in the sample batch; h a is the sentence vector corresponding to x a ; x p , x a are two different target sentence samples; use the training loss to perform back propagation on the trained sentence encoding model to update the parameters in the trained sentence encoding model; repeat the use of alias matching to extract the list of government entities contained in each sentence sample in the government text corpus, and label the sentence samples containing t or more same government entities as positive correlation; t is a positive integer, until each sentence sample in the government text corpus has been processed once, and finally obtain the optimized sentence encoding model.

[0067] Further, with reference to Figure 4 In some embodiments of the present application, according to the government work order set, the step of determining the first entity similarity can specifically include:

[0068] S401, extract all government entities contained in all sentences in each sub-government work order of the set of government work orders, and determine a government entity list corresponding to each sub-government work order;

[0069] S402, determine entity similarity of any two sub-government work orders according to any two government entity lists and a similarity calculation formula; wherein the similarity calculation formula comprises:

[0070]

[0071] wherein, EL i and EL j are respectively any two government entity lists corresponding to the sub-government work orders; ∩ is intersection operation, ∪ is union operation, and ESim(i,j) is the first entity similarity.

[0072] In this embodiment, the government entities contained in all sentences in each sub-government work order of the set of government work orders can be extracted, and a government entity list corresponding to each sub-government work order can be determined, which can include multiple government entities. According to any two government entity lists in the government entity list corresponding to each sub-government work order and a similarity calculation formula, the entity similarity of any two sub-government work orders can be determined; wherein the similarity calculation formula can comprise:

[0073]

[0074] wherein, EL i and EL j are respectively any two government entity lists corresponding to the sub-government work orders; ∩ is intersection operation, ∪ is union operation, and ESim(i,j) is the first entity similarity.

[0075] Further, in some embodiments of the present application, according to the first semantic similarity and the first entity similarity, the step of determining the target similarity can specifically comprise:

[0076] determining the target similarity according to the first semantic similarity, the first entity similarity and a fusion formula; wherein the fusion formula comprises:

[0077] FSim(i,j) = λ × SSim(i,j) + (1-λ) × ESim(i,j)

[0078] wherein, SSim(i,j) is the first semantic similarity, ESim(i,j) is the first entity similarity, FSim(i,j) is the target similarity, and 0 < λ < 1.

[0079] In the embodiment, after the first semantic similarity and the first entity similarity are obtained through the above embodiment, the first semantic similarity and the first entity similarity can be fused through a fusion formula, wherein the fusion formula includes:

[0080] FSim(i,j) = λ × SSim(i,j) + (1-λ) × ESim(i,j)

[0081] wherein SSim(i,j) is the first semantic similarity, ESim(i,j) is the first entity similarity, FSim(i,j) is the target similarity, and 0<λ<1.

[0082] Further, with reference to Figure 5 In some embodiments of the present application, the step of constructing a sentence association graph of a government work order can specifically include:

[0083] S601, splitting the government work order by sentence and filtering according to the number of words to determine a sentence list;

[0084] S602, determining the sentence vectors of any two sentences according to the sentence encoding model and the sentence list, and taking the cosine similarity of the any two sentence vectors as the first correlation degree of the any two sentences;

[0085] S603, determining the number of government entities according to the sentence list;

[0086] S604, determining the sentence association graph of the government work order according to the first correlation degree and the number of government entities;

[0087] In the embodiment, the government work order can be split by sentence and the sentences with a word number less than a preset threshold can be filtered to obtain a sentence list. The obtained sentence list can be input into a trained sentence encoding model to determine the sentence vectors of any two sentences. The cosine similarity corresponding to the any two sentence vectors can be taken as the first correlation degree of the any two sentences. The sentence list can also determine the number of government entities. Finally, the first correlation degree and the number of government entities can be used to determine the sentence association graph of the government work order. It should be noted that the preset threshold of the embodiment can be one or more, and the specific number is not limited. The trained sentence encoding model can be the same as the sentence encoding model used in the process of determining the first semantic similarity, or it can be other existing artificial intelligence models.

[0088] Further, in some embodiments of the present application, the step of determining the first semantic similarity according to any two sentence vectors in the plurality of sentence vectors can specifically include:

[0089] S701, extract the sentence vectors of any two sentences as a first vector and a second vector respectively;

[0090] S702, determine a first semantic similarity according to the first vector, the second vector, and a similarity calculation formula, the similarity calculation formula comprising:

[0091]

[0092] wherein SSim(i,j) is the first semantic similarity, h ia is the first vector, h jb is the second vector, and sim(,) is a similarity operator, is a maximum value operation.

[0093] In this embodiment, the sentence vectors of any two sentences can be extracted as a first vector and a second vector respectively. The first vector and the second vector are input into a similarity calculation formula, and a first semantic similarity can be determined, the similarity calculation formula comprising:

[0094]

[0095] wherein in the similarity calculation formula, SSim(i,j) is the first semantic similarity, h ia is the first vector, h jb is the second vector, and sim(,) is a similarity operator, is a maximum value operation.

[0096] The hot event mining method of the present application will be further described below in combination with Figure 6 and Figure 7 .

[0097] In this embodiment, the sentence encoding model is a BERT model, and the weight ranking is the top 3 ranking.

[0098] Firstly, a government entity knowledge base ES and a sentence encoding model can be constructed. The government entity knowledge base ES can be constructed by manual review according to an existing government term library, a department list, a current hot event, and / or a policy list. Each government entity is represented as a list of alias keywords; for example, government entity 1: {certificate handling, marriage certificate}, government entity 2: {tax processing, tax rate}, government entity 3: {living expenses, water fee}, and so on.

[0099] Secondly, the constructed sentence encoding model is trained and optimized. The government text corpus TC can be used to train the BERT model to obtain the trained sentence encoding model; referring to Figure 7 , the specific training steps of the sentence encoding model are:

[0100] Step 1: For each sentence sample x in the government affairs text corpus TC i Extracting x based on alias matching i The list of government entities included. Figure 6 In the examples shown, sentence sample (1) contains the government entity {Guangzhou, tax rate}, sentence sample (2) contains the government entity {Guangzhou, individual income tax}, and sentence sample (3) contains the government entity {housing provident fund, housing loan}. Here, “tax rate” and “individual income tax” are aliases for the same government entity.

[0101] Step 2: Given a pair of sentence samples x i and x j If x i and x j If there are more than one identical government entity, then x will be... i and x j Marked as positively correlated. If x i and x j If it does not contain any identical government entities, then x will be... i and x j Marked as negative correlation. Figure 3 In the example shown, sentence sample (1) and sentence sample (2) contain two identical government entities and are therefore labeled as positively correlated.

[0102] Step 3, refer to Figure 8 Select a target sentence sample x from the government affairs text corpus TC. k First, at most N / 2-1 texts related to x are randomly selected from the government affairs text corpus TC. k Positively correlated sentence samples (denoted as P(k)); then N-|P(k)| sentences related to x are randomly selected from the government text corpus TC. k Negatively correlated sentence samples P(-k), P(-k) and sentence sample x k The positively correlated sentence samples P(k) together form a sample batch BS(k).

[0103] Step 4: Extract each sentence sample x from the sample batch BS(k). i The sample sentence x is fed twice into the sentence encoding model Enc, which has a random dropout mechanism, to obtain the target sentence sample. k The corresponding encoding vector h k The encoding vector h corresponding to the positively correlated sentence samples p and the encoding vector h corresponding to the sentence samples of the sample batch BS(k). a According to the encoding vector h k h pand the encoding vector h corresponding to the sentence sample of the sample batch BS(k) a The loss function is calculated; the calculation formula is:

[0104]

[0105] Wherein, sim() is the cosine similarity function, τ is the hyperparameter, x a is the a-th sentence in BU(k), h a is the corresponding encoding vector, h k is the target sentence sample x k corresponding encoding vector, h p is the encoding vector corresponding to the positively correlated sentence sample;

[0106] Step five, use the loss function to back-propagate the sentence encoding model, update the parameters in the sentence encoding model Enc;

[0107] Step six, return to step one until each sentence sample in the government text corpus TC is processed once, and finally obtain the optimized sentence encoding model; in this embodiment, the sample batch is composed of positively correlated and non-positively correlated sentence samples, and the trained sentence encoding model is optimized, so the method is supervised contrast learning;

[0108] Then, after the model training is completed, all government work orders meeting the conditions can be retrieved from the database according to the time and space range of the to-be-mined hot event; the above-mentioned space range can be the location range of the to-be-mined hot event or other to-be-constrained ranges; a sentence association graph corresponding to each government work order meeting the conditions retrieved from the database is generated, and the TextRank algorithm is used to process each sentence association graph to obtain the final weight of all nodes in each sentence association graph; wherein, in the sentence association graph, each sentence is a node, and if the association degree between two nodes is greater than a specified threshold, an edge is created; the specific steps of generating a sentence association graph corresponding to any government work order Ok are as follows:

[0109] Step 1-1, split the government work order Ok by sentence, and filter out sentences with a word count less than a preset threshold to obtain a sentence list OLk.

[0110] Step 1-2, for any two sentences x i and x j in OLk, use the optimized sentence encoding model in step 3 to convert x i and x j into sentence vectors h i and h j , respectively, and calculate h i and h jcosine similarity between x i and x j .

[0111] Step 1-3, for each sentence x i in OLk, calculate the number of government entities contained in x i , denoted as num(i).

[0112] Step 1-4, construct a sentence association graph Gk for government work order Ok, each node v i in Gk represents a sentence; if the association degree between sentence x i and x j is greater than a specified threshold, there is an edge e i between nodes v j and v ij .

[0113] Step 1-5, for each node v i in Gk, if num(i) > 0, set the initial weight w i of v i to num(i), otherwise uniformly set it to a fixed value c; and the weight w ij of edge e ij is set to the association degree between sentence x i and x j .

[0114] In addition, after completing the sentence association graph, for any government work order, sort the maximum weights of all nodes in the sentence association graph of the government work order from large to small, and obtain the sentences represented by the top 3 weight nodes as the summary of the current government work order. After obtaining the summaries of the 3 current government work orders, use the optimized sentence encoding model to convert the 3 sentences corresponding to the summary of each government work order in step 6 into 3 sentence vectors, and calculate the semantic similarity between any two government work orders according to the sentence vectors corresponding to the two government work orders; in this embodiment, the formula for calculating the semantic similarity is:

[0115]

[0116] where SSim(i,j) is the semantic similarity, h ia is the first vector, h jb is the second vector, sim(,) is the similarity operator, is the maximum value operation.

[0117] Further, after the semantic similarity is calculated, all government entities contained in all sentences of each government work order meeting the condition retrieved from the database can be extracted to form a government entity list corresponding to each government work order, and the entity similarity of any two government work orders is calculated according to the government entity lists corresponding to the two government work orders.

[0118] In this embodiment, the entity similarity is the Jaccard coefficient between and, and the calculation formula is:

[0119]

[0120] wherein EL i and EL j are the government entity lists corresponding to any two sub-government work orders; ∩ is the intersection operation, ∪ is the union operation, and ESim(i,j) is the first entity similarity.

[0121] After the semantic similarity and the entity similarity are obtained, the semantic similarity and the entity similarity of any two government work orders can be fused to obtain the final similarity. In this embodiment, the calculation formula of the final similarity is:

[0122] FSim(i,j) = λ × SSim(i,j) + (1-λ) × ESim(i,j)

[0123] wherein SSim(i,j) is the first semantic similarity, ESim(i,j) is the first entity similarity, FSim(i,j) is the target similarity, and 0 < λ < 1.

[0124] Finally, according to the final similarity calculated for all government work orders, K-Means algorithm is used to perform clustering processing on all government work orders, and the cluster containing the number of government work orders exceeding the set threshold is taken as a hot event.

[0125] To sum up, the government entity is used to assist the contrast learning model to label the positive sample pair, so as to realize the supervised contrast learning without manual labeling. The government entity knowledge base is used to enhance the ability of summarizing and clustering the government work orders, and the adaptation ability of the model to the government field is improved.

[0126] Further, referring to Figure 9 , and Figure 1Corresponding to the method, the embodiments of the present application also provide a hotspot event mining system, which comprises: an acquisition unit 101 configured to acquire a set of government work orders of events to be mined; a first processing unit 102 configured to determine a first work order abstract of each of the sub government work orders according to the set of government work orders; a second processing unit 103 configured to determine a first semantic similarity according to the first work order abstract and a sentence coding model; the first semantic similarity is used to represent semantic similarity of any two sub government work orders corresponding to the first work order abstract; a third processing unit 104 configured to determine a first entity similarity according to the set of government work orders; the first entity similarity is used to represent entity similarity of government entities corresponding to any two sub government work orders; a fourth processing unit 105 configured to determine a target similarity according to the first semantic similarity and the first entity similarity; and a fifth processing unit 106 configured to perform clustering processing on all the sub government work orders in the set of government work orders according to the target similarity to determine a hotspot event.

[0127] In some embodiments of the present application, the acquisition module 101, the first processing unit 102, the second processing unit 103, the third processing unit 104, the fourth processing unit 105 and the fifth processing unit 106 can be arranged in the same host computer, and the set of government work orders of events to be mined in the database is acquired by the modules in the host computer, and then the acquired data is processed by the subsequent processors in the host computer. In some other embodiments of the present application, the acquisition module 101 can also be any module connected with the host computer, and the acquired data is transmitted to the host computer through wired or wireless connection with the host computer, and the data is processed by the modules in the host computer. In addition, in some embodiments, the first processing unit 102, the second processing unit 103, the third processing unit 104, the fourth processing unit 105 and the fifth processing unit 106 are the same, and the specific device connection mode and device arrangement are not limited.

[0128] It should be understood that the contents in the above-mentioned hotspot event mining method embodiments are applicable to the present hotspot event mining system embodiments, the functions realized by the present hotspot event mining system embodiments are the same as those of the above-mentioned hotspot event mining method embodiments, and the beneficial effects achieved by the present hotspot event mining system embodiments are also the same as those achieved by the above-mentioned hotspot event mining method embodiments.

[0129] Corresponding to the method, Figure 1 the embodiments of the present application also provide a hotspot event mining device, the specific structure of which can refer to Figure 10 , comprising:

[0130] at least one processor 1001;

[0131] at least one memory 1002, configured to store at least one program;

[0132] When the at least one program is executed by the at least one processor, the at least one processor implements the hotspot event mining method.

[0133] The contents in the above method embodiments are all applicable to the device embodiments, the device embodiments specifically implement the functions same as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0134] Corresponding to the method of Figure 1 Corresponding to the method of

[0135] The contents in the above method embodiments are all applicable to the storage medium embodiments, the storage medium embodiments specifically implement the functions same as the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0136] In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously with each other, or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example, and the purpose is to provide a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are independently executed.

[0137] Furthermore, although the present application is described in the context of functional modules, it is understood that one or more of the functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary to an understanding of the present application. Rather, the actual implementation is within the routine skill of engineers familiar with the property, function and internal relationships of the various functional modules disclosed herein. Accordingly, the present application is not limited to the specific details of the functional modules described herein. Rather, it is understood that one of ordinary skill in the art is able to practice the application as claimed without undue experimentation having regard to the property, function and internal relationships of the various functional modules disclosed herein. It is also understood that the specific concepts disclosed are merely illustrative and that the scope of the present application is determined by the appended claims and their equivalents.

[0138] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, and includes a plurality of programs used to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0139] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable program instructions for implementing logical functions, and can be specifically embodied in any computer readable medium for program execution system, device or equipment (such as computer-based system, system including processor or other system that can take program from program execution system, device or equipment and execute program) to use, or in conjunction with these program execution system, device or equipment. For the purpose of the present specification, "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for program execution system, device or equipment or in conjunction with these program execution system, device or equipment.

[0140] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0141] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above described embodiments, multiple steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable program executing system. For example, if implemented in hardware, and as in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0142] In the above description of the present specification, the description referring to the terms "one embodiment", "another embodiment", or "certain embodiments" or the like means that a specific feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. The illustrative expressions of the above terms do not necessarily refer to the same embodiment or example in the present specification. Also, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in an appropriate manner.

[0143] Although the embodiments of the present application have been shown and described, it would be appreciated by those skilled in the art that changes, modifications, alternatives and variations to these embodiments can be made without departing from the principles and spirit of the application, the scope of which is defined by the claims and their equivalents.

[0144] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and these equivalent modifications or substitutions are included in the scope defined by the claims of the present application.

Claims

1. A method for mining trending events, characterized in that, Includes the following steps: Obtain the set of government service orders for the events to be mined; the set of government service orders includes several sub-government service orders; Based on the set of government service work orders, determine the first work order summary for each of the sub-government service work orders; Based on the first work order summary and the sentence encoding model, a first semantic similarity is determined; the first semantic similarity is used to characterize the semantic similarity between any two sub-government work orders corresponding to the first work order summary. Based on the set of government service orders, a first entity similarity is determined; the first entity similarity is used to characterize the entity similarity between government entities corresponding to any two sub-government service orders. The target similarity is determined based on the first semantic similarity and the first entity similarity. Based on the target similarity, all the sub-government work orders in the government work order set are clustered to identify hot events; The step of determining the first semantic similarity based on the first work order summary and the sentence encoding model specifically includes: Based on the first work order summary and the sentence encoding model, several sentence vectors are determined; The first semantic similarity is determined based on any two sentence vectors from the given sentence vectors.

2. The hot topic event mining method according to claim 1, characterized in that, The step of determining the first work order summary of each sub-government work order based on the government work order set specifically includes: Construct a sentence association graph for the aforementioned sub-government work orders; Based on the sentence association graph, determine the weight of the first node corresponding to each sentence; The sentences corresponding to the top few nodes in the weight ranking of the first node are determined as the first work order summary.

3. The hot topic event mining method according to claim 1, characterized in that, The step of determining the first entity similarity based on the set of government service orders specifically includes: Extract all government entities contained in sentences in each sub-government work order set, and determine the list of government entities corresponding to each sub-government work order; Based on any two of the aforementioned government entity lists and the similarity calculation formula, determine the entity similarity between any two sub-government work orders; wherein the similarity calculation formula includes: in, and These are lists of government entities corresponding to any two sub-government work orders; For intersection operations, For the union operation, This represents the first entity similarity.

4. The hot topic event mining method according to claim 1, characterized in that, The step of determining the target similarity based on the first semantic similarity and the first entity similarity specifically includes: The target similarity is determined based on the first semantic similarity, the first entity similarity, and the fusion formula; wherein the fusion formula includes: in, The first semantic similarity, The first entity similarity, For target similarity, 0 < <1.

5. The hot topic event mining method according to claim 2, characterized in that, The step of constructing the sentence association graph of the sub-government work order specifically includes: The government work orders are split into sentences and filtered according to the number of characters to determine a list of sentences; Based on the sentence encoding model and the sentence list, determine the sentence vectors of any two sentences, and use the cosine similarity of the sentence vectors of any two sentences as the first correlation between the two sentences. The number of government entities is determined based on the list of sentences; Based on the first degree of relevance and the number of government entities, a sentence association graph for government work orders is determined.

6. The hot topic event mining method according to claim 1, characterized in that, The step of determining the first semantic similarity based on any two sentence vectors from the plurality of sentence vectors specifically includes: Extract the sentence vectors of any two sentences and use them as the first vector and the second vector, respectively; Based on the first vector, the second vector, and the similarity calculation formula, a first semantic similarity is determined, wherein the similarity calculation formula includes: in The first semantic similarity, Let be the first vector. For the second vector, For similarity operators, This is for calculating the maximum value.

7. A hot topic event mining system, characterized in that, include: The acquisition unit is used to acquire a set of government work orders for the events to be mined, wherein the set of government work orders includes several sub-government work orders; The first processing unit is used to determine the first work order summary of each of the sub-government work orders based on the government work order set; The second processing unit is used to determine a first semantic similarity based on the first work order summary and the sentence encoding model; the first semantic similarity is used to characterize the semantic similarity between any two sub-government work orders corresponding to the first work order summary. The third processing unit is used to determine the first entity similarity based on the set of government service orders; the first entity similarity is used to characterize the entity similarity between government entities corresponding to any two sub-government service orders. The fourth processing unit is used to determine the target similarity based on the first semantic similarity and the first entity similarity; The fifth processing unit is used to perform clustering processing on all the sub-government work orders in the government work order set according to the target similarity to determine hot events; The step of determining the first semantic similarity based on the first work order summary and the sentence encoding model specifically includes: Based on the first work order summary and the sentence encoding model, several sentence vectors are determined; The first semantic similarity is determined based on any two sentence vectors from the given sentence vectors.

8. A hot topic event mining device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a hotspot event mining method as described in any one of claims 1-6.

9. A computer-readable storage medium storing processor-executable instructions, characterized in that, The processor-executable instructions, when executed by the processor, are used to perform a hotspot event mining method as described in any one of claims 1-6.

Images