CN115983255B

Movatterモバイル変換

Info

Publication number: CN115983255B
Application number: CN202310275728.4A
Authority: CN
Inventors: 苏红梅; 袁戟; 陈嘉维; 郑泽涛; 黄甜甜
Original assignee: Shenzhen Wanwuyun Technology Co ltd
Current assignee: Shenzhen Wanwuyun Technology Co ltd; Shenzhen Vanke Property Service Co Ltd
Priority date: 2023-03-21
Filing date: 2023-03-21
Publication date: 2023-06-02
Anticipated expiration: 2043-03-21
Also published as: CN115983255A

Abstract

The invention discloses a sudden event management method, a device, computer equipment and a storage medium, wherein the method comprises the following steps: acquiring emergency information, and preprocessing the emergency information to obtain a model data set; extracting keywords from the model data set based on a preset classification range, and performing regular matching processing on the extracted keywords to obtain a first classification result; inputting the model data set into a multi-task learning model, and outputting a second classification result from the multi-task learning model based on a preset classification range; and voting integration is carried out on the first classification result and the second classification result, and the voting integration result is output as a final classification result. According to the invention, the emergency information is respectively classified through the regular matching and the multi-task learning model, and the classification results are voting integrated to obtain the final classification result, so that the emergency of different types and grades can be more accurately classified.

Description

Translated fromChinese

一种突发事件管理方法、装置、计算机设备及存储介质Emergency management method, device, computer equipment and storage medium

技术领域technical field

本发明涉及计算机软件技术领域，特别涉及一种突发事件管理方法、装置、计算机设备及存储介质。The present invention relates to the technical field of computer software, in particular to an emergency management method, device, computer equipment and storage medium.

背景技术Background technique

突发事件的类型很多，如果人工对突发事件信息进行分类分级，需要大量的人力的同时还可能会存在较多的误判，且响应效率满足不了需求。以历史数据作为标签，通过自然语言理解的方式对物业报警的突发事件文本信息进行分类，例如将突发事件分类等级划分为一级12大级，二级66小级，紧急程度3大级，可以提高分类分级的速度和响应效率，同时通过对历史数据的学习，能够不断优化模型，提高突发事件分类分级的准确率。There are many types of emergencies. If the emergency information is manually classified and graded, a large amount of manpower is required and there may be more misjudgments, and the response efficiency cannot meet the demand. Use historical data as tags to classify the emergency text information of property alarms through natural language understanding. For example, the classification of emergencies is divided into 12 major levels for the first level, 66 minor levels for the second level, and 3 major levels of urgency , can improve the speed and response efficiency of classification and grading, and at the same time, through the study of historical data, the model can be continuously optimized to improve the accuracy of emergency classification and grading.

现有技术对于突发事件的分类分级，都存在着各自的不足。例如基于正则、传统机器学习的自然语言的泛化能力不足，基于深度学习的自然语言进行标注需要的人力成本较大等，因此需要一种兼备泛化能力和标注效率的突发事件管理方法。The prior art has its own deficiencies in the classification and grading of emergencies. For example, the generalization ability of natural language based on regularization and traditional machine learning is insufficient, and the human cost of labeling natural language based on deep learning is relatively large. Therefore, an emergency management method with both generalization ability and labeling efficiency is needed.

发明内容Contents of the invention

本发明实施例提供了一种突发事件管理方法、装置、计算机设备及存储介质，旨在提高对突发事件的管理效率。Embodiments of the present invention provide an emergency management method, device, computer equipment, and storage medium, aiming at improving the management efficiency of emergency.

第一方面，本发明实施例提供了一种突发事件管理方法，包括：In a first aspect, an embodiment of the present invention provides an emergency management method, including:

获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集；Obtain emergency event information, and preprocess the emergency event information to obtain a model data set;

基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果；performing keyword extraction on the model data set based on a preset classification range, and performing regular matching processing on the extracted keywords to obtain a first classification result;

将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果；Inputting the model data set into a multi-task learning model, and outputting the multi-task learning model based on a preset classification range to obtain a second classification result;

对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出。Vote integration is performed on the first classification result and the second classification result, and the vote integration result is output as a final classification result.

第二方面，本发明实施例提供了一种突发事件管理装置，包括：In a second aspect, an embodiment of the present invention provides an emergency management device, including:

数据处理单元，用于获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集；A data processing unit, configured to obtain emergency event information, and preprocess the emergency event information to obtain a model data set;

第一分类单元，用于基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果；The first classification unit is configured to extract keywords from the model data set based on a preset classification range, and perform regular matching processing on the extracted keywords to obtain a first classification result;

第二分类单元，用于将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果；A second classification unit, configured to input the model data set into a multi-task learning model, and output the multi-task learning model based on a preset classification range to obtain a second classification result;

投票集成单元，用于对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出。A vote integration unit, configured to perform vote integration on the first classification result and the second classification result, and output the vote integration result as a final classification result.

第三方面，本发明实施例提供了一种计算机设备，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如第一方面所述的一种突发事件管理方法。In a third aspect, an embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and operable on the processor. When the processor executes the computer program, A method for managing emergencies as described in the first aspect is realized.

第四方面，本发明实施例提供了一种计算机可读存储介质，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现如第一方面所述的一种突发事件管理方法。In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the one described in the first aspect is implemented. Emergency management approach.

本发明实施例提供了一种突发事件管理方法、装置、计算机设备及存储介质，该方法包括：获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集；基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果；将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果；对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出。与现有技术相比，本发明实施例通过正则匹配和多任务学习模型分别对突发事件信息进行分类，并将分类结果进行投票集成，得到最终分类结果，不仅能够提高突发事件管理方法的泛化能力，而且还能对不同类型和等级的突发事件更准确地进行分类。An embodiment of the present invention provides an emergency management method, device, computer equipment, and storage medium. The method includes: acquiring emergency information, and preprocessing the emergency information to obtain a model data set; based on Extract keywords from the model data set within the preset classification range, and perform regular matching processing on the extracted keywords to obtain the first classification result; input the model data set into a multi-task learning model, and classify based on the preset The second classification result is obtained from the output of the multi-task learning model; vote integration is performed on the first classification result and the second classification result, and the vote integration result is output as a final classification result. Compared with the prior art, the embodiment of the present invention classifies emergency event information through regular matching and multi-task learning models, and votes and integrates the classification results to obtain the final classification result, which can not only improve the efficiency of the emergency management method Generalization ability, but also more accurate classification of different types and grades of emergencies.

附图说明Description of drawings

为了更清楚地说明本发明实施例技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. Ordinary technicians can also obtain other drawings based on these drawings on the premise of not paying creative work.

图1为本发明实施例提供的一种突发事件管理方法的流程示意图；Fig. 1 is a schematic flow chart of an emergency management method provided by an embodiment of the present invention;

图2为本发明实施例提供的一种突发事件管理方法的子流程示意图；FIG. 2 is a schematic subflow diagram of an emergency management method provided by an embodiment of the present invention;

图3为本发明实施例提供的一种突发事件管理方法另一子流程示意图；FIG. 3 is a schematic diagram of another sub-flow of an emergency management method provided by an embodiment of the present invention;

图4为本发明实施例提供的一种突发事件管理装置的示意性框图；FIG. 4 is a schematic block diagram of an emergency management device provided by an embodiment of the present invention;

图5为本发明实施例提供的一种突发事件管理装置的子示意性框图；FIG. 5 is a sub-schematic block diagram of an emergency management device provided by an embodiment of the present invention;

图6为本发明实施例提供的一种突发事件管理装置的另一子示意性框图。Fig. 6 is another sub-schematic block diagram of an emergency management device provided by an embodiment of the present invention.

具体实施方式Detailed ways

下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

应当理解，当在本说明书和所附权利要求书中使用时，术语“包括”和 “包含”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。It should be understood that when used in this specification and the appended claims, the terms "comprising" and "comprises" indicate the presence of described features, integers, steps, operations, elements and/or components, but do not exclude one or Presence or addition of multiple other features, integers, steps, operations, elements, components and/or collections thereof.

还应当理解，在此本发明说明书中所使用的术语仅仅是出于描述特定实施例的目的而并不意在限制本发明。如在本发明说明书和所附权利要求书中所使用的那样，除非上下文清楚地指明其它情况，否则单数形式的“一”、“一个”及“该”意在包括复数形式。It should also be understood that the terminology used in the description of the present invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the present invention. As used in this specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural referents unless the context clearly dictates otherwise.

还应当进一步理解，在本发明说明书和所附权利要求书中使用的术语“和/ 或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。It should also be further understood that the term "and/or" used in the description of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations .

S101、获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集；S101. Obtain emergency event information, and preprocess the emergency event information to obtain a model data set;

S102、基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果；S102. Extract keywords from the model data set based on a preset classification range, and perform regular matching processing on the extracted keywords to obtain a first classification result;

S103、将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果；S103. Input the model data set into a multi-task learning model, and output the multi-task learning model based on a preset classification range to obtain a second classification result;

S104、对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出。S104. Perform vote integration on the first classification result and the second classification result, and output the vote integration result as a final classification result.

本实施例中，首先根据突发事件信息得到模型数据集，其次对所述模型数据集提取关键词，对所述关键词进行正则匹配，并输出第一分类结果，然后将所述模型数据集输入所述多任务学习模型，通过所述多任务学习模型输出第二分类结果，最后将所述第一分类结果与所述第二分类结果进行投票集成，得到最终分类结果，通过所述最终分类结果对突发事件进行管理。本发明实施例通过正则匹配和多任务学习模型分别对突发事件信息进行分类，并将分类结果进行投票集成，得到最终分类结果，不仅能够提高突发事件管理方法的泛化能力，而且能对不同类型和等级的突发事件更准确地进行分类，便于针对突发事件的不同分类和等级进行快速响应，从而实现对突发事件的安全管理。In this embodiment, firstly, the model data set is obtained according to the emergency event information, and secondly, keywords are extracted from the model data set, regular matching is performed on the keywords, and the first classification result is output, and then the model data set is Input the multi-task learning model, output the second classification result through the multi-task learning model, and finally vote and integrate the first classification result and the second classification result to obtain the final classification result, and pass the final classification As a result, emergencies are managed. The embodiment of the present invention classifies emergency event information respectively through regular matching and multi-task learning models, and votes and integrates the classification results to obtain the final classification result, which can not only improve the generalization ability of the emergency event management method, but also Different types and grades of emergencies can be classified more accurately, which is convenient for rapid response to different classifications and grades of emergencies, so as to realize the safety management of emergencies.

在一实施例中，所述突发事件信息包括事件时间，事件主题，事件主要内容；所述步骤S101包括：In one embodiment, the emergency event information includes event time, event subject, and event main content; the step S101 includes:

获取突发事件信息对应的事件主表ID，根据所述事件主表ID关联事件类型，并根据所述事件类型对所述突发事件信息进行特征工程处理；Obtaining the event master table ID corresponding to the emergency event information, associating event types according to the event master table ID, and performing feature engineering processing on the emergency event information according to the event type;

对所述突发事件信息进行文本清洗，以及基于特征工程处理的结果对所述突发事件信息进行数据转换，以此构建模型数据集。Perform text cleaning on the emergency event information, and perform data conversion on the emergency event information based on the result of feature engineering processing, so as to construct a model data set.

在本实施例中，根据突发事件信息的事件主表ID关联事件类型，并对所述突发事件信息进行特征工程处理、文本清洗和数据转换，以此构建模型数据集。进一步的，还可将所述模型数据集划分为训练集、验证集和测试集。其中训练集用于更新输入的参数对模型进行训练；所述验证集用于评估多种不同模型或者是带着不同参数的同一模型的训练效果；测试集用于评估模型的泛化能力。In this embodiment, event types are associated according to the event master table ID of emergency event information, and feature engineering processing, text cleaning, and data conversion are performed on the emergency event information to construct a model data set. Further, the model data set can also be divided into training set, verification set and test set. The training set is used to update the input parameters to train the model; the verification set is used to evaluate the training effect of a variety of different models or the same model with different parameters; the test set is used to evaluate the generalization ability of the model.

在具体实施例中，所述特征工程处理以所述模型数据集中的突发事件的主题、详细信息和主要内容为特征，以突发事件的分类和风险等级作为目标值。In a specific embodiment, the feature engineering process is characterized by the theme, detailed information and main content of the emergencies in the model data set, and the classification and risk level of the emergencies are used as target values.

结合表1，在一具体实施例中，文本清洗操作可将文本中的无用信息去除，例如去除文本中的标点符号、地点、事件等。Referring to Table 1, in a specific embodiment, the text cleaning operation can remove useless information in the text, such as removing punctuation marks, locations, events, etc. in the text.

表1清洗前清洗后a市b小区突发火灾成功扑灭突发火灾成功扑灭关于c商场配电室电容柜自燃事件（更新版）配电室电容柜自燃关于××年d公司首创e项目物业管理区域负一层g公司厨房发现烟雾的突发事件厨房发现烟雾Table 1 before cleaning after cleaning Sudden fire in residential area B of city A successfully extinguished The sudden fire was successfully extinguished Regarding the spontaneous combustion incident of the capacitor cabinet in the power distribution room of the c shopping mall (updated version) Spontaneous combustion of capacitor cabinet in power distribution room Regarding the unexpected incident in which smoke was found in the kitchen of company g on the first floor of the property management area of d company's first e project in ×× smoke found in the kitchen

结合表2，所述数据转换为将突发事件分类等级名称转换为类别特征。In combination with Table 2, the data conversion is to convert the name of the classification level of the emergency into a category feature.

表2分类等级类别特征A10B11C12C23C34Table 2 classification level category features A1 0 B1 1 C1 2 C2 3 C3 4

结合图2，在一实施例中，步骤S102包括：步骤S201~S205。Referring to FIG. 2, in an embodiment, step S102 includes: steps S201-S205.

S201、将所述模型数据集中的所有突发事件信息合并为同一文档，并对所述文档进行切词处理；S201. Merge all emergency event information in the model data set into the same document, and perform word segmentation processing on the document;

S202、通过gensim工具中的doc2bow函数对切词处理后的文档进行词频统计，得到二维数组；S202. Perform word frequency statistics on the document after the word segmentation process through the doc2bow function in the gensim tool to obtain a two-dimensional array;

S203、利用所述二维数组输入至TF-IDF模型，并由所述TF-IDF模型输出所述关键词；S203. Use the two-dimensional array to input to the TF-IDF model, and output the keyword from the TF-IDF model;

S204、基于预置分类范围对所述关键词进行正则匹配，得到所述关键词的匹配结果；S204. Perform regular matching on the keyword based on a preset classification range to obtain a matching result of the keyword;

S205、当匹配结果成功时，获取所述关键词在所述文档中的位置信息，并结合所述分类结果和位置信息生成第一分类结果；当匹配结果失败时，则设置其他类为第一分类结果。S205. When the matching result is successful, obtain the position information of the keyword in the document, and combine the classification result and the position information to generate a first classification result; when the matching result fails, set other categories as the first classification results.

在本实施例中，首先将所述模型数据集中的所有突发事件信息合并的文档进行切词处理，其次通过gensim工具中的doc2bow函数对所述文档中的文本进行词频统计，以得到二维数组，再次将二维数组输入至TF-IDF模型并输出关键词，然后基于预置分类范围对所述关键词进行正则匹配，以得到第一分类结果。In this embodiment, firstly, word segmentation processing is performed on the document in which all emergency information in the model data set is merged, and secondly, word frequency statistics are performed on the text in the document through the doc2bow function in the gensim tool to obtain a two-dimensional array, input the two-dimensional array to the TF-IDF model again and output keywords, and then perform regular matching on the keywords based on the preset classification range to obtain the first classification result.

结合表3，在具体实施例中，所述二维数组的最小元素为[[(0, 1), (1, 1), (2,1)], [(3, 1), (4, 1)], [(5, 1), (6, 1), (7, 1), (8, 1), (9, 1)]]，其中（词的ID号，词频）。举例来说，将二维数组输入TF-IDF模型后，TF-IDF模型输出的关键词如表3所示，所述模型数据集中突发事件1的文本内容是：因为起火导致电梯停运，则输出的关键词包括：“起火”和“电梯停运”，对所述关键词进行正则匹配，其中关键词“起火”的分类结果属于A1类，关键词“电梯停运”的分类结果属于B2类，但“起火”在文本中出现的顺序早于“电梯停运”，所以最后选择突发事件1的分类类别为A1。In conjunction with Table 3, in a specific embodiment, the minimum element of the two-dimensional array is [[(0, 1), (1, 1), (2,1)], [(3, 1), (4, 1)], [(5, 1), (6, 1), (7, 1), (8, 1), (9, 1)]], where (word ID number, word frequency). For example, after inputting the two-dimensional array into the TF-IDF model, the keywords output by the TF-IDF model are shown in Table 3. The text content of emergency 1 in the model data set is: the elevator was out of service due to a fire, Then the keywords output include: "on fire" and "elevator out of service", and the keywords are regularly matched, wherein the classification result of the keyword "on fire" belongs to the A1 category, and the classification result of the keyword "elevator out of service" belongs to Category B2, but the order of "fire" in the text appears earlier than "elevator outage", so the classification category of emergency 1 is finally selected as A1.

表3A1着火1.8367浓烟1.6543火情1.4567自燃1.3456B1电梯故障1.6367停梯1.5543电梯滑梯1.3567电梯进水1.126table 3 A1 on fire 1.8367 thick smoke 1.6543 fire 1.4567 spontaneous combustion 1.3456 B1 Elevator failure 1.6367 stop elevator 1.5543 elevator slide 1.3567 Elevator flooded 1.126

结合图3，在一实施例中，步骤S103包括步骤S301~S304：Referring to FIG. 3, in one embodiment, step S103 includes steps S301~S304:

S301、采用AT-BMC模型中的预训练模型对所述模型数据集输出对应的分类标签；S301, using the pre-trained model in the AT-BMC model to output the corresponding classification label for the model data set;

S302、判断所述分类标签是否属于基于预置分类范围；S302. Determine whether the classification label belongs to a preset classification range;

S303、若所述分类标签不属于基于预置分类范围，则设置其他类为第二分类结果；S303. If the classification label does not belong to the preset classification range, set other classes as the second classification result;

S304、若所述分类标签属于基于预置分类范围，则利用CRF层解码器对所述分类标签进行解码处理，以提取所述分类标签的证据，并结合所述分类标签和证据生成第二分类结果。S304. If the classification label belongs to the preset classification range, use a CRF layer decoder to decode the classification label to extract the evidence of the classification label, and combine the classification label and evidence to generate a second classification result.

在本实施例中，采用AT-BMC模型作为所述多任务学习模型的基本框架，AT-BMC模型是以非监督学习的方式对文集的隐含语义结构进行聚类的统计模型，利用AT-BMC模型中的预训练模型输出所述模型数据集对应的分类标签，当所述分类标签属于基于预置分类范围时，则利用CRF层解码器对所述分类标签进行解码处理，并将解码处理结果和结合所述分类标签生成第二分类结果。举例来说，在具体应用场景中，所述模型数据集中突发事件2为小朋友玩单杠摔倒扭伤，得到的分类标签为“摔倒扭伤”，解码处理得到的编码为[0，0......1，1]，其中0代表非证据，1表示证据，通过所述证据结合分类标签生成突发事件2的第二分类结果为F8，F8即代表“摔倒扭伤”。In this embodiment, the AT-BMC model is adopted as the basic framework of the multi-task learning model. The AT-BMC model is a statistical model for clustering the implicit semantic structure of the corpus in an unsupervised learning manner. The pre-training model in the BMC model outputs the classification label corresponding to the model data set. When the classification label belongs to the preset classification range, the CRF layer decoder is used to decode the classification label, and the decoding process Resulting and combining the classification labels to generate a second classification result. For example, in a specific application scenario, the emergency event 2 in the model data set is a child who fell and sprained while playing on a horizontal bar, and the obtained classification label is "fall and sprain", and the code obtained by the decoding process is [0, 0.. ....1, 1], where 0 represents non-evidence, 1 represents evidence, and the second classification result of emergency 2 generated by combining the evidence with classification labels is F8, and F8 represents "fall and sprain".

在一实施例中，步骤S103之前还包括：In one embodiment, before step S103, it also includes:

通过对抗训练模型对所述模型数据集进行数据增强处理，并通过对抗训练迭代构建对抗嵌入；performing data enhancement processing on the model data set through an adversarial training model, and iteratively constructing an adversarial embedding through adversarial training;

采用start指针和end指针对所述多任务学习模型进行边界匹配约束。Boundary matching constraints are performed on the multi-task learning model by using the start pointer and the end pointer.

在本实施例中，首先通过对抗训练模型对所述模型数据集进行扰动，以扩充所述模型数据集，其次进行多次PGD迭代来构建对抗嵌入，并在每次迭代中迭代出累积参数梯度，然后通过虚拟创建一个采样小批次，利用累积梯度有效地逐一更新所述多任务学习模型的参数，并利用在嵌入空间上操作的对抗性训练作为有效的正则化，以改善共享编码器的泛化，从而提高多任务学习模型的泛化能力和鲁棒性。例如，突发事件2为小朋友玩单杠摔倒扭伤，对所述突发事件2进行扰动，则得到具体内容为老人走路摔倒扭伤的突发事件3，突发事件2和突发事件3的分类标签都是F8(扭伤跌倒)，将分类标签经过词嵌入后和所述模型数据集一起输入到CRF层解码器中对多任务学习模型进行训练，使同一类别突发事件的描述不同时多任务学习模型能正确进行结果分类，从而减少噪音文本对多任务学习模型的干扰，提高多任务学习模型的分类能力。In this embodiment, firstly, the model data set is perturbed through the confrontation training model to expand the model data set, and then multiple PGD iterations are performed to construct the confrontation embedding, and the cumulative parameter gradient is iterated in each iteration , then effectively update the parameters of the multi-task learning model one by one with cumulative gradients by virtually creating a sampled mini-batch, and utilize adversarial training operating on the embedding space as an effective regularization to improve the performance of the shared encoder. Generalization, thereby improving the generalization ability and robustness of the multi-task learning model. For example, emergency 2 is a child who fell and sprained while playing on a horizontal bar. If the emergency 2 is disturbed, the specific content is emergency 3, emergency 2 and emergency 3 whose specific content is the elderly walking, falling and spraining. The classification labels are all F8 (sprain and fall), and the classification labels are input into the CRF layer decoder together with the model data set after word embedding, and the multi-task learning model is trained, so that the description of the same category of emergencies is different at the same time. The task learning model can correctly classify the results, thereby reducing the interference of noisy text on the multi-task learning model and improving the classification ability of the multi-task learning model.

此外，本实施例还采用start指针和end指针作为所述多任务学习模型的边界限制，从而促进所述多任务学习模型更准确的定位证据边界，通过对边界位置进行边界匹配约束，使得所述多任务学习模型能够进一步关注模型边界相关的区域，以提高所述多任务学习模型分类的准确度。In addition, this embodiment also uses the start pointer and the end pointer as the boundary constraints of the multi-task learning model, thereby promoting the multi-task learning model to more accurately locate the evidence boundary, and by performing boundary matching constraints on the boundary position, the The multi-task learning model can further focus on regions related to model boundaries, so as to improve the classification accuracy of the multi-task learning model.

在一实施例中，步骤S103还包括：In one embodiment, step S103 also includes:

按照下式，计算总的损失函数L对所述多任务学习模型进行优化更新：According to the following formula, calculate the total loss function L to optimize and update the multi-task learning model:

，

,

式中，

表示预训练模型的损失函数，/>

表示CRF层的损失函数，/>

表示对抗训练模型的损失函数，/>

表示边界匹配约束的损失函数，/>

和/>

均表示超参数。In the formula,

Indicates the loss function of the pre-trained model, />

Indicates the loss function of the CRF layer, />

represents the loss function of the adversarial training model, />

A loss function representing the bounds matching constraint, />

and />

Both represent hyperparameters.

在本实施例中，通过计算多任务学习模型的损失函数，得到损失函数之后，多任务学习模型通过反向传播去更新各个参数，来降低输出结果的真实值与预测值之间的损失，使得多任务学习模型生成的预测值往真实值方向靠拢，从而达到学习的目的。损失函数可以很好地反映模型与实际数据的差距，更好地对后续优化工具（梯度下降等）进行分析与理解。In this embodiment, by calculating the loss function of the multi-task learning model, after obtaining the loss function, the multi-task learning model updates each parameter through backpropagation to reduce the loss between the real value and the predicted value of the output result, so that The predicted value generated by the multi-task learning model is close to the real value, so as to achieve the purpose of learning. The loss function can well reflect the gap between the model and the actual data, and better analyze and understand the subsequent optimization tools (gradient descent, etc.).

在具体实施例中，预训练模型、CRF层、对抗训练模型、边界约束共享同一编码pretained Encode。所述模型数据集

一方面用于预训练模型输出分类标签/>

，一方面用于CRF层输出证据/>

，按照下式，计算预训练模型的损失函数/>

：In a specific embodiment, the pre-training model, CRF layer, confrontation training model, and boundary constraints share the same code preserved Encode. The model dataset

On the one hand, it is used to output classification labels for pre-trained models />

, on the one hand for the CRF layer output evidence />

, according to the following formula, calculate the loss function of the pre-training model />

:

，

,

式中，i表示所述模型数据集中文字的序号，n表示所述模型数据集文字总数。In the formula, i represents the serial number of the text in the model data set, and n represents the total number of text in the model data set.

将预训练模型输出的分类标签经过词嵌入生成序列，并和所述模型数据集一起输入到CRF层解码器中对多任务学习模型进行训练，得到证据

，按照下式，计算对应预训练模型的损失函数/>

：The classification labels output by the pre-training model are generated through word embedding, and input into the CRF layer decoder together with the model data set to train the multi-task learning model and obtain evidence

, according to the following formula, calculate the loss function corresponding to the pre-trained model />

:

，

,

式中，j表示分类类别的序号，m表示分类类别的总数。In the formula, j represents the serial number of the classification category, and m represents the total number of classification categories.

按照下式，计算对抗训练模型的损失函数

：According to the following formula, calculate the loss function of the confrontation training model

:

其中：in:

，

,

式中，

是基于扰动的对抗训练PAT的损失，/>

是正则项，θ表示先验分布中的参数，/>

表示对抗嵌入的交叉熵损失，/>

表示扰动，/>

表示范数，/>

，/>

均表示交叉熵损失，

表示预训练模型的预测结果。In the formula,

is the loss of perturbation-based adversarial training PAT, />

is a regular term, θ represents the parameters in the prior distribution, />

represents the cross-entropy loss against embeddings, />

Indicates disturbance, />

Indicates the norm, />

, />

Both represent the cross-entropy loss,

Indicates the prediction result of the pre-trained model.

利用在嵌入空间上操作的对抗性训练作为有效的正则化，以改善共享编码器的泛化，减少鲁棒性错误。对抗性训练通过PGD（Projected gradient descent，投影梯度下降算法）监督进行，K步PGD 需要通过网络进行K次前向-后向传播，计算成本较高，此外，K步PGD之后只有最后一步的扰动用于对抗训练模型参数更新。因此本实施例通过遵循 FreeLB算法中的自由对抗训练框架，进行多次PGD 迭代来构建对抗嵌入，并在每次迭代中迭代出累积参数梯度，然后通过虚拟创建一个采样小批次，利用累积梯度有效地逐一更新模型参数。Utilizes adversarial training operating on the embedding space as an effective regularization to improve generalization of shared encoders and reduce robustness errors. Adversarial training is supervised by PGD (Projected gradient descent, projected gradient descent algorithm). K-step PGD requires K times of forward-backward propagation through the network, and the calculation cost is high. In addition, after K-step PGD, there is only the last step of disturbance Used for adversarial training model parameter updates. Therefore, this embodiment follows the free adversarial training framework in the FreeLB algorithm, performs multiple PGD iterations to build an adversarial embedding, and iterates out the cumulative parameter gradient in each iteration, and then creates a sampling small batch virtually, using the cumulative gradient Efficiently update model parameters one by one.

通过下式计算边界约束的损失函数

：Calculate the loss function of the boundary constraints by the following formula

:

其中：in:

，

,

式中，

表示在分类标签下，证据对应的预测起始索引与其对应的结束索引匹配的概率，W表示模型隐藏层的权重函数，/>

表示模型权重系数，/>

表示每个词作为起始词的概率分布，/>

表示每个词作为结束词的概率分布，i=1,...,L，j=1,...,L。In the formula,

Indicates the probability that the predicted start index corresponding to the evidence matches the corresponding end index under the classification label, W represents the weight function of the hidden layer of the model, />

Indicates the model weight coefficient, />

Indicates the probability distribution of each word as the starting word, />

Represents the probability distribution of each word as an end word, i=1,...,L, j=1,...,L.

本实施例中的CRF层存在生成非法标签序列的局限性，因为CRF层鼓励合理的标签序列，而对于不合理的过渡惩罚较低。因此，本实施例使用边界约束来鼓励CRF层在定位边界时更加准确。The CRF layer in this embodiment has the limitation of generating illegal tag sequences, because the CRF layer encourages reasonable tag sequences and has relatively low penalties for unreasonable transitions. Therefore, this embodiment uses boundary constraints to encourage the CRF layer to be more accurate in locating boundaries.

在一实施例中，步骤S104包括：In one embodiment, step S104 includes:

判断所述第一分类结果和第二分类结果是否为其他类；judging whether the first classification result and the second classification result are other classes;

若所述第一分类结果和第二分类结果均为其他类，则将最终分类结果输出为其他类；If the first classification result and the second classification result are both other classes, outputting the final classification results as other classes;

若所述第一分类结果和第二分类结果中的任意一个为其他类，则将另一个分类结果作为最终分类结果输出；If any one of the first classification result and the second classification result is another class, then output the other classification result as the final classification result;

若所述第一分类结果和第二分类结果均不为其他类，则获取第一分类结果和第二分类结果的风险等级，并选择风险等级高的分类结果作为最终分类结果输出。If neither the first classification result nor the second classification result belongs to other categories, the risk levels of the first classification result and the second classification result are obtained, and the classification result with a higher risk level is selected as the final classification result output.

在本实施例中，通过判断所述第一分类结果和第二分类结果是否为其他类，并将判断结果进行集成，得到最终分类结果，能够更准确地对突发事件进行分类分级，以提高突发事件安全管理的效率。在具体应用场景中，还可在突发事件报事人上报突发事件后，提供突发事件管理方法输出的分类结果作为参考，使报事人可参考分类结果，从而决定是否修改自己上报的突发事件的分类分级，通过报事人的反馈能够进一步提到突发事件分类分级的准确度，实现突发事件安全管理的进一步优化。In this embodiment, by judging whether the first classification result and the second classification result belong to other categories, and integrating the judging results to obtain the final classification result, emergencies can be classified and graded more accurately to improve Efficiency of emergency safety management. In a specific application scenario, after the emergency reporter reports the emergency, the classification result output by the emergency management method can be provided as a reference, so that the reporter can refer to the classification result to decide whether to modify the report The classification and grading of emergencies can further improve the accuracy of emergency classification and grading through the feedback of the reporter, and realize the further optimization of emergency safety management.

在具体应用场景中，本发明实施例还对模型的召回率和准确率进行评估。In a specific application scenario, the embodiment of the present invention also evaluates the recall rate and accuracy rate of the model.

按照下式计算精确率Precision，精确率为分类正确的正样本数 (TP) 占预测为正样本的总样本数 (TP + FP) 的比例：The accuracy rate Precision is calculated according to the following formula, and the accuracy rate is the ratio of the number of positive samples (TP) that are correctly classified to the total number of samples predicted to be positive samples (TP + FP):

按照下式计算召回率Recall ，召回率为分类正确的正样本数 (TP) 占实际的正样本总数 (TP + FN) 的比例：The recall rate Recall is calculated according to the following formula, and the recall rate is the ratio of the number of correctly classified positive samples (TP) to the actual total number of positive samples (TP + FN):

在一具体实施例中，第一分类结果的、第二分类结果和最终分类结果对于突发事件D1、D2、D3、D4的精确率如表4所示，结合表4可以得到投票集成所得最终分类结果相较于第一分类结果和第二分类结果的精确率更高的结论，可见投票集成的预测性能更优秀。In a specific embodiment, the accuracy rates of the first classification result, the second classification result, and the final classification result for emergencies D1, D2, D3, and D4 are shown in Table 4. In combination with Table 4, the final result obtained by vote integration can be obtained. The conclusion that the accuracy of the classification result is higher than that of the first classification result and the second classification result shows that the prediction performance of the voting ensemble is better.

表4D1D2D3D4第一分类结果0.900.960.950.92第二分类结果0.920.980.930.93最终分类结果0.9310.960.94Table 4 D1 D2 D3 D4 First Classification Results 0.90 0.96 0.95 0.92 Second Classification Results 0.92 0.98 0.93 0.93 final classification result 0.93 1 0.96 0.94

图4为本发明实施例提供的一种突发事件管理装置400的示意性框图，该装置400包括：FIG. 4 is a schematic block diagram of anemergency management device 400 provided by an embodiment of the present invention, and thedevice 400 includes:

数据处理单元401，用于获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集；Adata processing unit 401, configured to acquire emergency event information, and preprocess the emergency event information to obtain a model data set;

第一分类单元402，用于基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果；Thefirst classification unit 402 is configured to extract keywords from the model data set based on a preset classification range, and perform regular matching processing on the extracted keywords to obtain a first classification result;

第二分类单元403，用于将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果；Thesecond classification unit 403 is configured to input the model data set into a multi-task learning model, and output the multi-task learning model based on a preset classification range to obtain a second classification result;

投票集成单元404，用于对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出。The vote integration unit 404 is configured to perform vote integration on the first classification result and the second classification result, and output the vote integration result as a final classification result.

在一实施例中，所述突发事件信息包括事件时间，事件主题，事件主要内容，所述数据处理单元401包括：In one embodiment, the emergency event information includes event time, event subject, event main content, and thedata processing unit 401 includes:

特征工程单元，用于获取突发事件信息对应的事件主表ID，根据所述事件主表ID关联事件类型，并根据所述事件类型对所述突发事件信息进行特征工程处理；A feature engineering unit, configured to acquire the event master table ID corresponding to the emergency event information, associate event types according to the event master table ID, and perform feature engineering processing on the emergency event information according to the event type;

数据转换单元，用于对所述突发事件信息进行文本清洗，以及基于特征工程处理的结果对所述突发事件信息进行数据转换，以此构建模型数据集。The data conversion unit is configured to perform text cleaning on the emergency event information, and perform data conversion on the emergency event information based on the result of feature engineering processing, so as to construct a model data set.

结合图5，在一实施例中，所述第一分类单元402包括：Referring to FIG. 5, in an embodiment, thefirst classification unit 402 includes:

切词处理单元501，用于将所述模型数据集中的所有突发事件信息合并为同一文档，并对所述文档进行切词处理；A word segmentation processing unit 501, configured to combine all emergency event information in the model data set into the same document, and perform word segmentation processing on the document;

词频统计单元502，用于通过gensim工具中的doc2bow函数对切词处理后的文档进行词频统计，得到二维数组；The word frequency statistical unit 502 is used to carry out word frequency statistics to the document after word segmentation processing by the doc2bow function in the gensim tool to obtain a two-dimensional array;

关键词输出单元503，用于利用所述二维数组输入至TF-IDF模型，并由所述TF-IDF模型输出所述关键词；A keyword output unit 503, configured to use the two-dimensional array to input to the TF-IDF model, and output the keyword from the TF-IDF model;

正则匹配单元504，用于基于预置分类范围对所述关键词进行正则匹配，得到所述关键词的匹配结果；A regular matching unit 504, configured to perform regular matching on the keyword based on a preset classification range to obtain a matching result of the keyword;

匹配结果单元505，用于当匹配结果成功时，获取所述关键词在所述文档中的位置信息，并结合所述分类结果和位置信息生成第一分类结果；当匹配结果失败时，则设置其他类为第一分类结果。The matching result unit 505 is configured to obtain the position information of the keyword in the document when the matching result is successful, and combine the classification result and the position information to generate a first classification result; when the matching result fails, set The other classes are the results of the first classification.

结合图6，在一实施例中，所述第二分类单元403包括：Referring to FIG. 6, in an embodiment, thesecond classification unit 403 includes:

预训练单元601，用于采用AT-BMC模型中的预训练模型对所述模型数据集输出对应的分类标签；Apre-training unit 601, configured to output a corresponding classification label to the model data set using a pre-trained model in the AT-BMC model;

范围判断单元602，用于判断所述分类标签是否属于基于预置分类范围；Arange judging unit 602, configured to judge whether the classification label belongs to a preset classification range;

分类标签判断单元603，用于若所述分类标签不属于基于预置分类范围，则设置其他类为第二分类结果；A classificationlabel judging unit 603, configured to set other classes as the second classification result if the classification label does not belong to the preset classification range;

证据提取单元604，用于若所述分类标签属于基于预置分类范围，则利用CRF层解码器对所述分类标签进行解码处理，以提取所述分类标签的证据，并结合所述分类标签和证据生成第二分类结果。Theevidence extraction unit 604 is configured to use a CRF layer decoder to decode the classification label if the classification label belongs to a preset classification range, so as to extract the evidence of the classification label, and combine the classification label and The evidence generates a second classification result.

在一实施例中，所述第二分类单元403之前，还包括：In one embodiment, before thesecond classification unit 403, it also includes:

数据增强单元，用于通过对抗训练模型对所述模型数据集进行数据增强处理，并通过对抗训练迭代构建对抗嵌入；A data enhancement unit, configured to perform data enhancement processing on the model data set through an adversarial training model, and iteratively construct an adversarial embedding through an adversarial training;

边界约束单元，用于采用start指针和end指针对所述多任务学习模型进行边界匹配约束。The boundary constraint unit is used to perform boundary matching constraints on the multi-task learning model by using the start pointer and the end pointer.

在一实施例中，所述第二分类单元403还包括：In an embodiment, thesecond classification unit 403 further includes:

优化更新单元，用于按照下式，采用损失函数L对所述多任务学习模型进行优化更新：The optimization update unit is used to optimize and update the multi-task learning model using the loss function L according to the following formula:

，

,

式中，

表示预训练模型的损失函数，/>

表示CRF层的损失函数，/>

表示对抗训练模型的损失函数，/>

表示边界匹配约束的损失函数，/>

和/>

均表示超参数。In the formula,

Indicates the loss function of the pre-trained model, />

Indicates the loss function of the CRF layer, />

represents the loss function of the adversarial training model, />

A loss function representing the bounds matching constraint, />

and />

Both represent hyperparameters.

在一实施例中，所述投票集成单元404包括：In one embodiment, the voting integration unit 404 includes:

结果判断单元，用于判断所述第一分类结果和第二分类结果是否为其他类；a result judging unit, configured to judge whether the first classification result and the second classification result belong to other classes;

其他类结果输出单元，用于若所述第一分类结果和第二分类结果均为其他类，则将最终分类结果输出为其他类；Other class result output unit, configured to output the final classification result as other class if the first classification result and the second classification result are both other classes;

第一最终分类结果单元，用于若所述第一分类结果和第二分类结果中的任意一个为其他类，则将另一个分类结果作为最终分类结果输出；The first final classification result unit is configured to output another classification result as the final classification result if any one of the first classification result and the second classification result is another class;

第二最终分类结果单元，用于若所述第一分类结果和第二分类结果均不为其他类，则获取第一分类结果和第二分类结果的风险等级，并选择风险等级高的分类结果作为最终分类结果输出。The second final classification result unit is used to obtain the risk levels of the first classification result and the second classification result if the first classification result and the second classification result are not other classes, and select a classification result with a higher risk level output as the final classification result.

由于装置部分的实施例与方法部分的实施例相互对应，因此装置部分的实施例请参见方法部分的实施例的描述，这里暂不赘述。Since the embodiment of the device part corresponds to the embodiment of the method part, please refer to the description of the embodiment of the method part for the embodiment of the device part, and details will not be repeated here.

本发明实施例还提供了一种计算机可读存储介质，其上存有计算机程序，该计算机程序被执行时可以实现上述实施例所提供的步骤。该存储介质可以包括：U盘、移动硬盘、只读存储器（Read-Only Memory，ROM）、随机存取存储器（Random Access Memory，RAM）、磁碟或者光盘等各种可以存储程序代码的介质。An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed, the steps provided in the above-mentioned embodiments can be realized. The storage medium may include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk, and other media capable of storing program codes.

本发明实施例还提供了一种计算机设备，可以包括存储器和处理器，存储器中存有计算机程序，处理器调用存储器中的计算机程序时，可以实现上述实施例所提供的步骤。当然计算机设备还可以包括各种网络接口，电源等组件。An embodiment of the present invention also provides a computer device, which may include a memory and a processor. A computer program is stored in the memory. When the processor invokes the computer program in the memory, the steps provided in the above embodiments can be implemented. Of course, the computer equipment may also include components such as various network interfaces and power supplies.

说明书中各个实施例采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似部分互相参见即可。对于实施例公开的系统而言，由于其与实施例公开的方法相对应，所以描述的比较简单，相关之处参见方法部分说明即可。应当指出，对于本技术领域的普通技术人员来说，在不脱离本申请原理的前提下，还可以对本申请进行若干改进和修饰，这些改进和修饰也落入本申请权利要求的保护范围内。Each embodiment in the description is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for the related information, please refer to the description of the method part. It should be pointed out that those skilled in the art can make some improvements and modifications to the application without departing from the principles of the application, and these improvements and modifications also fall within the protection scope of the claims of the application.

还需要说明的是，在本说明书中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的状况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。It should also be noted that in this specification, relative terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations There is no such actual relationship or order between the operations. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

Claims

Translated fromChinese

1.一种突发事件管理方法，其特征在于，包括：1. A method for emergency management, characterized in that, comprising:

对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出；performing vote integration on the first classification result and the second classification result, and outputting the vote integration result as a final classification result;

所述突发事件信息包括事件时间，事件主题，事件主要内容；The emergency event information includes event time, event theme, and event main content;

所述获取突发事件信息，并对所述突发事件信息进行预处理，得到模型数据集，包括：The acquiring emergency event information, and preprocessing the emergency event information to obtain a model data set, including:

对所述突发事件信息进行文本清洗，以及基于特征工程处理的结果对所述突发事件信息进行数据转换，以此构建模型数据集；performing text cleaning on the emergency event information, and performing data conversion on the emergency event information based on the result of feature engineering processing, so as to construct a model data set;

所述基于预置分类范围对所述模型数据集进行关键词提取，并对提取的关键词进行正则匹配处理，得到第一分类结果，包括：The keyword extraction is performed on the model data set based on the preset classification range, and regular matching processing is performed on the extracted keywords to obtain the first classification result, including:

将所述模型数据集中的所有突发事件信息合并为同一文档，并对所述文档进行切词处理；Merging all emergency event information in the model data set into the same document, and performing word segmentation processing on the document;

通过gensim工具中的doc2bow函数对切词处理后的文档进行词频统计，得到二维数组；Use the doc2bow function in the gensim tool to perform word frequency statistics on the document after word segmentation to obtain a two-dimensional array;

利用所述二维数组输入至TF-IDF模型，并由所述TF-IDF模型输出所述关键词；Using the two-dimensional array to input to the TF-IDF model, and output the keywords from the TF-IDF model;

基于预置分类范围对所述关键词进行正则匹配，得到所述关键词的匹配结果；performing regular matching on the keyword based on a preset classification range to obtain a matching result of the keyword;

当匹配结果成功时，获取所述关键词在所述文档中的位置信息，并结合所述分类结果和位置信息生成第一分类结果；当匹配结果失败时，则设置其他类为第一分类结果；When the matching result is successful, obtain the position information of the keyword in the document, and combine the classification result and the position information to generate the first classification result; when the matching result fails, set other classes as the first classification result ;

所述将所述模型数据集输入至多任务学习模型中，并基于预置分类范围由所述多任务学习模型输出得到第二分类结果，包括：Said inputting said model data set into a multi-task learning model, and outputting said multi-task learning model based on a preset classification range to obtain a second classification result includes:

采用AT-BMC模型中的预训练模型对所述模型数据集输出对应的分类标签；Using the pre-training model in the AT-BMC model to output the corresponding classification label to the model data set;

判断所述分类标签是否属于基于预置分类范围；Judging whether the classification label belongs to a preset classification range;

若所述分类标签不属于基于预置分类范围，则设置其他类为第二分类结果；If the classification label does not belong to the range based on the preset classification, then setting other classes as the second classification result;

若所述分类标签属于基于预置分类范围，则利用CRF层解码器对所述分类标签进行解码处理，以提取所述分类标签的证据，并结合所述分类标签和证据生成第二分类结果；If the classification label belongs to a preset classification range, a CRF layer decoder is used to decode the classification label to extract evidence of the classification label, and combine the classification label and evidence to generate a second classification result;

所述对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出，包括：The performing vote integration on the first classification result and the second classification result, and outputting the vote integration result as the final classification result includes:

2.根据权利要求1所述的一种突发事件管理方法，其特征在于，所述将所述模型数据集输入至多任务学习模型中，并由所述多任务学习模型输出得到第二分类结果步骤之前，还包括：2. a kind of emergency management method according to claim 1, is characterized in that, described model data set is input in the multi-task learning model, and obtains the second classification result by the multi-task learning model output Before step, also include:

3.根据权利要求2所述的一种突发事件管理方法，其特征在于，所述将所述模型数据集输入至多任务学习模型中，并由所述多任务学习模型输出得到第二分类结果，还包括：3. a kind of emergency management method according to claim 2, is characterized in that, described model data set is input in the multi-task learning model, and obtains the second classification result by the multi-task learning model output ,Also includes:

按照下式，采用损失函数L对所述多任务学习模型进行优化更新：According to the following formula, the multi-task learning model is optimized and updated using the loss function L:

式中，/>

表示预训练模型的损失函数，

表示CRF层的损失函数，/>

表示对抗训练模型的损失函数，/>

表示边界匹配约束的损失函数，/>

和/>

均表示超参数。

In the formula, />

Represents the loss function of the pre-trained model,

Indicates the loss function of the CRF layer, />

represents the loss function of the adversarial training model, />

A loss function representing the bounds matching constraint, />

and />

Both represent hyperparameters.

4.一种突发事件管理装置，其特征在于，包括：4. An emergency management device, characterized in that it comprises:

投票集成单元，用于对所述第一分类结果与所述第二分类结果进行投票集成，并将投票集成结果作为最终分类结果输出；a vote integration unit, configured to perform vote integration on the first classification result and the second classification result, and output the vote integration result as the final classification result;

所述突发事件信息包括事件时间，事件主题，事件主要内容，所述数据处理单元包括：The emergency event information includes event time, event theme, event main content, and the data processing unit includes:

数据转换单元，用于对所述突发事件信息进行文本清洗，以及基于特征工程处理的结果对所述突发事件信息进行数据转换，以此构建模型数据集；A data conversion unit, configured to perform text cleaning on the emergency event information, and perform data conversion on the emergency event information based on the result of feature engineering processing, so as to construct a model data set;

所述第一分类单元包括：The first taxonomic unit includes:

切词处理单元，用于将所述模型数据集中的所有突发事件信息合并为同一文档，并对所述文档进行切词处理；A word segmentation processing unit, configured to combine all emergency event information in the model data set into the same document, and perform word segmentation processing on the document;

词频统计单元，用于通过gensim工具中的doc2bow函数对切词处理后的文档进行词频统计，得到二维数组；The word frequency statistics unit is used to perform word frequency statistics on the document after the word segmentation process through the doc2bow function in the gensim tool to obtain a two-dimensional array;

关键词输出单元，用于利用所述二维数组输入至TF-IDF模型，并由所述TF-IDF模型输出所述关键词；A keyword output unit, configured to use the two-dimensional array to input to the TF-IDF model, and output the keyword from the TF-IDF model;

正则匹配单元，用于基于预置分类范围对所述关键词进行正则匹配，得到所述关键词的匹配结果；A regular matching unit, configured to perform regular matching on the keyword based on a preset classification range to obtain a matching result of the keyword;

匹配结果单元，用于当匹配结果成功时，获取所述关键词在所述文档中的位置信息，并结合所述分类结果和位置信息生成第一分类结果；当匹配结果失败时，则设置其他类为第一分类结果；The matching result unit is used to obtain the position information of the keyword in the document when the matching result is successful, and combine the classification result and the position information to generate the first classification result; when the matching result fails, set other Class is the result of the first classification;

所述第二分类单元包括：The second taxa includes:

预训练单元，用于采用AT-BMC模型中的预训练模型对所述模型数据集输出对应的分类标签；The pre-training unit is used to adopt the pre-training model in the AT-BMC model to output the corresponding classification label to the model data set;

范围判断单元，用于判断所述分类标签是否属于基于预置分类范围；A range judging unit, configured to judge whether the classification label belongs to a preset classification range;

分类标签判断单元，用于若所述分类标签不属于基于预置分类范围，则设置其他类为第二分类结果；A classification label judging unit, configured to set other classes as the second classification result if the classification label does not belong to the preset classification range;

证据提取单元，用于若所述分类标签属于基于预置分类范围，则利用CRF层解码器对所述分类标签进行解码处理，以提取所述分类标签的证据，并结合所述分类标签和证据生成第二分类结果；The evidence extraction unit is configured to use a CRF layer decoder to decode the classification label if the classification label belongs to a preset classification range, so as to extract the evidence of the classification label, and combine the classification label and the evidence generating a second classification result;

所述投票集成单元包括：The voting integration unit includes:

第二最终分类结果单元，用于若所述第一分类结果和第二分类结果均不为其他类，则获取第一分类结果和第二分类结果的风险等级，并选择风险等级高的分类结果作为最终分类结果输出。The second final classification result unit is used to obtain the risk levels of the first classification result and the second classification result if the first classification result and the second classification result are not other classes, and select a classification result with a high risk level output as the final classification result.

5.一种计算机设备，其特征在于，包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如权利要求1至3任一项所述的一种突发事件管理方法。5. A computer device, characterized in that it comprises a memory, a processor, and a computer program stored on the memory and operable on the processor, and when the processor executes the computer program, the computer program according to the claim is realized. An emergency management method described in any one of 1 to 3.

6.一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现如权利要求1至3任一项所述的一种突发事件管理方法。6. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, a computer program according to any one of claims 1 to 3 is implemented. approach to emergency management.