Data processing method, apparatus, device, and method for generating an event extraction model

Abstract

The present application discloses a data processing method, including: obtaining a target natural language text; obtaining a summary text corresponding to an entity included in the target natural language text; generating a summary embedding vector of the target natural language text according to the summary text corresponding to the entity included in the target natural language text; and performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text. By adopting the above method, the problem of low performance of the event extraction system in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and specifically relates to a data processing method. This application also relates to a data processing device, an electronic device, and a storage device. This application also relates to a method for generating an event extraction model and a method for generating a training set. Background Art

[0002] In recent years, computer and Internet technologies have continued to develop at a high speed. The Internet has accumulated and continuously generated a huge amount of natural language text information, such as continuously updated news reports, opinions published by self-media, and social status information published by individuals. Behind these large amounts of natural language text information lies huge value, but currently most computer applications are difficult to directly process unstructured natural language text information. Therefore, it is necessary to develop technologies that can convert unstructured natural language text into structured information.

[0003] In the prior art, in the technology of converting unstructured natural language text into structured information, the method of manual annotation is usually used to obtain samples for training an event extraction system, and there is a problem of a small sample size, which affects the performance of the event extraction system to a certain extent. Summary of the Invention

[0004] This application provides a data processing method to solve the problem of the low performance of the event extraction system in the prior art.

[0005] This application provides a data processing method, including:

[0006] Obtain a target natural language text;

[0007] Obtain the summary text corresponding to the entity included in the target natural language text;

[0008] Generate a summary embedding vector of the target natural language text according to the summary text corresponding to the entity included in the target natural language text;

[0009] Perform event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text.

[0010] Optionally, it further includes:

[0011] Obtain the triple information corresponding to the entity included in the target natural language text from a second knowledge base; the triple information includes a head entity, a relationship between entities, and a tail entity;

[0012] Perform an embedded representation on the triple information to obtain an embedded expression of the entity;

[0013] Performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text, including:

[0014] Performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text and the embedded expression of the entity.

[0015] Optionally, performing embedded representation on the triple information to obtain the embedded expression of the entity, including:

[0016] Using a knowledge representation learning model to perform embedded representation on the triple information to obtain the embedded expression of the entity.

[0017] Optionally, the second knowledge base is a knowledge base recording triple information.

[0018] Optionally, obtaining the summary text corresponding to the entity included in the target natural language text, including:

[0019] Obtaining the entities included in the target natural language text;

[0020] According to the entities included in the target natural language text, obtaining the summary information of the entities included in the target natural language text.

[0021] Optionally, generating the summary embedding vector of the target natural language text according to the summary information of the entities included in the target natural language text, including:

[0022] According to the summary information of the entities included in the target natural language text, obtaining the embedding vectors of the entities included in the target natural language text;

[0023] Performing summary attention processing on the embedding vectors of all entities included in the target natural language text to generate the summary embedding vector of the target natural language text.

[0024] Optionally, obtaining the embedding vectors of the entities included in the target natural language text according to the summary information of the entities included in the target natural language text, including:

[0025] Encoding each sentence in the summary information of the entities included in the target natural language text to generate the sentence vector of each sentence;

[0026] Performing encoding processing on all sentence vectors to generate the embedding vectors of the entities included in the target natural language text.

[0027] Optionally, generating the summary embedding vector of the target natural language text according to the summary text of the entities included in the target natural language text, including:

[0028] Using at least one of the following networks, encode the summary information of the entities included in the target natural language text to generate a summary embedding vector of the target natural language text:

[0029] GRU network;

[0030] LSTM network;

[0031] CNN network;

[0032] Transformer network.

[0033] The present application also provides a data processing device, including:

[0034] A target natural language text acquisition unit, configured to acquire a target natural language text;

[0035] A summary text acquisition unit, configured to acquire summary text corresponding to the entities included in the target natural language text;

[0036] A summary embedding vector generation unit, configured to generate a summary embedding vector of the target natural language text according to the summary text corresponding to the entities included in the target natural language text;

[0037] An event extraction processing unit, configured to perform event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text.

[0038] The present application also provides an electronic device, including:

[0039] A processor; and

[0040] A memory, configured to store a program of a data processing method. After the device is powered on and runs the program of the data processing method through the processor, the following steps are performed:

[0041] Acquire a target natural language text;

[0042] Acquire summary text corresponding to the entities included in the target natural language text;

[0043] Generate a summary embedding vector of the target natural language text according to the summary text corresponding to the entities included in the target natural language text;

[0044] Perform event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text.

[0045] The present application also provides a storage device, storing a program of a data processing method. The program is run by a processor, and the following steps are performed: including:

[0046] Obtain the target natural language text;

[0047] Obtain the abstract text corresponding to the entity included in the target natural language text;

[0048] Generate an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text;

[0049] Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text.

[0050] The present application provides a method for generating an event extraction model, including:

[0051] Construct an initial event extraction model, wherein the parameters of the initial event extraction model are initialization data;

[0052] Obtain training data for training the initial event extraction model, wherein the training data is obtained according to any one of the foregoing methods;

[0053] Use the training data to train the initial event extraction model to obtain the target parameters of the initial event extraction model;

[0054] Generate a target event extraction model according to the target parameters.

[0055] Optionally, the generation method further includes:

[0056] Obtain test data including the target natural language text;

[0057] Input the test data into the target event extraction model to obtain event extraction information of the test data.

[0058] The present application provides a method for generating a training set, including:

[0059] Obtain entity objects in the triple knowledge base;

[0060] Retrieve in the knowledge base according to the entity objects to obtain the trigger words corresponding to the entity objects;

[0061] Process the trigger words using a remote supervision algorithm to obtain a labeled training set.

[0062] Optionally, the method further includes:

[0063] Obtain updated entity objects in the triple knowledge base;

[0064] Retrieve in the knowledge base according to the updated entity object to obtain the updated trigger word corresponding to the entity object;

[0065] Process the updated trigger word using a remote supervision algorithm to obtain an updated labeled training set.

[0066] Compared with the prior art, the present application has the following advantages:

[0067] The present application proposes a method for generating a summary embedding vector of a target natural language text using large-scale knowledge base data and using the summary embedding vector for event extraction processing of the target natural language text. The present application introduces a large amount of external information knowledge in a limited training set scenario to expand the information that can be represented by the limited training set, and solves the problem of limited performance of the event extraction system caused by the lack of labeled data. Description of the Drawings

[0068] Figure 1a It is a schematic diagram of an application scenario embodiment of a data processing method provided by the present application.

[0069] Figure 1b It is a schematic diagram of news event extraction in an application scenario of a data processing method provided by the present application.

[0070] Figure 1 It is a flowchart of a data processing method provided by the first embodiment of the present application.

[0071] Figure 2 It is a schematic diagram of a process for generating a summary embedding vector of a sentence by using a Wikipedia embedding module according to the summary information of entities included in the sentence provided by the first embodiment of the present application.

[0072] Figure 3 It is a schematic diagram of encoding the summary information of entities included in a target natural language text by using a GRU network to generate a summary embedding vector of the target natural language text provided by the first embodiment of the present application.

[0073] Figure 4 It is a schematic diagram of event extraction processing of a target natural language text according to the summary embedding vector of the target natural language text and the embedded expression of the entity provided by the first embodiment of the present application.

[0074] Figure 5 It is a schematic diagram of a data processing device provided by the first embodiment of the present application.

[0075] Figure 6 It is a schematic diagram of an electronic device provided by the first embodiment of the present application. Detailed Embodiments

[0076] Numerous specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific implementations disclosed below.

[0077] To enable those skilled in the art to better understand the solution of the present application, a specific application scenario embodiment of the present application will be described in detail first. As Figure 1a shown, it is a schematic diagram of an embodiment of an application scenario of a data processing method provided by the present application. In the specific implementation process, user 108-1 can send an instruction to obtain news data to news data server 109 through client application 107-1 on client device 106-1 by using network 105. As Figure 1b shown, the news data can be news data of the AB border confrontation event. News data server 109 returns the news data to client application 107-1 according to the instruction, and client application 107-1 sends the news data to event extraction and generation server 100 through network 105. After receiving the news data, event extraction and generation server 100 sends the news data to target natural language text obtaining unit 101, the target natural language text obtaining unit. Then, through summary text obtaining unit 102, the summary text corresponding to the entity included in the target natural language text is obtained. Next, through summary embedding vector generation unit 103, a summary embedding vector of the target natural language text is generated according to the summary text corresponding to the entity included in the target natural language text. Furthermore, through event extraction processing unit 104, event extraction processing is performed on the target natural language text according to the summary embedding vector of the target natural language text to obtain an event extraction result. Finally, event extraction and generation server 100 returns the generated event extraction result to client application 107-1. Please refer to Figure 1b shown, it is the event extraction result of the tracking event of the news of the AB border confrontation event.

[0078] In the prior art, in the technology of converting unstructured natural language text into structured information, the method of manual annotation is usually adopted to obtain samples for training the event extraction system, and there is a problem of a relatively small sample size, which to a certain extent affects the performance of the event extraction system.

[0079] By using the data processing method provided in this embodiment, a summary embedding vector of the target natural language text is generated using large-scale knowledge base data, and the summary embedding vector is used for event extraction processing of the target natural language text, thereby introducing a large amount of external information knowledge in a limited training set scenario and expanding the information that can be represented by the limited training set.

[0080] The first embodiment of the present application provides a data processing method, which will be described below in conjunction with Figures 1 to 4 for illustration.

[0081] As Figure 1 shown, in step S101, the target natural language text is obtained.

[0082] The target natural language text may include a target sentence or a target phrase. For example, the sentence "He was injured by a soldier's grenade attack" is a target natural language text.

[0083] As Figure 1 shown, in step S102, the summary text corresponding to the entity included in the target natural language text is obtained from the first knowledge base.

[0084] The first knowledge base is a knowledge base that records summary text. For example, the first knowledge base can be Wikipedia, which is a large-scale encyclopedia knowledge editing site that stores a large amount of knowledge represented in natural language text and can be automatically updated continuously. For example, Figure 2 "A hand Grenade is any small bomb..." in

[0085] is the summary text of Grenade recorded by Wikipedia.

[0086] The obtaining of the summary text corresponding to the entity included in the target natural language text from the first knowledge base includes:

[0087] Obtaining the entities included in the target natural language text;

[0088] According to the entities included in the target natural language text, obtaining the summary information of the entities included in the target natural language text in the first knowledge base.

[0089] For example, the natural language text is the sentence "...attack…fellow". The process of obtaining the summary text corresponding to the entities included in the above sentence from the first knowledge base is as follows: First, obtain the entities included in the above sentence as "grenade" and "soldier", and then obtain the summary information of the entities "grenade" and "soldier" in the first knowledge base.

[0090] As Figure 1 shown, in step S103, a summary embedding vector of the target natural language text is generated according to the summary text corresponding to the entities included in the target natural language text.

[0091] The summary embedding vector refers to a vector finally generated from the summary text corresponding to the entities included in the target natural language text.

[0092] Generating the summary embedding vector of the target natural language text according to the summary information of the entities included in the target natural language text includes:

[0093] According to the summary information of the entities included in the target natural language text, obtain the embedding vectors of the entities included in the target natural language text;

[0094] Perform summary attention processing on the embedding vectors of all entities included in the target natural language text to generate the summary embedding vector of the target natural language text.

[0095] Obtaining the embedding vectors of the entities included in the target natural language text according to the summary information of the entities included in the target natural language text includes:

[0096] Encode each sentence in the summary information of the entities included in the target natural language text to generate a sentence vector for each sentence;

[0097] Perform encoding processing on all sentence vectors to generate the embedding vectors of the entities included in the target natural language text.

[0098] As Figure 2 shown, it is a schematic diagram of a process of using a Wikipedia embedding module to generate a summary embedding vector of a sentence from the summary information of the entities included in the sentence. The goal of the Wikipedia embedding module is to generate a corresponding summary embedding vector for each sentence, taking the sentence as a unit, from the summary text corresponding to all entities included in the sentence.

[0099] The specific steps of using the Wikipedia embedding module to generate a summary embedding vector of a sentence from the summary information of the entities included in the sentence are as follows: In step S201, according to the entities included in the sentence, obtain the summary text corresponding to the entities included in the sentence in Wikipedia.Figure 2 The summary information of the entities "grenade" and "soldier" in Wikipedia was obtained. In step S202, for each sentence in the obtained Wikipedia summary, the module encodes all the words in a sentence into a sentence vector through the word encoding step. Figure 2 The summary information of the entities "grenade" and "soldier" in Wikipedia are two sentences respectively. Therefore, the summary information of the entity "grenade" in Wikipedia is encoded into two sentence vectors 201 and sentence vector 202 through the word encoding step; the summary information of the entity "soldier" in Wikipedia is encoded into two sentence vectors 203 and sentence vector 204 through the word encoding step; in step S203, the Wikipedia embedding module encodes all the sentence vectors in a summary into an entity embedding vector through the sentence encoding step. Figure 2 In it, sentence vector 201 and sentence vector 202 are encoded into entity embedding vector 205, and sentence vector 203 and sentence vector 204 are encoded into entity embedding vector 206. In step S204, for all the entity embedding vectors generated in the previous steps in a sentence, the Wikipedia embedding module generates a summary embedding vector through the summary attention step. Figure 2 In it, the summary embedding vector is generated according to entity embedding vector 205 and entity embedding vector 206 through the summary attention step.

[0100] Generating a summary embedding vector of the target natural language text according to the summary text of the entities included in the target natural language text includes:

[0101] Using at least one of the following networks to encode the summary information of the entities included in the target natural language text to generate a summary embedding vector of the target natural language text:

[0102] GRU network;

[0103] LSTM network;

[0104] CNN network;

[0105] Transformer network.

[0106] As Figure 3 shown, it is a schematic diagram of using a GRU network to encode the summary information of the entities included in the target natural language text to generate a summary embedding vector of the target natural language text. As Figure 3 shown, in the case where the input is a single sentence, the dimension of the input is 3, that is, the first dimension is how many entities are included in the sentence, the second dimension is how many sentences are included in the summary of the entity, and the third dimension is how many words are included in each sentence.Figure 3 In [the method of] [the present invention], three-digit subscripts are used to refer to vocabulary, two-digit subscripts are used to refer to sentences, and one-digit subscripts are used to refer to entities. Figure 3 Only the case where a sentence contains two entities is shown in [the method of] [the present invention].

[0107] Figure 3 In [the method of] [the present invention], a GRU network is used to encode the summary information of the entities contained in the target natural language text to generate a summary embedding vector of the target natural language text. In specific implementation, the GRU network can also be replaced with an LSTM, CNN, or Transformer structure. For different data and operating environments, a suitable network structure can be selected according to the algorithm effect and execution efficiency.

[0108] As Figure 1 shown, in step S104, according to the summary embedding vector of the target natural language text, event extraction processing is performed on the target natural language text.

[0109] The event extraction refers to extracting events from unstructured natural language text and converting them into structured information. It mainly includes the extraction of trigger words and elements. Among them, the trigger word is the core word indicating the occurrence of an event, mostly a verb or a noun, and the element refers to the element that composes the event and the role that composes the event.

[0110] Performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text can improve the performance of the event extraction system.

[0111] To further improve the performance of the event extraction system, the method of the first embodiment of the present application may further include:

[0112] Obtaining triple information corresponding to the entities contained in the target natural language text from a second knowledge base; the triple information includes a head entity, a relationship between entities, and a tail entity;

[0113] Performing an embedded representation on the triple information to obtain an embedded expression of the entity.

[0114] The second knowledge base is a knowledge base that records triple information. For example, the Freebase knowledge base.

[0115] Performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text includes:

[0116] Performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text and the embedded expression of the entity.

[0117] The basic forms of the triple information mainly include (entity 1 - relation - entity 2) and (entity - attribute - attribute value), etc. Each entity (the extension of a concept) can be represented by a globally unique determined pair, and the AVP can be used to depict the internal characteristics of the entity, while the relation can be used to connect two entities and depict the association between them.

[0118] As follows Figure 1 As shown in the knowledge graph example below, China is an entity, Beijing is an entity, and China - capital - Beijing is a triple of (entity - relation - entity).

[0119] In specific implementation, for the embedded representation of the triple information to obtain the embedded expression of the entity, a knowledge representation learning model can be used to perform the embedded representation of the triple information to obtain the embedded expression of the entity. For example, knowledge embedding libraries such as OpenKE can be used to perform the embedded representation of the triple to obtain the embedded expression of the entity. OpenKE is an open - source knowledge representation learning platform, including commonly used knowledge representation learning (KRL) methods.

[0120] According to the summary embedding vector of the target natural language text and the embedded expression of the entity, perform event extraction processing on the target natural language text, including:

[0121] According to the summary embedding vector of the entity in the summary embedding vector of the target natural language text and the embedded expression of the entity, obtain the word vector of the entity;

[0122] According to the word vector of the entity, perform event extraction processing on the target natural language text.

[0123] For example, as Figure 4 shown, the summary text of the entity "Grenade" generates the summary embedding vector 401 of the entity "Grenade" through the summary embedding module, and the triple information of the entity "Grenade" generates the embedded expression 402 of the entity "Grenade" through the Freebase knowledge embedding module. Combine the summary embedding vector 401 and the embedded expression 402 to generate the word vector 403 of the entity "Grenade", and perform event extraction processing on the target natural language text according to the word vector of the entity.

[0124] It should be noted here that for event extraction from large-scale knowledge bases, it can be applied to product reviews on e-commerce websites, video reviews on video websites, and ticket reviews on e-ticketing platforms. For example, in the promotional activities of e-commerce websites, events such as "Product XX is the same model as YY" can be obtained from popular product reviews.

[0125] So far, the first embodiment of this application has been introduced in detail. The data processing method provided by the first embodiment of this application expands the original training set information by introducing knowledge from the first knowledge base and the second knowledge base to make up for the problem of the high cost of manually labeled data and reduce the problem of reducing the systematic performance of event extraction. For example, in the scenario of news event extraction, the manually labeled training set contains the annotation of the event "President A visits China", but limited by the cost of manual annotation, events such as "President A visits the Czech Republic" are very likely to be missed. If information from knowledge bases such as Wikipedia and Freebase is used, the Czech Republic is labeled as a country, which can increase the probability of the event extraction system detecting this event. In addition, using the information of Wikipedia and Freebase knowledge bases can also improve the information utilization rate of the event extraction system. For example, in the following sentence: He was injured by a grenade attack by a soldier. The training data cannot provide detailed explanations for the two nouns "soldier" and "grenade", but in the Wikipedia and Freebase knowledge bases, the explanations of "soldier" and "grenade" are respectively: a small throwing bomb and an organized and disciplined armed force. Therefore, introducing Wikipedia and Freebase knowledge bases can provide more comprehensive information to the event extraction system. Another example is in the scenario of natural disaster event extraction. The manually labeled training set contains "A tornado appears in Florida", but similarly, manual annotation is limited by high costs and may miss "A tornado appears in Oklahoma". However, if the knowledge base is used to assist event extraction, Oklahoma will be recognized as a state name, thereby increasing the probability of this event being detected. This application introduces a large amount of external information knowledge in the scenario of a limited training set to expand the information that the limited training set can represent; at the same time, this application first uses the Freebase knowledge embedding representation module to introduce the knowledge in Freebase into the event extraction field in the form of discrete vectors; in addition, the Wikipedia abstract embedding module proposed in this application applies the knowledge in Wikipedia to the event extraction task and achieves good results.

[0126] Corresponding to the data processing method provided by the first embodiment of this application, the second embodiment of this application also provides a data processing device.

[0127] As Figure 5As shown, the data processing device includes:

[0128] A target natural language text acquisition unit 501 for acquiring a target natural language text;

[0129] An abstract text acquisition unit 502 for acquiring an abstract text corresponding to an entity included in the target natural language text;

[0130] An abstract embedding vector generation unit 503 for generating an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text;

[0131] An event extraction processing unit 504 for performing event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text.

[0132] Optionally, the data processing device further includes:

[0133] A triple information acquisition unit for acquiring triple information corresponding to an entity included in the target natural language text from a second knowledge base; the triple information includes a head entity, a relationship between entities, and a tail entity;

[0134] An embedded expression acquisition unit for performing an embedded representation on the triple information to obtain an embedded expression of the entity;

[0135] The event extraction processing unit is specifically configured to:

[0136] Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text and the embedded expression of the entity.

[0137] Optionally, the embedded expression acquisition unit is specifically configured to:

[0138] Adopt a knowledge representation learning model to perform an embedded representation on the triple information to obtain an embedded expression of the entity.

[0139] Optionally, the second knowledge base is a knowledge base recording triple information.

[0140] Optionally, the abstract text acquisition unit is specifically configured to:

[0141] Obtain the entities included in the target natural language text;

[0142] According to the entities included in the target natural language text, obtain abstract information of the entities included in the target natural language text.

[0143] Optionally, the summary embedding vector generation unit is specifically configured to:

[0144] Obtain the embedding vectors of the entities included in the target natural language text according to the summary information of the entities included in the target natural language text;

[0145] Perform summary attention processing on the embedding vectors of all the entities included in the target natural language text to generate the summary embedding vector of the target natural language text.

[0146] Optionally, the summary embedding vector generation unit is specifically configured to:

[0147] Encode each sentence in the summary information of the entities included in the target natural language text to generate the sentence vector of each sentence;

[0148] Perform encoding processing on all the sentence vectors to generate the embedding vectors of the entities included in the target natural language text.

[0149] Optionally, the summary embedding vector generation unit is specifically configured to:

[0150] Use at least one of the following networks to encode the summary information of the entities included in the target natural language text to generate the summary embedding vector of the target natural language text:

[0151] GRU network;

[0152] LSTM network;

[0153] CNN network;

[0154] Transformer network.

[0155] It should be noted that for the detailed description of the device provided in the second embodiment of the present application, reference may be made to the relevant description of the first embodiment of the present application, which will not be elaborated here.

[0156] Corresponding to the data processing method provided in the first embodiment of the present application, the third embodiment of the present application further provides an electronic device.

[0157] As Figure 6 shown, the electronic device includes:

[0158] A processor 601; and

[0159] A memory 602 for storing the program of the data processing method. After the device is powered on and the processor runs the program of the data processing method, the following steps are executed:

[0160] Obtain the target natural language text;

[0161] Obtain the abstract text corresponding to the entity included in the target natural language text;

[0162] Generate an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text;

[0163] Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text.

[0164] Optionally, the electronic device further performs the following steps:

[0165] Obtain the triple information corresponding to the entity included in the target natural language text from the second knowledge base; the triple information includes a head entity, a relationship between entities, and a tail entity;

[0166] Perform an embedded representation on the triple information to obtain an embedded expression of the entity;

[0167] Performing event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text includes:

[0168] Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text and the embedded expression of the entity.

[0169] Optionally, performing an embedded representation on the triple information to obtain an embedded expression of the entity includes:

[0170] Adopt a knowledge representation learning model to perform an embedded representation on the triple information to obtain an embedded expression of the entity.

[0171] Optionally, the second knowledge base is a knowledge base recording triple information.

[0172] Optionally, obtaining the abstract text corresponding to the entity included in the target natural language text includes:

[0173] Obtain the entities included in the target natural language text;

[0174] According to the entities included in the target natural language text, obtain the abstract information of the entities included in the target natural language text.

[0175] Optionally, generating an abstract embedding vector of the target natural language text according to the abstract information of the entities included in the target natural language text includes:

[0176] According to the abstract information of the entities included in the target natural language text, obtain the embedding vectors of the entities included in the target natural language text;

[0177] Perform abstract attention processing on the embedding vectors of all entities included in the target natural language text to generate an abstract embedding vector of the target natural language text.

[0178] Optionally, obtaining the embedding vectors of the entities included in the target natural language text according to the abstract information of the entities included in the target natural language text includes:

[0179] Encode each sentence in the abstract information of the entities included in the target natural language text to generate a sentence vector for each sentence;

[0180] Perform encoding processing on all sentence vectors to generate the embedding vectors of the entities included in the target natural language text.

[0181] Optionally, generating an abstract embedding vector of the target natural language text according to the abstract text of the entities included in the target natural language text includes:

[0182] Use at least one of the following networks to encode the abstract information of the entities included in the target natural language text to generate an abstract embedding vector of the target natural language text:

[0183] GRU network;

[0184] LSTM network;

[0185] CNN network;

[0186] Transformer network.

[0187] It should be noted that for the detailed description of the electronic device provided in the third embodiment of the present application, reference can be made to the relevant description of the first embodiment of the present application, which will not be elaborated here.

[0188] Corresponding to a data processing method provided in the first embodiment of the present application, the fourth embodiment of the present application further provides a storage device storing a program of the data processing method, and the program is run by a processor to execute the following steps: including:

[0189] Obtain the target natural language text;

[0190] Obtain the abstract text corresponding to the entities included in the target natural language text;

[0191] Generate an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entities included in the target natural language text;

[0192] Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text.

[0193] It should be noted that for the detailed description of the storage device provided in the fourth embodiment of the present application, reference may be made to the relevant description of the first embodiment of the present application, which will not be elaborated here.

[0194] The fifth embodiment of the present application provides a method for generating an event extraction model, including:

[0195] Construct an initial event extraction model, where the parameters of the initial event extraction model are initialization data.

[0196] For example, a DMCNN (Dynamic Multi-Pooling Convolutional Neural Networks) model can be constructed as the event extraction model. Obtain training data for training the initial event extraction model, where the training data is obtained according to any one of the methods provided in the first embodiment of the present application.

[0197] For example, the DMCNN model can be trained using the training data.

[0198] Use the training data to train the initial event extraction model to obtain the target parameters of the initial event extraction model.

[0199] For example, use the backpropagation algorithm to train the DMCNN model to obtain the target parameters of the initial event extraction model.

[0200] Generate a target event extraction model according to the target parameters.

[0201] For example, after obtaining the target parameters, a target event extraction model can be generated according to the target parameters.

[0202] In this embodiment, the generation method further includes:

[0203] Obtain test data including target natural language text;

[0204] Input the test data into the target event extraction model to obtain event extraction information of the test data.

[0205] For example, after obtaining the target event extraction model, the model can be used to process the test data to obtain test results.

[0206] The sixth embodiment of the present application provides a method for generating a training set, including:

[0207] Obtain entity objects in the triple knowledge base.

[0208] Retrieve in the knowledge base according to the entity object to obtain the trigger word corresponding to the entity object.

[0209] Process the trigger word using the remote supervision algorithm to obtain an annotated training set.

[0210] First, detect the core entities in Freebase, sort the entities according to role saliency, event relevance, and key rate, then use all the core entities to back-reference in Wikipedia, and perform trigger word detection according to trigger rate, trigger candidate frequency, and trigger event type frequency. The trigger word list obtained in this stage only contains verbs, lacks nouns, and there is also noise. Therefore, use FrameNet to filter the noise in the verb trigger words and expand the noun trigger words at the same time. Finally, use the remote supervision algorithm (SoftDistant Supervision) to automatically generate the annotated training set data.

[0211] In this embodiment, the method further includes:

[0212] Obtain the updated entity object in the triple knowledge base.

[0213] Retrieve in the knowledge base according to the updated entity object to obtain the updated trigger word corresponding to the entity object.

[0214] Process the updated trigger word using the remote supervision algorithm to obtain an updated annotated training set.

[0215] Similarly, when the triple knowledge base, such as Freebase, is updated, the updated trigger word corresponding to the entity object can be retrieved in the knowledge base according to the updated entity object. Process the updated trigger word using the remote supervision algorithm to obtain an updated annotated training set.

[0216] Although this application is disclosed above with preferred embodiments, it is not used to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be subject to the scope defined by the claims of this application.

[0217] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0218] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash RAM. The memory is an example of computer-readable media.

[0219] Computer-readable media includes both permanent and non-permanent, removable and non-removable media implemented by any method or technology for information storage. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0220] Those skilled in the art will appreciate that the embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A data processing method, characterized in that including: obtaining a target natural language text; obtaining a summary text corresponding to an entity included in the target natural language text; generating a summary embedding vector of the target natural language text according to the summary text corresponding to the entity included in the target natural language text; performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text; wherein, obtaining the summary text corresponding to the entity included in the target natural language text includes: acquiring the entity included in the target natural language text; obtaining the summary text corresponding to the entity in a first knowledge base, and the first knowledge base is a knowledge base recording the summary text.

2. The method according to claim 1, wherein further including: obtaining triple information corresponding to the entity included in the target natural language text from a second knowledge base; the triple information includes a head entity, a relationship between entities, and a tail entity; performing an embedded representation on the triple information to obtain an embedded expression of the entity; performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text includes: performing event extraction processing on the target natural language text according to the summary embedding vector of the target natural language text and the embedded expression of the entity.

3. The method according to claim 2, characterized in that, performing an embedded representation on the triple information to obtain an embedded expression of the entity includes: using a knowledge representation learning model to perform an embedded representation on the triple information to obtain an embedded expression of the entity.

4. The method according to claim 2, wherein the second knowledge base is a knowledge base recording triple information.

5. The method according to claim 1, wherein the obtaining the summary text corresponding to the entity in the first knowledge base according to the entity includes: obtaining summary information of the entity included in the target natural language text according to the entity.

6. The method according to claim 1, wherein the generating a summary embedding vector of the target natural language text according to the summary text corresponding to the entity included in the target natural language text includes: obtaining an embedding vector of the entity included in the target natural language text according to the summary text corresponding to the entity included in the target natural language text; performing summary attention processing on the embedding vectors of all entities included in the target natural language text to generate a summary embedding vector of the target natural language text.

7. The method according to claim 6, wherein the obtaining an embedding vector of the entity included in the target natural language text according to the summary text corresponding to the entity included in the target natural language text includes: encoding each sentence in the summary text corresponding to the entity included in the target natural language text to generate a sentence vector of each sentence; performing encoding processing on all sentence vectors to generate an embedding vector of the entity included in the target natural language text.

8. The method according to claim 1, characterized in that the generating a summary embedding vector of the target natural language text according to the summary text corresponding to the entity included in the target natural language text includes: using at least one of the following networks to encode the summary text corresponding to the entity included in the target natural language text to generate the summary embedding vector of the target natural language text: GRU network; LSTM network; CNN network; Transformer network.

9. A data processing device, characterized in that, including: A target natural language text acquisition unit for acquiring a target natural language text; An abstract text acquisition unit for acquiring an abstract text corresponding to an entity included in the target natural language text; An abstract embedding vector generation unit for generating an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text; An event extraction processing unit for performing event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text; Wherein, the abstract text acquisition unit is further configured to acquire an abstract text corresponding to the entity included in the target natural language text through the following steps: acquire the entity included in the target natural language text; acquire the abstract text corresponding to the entity in a first knowledge base, and the first knowledge base is a knowledge base recording the abstract text.

10. An electronic device, characterized in that, Including: A processor; And A memory for storing a program of a data processing method. After the device is powered on and runs the program of the data processing method through the processor, the following steps are executed: Acquire a target natural language text; Acquire an abstract text corresponding to the entity included in the target natural language text; Generate an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text; Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text; Wherein, acquiring the abstract text corresponding to the entity included in the target natural language text includes: acquiring the entity included in the target natural language text; acquiring the abstract text corresponding to the entity in a first knowledge base, and the first knowledge base is a knowledge base recording the abstract text.

11. A storage device, characterized in that, Store a program of a data processing method, and the program is run by a processor to execute the following steps: including: Acquire a target natural language text; Acquire an abstract text corresponding to the entity included in the target natural language text; Generate an abstract embedding vector of the target natural language text according to the abstract text corresponding to the entity included in the target natural language text; Perform event extraction processing on the target natural language text according to the abstract embedding vector of the target natural language text; Wherein, acquiring the abstract text corresponding to the entity included in the target natural language text: acquire the entity included in the target natural language text; acquire the abstract text corresponding to the entity in a first knowledge base, and the first knowledge base is a knowledge base recording the abstract text.

12. A method for generating an event extraction model, characterized in that, Including: Construct an initial event extraction model, wherein the parameters of the initial event extraction model are initialization data; Acquire training data for training the initial event extraction model, wherein the training data is acquired according to any one of the methods in claims 1-8; Use the training data to train the initial event extraction model to obtain target parameters of the initial event extraction model; Generate a target event extraction model according to the target parameters.

13. The generation method according to claim 12, wherein Further including: Acquire test data including a target natural language text; Input the test data into the target event extraction model to obtain the event extraction information of the test data.

Images