Event Extraction Method, Related Devices and Readable Storage Media

Abstract

The present application discloses an event extraction method, related devices, and readable storage media. In this solution, the text to be subjected to event extraction is input into a parameter extraction model. After processing the text, the parameter extraction model outputs the parameter information corresponding to the text. Based on the parameter information corresponding to the text, at least one event included in the text is determined. In this method, since the parameter extraction model uses the parameter content corresponding to each preset combination of event types and parameter types annotated with training texts as sample labels, joint learning of event detection and parameter recognition is achieved, reducing the impact of cascading errors. Therefore, the accuracy of event extraction results can be improved by adopting this solution.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and more specifically, to an event extraction method, related devices, and readable storage media. Background Art

[0002] An event refers to a series of activities that occur in a specific time and space (time, space), involve one or more roles (event subjects), and revolve around a certain theme. Event extraction (EE) refers to the structured extraction of event parameters. As a subtask of information extraction, its purpose is to extract events of interest to users from unstructured information and present them to users in a structured manner.

[0003] However, using the current event extraction scheme for event extraction will result in low accuracy of event extraction results.

[0004] Therefore, how to improve the accuracy of event extraction results has become a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, this application proposes an event extraction method, related devices, and readable storage media. The specific solutions are as follows:

[0006] An event extraction method, the method includes:

[0007] Obtain the text to be subjected to event extraction;

[0008] Input the text into a parameter extraction model. After the parameter extraction model processes the text, it outputs the parameter information corresponding to the text. The parameter information includes parameter content, parameter type, and event type. The parameter extraction model is trained with training text as training samples and the parameter content corresponding to each preset combination of event types and parameter types annotated in the training text as sample labels.

[0009] Based on the parameter information corresponding to the text, determine at least one event included in the text.

[0010] Optionally, the parameter extraction model includes an encoder and a decoder;

[0011] The encoder is used to encode the text to obtain a feature vector of the text;

[0012] The decoder is used to decode the feature vector of the text to obtain an output vector, and the output vector is used to indicate the parameter content corresponding to the preset combination of event types and parameter types in the text.

[0013] Optionally, determining at least one event included in the text based on the parameter information corresponding to the text includes:

[0014] Combining parameter information with the same event type to obtain at least one candidate event included in the text;

[0015] Determining a to-be-split event from the at least one candidate event;

[0016] Splitting the to-be-split event to obtain split events;

[0017] Determining the event that does not need to be split and the split events as at least one event included in the text.

[0018] Optionally, determining the to-be-split event from the at least one candidate event includes:

[0019] For each candidate event, determining whether the parameter information of the candidate event includes a core parameter type; the core parameter type is a parameter type corresponding to a preset event type, corresponding to multiple parameter contents and different parameter contents corresponding to different events;

[0020] If the parameter information of the candidate event includes a core parameter type and there are at least two parameters corresponding to the core parameter type, then determining the candidate event as the to-be-split event; otherwise, determining the candidate event as a non-to-be-split event.

[0021] Optionally, the parameter corresponding to the core parameter type is a core parameter, the parameter type other than the core parameter type in the parameter information of the candidate event is another parameter type, and the parameter corresponding to the other parameter type is another parameter. Then, splitting the to-be-split event to obtain split events includes:

[0022] For each core parameter, determining the related parameter of the core parameter from the other parameters; the related parameter is a parameter having an association relationship with the core parameter;

[0023] Based on each core parameter and the related parameter of the core parameter, splitting the to-be-split event to obtain split events.

[0024] Optionally, determining the related parameter of the core parameter from the other parameters includes:

[0025] For each other parameter, inputting the feature vector of the text, the core parameter, and the other parameter into a parameter association model, and the parameter association model outputs a parameter association result, and the parameter association result is used to indicate whether the core parameter is related to the other parameter;

[0026] The parameter correlation model is trained with the feature vectors of the training texts, the core parameters for training, and the other parameters for training as training samples, and the other parameters for training labeled by the training texts as sample labels.

[0027] Optionally, the parameter correlation model includes a feature extraction layer, a parameter type encoding layer, a feature fusion layer, and a decoding layer;

[0028] The feature extraction layer is used to extract the feature vectors of the core parameters and the feature vectors of the other parameters from the feature vectors of the texts;

[0029] The parameter type encoding layer is used to encode the parameter types corresponding to the core parameters to obtain the feature vectors of the parameter types corresponding to the core parameters, and to encode the parameter types corresponding to the other parameters to obtain the feature vectors of the parameter types corresponding to the other parameters;

[0030] The feature fusion layer is used to obtain the feature vectors of the parameter types corresponding to the core parameters and the feature vectors of the parameter types corresponding to the other parameters, and fuse the feature vectors of the core parameters, the feature vectors of the other parameters, the feature vectors of the parameter types corresponding to the core parameters, and the feature vectors of the parameter types corresponding to the other parameters to obtain the fused feature vectors;

[0031] The decoding layer is used to decode the fused feature vectors to obtain the parameter correlation results.

[0032] An event extraction device, the device includes:

[0033] An acquisition unit, configured to acquire the text to be subjected to event extraction;

[0034] A parameter information determination unit, configured to input the text into a parameter extraction model, and after the parameter extraction model processes the text, output the parameter information corresponding to the text, where the parameter information includes parameter content, parameter type, and event type; the parameter extraction model is trained with the training texts as training samples and the parameter content corresponding to each preset combination of event types and parameter types labeled by the training texts as sample labels;

[0035] An event determination unit, configured to determine at least one event included in the text based on the parameter information corresponding to the text.

[0036] Optionally, the parameter extraction model includes an encoder and a decoder;

[0037] The encoder is used to encode the text to obtain the feature vector of the text;

[0038] The decoder is used to decode the feature vector of the text to obtain an output vector, and the output vector is used to indicate the parameter content in the text corresponding to the combination of a preset event type and parameter type.

[0039] Optionally, the event determination unit includes:

[0040] A candidate event determination unit, configured to combine parameter information with the same event type to obtain at least one candidate event included in the text;

[0041] A to-be-split event determination unit, configured to determine a to-be-split event from the at least one candidate event;

[0042] An event splitting unit, configured to split the to-be-split event to obtain a split event;

[0043] An event determination subunit, configured to determine the event that does not need to be split and the split event as at least one event included in the text.

[0044] Optionally, the to-be-split event determination unit is specifically configured to:

[0045] For each candidate event, determine whether the parameter information of the candidate event includes a core parameter type; the core parameter type is a parameter type corresponding to a preset event type, which corresponds to multiple parameter contents and different parameter contents correspond to different events; if the parameter information of the candidate event includes a core parameter type and there are at least two parameters corresponding to the core parameter type, then determine the candidate event as a to-be-split event; otherwise, determine the candidate event as a non-to-be-split event.

[0046] Optionally, the parameter corresponding to the core parameter type is a core parameter, the parameter type other than the core parameter type in the parameter information of the candidate event is an other parameter type, and the parameter corresponding to the other parameter type is an other parameter. Then, the event splitting unit includes:

[0047] A relevant parameter determination unit, configured to, for each core parameter, determine a relevant parameter of the core parameter from the other parameters; the relevant parameter is a parameter having an association relationship with the core parameter;

[0048] A splitting unit, configured to split the to-be-split event based on each core parameter and the relevant parameter of the core parameter to obtain a split event.

[0049] Optionally, the relevant parameter determination unit is specifically configured to:

[0050] For each of the other parameters, input the feature vector of the text, the core parameter, and the other parameter into a parameter correlation model, and the parameter correlation model outputs a parameter correlation result, which is used to indicate whether the core parameter is related to the other parameter;

[0051] The parameter correlation model is trained with the feature vector of the training text, the training core parameter, and the training other parameters as training samples, and the training other parameters annotated by the training text as sample labels.

[0052] Optionally, the parameter correlation model includes a feature extraction layer, a parameter type encoding layer, a feature fusion layer, and a decoding layer;

[0053] The feature extraction layer is used to extract the feature vector of the core parameter and the feature vector of the other parameter from the feature vector of the text;

[0054] The parameter type encoding layer is used to encode the parameter type corresponding to the core parameter to obtain the feature vector of the parameter type corresponding to the core parameter, and to encode the parameter type corresponding to the other parameter to obtain the feature vector of the parameter type corresponding to the other parameter;

[0055] The feature fusion layer is used to obtain the feature vector of the parameter type corresponding to the core parameter, the feature vector of the parameter type corresponding to the other parameter, and fuse the feature vector of the core parameter, the feature vector of the other parameter, the feature vector of the parameter type corresponding to the core parameter, and the feature vector of the parameter type corresponding to the other parameter to obtain a fused feature vector;

[0056] The decoding layer is used to decode the fused feature vector to obtain a parameter correlation result.

[0057] An event extraction device includes a memory and a processor;

[0058] The memory is used to store a program;

[0059] The processor is used to execute the program to implement each step of the event extraction method described above.

[0060] A readable storage medium stores a computer program, and when the computer program is executed by a processor, each step of the event extraction method described above is implemented.

[0061] With the above technical solution, the present application discloses an event extraction method, related devices and readable storage media. In this solution, the text to be subjected to event extraction is input into a parameter extraction model. After processing the text, the parameter extraction model outputs the parameter information corresponding to the text. Based on the parameter information corresponding to the text, at least one event included in the text is determined. In this method, since the parameter extraction model uses the parameter content corresponding to each preset combination of event types and parameter types annotated with training texts as sample labels, joint learning of event detection and parameter recognition is achieved, reducing the influence of cascading errors. Therefore, adopting this solution can improve the accuracy of event extraction results. Description of the Drawings

[0062] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present application. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0063] Figure 1 is a schematic flowchart of the event extraction method disclosed in the embodiments of the present application;

[0064] Figure 2 is a schematic structural diagram of the parameter extraction model implemented based on BERT disclosed in the embodiments of the present application;

[0065] Figure 3 is a schematic structural diagram of the parameter association model disclosed in the embodiments of the present application;

[0066] Figure 4 is a schematic structural diagram of an event extraction device disclosed in the embodiments of the present application;

[0067] Figure 5 is a hardware structural block diagram of an event extraction device disclosed in the embodiments of the present application. Detailed Embodiments

[0068] Next, the accompanying drawings in the embodiments of the present application will be used to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0069] First, the present case will elaborate in detail on the reasons why the current event extraction solution for text leads to low accuracy of event extraction results:

[0070] First, the current event extraction solution is two-stage, namely event detection and argument identification. The purpose of event detection is to identify all event types contained in the input text (if there is no event in the text, the event type is empty). The purpose of argument identification is to identify all arguments under the argument type corresponding to the event type. The common practice is to use predefined event types as labels to extract trigger words in the text to obtain trigger words and event types; then use the argument types corresponding to the event types as labels to implement event argument extraction. However, since the best result of current trigger word extraction is only seventy or eighty percent, the event arguments extracted based on such trigger word extraction results will have cascading errors, which will lead to low accuracy of event extraction results.

[0071] In some cases, there are often complex events in the text, such as argument overlap, argument nesting, and multi-argument association. Among them, argument overlap means that there are multiple events in the text, and different events share the same argument. Argument nesting means that there is a nesting problem between the arguments of different events in the text. Multi-argument association means that one argument type corresponds to multiple different arguments.

[0072] For easy understanding, please refer to Table 1, which is an example of complex events provided by the embodiments of the present application.

[0073] Table 1

[0074]

[0075] As shown in Table 1, in the example of argument overlap, "Raptors" is the same argument shared by different events. In the example of argument nesting, there is a nesting between the argument "Apple" and the argument "Some models of Apple computers". In the example of multi-argument association, the argument type "financial report period" corresponds to two different arguments, "the third quarter" and "the first three quarters".

[0076] For the above complex events, the current event extraction solution cannot handle them well. Using the current event extraction solution to extract events from the text will lead to low accuracy of event extraction results. For example, the current event extraction method can only assign one label to each word, so it cannot solve the problems of argument overlap and argument nesting. For example, the label of "Raptors" can only be "winner" or "advancing party"; another example is that for the example of multi-argument association, the trigger words of the two events it contains are both "loss" at the same position in the original text. In this case, the current event extraction solution only regards it as one event, that is, the event of loss. Its argument types are financial report period and loss amount. Among them, the arguments corresponding to the argument type of financial report period are "the first three quarters" and "the third quarter", and the arguments corresponding to the argument type of loss amount are "9.31 million yuan" and "5.2049 million yuan". It cannot distinguish the loss amounts corresponding to the two financial report periods.

[0077] For the above reasons, using the current event extraction solution for event extraction will result in a low accuracy of the event extraction results.

[0078] In view of the above problems existing in the current event extraction solution, the inventors of this case conducted in-depth research and finally proposed a new event extraction solution. Next, the event extraction method provided in this application will be introduced through the following embodiments.

[0079] Refer to Figure 1 , Figure 1 which is a schematic flowchart of the event extraction method disclosed in the embodiments of this application. The method may include:

[0080] Step S101: Obtain the text to be subjected to event extraction.

[0081] In this application, the text to be subjected to event extraction may be a sentence-level text or a passage-level text, and this application does not make any limitation thereto.

[0082] Step S102: Input the text into a parameter extraction model. After processing the text, the parameter extraction model outputs the parameter information corresponding to the text.

[0083] It should be noted that the parameter extraction model is trained with training texts as training samples and the parameter contents corresponding to the preset combinations of various event types and parameter types annotated in the training texts as sample labels; wherein, the training texts are texts in the same field as the text to be subjected to event extraction. The parameter information includes parameter content, parameter type, and event type. As an implementable manner, the parameter information includes the combination of event type and parameter type, and the parameter content corresponding to the combination of the event type and parameter type.

[0084] It should be noted that in this application, when performing annotation, the combination of "event type" and "parameter type" is used as the sample label, realizing the joint learning of the steps of "event detection" and "parameter recognition", and reducing the influence of cascading errors.

[0085] In this application, multiple event types can be specified in advance according to the scenario requirements, and several parameter types are defined for each event type, and the corresponding annotation criteria are determined. Annotators annotate a certain number of texts as training texts for training the parameter extraction model.

[0086] For easy understanding, please refer to Table 2, which is an example of event types and parameter types in the Chinese financial field provided in this application.

[0087] Table 2

[0088]

[0089] As shown in Table 2, the first column is a set of predefined event types, such as company listing, shareholder increase in holdings, etc. The subsequent columns represent the parameter types corresponding to the event types. For example, the parameter types corresponding to the event type of "company listing" include "listed company, stock code, disclosure time, etc.".

[0090] Please refer to Table 3, which is an example of a training text in the Chinese financial field provided by an embodiment of the present application.

[0091] Table 3

[0092]

[0093] Step S103: Based on the parameter information corresponding to the text, determine at least one event included in the text.

[0094] After determining the parameter information corresponding to the text, different parameter information can be clustered to determine at least one event included in the text. The specific implementation will be described in detail in the following embodiments.

[0095] In this embodiment, an event extraction method is disclosed. In this method, the text to be subjected to event extraction is input into a parameter extraction model. After the parameter extraction model processes the text, it outputs the parameter information corresponding to the text. Based on the parameter information corresponding to the text, at least one event included in the text is determined. In this method, since the parameter extraction model uses the parameter content corresponding to the preset combination of each event type and parameter type annotated with the training text as the sample label, the joint learning of event detection and parameter recognition is realized, and the influence of cascading errors is reduced. Therefore, using this method can improve the accuracy of the event extraction result.

[0096] In another embodiment of the present application, the structure of the parameter extraction model is described. The parameter extraction model may include: an encoder and a decoder; wherein, the encoder is used to encode the text to obtain the feature vector of the text; the decoder is used to decode the feature vector of the text to obtain an output vector, and the output vector is used to indicate the parameter content corresponding to the preset combination of event type and parameter type in the text.

[0097] It should be noted that in the present application, the encoder can be implemented based on the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers).

[0098] For ease of understanding, refer to Figure 2 , Figure 2Schematic diagram of the parameter extraction model implemented based on BERT disclosed in the embodiments of the present application.

[0099] After the original text is input into the BERT Encoder, the output vector obtained is h N , and after passing through the Decoder, the dimension of the output vector is R×L×2, where R represents the number of combinations of labels (event type, parameter type), L is the length of the input sequence, and each sequence is represented by two R×L vectors.

[0100] The whole process can be modeled as:

[0101]

[0102]

[0103] Among them, and respectively represent the probabilities of the start (start) and end (end) of the i-th word in the sentence corresponding to the event parameter under the label l j , and j ∈ (1, 2,..., R). If the probability is greater than a pre-set threshold, the corresponding word mark in the figure is set to 1, otherwise it is set to 0. x i represents the vector representation corresponding to the i-th token in h N . and are the weight matrix and bias vector of the decoder respectively, and their weights will be continuously updated during the training process. σ is the Sigmoid activation function.

[0104] As Figure 2 shown, the input text contains a product-related event, and the event type of this event is "recall". Under the label (recall, recall party), the parameter "Apple" can be found, and here "Apple" represents a company name; under the label (recall, recall content), the parameter "some models of Apple computers" can be obtained, and here "Apple" represents a product. Therefore, in this event, the two words "Apple" correspond to multiple different labels. It can be seen that the parameter extraction module proposed in this solution can solve the problems of parameter overlap and nesting.

[0105] After obtaining all possible parameter information in the text, the common practice is to use the "event type" as the criterion for distinguishing different events, and formulate post-processing rules based on the event type to achieve the distinction of different events. In another embodiment of the present application, a specific implementation manner for determining at least one event included in the text based on the parameter information corresponding to the text in step S103 is described. This manner may include: combining parameter information with the same event type to obtain at least one event included in the text.

[0106] For ease of understanding, referring to Table 4, Table 4 shows parameter information with the event type of "pledge".

[0107] Table 4

[0108]

[0109] The event obtained after post-processing the content of Table 4 based on the event type is:

[0110] Event type: Pledge

[0111] Pledgor: Zhang ××, Company to which the pledged property belongs: Fang × Pharmaceutical, Event time: As of the disclosure date of this announcement, Event time: July 7th, Number of pledged shares: Approximately 140 million, Number of pledged shares: 19.9 million.

[0112] However, in actual scenarios, examples where different events correspond to the same event type are common. As shown in Table 4, multiple events correspond to the same event type of "pledge". Therefore, only using whether the "event type" is the same as the criterion for distinguishing events cannot distinguish two events.

[0113] To solve the above problems, in another embodiment of the present application, a specific implementation manner for determining at least one event included in the text based on the parameter information corresponding to the text in step S103 is described. This manner may include the following steps:

[0114] Step S201: Combine parameter information with the same event type to obtain at least one candidate event included in the text.

[0115] For example, the candidate event obtained after processing the content of Table 4 using this step is:

[0116] Event type: Pledge

[0117] Pledgor: Zhang ××, Company to which the pledged property belongs: Fang × Pharmaceutical, Event time: As of the disclosure date of this announcement, Event time: July 7th, Number of pledged shares: Approximately 140 million, Number of pledged shares: 19.9 million.

[0118] Step S202: Determine the event to be split from the at least one candidate event.

[0119] As an implementable manner, the determining the event to be split from the at least one candidate event includes: for each candidate event, determining whether the parameter information of the candidate event includes a core parameter type; if the parameter information of the candidate event includes a core parameter type and there are at least two parameters corresponding to the core parameter type, then determining the candidate event as the event to be split; otherwise, determining the candidate event as a non-event to be split.

[0120] In this application, for each predefined event type, a core parameter type (Corearguments role, CAR) can be selected as the standard for distinguishing different events. The core parameter type is a parameter type corresponding to a predefined event type, in which there are multiple parameter contents and different parameter contents correspond to different events. The parameter corresponding to the core parameter type is a core parameter (Core arguments, CA), and the parameter types other than the core parameter type in the parameter information of the candidate event are other parameter types, and the parameters corresponding to the other parameter types are other parameters (Remain arguments, RA).

[0121] For easy understanding, taking Table 4 as an example, it contains two events and the event types of the two events are the same. However, in the two events, the parameters corresponding to "pledged stock / share quantity" are different, which are "about 140 million" and "19.9 million" respectively. Therefore, for the event type "pledge", the CAR can be "pledged stock / share quantity", and the parameter corresponding to it is called the core parameter, and the parameters corresponding to non-CAR (company to which the pledged property belongs, pledgor, event time) are called other parameters.

[0122] Step S203: Split the event to be split to obtain the split events.

[0123] As an implementable manner, the splitting the event to be split to obtain the split events includes: for each core parameter, determining the relevant parameter of the core parameter from the other parameters; the relevant parameter is a parameter having an association relationship with the core parameter; based on each core parameter and the relevant parameter of the core parameter, splitting the event to be split to obtain the split events. Specifically, the event type corresponding to the core parameter, the parameter type corresponding to the core parameter, the parameter type corresponding to the relevant parameter, the core parameter, and the relevant parameter of the core parameter can be determined as the same event.

[0124] For ease of understanding, taking Table 4 as an example, the relevant parameters for "19.9 million" are determined to be {Fang × Pharmaceutical, Zhang ××, July 7th}, and the relevant parameters for "about 140 million" are {Fang × Pharmaceutical, Zhang ××, as of the disclosure date of this announcement}. Then, the split Event 1 is:

[0125] Event type: Pledge

[0126] Pledger: Zhang ××, Company to which the pledged property belongs: Fang × Pharmaceutical, Event time: as of the disclosure date of this announcement, Number of pledged shares: about 140 million.

[0127] The split Event 2 is:

[0128] Event type: Pledge

[0129] Pledger: Zhang ××, Company to which the pledged property belongs: Fang × Pharmaceutical, Event time: July 7th, Number of pledged shares: 19.9 million.

[0130] Step S204: Determine the events that do not need to be split and the split events as at least one event included in the text.

[0131] In another embodiment of the present application, a specific implementation manner for determining the relevant parameters of the core parameter from the above other parameters is introduced. This manner may include: for each other parameter, input the feature vector of the text, the core parameter, and the other parameter into a parameter association model. The parameter association model outputs a parameter association result, and the parameter association result is used to indicate whether the core parameter is related to the other parameter. The parameter association model is trained with the feature vector of the training text, the training core parameter, and the training other parameter as training samples, and the training other parameter labeled with the training text as a sample label.

[0132] In another embodiment of the present application, the structure of the parameter association model is described. Refer to Figure 3 , Figure 3 which is a schematic structural diagram of the parameter association model disclosed in the embodiments of the present application. The parameter association model may include a feature extraction layer, a parameter type encoding layer, a feature fusion layer, and a decoding layer.

[0133] Among them, the feature extraction layer is used to extract the feature vector of the core parameter and the feature vector of the other parameter from the feature vector of the text;

[0134] The parameter type encoding layer is used to encode the parameter type corresponding to the core parameter to obtain the feature vector of the parameter type corresponding to the core parameter, and, encode the parameter type corresponding to the other parameter to obtain the feature vector of the parameter type corresponding to the other parameter;

[0135] The feature fusion layer is used to obtain the feature vectors of the parameter types corresponding to the core parameters and the feature vectors of the parameter types corresponding to the other parameters, and fuse the feature vectors of the core parameters, the feature vectors of the other parameters, the feature vectors of the parameter types corresponding to the core parameters, and the feature vectors of the parameter types corresponding to the other parameters to obtain the fused feature vectors;

[0136] The decoding layer is used to decode the fused feature vectors to obtain the parameter association result.

[0137] Specifically, during the training process of the parameter association model, the vector at the start position of the feature vector h of the core parameter in the text N and the vector at the end position are combined to obtain the feature vector of the core parameter The vector at the start position of the feature vector h of the other parameter in the text and the vector at the end position N are combined to obtain the feature vector of the other parameter and the vector at the end position are combined to obtain the feature vector of the other parameter Then, the feature vectors of the parameter types corresponding to the core parameters are obtained The feature vectors of the parameter types corresponding to the other parameters It should be noted that the feature vectors of the parameter types corresponding to the core parameters The feature vectors of the parameter types corresponding to the other parameters are obtained by random initialization, with the same dimension as the feature vectors of the core parameters and the feature vectors of the other parameters, and are updated during the training phase.

[0138] The whole process can be modeled as:

[0139]

[0140]

[0141] Among them, and respectively represent the probability values that the i-th token in the sentence is the start (start) and end (end) of the predicted parameter. x i represents the vector representation corresponding to the i-th token in h N . and They are the weight matrix and bias vector of the decoding layer, and their weights will be continuously updated during the training process. σ is the Sigmoid activation function. If the probability is greater than a pre-set threshold, the label of the corresponding word in the graph is set to 1, otherwise it is set to 0. If the predicted value is consistent with the RA input to the model, it indicates that CA is related to RA, otherwise it is not related.

[0142] It should be noted that in this application, the parameter extraction model and the parameter association model can be jointly trained. When jointly training the parameter extraction model and the parameter association model, the binary cross-entropy loss function (Binary Cross Entropy Loss, BCE Loss) can be used as the loss function for training. Denote the target loss function for training the parameter extraction model as L model1 , and the target loss function for training the parameter association model as L model2 . Then the total loss in the training stages of the two models is L total , then: L total = L model1 + L model2

[0143] The optimization function can be selected as stochastic gradient descent or Adam, etc.

[0144] Next, the event extraction device disclosed in the embodiments of the present application will be described. The event extraction device described below can be correspondingly referred to the event extraction method described above.

[0145] Refer to Figure 4 , Figure 4 which is a schematic structural diagram of an event extraction device disclosed in an embodiment of the present application. As Figure 4 shown, the event extraction device may include:

[0146] An acquisition unit 41, configured to acquire the text to be subjected to event extraction;

[0147] A parameter information determination unit 42, configured to input the text into a parameter extraction model. After the parameter extraction model processes the text, it outputs the parameter information corresponding to the text. The parameter information includes parameter content, parameter type, and event type. The parameter extraction model is trained with training texts as training samples and the parameter content corresponding to each preset combination of event types and parameter types annotated in the training texts as sample labels;

[0148] An event determination unit 43, configured to determine at least one event included in the text based on the parameter information corresponding to the text.

[0149] As an implementable manner, the parameter extraction model includes an encoder and a decoder;

[0150] The encoder is used to encode the text to obtain a feature vector of the text;

[0151] The decoder is used to decode the feature vector of the text to obtain an output vector, and the output vector is used to indicate parameter content corresponding to a combination of a preset event type and parameter type in the text.

[0152] As an implementable manner, the event determination unit includes:

[0153] A candidate event determination unit, configured to combine parameter information with the same event type to obtain at least one candidate event included in the text;

[0154] An event to be split determination unit, configured to determine an event to be split from the at least one candidate event;

[0155] An event splitting unit, configured to split the event to be split to obtain a split event;

[0156] An event determination subunit, configured to determine the event that does not need to be split and the split event as at least one event included in the text.

[0157] As an implementable manner, the event to be split determination unit is specifically configured to:

[0158] For each candidate event, determine whether the parameter information of the candidate event includes a core parameter type; the core parameter type is a parameter type corresponding to a preset event type, in which there are multiple parameter contents and different parameter contents correspond to different events; if the parameter information of the candidate event includes a core parameter type and there are at least two parameters corresponding to the core parameter type, then determine the candidate event as an event to be split; otherwise, determine the candidate event as a non-event to be split.

[0159] As an implementable manner, the parameter corresponding to the core parameter type is a core parameter, the parameter type other than the core parameter type in the parameter information of the candidate event is an other parameter type, and the parameter corresponding to the other parameter type is an other parameter. Then, the event splitting unit includes:

[0160] A relevant parameter determination unit, configured to, for each core parameter, determine a relevant parameter of the core parameter from the other parameters; the relevant parameter is a parameter having an association relationship with the core parameter;

[0161] A splitting unit, configured to split the event to be split based on each core parameter and the relevant parameter of the core parameter to obtain a split event.

[0162] As an implementable manner, the relevant parameter determination unit is specifically configured to:

[0163] For each other parameter, input the feature vector of the text, the core parameter, and the other parameter into a parameter correlation model, and the parameter correlation model outputs a parameter correlation result, where the parameter correlation result is used to indicate whether the core parameter is relevant to the other parameter;

[0164] The parameter correlation model is trained with the feature vector of the training text, the training core parameter, and the training other parameter as training samples, and the training other parameter labeled by the training text as a sample label.

[0165] As an implementable manner, the parameter correlation model includes a feature extraction layer, a parameter type encoding layer, a feature fusion layer, and a decoding layer;

[0166] The feature extraction layer is used to extract the feature vector of the core parameter and the feature vector of the other parameter from the feature vector of the text;

[0167] The parameter type encoding layer is used to encode the parameter type corresponding to the core parameter to obtain the feature vector of the parameter type corresponding to the core parameter, and to encode the parameter type corresponding to the other parameter to obtain the feature vector of the parameter type corresponding to the other parameter;

[0168] The feature fusion layer is used to obtain the feature vector of the parameter type corresponding to the core parameter and the feature vector of the parameter type corresponding to the other parameter, and fuse the feature vector of the core parameter, the feature vector of the other parameter, the feature vector of the parameter type corresponding to the core parameter, and the feature vector of the parameter type corresponding to the other parameter to obtain a fused feature vector;

[0169] The decoding layer is used to decode the fused feature vector to obtain a parameter correlation result.

[0170] Refer to Figure 5 , Figure 5 which is the hardware structure block diagram of the event extraction device provided by the embodiment of the present application. Refer to Figure 5 ,the hardware structure of the event extraction device may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4;

[0171] In the embodiment of the present application, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 complete communication with each other through the communication bus 4;

[0172] The processor 1 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, etc.;

[0173] The memory 3 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0174] Wherein, the memory stores a program, and the processor can call the program stored in the memory, and the program is used for:

[0175] Obtain the text to be subjected to event extraction;

[0176] Input the text into a parameter extraction model, and after processing the text, the parameter extraction model outputs parameter information corresponding to the text, where the parameter information includes parameter content, parameter type, and event type; the parameter extraction model is trained with training text as training samples and the parameter content corresponding to the preset combination of each event type and parameter type marked by the training text as sample labels;

[0177] Based on the parameter information corresponding to the text, determine at least one event included in the text.

[0178] Optionally, the refinement function and expansion function of the program can be referred to the above description.

[0179] The embodiment of the present application further provides a readable storage medium, which can store a program suitable for execution by a processor, and the program is used for:

[0180] Obtain the text to be subjected to event extraction;

[0181] Input the text into a parameter extraction model, and after processing the text, the parameter extraction model outputs parameter information corresponding to the text, where the parameter information includes parameter content, parameter type, and event type; the parameter extraction model is trained with training text as training samples and the parameter content corresponding to the preset combination of each event type and parameter type marked by the training text as sample labels;

[0182] Based on the parameter information corresponding to the text, determine at least one event included in the text.

[0183] Optionally, the refinement function and expansion function of the program can be referred to the above description.

[0184] Finally, it should also be noted that in this text, relational 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 any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0185] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0186] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An event extraction method, characterized in that, the method includes: Obtain the text to be subjected to event extraction; Input the text into a parameter extraction model. After processing the text, the parameter extraction model outputs parameter information corresponding to the text. The parameter information includes parameter content, parameter type, and event type. The parameter extraction model is trained with training text as training samples and the parameter content corresponding to each preset combination of event types and parameter types annotated in the training text as sample labels; Combine the parameter information with the same event type to obtain at least one candidate event included in the text; Determine the event to be split from the at least one candidate event; Split the event to be split to obtain the split event; Determine the event that does not need to be split and the split event as at least one event included in the text.

2. The method according to claim 1, characterized in that, the parameter extraction model includes an encoder and a decoder; The encoder is used to encode the text to obtain a feature vector of the text; The decoder is used to decode the feature vector of the text to obtain an output vector, and the output vector is used to indicate the parameter content corresponding to the preset combination of event types and parameter types in the text.

3. The method according to claim 1, characterized in that, the determining the event to be split from the at least one candidate event includes: For each candidate event, determine whether the parameter information of the candidate event includes a core parameter type. The core parameter type is a parameter type corresponding to the preset event type, in which there are multiple parameter contents and different parameter contents correspond to different events; If the parameter information of the candidate event includes a core parameter type and there are at least two parameters corresponding to the core parameter type, then determine the candidate event as the event to be split; otherwise, determine the candidate event as a non-event to be split.

4. The method according to claim 3, characterized in that, the parameter corresponding to the core parameter type is the core parameter, the parameter type other than the core parameter type in the parameter information of the candidate event is the other parameter type, and the parameter corresponding to the other parameter type is the other parameter. Then, the splitting the event to be split to obtain the split event includes: For each core parameter, determine the relevant parameter of the core parameter from the other parameters. The relevant parameter is a parameter having an association relationship with the core parameter; Based on each core parameter and the relevant parameter of the core parameter, split the event to be split to obtain the split event.

5. The method according to claim 4, characterized in that, the determining the relevant parameter of the core parameter from the other parameters includes: For each other parameter, input the feature vector of the text, the core parameter, and the other parameter into a parameter association model. The parameter association model outputs a parameter association result, and the parameter association result is used to indicate whether the core parameter is relevant to the other parameter; The parameter association model is trained using the feature vectors of the training texts, the core parameters for training, and the other parameters for training as training samples, and the other parameters for training labeled by the training texts as sample labels.

6. The method according to claim 5, wherein, the parameter association model includes a feature extraction layer, a parameter type encoding layer, a feature fusion layer, and a decoding layer; the feature extraction layer is configured to extract the feature vectors of the core parameters and the feature vectors of the other parameters from the feature vectors of the texts; the parameter type encoding layer is configured to encode the parameter types corresponding to the core parameters to obtain the feature vectors of the parameter types corresponding to the core parameters, and encode the parameter types corresponding to the other parameters to obtain the feature vectors of the parameter types corresponding to the other parameters; the feature fusion layer is configured to obtain the feature vectors of the parameter types corresponding to the core parameters, the feature vectors of the parameter types corresponding to the other parameters, and fuse the feature vectors of the core parameters, the feature vectors of the other parameters, the feature vectors of the parameter types corresponding to the core parameters, and the feature vectors of the parameter types corresponding to the other parameters to obtain the fused feature vectors; the decoding layer is configured to decode the fused feature vectors to obtain the parameter association result.

7. An event extraction device, wherein, the device includes: an acquisition unit configured to acquire the text to be subjected to event extraction; a parameter information determination unit configured to input the text into a parameter extraction model, and after the parameter extraction model processes the text, output the parameter information corresponding to the text, where the parameter information includes parameter content, parameter type, and event type; the parameter extraction model is trained using the training texts as training samples and the parameter content corresponding to each preset combination of event types and parameter types labeled by the training texts as sample labels; an event determination unit configured to determine at least one event included in the text based on the parameter information corresponding to the text; wherein, the event determination unit includes: a candidate event determination unit configured to combine the parameter information with the same event type to obtain at least one candidate event included in the text; a to-be-split event determination unit configured to determine the to-be-split event from the at least one candidate event; an event splitting unit configured to split the to-be-split event to obtain the split events; an event determination subunit configured to determine the events that do not need to be split and the split events as at least one event included in the text.

8. An event extraction device, wherein, it includes a memory and a processor; the memory is configured to store a program; the processor is configured to execute the program to implement each step of the event extraction method according to any one of claims 1 to 6.

9. A readable storage medium, on which a computer program is stored, wherein, when the computer program is executed by a processor, it implements each step of the event extraction method according to any one of claims 1 to 6.

Images