Event extraction method, event extraction network training method and device
Through the event extraction network, the natural language text is divided into multiple sub-events and the tag information is merged, which solves the problem of low event extraction efficiency and accuracy in the prior art, and achieves efficient and accurate event recognition.
Patent Information
- Application Number
- CN202011537728.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-23
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2040-12-23
AI Technical Summary
The existing event extraction technology is inefficient and has low accuracy, especially in the fields of finance and public opinion monitoring. Manual news and announcements are time-consuming and labor-intensive, and two-stage tasks based on pattern matching or neural networks lead to insufficient event extraction efficiency and accuracy. .
Through event extraction network, the natural language text of a single event is divided into multiple non-conflicting sub-events, and the label information of the sub-events is merged to directly output event identification information, avoiding the complexity of pattern construction and two-stage recognition, and improving the Efficiency and accuracy.
It realizes efficient and accurate event extraction, simplifies the event extraction process, improves the work efficiency in fields such as finance and public opinion monitoring, and accurately identify a single event in the presence of multiple event arguments or overlaps.
Smart Images

Figure CN114662481B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of natural language processing technology, and in particular to an event extraction method, a training method, a device, an electronic device and a storage medium for an event extraction network. Background Art
[0002] Event extraction technology is an important basic technology for understanding language, and is widely used in finance, public opinion monitoring, data mining, and text mining. Before the application of event extraction technology, financial practitioners usually collected news and / or announcements manually, and it was very time-consuming and laborious to find some risk signals or marketing opportunities from long news or announcements; event extraction technology can directly extract important events such as company names from news and / or announcements, which can greatly improve work efficiency.
[0003] In related technologies, event extraction technology usually extracts events based on pattern matching, or divides event extraction into two-stage tasks of event type recognition and argument role recognition based on neural networks to obtain event recognition information. However, in the above schemes, on the one hand, the event extraction mode setting process is relatively cumbersome, resulting in low event extraction efficiency; on the other hand, the event extraction accuracy is low. Summary of the invention
[0004] The embodiments of the present invention provide an event extraction method, a training method and device of an event extraction network, an electronic device and a storage medium.
[0005] The technical solution of the embodiment of the present invention is achieved as follows:
[0006] In a first aspect, an embodiment of the present invention provides an event extraction method, comprising:
[0007] Inputting the natural language text of a single event into an event extraction network, and obtaining labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other;
[0008] The event arguments and event roles described by the tags of the multiple sub-events are combined to obtain the event identification information of the single event.
[0009] In a second aspect, an embodiment of the present invention provides a training method for an event extraction network, comprising:
[0010] Acquire a training sample set of natural language text and label information of the training sample set; the label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other;
[0011] Inputting the text of the training sample set into the initial network to obtain the predicted label output by the initial network; the predicted label is formed by the event arguments and event roles corresponding to the multiple sub-events of a single event;
[0012] Determining a loss function value of the initial network according to a difference between the predicted label and the label information;
[0013] If the loss function value does not meet the training stop condition, continue training the initial network.
[0014] In a third aspect, an embodiment of the present invention provides an event extraction device, the device comprising:
[0015] An event extraction module, used to input the natural language text of a single event into an event extraction network, and obtain labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other;
[0016] The merging module is used to merge the event arguments and event roles described by the labels of the multiple sub-events to obtain the event identification information of the single event.
[0017] In a fourth aspect, an embodiment of the present invention provides a training device for an event extraction network, the training device comprising:
[0018] An acquisition module, used to acquire a training sample set of natural language text and label information of the training sample set; the label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other;
[0019] Network acquisition module, used to:
[0020] Inputting the text of the training sample set into the initial network to obtain the predicted label output by the initial network; the predicted label is formed by the event arguments and event roles corresponding to the multiple sub-events of a single event;
[0021] Determining a loss function value of the initial network according to a difference between the predicted label and the label information;
[0022] If the loss function value does not meet the training stop condition, continue training the initial network.
[0023] In a fifth aspect, an embodiment of the present invention provides an electronic device, including:
[0024] A memory for storing executable instructions;
[0025] The processor is used to implement the event extraction method provided by one or more of the above-mentioned technical solutions when executing the executable instructions stored in the memory.
[0026] In a sixth aspect, an embodiment of the present invention provides an electronic device, including:
[0027] A memory for storing executable instructions;
[0028] The processor is used to implement the training method of the event extraction network provided by one or more of the above-mentioned technical solutions when executing the executable instructions stored in the memory.
[0029] In the seventh aspect, an embodiment of the present invention provides a computer-readable storage medium, which stores computer-executable instructions; after the computer-executable instructions are executed by a processor, the event extraction method provided by one or more of the aforementioned technical solutions can be implemented.
[0030] In an eighth aspect, an embodiment of the present invention provides a computer-readable storage medium, which stores computer-executable instructions; after the computer-executable instructions are executed by a processor, the training method of the event extraction network provided by one or more of the aforementioned technical solutions can be implemented.
[0031] The embodiment of the present invention provides an event extraction method, a training method, a device, an electronic device and a storage medium for an event extraction network. By inputting the natural language text of a single event into the event extraction network, labels of multiple sub-events in the single event are obtained, and the labels of the multiple sub-events do not conflict with each other; by merging the event arguments and event roles described by the labels of the multiple sub-events, event identification information of the single event is obtained.
[0032] First, the present invention does not need to construct an event extraction model and perform event extraction based on pattern matching. Instead, it directly outputs labels of multiple sub-events in a single event through an event extraction network, and obtains event identification information of a single event by merging event arguments and event roles described by the labels, thereby improving event extraction efficiency.
[0033] Secondly, the present invention outputs labels of multiple sub-events in a single event through an event extraction network, and determines the event identification information of a single event by merging the event arguments and event roles described by the labels of multiple non-conflicting sub-events, so that when there are multiple event arguments in a single event and the event arguments overlap or partially overlap, event extraction can be accurately performed on the single event. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1It is a flowchart of an event extraction method provided by an embodiment of the present invention;
[0035] Figure 2 yes Figure 1 A detailed flowchart of step 101 in the corresponding embodiment;
[0036] Figure 3 It is a flow chart of a training method for an event extraction network provided by an embodiment of the present invention;
[0037] Figure 4 for Figure 3 Detailed flowchart of step 301 in the corresponding embodiment;
[0038] Figure 5 is a schematic diagram of the structure of an event extraction device provided by an embodiment of the present invention;
[0039] Figure 6 It is a structural schematic diagram of a training device for an event extraction network provided by an embodiment of the present invention;
[0040] Figure 7 A method for constructing a role conflict matrix provided by an embodiment of the present invention;
[0041] Figure 8 is a schematic diagram of a conflict matrix update provided by an embodiment of the present invention;
[0042] Fig. 9 It is a schematic diagram of an event extraction network architecture provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0044] There are three main event extraction methods:
[0045] (1) Meta-event extraction based on pattern matching. This mainly refers to defining various event extraction patterns before extraction, and finding information about pattern constraints based on the pattern matching algorithm. The core of this method is the construction of extraction patterns, but the construction of patterns requires the guidance of experts who have both event type domain knowledge and natural language knowledge, which consumes a lot of manpower and time, resulting in low event extraction efficiency.
[0046] (2) Event extraction based on machine learning or deep learning. It can be mainly divided into two stages: event type recognition and argument role recognition. Usually, events are detected based on trigger word recognition, and argument roles are recognized using methods such as named entities. Finally, the results of the two tasks are combined to complete the event extraction task. However, this method requires two recognitions, and the final event extraction result will be gradually amplified due to the errors in the event type recognition task, resulting in a low event extraction accuracy.
[0047] (3) Joint learning of two tasks based on machine learning or deep learning. The machine learning method mainly uses a structured perceptron model to perform joint learning of two tasks, and uses a column search strategy to search for the optimal result. The deep learning method avoids a large number of artificial features, uses a neural network model for feature learning and joint learning, and avoids complex feature engineering. However, this method cannot avoid the problem of multiple events and the overlap or partial overlap of arguments between events, and the accuracy of event extraction is low.
[0048] The embodiment of the present invention provides an event extraction method. Figure 1 is a flow chart of an event extraction method provided by an embodiment of the present invention, such as Figure 1 As shown, the following steps are included:
[0049] Step 101, inputting the natural language text of a single event into an event extraction network, and obtaining labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other;
[0050] Step 102: merge the event arguments and event roles described by the tags of the multiple sub-events to obtain event identification information of the single event.
[0051] In the embodiment of the present invention, an event refers to a form of information, which is defined as an objective fact that a specific subject (the subject includes but is not limited to: a person, an object or a phenomenon) interacts at a specific time and a specific place, generally at the sentence level. The elements that make up an event include a trigger word (Trigger), an event type (Event Type), an event argument (Argument) and an event role (Role). Among them, the trigger word is a core word indicating the occurrence of an event, which is used to identify the event type, mostly a verb or a noun. The event type refers to the category to which different events belong, which can be determined by the event argument and the trigger word; the Automatic Content Extraction Evaluation (ACE) 2005 conference defined 8 event types and 33 subtypes. The event argument refers to the participant of the event, generally an entity. The event role refers to the role played by the event argument in the event.
[0052] In practical applications, the natural language text of a single event is input into a pre-trained event extraction network, which divides the single event into multiple non-conflicting sub-events and outputs the labels of the multiple sub-events.
[0053] Here, the natural language text of a single event can be any type of natural language used to describe a single event; for example, a report describing a company's sales in a quarter, or a news report describing a sports event. The event extraction network is a pre-trained neural network, which may include a feature extraction layer and a classification layer, which are used to extract features from the text and identify arguments and roles, respectively.
[0054] Since a single event contains multiple event roles, the event arguments corresponding to the event roles may overlap or partially overlap. The single event is divided into multiple sub-events through the event extraction network, and the event arguments in the multiple sub-events do not conflict with each other. Then, event extraction is performed on the multiple sub-events respectively to obtain labels for the multiple sub-events.
[0055] Here, the event arguments in the plurality of sub-events not conflicting with each other refers to a situation where the event arguments corresponding to the event roles included in the sub-events do not overlap.
[0056] For example, the text of a single event "The collapse of the Qingdao subway construction section caused 3 deaths and 2 missing persons are still being searched and rescued" contains three event roles: "disaster / accident-collapse / collapse-collapse subject", "disaster / accident-collapse / collapse-number of deaths" and "life-death-location", and the event arguments corresponding to the event role "disaster / accident-collapse / collapse-collapse subject" and the event role "life-death-location" overlap. The single event is divided into 2 sub-events through the event extraction network. The event roles contained in sub-event 1 are "disaster / accident-collapse / collapse-collapse subject" and "disaster / accident-collapse / collapse-number of deaths", and the event role contained in sub-event 2 is "life-death-location"; then the event arguments "Qingdao subway construction section" and "3 people" corresponding to the event roles in sub-event 1 do not overlap; the event argument "Qingdao" corresponding to the event role in sub-event 2 also does not overlap.
[0057] According to the event roles described by the labels of multiple sub-events, event arguments corresponding to the event roles of the multiple sub-events are merged to obtain event identification information of a single event.
[0058] In this way, the labels of multiple sub-events in a single event are directly output through the event extraction network, and the event identification information of a single event is obtained by merging the event arguments and event roles described by the labels, thereby improving the efficiency of event extraction; and, by dividing a single event task into multiple non-conflicting sub-event tasks, when there are multiple event arguments in a single event and the event arguments overlap or partially overlap, event extraction can be accurately performed on a single event.
[0059] Alternatively, if Figure 2 As shown, Figure 2 yes Figure 1 The detailed flow chart of step 101 in the corresponding embodiment. Step 101 includes:
[0060] Step 1011, preprocessing the natural language text of the single event to obtain a word vector;
[0061] Step 1012, inputting the word vector into the event extraction network;
[0062] Step 1013, extracting features from the word vector through the shared layer in the event extraction network to obtain feature information;
[0063] Step 1014, through a plurality of parallel task layers located at the output end of the shared layer in the event extraction network, arguments and roles are respectively identified based on the feature information to obtain labels of a plurality of sub-events in the single event.
[0064] In practical applications, before extracting events from the natural language text of a single event, the natural language text of the single event needs to be preprocessed. The preprocessing includes word segmentation of the text and vector conversion of the word segmentation to obtain word vectors.
[0065] In step 1013, the feature information may include semantic feature information and context-based associated feature information of each word vector.
[0066] Here, the context-based associated feature information refers to the semantic feature information between the word vector and the associated context word vector.
[0067] In practical applications, the shared layer can be the Chinese pre-trained model RoBERTa or a customized deep network model.
[0068] It should be noted that the RoBERTa model obtains multi-level grammatical and semantic information in the input data by stacking 12 layers of transformer model structure; while the transformer model obtains the context-based associated feature information of each word in the input data through a multi-head attention mechanism.
[0069] In step 1014, the multiple parallel task layers are independent of each other and correspond to each sub-event respectively. The multiple task layers identify arguments and roles based on the feature information of the word vectors, determine the probability that the word vectors correspond to each label, and thus determine the labels corresponding to the multiple sub-events.
[0070] In practical applications, FFN+CRF / softmax can be used as the task layer, where the FNN layer is a two-layer fully connected layer, and the feature information output by the shared layer is subjected to two linear transformations and a ReLU activation through the FNN layer; the feature information is annotated through the CRF layer or the softmax layer to obtain the label of the word vector.
[0071] Optionally, step 1011 includes:
[0072] Filtering illegal characters from the natural language text of the single event;
[0073] The processed natural language text is then segmented, and each word obtained by the segmentation is mapped into a word vector.
[0074] Here, the illegal characters refer to noise characters that interfere with event extraction from the natural language text, such as emoticons, special characters, and unrecognizable garbled information.
[0075] In practical applications, illegal characters in the natural language text of a single event can be removed, and uppercase English letters in the text can be converted to lowercase English letters. The natural language text can also be segmented using predetermined segmentation rules.
[0076] In actual implementation, regular expressions and other methods can be used to remove garbled characters in the text, convert uppercase English letters in the text into lowercase English letters, and segment the text in the same way that Chinese characters and English words form separate words. The segmentation results are converted into vectors according to the word2vec algorithm to obtain word vectors.
[0077] It should be noted that a regular expression is a regular string composed of some pre-defined specific characters and combinations of these specific characters.
[0078] Optionally, step 1014 includes:
[0079] The plurality of parallel task layers respectively select characteristic information associated with the sub-event corresponding to the task layer from the characteristic information;
[0080] For the preset event role label information, arguments and roles are identified based on the characteristic information associated with the sub-event corresponding to the task layer, and labels of event arguments and event roles describing the sub-event are obtained.
[0081] Here, the event role label information is generated according to the event type and the event role corresponding to the event type. For example, the event role label information (ie, event type-event role) can be obtained by directly concatenating the event type and the event role corresponding to the event type.
[0082] The embodiment of the present invention does not limit the determination method and specific format of the event role label information, as long as the determined event role label information can ensure the uniqueness of the event role label information and can directly determine the event type and event role corresponding to the event role label based on the event role label information.
[0083] In practical applications, through multiple parallel task layers, the feature information associated with each sub-event is respectively identified for arguments and roles, and the event extraction task of a single event is split into event extraction tasks of multiple sub-events. Moreover, the event extraction tasks of multiple sub-events are performed independently to avoid the problem of argument conflicts.
[0084] In this way, event extraction is performed on multiple sub-events of a single event through multiple independent task layers, thereby avoiding the situation of identifying argument conflicts and improving the accuracy of event extraction; and, the event role label information is determined by the event type and the event role, and based on the event role label information, the event type and argument role are identified at one time, thereby simplifying the event extraction method.
[0085] Next, an embodiment of the present invention provides a training method for an event extraction network, such as Figure 3 As shown, Figure 3 1 is a flow chart of a training method for an event extraction network provided by an embodiment of the present invention. The method comprises:
[0086] Step 301, obtaining a training sample set of natural language text and label information of the training sample set; the label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other;
[0087] Step 302: input the text of the training sample set into the initial network to obtain the predicted label output by the initial network; the predicted label is formed by the event arguments and event roles corresponding to the multiple sub-events of a single event;
[0088] Step 303, determining a loss function value of the initial network according to the difference between the predicted label and the label information;
[0089] Step 304: If the loss function value does not meet the training stop condition, continue training the initial network.
[0090] In step 301, a large amount of natural language texts are obtained to form a training sample set; and the event type, event arguments, and event roles corresponding to the event arguments contained in each text in the training sample set are annotated to obtain label information of the text.
[0091] In practical applications, the training sample set is the text to be trained in the initial network. The single event corresponding to the text in the training sample set is split into multiple sub-events. Among them, the event arguments between the multiple sub-events do not conflict with each other, and the multiple sub-events are respectively labeled to obtain the label information of the multiple sub-events corresponding to the single event. The label information at least includes the event type and argument role corresponding to the text.
[0092] The training stop condition may be that the training reaches the number of iterations and / or the loss function converges. Here, the number of iterations may be determined according to the number of texts in the training sample set and the size of the initial network, and the present invention does not specifically limit this.
[0093] In an embodiment of the present invention, the text in the training sample is input into the initial network, and the predicted labels of the event types and event roles corresponding to the multiple sub-events of a single event output by the initial network are obtained. And according to the difference between the predicted label and the label information of the text, the loss function value of the initial network is calculated, and the back propagation gradient is calculated according to the loss function value, and the parameters to be trained in the initial network are updated, and then the above steps are repeated until the number of iterations is reached or the loss function converges, and the event extraction network is obtained.
[0094] Exemplarily, when training the initial network, first obtain any training text, preprocess the training text, obtain word vectors, input the word vectors into the initial network, extract features from the word vectors through the shared layer contained in the initial network, and obtain feature information. Multiple parallel task layers select feature information associated with sub-events corresponding to the task layer from the feature information output by the shared layer to identify arguments and roles, and obtain predicted labels of event arguments and event roles corresponding to multiple sub-events of a single event in the description text. Input the predicted label and the label information corresponding to the training text into the loss function of the initial network, calculate the loss function value of the initial network, and determine whether the loss function value meets the stop training condition. If the loss function value does not meet the stop training condition, optimize the parameters to be trained of the initial function, and repeat the above training steps until the loss function value meets the stop training condition, and then the event extraction network can be obtained.
[0095] Alternatively, if Figure 4 As shown, Figure 4 for Figure 3 The detailed flow chart of step 301 in the corresponding embodiment, step 301 includes:
[0096] Step 3011, obtaining a training sample set of natural language text;
[0097] Step 3012, based on the preset event role label information, determine the label information corresponding to the single event in the training sample set;
[0098] Step 3013, constructing a role conflict matrix according to the event arguments and event role labels in the label information corresponding to the single event;
[0099] Step 3014: based on the role conflict matrix, group the label information corresponding to a single event in the training sample set into label information corresponding to the multiple sub-events.
[0100] In the embodiment of the present invention, the preset event role label information is determined according to the event type and the event role corresponding to the event type. For example, the event role label information (i.e., event type-event role) can be obtained by directly concatenating the event type and the event role corresponding to the event type.
[0101] Here, the method and specific format of determining the event role label information are not limited, as long as the determined event role label information can ensure the uniqueness of the event role label information and can directly determine the event type and event role corresponding to the event role label based on the event role label information.
[0102] In practical applications, a single event in a training sample set is labeled according to preset event role label information to obtain all event arguments contained in the single event and the event role labels corresponding to the event arguments.
[0103] In an embodiment of the present invention, a role conflict matrix is constructed based on all event role labels contained in the label information corresponding to the single event; based on the role conflict matrix, the number of conflicts of the event role labels corresponding to the single event is analyzed, and based on the number of conflicts of each event role label, the label information of the single event is grouped to obtain multiple sub-events. One sub-event corresponds to one label information, and the event role labels corresponding to the label information of the sub-event do not conflict with each other.
[0104] In this way, by constructing a role conflict matrix, the event role labels contained in a single event are dynamically grouped into non-conflicting event role labels corresponding to multiple sub-events, thereby splitting a single event extraction task into multiple non-conflicting sub-event tasks to avoid the problem of argument conflict.
[0105] Optionally, the step 3013 includes:
[0106] Constructing the role conflict matrix according to the event role label in the label information corresponding to a single event;
[0107] Based on the event arguments and event role labels in the label information, the role conflict matrix is updated according to a preset update rule;
[0108] Determining whether the role conflict matrix satisfies a no-conflict condition;
[0109] If the role conflict matrix does not meet the non-conflict condition, the event role labels in the label information associated with the conflicting items in the role conflict matrix are deleted, and the role conflict matrix is continuously updated based on the processed label information until the role conflict matrix meets the non-conflict condition.
[0110] In the embodiment of the present invention, the size of the role conflict matrix is determined according to the number of event role labels contained in the label information corresponding to a single event, and the role conflict matrix is constructed, and the initial values of the elements in the role conflict matrix are set to zero.
[0111] For example, the label information corresponding to the single event includes three event role labels: "disaster / accident-collapse / collapse-collapse subject", "disaster / accident-collapse / collapse-number of deaths" and "life-death-location". Then the role conflict matrix is a 3×3 matrix, and a 3×3 zero matrix is constructed.
[0112] The preset update rule may be to generate a role label list based on the event role label contained in the label information; determine whether there is a conflict in the event arguments in the label information; when there is a conflict, obtain the index value of the event role label corresponding to the event argument in the role label list, and update the element value associated with the index value in the role conflict matrix.
[0113] In practical applications, a role label list is generated based on all event role labels contained in the label information corresponding to a single event; and a role conflict matrix is generated based on the role label list. Obtain the label information of a single event. If the event arguments described by the label information overlap or partially overlap, the index value of the event role label corresponding to the event argument in the role label list is added by one to the element value associated with the index value in the role conflict matrix to complete the update of the role conflict matrix.
[0114] For example, according to the label information corresponding to the text "The collapse of the Qingdao subway construction section caused 3 deaths and 2 missing persons are still being searched and rescued", there are 3 event role labels, and the role label list generated by this is {disaster / accident-collapse / collapse-collapse subject, disaster / accident-collapse / collapse-death count, life-death-location}. Since the event argument "Qingdao subway construction section" corresponding to "disaster / accident-collapse / collapse-collapse subject" and the event argument "Qingdao" corresponding to "life-death-location" partially overlap; among them, the index value of the role label "disaster / accident-collapse / collapse-collapse subject" is 1, and the index value of the role label "life-death-location" is 3; add 1 to the values of M[1][3] and M[3][1] in the role conflict matrix M to obtain the updated role conflict matrix M.
[0115] In an embodiment of the present invention, the element value M[i][j] in the role conflict matrix is used to indicate that there is a conflict between the event argument corresponding to the i-th event role label in the role label list and the event argument corresponding to the j-th event role label.
[0116] In practical applications, the number of conflicts of each event role label in the role label list can be determined based on the updated role conflict matrix; the event role label with the highest number of conflicts can be deleted from the role label list; and the role conflict matrix can be rebuilt and updated based on the deleted role label list until there are no conflicting items in the role conflict matrix.
[0117] For example, the element values in each row or column of the updated role conflict matrix M are accumulated to obtain the matrix MR, where the element value MR[i] of the matrix MR represents the number of conflicts of the i-th event role label in the role label list. The largest element value MR[k] in the matrix MR is obtained. If the element value MR[k] is greater than 0, the k-th event role label is deleted from the role label list. The role conflict matrix is reconstructed based on the new role label list, and the above role conflict matrix update steps are repeated until there are no conflicting items in the role conflict matrix, and the role label list is a conflict-free role list.
[0118] In some embodiments, the method further comprises:
[0119] A new role label list is generated according to the event role labels deleted from the role label list; and a role conflict matrix is constructed based on the new role label list.
[0120] In practical applications, the role conflict matrix is reconstructed based on the new role label list, and is updated according to preset update rules until all event role labels in the label information of a single event are included in the non-conflicting role list.
[0121] For example, based on the event role labels deleted in the role label list L1, a role label list L2 is generated, and a role conflict matrix M2 is constructed based on the role label list L2. Obtain the label information of a single event. If the event arguments described in the label information overlap or partially overlap, according to the index value of the event role label corresponding to the event argument in the role label list L2; update the element value associated with the index value in the role conflict matrix M2, and based on the updated role conflict matrix M2, delete the conflicting event role labels in the role label list L2. And according to the deleted event role labels in the role label list L2, a role label list L3 is formed, and the above steps are repeated until all the event role labels contained in the label information of a single event are contained in the role label lists L1, L2, L3...Ln, and the role label lists L1, L2, L3...Ln are used as the final label lists.
[0122] In this way, by constructing a role conflict matrix, all event role labels contained in the label information of a single event are dynamically grouped to ensure that event role labels in the same group do not conflict with each other.
[0123] Optionally, the initial network includes:
[0124] A shared layer, used for outputting feature information based on the input text of the training sample set;
[0125] A plurality of task layers arranged in parallel are located at the output end of the shared layer, and are used to identify arguments and roles according to the feature information, and output event arguments and event roles corresponding to the text.
[0126] In practical applications, the Chinese pre-trained model RoBERTa can be used as a shared layer to extract features from the input word vectors to obtain the feature information of the word vectors.
[0127] In other embodiments, feature extraction of input word vectors may also be achieved by building a custom deep network.
[0128] In an embodiment of the present invention, the multiple parallel task layers are independent of each other and correspond to respective sub-events. The multiple task layers extract feature information associated with the sub-events corresponding to the task layers from the feature information output by the shared layer, and annotate based on the feature information to obtain prediction labels for the multiple sub-events.
[0129] In practical applications, FFN+CRF / softmax can be used as the task layer, where the FNN layer is a two-layer fully connected layer, and the feature information output by the shared layer is subjected to two linear transformations and a ReLU activation through the FNN layer; the feature information is labeled through the CRF layer or the softmax layer to obtain the predicted label of the word vector.
[0130] In this way, features of a single event are extracted through the shared layer, and an independent sub-network is built for each sub-event to identify arguments and roles. The learning of all networks can be completed in one training, reducing the training cost.
[0131] Next, an embodiment of the present invention provides an event extraction device 50, such as Figure 5 As shown, Figure 5 1 is a schematic diagram of the structure of an event extraction device provided by an embodiment of the present invention. The device comprises:
[0132] An event extraction module 51 is used to input the natural language text of a single event into an event extraction network, and obtain labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other;
[0133] The merging module 52 is used to merge the event arguments and event roles described by the tags of the multiple sub-events to obtain the event identification information of the single event.
[0134] Optionally, the event extraction module 51 includes:
[0135] A preprocessing module 511 is used to preprocess the natural language text of the single event to obtain a word vector; and input the word vector into the event extraction network;
[0136] A feature extraction module 512 is used to extract features from the word vector through a shared layer in the event extraction network to obtain feature information;
[0137] The entity extraction module 513, through a plurality of parallel task layers located at the output end of the shared layer in the event extraction network, respectively identifies arguments and roles based on the feature information to obtain labels of a plurality of sub-events within the single event.
[0138] Optionally, the preprocessing module 511 is specifically used for:
[0139] Filtering illegal characters from the natural language text of the single event;
[0140] The processed natural language text is then segmented, and each word obtained by the segmentation is mapped into a word vector.
[0141] Optionally, the entity extraction module 513 is specifically used for:
[0142] The plurality of parallel task layers respectively select characteristic information associated with the sub-event corresponding to the task layer from the characteristic information;
[0143] For the preset event role label information, arguments and roles are identified based on the characteristic information associated with the sub-event corresponding to the task layer, and labels of event arguments and event roles describing the sub-event are obtained.
[0144] Optionally, the event role label information is generated according to the event type and the event role.
[0145] Next, an embodiment of the present invention provides a training device 60 for an event extraction network, such as Figure 6 As shown, Figure 6 1 is a schematic diagram of a training device for an event extraction network provided by an embodiment of the present invention. The device comprises:
[0146] The acquisition module 61 is used to acquire a training sample set of natural language text and label information of the training sample set; the label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other;
[0147] The network acquisition module 62 is used to:
[0148] Inputting the text of the training sample set into the initial network to obtain the predicted label output by the initial network; the predicted label is formed by the event arguments and event roles corresponding to the multiple sub-events of a single event;
[0149] Determining a loss function value of the initial network according to a difference between the predicted label and the label information;
[0150] If the loss function value does not meet the training stop condition, continue training the initial network.
[0151] Optionally, the acquisition module 61 includes:
[0152] The determination module 611 is used to obtain a training sample set of natural language text; based on the preset event role label information, determine the label information corresponding to a single event in the training sample set;
[0153] The construction module 612 is used to construct a role conflict matrix according to the event arguments and event role labels in the label information corresponding to the single event; based on the role conflict matrix, the label information corresponding to the single event in the training sample set is grouped into label information corresponding to the multiple sub-events.
[0154] Optionally, the construction module 612 is specifically used for:
[0155] Constructing the role conflict matrix according to the event role label in the label information corresponding to a single event;
[0156] Based on the event arguments and event role labels in the label information, the role conflict matrix is updated according to a preset update rule;
[0157] Determining whether the role conflict matrix satisfies a no-conflict condition;
[0158] If the role conflict matrix does not meet the non-conflict condition, the event role labels in the label information associated with the conflicting items in the role conflict matrix are deleted, and the role conflict matrix is continuously updated based on the processed label information until the role conflict matrix meets the non-conflict condition.
[0159] Optionally, the event role label information is generated according to the event type and the event role.
[0160] Optionally, the initial network includes:
[0161] A shared layer, used for outputting feature information based on the input text of the training sample set;
[0162] A plurality of task layers arranged in parallel are located at the output end of the shared layer, and are used to identify arguments and roles according to the feature information, and output event arguments and event roles corresponding to the text.
[0163] In combination with the above-mentioned embodiments of the present invention, an exemplary application of the embodiments of the present invention in a practical application scenario will be described below.
[0164] This example provides an event extraction method. The specific steps are as follows:
[0165] Step 701, generating a role tag according to the event type and event role;
[0166] In practical applications, a new role label is formed by binding the event type and the event role to ensure the uniqueness of the role label.
[0167] In actual implementation, the event type and the event role corresponding to the event type can be directly concatenated to generate a role label. For example, if the event type is "life-death", and its corresponding event roles are "place", "time" and "dead person", the new role labels generated are "life-death-place", "life-death-time" and "life-death-dead person".
[0168] Step 702, obtaining a training sample set of natural language text, and annotating the training sample set based on the role label to obtain annotation information;
[0169] In practical applications, event arguments in a training sample set are manually labeled according to the role labels to obtain labeling information.
[0170] The annotation information may be a file in Json format, which includes text, arguments, index, event argument and role.
[0171] For example, the annotation information for the text "The collapse of the Qingdao subway construction section caused 3 deaths and 2 missing persons are still being searched and rescued" is as follows:
[0172]
[0173]
[0174] Step 703, constructing a role conflict matrix according to the annotation information of the training sample set;
[0175] In practical applications, by constructing a role conflict matrix, the event roles in the annotation information of the training sample set are dynamically grouped to ensure that event roles in the same group do not conflict with each other.
[0176] Alternatively, if Figure 7 As shown, Figure 7 A method for constructing a role conflict matrix is provided in an embodiment of the present invention. The method for constructing a role conflict matrix includes:
[0177] Step 7031, generating a role list L0 based on the event roles in the annotation information of the training sample set, obtaining the length k of the role list L0, and constructing a conflict matrix M[k][k];
[0178] Step 7032, obtaining texts in the training sample set, and updating the conflict matrix M according to the annotation information corresponding to the texts and the preset updating rules;
[0179] Step 7033, accumulate each row of the conflict matrix M to obtain a matrix MR; wherein MR[i] is the number of times the i-th event role conflicts with other event roles;
[0180] Step 7034, based on the matrix MR, obtain the event role K with the most conflicts;
[0181] Step 7035, if MR[K] is greater than zero, delete the Kth event role in the role list L0; reconstruct the role conflict matrix M and reset the training data set; jump to step 7032 until there is no conflict in the conflict matrix M.
[0182] In step 7031, the rows and columns of the conflict matrix M[k][k] correspond one-to-one to the event roles in the role list L0, and the initial value of the conflict matrix M is 0.
[0183] In step 7032, it is determined whether there is a conflict based on the event arguments and event roles in the annotation information corresponding to the text.
[0184] The conflict judgment of the text may be to judge whether there are multiple event arguments in the annotation information corresponding to the text, and whether the event arguments overlap in position.
[0185] In actual applications, when there are multiple event arguments in the annotation information corresponding to the text, and the event arguments overlap in position, multiple event roles corresponding to the multiple event arguments are obtained, and the index values of the multiple event roles in the role list are calculated, and the value corresponding to the index value in the conflict matrix M is increased by one.
[0186] When there are two overlapping event arguments in the text, obtain the event roles corresponding to the two event arguments, and according to the index values i and j of the event roles in the role list, add 1 to the conflict matrix M[i-1][j-1] and 1 to M[j-1][i-1].
[0187] For example, Figure 8 As shown, Figure 8 : This is a schematic diagram of a conflict matrix update provided by an embodiment of the present invention. In which, {R_1, R_2, ... R_N} refers to the role list included in the training sample set, and each row / column of the conflict matrix corresponds to an event role label in the role list; M[2][3] represents the element value corresponding to the 3rd row and 4th column of the conflict matrix, and M[3][2] represents the element value corresponding to the 4th row and 3rd column of the conflict matrix.
[0188] For the text "The collapse of a Qingdao subway construction section resulted in the death of three people and the search for two missing persons is still ongoing", there is an overlap between the event argument "Qingdao" corresponding to the event role "life-death-place" and the event argument "Qingdao subway construction section" corresponding to the event role "disaster / accident-collapse / collapse-collapse subject".
[0189] The event role "Life-Death-Place" has an index value of 3 in the role list; the event role "Disaster / Accident-Collapse / Collapse-Collapse Subject" has an index value of 4 in the role list; therefore, add 1 to the conflict matrix M[2][3] and M[3][2].
[0190] In practical applications, when there is no conflict in the conflict matrix M, the method may further include: reconstructing the role list L according to the event roles deleted in the role list L0; reconstructing the conflict matrix based on the role list L until all event roles in the training sample set are included in the role lists L0, L1, L2, ... Ln. The role lists L0, L1, L2, ... Ln are used as the final role label lists.
[0191] Step 704, building an initial network based on the role conflict matrix;
[0192] like Fig. 9 As shown, Fig. 9 This is a schematic diagram of an event extraction network architecture provided by an embodiment of the present invention. The shared layer is a RoBERTa network model; the task layer Task is composed of an FNN layer and a CRF / Softmax layer; Label is the predicted label corresponding to the text of the training sample set output by the task layer; and L is the label information corresponding to the text of the training sample set. Specifically:
[0193] An input layer is set at the beginning of the initial network, and the input text is converted into an input format acceptable to the network through the input layer.
[0194] In practical applications, the input text can be processed to filter out illegal characters in the text, and all English letters in the text can be converted to lowercase letters; the text can be segmented, and each word after segmentation can be converted into a vector V through a dictionary. The text input matrix is Input[n,v], where n is the number of words in the text and v is the vector dimension.
[0195] Here, a single Chinese character can be made into a word, and English characters can be made into separate words. For example, the text "I LOVE Suzhou" can be split into "i", "love", "Suzhou" and "zhou".
[0196] A shared layer is provided at the output end of the input layer, and the shared layer is used to extract features from the input word vector to obtain feature data.
[0197] In practical applications, the shared layer may adopt the RoBERTa model, or the feature extraction function of the shared layer may be realized by constructing a deep network.
[0198] A plurality of parallel task layers are arranged at the output end of the shared layer, wherein the task layers are used to identify arguments and roles based on the feature data.
[0199] In practical applications, a task group Task can be constructed according to the final role label list, where one task group Task corresponds to one role label list, and event roles in role label lists corresponding to different task groups Task do not conflict with each other.
[0200] The task layer can be a FFN+CRF / Softmax model. The FFN is a two-layer fully connected network, in which the first layer of the fully connected network uses the Relu activation function, and CRF / Softmax is mainly used for decoding to convert the neural network into corresponding label probabilities.
[0201] Step 705, determining the label corresponding to the training sample set according to the role conflict matrix;
[0202] In practical applications, a role label list corresponding to each task group Task is obtained, and an entity index is constructed according to the BIO annotation method; and based on the entity index, the label corresponding to the text in the training sample set is determined.
[0203] Here, the BI0 notation is to mark non-"entity / role" as 0, and "entity / role" starts with B, and I connects the notation.
[0204] The following is an introduction using the text "The collapse of a Qingdao subway construction section resulted in the death of three people and the search and rescue of two missing persons is still ongoing" as an example.
[0205] Get the role label list corresponding to the task group Task, which contains a group of non-conflicting event roles: "disaster / accident-collapse / collapse-collapse subject" and "disaster / accident-collapse / collapse-death count"; according to the BI0 labeling method, it can be determined that the number of labels corresponding to the task group Task is 5, namely {0, B-disaster / accident-collapse / collapse-collapse subject, I-disaster / accident-collapse / collapse-collapse subject, B-disaster / accident-collapse / collapse-death count, I-disaster / accident-collapse / collapse-death count}. Therefore, it is determined that the entity indexes corresponding to the labels are 0-4 respectively.
[0206] According to the entity index, the label corresponding to the text is determined as follows:
[0207] i.Text: The collapse of a subway construction section in Qingdao has left three people dead and two missing persons are still being searched for.
[0208] ii.Label: 1 2 2 2 2 2 2 0 0 0 3 4 0 0 0 0 0 0 0 0 0 0 0
[0209] Step 706, inputting the text of the training sample set into the initial network to obtain the predicted label of the training sample set output by the initial network; determining the loss function according to the difference between the label corresponding to the training sample set and the predicted label;
[0210] In practical applications, the input of each task layer is the output of the shared layer:
[0211] Task output =RoBERTa(Input)
[0212] Task output Represents the feature information of the task layer input; Input represents the word vector of the shared layer input; RoBERTa() represents the function corresponding to the RoBERTA network model.
[0213] For different Tasks i Construct an independent network and the final output is DP i , the loss function is loss i ;
[0214] DP i =CRF(FNN(Task output ))
[0215] loss i =NLL(DP i ,target)
[0216] Among them, the loss function mainly uses the negative log-likelihood function (NLL); FNN() represents the function corresponding to the feedforward neural network (FNN) model; CRF() represents the function DP corresponding to the conditional random field (CRF) network model i represents the predicted label output by the i-th task layer; target represents the label corresponding to the text of the training sample set.
[0217] The loss function of each task layer is accumulated as the loss function of the initial network.
[0218]
[0219] Wherein, Loss represents the loss function value of the initial network; lossi represents the loss function value of the i-th task layer; n represents the number of task layers in the initial network.
[0220] Step 707, according to the loss function value, the parameters to be trained of the initial network are adjusted until the loss function value meets the training stop condition, thereby obtaining an event extraction network.
[0221] In practical applications, the number of training iterations can be pre-set, and when the loss function value converges or the training reaches the number of training iterations, the training of the initial network is stopped.
[0222] Step 708, input the natural language text of the single event to be identified into the event extraction network, obtain the labels of multiple sub-events of the single event output by the event extraction network; merge the labels of the multiple sub-events to obtain the event recognition result.
[0223] In practical applications, the final recognition results [DP1, DP2, …, DP n ] to disassemble and merge events to obtain event recognition results.
[0224] Below, we will use the text "The collapse of a Qingdao subway construction section resulted in the death of three people and the search and rescue of two missing persons is still ongoing" as an example.
[0225] The final recognition result output by the event recognition network:
[0226] Text: The collapse of the Qingdao subway construction section caused 3 deaths and 2 missing persons are still being searched and rescued.
[0227] DP1: 1 2 2 2 2 2 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[0228] DP2: 0 0 0 0 0 0 0 0 0 0 3 4 0 0 0 0 0 0 0 0 0 0 0
[0229] …
[0230] DP n :5 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[0231] The event arguments finally identified are: Qingdao Metro Construction Section (DP1), 3 people (DP2) and Qingdao (DP n ), the event roles corresponding to the event arguments are: disaster / accident-collapse-collapse subject, disaster / accident-collapse-death number and life-death-place.
[0232] According to the role label type, the event roles are merged to complete event identification:
[0233]
[0234] An embodiment of the present invention further provides an electronic device, the electronic device comprising:
[0235] A memory for storing executable instructions;
[0236] The processor is used to implement the event extraction method provided by one or more of the above-mentioned technical solutions when executing the executable instructions stored in the memory.
[0237] The memory may include: various types of storage media that can be used for data storage. In this embodiment, at least part of the storage media included in the memory is a non-volatile storage medium that can be used to store the computer program.
[0238] The processor may include: a central processing unit, a microprocessor, a digital signal processor, an application processor, a dedicated integrated circuit or a programmable array, etc., which can be used for the event extraction method of one or more of the above-mentioned technical solutions through a computer program.
[0239] In this embodiment, the processor may be connected to the memory via an internal bus of a device such as an integrated circuit bus.
[0240] An embodiment of the present invention further provides an electronic device, the electronic device comprising:
[0241] A memory for storing executable instructions;
[0242] The processor is used to implement the event extraction network training method provided by one or more of the above-mentioned technical solutions when executing the executable instructions stored in the memory.
[0243] The embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium stores a computer program. After the computer program is executed by a processor, the event extraction method provided by one or more of the above technical solutions is executed. For example, the following can be executed: Figure 1 The method shown.
[0244] The embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium stores a computer program, and after the computer program is executed by a processor, the training method of the event extraction network provided by one or more of the above technical solutions is executed, for example, the following can be executed: Figure 3 The method shown.
[0245] The computer storage medium provided in the embodiment of the present invention includes: a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes. Optionally, the computer storage medium can be a non-transient storage medium. The non-transient storage medium here can also be called a non-volatile storage medium.
[0246] The above description is only an embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement and improvement made within the spirit and scope of the present invention are included in the protection scope of the present invention.
Claims
1. An event extraction method, characterized in that: include: Inputting the natural language text of a single event into an event extraction network, and obtaining labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other; Merging event arguments and event roles described by the tags of the plurality of sub-events to obtain event identification information of the single event; The method of inputting the natural language text of a single event into an event extraction network to obtain labels of multiple sub-events within the single event output by the event extraction network includes: processing the text of the natural language text of the single event input into the event extraction network through a shared layer in the event extraction network and a plurality of parallel task layers located at the output end of the shared layer to obtain labels of multiple sub-events within the single event output by the event extraction network.
2. The method according to claim 1, characterized in that The step of inputting the natural language text of a single event into an event extraction network to obtain labels of multiple sub-events in the single event output by the event extraction network includes: Preprocessing the natural language text of the single event to obtain a word vector; Input the word vector into the event extraction network; Extracting features from the word vectors through a shared layer in the event extraction network to obtain feature information; Through a plurality of task layers arranged in parallel at the output end of the shared layer in the event extraction network, arguments and roles are identified based on the feature information respectively, so as to obtain labels of a plurality of sub-events in the single event.
3. The method according to claim 2, characterized in that The preprocessing of the natural language text of the single event to obtain a word vector includes: Filtering illegal characters from the natural language text of the single event; The processed natural language text is then segmented, and each word obtained by the segmentation is mapped into a word vector.
4. The method according to claim 2, characterized in that: The multiple task layers located at the output end of the shared layer in the event extraction network respectively identify arguments and roles based on the feature information to obtain labels of multiple sub-events in the single event, including: The plurality of parallel task layers respectively select characteristic information associated with the sub-event corresponding to the task layer from the characteristic information; For the preset event role label information, arguments and roles are identified based on the characteristic information associated with the sub-event corresponding to the task layer, and labels of event arguments and event roles describing the sub-event are obtained.
5. The method according to claim 4, characterized in that The event role label information is generated according to the event type and the event role.
6. A training method for an event extraction network, characterized in that: include: Obtaining a training sample set of natural language text and label information of the training sample set; The label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other; Inputting the text of the training sample set into the initial network, processing the text of the training sample set through a shared layer in the initial network and a plurality of parallel task layers located at the output end of the shared layer, and obtaining a predicted label output by the initial network; the predicted label is formed by event arguments and event roles corresponding to the plurality of sub-events of a single event; Determining a loss function value of the initial network according to a difference between the predicted label and the label information; If the loss function value does not meet the training stop condition, continue training the initial network.
7. The training method according to claim 6, characterized in that: The step of obtaining a training sample set of natural language text and label information of the training sample set includes: Obtain a training sample set of natural language text; Based on the preset event role label information, determine the label information corresponding to the single event in the training sample set; Constructing a role conflict matrix according to event arguments and event role labels in the label information corresponding to the single event; Based on the role conflict matrix, the label information corresponding to a single event in the training sample set is grouped into label information corresponding to the multiple sub-events.
8. The training method according to claim 7, characterized in that: The step of constructing a role conflict matrix according to the event arguments and event role labels in the label information corresponding to the single event includes: Constructing the role conflict matrix according to the event role label in the label information corresponding to a single event; Based on the event arguments and event role labels in the label information, the role conflict matrix is updated according to a preset update rule; Determining whether the role conflict matrix satisfies a no-conflict condition; If the role conflict matrix does not meet the non-conflict condition, the event role labels in the label information associated with the conflicting items in the role conflict matrix are deleted, and the role conflict matrix is continuously updated based on the processed label information until the role conflict matrix meets the non-conflict condition.
9. The training method according to claim 6 or 7, characterized in that: The event role label information is generated according to the event type and the event role.
10. The training method according to claim 6, characterized in that: The initial network includes: The shared layer is used to output feature information based on the input text of the training sample set; The multiple parallel task layers are located at the output end of the shared layer, and are used to identify arguments and roles according to the feature information, and output event arguments and event roles for describing the event corresponding to the text.
11. An event extraction device, characterized in that: include: An event extraction module, used to input the natural language text of a single event into an event extraction network, and obtain labels of multiple sub-events in the single event output by the event extraction network, wherein one sub-event corresponds to one label, and labels of different sub-events do not conflict with each other; A merging module, used for merging event arguments and event roles described by the labels of the plurality of sub-events to obtain event identification information of the single event; The event extraction module is used to process the pre-processed natural language text of the single event input into the event extraction network through a shared layer in the event extraction network and a plurality of parallel task layers located at the output end of the shared layer, and obtain labels of a plurality of sub-events within the single event output by the event extraction network.
12. A training device for an event extraction network, characterized in that: include: An acquisition module, used to acquire a training sample set of natural language text and label information of the training sample set; The label information includes label information of multiple sub-events corresponding to a single event in the training sample set; wherein one sub-event corresponds to one label information, and label information of different sub-events does not conflict with each other; Network acquisition module, used to: Inputting the text of the training sample set into the initial network, processing the text of the training sample set through a shared layer in the initial network and a plurality of parallel task layers located at the output end of the shared layer, and obtaining a predicted label output by the initial network; the predicted label is formed by event arguments and event roles corresponding to the plurality of sub-events of a single event; Determining a loss function value of the initial network according to a difference between the predicted label and the label information; If the loss function value does not meet the training stop condition, continue training the initial network.
13. An electronic device, comprising: A memory for storing executable instructions; The processor is used to implement the event extraction method according to any one of claims 1 to 5 when executing the executable instructions stored in the memory.
14. An electronic device comprising: A memory for storing executable instructions; The processor is used to implement the training method of the event extraction network as described in any one of claims 6 to 10 when executing the executable instructions stored in the memory.
15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores executable instructions, and when the executable instructions are executed by a processor, the event extraction method according to any one of claims 1 to 5 is implemented.
16. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores executable instructions, and when the executable instructions are executed by the processor, the training method of the event extraction network as described in any one of claims 6 to 10 is implemented.
