CN116028812A - A Construction Method of Pipeline Multi-event Extraction Model - Google Patents

Abstract

The invention relates to a construction method of a pipeline type multi-event extraction model, belongs to the technical field of natural language processing, and solves the problems that the existing event extraction model is easy to identify missing and the event elements cannot be matched under the condition that more events or multi-event overlapping exists in corpus, so that the accuracy is low. The method comprises the steps of constructing an event characteristic data set based on an original data set, further constructing a training set containing positive and negative samples of event types and event elements, and training a T5 model by using the training set, so that the model effectively learns internal relations among the event types, the event roles, the event elements and trigger words, and particularly improves understanding and predicting capabilities of the model on multiple events.

Description

A method for constructing a pipeline multi-event extraction model

Technical Field

The present invention relates to the technical field of natural language processing, and in particular to a method for constructing a pipeline multi-event extraction model.

Background Art

Event extraction (EE) is one of the important tasks in the field of natural language processing (NLP). The purpose of event extraction is to identify the event type (eventtype), event trigger (trigger), event element (argument) and element role (argumentrole) contained in a given corpus. At present, the application scenarios of event extraction technology are very wide. It can efficiently extract useful information from massive texts and provide strong data support for the construction of knowledge graphs.

The existing mainstream event extraction model extraction methods include sequence labeling, pointer discrimination and generative methods. Sequence labeling is essentially a multi-label multi-classification method, which predicts the possible label for each token; pointer discrimination extracts events by predicting the start and end positions of the text corresponding to each label; generative method is an end-to-end method, which extracts context information through a deeper network and directly outputs event information in text format. The above three methods have good performance in single events or non-overlapping multi-event corpora with a small number of events, but when there are more events in the corpus, especially when one or several elements overlap, it is very easy to have problems such as missing recognition, recognition errors, and event elements cannot be matched, resulting in very low accuracy. Since overlapping multi-events are ubiquitous in actual corpora, there is an urgent need for a more optimized method for constructing a multi-event extraction model to solve the problem that the event extraction model in the prior art has low extraction accuracy due to missing recognition and event elements cannot be matched in multi-event and overlapping multi-event extraction tasks.

Summary of the invention

In view of the above analysis, an embodiment of the present invention aims to provide a method for constructing a pipeline multi-event extraction model, so as to solve the problem of low extraction accuracy of event extraction models in the prior art in multi-event and overlapping multi-event extraction tasks due to missing recognition and inability to match event elements.

On the one hand, an embodiment of the present invention provides a method for constructing a pipeline multi-event extraction model, comprising the following steps:

Obtain the labeled text data as the original data set;

Based on the original data set, the event feature data set is obtained, and the event type positive sample data set D ₊₁ , event element positive sample data set D ₊₂ , event type all negative sample data set D _-1 and event element random negative sample data set D _-2 are further constructed, and finally the model training data set D _all is obtained;

Use the training data set D _all to train the T5 model and obtain the trained pipeline multi-event extraction model M _trained ;

When extracting multiple events, a prediction sample set for each step is gradually constructed, and the trained model M _{trained is} used to obtain the prediction result of each step based on the prediction sample set of each step, and the final extraction result is obtained by integration.

Furthermore, the step of obtaining the annotated text data includes:

Get the original text data;

The original text data is annotated; wherein the annotation includes: determining the event type contained in the sentences in the text data; extracting trigger words, event elements and their positions according to the event type; and labeling the event elements with appropriate event role labels.

Furthermore, the event feature data set includes:

The schema is the correspondence between event types and all event roles, the corresponding set S _{type_role} between event types and individual event roles, the set S _type of all event types, the set S _trigger of all trigger words, and the set S _argument of all event elements. The schema records all event types in the original dataset and all the event roles corresponding to them. S _{type_role} is derived from the schema, including the event type of each event in the schema and the pairwise combination of all event roles, as well as the event role to which the event role belongs in the schema. S _type records all event types. S _trigger records all trigger words that appear in the original dataset. S _argument records all event elements contained in the original dataset.

Furthermore, the model training dataset D _all is constructed by the following steps:

Summarize and organize the annotation information of the original data set to obtain three event feature data sets: the corresponding relationship schema between event types and all event roles, the corresponding set S _{type_role} between event types and single event roles, and the set S _{type of} all event types;

Use the original dataset and dataset schema to construct event type positive sample dataset D ₊₁ and event element positive sample dataset D ₊₂ , as well as two event feature datasets: the set of all trigger words S _trigger and the set of all event elements S _{argument that} appear in the original dataset;

Use the event type positive sample dataset D ₊₁ and the event type dataset S _type to construct the event type full negative sample dataset D _-1 ;

Use the event element positive sample dataset D ₊₂ , the trigger word set S _trigger , the event element set S _argument , and the corresponding set S _{type_role} of event type and single event role to construct the event element random negative sample dataset D _-2 ;

D ₊₁ , D ₊₂ , D _-1 , and D _-2 are mixed and shuffled to finally obtain the model training data set D _all .

Furthermore, the event type positive sample dataset D ₊₁ and the event element positive sample dataset D ₊₂ are constructed by the following steps:

A1. Extract the event type e _type , trigger word w _trigger , event roles e _{role_1} ～e _{role_n} , and corresponding event elements w _{arg_1} ～w _{arg_n} (n is the number of event roles contained in the event, which is also equal to the number of event elements) corresponding to a certain event contained in the original data set text data text_p; the input for constructing the event type positive sample of the event is text_p+e _type + "trigger word", and the output is w _trigger ; the input for constructing the event element positive sample of the event is text_p+prompt _arg , and the output is w _{arg_1} ～w _{arg_n} ; wherein, the event element prompt prompt _arg can be obtained by the following formula:

A2. For each event in the text data text_p, use the method in (1) to construct event type positive samples and event element positive samples, and obtain event type positive sample dataset D ₊₁ and event element positive sample dataset D ₊₂ ;

Furthermore, the event type all-negative sample dataset D _-1 is constructed by the following steps:

B1. Replace the e _type of a positive sample of a certain event type with other event types of the event in the event type dataset S _type in sequence, and the target output is empty to obtain all negative samples of the event type of the event;

B2. Use method (1) for all events in the event type positive sample dataset D ₊₁ to construct the event type full negative sample dataset D _-1 .

Furthermore, the event element random negative sample dataset D _-2 is constructed by the following steps:

(1) Find all positive event element samples of a certain event in D _{+ 2} , and find all event element prompts prompt _arg from the positive event element samples to form a set S _prompt ;

(2) Randomly select a trigger word from S _trigger to obtain w _{trigger_random} ; randomly select an element from S _{type_role} to obtain an event type e _{type_random} , an event role e _{role_random} and the position p of the event role;

(3) Randomly select p event elements from the event element set S _argument to obtain w _{arg_r_1} to w _{arg_r_p} , and combine them in the following format to obtain the event element random prompt prompt _{arg_random} ;

prompt _{arg_random} =e _{type_random} +w _{trigger_random} +w _{arg_r_1} +…+w _{arg_r_p} +e _{role_random}

(4) Determine whether prompt _{arg_random} exists in S _prompt . If so, repeat steps 2, 3, and 4. If not, use prompt _{arg_random} to construct a negative sample and add prompt _{arg_random} to S _prompt .

(5) Repeat steps (1) to (4) until 5n random negative samples of event elements are obtained.

(6) All event samples in D ₊₂ are constructed using the methods in (1) to (5) to obtain the event element random negative sample dataset D _-2 .

Furthermore, the training of the T5 model includes:

Divide the model training data set D _all into a training set D _train , a validation set _Deval , and a test set D _test according to a certain ratio;

Fine-tune the T5 model using the training set D _train for n rounds. After each round of training, use the validation set _Deval for validation. Take the model with the best validation set result as the final model and test it with the test set D _test to finally get the trained model M _trained ;

During training, the following formula is used to calculate the model loss and update the parameters:

Loss=CrossEntropy(x _pred ,x _gold )

Among them, x _pred is the prediction result and x _gold is the target output.

Furthermore, the stepwise construction of the prediction sample set for each step includes:

Construct the first step prediction sample set D _{step_1} based on the text to be extracted and the event feature data set;

Based on the text to be extracted, the event feature data set and the prediction result of the previous model M _trained, the prompt information prompt is constructed, and the prediction sample set of the next model is constructed with the text+prompt structure, so as to realize the step-by-step construction of the prediction sample sets D _{step_2} to D _{step_(n+1)} of the 2nd to n+1st steps.

Furthermore, the trained model M _{trained is} used to obtain the prediction result of each step based on the prediction sample set of each step, and integrate to obtain the final extraction result, including:

Input D _{step_1} into the model M _trained to obtain all trigger words p _trigger contained in the first step prediction result text;

对应事件类型

的第x-1事件角色

对应的第x-1事件要素

其中x∈[2,n+1]；其中prompt_x表示为：Construct the prediction sample set D _{step_2} ~D _{step_(n+1)} for the 2nd to n+1st steps in the format text+prompt_x, input D _{step_x} into the model M _trained , and obtain each trigger word

Corresponding event type

The x-1 event character

The corresponding x-1 event element

where x∈[2,n+1]; prompt_x is represented as:

Combine the prompt information of the last step with the extraction result to get the complete event.

Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

1. By constructing an event feature data set based on the original data set, and further constructing a training set containing positive and negative samples of event types and event elements, the T5 model is trained using the training set, so that the model can effectively learn the intrinsic connections between various event types, event roles, event elements and trigger words, especially improving the model's understanding and prediction capabilities for multiple events. The overall training process uses the prompt method to ensure the extraction accuracy and fidelity to a certain extent, and obtain an event extraction model with a high recognition rate for event text.

2. Extract event text based on the trained model. Events can be extracted in a layered and progressive manner by using prompt information. All event types are used as prompt information to extract corresponding trigger words. Then the trigger words and the element roles to be extracted are added step by step to the prompt to extract event elements. After all event elements contained in the event type are extracted, the prompt information of the last step is combined with the extraction result to obtain a complete event. This pipeline extraction method provides a separate extraction path for each possible event, focusing on solving the problems of missing identification and event element matching when extracting multiple events and overlapping events, greatly improving the extraction accuracy.

In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

FIG1 is a schematic flow chart of a method for constructing a pipeline multi-event extraction model according to an embodiment of the present invention;

FIG2 is a schematic diagram of the overall implementation process of the method for constructing a pipeline multi-event extraction model including actual prediction according to an embodiment of the present invention;

FIG3 is a schematic diagram of a flow chart of constructing training data according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a process for obtaining prediction results provided by an embodiment of the present invention.

DETAILED DESCRIPTION

The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

A specific embodiment of the present invention discloses a method for constructing a pipeline multi-event extraction model, as shown in FIG1 , comprising the following steps:

Step S110, obtaining the annotated text data as the original data set;

Step S120, obtaining an event feature data set based on the original data set, and further constructing an event type positive sample data set D ₊₁ , an event element positive sample data set D ₊₂ , an event type full negative sample data set D _-1 and an event element random negative sample data set D _-2 , and finally obtaining a model training data set D _all ;

Step S130: Use the training data set D _all to train the T5 model to obtain a trained pipeline multi-event extraction model M _trained ;

When extracting multiple events, a prediction sample set for each step is gradually constructed, and the trained model M _{trained is} used to obtain the prediction result of each step based on the prediction sample set of each step, and the final extraction result is obtained by integration.

The embodiment of the present invention uses a training set containing positive and negative samples of event types and event elements to train the T5 model to obtain a multi-event extraction model. By constructing an event feature data set based on the original data set, and further constructing a training set containing positive and negative samples of event types and event elements, the T5 model is trained using the training set, so that the model effectively learns the internal connections between various event types, event roles, event elements and trigger words, and especially improves the model's understanding and prediction capabilities for multiple events. The overall training process uses a prompt method to ensure extraction accuracy and fidelity to a certain extent, and obtains an event extraction model with a high recognition rate for event texts.

Based on the above embodiment, specifically, the annotated text data in the above step S110 is obtained by the following method:

Directly use Baidu's event extraction dataset;

Label the original text data by yourself; the labeling method is: determine the event type contained in the sentence in the text data; extract the trigger words, event elements and their positions according to the event type; label the event elements with appropriate event role labels;

Specifically, the above step S120 can also be optimized into the following steps:

Step S210, summarize and organize the annotation information of the original data set to obtain three event feature data sets: the corresponding relationship schema between event type and all event roles, the corresponding set S _{type_role} between event type and a single event role, and the set S _type of all event types;

Specifically, all event types and event roles in the original data set are summarized and organized to construct data set schema, S _{type_role} and S _type ; wherein the schema records all event types in the original data set and all event roles corresponding thereto; S _{type_role} is obtained according to the schema, including the event type of each event in the schema and the pairwise combination of all event roles, as well as the event role to which the event role belongs in the schema; S _type records all event types; preferably, the schema is stored in a file json, and both S _{type_role} and S _type are stored using a set.

For example, for an event whose event type is "acquisition" and whose event roles include "acquisition time, acquirer, acquired party", its records in schema, S _{type_role} and S _type are shown in Table 1.

Table 1 Example of records of events of type "acquisition" in schema, S _{type_role} and S _type

Step S220, using the original data set and the data set schema to construct an event type positive sample data set D ₊₁ and an event element positive sample data set D ₊₂ , as well as two event feature data sets, namely, a set of all trigger words S _trigger and a set of all event elements S _argument that appear in the original data set;

Specifically, the constructing of the event type positive sample data set D ₊₁ and the event element positive sample data set D ₊₂ , as well as all trigger word sets S _trigger and all event element sets S _argument appearing in the original data set includes:

(1) Extract the event type e _type , trigger word w _trigger , event roles e _{role_1} to e _{role_n} , and corresponding event elements w _{arg_1} to w _{arg_n} (n is the number of event roles contained in the event, which is also equal to the number of event elements) corresponding to a certain event contained in the original data set text data text_p; the input for constructing the event type positive sample of the event is text_p + e _type + "trigger word", and the output is w _trigger ; the input for constructing the event element positive sample of the event is text_p + prompt _arg , and the output is w _{arg_1} to w _{arg_n} ; where the event element prompt prompt _arg can be obtained by the following formula:

(2) For each event in the text data text_p, use the method in (1) to construct event type positive samples and event element positive samples, and obtain event type positive sample dataset D ₊₁ and event element positive sample dataset D ₊₂ ;

(3) Save the trigger words w _trigger of all events in the text data text_p in the trigger word set S _trigger , and save all event elements w _{arg_1} to w _{arg_n} in the event element set S _argument , to obtain the trigger word dataset S _trigger and the event element dataset S _argument .

Exemplarily, for an event of type “acquisition”, the constructed event type positive samples and event element positive samples and the stored examples in S _trigger and S _argument are shown in Table 2.

Table 2. Examples of positive samples of event types and event elements for an event of type "acquisition" and examples of their storage in S _trigger and S _argument

It should be noted that in this example, the prompt information elements are separated by "-", but other symbols or spaces can also be used. When constructing event element positive samples, the order of event elements in prompt _arg must be consistent with that recorded in the schema.

For events in complex situations, the output may be multiple event elements. When constructing the input, event element prompts need to be constructed separately. For example, Table 3 shows an example of positive event element samples of multiple events with a common trigger word:

Table 3. Examples of positive samples of event elements of multiple events with common trigger words

Step S230, constructing an event type all-negative sample dataset D _-1 using the event type positive sample dataset D ₊₁ and the event type dataset S _type ;

Specifically, the construction of the event type all-negative sample dataset D _-1 includes:

(1) Replace the e _type of a certain event type positive sample with other event types of the event in the event type dataset S _type in sequence, and the target output is empty to obtain all negative samples of the event type of the event;

(2) The method (1) is applied to all events in the event type positive sample dataset D ₊₁ to construct the event type full negative sample dataset D _-1 .

For example, there are m event types in the event type dataset S _type , and there are m-1 full negative samples of each event type;

For model training, positive samples are samples with target output results, and negative samples are samples without output results. Adding negative samples during training can effectively improve the model recognition accuracy.

Step S240, constructing an event element random negative sample dataset D _-2 using the event element positive sample dataset D ₊₂ , the trigger word set S _trigger , the event element set S _argument , and the corresponding set S _{type_role} of event types and single event roles;

The input format of event element random negative samples is the same as that of event element positive samples. The difference is that the prompt information of event element random negative samples is different, and the output results are empty. For a certain event, the number of event element random negative samples is generally recommended to be 5 times the number of event element positive samples;

Specifically, the steps of constructing the event element random negative sample data set D _-2 are as follows:

(1) Find all positive event element samples of a certain event in D _{+ 2} , and find all event element prompts prompt _arg from the positive event element samples to form a set S _prompt ;

(2) Randomly select a trigger word from S _trigger to obtain w _{trigger_random} ; randomly select an element from S _{type_role} to obtain an event type e _{type_random} , an event role e _{role_random} and the position p of the event role;

(3) Randomly select p event elements from the event element set S _argument to obtain w _{arg_r_1} to w _{arg_r_p} , and combine them in the following format to obtain the event element random prompt prompt _{arg_random} ;

prompt _{arg_random} =e _{type_random} +w _{trigger_random} +w _{arg_r_1} +…+w _{arg_r_p} +e _{role_random}

(4) Determine whether prompt _{arg_random} exists in S _prompt . If so, repeat steps 2, 3, and 4. If not, use prompt _{arg_random} to construct a negative sample and add prompt _{arg_random} to S _prompt .

(5) Repeat steps (1) to (4) until 5n random negative samples of event elements are obtained.

(6) Using the methods (1) to (5) for all event samples in D ₊₂ , we construct a random negative sample dataset D _-2 of event elements;

Step S250, D ₊₁ , D ₊₂ , D _-1 , and D _-2 are mixed and shuffled to finally obtain a model training data set D _all ;

Specifically, the training of the T5 model in the above step S130 includes:

The model training data set D _all is divided into a training set D _train , a validation set _Deval and a test set D _test according to a certain ratio; preferably, the ratio is 8:1:1; the T5 model is fine-tuned and trained for n rounds using the training set D _train , and the validation set _Deval is used for validation after each round of training, and the model with the best validation set result is taken as the final model, and the model is tested with the test set D _test to finally obtain a trained model M _trained ; preferably, the training set n is 20;

Furthermore, the following formula is used to calculate the model loss and update the parameters during training:

Loss=CrossEntropy(x _pred ,x _gold )

Among them, x _pred is the prediction result and x _gold is the label.

Furthermore, using the trained model M _trained ; extracting the actual event text includes the following steps:

Step S310, obtaining the text to be extracted; wherein the text to be extracted may be news text data crawled from a website;

Step S320: Based on the text to be extracted, the event feature data set obtained from the original data set and the prediction result of the previous step model M _train , the prediction sample sets D _{step_1} to D _{step_(n+1)} of the first to n+1 steps are constructed step by step with the text+prompt structure, and D _{step_1} to D _{step_(n+1)} are input into the model M _train step by step to obtain the prediction results of the first to n+1 steps model M _train ; n is the number of event roles of the event type corresponding to the prediction result of the first step;

Specifically, the construction of the prediction sample sets D _{step_1} to D _{step_(n+1)} of the first to n+1 steps and the acquisition of the prediction results of the models M _train of the first to n+1 steps include the following steps:

向第一步预测样本集D_step1中加入样本：

遍历结束后，D_step1中的样本数为m(m为事件类型数，k∈[1,m])；(A) Traverse all event types e _type in S _type in turn. For any event type

Add samples to the first step prediction sample set D _step1 :

After the traversal is completed, the number of samples in D _step1 is m (m is the number of event types, k∈[1,m]);

的触发词

记为

从schema中查找

对应的第一事件角色

并以格式text+prompt_{_2}将输出结果加入下一步预测样本集D_step2；其中

(B) Input the first step prediction sample set D _step1 into M _trained . When a sample has an output result, its output result is the event type in the text to be extracted.

Trigger word

Recorded as

Search from schema

Corresponding first event role

And add the output result to the next prediction sample set D _step2 in the format of text+prompt _{_2} ;

的触发词，即文本text中不包含事件类型为

的事件。For samples with no output results, it means that there is no input event type in the text text

The trigger word of the event type is not contained in the text text.

of events.

对应的第一事件角色

的事件要素，记为

通过查看schema判断该事件类型

是否有其他事件角色，若没有进行步骤S330；(C) Input D _step2 into M _trained to predict each trigger word

Corresponding first event role

The event element is recorded as

Determine the event type by viewing the schema

Whether there are other event roles, if not, proceed to step S330;

在schema中存在其他事件角色

则对该事件类型

的其他事件角色

按步骤依次构建下一步预测样本集D_{step_3}～D_{step_(n+1)}，并将D_{step_3}～D_{step_(n+1)}按步依次输入模型M_trained进行事件要素

的抽取，直到

包含的全部事件角色对应的事件要素都被模型M_trained抽取，进行步骤S330；If the event type

Other event roles exist in the schema

For this event type

Other event roles

Construct the next prediction sample set D _{step_3} ~D _{step_(n+1) step} by step, and input D _{step_3} ~D _{step_(n+1)} into the model M _trained step by step to perform event factor

of extraction until

The event elements corresponding to all the event roles are extracted by the model M _trained , and step S330 is performed;

More specifically, the method for constructing the next prediction sample set D _{step_3} ~D _{step_(n+1)} is:

为prompt_{_(x-1)}基础上将

替换为

并在最后加入

其中x∈[3,n+1]，n为schema中该事件类型所包含的事件角色数；Construct samples in the format of text+prompt _{_X} and add them to the next prediction sample set D _{step_(x)} ;

Based on prompt _{_(x-1)}

Replace with

And add at the end

Where x∈[3,n+1], n is the number of event roles contained in the event type in the schema;

其确定方法如下：The event elements used in the prompt include:

The determination method is as follows:

中包含多个预测结果，需按照本步中的格式将多个结果分开构建预测样本。Where j∈[1,n-1], n is the number of event roles contained in the event type in the schema; if

contains multiple prediction results, and the multiple results need to be separated into prediction samples according to the format in this step.

Step S330, based on the prediction results of the n+1-step prediction sample set D _{step_(n+1)} and the n+1-step model M _train , integrate to obtain the final recognition result;

Specifically, the integration to obtain the final recognition result includes:

According to D _{step_n+1} and the nth event element of the prediction result, the event extraction results are sorted as follows:

Event Type:

Trigger words:

Event role/event element (role/argument):

Exemplarily, the event extraction results may be integrated using the format shown in Table 4.

Table 4 Example of event extraction result integration

In summary, the beneficial effects of this embodiment are as follows:

Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

1. By constructing an event feature data set based on the original data set, and further constructing a training set containing positive and negative samples of event types and event elements, the T5 model is trained using the training set, so that the model can effectively learn the intrinsic connections between various event types, event roles, event elements and trigger words, especially improving the model's understanding and prediction capabilities for multiple events. The overall training process uses the prompt method to ensure the extraction accuracy and fidelity to a certain extent, and obtain an event extraction model with a high recognition rate for event text.

2. Extract event text based on the trained model. Events can be extracted in a layered and progressive manner by using prompt information. All event types are used as prompt information to extract corresponding trigger words. Then the trigger words and the element roles to be extracted are added step by step to the prompt to extract event elements. After all event elements contained in the event type are extracted, the prompt information of the last step is combined with the extraction result to obtain a complete event. This pipeline extraction method provides a separate extraction path for each possible event, focusing on solving the problems of missing identification and event element matching when extracting multiple events and overlapping events, greatly improving the extraction accuracy.

Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims (10)

1. The method for constructing the pipeline type multi-event extraction model is characterized by comprising the following steps of:

acquiring marked text data as an original data set;

obtaining an event feature data set based on the original data set, and further constructing an event type positive sample data set D ₊₁ Event element positive sample dataset D ₊₂ Event type all negative sample dataset D _-1 And event element random negative sample dataset D _-2 Finally, a model training data set D is obtained _all ；

Using training dataset D _all Training the T5 model to obtain a trained pipeline type multi-event extraction model M _trained ；

Gradually constructing a prediction sample set of each step during multi-event extraction, wherein the trained model M _trained The method is used for obtaining the prediction result of each step based on the prediction sample set of each step, and integrating to obtain the final extraction result.

2. The method of claim 1, wherein the obtaining the annotated text data comprises:

acquiring original text data;

labeling the original text data; wherein, the labeling includes: determining the event type contained in sentences in the text data; extracting trigger words, event elements and positions thereof according to the event types; the event elements are labeled with the appropriate event roles.

3. The method of claim 1, wherein the set of event feature data comprises:

event type and all event role corresponding relation schema, event type and single event role corresponding set S _{type_role} All event type set S _type All trigger word sets S _trigger And all event element set S _argument The method comprises the steps of carrying out a first treatment on the surface of the Wherein the schema records all event types in the original data set and all event roles corresponding to the event types respectively; s is S _{type_role} Obtaining according to the schema, including the event type of each event in the schema and the combination of all event roles, and the event roles belonging to the same event role in the schema; s is S _type All event types are recorded; s is S _trigger Recording all trigger words appearing in the original data set; s is S _argument All event elements contained in the original dataset are recorded.

4. The method according to claim 1, characterized in that the model training dataset D _all The method is constructed by the following steps:

summarizing and sorting the labeling information of the original data set to obtain a corresponding relation schema of the event types and all event roles and a corresponding set S of the event types and single event roles _{type_role} All event type set S _type Three event feature data sets;

constructing event type positive samples using raw data sets and data set schemasData set D ₊₁ And event element positive sample dataset D ₊₂ And all trigger word sets S that occur in the original dataset _trigger And all event element set S _argument Two event feature data sets;

using event type positive sample dataset D ₊₁ And event type dataset S _type Constructing event type all negative sample dataset D _-1 ；

Using event element positive sample dataset D ₊₂ Trigger word set S _trigger Event element set S _argument And a corresponding set S of event types and individual event roles _{type_role} Construction of event element random negative sample dataset D _-2 ；

Will D ₊₁ 、D ₊₂ 、D _-1 、D _-2 Mixing and scrambling to finally obtain a model training data set D _all 。

5. The method of claim 4, wherein the event type positive sample dataset D ₊₁ And event element positive sample dataset D ₊₂ The method is constructed by the following steps:

A1. extracting an event type e corresponding to a certain event contained in the text data text_p of the original data set _type Trigger word w _trigger Event role e _{role_1} ～e _{role_n} Corresponding event element w _{arg_1} ～w _{arg_n} (n is the number of event roles that the event contains, also equal to the number of event elements); the input of the event type positive sample for constructing the event is text_p+e _type The "+" trigger word "is output as w _trigger The method comprises the steps of carrying out a first treatment on the surface of the The input of the positive sample of the event element for constructing the event is text_p+sample _arg Output is w _{arg_1} ～w _{arg_n} The method comprises the steps of carrying out a first treatment on the surface of the Wherein the event element prompts prompt _arg Obtainable by the following formula:

A2. constructing positive samples of event types and positive samples of event elements by using the method in (1) for each event in text_p of the text data to obtain a positive sample data set D of the event types ₊₁ And event element positive sample dataset D ₊₂ 。

6. The method of claim 4, wherein the event type full negative sample dataset D _-1 The method is constructed by the following steps:

B1. e of positive samples of a certain event type _type Sequentially changing into event type data sets S _type The target output is null for other event types of the event, so as to obtain an event type full negative sample of the event;

B2. positive sample dataset D for event type ₊₁ The method of (1) is used for all events in the database, and an event type full negative sample data set D is constructed _-1 。

7. The method of claim 4, wherein the event element randomizes a negative sample dataset D _-2 The method is constructed by the following steps:

(1) At D ₊₂ Finding out positive samples of all event elements of a certain event, and finding out prompt of all event elements in the positive samples of the event elements _arg Form a set S _prompt ；

(2) From S _trigger Randomly selecting a trigger word to obtain w _{trigger_random} The method comprises the steps of carrying out a first treatment on the surface of the From S _{type_role} Randomly selecting an element to obtain an event type e _{type_random} An event role e _{role_random} The position p of the event role;

(3) From the event element set S _argument Randomly selecting p event elements to obtain w _{arg_r_1} ～w _{arg_r_p} The event element random prompt is obtained by combining the following formats _{arg_random} ；

prompt _{arg_random} ＝e _{type_random} +w _{trigger_random} +w _{arg_r_1} +…+w _{arg_r_p} +e _{role_random}

(4) Judging prompt _{arg_random} Whether or not to exist in S _prompt If so, repeating the steps 2, 3 and 4, and if not, using promt _{arg_random} Negative samples were constructed and promtt was performed _{arg_random} Adding S _prompt ；

(5) Repeating the steps (1) - (4) until 5n event element random negative samples are obtained;

(6) Pair D ₊₂ All event samples in the sequence are constructed by using the methods in (1) to (5) to obtain an event element random negative sample data set D _-2 。

8. The method of claim 1, wherein training the T5 model comprises:

training the model into a data set D _all Dividing according to a certain proportion to obtain a training set D _train Verification set D _eval And test set D _test ；

Using training set D _train Fine-tuning the T5 model for training n rounds, and finishing each round of training by using the verification set D _eval Performing verification, taking a round of model with the best verification result as a final model, and using a test set D _test Testing to finally obtain a trained model M _trained ；

Model loss is calculated and parameters are updated during training using the following formula:

Loss＝CrossEntropy(x _pred ,x _gold )

wherein x is _pred To predict the result, x _gold Is output as a target.

9. The method of claim 1, wherein gradually constructing the prediction sample set for each step comprises:

constructing a first-step prediction sample set D based on text to be extracted and event feature data set _{step_1} ；

Based on text to be extracted, event feature data set and previous step model M _trained The prediction result of (1) constructs prompt information promtt by text+promptStructure construction of prediction sample set of next model, realizing construction of 2 nd-n+1 th step prediction sample set D in sequence _{step_2} ～D _{step_(n+1)} 。

10. The method according to claim 9, characterized in that the trained model M _trained The method is used for obtaining the prediction result of each step based on the prediction sample set of each step, and integrating to obtain a final extraction result, and comprises the following steps:

will D _{step_1} Input model M _trained Obtaining all trigger words p contained in the text of the first-step prediction result _trigger ；

In the format text+sample _{_X} Constructing a2 nd to n+1 th step prediction sample set D _{step_2} ～D _{step_(n+1)} D is to _{step_x} Input model M _trained Obtaining each trigger word

Corresponding event type->

Is the x-1 th event role->

Corresponding x-1 event element->

Wherein x is ∈ [2, n+1 ]]The method comprises the steps of carrying out a first treatment on the surface of the Wherein prompt is _{_X} Expressed as:

and combining the prompting information of the last step with the extraction result to obtain a complete event.

Images