Event argument extraction method and system based on prediction iteration bidirectional span prediction

Abstract

The present invention proposes an event argument extraction method and system based on prediction iteration bidirectional span prediction, the method comprising: inputting a chapter and giving a prompt template, using a pre-trained language model to perform encoding and decoding processing on the chapter and the prompt template in sequence, so as to obtain the embedding representation of the chapter and the embedding representation of the role respectively, and based on the role learning matrix mechanism, the embedding representation of the role is interacted with the learning vectors of different properties in different directions respectively, so as to obtain the role representation of the predictor of different properties in different directions. Under the framework of the prediction iteration strategy and the bidirectional span prediction strategy, the present invention formulates a role fusion update mechanism, which enables the spans of bidirectional prediction to be fused with each other in one iteration process for each role, and all the prediction spans of multiple iterations can also interact with each other, so as to obtain a semantically rich role representation.

Description

Technical Field

[0001] The present invention relates to the technical field of information extraction, and in particular to an event argument extraction method and system based on prediction iteration bidirectional span prediction. Background Art

[0002] In order to expand the scope of application of the task and better apply it to practical scenarios, the research task of extracting event arguments at the chapter level has shifted from entity arguments to complex arguments. At present, some research results have been achieved in span-oriented event argument extraction methods, including role classification models based on candidate spans, sequence-to-sequence generation models based on reading comprehension, and role-based span prediction models.

[0003] In the role-based span prediction strategy, the existing models have the following shortcomings: 1. The role semantics are only enriched at the event ontology pattern level, so that the semantics of role representation have no difference in different event samples, or the chapter semantics are simply integrated into the role representation through the event type, and the role semantics contained in the event samples contained in the chapter are not fully utilized; 2. The span prediction method is single, and span prediction is not performed from different directions; 3. All roles are modeled through a shared learning vector, which increases the difficulty of model learning. Summary of the invention

[0004] In view of the above situation, the main purpose of the present invention is to propose an event argument extraction method and system based on prediction iteration bidirectional span prediction to solve the above technical problems.

[0005] The present invention proposes an event argument extraction method based on prediction iteration bidirectional span prediction, the method comprising the following steps:

[0006] Step 1: construct a prediction iteration module based on the prediction iteration mechanism, construct a prediction span interaction module based on the bidirectional span prediction mechanism and the role fusion update mechanism, and construct a span prediction module based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model;

[0007] Step 2: Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and the prompt template in sequence to obtain the embedded representation of the passage and the embedded representation of the character respectively;

[0008] Based on the role learning matrix mechanism, the role embedding representation interacts with the learning vectors of different directions and properties to obtain the role representation of the predictor of different directions and properties;

[0009] Step 3: Use the prediction iteration module to fuse the role representations of predictors of different directions and properties to obtain the iteratively updated role representation;

[0010] Step 4: Based on the prediction span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character;

[0011] Step 5: Use the span prediction module to predict the span distribution of the final representation of the character and the embedded representation of the chapter to obtain the prediction result, and construct the cross entropy loss based on the prediction result;

[0012] The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model;

[0013] The argument results are obtained based on the optimized event argument extraction model.

[0014] The present invention also proposes an event argument extraction system based on prediction iteration bidirectional span prediction, the system comprising:

[0015] Building blocks for:

[0016] A prediction iteration module is constructed based on the prediction iteration mechanism, a prediction span interaction module is constructed based on the bidirectional span prediction mechanism and the role representation fusion update mechanism, and a span prediction module is constructed based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model.

[0017] Represents a module for:

[0018] Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and prompt template in turn to obtain the embedded representation of the passage and the embedded representation of the character respectively;

[0019] Predictor discriminator module for:

[0020] Based on the role learning matrix mechanism, the role embedding representation interacts with the learning vectors of different directions and properties to obtain the role representation of the predictor of different directions and properties;

[0021] Iteration module, used to:

[0022] The prediction iteration module is used to fuse the role representations of predictors with different properties in different directions to obtain the iteratively updated role representation;

[0023] Span Interaction Module for:

[0024] Based on the prediction span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character;

[0025] Prediction module for:

[0026] The span prediction module is used to predict the span distribution of the final representation of the character and the embedded representation of the chapter, and the prediction results are obtained. The cross entropy loss is constructed based on the prediction results.

[0027] The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model;

[0028] The argument results are obtained based on the optimized event argument extraction model.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The present invention develops a role fusion update mechanism under the framework of the prediction iteration strategy and the bidirectional span prediction strategy. This mechanism enables the spans of bidirectional prediction to be fused with each other during one iteration for each role, and all prediction spans of multiple iterations can also interact with each other, thereby obtaining a semantically rich role representation;

[0031] 2. Based on the recognition characteristics of span, the present invention proposes a bidirectional span prediction strategy, which is combined with the prediction iteration strategy to enable the model to determine the start and end positions of spans from different directions according to different roles and chapter instances, thereby reducing the difficulty of span prediction of the model;

[0032] 3. The present invention allocates an independent learning vector to each role through the role learning matrix mechanism, which is specifically responsible for learning the role's current forward or reverse and start or end span prediction, thereby achieving parameter independence, reducing conflicts in model training, and reducing the difficulty of model learning.

[0033] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of the steps of the event argument extraction method based on prediction iteration bidirectional span prediction proposed by the present invention;

[0035] Figure 2 This is a model framework diagram of the event argument extraction method based on prediction iteration bidirectional span prediction proposed by the present invention;

[0036] Figure 3 This is a system structure diagram of the event argument extraction system based on prediction iteration and bidirectional span prediction proposed by the present invention. DETAILED DESCRIPTION

[0037] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0038] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0039] See also Figure 1 This embodiment provides an event argument extraction method based on prediction iteration bidirectional span prediction, the method comprising the following steps:

[0040] Step 1: Construct a prediction iteration module based on the prediction iteration mechanism, construct a prediction span interaction module based on the bidirectional span prediction mechanism and the role representation fusion update mechanism, and construct a span prediction module based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model.

[0041] Step 2: Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and the prompt template in sequence to obtain the embedded representation of the passage and the embedded representation of the character respectively;

[0042] Based on the role learning matrix mechanism, the embedded representation of the role is interacted with the learning vectors of different properties in different directions to obtain the role representation of predictors of different properties in different directions.

[0043] See also Figure 2 In step 2, the passage is input and a prompt template is given. The passage and the prompt template are encoded and decoded in turn using the pre-trained language model to obtain the embedded representation of the passage and the embedded representation of the role respectively. The following relationship exists in the corresponding process:

[0044] ;

[0045] in, Indicates chapter Pre-trained language model encoder The output chapter shows that Indicates chapter Pre-trained language model encoder Decoder with pre-trained language model The embedded representation generated after processing, Indicates that the prompt template has been pre-trained by the language model The embedded representation generated after processing, Represents a prompt template, Indicates the first kind of role, Represents the index of the role in the prompt template, using express Middle role The embedded representation of .

[0046] Based on the role learning matrix mechanism, the role embedding representation is interacted with the learning vectors of different directions and properties respectively to obtain the role representation of the predictor of different directions and properties. Specifically, the following sub-steps are included:

[0047] Using the learning vector of the starting position of the character's forward span predictor, the starting position of the character's forward span predictor is predicted to obtain the character representation of the starting position of the first iteration of the forward span predictor. The following relationship exists in the corresponding process:

[0048] ;

[0049] in, The role representation representing the starting position of the first iteration of the forward span predictor, Representing roles The learning vector of the starting position of the forward span predictor, Representing roles The embedding representation of Indicates the positive starting position, Indicates the role representation of the positive starting position, A character representation representing the forward starting position of the first iteration.

[0050] Using the learning vector of the end position of the character's forward span predictor, the end position of the character's forward span predictor is predicted to obtain the character representation of the end position of the first iteration of the forward span predictor. The following relationship exists in the corresponding process:

[0051] ;

[0052] in, The role representation representing the end position of the first iteration of the forward span predictor, Representing roles The learned vector of the end position of the forward span predictor, Indicates the positive end position, Indicates the role representation of the positive end position, A character representation representing the positive end position of the first iteration.

[0053] Using the learning vector of the starting position of the character's reverse span predictor, the starting position of the character's reverse span predictor is predicted to obtain the character representation of the starting position of the first iteration of the reverse span predictor. The following relationship exists in the corresponding process:

[0054] ;

[0055] in, The role representation representing the starting position of the first iteration of the backward span predictor, Representing roles The learning vector of the starting position of the reverse span predictor, Indicates the reverse start position, The character representation indicating the reverse start position, A character representation indicating the reverse start position of the first iteration.

[0056] Using the learning vector of the end position of the character's reverse span predictor, the end position of the character's reverse span predictor is predicted to obtain the character representation of the end position of the first iteration of the reverse span predictor. The following relationship exists in the corresponding process:

[0057] ;

[0058] in, The role representation representing the end position of the first iteration of the backward span predictor, Representing roles The learned vector of the end position of the reverse span predictor, Indicates the reverse end position, The character representation indicating the reverse end position, A character representation indicating the reverse end position of the first iteration.

[0059] Specifically, after the role's embedded representation interacts with the role learning vector, the role's embedded representation can be made discriminative for different span predictors. Since existing studies use the same learning vector for all roles under all event types, it is not conducive to learning the role's discriminative ability. For this reason, the present invention sets a learning vector for each role, which is specifically responsible for learning the role's span prediction for different span predictors. The learning vectors of all roles are combined into a role learning matrix.

[0060] Furthermore, for the role , the learning vector of the start position of the forward span predictor is expressed as , then the learning vectors of the starting position of the forward span predictor of all characters can form a matrix, that is, the character learning matrix of the starting position of the forward span predictor. Similarly, the character learning matrix of the ending position of the forward span predictor, the character learning matrix of the starting position of the reverse span predictor, and the character learning matrix of the ending position of the reverse span predictor can be formed.

[0061] It should be noted that the attached Figure 2 middle Represents dot product.

[0062] Step 3: Use the prediction iteration module to fuse the role representations of predictors of different directions and properties to obtain the iteratively updated role representation.

[0063] In step 3, the prediction iteration module is used to fuse the role representations of predictors of different directions and properties to obtain the iteratively updated role representation, which specifically includes the following sub-steps:

[0064] In one iteration cycle, the forward span predictor The role representation and forward span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the forward span predictor. The span prediction result of the iteration has the following relationship in the corresponding process:

[0065] ;

[0066] in, represents the forward span predictor Second iteration role The distribution of the starting position in the entire chapter, represents the forward span predictor Second iteration role The distribution of ending positions throughout the chapter, represents taking the logarithm, represents the forward span predictor The role representation of the starting position of the iteration, represents the forward span predictor The role representation of the ending position of the iteration, represents the embedding representation of the passage, Indicates that it has been processed by the function that obtains the span. Indicates Iterations, Indicates the role in the forward span predictor The span prediction results of the iteration.

[0067] It should be noted that the total number of iterations in one iteration cycle is times, and .

[0068] Put the role first in the forward span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the forward span predictor. Iterations for roles The predicted span indicates that the following relationship exists in the corresponding process:

[0069] ;

[0070] in, represents the forward span predictor Iterations for roles The predicted span represents, represents the mean operation, Indicates that it has been processed by bidirectional LSTM. The role representation and reverse span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the reverse span predictor The span prediction result of the iteration has the following relationship in the corresponding process:

[0071] ;

[0072] in, Represents the reverse span predictor Second iteration role The distribution of the starting position in the entire chapter, Represents the reverse span predictor Second iteration role The distribution of ending positions throughout the chapter, Represents the reverse span predictor The role representation of the starting position of the iteration, Represents the reverse span predictor The role representation of the ending position of the iteration, Indicates the role in the reverse span predictor The span prediction results of the iteration.

[0073] The role is first in the reverse span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the reverse span predictor. Iterations for roles The predicted span indicates that the following relationship exists in the corresponding process:

[0074] ;

[0075] in, Represents the reverse span predictor Iterations for roles The span representation of the prediction.

[0076] The forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the starting position of the iteration is the same as the reverse span predictor The role representation of the end position of the iteration is fused to obtain the forward span predictor The role representation of the starting position of the iteration, the corresponding process has the following relationship:

[0077] ;

[0078] in, represents the forward span predictor The role representation of the starting position of the iteration.

[0079] The forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the end position of the iteration is the same as the reverse span predictor The role representation of the starting position of the iteration is fused to obtain the forward span predictor The role representation of the end position of the iteration, the corresponding process has the following relationship:

[0080] ;

[0081] in, represents the forward span predictor The role representation of the ending position of the iteration.

[0082] Step 4: Based on the predicted span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character.

[0083] In step 4, based on the predicted span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character, which specifically includes the following sub-steps:

[0084] The span representation predicted by the forward span predictor for each iteration and the span representation predicted by the reverse span predictor for each iteration are combined and sent to the Bi-LSTM network to obtain the representation of the predicted span after one iteration cycle of the role. The following relationship exists in the corresponding process:

[0085] ;

[0086] in, Representing roles Representation of the prediction span after one iteration cycle, represents the first iteration of the forward span predictor for the role The predicted span represents, Represents the first iteration of the reverse span predictor for the role The predicted span represents, represents the forward span predictor Iterations for roles The predicted span represents, Represents the reverse span predictor Iterations for roles The span representation of the prediction.

[0087] The representation of the predicted span after one iteration of the character is fused with the representation of the character at the start position of the iteration completed in the forward predictor and the representation of the character at the end position of the iteration completed in the forward predictor, respectively, to obtain the character representation of the final start position of the character and the character representation of the final end position of the character. The following relationship exists in the corresponding process:

[0088] ;

[0089] in, Representing roles The character representation of the final starting position, Representing roles The final ending position of the character is indicated. The character representing the starting position, The character indicating the end position, represents the forward predictor The role representation of the starting position of the iteration, represents the forward predictor The role representation of the starting position of the iteration.

[0090] Step 5: Use the span prediction module to predict the span distribution of the final representation of the character and the embedded representation of the chapter to obtain the prediction result, and construct the cross entropy loss based on the prediction result;

[0091] The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model;

[0092] The argument results are obtained based on the optimized event argument extraction model.

[0093] In step 5, the span prediction module is used to predict the span distribution of the final representation of the character and the embedded representation of the chapter to obtain the prediction result, and the cross entropy loss is constructed based on the prediction result, which specifically includes the following sub-steps:

[0094] The character representation of the final starting position of the character and the character representation of the final ending position of the character are multiplied with the embedding representation of the passage to obtain the distribution of the predicted span starting position of the character in the passage and the distribution of the predicted span ending position of the character in the passage. The following relationship exists in the corresponding process:

[0095] ;

[0096] in, Representing roles In the chapter The distribution of the predicted span start positions, Representing roles In the chapter The distribution of the span end positions predicted above.

[0097] The cross entropy loss is constructed based on the distribution of the span start position predicted by the character in the chapter and the distribution of the span end position predicted by the character in the chapter. The following relationship exists in the corresponding process:

[0098] ;

[0099] in, Representing roles The probability distribution of the span start position, Representing roles The probability distribution of the span end position, Indicates the number of chapters contained in the corpus, Indicates the number of roles in the prompt template. Indicates the index of the chapter. represents the cross entropy loss.

[0100] Please refer to Figure 3 The present invention also provides an event argument extraction system based on prediction iteration bidirectional span prediction, the system applies the event argument extraction method based on prediction iteration bidirectional span prediction described above, and the system comprises:

[0101] Building blocks for:

[0102] A prediction iteration module is constructed based on the prediction iteration mechanism, a prediction span interaction module is constructed based on the bidirectional span prediction mechanism and the role representation fusion update mechanism, and a span prediction module is constructed based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model.

[0103] Represents a module for:

[0104] Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and prompt template in turn to obtain the embedded representation of the passage and the embedded representation of the character respectively;

[0105] Predictor discriminator module for:

[0106] Based on the role learning matrix mechanism, the role embedding representation interacts with the learning vectors of different directions and properties to obtain the role representation of the predictor of different directions and properties;

[0107] Iteration module, used to:

[0108] The prediction iteration module is used to fuse the role representations of predictors with different properties in different directions to obtain the iteratively updated role representation;

[0109] Span Interaction Module for:

[0110] Based on the prediction span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character;

[0111] Prediction module for:

[0112] The span prediction module is used to predict the span distribution of the final representation of the character and the embedded representation of the chapter, and the prediction results are obtained. The cross entropy loss is constructed based on the prediction results.

[0113] The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model;

[0114] The argument results are obtained based on the optimized event argument extraction model.

[0115] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0116] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0117] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for extracting event arguments based on prediction iteration bidirectional span prediction, characterized in that: The method comprises the following steps: Step 1: construct a prediction iteration module based on the prediction iteration mechanism, construct a prediction span interaction module based on the bidirectional span prediction mechanism and the role representation fusion update mechanism, and construct a span prediction module based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model; Step 2: Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and the prompt template in sequence to obtain the embedded representation of the passage and the embedded representation of the character respectively; Based on the role learning matrix mechanism, the role embedding representation interacts with the learning vectors of different directions and properties to obtain the role representation of the predictor of different directions and properties; Step 3: Use the prediction iteration module to fuse the role representations of predictors of different directions and properties to obtain the iteratively updated role representation; Step 4: Based on the prediction span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character; Step 5: Use the span prediction module to predict the span distribution of the final representation of the character and the embedded representation of the chapter to obtain the prediction result, and construct the cross entropy loss based on the prediction result; The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model; The argument results are obtained based on the optimized event argument extraction model.

2. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 1 is characterized in that: In step 2, based on the role learning matrix mechanism, the role embedding representation is interacted with the learning vectors of different directions and properties respectively to obtain the role representation of the predictors of different directions and properties, which specifically includes the following sub-steps: Using the learning vector of the start position of the forward span predictor of the character, predict the start position of the forward span predictor of the character to obtain the character representation of the start position of the first iteration of the forward span predictor; Using the learning vector of the end position of the character's forward span predictor, predict the end position of the character's forward span predictor to obtain the character representation of the end position of the first iteration of the forward span predictor; Using the learning vector of the start position of the reverse span predictor of the character, predict the start position of the reverse span predictor of the character, and obtain the character representation of the start position of the first iteration of the reverse span predictor; The end position of the reverse span predictor of the character is predicted using the learning vector of the end position of the reverse span predictor of the character, and the character representation of the end position of the first iteration of the reverse span predictor is obtained.

3. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 2 is characterized in that: Using the learning vector of the starting position of the character's forward span predictor, the starting position of the character's forward span predictor is predicted to obtain the character representation of the starting position of the first iteration of the forward span predictor. The following relationship exists in the corresponding process: ； in, The role representation representing the starting position of the first iteration of the forward span predictor, Representing roles The learning vector of the starting position of the forward span predictor, Representing roles The embedding representation of Indicates the first kind of role, Represents the index of the role in the prompt template. Indicates the positive starting position, Indicates the role representation of the positive starting position, The role representation representing the forward starting position of the first iteration; In the step of predicting the end position of the character's forward span predictor using the learning vector of the end position of the character's forward span predictor to obtain the character representation of the end position of the first iteration of the forward span predictor, the following relationship exists in the corresponding process: ； in, The role representation representing the end position of the first iteration of the forward span predictor, Representing roles The learned vector of the end position of the forward span predictor, Indicates the positive end position, Indicates the role representation of the positive end position, The character representation indicating the positive end position of the first iteration; In the step of predicting the starting position of the reverse span predictor of the character using the learning vector of the starting position of the reverse span predictor of the character to obtain the character representation of the starting position of the first iteration of the reverse span predictor, the following relationship exists in the corresponding process: ； in, The role representation representing the starting position of the first iteration of the backward span predictor, Representing roles The learning vector of the starting position of the reverse span predictor, Indicates the reverse start position, The character representation indicating the reverse start position, A character representation indicating the reverse start position of the first iteration; In the step of predicting the end position of the reverse span predictor of the character using the learning vector of the end position of the reverse span predictor of the character to obtain the character representation of the end position of the first iteration of the reverse span predictor, the following relationship exists in the corresponding process: ； in, The role representation representing the end position of the first iteration of the backward span predictor, Representing roles The learned vector of the end position of the reverse span predictor, Indicates the reverse end position, The character representation indicating the reverse end position, A character representation indicating the reverse end position of the first iteration.

4. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 3 is characterized in that: In step 3, the role representations of predictors of different directions and properties are fused using a prediction iteration module to obtain an iteratively updated role representation, which specifically includes the following sub-steps: In one iteration cycle, the forward span predictor The role representation and forward span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the forward span predictor. The span prediction result of the iteration; Put the role first in the forward span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the forward span predictor. Iterations for roles The predicted span representation; The reverse span predictor The role representation and reverse span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the reverse span predictor The span prediction result of the iteration; The role is first in the reverse span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the reverse span predictor. Iterations for roles The predicted span representation; The forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the starting position of the iteration is the same as the reverse span predictor The role representation of the end position of the iteration is fused to obtain the forward span predictor The role representation of the starting position of the iteration; The forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the end position of the iteration is the same as the reverse span predictor The role representation of the starting position of the iteration is fused to obtain the forward span predictor The role representation of the ending position of the iteration.

5. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 4 is characterized in that: In one iteration cycle, the forward span predictor The role representation and forward span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the forward span predictor. The span prediction result of the iteration has the following relationship in the corresponding process: ； in, represents the forward span predictor Second iteration role The distribution of the starting position in the entire chapter, represents the forward span predictor Second iteration role The distribution of ending positions in the entire chapter, represents taking the logarithm, represents the forward span predictor The role representation of the starting position of the iteration, represents the forward span predictor The role representation of the ending position of the iteration, represents the embedding representation of the passage, Indicates that it has been processed by the function that obtains the span. Indicates Iterations, Indicates the role in the forward span predictor The span prediction result of the iteration; In the forward span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the forward span predictor. Iterations for roles In the steps of the predicted span representation, the following relationship exists in the corresponding process: ； in, represents the forward span predictor Iterations for roles The predicted span represents, represents the averaging operation, Indicates that it has been processed by bidirectional LSTM; Before setting the reverse span predictor The role representation and reverse span predictor of the starting position of the iteration The role representation at the end position of the iteration is multiplied with the embedding representation of the passage, and then Function is used to process and obtain the role in the reverse span predictor In the step of predicting the span result of the iteration, the following relationship exists in the corresponding process: ； in, Represents the reverse span predictor Second iteration role The distribution of the starting position in the entire chapter, Represents the reverse span predictor Second iteration role The distribution of ending positions throughout the chapter, Represents the reverse span predictor The role representation of the starting position of the iteration, Represents the reverse span predictor The role representation of the ending position of the iteration, Indicates the role in the reverse span predictor The span prediction result of the iteration; In the reverse span predictor The span prediction results obtained in the first iteration are sent to the Bi-LSTM network and averaged to obtain the reverse span predictor. Iterations for roles In the steps of the predicted span representation, the following relationship exists in the corresponding process: ； in, Represents the reverse span predictor Iterations for roles The predicted span representation; In the forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the starting position of the iteration is the same as the reverse span predictor The role representation of the end position of the iteration is fused to obtain the forward span predictor In the step of character representation of the starting position of the iteration, the following relationship exists in the corresponding process: ； in, represents the forward span predictor The role representation of the starting position of the iteration; In the forward span predictor Iterations for roles Predicted span representation, reverse span predictor Iterations for roles Predicted span representation, forward span predictor The role representation of the end position of the iteration is the same as the reverse span predictor The role representation of the starting position of the iteration is fused to obtain the forward span predictor In the step of character representation of the end position of the iteration, the following relationship exists in the corresponding process: ； in, represents the forward span predictor The role representation of the ending position of the iteration.

6. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 5 is characterized in that: In step 4, based on the predicted span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character, which specifically includes the following sub-steps: The span representation predicted by the forward span predictor at each iteration and the span representation predicted by the reverse span predictor at each iteration are combined and sent to the Bi-LSTM network to obtain the representation of the predicted span after one iteration cycle of the role; The representation of the predicted span after one iteration cycle of the character is fused with the character representation of the starting position of the iteration completed in the forward predictor and the character representation of the ending position of the iteration completed in the forward predictor, respectively, to obtain the character representation of the final starting position of the character and the character representation of the final ending position of the character.

7. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 6 is characterized in that: The span representation predicted by the forward span predictor for each iteration and the span representation predicted by the reverse span predictor for each iteration are combined and sent to the Bi-LSTM network to obtain the representation of the predicted span after one iteration cycle of the role. The following relationship exists in the corresponding process: ； in, Representing roles Representation of the prediction span after one iteration cycle, represents the first iteration of the forward span predictor for the role The predicted span represents, Represents the first iteration of the reverse span predictor for the role The predicted span represents, represents the forward span predictor Iterations for roles The predicted span represents, Represents the reverse span predictor Iterations for roles The predicted span represents, Indicates the total number of iterations in an iteration cycle; In the step of fusing the representation of the predicted span after one iteration cycle of the character with the representation of the character at the starting position of the iteration completed in the forward predictor and the representation of the character at the ending position of the iteration completed in the forward predictor, respectively, to obtain the representation of the character at the final starting position of the character and the representation of the character at the final ending position of the character, the following relationship exists in the corresponding process: ； in, Representing roles The character representation of the final starting position, Representing roles The final ending position of the character is indicated. The character representing the starting position, The character indicating the end position, represents the forward predictor The role representation of the starting position of the iteration, represents the forward predictor The role representation of the starting position of the iteration, Indicates the total number of iterations in one iteration cycle.

8. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 7 is characterized in that: In step 5, the span prediction module is used to predict the span distribution of the final representation of the character and the embedded representation of the chapter to obtain a prediction result, and a cross entropy loss is constructed based on the prediction result, which specifically includes the following sub-steps: The character representation of the character's final start position and the character representation of the character's final end position are respectively multiplied by the embedding representation of the passage to obtain the distribution of the character's predicted span start position and the distribution of the character's predicted span end position in the passage; The cross entropy loss is constructed based on the distribution of the span start position predicted by the character in the passage and the distribution of the span end position predicted by the character in the passage.

9. The event argument extraction method based on prediction iteration bidirectional span prediction according to claim 8, characterized in that: The character representation of the final starting position of the character and the character representation of the final ending position of the character are multiplied with the embedding representation of the passage to obtain the distribution of the predicted span starting position of the character in the passage and the distribution of the predicted span ending position of the character in the passage. The following relationship exists in the corresponding process: ； in, Representing roles In the chapter The distribution of the predicted span start positions, Representing roles In the chapter The distribution of the span end positions predicted above; In the step of constructing the cross entropy loss based on the distribution of the span start position predicted by the character in the passage and the distribution of the span end position predicted by the character in the passage, the following relationship exists in the corresponding process: ； in, Representing roles The probability distribution of the span start position, Representing roles The probability distribution of the span end position, Indicates the number of chapters contained in the corpus, Indicates the number of roles in the prompt template. Indicates the index of the chapter. Represents the index of the role in the prompt template. represents the cross entropy loss.

10. An event argument extraction system based on prediction iteration bidirectional span prediction, characterized in that: The system applies the event argument extraction method based on prediction iteration bidirectional span prediction according to any one of claims 1 to 9, and the system comprises: Building blocks for: A prediction iteration module is constructed based on the prediction iteration mechanism, a prediction span interaction module is constructed based on the bidirectional span prediction mechanism and the role representation fusion update mechanism, and a span prediction module is constructed based on the span prediction mechanism. The prediction iteration module, the prediction span interaction module and the span prediction module constitute an event argument extraction model. Represents a module for: Input a passage and a given prompt template, and use the pre-trained language model to encode and decode the passage and prompt template in turn to obtain the embedded representation of the passage and the embedded representation of the character respectively; Predictor discriminator module for: Based on the role learning matrix mechanism, the role embedding representation interacts with the learning vectors of different directions and properties to obtain the role representation of the predictor of different directions and properties; Iteration module, used to: The prediction iteration module is used to fuse the role representations of predictors with different properties in different directions to obtain the iteratively updated role representation; Span Interaction Module for: Based on the prediction span interaction module, all span representations predicted within one iteration cycle of the character and the iteratively updated character representation are processed to obtain the final representation of the character; Prediction module for: The span prediction module is used to predict the span distribution of the final representation of the character and the embedded representation of the chapter, and the prediction results are obtained. The cross entropy loss is constructed based on the prediction results. The event argument extraction model is optimized using cross entropy loss to obtain an optimized event argument extraction model; The argument results are obtained based on the optimized event argument extraction model.

Images