Event theory role classification method and system based on role consistency

Abstract

The application provides an event argument role classification method and system based on role consistency, adopts a graph neural network structure, integrates the estimated role probability distribution of each argument into the modeling of the association features of an event trigger and an event argument, performs multi-order graph neural network interaction on a star-shaped graph connected by the trigger and each argument, enables each argument to perceive the role information of other arguments, and thus improves the overall consistency of the roles of all arguments in an event.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information extraction, and particularly relates to a method for improving the prediction effect of event argument roles in an event extraction task. BACKGROUND

[0002] Event, as a structured representation of information, refers to an actual occurrence involving certain participants. As a special information extraction task, the goal of event extraction is to extract instances of predefined event types from a given text. An event is generally composed of a trigger and arguments, where the trigger is the word in the text that most clearly expresses the occurrence of the event, and is generally the core verb of the sentence in which the event occurs; the arguments are entities related to the event and play a certain role in the event. Unlike the above example, which extracts structured events from text, the structure of the event conforms to the structure of a predefined event type; an entity is an instance of a concept, such as "Beijing", which is an instance of "place" or "city". The two are related, and the arguments in the event instance must be entities. Generally, event extraction can be divided into two tasks: trigger extraction and argument extraction. The goal of the trigger extraction task is to find the trigger of the event and determine the event type; the goal of the argument extraction task is to determine whether an entity in the text is a relevant argument of the event, and if so, to determine the role the entity plays in the event. Predicting the role is actually classifying the argument, where the role can be "place", "time", "buyer", etc. For example, in the sentence "A arrived in Paris on the 29th", "arrived" is the trigger, "A" plays the role of "visitor", "29th" plays the role of "time", and "Paris" plays the role of "place". Through the advancement of event extraction research in recent years, existing methods have good results for the trigger extraction task, while the argument extraction task is more difficult due to the complex structure of the argument role ontology, and has become a bottleneck for the entire event extraction task. The focus of the argument role prediction task is on the context semantic modeling of candidate argument entities and the association feature modeling between the argument and the event trigger and other arguments. By accurately predicting the argument role, the following technologies will have a significant performance improvement: 1. Information retrieval: a new index can be constructed in the dimension of extracted argument entities (persons, organizations, institutions, etc.) to quickly retrieve documents containing a certain specific event type or containing a specific entity argument involved in the event; 2. Event knowledge graph: the extracted argument roles and triggers form an event, which participates in the construction of an event knowledge graph, providing a foundation for downstream tasks such as event prediction and event relationship reasoning.

[0003] For the context semantic modeling of candidate argument entities, with the development of deep learning, existing methods use increasingly structured text semantic encoding network structures, from convolutional neural networks to recurrent neural networks, to recurrent neural networks with fast links to alleviate the problem of long-distance dependence information loss, and to the latest graph neural network structure, which converts sequence modeling into graph information modeling. Graph neural networks can well solve the problem of poor long-distance dependence modeling in text semantic encoding due to their unordered information transmission. However, existing methods based on graph neural networks are based on dependency relationship graphing to model semantic information, and dependency relationship needs to be parsed based on third-party tools, which will introduce noise from third-party tools into argument role context semantic encoding.

[0004] For the modeling of the association features between arguments and trigger words and other arguments, existing methods model the association features between arguments and other event elements through memory matrix, attention mechanism and other methods, but the problem is that when role prediction classification is performed after feature modeling, it is independently predicted, and whether different argument roles are correct, complete and compatible is not considered, that is, the consistency between roles is not modeled, which makes the final predicted argument roles prone to inconsistent results such as contradictions and omissions.

[0005] In summary, the prior art has the problems of not modeling the consistency of different argument roles in the same event and the easy introduction of third-party tool noise in graph neural network semantic modeling. SUMMARY

[0006] In order to overcome the deficiencies in the prior art, the present application provides an event argument role prediction method based on role consistency modeling. The method uses a graph neural network structure to integrate the estimated role probability distribution of each argument into the modeling of the association features between the event trigger word and the event argument, and performs multi-order graph neural network interaction on the star graph connected by the trigger word and each argument, so that each argument can perceive the role information of other arguments, thereby improving the overall consistency of all argument roles in the event.

[0007] In view of the deficiencies in the prior art, the present application proposes an event argument role classification method based on role consistency, which comprises:

[0008] A training corpus preprocessing step is used to obtain a training corpus of annotated event argument role categories and event trigger words, perform word segmentation on the text of the training corpus, and obtain the ID of each word in the pre-training dictionary of the language representation model BERT according to the pre-training dictionary of the language representation model BERT; input the word sequence composed of all word IDs into the multi-layer translation model Transformer of BERT to pre-encode the word sequence and obtain the word representation vector;

[0009] The candidate argument semantic encoding step aggregates the event trigger word and the word representation vector to obtain a context semantic expression vector of the trigger word, splices the context semantic expression vector and the word representation vector and fuses them through a linear network to obtain a semantic expression vector of each candidate argument;

[0010] The candidate argument feature encoding step encodes the semantic expression vector of the candidate argument according to the event type to select a corresponding feature modeling network to obtain a feature expression vector of each candidate argument;

[0011] The candidate argument role decoding step decodes the feature expression vector of each candidate argument from two dimensions of global and local to obtain a role probability distribution matrix in the two dimensions;

[0012] The role consistency information interaction step performs multi-order graph neural network interaction on the trigger-argument connection graph, and fuses the role probability distribution matrix in the two dimensions into the information interaction based on the trigger-argument connection graph to obtain a global expression vector of each candidate argument fused with the role probability information as a new semantic expression vector of the candidate argument;

[0013] The iterative role classification step executes the candidate argument feature encoding step, the candidate argument role decoding step and the role consistency information interaction step again until the role probability distribution matrix obtained by decoding converges, and the role category with the highest probability is taken as the output of the final argument role prediction.

[0014] The event argument role classification method based on role consistency, wherein the training corpus preprocessing step comprises:

[0015] The text is segmented, and the input of the same batch of words is padded to the same length according to the longest text length of the batch of words;

[0016] The input word, the word sequence T is pre-encoded by the BERT language model to obtain a word representation vector with dynamic semantic expression ;

[0017] C = BERT (T)

[0018] The candidate argument semantic encoding step comprises:

[0019] The event trigger word and the word representation vector are aggregated to obtain respective context semantic expression vectors, the context semantic expression vectors and the word representation vectors are spliced and fused through a linear network to obtain a semantic expression vector of each candidate argument ;

[0020] The candidate argument feature encoding step includes: feature encoding of each candidate argument A by an event type E related feature encoder to obtain its feature expression vector U = {𝑢0, 𝑢1,..., 𝑢 𝐾};

[0021] The event argument role classification method based on role consistency, wherein the candidate argument role decoding step includes:

[0022] The feature vectors U local and U global

[0023]

[0024] The decoding is performed by the local and global dimension softmax to obtain the probability distribution matrix P k of each argument A m under different roles R A , and the probability distribution matrix P m of each role R k under different arguments A R .

[0025]

[0026] Wherein represents the probability of the kth argument belonging to the mth role in the E event, represents the probability of the mth role being expressed by the kth argument; according to the role label L, the cross entropy loss function is made:

[0027]

[0028]

[0029]

[0030] Wherein, represents the true probability label of the kth argument belonging to the mth role, and if there is a role overlap phenomenon, the probability is equally divided by multiple roles; represents the true probability label of the mth role being assumed by the kth argument, and the probability is equally divided when the same role is assumed by multiple arguments;

[0031] The event argument classification method based on role consistency, wherein the role consistency information interaction step comprises: using a graph neural network RGCN to perform message passing on the trigger-argument connection graph, and each node obtains the feature and role probability distribution information of its neighbor in each layer of the graph neural network; the event trigger is connected with all arguments in the trigger-argument connection graph, and information is passed through two layers of graph neural networks, and each argument includes all node information including the trigger; for each event type E, corresponding independent model parameters are set, and in the graph message passing stage, the trigger representation and the semantic expression vector of the argument are obtained as the initialized representation of the graph node, the edge weight matrix is obtained by the label probability matrix of the role label information aggregation module and , represents a cross-node edge, represents an internal self-loop edge of each node; the self-loop edge is set as and the tensor is set as the linear transformation matrix corresponding to the trigger type and the linear transformation matrix set corresponding to all argument roles under the type; the edge weight matrix of the self-loop edge and the non-self-loop edge is obtained by weighting the tensor by the global and local label probabilities and :

[0032]

[0033]

[0034] The argument and trigger representation are mapped into a new space fused with label probability information, and a first layer of graph network information interaction is performed to obtain the trigger and argument node representation after interaction and ;

[0035]

[0036]

[0037] The same operation is performed again using the new node representation to complete the second-order graph information interaction:

[0038]

[0039]

[0040] The iterative role classification step comprises: taking as a new fusion vector representation performing the candidate argument feature encoding step again, then return to the first layer of the graph network information interaction position to perform the cycle again until the role probability distribution matrix obtained by decoding converges

[0041] The application also provides an event argument role classification system based on role consistency, which comprises the following modules:

[0042] The training corpus preprocessing module obtains training corpus of annotated event argument role categories and event trigger words, performs word segmentation on the training corpus text, and obtains the ID of each word in the pre-training dictionary of the language representation model BERT according to the pre-training dictionary of the language representation model BERT; the word sequence composed of all word IDs is input into the multi-layer translation model Transformer of BERT to pre-encode the word sequence and obtain word representation vectors;

[0043] The candidate argument semantic encoding module aggregates the event trigger word and the word representation vectors to obtain a context semantic expression vector of the trigger word, splices the context semantic expression vector and the word representation vectors and fuses them through a linear network to obtain a semantic expression vector of each candidate argument;

[0044] The candidate argument feature encoding module encodes the semantic expression vector of the candidate argument according to the corresponding feature modeling network selected according to the event type to obtain a feature expression vector of each candidate argument;

[0045] The candidate argument role decoding module decodes the feature expression vector of each candidate argument from two dimensions of global and local to obtain role probability distribution matrices in the two dimensions;

[0046] The role consistency information interaction module performs multi-order graph neural network interaction on the trigger-argument connection graph, integrates the role probability distribution matrices in the two dimensions into the information interaction based on the trigger-argument connection graph, and obtains a global expression vector of each candidate argument fused with role probability information as a new semantic expression vector of the candidate argument;

[0047] The iterative role classification module performs the candidate argument feature encoding module, the candidate argument role decoding module and the role consistency information interaction module again until the role probability distribution matrix obtained by decoding converges, and the role category with the largest probability is taken as the final argument role prediction output.

[0048] The event argument role classification system based on role consistency, wherein the training corpus preprocessing module comprises the following modules:

[0049] The text is segmented, and the input of the same batch of words is padded to the same length according to the longest text length of the batch of words;

[0050] The input word, the word sequence T is pre-encoded by the BERT language model to obtain a word representation vector with dynamic semantic expression

[0051] C = BERT (T)

[0052] The candidate argument semantic encoding module comprises:

[0053] The event trigger word and the word representation vector are aggregated to obtain respective context semantic expression vectors, the context semantic expression vectors and the word representation vectors are spliced and fused through a linear network to obtain a semantic expression vector of each candidate argument

[0054] The candidate argument feature encoding module comprises: encoding each candidate argument A through an event type E related feature encoder to obtain a feature expression vector U = {u0, u1,..., u 𝐾};

[0055]

[0056] The event argument role classification system based on role consistency, wherein the candidate argument role decoding module comprises:

[0057] The feature vectors U local and U global in local and global dimensions are learned through two fully connected networks FNN:

[0058]

[0059]

[0060] The local dimension of each argument A k under different roles R m The probability distribution matrix P A , and the probability distribution matrix P R of each role R m under different arguments A k in the global dimension are obtained by decoding through the local and global dimensions of softmax:

[0061]

[0062]

[0063] Wherein represents the probability of the kth argument belonging to the mth role in the E event, ​​Let the probability of the nth role being expressed by the nth argument be represented; calculate the cross-entropy loss function based on the role's label φ:

[0064]

[0065]

[0066]

[0067] in, This represents the true probability label of the k-th argument belonging to the m-th role. If there is role overlap, the probability is equally divided among multiple roles. This indicates the true probability label of the m-th role being borne by the k-th argument; when the same role is borne by multiple arguments, the probability is equally distributed.

[0068] The aforementioned event argument role classification system based on role consistency includes a role consistency information interaction module comprising: using a graph neural network (RGCN) to perform message passing on the trigger word-argument connection graph, where each layer of the RGCN allows each node to obtain the features and role probability distribution information of its neighbors; in the trigger word-argument connection graph, the event trigger word is connected to all arguments, and information is passed through a two-layer RGCN, with each argument including information from all nodes, including the trigger word; for each event type E, corresponding independent model parameters are set, and during the graph message passing phase, the trigger word is represented... and the semantic representation vector of arguments As the initial representation of graph nodes, the edge weight matrix Tag probability matrix from the character tag information aggregation module and get, This represents an edge that connects nodes. Represents the self-loop edges within each node; is set with and tensor This serves as the linear transformation matrix corresponding to the trigger word type and the set of linear transformation matrices corresponding to all argument roles under that type; through the label probabilities in both global and local dimensions, the tensor is... The weighted average yields the edge weight matrix for both self-loop and non-self-loop edges. and :

[0069]

[0070]

[0071] The argument unit and trigger word representation mapping is mapped into a new space integrated with label probability information, and the first layer graph network information interaction is carried out to obtain the interactive trigger word and argument unit node representation and ;

[0072]

[0073]

[0074] The same operation is carried out again using the new node representation to complete the second-order graph information interaction:

[0075]

[0076]

[0077] The iterative role classification module includes, as a new fusion vector representation The candidate argument feature encoding module is executed again, Then, the cycle is carried out again at the first layer of the graph network information interaction until the role probability distribution matrix obtained by decoding converges.

[0078] The application further provides a storage medium for storing a program for executing the event argument role classification method with the role consistency.

[0079] The application further provides a client for the event argument role classification system with the role consistency.

[0080] As can be seen from the above scheme, the application has the following advantages: first, the graph convolutional neural network is used to model the features on the trigger word and argument connection graph, avoiding the introduction of noise caused by the use of third-party tools to predict the dependency relationship; second, the role consistency between the candidate arguments is modeled, relieving the role inconsistency problems such as role contradiction, omission and the like that are prone to occur in the previous methods; third, the performance on the event argument role prediction task is improved, and the argument role prediction F1 value of the application reaches 79.2% on the RAMS public data test set. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is the overall flowchart of the application;

[0082] Figure 2 is a detailed introduction diagram of the candidate argument multi-dimensional feature decoding in the application;

[0083] Figure 3 is a detailed introduction diagram of the candidate argument role consistency interaction in the application;

[0084] Figure 4The overall method structure in the application. DETAILED DESCRIPTION

[0085] The application comprises the following steps: 1) training corpus preprocessing, 2) candidate argument semantic encoding, 3) candidate argument feature encoding, 4) candidate argument role decoding, 5) role consistency information interaction, and 6) iterative role prediction. Specifically, the application comprises the following steps:

[0086] 1) Training corpus preprocessing: The training corpus used in the application is selected from the Roles Across Multiple Sentence (RAMS) dataset, which includes multiple events annotated in a news corpus. The processing process is to perform word segmentation on the text using the wordpiece method, then convert each word into an ID corresponding to the BERT pretraining dictionary of the language representation model, and finally add [CLS] and [SEP] special tags consistent with BERT and the training task at the beginning and end of the sentence, respectively. [CLS] is placed at the beginning of the first sentence, and the representation vector obtained by BERT can be used for subsequent classification tasks. [SEP] is used to separate two input sentences, and needs to be added at the end of each sentence.

[0087] 2) Candidate argument semantic encoding: the BERT pretraining language model is used for pre-encoding, and the word id sequence processed in the previous step is input into the multi-layer translation model Transformer of BERT to pre-encode the word sequence. Compared with traditional static word and word vectors, BERT trained on a large corpus can provide more dynamic semantic expression. Then, the BERT encoded word representation vectors corresponding to the trigger word and each given candidate argument entity are aggregated to obtain the context semantic expression vectors of the trigger word and each entity. Then, the trigger word corresponding vector and each entity corresponding vector are spliced and fused through a linear network to obtain the semantic expression vector of each candidate argument. The trigger word corresponding vector refers to the word vector representation corresponding to the trigger word in the text after BERT pre-encoding.

[0088] 3) Candidate argument feature encoding: according to the event type, the corresponding feature modeling network is selected to automatically construct the feature expression vector of each candidate argument. These network structures are completely the same, but the parameters are not shared with each other. The event type can be birth, meeting, transportation, etc.

[0089] 4) Candidate argument role decoding: the feature expression vector of each argument is decoded from the global and local dimensions to obtain the role probability distribution matrix in the two dimensions.

[0090] 5) Role consistency information interaction: the trigger word is connected with all arguments to form a trigger-argument connection graph, and multi-stage graph neural network interaction is performed on the trigger-argument connection graph, the role probability information obtained in the previous step is integrated into the graph-based information interaction, and the global expression vector of each argument fused with the role probability information is obtained as the new feature expression vector of the candidate argument.

[0091] 6) Iterative role prediction: the steps 4 and 5 are iteratively performed until the predicted roles of all arguments are stable, and the stable roles are taken as the output of the final argument role prediction. The stability refers to the stability of the role probability distribution matrix in two dimensions. The final prediction result is obtained by adding the role probabilities in two dimensions, and when the added value exceeds a certain threshold, the role is determined.

[0092] In order to make the above features and effects of the present application more clear and easy to understand, the following embodiments are described in detail below, and the drawings are described as follows.

[0093] The present application proposes an event argument role prediction method based on role consistency modeling, and the overall process of the method is shown in Figure 1 The method mainly includes argument semantic encoding, argument feature encoding, multi-dimensional argument role decoding, and role consistency interaction. The argument semantic encoding is based on the BERT pre-training language model, and the semantic information of the trigger word and the candidate argument is fused by combining the full connection network to obtain the context semantic representation vector of the candidate argument. The argument feature encoding is to extract the features of each candidate argument through the event type related feature encoder. The multi-dimensional argument role decoding is to decode the argument feature vector from the local and global dimensions to obtain the argument role probability distribution in multiple dimensions. The role consistency interaction is to model the role information into the feature vector encoding of the argument. The feature encoding, role decoding and role consistency interaction are iteratively performed to continuously improve the consistency between the predicted roles. The specific method steps include:

[0094] S1, using the BERT pre-training language model to pre-encode the input word sequence to obtain the distributed semantic expression of each word after the BERT pre-training model encoding. Then the average vector expression of each trigger word and candidate argument is obtained by taking the average method, and then the trigger word vector and the candidate argument vector are spliced, and the information is fused by using the full connection network to obtain the final semantic expression of the candidate argument.

[0095] S2, the features of the candidate argument are extracted by the event type related feature encoder based on the connection full network, which is used for subsequent classification and prediction of roles. Here, the ReLU activation function is used.

[0096] S3. Decode the feature vectors of the above candidate arguments using softmax from both local and global dimensions to predict the probability distribution of the argument roles in multiple dimensions.

[0097] S4. Use a graph neural network to model the role probability into the feature vector of the argument in the form of edge weights to obtain a new candidate argument feature vector. Iterate with step S4 until the roles decoded from all candidate arguments tend to be stable.

[0098] Specifically, S1 includes three sub-steps as follows.

[0099] S101. Training data preprocessing. The text is segmented using the WordPieceTokenizer module in the Transformers library, and the input of the same batch of words is padded to the same length according to the longest text length of that batch of words.

[0100] S102, BERT pre-trained model encoding. The input characters and word sequences T are pre-encoded using a BERT language model pre-trained on a large-scale corpus. Compared to traditional static word vectors, this method can obtain richer dynamic semantic expressions. .

[0101] 𝐶 = 𝐵𝐸𝑅𝑇 (𝑇 )

[0102] S103. Aggregate the semantic vectors of the trigger words and the semantic vectors of each argument to obtain their respective semantic vectors. Then, fuse the trigger word vectors with the candidate argument vectors through a fully connected network to obtain the fused vector representation of each candidate argument. .

[0103]

[0104]

[0105]

[0106] Furthermore, in step S2, each candidate argument A is feature-encoded using a feature encoder related to event type E, resulting in its feature vector U = {𝑢0, 𝑢1, ..., 𝑢𝐾}.

[0107]

[0108] S3 also contains 3 sub-steps, the specific structure of which is as follows: Figure 2 As shown.

[0109] S301. Learning feature vectors U in local and global dimensions through two different fully connected networks. local and U global .

[0110]

[0111]

[0112] S302. Decode each argument A in the local dimension by using softmax in both local and global dimensions. k In different roles R m The probability distribution P under the following conditions A And R for each role in the global dimension m In different arguments A k The probability distribution P under the following conditions R .

[0113]

[0114]

[0115] in Let represent the probability that the ith argument belongs to the ith role in event . This probability characterizes the local role classification probability estimate made by the model based on information from candidate arguments. Let represent the probability that the nth role is represented by the nth argument. This probability characterizes the model's probability estimate of which argument should represent a certain role from a global perspective. Combining the above two probabilities, this invention calculates the following cross-entropy loss function based on the role's label :

[0116]

[0117]

[0118]

[0119] in, This represents the true probability label of the k-th argument belonging to the m-th role. If role overlap exists, the probability is equally divided among multiple roles. Similarly, This indicates the true probability label of the m-th role being borne by the k-th argument, where the probability is equally distributed when the same role is borne by multiple arguments.

[0120] Finally, for step S4, to enable better global information aggregation of the label probability distribution information of each argument and to allow each argument to perceive the label information of all other arguments, a two-layer network structure similar to a multi-relation graph neural network (RGCN) is used to perform message passing on the corresponding event star graph. Its structure is as follows: Figure 3The message passing in each layer of the graph neural network enables each node to obtain the feature and role probability distribution information of its neighbors. In the star-shaped event graph, the event trigger word is connected with all arguments, and the arguments are not connected with each other, so the arguments are second-order neighbors of each other. Therefore, through the information transmission of two layers of graph neural networks, each argument includes all node information including the trigger word. Specifically, for each event type E, a corresponding independent model parameter is set to learn the event type-specific interaction mode. In the graph message passing stage, the present application represents the trigger word as and the argument as The edge weight matrix is initialized as the role label information aggregation module and . represents the cross-node edge, represents the self-loop edge of each node. Meanwhile, the present application sets and the tensor as the linear transformation matrix corresponding to the trigger word type and the linear transformation matrix set corresponding to all argument roles under this type. The specific graph neural network information interaction calculation process is as follows: first, the present application weights the tensor by the global and local two-dimensional label probability to obtain the edge weight matrices of the self-loop edge and the non-self-loop edge and .

[0121]

[0122]

[0123] After that, the present application can map the argument and trigger word representation into a new space that integrates label probability information, and can perform the first layer of graph network information interaction to obtain the interactive trigger word and argument node representation and . The calculation is as follows:

[0124]

[0125]

[0126] After that, the present application uses the new node representation to perform the same operation again to complete the second-order graph information interaction:

[0127]

[0128]

[0129] Then the present application fuses the The role label information prediction module is input again, and specifically The fusion vector obtained in step S103 represents Step S2 is performed again, The information interaction position of the first layer is returned to for circulation. This iteration is performed until the argument role classification output by the role label information prediction module tends to be stable, and a higher-consistency argument role label result is finally obtained. The result is obtained by the label probability matrix And Together, for the i-th argument, if The role output by the present application is , Is a hyperparameter. Figure 4 Is the overall method structure diagram of the present application.

[0130] The following is a system embodiment corresponding to the above-mentioned method embodiment. The present embodiment can be implemented in cooperation with the above-mentioned embodiments. The related technical details mentioned in the above-mentioned embodiments are still valid in the present embodiment. In order to reduce repetition, they will not be described here. Correspondingly, the related technical details mentioned in the present embodiment can also be applied in the above-mentioned embodiments.

[0131] The present application also proposes an event argument role classification system based on role consistency, which comprises:

[0132] The training corpus preprocessing module obtains the training corpus of the annotated event argument role category and event trigger word, performs word segmentation on the text of the training corpus, and obtains the ID of each word in the pre-training dictionary of the language representation model BERT according to the pre-training dictionary of the language representation model BERT; input the word sequence composed of all word IDs into the multi-layer translation model Transformer of BERT to pre-encode the word sequence and obtain the word representation vector;

[0133] The candidate argument semantic encoding module aggregates the event trigger word and the word representation vector to obtain a context semantic expression vector of the trigger word, splices the context semantic expression vector and the word representation vector and fuses them through a linear network to obtain a semantic expression vector of each candidate argument;

[0134] The candidate argument feature encoding module encodes the semantic expression vector of the candidate argument according to the corresponding feature modeling network selected according to the event type to obtain a feature expression vector of each candidate argument;

[0135] The candidate argument role decoding module decodes the feature expression vector of each candidate argument from two dimensions of global and local to obtain a role probability distribution matrix in two dimensions;

[0136] The role consistency information interaction module performs multi-stage graph neural network interaction on the trigger-argument connection graph, and fuses the role probability distribution matrix in the two dimensions into information interaction based on the trigger-argument connection graph to obtain a global expression vector of each candidate argument fused with role probability information as a new semantic expression vector of the candidate argument.

[0137] The iterative role classification module executes the candidate argument feature encoding module, the candidate argument role decoding module and the role consistency information interaction module again until the role probability distribution matrix obtained by decoding converges, and the role category with the maximum probability is taken as the output of the final argument role prediction.

[0138] The event argument role classification system based on role consistency, wherein the training corpus preprocessing module comprises:

[0139] The text is segmented, and the input of the same batch of words is padded to the same length according to the longest text length of the batch of words;

[0140] The input word, the word sequence T is pre-encoded by the BERT language model to obtain a word representation vector with dynamic semantic expression ;

[0141] C = BERT (T)

[0142] The candidate argument semantic encoding module comprises:

[0143] The event trigger word and the word representation vector are aggregated to obtain respective context semantic expression vectors, the context semantic expression vectors and the word representation vectors are spliced and fused through a linear network to obtain a semantic expression vector of each candidate argument ;

[0144] The candidate argument feature encoding module comprises: encoding the features of each candidate argument A through an event type E related feature encoder to obtain a feature expression vector U = {u0, u1,..., u 𝐾};

[0145]

[0146] The event argument role classification system based on role consistency, wherein the candidate argument role decoding module comprises:

[0147] The two fully connected networks FNN learn the feature vectors U local and U global in the local and global dimensions:

[0148]

[0149]

[0150] Decoding by local and global dimension softmax, get each argument A k under different roles R m Probability distribution matrix P A , and each role R m under different arguments A k Probability distribution matrix P R ;

[0151]

[0152]

[0153] Wherein The probability of the kth argument belonging to the mth role in the E event, The probability of the mth role being expressed by the kth argument; According to the label L of the role, the cross entropy loss function is made:

[0154]

[0155]

[0156]

[0157] Wherein, The true probability label of the kth argument belonging to the mth role, if there is role overlap, then it is divided by multiple roles; The true probability label of the mth role being assumed by the kth argument, and the probability is divided when the same role is assumed by multiple arguments;

[0158] The event argument role classification system based on role consistency, wherein the role consistency information interaction module comprises: using graph neural network RGCN to perform message passing on the trigger-argument connection graph, and each node obtains the feature and role probability distribution information of its neighbor in each layer of the graph neural network; In the trigger-argument connection graph, the event trigger is connected with all arguments, and all node information including the trigger is included in each argument through two layers of graph neural network information transmission; For each event type E, corresponding independent model parameters are set, and in the graph message passing stage, the trigger representation And the semantic expression vector of argument The edge weight matrix Is initialized by the label probability matrix of the role label information aggregation module And obtained, representing the edges across nodes, representing the self-loop edges inside each node; provided with and tensor as the linear transformation matrix corresponding to the trigger word type and the linear transformation matrix set corresponding to all argument roles under this type; through the label probability of two dimensions of global and local, the tensor weighted to obtain the edge weight matrix of the self-loop edge and the non-self-loop edge and :

[0159]

[0160]

[0161] The argument and trigger word representation is mapped into a new space fused with label probability information, and first-order graph network information interaction is carried out to obtain the interactive trigger word and argument node representation and ;

[0162]

[0163]

[0164] The same operation is carried out again using the new node representation to complete the second-order graph information interaction:

[0165]

[0166]

[0167] The iterative role classification module includes, taking as a new fusion vector representation the candidate argument feature encoding module is executed again, then the cycle is carried out again at the first-order graph network information interaction position until the role probability distribution matrix obtained by decoding converges.

[0168] The application further provides a storage medium for storing a program for executing the event argument role classification method with any one of the role consistency.

[0169] The application further provides a client for an event argument role classification system with role consistency.

Claims

1. A role-consistency-based event argument role classification method, characterized in that, Comprise: The training corpus preprocessing step, obtaining the training corpus of annotated event argument role category and event trigger word, performing word segmentation on the text of the training corpus, and obtaining the ID of each word in the pre-training dictionary of the language representation model BERT according to the pre-training dictionary; input the word sequence composed of all word IDs into the multi-layer translation model Transformer of BERT to pre-encode the word sequence and obtain the word representation vector; The candidate argument semantic encoding step, aggregating the event trigger word and the word representation vector to obtain the context semantic expression vector of the trigger word, splicing the context semantic expression vector and the word representation vector and fusing them through a linear network to obtain the semantic expression vector of each candidate argument; The candidate argument feature encoding step, according to the event type selection corresponding feature modeling network encodes the semantic expression vector of the candidate argument to obtain the feature expression vector of each candidate argument; The candidate argument role decoding step, decoding the feature expression vector of each candidate argument from the global and local two dimensions to obtain the role probability distribution matrix in the two dimensions; The role consistency information interaction step, performing multi-order graph neural network interaction on the trigger-argument connection graph, and fusing the role probability distribution matrix in the two dimensions into the information interaction based on the trigger-argument connection graph to obtain the global expression vector of each candidate argument fused with the role probability information as the new semantic expression vector of the candidate argument; The iterative role classification step, again performing the candidate argument feature encoding step, the candidate argument role decoding step and the role consistency information interaction step until the role probability distribution matrix obtained by decoding converges, and the role category with the largest probability is taken as the final argument role prediction output; The role consistency information interaction step includes: using a graph neural network RGCN to perform message passing on the trigger-word-argument connection graph, and each layer of the graph neural network performs message passing to enable each node to obtain feature and role probability distribution information of neighbors; the event trigger word is connected with all arguments in the trigger-word-argument connection graph, information is passed through two layers of the graph neural network, and each argument includes all node information including the trigger word; a corresponding independent model parameter is set for each event type E, and in the graph message passing stage, the trigger word representation and the semantic expression vector of the argument is initialized as a graph node, and an edge weight matrix is initialized by the label probability matrix of the role label information aggregation module and is obtained, represents a cross-node edge, represents an internal self-loop edge of each node; the edge weight matrix and the tensor are set as a linear transformation matrix corresponding to the trigger word type and a linear transformation matrix set corresponding to all argument roles under the type; the edge weight matrix of the self-loop edge and the non-self-loop edge is obtained by weighting the tensor and : 。 2. The event argument role classification method based on role consistency of claim 1, wherein, The training corpus preprocessing step comprises: Performing word segmentation on the text, and padding the input of the same batch of words to the same length according to the longest text length of the batch of words; The input word, the word sequence T is pre-encoded by the BERT language model to obtain a word representation vector with dynamic semantic expression ; C = BERT (T) The candidate argument semantic encoding step comprises: The event trigger word and the word representation vector are aggregated to obtain respective context semantic expression vectors, the context semantic expression vectors and the word representation vectors are spliced and fused through a linear network to obtain a semantic expression vector of each candidate argument ; The candidate argument feature encoding step includes: feature encoding of each candidate argument A by a feature encoder related to the event type E to obtain its feature expression vector U = {u0, u1,..., u 𝐾} 。 3. The event argument role classification method based on role consistency of claim 2, wherein, The candidate argument role decoding step comprises: The feature vectors U under local and global dimensions are learned by two fully connected networks FNN local and U global : Decoding by local and global dimension softmax, get each argument A k Under the probability distribution matrix P m of different roles R A , and the probability distribution matrix P m of each role R k under different arguments A R ; wherein represents the probability that the kth argument belongs to the mth role in the E event, represents the probability that the mth role is expressed by the kth argument; the cross-entropy loss function is made according to the label L of the role: wherein, represents the true probability label of the kth argument belonging to the mth role, and if there is a role overlap phenomenon, it is equally divided by multiple roles; represents the true probability label of the mth role being assumed by the kth argument, and when the same role is assumed by multiple arguments, the probability is equally divided.

4. The event argument role classification method based on role consistency of claim 3, wherein, The role consistency information interaction step comprises: The argument unit and trigger word representation are mapped into a new space fused with label probability information, and the first layer graph network information interaction is performed to obtain the interactive trigger word and argument unit node representation and ; Using the new node representation to perform the same operation again to complete the two-order graph information interaction: The iterative role classification step includes as a new fusion vector representation The candidate argument feature encoding step is performed again, Then, it is cycled again at the position of the first layer of the graph network information interaction until the role probability distribution matrix obtained by decoding converges.

5. An event argument role classification system based on role consistency, characterized in that, Comprise: The training corpus preprocessing module, obtaining the training corpus of annotated event argument role category and event trigger word, performing word segmentation on the text of the training corpus, and obtaining the ID of each word in the pre-training dictionary of the language representation model BERT according to the pre-training dictionary; input the word sequence composed of all word IDs into the multi-layer translation model Transformer of BERT to pre-encode the word sequence and obtain the word representation vector; The candidate argument semantic encoding module, aggregating the event trigger word and the word representation vector to obtain the context semantic expression vector of the trigger word, splicing the context semantic expression vector and the word representation vector and fusing them through a linear network to obtain the semantic expression vector of each candidate argument; The candidate argument feature encoding module, according to the event type selection corresponding feature modeling network encodes the semantic expression vector of the candidate argument to obtain the feature expression vector of each candidate argument; The candidate argument role decoding module decodes the feature expression vector of each candidate argument from two dimensions of global and local to obtain role probability distribution matrices in the two dimensions; The role consistency information interaction module performs multi-stage graph neural network interaction on the trigger-argument connection graph, and integrates the role probability distribution matrices in the two dimensions into the information interaction based on the trigger-argument connection graph to obtain a global expression vector of each candidate argument fused with role probability information as a new semantic expression vector of the candidate argument; The iterative role classification module executes the candidate argument feature encoding module, the candidate argument role decoding module and the role consistency information interaction module again until the role probability distribution matrix obtained by decoding converges, and the role category with the maximum probability is taken as the output of the final argument role prediction. The role consistency information interaction module includes: using a graph neural network RGCN to perform message passing on the trigger-word-argument connection graph, and each node obtains the feature and role probability distribution information of its neighbor in each layer of the graph neural network; the event trigger is connected with all arguments in the trigger-word-argument connection graph, and the information is passed through two layers of the graph neural network, and each argument includes all node information including the trigger; a corresponding independent model parameter is set for each event type E, and in the graph message passing stage, the trigger representation and the semantic expression vector of the argument is obtained. The edge weight matrix is initialized as the graph node representation and the label probability matrix of the role label information aggregation module . The cross-node connection edge is represented as The internal self-loop edge of each node is represented as and the tensor is set as the linear transformation matrix corresponding to the trigger type and the linear transformation matrix set corresponding to all argument roles in the type; the self-loop edge and the non-self-loop edge are weighted by the label probability in the global and local two dimensions to obtain the edge weight matrix and :​ 。 6. The event argument role classification system based on role consistency of claim 5, wherein, The training corpus preprocessing module comprises: The text is segmented, and the input of the same batch of words is padded to the same length according to the longest text length of the batch of words; The input word, the word sequence T is pre-encoded by the BERT language model to obtain a word representation vector with dynamic semantic expression ; C = BERT (T) The candidate argument semantic encoding module comprises: The event trigger word and the word representation vector are aggregated to obtain respective context semantic expression vectors, the context semantic expression vectors and the word representation vectors are spliced and fused through a linear network to obtain a semantic expression vector of each candidate argument ; The candidate argument feature encoding module comprises: performing feature encoding on each candidate argument A by a feature encoder related to an event type E to obtain a feature expression vector U = {u0, u1,..., u 𝐾}. 。 7. The event argument role classification system based on role consistency of claim 6, wherein, The candidate argument role decoding module comprises: The feature vectors U under local and global dimensions are learned by two fully connected networks FNN local and U global : Decoding by local and global dimension softmax, get each argument A k under different roles R m Probability distribution matrix P A , and global dimension each role R m under different arguments A k Probability distribution matrix P R ; wherein represents the probability that the kth argument belongs to the mth role in the E event, represents the probability that the mth role is expressed by the kth argument; the cross-entropy loss function is made according to the label L of the role: wherein, represents the true probability label of the kth argument belonging to the mth role, and if there is a role overlap phenomenon, it is equally divided by multiple roles; represents the true probability label of the mth role being assumed by the kth argument, and when the same role is assumed by multiple arguments, the probability is equally divided.

8. The event argument role classification system based on role consistency of claim 7, wherein, The role consistency information interaction module comprises: using a graph neural network RGCN to perform message passing on the trigger-word-argument connection graph, and message passing of each layer of the graph neural network enables each node to obtain feature and role probability distribution information of neighbors of the node; in the trigger-word-argument connection graph, an event trigger word is connected with all arguments, information is passed through two layers of the graph neural network, and each argument includes all node information of the trigger word; a corresponding independent model parameter is set for each event type E, in the graph message passing stage, the trigger word representation and the semantic expression vector of the argument are obtained. An edge weight matrix is obtained by the role label information aggregation module. and represent cross-node edges. represent internal self-loop edges of each node; the edge weight matrix and the tensor are set as a linear transformation matrix corresponding to the trigger word type and a linear transformation matrix set corresponding to all argument roles under the type; the edge weight matrix is obtained by weighting the tensor and :​ The argument unit and trigger word representation are mapped into a new space fused with label probability information, and the first layer graph network information interaction is performed to obtain the interactive trigger word and argument unit node representation and ; The same operation is performed again using the new node representation to complete the second-stage graph information interaction: The iterative role classification module includes, as a new fusion vector representation The candidate argument feature encoding module is executed again, Then, the cycle is performed again at the information interaction position of the first layer of the graph network until the role probability distribution matrix obtained by decoding converges.

9. A storage medium for storing a program for executing the role consistency event argument role classification method according to any one of claims 1 to 4.

10. A client for the role consistency event argument role classification system according to any one of claims 5 to 8.

Images