Event argument extraction method for military field

Abstract

The invention discloses a military field-oriented event argument extraction method, and relates to the technical field of event argument extraction, and the method comprises the following specific steps: 1, obtaining data; step 2, defining an annotation rule; step 3, introducing a multi-label entity classifier: labeling the entity as a single-label or multi-label type; 4, introducing a multi-label entity guiding attention mechanism; step 5, sequence labeling; step 6, joint training; step 7, coding is carried out by using a BERT model; 8, capturing a context and encoding the document; and step 9, constructing a global AMR graph and a local AMR graph: stimulating interaction between concepts in the document through an AMR guide module, and fusing double-flow representation by using an information fusion module. According to the event argument extraction method, a multi-label entity classifier and a multi-task learning mechanism are combined, and higher accuracy and recall rate are provided by improving the recognition capability of the event argument, especially in the application of the military field.

Description

Technical Field

[0001] The present invention relates to the technical field of event argument extraction, and in particular to an event argument extraction method for the military field. Background Art

[0002] In the field of natural language processing, event argument extraction is a core task, which aims to identify and obtain the participating element information related to a specific event from the text content. In many application scenarios, such as news text analysis and information extraction, accurate extraction of event arguments is crucial to understanding the event information in the text.

[0003] There are three main methods for event argument extraction at present. The first is rule-based event argument extraction method, the second is statistics-based event argument extraction method, and the third is a hybrid event argument extraction method that combines the two. Rule-based event argument extraction methods often require a large amount of manually pre-labeled data and cannot cope well with language diversity. Statistics-based event argument extraction methods, such as hidden Markov models (HMMs) and conditional random fields (CRFs), although they can better capture complex language characteristics, still face challenges in terms of accuracy and recall.

[0004] In recent years, deep learning-based models, such as BERT, BiLSTM-CRF, and Transformer models, have made significant progress. However, these models still face some challenges when dealing with domain-specific event extraction tasks. For example, although BERT can capture contextual information, it may still not be able to fully solve the problem of event argument extraction in specific fields, such as the military field. At the same time, the application of multi-label entity recognition and multi-task learning has also shown its potential in improving model performance, but how to effectively combine these methods to meet the needs of specific fields is still a problem to be solved. To this end, we propose an event argument extraction method for the military field. Summary of the invention

[0005] The purpose of the present invention is to provide an event argument extraction method for the military field, which combines a multi-label entity classifier and a multi-task learning mechanism to improve the recognition ability of event arguments, especially in applications in the military field, and provides higher accuracy and recall rate.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for extracting event arguments in the military field, comprising the following specific steps:

[0007] Step 1: Obtain data: obtain a certain amount of data and process the acquired data;

[0008] Step 2: Define annotation rules: To standardize the annotation of controversial corpus, ensure the annotation quality and reduce the difficulty of annotation, and annotate events that are consistent with the information analysis theme and have clear semantics;

[0009] Step 3: Introduce a multi-label entity classifier: label entities as single-label or multi-label types;

[0010] Step 4: Introduce a multi-label entity-guided attention mechanism: Input sentences with multi-label entities into the document-level encoder, use the multi-label entity-guided attention mechanism to distinguish the importance of entities, and combine sentence-level contextual information and document-level features;

[0011] Step 5: Sequence labeling: Use the fusion mechanism to input the obtained fusion features into the conditional random field (CRF) layer for sequence labeling to extract event arguments;

[0012] Step 6: Joint training: Event argument extraction is the main task, and named entity boundary detection is used as an auxiliary task for joint training;

[0013] Step 7: Use the BERT model for encoding: Improve the performance of event argument extraction by combining the parameter sharing mechanism of multi-task learning and the named entity boundary information;

[0014] Step 8: Capture context and encode the document: Use global and local encoders to capture contexts of different scopes and encode the document;

[0015] Step 9: Construct global AMR graph and local AMR graph: stimulate the interaction between concepts in the document through the AMR guidance module, and use the information fusion module to fuse the two-stream representation.

[0016] Preferably, when defining the labeling rules in step 2, data is preprocessed by word form restoration and a trusted trigger word set.

[0017] Preferably, in step 4, a multiple encoding mechanism is specifically used to dynamically fuse sentence-level features and document-level features of the text.

[0018] Preferably, in step 4, an improved first-layer graph attention network model is used to process the instance embedding initialization matrix and the relationship embedding initialization matrix to obtain a first-layer instance embedding optimization matrix.

[0019] Preferably, in step 1, the specific method for obtaining data is: crawling data from public corpus or specific websites, integrating the data, and then cleaning the data by deleting invalid or missing data.

[0020] Preferably, in step nine, the two stream representations are fused through an information fusion module, and boundary information is enhanced through boundary loss.

[0021] Preferably, the calculation formula of the multi-label entity-guided attention mechanism is:

[0022] Attention(t,e)=Softmax(LeakyReLU(Watt·[t;e]))

[0023] Among them, Watt is the linear transformation matrix, and [t; e] represents the concatenation operation of the trigger word and the entity.

[0024] Preferably, the loss function of the multi-task learning is:

[0025] Loss=Lossmain+λ·Lossaux

[0026] Among them, Lossmain is the main task loss of event argument extraction, and Lossaux is the auxiliary task loss of named entity boundary detection.

[0027] Preferably, the calculation formula for fusion of global and local context information of the dual-stream encoding is:

[0028] FusedRepresentation=Concat(GlobalEncoder(X),LocalEncoder(X))

[0029] Where Concat means concatenating the output results of the global encoder and the local encoder, GlobalEncoder(X) and LocalEncoder(X) represent the results of global and local encoding of the document, respectively.

[0030] Preferably, the construction formula of the AMR graph is:

[0031] AMRGraph=ConstructAMR(X,Interactions)

[0032] The ConstructAMR function is used to construct global and local AMR graphs based on the concept interactions in the document, and Interactions represents the interaction information between concepts.

[0033] Technical effects and advantages of the present invention:

[0034] (1) The entities in the document are classified into multi-label entities and single-label entities through a multi-label entity classifier to solve the label confusion problem caused by inconsistent expressions of the same military entity in different sentences. Secondly, a model based on multi-task learning is adopted, and the named entity boundary detection task is used as an auxiliary task. The role information is integrated into the main task of event argument extraction. The auxiliary task guides and optimizes the main task by learning the entity boundary information. The accurate identification of entity boundaries can directly promote the correct allocation of argument roles, thereby improving the accuracy of event argument extraction. Finally, a model based on AMR and dual-stream encoding is used to capture semantic features, and the long-range dependency between arguments and triggers is reflected through the AMR graph. Combined with the dual-stream encoding of the global encoder and the local encoder, the context information is fully utilized to enhance the extraction effect of event arguments. This method aims to improve the accuracy and robustness of military event argument extraction, especially in complex text environments.

[0035] (2) Through the multi-label entity-based paragraph-level argument extraction method, it is convenient to introduce a multi-label entity classifier into the model, and mark the entities in the sentence as single-label and multi-label types. Secondly, the sentence with multi-label entities is input into the document-level encoder, and the multi-label entity-guided attention mechanism is used to distinguish the importance of different entities in the document. Finally, a fusion mechanism is used to dynamically fuse sentence-level context information and document-level features, and the obtained fusion features are input into the CRF layer for sequence labeling to extract event arguments;

[0036] (3) Through the chapter-level ship event argument extraction model based on the multi-label entity-guided attention mechanism, it is convenient to jointly train the event argument extraction task and the named entity boundary detection task. By adapting to different tasks to obtain shared representations, the entity boundary information is used to strengthen the understanding of event arguments, guide the generation of shared semantic information encoding, and improve the accuracy of the model's event argument extraction;

[0037] (4) Through the event argument extraction process based on dual-stream encoding and AMR, it is convenient to use global and local encoders with different attention reception fields to capture the context in different ranges of the document to be extracted and encode the document. The global AMR graph and the local AMR graph are constructed through the AMR guidance module to stimulate the interaction between concepts in the document, especially those that are far apart. The two stream representations are fused using the information fusion module, and the boundary information is enhanced through the boundary loss. Finally, the classification module is used to predict the candidate span. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flowchart of the event argument extraction method of the present invention.

[0039] Figure 2This is the architecture diagram of the chapter-level ship event argument extraction model based on the multi-label entity-guided attention mechanism of the present invention.

[0040] Figure 3 This is the architecture diagram of the chapter-level ship event argument extraction model based on multi-task assistance of the present invention.

[0041] Figure 4 This is a flow chart of event argument extraction based on dual-stream coding and AMR of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The present invention provides Figure 1-4The method for extracting event arguments for the military field shown in the figure includes the following specific steps: Step 1, data acquisition: obtain a certain amount of data and process the acquired data; Step 2, define annotation rules: standardize the annotation of controversial corpus, ensure the annotation quality and reduce the annotation difficulty, and annotate events that meet the information analysis theme and have clear semantics; Step 3, introduce multi-label entity classifier: annotate entities as single-label or multi-label types; Step 4, introduce multi-label entity-guided attention mechanism: input sentences with multi-label entities into the document-level encoder, use the multi-label entity-guided attention mechanism to distinguish the importance of entities, and combine sentence-level context information and document-level features; Step 5, sequence annotation: use the fusion mechanism to input the obtained fusion features into the conditional random field (CRF) layer for sequence annotation to extract event arguments; Step 6, joint training: take event argument extraction as the main task, and introduce named entity boundary detection As an auxiliary task for joint training; Step 7, use the BERT model for encoding: through the parameter sharing mechanism of multi-task learning, combined with the named entity boundary information to improve the performance of event argument extraction; Step 8, capture context and encode the document: use global and local encoders to capture contexts of different ranges and encode the document; Step 9, construct global AMR graph and local AMR graph: stimulate the interaction between concepts in the document through the AMR guidance module, and use the information fusion module to fuse the dual-stream representation; The specific process of the whole method is: the first step, obtain data, the second step, define the annotation rules, the third step, introduce the multi-label entity classifier, the fourth step, introduce the multi-label entity guided attention mechanism, the fifth step, sequence annotation, the sixth step, joint training, the seventh step, use the BERT model for encoding, the eighth step, capture the context and encode the document, the ninth step, construct the global AMR graph and the local AMR graph.

[0044] Furthermore, when defining the annotation rules in step 2, the data is preprocessed through word form restoration and a trusted trigger word set; through word form restoration and a trigger word set, the data can be more accurately grasped when defining the annotation rules, thereby ensuring the annotation quality and reducing the difficulty of annotation.

[0045] Furthermore, in step four, a multiple encoding mechanism is specifically used to dynamically fuse the sentence-level features and document-level features of the text; by utilizing a multiple encoding mechanism to dynamically fuse the sentence-level features and document-level features of the text, the efficiency of information extraction can be improved; multi-channel retrieval: first, multiple encoding methods provide a multi-channel retrieval approach. When information needs to be extracted, retrieval can be performed through any one or more encoding clues. Second, it can flexibly respond to interference. When one retrieval clue is interfered with or forgotten, other encoding clues may still help extract information.

[0046] Furthermore, in step 4, the improved first-layer graph attention network model is used to process the instance embedding initialization matrix and the relationship embedding initialization matrix to obtain the first-layer instance embedding optimization matrix.

[0047] Furthermore, in step one, the specific method for obtaining data is: crawling data from public corpus or specific websites, and integrating the data, and then cleaning the data by deleting invalid or missing data: for example, by extracting data from non-public military documents, and then integrating the data, and then deleting some incomplete or missing and invalid data, so as to achieve the effect of cleaning the data, thereby ensuring the accuracy of subsequent data extraction and avoiding affecting the accuracy of the implementation results of the present invention.

[0048] Furthermore, in step nine, the two stream representations are fused through the information fusion module, and the boundary information is enhanced through the boundary loss; by enhancing the boundary information through the boundary loss, the data quality can be improved: first, accurately locate the boundary: the boundary loss can guide the model to more accurately determine the boundary position of the target object, reduce the ambiguity and error of the boundary, thereby improving the accuracy of data annotation and improving the overall data quality; second, enhance data consistency: by emphasizing the boundary information, the characteristics of the data at the boundary are clearer and more consistent, reducing the contradictions and ambiguities within the data, improving the reliability and stability of the data, and facilitating subsequent analysis and processing. The model performance can also be optimized: first, improve the feature learning ability: the increase of boundary information helps the model to better learn the local features of the data, especially the subtle features at the boundary, so that the model can capture richer feature information, thereby improving the model's ability to represent the data and the feature learning effect; second, improve the model convergence speed: clear boundary information can provide more effective guidance for model training, help the model converge to the optimal solution faster, reduce training time and consumption of computing resources, and improve the efficiency of model training. It can also improve the classification and recognition effects: First, enhance the category distinction: In the classification task, boundary information can highlight the differences between different categories, especially in the boundary area, making it easier for the model to distinguish similar categories and improve the accuracy and recall of classification. Second, improve the recognition ability of small targets: For small targets or targets with unclear boundaries, adding boundary information can make the model pay more attention to the boundary characteristics of the target, improve the detection and recognition ability of small targets, and reduce missed detection and false detection. At the same time, it can enhance data security: First, detect anomalies and attacks: In some security-sensitive data processing scenarios, clear boundary information helps to timely discover anomalies and potential attacks in the data, and improve the security and risk prevention capabilities of the data through monitoring and analysis of the boundaries. Second, protect privacy information: When processing data involving privacy, the scope and boundaries of privacy data can be better defined through boundary loss, and more effective encryption and protection measures can be taken to prevent the leakage of privacy information and enhance the privacy protection capabilities of data. It can also assist in data visualization and understanding: First, clearly display the data structure: Adding boundary information can make the data more clearly display its internal structure and boundary characteristics when it is visualized, and more intuitively understand the distribution and characteristics of the data, providing more powerful support for data analysis and decision-making.

[0049] It should be further explained that the calculation formula of the multi-label entity-guided attention mechanism is:

[0050] Attention(t,e)=Softmax(LeakyReLU(Watt·[t;e]))

[0051] Among them, Watt is the linear transformation matrix, [t; e] represents the concatenation operation of the trigger word and the entity; by introducing the multi-label entity-guided attention mechanism, the effectiveness of feature extraction can be improved: First, focus on key information: it can guide the model to automatically focus on the key areas or features in the data, ignore irrelevant interference information, and more accurately extract effective features related to the task; second, enhance feature expression: by highlighting key features, the model can learn more discriminative feature representations, enrich the semantic information of the features, and improve the ability of the features to characterize the essence of the data. It can also improve model performance and efficiency: First, improve model accuracy: effective use of the attention mechanism can enable the model to better understand the data, thereby making more accurate predictions in tasks such as classification and regression, and improving the model's performance indicators such as accuracy and recall rate; third, accelerate model convergence: because the attention mechanism can help the model quickly locate key Key information reduces the search space and interference factors of the model during the learning process, which can usually speed up the training convergence of the model, save training time and computing resources, and improve the adaptability and robustness of the model: First, adapt to different data distributions: When facing data with different distributions, the guided attention mechanism can dynamically adjust the allocation of attention according to the characteristics of the data, so that the model can better adapt to data changes and improve the generalization ability of the model. Second, enhance anti-interference ability, promote multimodal data fusion, and align different modal information: When processing multimodal data, the attention mechanism can be used to align information between different modalities, find the correlation and correspondence between them, and achieve more effective multimodal data fusion, and can improve the fusion effect: By focusing on and integrating the key parts of different modal data, it can give full play to the advantages of each modality and improve the performance and effect of the multimodal fusion model.

[0052] Furthermore, the loss function of multi-task learning is:

[0053] Loss=Lossmain+λ·Lossaux

[0054] Among them, Lossmain is the main task loss of event argument extraction, and Lossaux is the auxiliary task loss of named entity boundary detection.

[0055] Furthermore, the calculation formula for the fusion of global and local context information of dual-stream encoding is:

[0056] FusedRepresentation=Concat(GlobalEncoder(X),LocalEncoder(X))

[0057] Where Concat means concatenating the output results of the global encoder and the local encoder, GlobalEncoder(X) and LocalEncoder(X) represent the results of global and local encoding of the document, respectively.

[0058] Furthermore, the construction formula of the AMR graph is:

[0059] AMRGraph=ConstructAMR(X,Interactions)

[0060] In the formula, the ConstructAMR function is used to construct global and local AMR graphs according to the concept interactions in the document, and Interactions represents the interactive information between concepts. AMR graphs can facilitate knowledge extraction and integration: First, knowledge extraction is convenient: AMR graphs provide a standardized semantic representation form. By constructing AMR graphs for a large amount of text, various knowledge can be easily extracted from them, such as entities, events, attributes, etc. Second, knowledge integration is efficient: text data from different sources can be uniformly semantically represented by constructing AMR graphs, and then these AMR graphs are integrated to achieve knowledge fusion and sharing, providing a rich knowledge foundation for subsequent knowledge reasoning and application. It can also support a variety of natural language processing tasks: First, machine translation assistance: In machine translation, AMR graphs can be used as an intermediate representation to convert source language text into AMR graphs, and then generate target language text from AMR graphs according to the characteristics of the target language, which helps to improve the accuracy and fluency of translation. Second, information retrieval optimization: By converting text into AMR graphs, the user's query intention can be understood more accurately, and semantic matching and retrieval can be performed on the graphs to improve the relevance and accuracy of retrieval results. Third, language variant adaptation: Languages ​​in different regions or fields may have differences in vocabulary, sentence structure, etc., but the semantics they express are often similar. AMR graphs focus on the representation of the semantic level, and can ignore these superficial language differences to a certain extent, and have good adaptability to various language variants. Fourth, cross-language processing advantages: Because AMR graphs have a relatively unified semantic representation, they have great advantages in cross-language processing. Texts in different languages ​​can be converted into AMR graphs, and then various processing and analysis can be performed at the AMR graph level to achieve cross-language information exchange and knowledge sharing.

[0061] By adopting multi-label entity classifiers and multi-task learning mechanisms, the model's ability to recognize event arguments in the military field is improved. In the multi-label entity classifier, the event arguments can be captured more accurately by classifying multi-label entities and single-label entities; in multi-task learning, the auxiliary task of named entity boundary detection is combined to significantly improve the event argument extraction effect in the main task, and the multi-task learning framework is used to integrate the two tasks; then, by constructing an abstract semantic graph at the chapter level, comprehensive modeling of text semantics is achieved, which significantly reduces the complexity of model training.

[0062] Figure 2 A method for extracting paragraph-level arguments based on multi-label entities is provided. A multi-label entity classifier is introduced into the model to mark entities in sentences as single-label and multi-label types. Secondly, the sentences with multi-label entities are input into the document-level encoder, and the multi-label entity-guided attention mechanism is used to distinguish the importance of different entities in the document. Finally, a fusion mechanism is used to dynamically fuse sentence-level contextual information and document-level features. The obtained fused features are input into the CRF layer for sequence labeling to extract event arguments.

[0063] Figure 3 A chapter-level ship event argument extraction model based on a multi-label entity-guided attention mechanism is provided. Event argument extraction tasks and named entity boundary detection tasks are jointly trained. Shared representations are obtained by adapting to different tasks, and entity boundary information is used to enhance the understanding of event arguments. This guides the generation of shared semantic information encoding and improves the accuracy of event argument extraction of the model.

[0064] Figure 4 This paper proposes an event argument extraction process based on dual-stream encoding and AMR. It uses global and local encoders with different attention reception fields to capture the context of different ranges of the document to be extracted and encode the document. The global AMR graph and local AMR graph are constructed through the AMR guidance module to stimulate the interaction between concepts in the document, especially those that are far apart. The two stream representations are fused using the information fusion module, and the boundary information is enhanced through the boundary loss. Finally, the classification module is used to predict the candidate span.

[0065] Working principle of the present invention:

[0066] Public corpus and specific website data are collected and cleaned, annotation rules are defined to improve the annotation quality, and data are preprocessed through word form restoration and trusted trigger word sets; the improved first-layer graph attention network model is used to process the instance embedding initialization matrix and the relationship embedding initialization matrix to obtain the first-layer instance embedding optimization matrix; a multi-label entity classifier is used to divide entities into single-label and multi-label categories; sentences with multi-label entities are input into the document-level encoder, and the multi-label entity-guided attention mechanism is used to distinguish the importance of different entities; a multiple encoding mechanism is used to dynamically fuse the sentence-level features and document-level features of the text; sequence annotation is performed through the CRF layer, encoded using the BERT model, and entity boundaries are utilized through multi-task learning to improve the performance of event argument extraction; global and local encoders are used to capture semantic information at different positions in the text respectively; global and local AMR graphs are constructed, and the two stream representations are fused through the information fusion module.

[0067] The event argument extraction method of the present invention is described below in conjunction with specific embodiments:

[0068] The model in this paper is experimentally verified on the non-public Chinese military event dataset CEER, which contains 13,000 data, including 7,000 training samples, 1,500 validation sets, and 4,500 test sets. The dataset defines 28 event types in total, and the number of arguments corresponding to each event type ranges from 4 to 8.

[0069] In the experimental setting of the CIME model proposed in this paper, the dimensions of the hidden layer and the feed-forward layer are 768 and 1024 respectively. The model uses a three-layer graph convolutional network, the dropout rate is set to 0.1, and a batch size of 16 is used. The learning rate is set to 1e-4, and the model training cycle is set to 100 epochs, from which the optimal model is saved.

[0070] In the field of chapter-level event extraction, the DCFEE model proposed by Yang et al. believes that chapter-level event arguments are mainly concentrated in the central sentence of the article. The model first extracts event arguments from the central sentence of the article, and then checks the sentences around the central sentence to supplement more argument information. The DCFEE model has two versions, DCFEE-S and DCFEE-M. DCFEE-S only extracts a single event record at a time, while DCFEE-M can generate multiple potential event argument combinations at the same time. In addition, the Doc2EDAG model uses two Transformer encoders to obtain sentence-level and entity-level feature representations and fuse contextual information to support the simultaneous extraction of multiple events. Greedy-Dec is a variant of Doc2EDAG. It continues to extract other events after completing the extraction of one event. This paper uses the MicroF1 value to evaluate the performance of the model. The results are shown in Table 1.

[0071] Table 1 Comparison of Micro F1 value results of models

[0072]

[0073] It can be seen that the CIME model shows better performance than other integrated models. Its advantage is mainly attributed to the abstract processing of the relationship between arguments when constructing the paragraph-level abstract semantic graph, which enables the model to more effectively capture and understand the complex interactions between event arguments within the paragraph.

[0074] The ablation experiment results of the model in this paper are shown in Table 2. It can be seen that this abstract processing significantly enhances the model's ability to understand and parse chapter-level events.

[0075] Table 2 Model ablation experiment results

[0076]

[0077] First, the CIME model does not perform semantic boundary information detection, and only constructs an abstract semantic graph for event arguments. This weakens the model's ability to understand text semantics, which is reflected in the performance indicators. The MicroF1 value of the model drops by 1.

[0078] Then, this paper reforms the argument extraction method of the dataset and CIME model. The event classification in the dataset is standardized into a unified template and used as the input of the model. At the same time, a tree structure based on this unified template is used to extract event arguments. This method frees the model from dependence on specific event types, simplifies the extraction process, and solves the problem of having to extract according to event types in practical applications, thereby improving the model's Micro F1 value to 89.7.

[0079] Event extraction is a key step in converting unstructured text into structured data, which helps to filter out irrelevant information and extract valuable data. As the core part of this process, the extraction of event arguments directly affects the accuracy of event information extraction. Therefore, this paper proposes a new chapter-level event extraction method CIME, which combines event argument extraction and entity detection tasks. First, it uses a multi-task learning framework to integrate these two tasks; second, it achieves comprehensive modeling of text semantics by constructing a chapter-level abstract semantic graph; finally, it integrates text semantics and argument entity information to enhance the overall performance of the model. Experimental results show that this method achieves good results on the dataset and significantly reduces the complexity of model training. Future work will focus on further simplifying the application of abstract semantic graphs in CIME and exploring its application potential in resource-constrained environments.

[0080] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting event arguments in the military field, characterized in that: The specific steps include: Step 1: Obtain data: obtain a certain amount of data and process the acquired data; Step 2: Define annotation rules: To standardize the annotation of controversial corpus, ensure the annotation quality and reduce the difficulty of annotation, and annotate events that are consistent with the information analysis theme and have clear semantics; Step 3: Introduce a multi-label entity classifier: label entities as single-label or multi-label types; Step 4: Introduce a multi-label entity-guided attention mechanism: Input sentences with multi-label entities into the document-level encoder, use the multi-label entity-guided attention mechanism to distinguish the importance of entities, and combine sentence-level contextual information and document-level features; Step 5: Sequence labeling: Use the fusion mechanism to input the obtained fusion features into the conditional random field layer for sequence labeling to extract event arguments; Step 6: Joint training: Event argument extraction is the main task, and named entity boundary detection is used as an auxiliary task for joint training; Step 7: Use the BERT model for encoding: Improve the performance of event argument extraction by combining the parameter sharing mechanism of multi-task learning and the named entity boundary information; Step 8: Capture context and encode the document: Use global and local encoders to capture contexts of different scopes and encode the document; Step 9: Construct global AMR graph and local AMR graph: stimulate the interaction between concepts in the document through the AMR guidance module, and use the information fusion module to fuse the two-stream representation.

2. The event argument extraction method for military field according to claim 1, characterized in that: When defining the annotation rules in step 2, the data is also preprocessed through word form restoration and a trusted trigger word set.

3. The event argument extraction method for military field according to claim 1 is characterized in that: In the step 4, a multiple encoding mechanism is specifically used to dynamically fuse the sentence-level features and document-level features of the text.

4. The event argument extraction method for military field according to claim 1, characterized in that: In the step 4, an improved first-layer graph attention network model is used to process the instance embedding initialization matrix and the relationship embedding initialization matrix to obtain a first-layer instance embedding optimization matrix.

5. The event argument extraction method for military field according to claim 1, characterized in that: In the step 1, the specific method of obtaining data is: crawling data from public corpus or specific websites, integrating the data, and then cleaning the data by deleting invalid or missing data.

6. The event argument extraction method for military field according to claim 1, characterized in that: In the step nine, the two stream representations are fused through an information fusion module, and the boundary information is enhanced through boundary loss.

7. The event argument extraction method for military field according to claim 1 is characterized in that: The calculation formula of the multi-label entity-guided attention mechanism is: Attention(t,e)=Softmax(LeakyReLU(Watt·[t;e])) Among them, Watt is the linear transformation matrix, and [t; e] represents the concatenation operation of the trigger word and the entity.

8. The method for extracting event arguments for military fields according to claim 7, characterized in that: The loss function of the multi-task learning is: Loss=Lossmain+λ·Lossaux Among them, Lossmain is the main task loss of event argument extraction, and Lossaux is the auxiliary task loss of named entity boundary detection.

9. The method for extracting event arguments for military fields according to claim 8, characterized in that: The calculation formula for the fusion of global and local context information of the dual-stream encoding is: FusedRepresentation = Concat(GlobalEncoder(X), LocalEncoder(X)) where Concat means concatenating the output results of the global encoder and the local encoder, and GlobalEncoder(X) and LocalEncoder(X) respectively represent the results of global and local encoding of the document.

10. The event argument extraction method for military field according to claim 9, characterized in that: The construction formula of the AMR graph is: AMRGraph=ConstructAMR(X,Interactions) The ConstructAMR function is used to construct global and local AMR graphs based on the concept interactions in the document, and Interactions represents the interaction information between concepts.

Images