Method for extracting chapter-level fine-grained events in military news field

Abstract

The application discloses a chapter-level fine-grained event extraction method in the field of military news, relates to the related technical field of military information extraction, and comprises the following steps: S1: data collection and preprocessing, S2: event trigger word recognition, S3: construction and training of a PAIE event argument extraction model based on MRC, S4: fine-grained event argument extraction, S5: application of transfer learning, S6: comprehensive optimization processing of event arguments, S7: structured output and application, and output of the event arguments after the comprehensive optimization processing in a structured form. The application introduces a PAIE model, improves the efficiency and accuracy of chapter-level event argument extraction in the field of military news, introduces a bipartite graph matching and reasoning stage in the method, avoids the complexity of exhaustive threshold adjustment in a traditional method, and the series of innovative designs not only enhance the accuracy of the model, but also ensure that the model can still effectively perform event argument extraction in the case of data scarcity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of military information extraction, and particularly relates to a military news field chapter-level fine-grained event extraction method. BACKGROUND

[0002] With the rapid development of big data and natural language processing technology, information extraction technology has shown great application potential in multiple fields such as military intelligence analysis and public opinion monitoring. Information extraction automatically identifies and extracts entities, relationships and events in text, effectively converting unstructured text information into structured data, thereby providing support for downstream analysis and decision-making. However, most current event extraction research in the field of military news mainly focuses on the sentence level, ignoring the deep mining of event arguments in chapter-level text. These methods often only focus on simple event relationships within a sentence, making it difficult to handle complex event arguments and long-distance dependency relationships across sentences.

[0003] Event extraction, as a key technology for building knowledge graphs and intelligent analysis systems, aims to automatically identify events and their related argument information from text. In the field of military news, fine-grained event extraction not only helps to accurately understand and analyze military situations, but also provides strong support for subsequent intelligence reasoning and decision-making. Traditional event extraction methods rely heavily on manually annotated features or rules, which are not only time-consuming and labor-intensive, but also difficult to handle complex event structures in chapter-level text. In recent years, event extraction based on MRC has gradually become a research hotspot, especially in the field of fine-grained argument extraction. By combining event recognition with question answering, the accuracy and breadth of event extraction have been significantly improved.

[0004] However, current event extraction methods based on MRC still have some limitations. On the one hand, most methods only perform event recognition and argument extraction at the sentence level, ignoring the complex event relationships across sentences in chapter-level text, which limits the extraction effect. On the other hand, event information in military news often has high professional and secretive characteristics, and the context dependence and noise data between different texts also pose challenges to event extraction. Therefore, how to effectively utilize chapter-level context information to improve the accuracy and robustness of event argument extraction has become a key problem in the field of military news that needs to be solved. SUMMARY

[0005] The present application aims to provide a military news field chapter-level fine-grained event extraction method to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a military news field chapter-level fine-grained event extraction method, comprising the following steps:

[0007] S1: Data collection and preprocessing, a large number of military news articles are collected as sample data, and these data are cleaned and preprocessed to prepare for subsequent event extraction work;

[0008] S2: Identification of event trigger words, using an event detection model based on machine reading comprehension, the preprocessed military news articles are analyzed sentence by sentence, and the event trigger words are accurately identified;

[0009] S3: Construction and training of PAIE event argument extraction model based on MRC, a PAIE event argument extraction model based on MRC is constructed, and a large number of annotated military news articles are used for training, and the parameters and structure of the model are continuously adjusted;

[0010] S4: Extraction of fine-grained event arguments, the trained PAIE event argument extraction model based on MRC is applied to military news articles, and the fine-grained event arguments associated with each event trigger word are extracted, so as to mine the complex interaction between trigger words, event types and arguments, and capture the long-distance dependency relationship between event arguments in long text;

[0011] S5: Application of transfer learning, using the transfer learning ability of the pre-trained language model, the language model parameters trained on other large-scale text data are transferred to the event extraction task of military news articles;

[0012] S6: Comprehensive optimization processing of event arguments;

[0013] S7: Structured output and application, the event arguments after comprehensive optimization processing are output in a structured form.

[0014] Preferably, the preprocessing in step S1 includes the following steps:

[0015] S11: Use a Chinese word segmentation tool to perform word segmentation processing on the military news articles, and divide the articles into individual word units;

[0016] S12: Perform part-of-speech tagging, and tag each word with its part-of-speech information;

[0017] S13: Build a vocabulary, collect all the words in the article, remove duplicate words, and assign a unique identifier to each word.

[0018] Preferably, the event detection technology of the event detection model in step S2 includes:

[0019] S21: Event trigger word identification stage, the event trigger word identification stage identifies the event trigger word through the MRC question and answer template;

[0020] S22: an event type classification stage, which performs accurate classification of the event type based on the identified trigger words.

[0021] Preferably, the event argument extraction process in step S3 comprises:

[0022] S31: a manual template, in which all roles are manually connected using natural language;

[0023] S32: a soft prompt template, in which different argument roles are combined with pseudo-tokens specific to the roles that can be learned by a pre-trained language model;

[0024] S33: a connection-based template, in which all role names belonging to one event type are connected using natural language.

[0025] Preferably, the PAIE event argument extraction model in step S3 performs event argument extraction through model creation, prompt template, event argument selection decoder part, and event argument prediction.

[0026] Preferably, the event argument prediction combines bipartite graph matching and reasoning to solve the problem of optimal allocation of multiple similar role arguments in the same event:

[0027] By introducing the bipartite graph matching technique, multiple slots are inserted in the model template setting using the Hungarian algorithm to achieve global optimal allocation with minimum cost matching, and in the reasoning stage, the candidate event argument set is defined and scored;

[0028] To perform reasoning, the candidate event argument set is defined as:

[0029]

[0030] The candidate event argument set contains all spans shorter than the threshold L and special spans (0, 0) without extracted event arguments, and the model extracts the argument of each span selector θ k by enumerating all candidate spans and scoring them to extract the argument of each span selector θ k :

[0031]

[0032] And the predicted span of time slot k is given by:

[0033]

[0034] Preferably, the extraction of fine-grained event arguments in step S4 comprises the following steps:

[0035] S41: Chapter analysis, detailed analysis of military news chapter by sentence, clear semantics and structure of each sentence;

[0036] S42: Trigger word positioning, in the analyzed chapter, accurately positioning each event trigger word, accurately identifying these key words through a model based on machine reading comprehension;

[0037] S43: Argument association, for each located event trigger word, associate relevant fine-grained event arguments according to its semantics and context information;

[0038] S44: Interaction mining, further mine the complex interaction between trigger words, event types and arguments in the process of associating arguments;

[0039] S45: Long-distance dependency capture, through overall understanding of the chapter and grasping of the context, capture the long-distance dependency relationship between event arguments in long text.

[0040] Preferably, the application of transfer learning in step S5 includes the following steps:

[0041] S51: Select a pre-trained language model, select a suitable pre-trained language model according to the characteristics and needs of military news chapters;

[0042] S52: Model parameter transfer preparation, obtain the parameters of the pre-trained language model trained on other large-scale text data;

[0043] S53: Adapt to military news data, adapt the selected pre-trained language model to adapt to the specific field and language style of military news chapters;

[0044] S54: Parameter transfer and fine-tuning, transfer the parameters trained on other large-scale text data to the adapted military news chapter event extraction model, and then fine-tune the model using the labeled data of the military news chapter;

[0045] S55: Evaluation and optimization, evaluate the military news chapter event extraction model after transfer learning, evaluate the extraction effect of the model by comparing with the artificial labeled data, and according to the evaluation result, further optimize and adjust the model.

[0046] Preferably, the comprehensive optimization processing of event arguments in step S6 includes the following steps:

[0047] S61: De-duplication processing, carefully check all extracted event arguments, use specific algorithms and rules to find and remove duplicate event arguments;

[0048] S62: error correction, carefully review each event argument, and correct the error classification;

[0049] S63: semantic consistency check, check and correct the semantic consistency between event arguments;

[0050] S64: quality evaluation and screening, quality evaluation criteria are formulated, and the event arguments after deduplication and error correction are quality evaluated.

[0051] Preferably, the structured output and application in step S7 include the following steps:

[0052] S71: determine the structured format, and determine the structured output format to be used;

[0053] S72: field extraction and filling, extracting event type, trigger word and information of each argument from the event arguments after comprehensive optimization;

[0054] S73: organize the output order, determine the output order of the event arguments according to the specific application requirements and logic;

[0055] S74: interface with the application system, and interface the structured output event arguments with the specific application system.

[0056] Technical effects and advantages of the present application:

[0057] (1) The present application improves the efficiency and accuracy of event argument extraction in the field of chapter-level military news by introducing the PAIE model. The method applies the event argument extraction method based on MRC to chapter-level military news text. According to the characteristics that event arguments may be scattered in different sentences, a model creation prompt template part is designed. In this part, a set of prompts is created for each event type. These prompts help the model better capture the implicit information and semantic association between different roles through different forms of templates, such as manual templates, soft prompt templates and connection-based templates. This design effectively improves the information extraction ability of the model in long text, making the event argument extraction more accurate.

[0058] (2) In order to further improve the generalization ability of the model in the case of few samples, the method also introduces a bipartite graph matching and reasoning stage. In the bipartite graph matching stage, the Hungarian algorithm is used to handle the problem of multiple argument roles of the same role, ensuring the globally optimal matching result, thereby reducing the prediction error. The reasoning stage scores the candidate event argument set and uses the enumeration strategy to select the best argument span, avoiding the complexity of exhaustive threshold adjustment in traditional methods. This series of innovative designs not only enhances the accuracy of the model, but also ensures that the model can still effectively extract event arguments in the case of data scarcity. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of the method for extracting events at the paragraph level and in fine granularity based on military news.

[0060] Figure 2 An architectural diagram of event detection based on the QAED model.

[0061] Figure 3 A flowchart of the preprocessing method.

[0062] Figure 4 A flowchart of the event detection technology of the event detection model.

[0063] Figure 5 A flowchart of the event argument extraction process.

[0064] Figure 6 A flowchart of the extraction of fine-grained event arguments.

[0065] Figure 7 A flowchart of the application of transfer learning.

[0066] Figure 8 A flowchart of the comprehensive optimization processing of event arguments.

[0067] Figure 9 A flowchart of the structured output and application. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0069] The present application provides a method for extracting events at the paragraph level and in fine granularity in the field of military news, as shown in Figures 1-9 The method comprises the following steps:

[0070] S1: Data collection and preprocessing, a large number of military news articles are collected as sample data, and these data are cleaned and preprocessed to prepare for subsequent event extraction work;

[0071] S2: Recognition of event trigger words, using an event detection model based on machine reading comprehension, the preprocessed military news articles are analyzed sentence by sentence, and the event trigger words therein are accurately identified;

[0072] S3: Construction and training of MRC-based PAIE event argument extraction model, construct an MRC-based PAIE event argument extraction model, and train it using a large amount of annotated military news text data, and continuously adjust the parameters and structure of the model to better understand the relationship and context between event arguments;

[0073] S4: Extraction of fine-grained event arguments, apply the trained MRC-based PAIE event argument extraction model to military news texts, and extract fine-grained event arguments associated with each event trigger, thereby deeply mining the complex interactions between triggers, event types and arguments, and capturing long-distance dependencies between event arguments in long texts;

[0074] S5: Application of transfer learning, use the transfer learning capability of the pre-trained language model to migrate the language model parameters trained on other large-scale text data to the event extraction task of military news texts;

[0075] S6: Comprehensive optimization processing of event arguments, comprehensive optimization processing of extracted event arguments;

[0076] S7: Structured output and application, output the event arguments after comprehensive optimization in a structured form.

[0077] By introducing the PAIE model, the efficiency and accuracy of the event argument extraction in the field of military news texts are improved. The MRC-based event argument extraction method is applied to the military news texts, and a model creation prompt template part is designed according to the characteristics that event arguments may be scattered in different sentences. In this part, a set of prompts is created for each event type. These prompts help the model better capture the implicit information and semantic association between different roles through different forms of templates, such as manual templates, soft prompt templates and connection-based templates. This design effectively improves the information extraction capability of the model in long texts, making the event argument extraction more accurate.

[0078] Specifically, the preprocessing in step S1 includes the following steps:

[0079] S11: Use a Chinese word segmentation tool to perform word segmentation processing on the military news texts, and divide the texts into individual word units. According to the characteristics of military news, appropriate parameter adjustment and optimization are performed to improve the accuracy of word segmentation;

[0080] S12: Perform part-of-speech tagging to label each word with its part-of-speech information, such as noun, verb, adjective, etc. This helps better understand the semantics and grammatical functions of words in the text, providing more rich language information for subsequent event extraction;

[0081] S13: Construct a vocabulary table to collect all the words in the chapter, remove duplicate words, and assign a unique identifier to each word. The construction of the vocabulary table helps improve the efficiency and processing speed of the data representation, and also provides a basic vocabulary resource for subsequent model training and event extraction.

[0082] Specifically, the event detection technology of the event detection model in step S2 includes:

[0083] S21: Event trigger word recognition stage, which identifies event trigger words through MRC question and answer templates;

[0084] S22: Event type classification stage, which accurately classifies event types based on the identified trigger words;

[0085] In event detection, identifying event trigger words through MRC question and answer templates can accurately locate key information and provide clear basis for subsequent event type classification. Based on this, accurate classification of event types can make event detection more targeted and accurate, effectively improve the recognition and classification efficiency of various events in the field of military news and other fields, avoid ambiguous and incorrect judgments, and lay a solid foundation for subsequent event analysis and processing.

[0086] Further, the event argument extraction process in step S3 includes:

[0087] S31: Manual template, all roles are connected using natural language manually;

[0088] S32: Soft prompt template, combining different argument roles with pre-trained language model learnable and specific role pseudo tokens;

[0089] S33: Connection-based template, using natural language to connect all role names belonging to one event type;

[0090] Combining various types of prompt structures can guide and constrain event arguments from different angles. Manual templates can provide clear rules and structures, soft prompt templates increase flexibility and adaptability, and connection-based templates can better capture the relationships between event arguments. This way, you can take full advantage of various templates and avoid the limitations of a single template, so you can more comprehensively and accurately extract event arguments, significantly enhance the generalization ability of event argument extraction, and better handle different types and styles of text, improving the accuracy and stability of extraction.

[0091] Further, the PAIE event argument extraction model in step S3 creates prompt templates, event argument selection decoders, and event argument prediction to extract event arguments.

[0092] Further, the event argument prediction combines bipartite graph matching and reasoning to solve the problem of optimal assignment of multiple similar argument entities in the same event:

[0093] By introducing bipartite graph matching technology, multiple slots are inserted in the model template setting using the Hungarian algorithm to achieve global optimal assignment with minimum cost matching, and in the reasoning stage, the candidate event argument set is defined and scored;

[0094] In order to reason, the candidate event argument set is defined as:

[0095]

[0096] The candidate event argument set contains all spans shorter than the threshold L and special spans (0, 0) without extracted event arguments. The model extracts the argument of each span selector θ k by enumerating all candidate spans and scoring them: k

[0097]

[0098] And the predicted span of time slot k is given by:

[0099]

[0100] Since each slot in the prompt predicts at most one span, this strategy avoids exhaustive threshold adjustment.

[0101] Figure 2 is the architecture diagram of event detection based on QAED model.

[0102] ​By introducing the event detection technology based on the QAED (Question Answering Event Detection) model, the accuracy and generalization ability of the fine-grained event detection in the military news field are improved. First, the method converts the traditional event detection task into a multi-round machine reading comprehension (MRC) question answering problem, which is divided into two stages: trigger word recognition and trigger word classification. Specifically, the trigger word recognition stage expresses the trigger word extraction task in the unstructured text as an extractive MRC question, and uses the multi-round question answering technology based on BERT to identify the continuous trigger word phrase from the sentence or paragraph. The trigger word classification stage expresses the trigger word classification as a yes / no MRC question by further constructing questions to determine whether the identified trigger word belongs to a specific event type. This method effectively utilizes the semantic interaction between trigger words, event types, and event arguments, thereby gradually learning the relevant knowledge of event types in rich training and improving the accuracy of event detection.

[0103] The method also introduces the strategy of converting event detection problems into two consecutive MRC questions through multi-round question answering, which further solves the generalization problem of traditional methods when facing a small amount of training data. This not only enhances the scalability of the model in handling new event types, but also avoids the limitations of traditional sequence labeling-based methods, allowing the model to maintain good detection performance on unseen event types. By explicitly encoding event types as a binary classification problem, the QAED model forms a closer relationship between trigger word recognition and classification, enabling the model to better handle complex and variable military news event detection requirements with the support of rich corpus.

[0104] By introducing the PAIE model, the efficiency and accuracy of event argument extraction in the military news field at the document level are improved. First, the method innovatively applies the MRC-based event argument extraction method to the military news text at the document level, and designs a model creation prompt template part for the characteristics that event arguments may be scattered in different sentences. In this part, a set of prompts is created for each event type, which helps the model better capture the implicit information and semantic association between different roles through different forms of templates, such as manual templates, soft prompt templates, and connection-based templates. This design effectively improves the model's information extraction ability in long texts, making event argument extraction more accurate.

[0105] To further improve the generalization ability of the model in the case of few samples, the method also introduces a bipartite graph matching and reasoning stage. In the bipartite graph matching stage, the problem of multiple argument roles of the same role is solved by using the Hungarian algorithm, ensuring a globally optimal matching result and reducing prediction errors. The reasoning stage scores the candidate event argument set and uses an enumeration strategy to select the best argument span, avoiding the complexity of exhaustive threshold adjustment in traditional methods. This series of innovative designs not only enhances the accuracy of the model, but also ensures that the model can still effectively perform event argument extraction in the case of data scarcity.

[0106] The data source of the military news field passage-level fine-grained event extraction method is the public CMNEE data set. The CMNEE data set is a large-scale, open-source Chinese military news event extraction data set at the file level, containing 17,000 documents and 29,223 events, which are manually annotated according to the pre-defined military field mode, including 8 event types and 11 parameter role types. The event and argument information of the CMNEE data set is used for event extraction in this application. In the evaluation stage, the data set is generally divided into a training set and a test set. The training set is used for model training, and the test set is used for final performance evaluation.

[0107] The performance of each model is compared based on the value of the evaluation index. The evaluation index used is Precision, Recall and F1 value as the main evaluation index to test the performance and effect of the named entity recognition model on the three data sets. Among them, precision (Precision) and recall (Recall) are used as auxiliary evaluation indexes, and F1 value is calculated by considering both accuracy and recall.

[0108] To evaluate the proposed model, some competitive entity alignment models are selected for comparison. Mainly including: GLACIER, TIER, Multi-Granularity Reader. The following is an introduction to the three models of GLACIER, TIER and Multi-Granularity Reader:

[0109] GLACIER uses a unified probabilistic model for event extraction, which jointly considers sentence features and phrase features when extracting each role filling. The model consists of a sentence event classifier and a set of event role identifiers. The final extraction decision is based on the product of normalized sentence and phrase probabilities.

[0110] TIER is a multi-layer architecture for event extraction. Documents are processed in a pipeline, analyzed at different levels of textual granularity, including document, sentence, and phrase. Without event detection, TIER aims to extract event arguments by using a document type classifier and a set of event role-specific sentence classifiers.

[0111] Muti-Granularity Reader aims to explore the impact of context length on event role extraction models and proposes a multi-granularity reader that combines sentence-level and context-level granularity. Muti-Granularity Reader achieved the best results to date in the event role extraction task based on the MUC-4 dataset.

[0112] The experimental results show that the model in this chapter performs almost better than all other methods on the CMNEE dataset.

[0113] Table 1 Performance comparison of models

[0114] Model Precision (%) Recall (%) F1(％) GLACIER 48.7 58.3 53.2 TIER 51.2 62.5 57.2 Multi-Granularity Reader 58.3 60.4 59.2 This Chapter Model 74.3 70.2 73.5

[0115] It can be seen that the performance of the model in this paper is better than that of other models, because the model effectively solves the limitations of traditional methods in label semantic modeling and generalization ability by converting the event detection and argument extraction task into a multi-round question answering problem based on machine reading comprehension (MRC). At the same time, by designing a prompt template and a bipartite graph matching algorithm, the complex semantic relationships and dependency information between event triggers, event types, and arguments are deeply mined, especially in long text processing and few-shot learning. In addition, the model uses the transfer learning ability of pre-trained language models to maintain high efficiency and accuracy in event extraction even in the case of data scarcity, thereby improving the overall performance.

[0116] Specifically, the step S4 of extracting fine-grained event arguments includes the following steps:

[0117] S41: paragraph analysis, a detailed analysis of each sentence in the military news paragraph is performed, the semantic and structure of each sentence is determined, and the foundation for subsequent accurate extraction of fine-grained event arguments is laid;

[0118] S42: trigger word positioning, in the analyzed paragraph, each event trigger word is accurately positioned, and these key words are accurately identified by a model based on machine reading comprehension, laying the foundation for subsequent accurate extraction of fine-grained event arguments;

[0119] S43: argument association, for each positioned event trigger word, according to its semantic and context information, associate the fine-grained event arguments related to it, which requires a deep understanding of the internal logical relationship between the trigger word and the argument;

[0120] S44: Interaction mining, in the process of argument element association, further mining the complex interaction between trigger words, event types and argument elements, such as the difference in the influence of different trigger words on argument elements, the characteristics of argument elements under the same event type, etc.

[0121] S45: Long-distance dependency capture, through overall understanding of the article and grasp of the context, capture the long-distance dependency relationship between event argument elements in long text, which may need to span multiple sentences or paragraphs, and consider various semantic associations.

[0122] Step S4 includes fine-grained event argument extraction, including article analysis to lay the foundation, trigger word positioning precision foundation, argument association in-depth understanding of internal logic, interaction mining to grasp complex situations, and long-distance dependency capture to consider semantics. The advantage is that it can accurately extract fine-grained event arguments from military news articles. Through comprehensive analysis of the article and mining of various relationships, the accuracy and comprehensiveness of event argument extraction are improved, and the details and internal logic of events in military news are better understood, providing higher quality event argument data for subsequent related research and application.

[0123] Specifically, the application of transfer learning in step S5 includes the following steps:

[0124] S51: Select a pre-trained language model, select a suitable pre-trained language model such as BERT, GPT, etc. according to the characteristics and needs of military news articles. These models have been trained on large-scale text data and have rich language knowledge and semantic understanding ability;

[0125] S52: Model parameter transfer preparation, obtain the parameters of the pre-trained language model trained on other large-scale text data, which contain language general patterns, vocabulary semantics, etc. These parameters are the basis for transfer learning;

[0126] S53: Adapt to military news data, adapt the selected pre-trained language model to the specific field and language style of military news articles, which may include adjusting the input and output format, vocabulary table, etc. of the model to better handle military news related content;

[0127] S54: Parameter transfer and fine-tuning, transfer the parameters trained on other large-scale text data to the adapted military news article event extraction model, and then fine-tune the model using the labeled data of military news articles. By adjusting the weights and structure of the model, it can better adapt to the military news event extraction task and improve the accuracy and performance of extraction;

[0128] S55: Evaluation and optimization, evaluate the military news event extraction model after transfer learning, compare with the manually annotated data to evaluate the extraction effect of the model, and according to the evaluation result, further optimize and adjust the model, such as adjusting parameters, improving model structure, etc., to continuously improve the performance of the model in the military news event extraction task.

[0129] Step S5 utilizes the rich language knowledge and semantic understanding ability of the pre-trained language model, and through adaptation and fine-tuning, the model can quickly adapt to the military news field, improve the efficiency and quality of event extraction, save training time and cost, and improve the application effect of the model in this specific field

[0130] Specifically, the comprehensive optimization processing of the event arguments in step S6 includes the following steps:

[0131] S61: De-duplication processing, carefully check all the extracted event arguments, use specific algorithms and rules to find and remove duplicate event arguments, for example, by comparing the text content, semantic features, etc. of the arguments, merge the completely same or similar arguments, and only keep one copy to reduce redundant information;

[0132] S62: Error correction, carefully review each event argument and correct the error classification, such as correcting a certain event argument that is misclassified into the wrong event type according to relevant knowledge and rules, to ensure that each event argument can be accurately classified into the corresponding event type;

[0133] S63: Semantic consistency check, check and correct the semantic consistency between event arguments, for example, for related arguments in the same event, they should be semantically related and logically reasonable. If there is a semantic inconsistency, such as a certain argument is semantically contradictory or unreasonable with other related arguments, adjust and correct it to ensure the semantic coherence and reasonableness of the entire event argument set;

[0134] S64: Quality evaluation and screening, develop quality evaluation standards, evaluate the quality of the event arguments after de-duplication and error correction, and select high-quality event arguments that meet the requirements according to the standards, remove those with poor quality and possible problems, and further improve the quality and reliability of the event arguments.

[0135] Step S6 reduces redundant information, ensures accurate argument classification, ensures semantic coherence and reasonableness, and improves the overall quality and reliability of event arguments, providing a better data foundation for subsequent related work.

[0136] Specifically, the structured output and application in step S7 includes the following steps:

[0137] S71: Determine the structured format, and determine the structured output format to be used, such as defining a table structure containing event type, trigger word, argument, etc. field, or using JSON format, etc. Determine the specific format specification so that the event arguments can be organized according to the format in the subsequent step;

[0138] S72: Field extraction and filling, from the event arguments after comprehensive optimization, the event type, trigger word and information of each argument are extracted in turn, and these information is accurately filled into the corresponding field according to the structured format determined before. For example, in the table structure, the event type is filled into the corresponding event type field, the trigger word is filled into the trigger word field, and the argument is filled into the corresponding argument field;

[0139] S73: Organize the output order, determine the output order of event arguments according to the specific application requirements and logic, which can be classified and output according to event type, or sorted according to time sequence, importance and other logic. Ensure that the output order can meet the requirements of subsequent applications, and facilitate the analysis and utilization of event arguments;

[0140] S74: Interface with application system, interface the structured output event arguments with specific application system, if it is a data analysis system, the event arguments are input as data for further statistics, analysis and mining; if it is a visualization system, the event arguments are converted into visual charts or graphs to display the event information more intuitively; if it is other related application system, according to the characteristics and requirements of the system, the corresponding adaptation and interface are carried out, so that the event arguments can play a role in the system;

[0141] Step S7 improves data specification, provides reliable data basis, meets application requirements, improves data value, expands application range, improves business efficiency and coordination.

[0142] Finally, it should be pointed out that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A method for extracting events at the chapter level in the field of military news, characterized in that, Includes the following steps: S1: Data collection and preprocessing. Collect a large number of military news articles as sample data, clean and preprocess this data to prepare for subsequent event extraction. S2: Event trigger word identification. Using an event detection model based on machine reading comprehension, the pre-processed military news articles are analyzed sentence by sentence to accurately identify event trigger words. S3: Construction and training of PAIE event argument extraction model based on MRC. The PAIE event argument extraction model based on MRC is constructed and trained using a large amount of labeled military news articles. The parameters and structure of the model are continuously adjusted. S4: Fine-grained event argument extraction. The trained MRC-based PAIE event argument extraction model is applied to military news articles to extract fine-grained event arguments associated with each event trigger word, thereby mining the complex interactions between trigger words, event types and arguments, and capturing the long-distance dependencies between event arguments in long texts. S5: Application of transfer learning: Utilizing the transfer learning capability of pre-trained language models, the parameters of language models trained on other large-scale text data are transferred to the event extraction task of military news articles. S6: Comprehensive optimization of event arguments, which involves comprehensively optimizing the extracted event arguments; S7: Structured output and application, outputting event arguments in a structured form after comprehensive optimization; In step S3, the PAIE event argument extraction model extracts event arguments through model-created prompt templates, an event argument selection decoder, and event argument prediction. The event argument prediction, combined with bipartite graph matching and reasoning, is used to solve the problem of optimal allocation of multiple similar role arguments in the same event. By introducing bipartite graph matching technology and using the Hungarian algorithm to insert multiple slots in the model template settings, the globally optimal allocation with minimum cost matching is achieved. During the inference phase, a set of candidate event arguments is defined and scored. For the purpose of reasoning, the set of candidate event arguments is defined as: The candidate event argument set contains all spans shorter than the threshold L and the special spans 0,0 from which no event arguments are extracted. The model extracts arguments for each span selector θk by enumerating all candidate spans and scoring them. Furthermore, the prediction span of time slot k is given by the following formula: 。 2. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The preprocessing in step S1 includes the following steps: S11: Use Chinese word segmentation tools to segment military news articles into individual word units; S12: Perform part-of-speech tagging, tagging the part-of-speech information for each word; S13: Construct a vocabulary list, summarizing all words appearing in the text, removing duplicate words, and assigning a unique identifier to each word.

3. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The event detection techniques of the event detection model in step S2 include: S21: Event trigger word identification stage, wherein the event trigger word identification stage identifies event trigger words through MRC question and answer template; S22: Event type classification stage, which performs precise classification of event types based on the identified trigger words.

4. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The event argument extraction process in step S3 includes: S31: Manual template, all characters are manually connected using natural language; S32: Soft cue template, which combines different argument roles with role-specific pseudo tokens that can be learned by the pre-trained language model; S33: A connection-based template that uses natural language to connect all role names belonging to an event type.

5. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The extraction of fine-grained event arguments in step S4 includes the following steps: S41: Text analysis, a detailed sentence-by-sentence analysis of military news articles to clarify the semantics and structure of each sentence; S42: Trigger word localization. In the analyzed text, accurately locate the trigger words for each event and accurately identify these keywords through a machine reading comprehension-based model. S43: Argument association: For each located event trigger word, associate it with related fine-grained event arguments based on its semantics and context information; S44: Interactive mining, further exploring trigger words, event types, and complex interactions between arguments during the process of associating arguments; S45: Long-distance dependency capture, capturing long-distance dependencies between event arguments in long texts through a holistic understanding of the text and grasp of the context.

6. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The application of transfer learning in step S5 includes the following steps: S51: Select a pre-trained language model. Based on the characteristics and needs of military news articles, select a suitable pre-trained language model. S52: Model parameter transfer preparation, obtaining the parameters of a pre-trained language model trained on other large-scale text data; S53: Adapt military news data by adapting the selected pre-trained language model to the specific domain and language style of military news articles. S54: Parameter transfer and fine-tuning. The parameters trained on other large-scale text data are transferred to the adapted military news article event extraction model, and then the model is fine-tuned using the labeled data of military news articles. S55: Evaluation and Optimization. Evaluate the military news article event extraction model after transfer learning. Evaluate the model's extraction effect by comparing it with manually labeled data. Based on the evaluation results, further optimize and adjust the model.

7. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The comprehensive optimization processing of event arguments in step S6 includes the following steps: S61: Deduplication process. Carefully examine all extracted event arguments and use specific algorithms and rules to find and remove duplicate event arguments. S62: Error correction, a detailed examination of each event argument and correction of incorrect classifications; S63: Semantic consistency check, checks and corrects semantic consistency between event arguments; S64: Quality assessment and screening, establishing quality assessment standards, and conducting quality assessments on event arguments after deduplication and error correction.

8. The method for extracting events at the chapter level in the field of military news according to claim 1, characterized in that, The structured output and application in step S7 includes the following steps: S71: Determine the structured format and specify the required structured output format; S72: Field extraction and filling. Extract the event type, trigger words, and information of each argument from the event arguments after comprehensive optimization. S73: Organize the output order and determine the output order of event arguments based on specific application requirements and logic; S74: Interface with application systems, connecting the structured output event arguments with specific application systems.

Images