Event argument extraction task data synthesis method and system based on large language model

Abstract

The invention discloses an event argument extraction task data synthesis method and system based on a large language model, belongs to the technical field of event argument extraction, and solves the problems of event logic breakage, concentrated argument distribution, single synthesized data style and the like in the existing event argument extraction task which synthesizes data by using the large language model. The method comprises the following steps: determining an argument type to be contained in the to-be-synthesized original data according to an event type to which an event in the obtained to-be-synthesized original data belongs and an event argument contained in the corresponding event type, and constructing a cue word A to generate a single sentence only containing a single event; constructing a prompt word B to generate a paragraph text containing a plurality of events; generating a single sentence or paragraph based on the cue word C; filtering results with the formal problem, and constructing a prompt word D for marking or adding an event trigger word; constructing a cue word E for reconstruction; constructing a cue word F input large language model to generate diversified and multi-stylized output; and constructing a cue word G to obtain final synthetic data. The method is used for event argument extraction.

Description

Technical Field

[0001] A method and system for synthesizing event argument extraction task data based on a large language model are provided, which are used for event argument extraction and belong to the technical field of event argument extraction. Background Art

[0002] As a core task in event structured parsing, event argument extraction is highly dependent on the scale and quality of labeled data. Traditional data synthesis methods, such as those based on heuristic rules and generative models, have the following technical problems:

[0003] 1. Generation models based on template replacement or entity substitution cannot guarantee semantic consistency between event trigger words and argument roles;

[0004] 2. The single-event, single-sentence synthesis logic cannot cover the complex scenarios of multiple nested events and shared arguments in real texts;

[0005] 3. The generated data lacks style diversity and domain adaptability, which can easily lead to overfitting of the training model.

[0006] Although large language models provide new possibilities for data synthesis, the following technical issues still exist:

[0007] Using large language models to synthesize data has problems such as discontinuity in event logic, concentrated argument distribution, and a single style of synthesized data. Summary of the Invention

[0008] In response to the above research problems, the purpose of the present invention is to provide a data synthesis method and system for event argument extraction tasks based on a large language model, so as to solve the problems of event logic discontinuity, concentrated argument distribution, and single synthesized data style in existing event argument extraction tasks using large language models to synthesize data.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A data synthesis method for event argument extraction task based on a large language model, including the following steps:

[0011] Step 1: According to the event type of the event in the acquired raw data to be synthesized and the event arguments contained in the corresponding event type, determine the argument types to be contained in the raw data to be synthesized, including participant 1, participant 2, location of occurrence, and event summary;

[0012] Step 2: Construct a large language model that includes the event summary and argument type prompt word A and the event argument obtained in step 1 as input to generate a single sentence containing only a single event. After judging the single sentence based on the judgment mechanism, the next round of dialogue generates all the single sentences.

[0013] Step 3: Use the obtained single sentence as the context content to construct the prompt word B and the event argument of the next raw data to be synthesized obtained in step 1 to input into the large language model to generate a paragraph text containing multiple events;

[0014] Step 4: Based on the prompt word C, multiple raw data to be synthesized are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph;

[0015] Step 5: Filter out the results obtained in steps 3 and 4 that still have format problems, and construct a prompt word D based on the filtered results. Input the prompt word D and the filtered results into the large language model, and let it mark or add event trigger words to the corresponding results;

[0016] Step 6: Construct a prompt word E and input it into the large language model. The large language model is instructed to merge or refer to the event arguments that appear repeatedly in the text, and reconstruct the result of step 5 using different event arguments as the main body. The prompt word E includes the identity information of the large language model, the result obtained in step 5, the generated content requirements, and the generated format requirements.

[0017] Step 7: Construct a prompt word F and input it into the large language model to generate a diversified and multi-styled output. The prompt word F includes the identity information of the large language model, the writing style requirements, the generation format requirements, and the reconstructed generation data obtained in step 6.

[0018] Step 8. Construct a prompt word G and input it into the large language model. The large language model determines whether the result obtained in step 7 changes the core meaning of the event summary of the original data, filters the synthesis results that change the core meaning, and integrates the synthesis results that meet the requirements to obtain the final synthesized data. The prompt word G includes the identity information of the large language model, the result obtained in step 7, the event summary of the original data in step 1 corresponding to the result obtained in step 7, the requirements for judging the core content, and the requirements for outputting the judgment result.

[0019] Furthermore, the specific steps of step 2 are:

[0020] Step 2.1: Construct a prompt word A, input the event arguments from step 1 into the large language model, and generate a single sentence containing only a single event. The prompt word A includes the identity information of the large language model, generation content requirements, the argument type to be synthesized, and generation format requirements. The generation content requires the large language model to generate a single sentence containing only a single event based on the event arguments and the event summary. The generation format requirements include adding XML tags to the start and end positions of the event arguments, a word limit, and prohibiting modification of the event arguments.

[0021] Step 2.2: Establish a judgment mechanism for generating a single sentence, that is, determine whether the single sentence generated by the large language model for the first time meets the requirements for generated content and generation format. Use the regularization judgment method to determine whether it contains all the argument types determined by the original data to be synthesized, whether each argument type appears only once, whether each event argument has a corresponding XML tag added, and whether the marked argument type is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record the round of dialogue as the dialogue history and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate single sentences that meet the rules until all rounds of dialogue are completed. If it does not meet the requirements, then provide solutions to the corresponding problems of the large language model matching for the problems that arise. Use the dialogue history and solutions from step 2.1 as input, proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate a single sentence that meets the regulations until all rounds of dialogue are completed.

[0022] Furthermore, the specific steps of step 3 are:

[0023] Step 3.1: Construct the prompt word B, including the identity information of the large language model, the generated content requirements, the argument type to be synthesized, the single sentence as the context content requirement, and the generated format requirements;

[0024] Step 3.2: Based on the prompt word B, the event arguments in step 1 are input into the large language model to generate a paragraph text containing multiple events;

[0025] Step 3.3: Establish a judgment mechanism for generating paragraph text, that is, determine whether the paragraph text generated by the large language model meets the generation format requirements, use the regularization judgment method to determine whether the argument type in the paragraph text contains the argument type determined in the original data to be synthesized, whether XML is used to mark each event argument, and whether the argument type in the tag is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate paragraph text that meets the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give solutions to the corresponding problems of the large language model matching for the problems that arise, use the dialogue history and solutions in step 3.2 as input, and proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate paragraph text that meets the regulations until all rounds of dialogue are completed.

[0026] Furthermore, the specific steps of step 4 are:

[0027] Step 4.1: Construct a prompt word C, including the identity information of the large language model, generation content requirements, argument type groups 1, ..., and N to be synthesized, and generation format requirements. The generation format requirements include generating single sentences and paragraphs based on the same event arguments.

[0028] Step 4.2: Input the prompt word C and two or more groups of data to be synthesized into the large language model to obtain a single sentence or paragraph text.

[0029] Furthermore, the specific steps of step 5 are:

[0030] Step 5.1: Use a regularization method to detect whether there are formal issues in the results obtained in Steps 3 and 4. If so, filter them out; otherwise, retain them. Formal issues include missing event arguments, XML tags, and each event argument only appears once in the text generated by the large language model in Steps 3 or 4.

[0031] Step 5.2: Construct a prompt word D and input it into the large language model. The large language model is asked to find the most suitable and meaningful verb that best represents the event type trigger word in the filtered generated data. If it exists, the selected trigger word is XML-tagged in the generated data. Otherwise, all event types and their sub-event types of the original data to be synthesized are input into the large language model. It is asked to generate a group of meaningful verbs suitable for the corresponding event types as event trigger words. Then, the group of event trigger words and the filtered generated data are input into the large language model. The large language model is asked to select a meaningful verb that represents the event type trigger word from the group of event trigger words as the trigger word, and insert it into the appropriate position in the generated data with the correct tense of the meaningful verb. Among them, the prompt word D includes the identity information of the large language model, the requirements for marking trigger words, and the results obtained after filtering in step 5.1.

[0032] Furthermore, when reconstructing with different event arguments as the main body in step 6, the reconstruction methods adopted include out-of-order description and description according to time or event development.

[0033] A data synthesis system for event argument extraction tasks based on a large language model, comprising:

[0034] Type acquisition module: determines the argument types to be included in the original data to be synthesized, including participant 1, participant 2, location of occurrence, and event summary, based on the event type of the event in the original data to be synthesized and the event arguments contained in the corresponding event type;

[0035] Sentence generation module: This module constructs a large language model that includes the event summary and argument type prompt word A and the event arguments obtained in step 1 as input to generate a single sentence containing only a single event. The module then uses the judgment mechanism to judge the single sentence and then proceeds to the next round of dialogue to generate all the single sentences.

[0036] Paragraph generation module: The obtained single sentence is used as the context to construct the prompt word B and the event argument of the next raw data to be synthesized obtained in step 1, and input into the large language model to generate a paragraph text containing multiple events;

[0037] Sentence or paragraph generation module: Based on the prompt word C, multiple raw data to be synthesized are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph;

[0038] Trigger word addition module: Filters the results from steps 3 and 4 that still have format problems, constructs prompt word D based on the filtered results, inputs prompt word D and the filtered results into the large language model, and instructs it to mark or add event trigger words to the corresponding results;

[0039] Reconstruction module: Construct a prompt word E and input it into the large language model. The large language model is instructed to merge or refer to the event arguments that appear repeatedly in the text, and reconstruct the result of step 5 using different event arguments as the main body. The prompt word E includes the identity information of the large language model, the result obtained in step 5, the generated content requirements, and the generated format requirements.

[0040] Diversified Generation Module: This module constructs a prompt word F and inputs it into the large language model to generate diverse and multi-styled outputs. The prompt word F includes the identity information of the large language model, the writing style requirements, the generation format requirements, and the reconstructed generation data obtained in step 6.

[0041] Synthetic data generation module: Construct a prompt word G, input it into the large language model, let the large language model determine whether the result obtained in step 7 changes the core meaning of the event summary of the original data, filter out the synthetic results that change the core meaning, and integrate the synthetic results that meet the requirements to obtain the final synthetic data, where the prompt word G includes the identity information of the large language model, the result obtained in step 7, the event summary of the original data in step 1 corresponding to the result in step 7, the requirements for judging the core content, and the requirements for outputting the judgment result.

[0042] Furthermore, the specific implementation steps of the sentence generation module are:

[0043] Step 2.1: Construct a prompt word A, input the event arguments from step 1 into the large language model, and generate a single sentence containing only a single event. The prompt word A includes the identity information of the large language model, generation content requirements, the argument type to be synthesized, and generation format requirements. The generation content requires the large language model to generate a single sentence containing only a single event based on the event arguments and the event summary. The generation format requirements include adding XML tags to the start and end positions of the event arguments, a word limit, and prohibiting modification of the event arguments.

[0044] Step 2.2: Establish a judgment mechanism for generating a single sentence. That is, determine whether the single sentence generated by the large language model for the first time meets the requirements for generated content and generation format. Use the regularization judgment method to determine whether it contains all argument types determined by the original data to be synthesized, whether each argument type appears only once, whether each event argument has a corresponding XML tag added, and whether the marked argument type is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate single sentences that meet the rules until all rounds of dialogue are completed. If it does not meet the requirements, then provide solutions to the corresponding problems of matching the large language model for the problems that arise. Use the dialogue history and solutions from step 2.1 as input, proceed to the next round of dialogue, and inform the large language model to modify the current text that does not meet the requirements. Finally, generate a single sentence that meets the regulations until all rounds of dialogue are completed.

[0045] Furthermore, the specific implementation steps of the paragraph generation module are:

[0046] Step 3.1: Construct the prompt word B, including the identity information of the large language model, the generated content requirements, the argument type to be synthesized, the single sentence as the context content requirement, and the generated format requirements;

[0047] Step 3.2: Based on the prompt word B, the event arguments in step 1 are input into the large language model to generate a paragraph text containing multiple events;

[0048] Step 3.3: Establish a judgment mechanism for generating paragraph text, that is, to judge whether the paragraph text generated by the large language model meets the generation format requirements, and use the regularization judgment method to determine whether the argument type in the paragraph text contains the argument type determined in the original data to be synthesized, whether XML is used to mark each event argument, and whether the argument type in the tag is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, input the dialogue history and the event arguments to be synthesized in the next round into the large language model, and continue to generate paragraph text that meets the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give solutions to the corresponding problems of the large language model matching for the problems that arise, use the dialogue history and solutions in step 3.2 as input, and proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate paragraph text that meets the regulations until all rounds of dialogue are completed.

[0049] Furthermore, the specific implementation steps of the sentence or paragraph generation module are as follows:

[0050] Step 4.1: Construct a prompt word C, including the identity information of the large language model, generation content requirements, argument type groups 1, ..., and N to be synthesized, and generation format requirements. The generation format requirements include generating single sentences and paragraphs based on the same event arguments.

[0051] Step 4.2: Input the prompt word C and two or more groups of data to be synthesized into the large language model to obtain a single sentence or paragraph text;

[0052] The specific implementation steps of the trigger word adding module are:

[0053] Step 5.1: Use a regularization method to detect whether there are formal issues in the results obtained in Steps 3 and 4. If so, filter them out; otherwise, retain them. Formal issues include missing event arguments, XML tags, and each event argument only appears once in the text generated by the large language model in Steps 3 or 4.

[0054] Step 5.2: Construct a prompt word D and input it into the large language model. The large language model is asked to find the most suitable and meaningful verb that best represents the event type trigger word in the filtered generated data. If it exists, the selected trigger word is XML-tagged in the generated data. Otherwise, all event types and their sub-event types of the original data to be synthesized are input into the large language model, and it is asked to generate a group of meaningful verbs suitable for the corresponding event types as event trigger words. Then, this group of event trigger words and the filtered generated data are input into the large language model, and the large language model is asked to select a meaningful verb that represents the event type trigger word from the group of event trigger words as the trigger word, and insert the meaningful verb into the appropriate position in the generated data in the correct tense. The prompt word D includes the identity information of the large language model, the requirements for marking trigger words, and the results obtained after filtering in step 5.1.

[0055] When reconstructing with different event arguments as the main body in step 6, the reconstruction methods adopted include out-of-order description, description according to time or event development. Compared with the prior art, the present invention has the following advantages:

[0056] First, this paper proposes a data synthesis method for event argument extraction tasks based on a large language model. This method constructs a hierarchical and progressive controllable generation framework. This method not only generates traditional data synthesis types for single-event single sentences, but also achieves progressive synthesis from single events to multi-event paragraphs. Furthermore, it dynamically corrects generation deviations through a multi-round dialogue mechanism.

[0057] Second, this invention effectively improves data diversity by giving a large language model rewriting instructions and different identity features to generate event argument arrangement sequences and text variants containing diverse expression styles;

[0058] Third, the present invention ensures the integrity of argument tags and the semantic adaptability of trigger words, and can filter out synthesis results with distorted information;

[0059] 4. The present invention does not require local deployment of the model, and can call the existing large language model API. It only requires zero learning to complete data synthesis using the large language model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The figure is a flow chart of the steps of the method of the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0062] See also Figure 1 , the present invention provides a technical solution:

[0063] The present invention provides the following technical solutions:

[0064] A data synthesis method for event argument extraction task based on a large language model, including the following steps:

[0065] Step 1: According to the event type of the event in the acquired raw data to be synthesized and the event arguments contained in the corresponding event type, determine the argument types to be contained in the raw data to be synthesized, including actor 1 (actor 1), actor 2 (actor 2), location (location), and event summary (notes);

[0066] Step 2: Construct a large language model that includes a summary of the event and a prompt word A for the argument type. This input, along with the event arguments obtained in Step 1, generates a single sentence containing only a single event. The judgment mechanism then determines the sentence and generates all the individual sentences for the next round of dialogue. Specifically, to make the output of the large language model more diverse, set the temperature of the large language model to 1.5.

[0067] The specific steps are:

[0068] Step 2.1: Construct a prompt word A. Input the event arguments from step 1 into the large language model and generate a single sentence containing only a single event. The prompt word A includes the identity information of the large language model, generation content requirements, argument types to be synthesized, and generation format requirements. The generation content requires the large language model to generate a single sentence containing only a single event based on the event arguments and the event summary. The generation format requirements include adding XML tags to the start and end positions of the event arguments, word count limits, and prohibiting modification of event arguments. The identity information of the large language model may be {"role":"system","content":"You area news writer."}, which assigns the identity of the news author to the large language model, guiding the large language model to output content with higher similarity to the news. In addition to the identity information of the large language model, the prompt word A may be {"role":"user","content":(generation content requirements),(argument types to be synthesized),(generation format requirements)}.

[0069] Step 2.2, establish a judgment mechanism for generating a single sentence, that is, to judge whether the single sentence generated by the large language model for the first time meets the requirements of the generated content and the generated format, and use the regularization judgment method to judge whether it contains all the argument types determined by the original data to be synthesized, whether each argument type appears only once, whether each event argument has added the corresponding XML tag, and whether the marked argument type is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record the round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate single sentences that meet the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give the large language model matching corresponding problem solutions for the problems that arise, and use the dialogue history of step 2.1 (that is, the result record obtained in step 2.1) and the solution as input, that is,

[0070] {"role":"assistant","content":<the big language model's solution to the problem>}, then proceed to the next round of dialogue to inform the big language model to modify the text that does not meet the requirements, and finally generate a single sentence that meets the requirements until all rounds of dialogue are completed.

[0071] Step 3: Use the obtained single sentence as the context content to construct the prompt word B and the event argument of the next raw data to be synthesized obtained in step 1 to input into the large language model to generate a paragraph text containing multiple events;

[0072] The specific steps are:

[0073] Step 3.1. Construct a prompt word B, including the identity information of the large language model, the generated content requirements, the argument type to be synthesized, the single sentence as the context content requirement, and the generated format requirements. The prompt word B is: the identity information of the large language model, such as {"role":"system","content":"You are a news writer."}; in addition to the identity information of the large language model, the prompt word B is {"role":"user","content":, (generated content requirements), (argument type to be synthesized), (context: <the single sentence generated in step 2 as the context content requirement>), (generated format requirements), '};

[0074] Step 3.2: Based on the prompt word B, the event arguments in step 1 are input into the large language model to generate a paragraph text containing multiple events;

[0075] Step 3.3, establish a judgment mechanism for generating paragraph text, that is, to judge whether the paragraph text generated by the large language model meets the generation format requirements, and use the regularization judgment method to judge whether the argument type in the paragraph text contains the argument type determined in the original data to be synthesized, whether XML is used to mark each event argument, and whether the argument type in the tag is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record the round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate paragraph text that meets the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give solutions to the corresponding problems of the large language model matching for the problems that arise, and use the dialogue history and solutions of step 3.2 as input, that is,

[0076] {"role":"assistant","content":<the big language model's solution to the problem>}, and then proceed to the next round of dialogue to inform the big language model to modify the current text that does not meet the requirements, and finally generate a paragraph text that meets the requirements until all rounds of dialogue are completed.

[0077] Step 4: Based on the prompt word C, multiple raw data to be synthesized are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph;

[0078] The specific steps are:

[0079] Step 4.1. Construct a prompt word C, including the large language model identity information, generation content requirements, argument type group 1 to be synthesized, ..., argument type group N to be synthesized, and generation format requirements. The generation format requirements include generating single sentences and paragraph text based on the same event argument. In addition to the large language model identity information, prompt word C is {"role":"user","content":(generation content requirements),(argument type group 1 to be synthesized),(argument type group 2 to be synthesized)...,(generation format requirements)};

[0080] Step 4.2: Input the prompt word C and two or more different sets of raw data to be synthesized into the large language model to obtain a single sentence or paragraph text. Specifically, within a given argument type group, there are overlapping event arguments. For example, participant 1 in argument type group 1 is the same as participant 1 in argument type group 2 (similarly for participant 2 and the location of the event), and participant 1 in argument type group 1 is the same as participant 2 in argument type group 2 (the same entity plays different roles in different events).

[0081] Step 5: Filter out the results obtained in steps 3 and 4 that still have format problems, and construct a prompt word D based on the filtered results. Input the prompt word D and the filtered results into the large language model, and let it mark or add event trigger words to the corresponding results;

[0082] The specific steps are:

[0083] Step 5.1: Use a regularization method to detect whether there are formal issues in the results obtained in Steps 3 and 4. If so, filter them out; otherwise, retain them. Formal issues include missing event arguments, XML tags, and each event argument only appears once in the text generated by the large language model in Steps 3 or 4.

[0084] Step 5.2: Construct a prompt word D and input it into the large language model. The large language model is asked to find the most suitable and meaningful verb that best represents the event type trigger word in the filtered generated data. If it exists, the selected trigger word is XML-tagged in the generated data. Otherwise, all event types and their sub-event types of the original data to be synthesized are input into the large language model. It is asked to generate a group of meaningful verbs suitable for the corresponding event types as event trigger words. Then, the group of event trigger words and the filtered generated data are input into the large language model. The large language model is asked to select a meaningful verb that represents the event type trigger word from the group of event trigger words as the trigger word, and insert it into the appropriate position in the generated data with the correct tense of the meaningful verb. Among them, the prompt word D includes the identity information of the large language model, the requirements for marking trigger words, and the results obtained after filtering in step 5.1.

[0085] Step 6. Since the paragraphs generated after step 5 may contain the same argument entities in different events, for the sake of paragraph coherence, the writing of the large language model prompt words requires that repeated event arguments be merged or referred to by pronouns. At the same time, when providing event arguments in steps 2-5, the order in which event arguments are provided is fixed, and in the event summary, the writing format is basically {[date], <participant 1>...(<participant 2>) at <location>, (consequences caused)}. This results in a similar writing style for the text generated by the large language model, requiring the large language model to rewrite the paragraph without changing the original arguments, tags, or meaning, and provide a structure that it can refer to, such as: causal order, chronological order, and the order of location changes. It is necessary to construct a prompt word E and input it into the large language model, so that the large language model can merge or refer to the event arguments that appear repeatedly in the text, and reconstruct the result of step 5 with different event arguments as the main body. The prompt word E includes the identity information of the large language model (such as {"role":"system","content":"You are a news editor."}, which gives the large language model the identity of a news editor, allowing the large language model to better rewrite the original paragraph), the result obtained in step 5, the generated content requirements and the generated format requirements; among them, when reconstructing with different event arguments as the main body, the reconstruction methods adopted include out-of-order description and description according to time or event development.

[0086] Step 7: Construct prompt words F to be input into the large language model to generate diverse and multi-styled outputs. Prompt words F include the large language model's identity information, writing style requirements, generation format requirements, and the reconstructed generated data from Step 6. Using prompt words, the large language model is given different identity characteristics, resulting in results with different linguistic styles. Some require the addition of emotionally charged vocabulary to express the emotional tone of the expanded text, as well as different information sources to enrich the expanded text's style. The large language model is instructed to rewrite the text while preserving the core meaning of the text, taking into account factors such as source material (social media, newspapers, academic journals), writing format (meeting minutes, speech transcripts), and descriptive techniques (detailed descriptions).

[0087] Step 8: Construct a prompt word G and input it into the large language model. The large language model is then asked to determine whether the result from step 7 changes the core meaning of the event summary of the original data, filter out the synthesis results that change the core meaning, and integrate the synthesis results that meet the requirements to obtain the final synthesized data. The prompt word G includes the identity information of the large language model, the result from step 7, the event summary of the original data corresponding to the result from step 7 in step 1, the core content judgment requirements, and the judgment result output requirements. The text generated in step 7 and the event summary of all the original data it covers are used as input. The large language model is asked to determine whether the core meaning of the generated text is different from the event summary of the original data. Depending on the quantity, manual judgment can be performed in another step. Finally, the generated text that changes the original meaning is filtered out. The filtered text is organized and converted into JSON file format, marking the event type, event argument, and the starting and ending positions of the event argument in the text for each text.

[0088] The present invention constructs a progressive data synthesis framework based on a large language model. Its main contents include argument expansion, sentence superposition, text shuffling, text diversification, filtering and trigger word addition, and comparative organization. Specifically:

[0089] Use prompt words to guide the large language model to initially synthesize event arguments into a sentence;

[0090] Multiple synthesized sentences are superimposed to form paragraphs, enriching the diversity of data forms and thus improving the applicability of model training;

[0091] Change the order of the text to increase data diversity in terms of format;

[0092] Changing the source, format, and description of texts to create multi-style data, moving closer to the diversity found in manually annotated datasets;

[0093] Filter out noise data and add trigger words to complete the data synthesis process;

[0094] Use the large language model to determine whether the generated text changes the meaning of the original data, organize the data that meets the rules, and use it as the training input of the model; (English data is used in the experiment, so the example data is in English, and the prompt words are also in English)

[0095] Step 1. Determine the argument type to be synthesized based on the event type and event arguments contained in the acquired raw data. Here, the argument types to be synthesized are determined to be participant 1, participant 2, location, and event summary. The raw data to be synthesized are: participant 1: Unidentified Armed Group (Russia), participant 2: Civilians (Ukraine), location: Edush, event summary: Property destruction: On July 19, 2024, according to local sources, unidentified individual(s) turned off the valve on the water pipeline near Edush village, Rostov region. This led to a rupture in the pipeline, and the water supply to the parts of the Russian-occupied Donetsk region of Ukraine was ceased. This act was suspected of sabotage.

[0096] Step 2: Write the prompt word A and let the large language model synthesize the event summary and event arguments.

[0097] Step 2.1, write prompt word A,

[0098] Assign the big model an identity, {"role":"system","content":"You are a news writer."}. Here, the big language model is given the identity of a news writer, guiding the big language model to output content that is more similar to the news.

[0099] {"role": "user", "content": 'Please generate a sentence based on the given Note and Elements, strictly adhering to the following requirements:

[0100] Elements to include:

[0101] <actor1>:Unidentified Armed Group(Russia)

[0102] <actor2>:Civilians(Ukraine)

[0103] <location>:Edush

[0104] notes:Property destruction: On 19 July 2024,according to localsources,unidentified individual(s)turned off the valve on the water pipelinenear Edush village,Rostov region.This led to a rupture in the pipeline andthe water supply to the parts of Russian-occupied Donetsk region of Ukrainewas ceased.This act was suspected as a sabotage activity.

[0105] Formatting Rules:

[0106] Never modify the phrasing of provided elements;

[0107] Enclose each element(actor1,actor2,location)with its designated XML-style tags;

[0108] Each tagged element(actor1,actor2,location)must appear exactly oncein the entire output if <actor2>is null do not tag it;

[0109] No more than 150 words;

[0110] Please return only the main sentence text;

[0111] Prhibitions: Avoid repetitve tagging.′}

[0112] The generated content requirement here is to generate a sentence based on the event summary that must contain all the given event arguments, and the given event arguments cannot be changed; the generated format requires that each argument needs to be XML-tagged and can only appear once in the original text. Since the large language model will add annotations such as "OK, the sentence has been generated according to your requirements" after generating data, the large language model is required to only return the generated sentence in the prompt word; the word count requirement is no more than 150 words; since the sentences generated by the large model may contain content with similar meanings to the event arguments, repeated marking is prohibited here to prevent the model from marking them in the same way. Input the prompt word A into the large model and get the output of the large language model: On 19 July 2024, an <actor1>Unidentified Armed Group (Russia) sabotaged a waterpipeline near <location>Edush< / location> , causing a rupture that cut off water supply to <actor2>Civilians (Ukraine) in the Russian-occupied Donetskregion.

[0113] Step 2.2: After obtaining the output of the large language model, use the regularization method to determine whether it contains all the argument types determined by the original data to be synthesized, whether each argument is marked with XML, etc. For example, if the output obtained is On 19July 2024, an <actor1>Unidentified Armed Group (Russia) sabotaged awater pipeline near Edush, causing a rupture that cut off water supply to <actor2>Civilians (Ukraine) in the Russian-occupied <location>Donetsk< / location> region., regularization determines that the output incorrectly marks the location argument, and gives the corresponding error message based on the regularization judgment result: location content mismatch: expected 'Edush', got 'Donetsk'. Integrate it with the output and input of the previous round of dialogue into [prompt word A, {"role":"assistant","content": obtain the incorrect output of the large language model}, {"role":"user","content": give the corresponding error message}] and input it into the large model again, pointing out the error in the output of the large language model, and let it generate the result after correcting the problem. If the output meets the requirements, the record is integrated into [prompt word A, prompt word A' containing the next original data to be synthesized] and input into the large language model, where prompt word A' does not contain the generation format requirements, but only contains the next original data to be synthesized. <actor1> 、 <actor2> 、 <location>This can effectively reduce the number of incoming tokens while ensuring that the output of a large language model meets the format requirements.

[0114] Step 3: Instruct the large language model to use the single sentence obtained in step 2 as context and superimpose it with the next set of raw data to be synthesized to form a paragraph containing two events;

[0115] Step 3.1, write prompt word B,

[0116] {"role":"system","content":"You are a news editor."}, giving the large model the identity of a news editor.

[0117] {"role":"user","content":"Generate a paragraph based on providedelements and combined with the context, strictly adhering to the following requirements:

[0118] Elements to Include(must appear exactly once unless null):

[0119] <actor1>:Military Forces of Ukraine(2019-)Air Force

[0120] <actor2>:Civilians(Russia)

[0121] <location>:Gornal

[0122] notes:On 19 July 2024,Ukrainian forces carried out FPV-kamikaze dronestrikes,dropped drone-borne explosives and shelled unidentified targets inGornal village,Kursk region,with(likely)artillery.1 local civilian man waskilled.

[0123] context:On 19 July 2024,an <actor1>Unidentified Armed Group (Russia) sabotaged a water pipeline near <location>Edush< / location> ,causing arupture that cut off water supply to <actor2>Civilians (Ukraine) in theRussian-occupied Donetsk region.

[0124] Formatting Rules:

[0125] Enclose each element(actor1,actor2,location)with its designated XML-style tags

[0126] Never modify the phrasing of provided elements

[0127] Keep the tags from context.

[0128] Stylistic Flexibility:

[0129] Ensure logical coherence when combining with context input.

[0130] Prohibitions:

[0131] Avoid repetitive tagging.″}

[0132] The generated content here requires the single sentence generated in step 2 to be used as context, and the second data set synthesized is concatenated with the context to form a paragraph containing two events. The generated format requires that all arguments must appear only once and be marked with the correct XML tags.

[0133] Step 3.2: Input the above prompt words into the large language model and get the output as follows:

[0134] On 19 July 2024, the <actor1>Military Forces of Ukraine (2019 - ) Air Force conducted FPV-kamikaze drone strikes, dropped drone-borne explosives, and shelled unidentified targets in <location>Gornal< / location> village, Kurskregion, likely using artillery.The attack resulted in the death of 1 <actor2>Civilians (Russia) .Meanwhile, in a separate incident, an <actor1>Unidentified Armed Group (Russia) sabotaged a water pipeline near <location>Edush< / location> , disrupting water supply to <actor2>Civilians (Ukraine) in the Russian-occupied Donetsk region.

[0135] Step 3.3: After obtaining the output of the large language model, use the regularization method to determine whether the argument types in the paragraph text contain the argument types determined in the original data to be synthesized, whether each event argument is marked with XML, etc. The output Output3 is as follows:

[0136] On 19 July 2024, the <actor1>Military Forces of Ukraine (2019 - ) Air Force launched FPV-kamikaze drone strikes, dropped drone-borne explosives, and shelled unidentified targets in <location>Gornal< / location> village, Kurskregion, likely using artillery.The attack killed one <actor2>Civilians (Russia) and injured several others, marking a tragic escalation for <actor2>Civilians (Russia) in the area.Meanwhile, in a separate incident, an <actor1>Unidentified Armed Group (Russia) sabotaged a water pipelinenear <location>Edush< / location> , cutting off supply to <actor2>Civilians (Ukraine) in Russian-occupied Donetsk. The dual incidents underscored the worsening humanitarian toll on civilians in the conflict.

[0137] Regularization determines that the argument Civilians(Russia) appears repeatedly and gives the corresponding error message: Tagged element Civilians(Russia) must appear exactly once. It is integrated with the output and input of the previous round of dialogue into [prompt word B, {″role″:″assistant″,″content″:Output3}, {″role″:″user″,″content″: (give the corresponding error message)}] and input into the large language model again, pointing out the errors in the output of the large language model and making it generate the result after correcting the problem.

[0138] If there is no problem, just organize and store it in a CSV file.

[0139] Step 4: In addition to superimposing the synthesized data in steps 2 and 3, in order to diversify the data forms, based on the prompt word C, multiple raw data to be synthesized in step 1 are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph. This step is more suitable for situations where there is overlap in event arguments in the two sets of data.

[0140] Step 4.1: Construct the prompt word C as follows:

[0141] {"role":"system","content":"You are a news writer and editor."}, which gives the large language model the identities of news writer and news editor;

[0142] {″role″:″user″,″content″:″i will give you two groups of elements,eachgroup has″actor1″,″actor2″,″location″elements and a Notes.According to thenotes,generate a sentence to combine and expand the two notes and put theelements into an brief sentence.

[0143] Formatting Rules:

[0144] Enclose each element(actor1,actor2,location)with its designated XML-style tags

[0145] Never modify the phrasing of provided elements

[0146] Keep the tags from context.

[0147] Prohibitions:

[0148] Avoid repetitive tagging.

[0149] The followings are the two group:

[0150] Group1:

[0151] <actor1>:Unidentified Armed Group(Ukraine)

[0152] <actor2>:Civilians(Ukraine)

[0153] <location>:Lviv

[0154] notes:On 19 July 2024,an unidentified individual shot a Ukrainianlinguist and a former Parliament member from Svoboda party in Lviv.As aresult,the woman was hospitalized in critical condition and died beforemidnight.

[0155] Group2:

[0156] <actor1>:Military Forces of Russia (2000-)

[0157] <actor2>:Civilians(Ukraine)

[0158] <location>: Mykolaiv

[0159] notes: On 19 July 2024, Russian forces shelled Mykolaiv, killing 3 civilians, including a child. 15 civilians were injured.″}

[0160] Step 4.2: Get the large language model output as follows:

[0161] On 19July 2024, an <actor1>Unidentified Armed Group (Ukraine) shot a Ukrainian linguist and former Parliament member in <location>Lviv< / location> , leaving the woman dead after hospitalization, while the <actor1>Military Forces of Russia (2000 - ) shelled <location>Mykolaiv< / location> , killing 3%actor2>Civilians(Ukraine)< / location> , including a child, and injuring 15 others.

[0162] Since Participant 2 is the same in both sets of data, Participant 2 appears only once in the output result, preventing arguments from being repeated and increasing the difficulty of model training.

[0163] Step 5: Filter out the results from Steps 3 and 4 that still have format problems, and construct a prompt word D based on the results. Input the prompt word D and the filtered results into the large language model, and let it mark or add event trigger words to the corresponding results.

[0164] Step 5.1: Use the regularization method to filter out the results in steps 3 and 4 that still have formal problems, clean up the annotation text in the output results of the large language model, and fix some problems in the erroneous results (such as correcting errors in XML tag content, deleting unnecessary duplicate tags, etc.), and organize the filtered results.

[0165] Step 5.2: First, based on all event types and sub-event types in the original data, let the large language model output a meaningful verb phrase that is suitable as the trigger word for that event type. For example, in the Battles event, the Armed clash type, the large language model outputs the following alternative trigger word phrases: ["clash", "fight", "shoot", "engage", "battle", "confront"].

[0166] Construct the prompt word D,

[0167] {"role":"system","content":"Now you are a professional annotator for event data."}, giving the large language model the identity of a professional annotator for event data.

[0168] {"role":"user","content":"The text below has two events.

[0169] Please mark one notional verb for each event in the text which is most relevant to event_type and use XML <trigger>< / trigger> to mark it.Anddeclare the trigger belongs to which event.

[0170] Attention:If there is no such verb,please output Eventl / Event2:trigger False.

[0171] Use Event1:trigger Event2:trigger to output.

[0172] One notional verb can be trigger,not a phrase.

[0173] Do not output any content beyond result.

[0174] Eventl:event_type:Strategic developments,

[0175] sub_event_type:Looting / property destruction.

[0176] actor1:Unidentified Armed Group(Russia),

[0177] actor2:Civilians(Ukraine),

[0178] location:Edush.

[0179] Event2:event_type:Explosions / Remote violence,

[0180] sub_event_type:Air / drone strike,

[0181] actor1:Military Forces of Ukraine(2019-)Air Force,

[0182] actor2:Civilians(Russia),

[0183] location:Gornal

[0184] Here is the text:On 19July 2024,the <actor1>Military Forces of Ukraine (2019 - ) Air Force conducted FPV-kamikazedrone strikes,droppeddrone-borne explosives,and shelled unidentified targets in<1ocation>Gornal < / location> village, Kursk region, likely using artillery. The attack resulted in the death of 1 <actor2>Civilians (Russia) .Meanwhile, in a separate incident, an <actor1>Unidentified Armed Group (Russia) sabotaged a waterpipeline near <location>Edush< / location> , disrupting water supply to <actor2>Civilians (Ukraine) in the Russian-occupied Donetsk region.″}

[0185] This includes the data filtered out in step 5.1, and the trigger word labeling requirement. The large language model is required to label the content verbs in the text compiled in step 5.1 that are suitable as event triggers. If a specific event does not have a content verb in the text that is suitable as a trigger word, the output is Event1 / Event2: trigger False. After receiving the large language model's response of Event1 / Event2: trigger False, add the prompt word:

[0186] {"role": "user", "content": "Select the trigger word for Eventl / 2 inTriggers: [corresponding trigger phrase] that best matches the text meaning and event_type from the given trigger phrases, and add it to the appropriate position in the text while preserving all facts, tags and core meaning.

[0187] Trigger must be a notional verb.

[0188] Please use XML <trigger>< / trigger> to mark the trigger.

[0189] Output the text after you add the trigger.

[0190] Do not output any content beyond result.″}

[0191] The large language model is required to find the most appropriate content verb as the trigger word of the event in the previously generated trigger phrases of the corresponding event type, add it to the appropriate position of the original text with the correct tense, and mark it with XML tags.

[0192] The output of the large language model for the example prompt word D is:

[0193] Event 1: trigger <trigger>sabotaged< / trigger>

[0194] Event 2: trigger <trigger>shelled< / trigger>

[0195] Just mark the data filtered out in step 5.-1 at the location of the trigger word.

[0196] Step 6: Construct prompt word E:

[0197] {″role″:″system″,″content″:″Now you are a professional texteditor.″}, giving the large model the identity of a professional text editor.

[0198] {″role″:″user″,″content″:″Please rewrite the following text to transform it into Completely different structures while keeping all tags and core meaning.

[0199] Do not change the words in the tags.

[0200] Do not generate tags that origin text does not have.

[0201] Do not output any content beyond your written results.

[0202] Make sure the numbers of tags in your output are the same as the following text.

[0203] Text:On 19 July 2024,the <actor1>Military Forces of Ukraine (2019 - ) Air Force conducted FPV-kamikaze drone strikes, dropped drone-borne explosives, and <trigger>shelled< / trigger> unidentified targets in <location>Gornal< / location> village, Kursk region, likely using artillery.The attack resulted in the death of 1 <actor2>Civilians (Russia) .Meanwhile, in separate incident, an <actor1>Unidentified Armed Group (Russia) <trigger>sabotaged< / trigger> a water pipeline near <location>Edush< / location> , disrupting water supply to <actor2>Civilians (Ukraine) in the Russian-occupied Donetsk region.″}

[0204] This includes the result obtained in step 5, the generated content requirements, the generated format requirements, and inputs them into the large language model. The large language model is instructed to merge or refer to event arguments that appear repeatedly in the text, and to use different event arguments as the main body to describe them in disorder or reconstruct the result of step 5 in the order of time, event development, etc., so that the synthesized data contains a certain logic.

[0205] For example, the prompt word E, the large language model result is:

[0206] The <actor1>Military Forces of Ukraine (2019 - ) Air Force carried out coordinated strikes in <location>​< / location> village, Kursk region on19 July 2024, <trigger>​< / trigger> unidentified targets with artillerywhiledeploying FPV-kamikaze drones and dropping explosives, which killed one <actor2>​ .In response, an <actor1>​ <trigger>​< / trigger> a water pipeline near <location>​< / location> , disrupting supplies to <actor2>​ in the occupied Donetsk region later that same day.

[0207] Step 7: Use the prompt word F to assign different identity characteristics and writing style requirements to the large language model, thereby obtaining expansion results with different language styles. Emotionally charged words are added to express the emotional tone of the expanded text, and different information sources are added to enrich the style of the expanded text. By considering the source medium, writing tone, and descriptive techniques, the large language model is instructed to rewrite the text generated in Step 6 while preserving the core meaning of the text.

[0208] Give large language models different identity information, such as: reporters, magazine authors, essayists, meeting minutes keeper, social media news writers, etc.

[0209] Construct the prompt word F:

[0210] {"role":"system","content":"Now you are ajournalist."}, here the large language model is given the identity of a journalist;

[0211] {″role″:″user″,″content″:″Please rewrite the following text to transform it so that make its style different by mimicking different platforms / tone-of-voice while keeping all tags and core meaning.

[0212] Do not generate tags that origin text does not have.

[0213] Do not output any content beyond your results.

[0214] Make sure the numbers of tags in your output are the same as the following text.

[0215] Do not appear emoji.

[0216] Text:The <actor1>​ carried out coordinated strikes in <location>​< / location> village, Kurskregion on 19 July 2024, <trigger>​< / trigger> unidentified targets with artillery while deploying FPV-kamikaze drones and dropping explosives, which killed one <actor2>​ .In response,an <actor1>Unidentified Armed Group(Russia) < / actor1> <trigger> sabotaged < / trigger> a waterpipeline near <location> Edush < / location> , disrupting supplies to <actor2> Civilians(Ukraine) < / actor2> in the occupied Donetsk region later that same day.″},

[0217] This includes the data after word order adjustment in step 6, content generation requirements, and format generation requirements. The content generation requirement is to rewrite the text without changing the core meaning of the text and retaining all arguments. The format generation requirement is to retain all original XML tags. In addition, due to the diversity of styles, the large language model may output emoticons, so the prompt word specifically prohibits emoticons. The result of inputting the prompt word of the above example into the large language model is:

[0218] In a significant escalation of hostilities, the <actor1> Military Forcesof Ukraine(2019-)Air Force < / actor1> executed precision strikes near <location> Gornal < / location> village in Russia′s Kursk region on 19 July 2024, <trigger> shelling < / trigger> suspected military positions with artillery and kamikazedrones. Tragically, these operations claimed the life of one <actor2> Civilians(Russia) < / actor2> .Hours later,retaliatory actions emerged as an <actor1> Unidentified Armed Group(Russia) < / actor1> <trigger> sabotaged < / trigger> criticalwater infrastructure outside <location> Edush < / location> , leaving <actor2> Civilians(Ukraine)< / actor2> without running water in occupied Donetsk.

[0219] Step 8: Compare the text with the trigger words added to the language model in step 7 with the original data, and organize the synthesized data that meets the requirements;

[0220] Step 8.1, construct the prompt word G:

[0221] {″role″:″system″,″content″:″You are a text similarity judgment expert″}, here gives the large language model the identity of a text similarity judgment expert

[0222] {"role": "user", "content": "Please check if the core meaning of the text is the same as that of the two notes. If they are the same, output True; if they are not the same, output False.

[0223] Text: [Step 7 generates data]

[0224] Notesl: [Raw Data 1 Event Summary]

[0225] Notes 2: [Raw Data 2 Event Summary]"}

[0226] The large language model determines whether the text generated in step 7 changes the general meaning of the event summary in the original data. If so, it outputs False; if not, it outputs True. Data with a False result after the large language model's judgment requires manual review to prevent filtering out incorrect results from the large language model.

[0227] Step 8.2: Organize the filtered text and convert it into JSON file format, marking the event type, argument, and the starting and ending positions of each text.

[0228] The above are only representative embodiments of the present invention in many specific application scopes and do not constitute any limitation on the protection scope of the present invention. Any technical solutions formed by transformation or equivalent replacement fall within the scope of protection of the present invention. < / location> < / location> < / actor2> < / actor1> < / location>

Claims

1. A data synthesis method for event argument extraction task based on a large language model, characterized in that: The steps are as follows: Step 1: Determine the argument types to be included in the original data to be synthesized, including participant 1, participant 2, location of occurrence, and event summary, based on the event type of the event in the original data to be synthesized and the event arguments contained in the corresponding event type; Step 2: Construct a large language model that includes the event summary and argument type prompt word A and the event argument obtained in step 1 as input to generate a single sentence containing only a single event. After judging the single sentence based on the judgment mechanism, the next round of dialogue generates all the single sentences. Step 3: Use the obtained single sentence as the context content to construct the prompt word B and the event argument of the next raw data to be synthesized obtained in step 1 to input into the large language model to generate a paragraph text containing multiple events; Step 4: Based on the prompt word C, multiple raw data to be synthesized are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph; Step 5: Filter out the results obtained in steps 3 and 4 that still have format problems, and construct a prompt word D based on the filtered results. Input the prompt word D and the filtered results into the large language model, and let it mark or add event trigger words to the corresponding results; Step 6: Construct a prompt word E and input it into the large language model. The large language model is instructed to merge or refer to the event arguments that appear repeatedly in the text, and reconstruct the result of step 5 using different event arguments as the main body. The prompt word E includes the identity information of the large language model, the result obtained in step 5, the generated content requirements, and the generated format requirements. Step 7: Construct a prompt word F and input it into the large language model to generate a diversified and multi-styled output. The prompt word F includes the identity information of the large language model, the writing style requirements, the generation format requirements, and the reconstructed generation data obtained in step 6. Step 8. Construct a prompt word G and input it into the large language model. The large language model determines whether the result obtained in step 7 changes the core meaning of the event summary of the original data, filters the synthesis results that change the core meaning, and integrates the synthesis results that meet the requirements to obtain the final synthesized data. The prompt word G includes the identity information of the large language model, the result obtained in step 7, the event summary of the original data in step 1 corresponding to the result obtained in step 7, the requirements for judging the core content, and the requirements for outputting the judgment result.

2. The method for synthesizing event argument extraction task data based on a large language model according to claim 1, characterized in that: The specific steps of step 2 are: Step 2.1: Construct a prompt word A, input the event arguments from step 1 into the large language model, and generate a single sentence containing only a single event. The prompt word A includes the identity information of the large language model, generation content requirements, the argument type to be synthesized, and generation format requirements. The generation content requires the large language model to generate a single sentence containing only a single event based on the event arguments and the event summary. The generation format requirements include adding XML tags to the start and end positions of the event arguments, a word limit, and prohibiting modification of the event arguments. Step 2.2: Establish a judgment mechanism for generating a single sentence, that is, determine whether the single sentence generated by the large language model for the first time meets the requirements for generated content and generation format. Use the regularization judgment method to determine whether it contains all the argument types determined by the original data to be synthesized, whether each argument type appears only once, whether each event argument has a corresponding XML tag added, and whether the marked argument type is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record the round of dialogue as the dialogue history and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate single sentences that meet the rules until all rounds of dialogue are completed. If it does not meet the requirements, then provide solutions to the corresponding problems of the large language model matching for the problems that arise. Use the dialogue history and solutions from step 2.1 as input, proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate a single sentence that meets the regulations until all rounds of dialogue are completed.

3. The method for synthesizing event argument extraction task data based on a large language model according to claim 2, characterized in that: The specific steps of step 3 are: Step 3.1: Construct the prompt word B, including the identity information of the large language model, the generated content requirements, the argument type to be synthesized, the single sentence as the context content requirement, and the generated format requirements; Step 3.2: Based on the prompt word B, the event arguments in step 1 are input into the large language model to generate a paragraph text containing multiple events; Step 3.3: Establish a judgment mechanism for generating paragraph text, that is, determine whether the paragraph text generated by the large language model meets the generation format requirements, use the regularization judgment method to determine whether the argument type in the paragraph text contains the argument type determined in the original data to be synthesized, whether XML is used to mark each event argument, and whether the argument type in the tag is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate paragraph text that meets the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give solutions to the corresponding problems of the large language model matching for the problems that arise, use the dialogue history and solutions in step 3.2 as input, and proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate paragraph text that meets the regulations until all rounds of dialogue are completed.

4. The method for synthesizing event argument extraction task data based on a large language model according to claim 3, characterized in that: The specific steps of step 4 are: Step 4.1: Construct a prompt word C, including the identity information of the large language model, generation content requirements, argument type groups 1, ..., and N to be synthesized, and generation format requirements. The generation format requirements include generating single sentences and paragraphs based on the same event arguments. Step 4.2: Input the prompt word C and two or more groups of data to be synthesized into the large language model to obtain a single sentence or paragraph text.

5. The method for synthesizing event argument extraction task data based on a large language model according to claim 4, characterized in that: The specific steps of step 5 are: Step 5.1: Use a regularization method to detect whether there are formal issues in the results obtained in Steps 3 and 4. If so, filter them out; otherwise, retain them. Formal issues include missing event arguments, XML tags, and each event argument only appears once in the text generated by the large language model in Steps 3 or 4. Step 5.2: Construct a prompt word D and input it into the large language model. The large language model is asked to find the most suitable and meaningful verb that best represents the event type trigger word in the filtered generated data. If it exists, the selected trigger word is XML-tagged in the generated data. Otherwise, all event types and their sub-event types of the original data to be synthesized are input into the large language model. It is asked to generate a group of meaningful verbs suitable for the corresponding event types as event trigger words. Then, the group of event trigger words and the filtered generated data are input into the large language model. The large language model is asked to select a meaningful verb that represents the event type trigger word from the group of event trigger words as the trigger word, and insert it into the appropriate position in the generated data with the correct tense of the meaningful verb. Among them, the prompt word D includes the identity information of the large language model, the requirements for marking trigger words, and the results obtained after filtering in step 5.

1.

6. The method for synthesizing event argument extraction task data based on a large language model according to claim 1, characterized in that: When reconstructing with different event arguments as the main body in step 6, the reconstruction methods adopted include out-of-order description and description according to time or event development.

7. A data synthesis system for event argument extraction tasks based on a large language model, characterized in that: include: Type acquisition module: determines the argument types to be included in the original data to be synthesized, including participant 1, participant 2, location of occurrence, and event summary, based on the event type of the event in the original data to be synthesized and the event arguments contained in the corresponding event type; Sentence generation module: This module constructs a large language model that includes the event summary and argument type prompt word A and the event arguments obtained in step 1 as input to generate a single sentence containing only a single event. The module then uses the judgment mechanism to judge the single sentence and then proceeds to the next round of dialogue to generate all the single sentences. Paragraph generation module: The obtained single sentence is used as the context to construct the prompt word B and the event argument of the next raw data to be synthesized obtained in step 1, and input into the large language model to generate a paragraph text containing multiple events; Sentence or paragraph generation module: Based on the prompt word C, multiple raw data to be synthesized are input into the large language model at one time, and all of them are synthesized into a single sentence or paragraph; Trigger word addition module: Filters the results from steps 3 and 4 that still have format problems, constructs prompt word D based on the filtered results, inputs prompt word D and the filtered results into the large language model, and instructs it to mark or add event trigger words to the corresponding results; Reconstruction module: Construct a prompt word E and input it into the large language model. The large language model is instructed to merge or refer to the event arguments that appear repeatedly in the text, and reconstruct the result of step 5 using different event arguments as the main body. The prompt word E includes the identity information of the large language model, the result obtained in step 5, the generated content requirements, and the generated format requirements. Diversified Generation Module: This module constructs a prompt word F and inputs it into the large language model to generate diverse and multi-styled outputs. The prompt word F includes the identity information of the large language model, the writing style requirements, the generation format requirements, and the reconstructed generation data obtained in step 6. Synthetic data generation module: Construct a prompt word G, input it into the large language model, let the large language model determine whether the result obtained in step 7 changes the core meaning of the event summary of the original data, filter out the synthetic results that change the core meaning, and integrate the synthetic results that meet the requirements to obtain the final synthetic data, where the prompt word G includes the identity information of the large language model, the result obtained in step 7, the event summary of the original data in step 1 corresponding to the result in step 7, the requirements for judging the core content, and the requirements for outputting the judgment result.

8. The event argument extraction task data synthesis system based on a large language model according to claim 7 is characterized in that: The specific implementation steps of the sentence generation module are: Step 2.1: Construct a prompt word A, input the event arguments from step 1 into the large language model, and generate a single sentence containing only a single event. The prompt word A includes the identity information of the large language model, generation content requirements, the argument type to be synthesized, and generation format requirements. The generation content requires the large language model to generate a single sentence containing only a single event based on the event arguments and the event summary. The generation format requirements include adding XML tags to the start and end positions of the event arguments, a word limit, and prohibiting modification of the event arguments. Step 2.2: Establish a judgment mechanism for generating a single sentence. That is, determine whether the single sentence generated by the large language model for the first time meets the requirements for generated content and generation format. Use the regularization judgment method to determine whether it contains all argument types determined by the original data to be synthesized, whether each argument type appears only once, whether each event argument has a corresponding XML tag added, and whether the marked argument type is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history and proceed to the next round of dialogue. In the next round of dialogue, the dialogue history and the event arguments to be synthesized in the next round are input into the large language model, and continue to generate single sentences that meet the rules until all rounds of dialogue are completed. If it does not meet the requirements, then provide solutions to the corresponding problems of matching the large language model for the problems that arise. Use the dialogue history and solutions from step 2.1 as input, proceed to the next round of dialogue, and inform the large language model to modify the current text that does not meet the requirements. Finally, generate a single sentence that meets the regulations until all rounds of dialogue are completed.

9. The event argument extraction task data synthesis system based on a large language model according to claim 8, characterized in that: The specific implementation steps of the paragraph generation module are: Step 3.1: Construct the prompt word B, including the identity information of the large language model, the generated content requirements, the argument type to be synthesized, the single sentence as the context content requirement, and the generated format requirements; Step 3.2: Based on the prompt word B, the event arguments from step 1 are input into the large language model to generate a paragraph text containing multiple events: Step 3.3: Establish a judgment mechanism for generating paragraph text, that is, to judge whether the paragraph text generated by the large language model meets the generation format requirements, and use the regularization judgment method to determine whether the argument type in the paragraph text contains the argument type determined in the original data to be synthesized, whether XML is used to mark each event argument, and whether the argument type in the tag is the same as the argument type of the original data to be synthesized. If it meets the requirements, then record this round of dialogue as the dialogue history, and proceed to the next round of dialogue. In the next round of dialogue, input the dialogue history and the event arguments to be synthesized in the next round into the large language model, and continue to generate paragraph text that meets the rules until all rounds of dialogue are completed. If it does not meet the requirements, then give solutions to the corresponding problems of the large language model matching for the problems that arise, use the dialogue history and solutions in step 3.2 as input, and proceed to the next round of dialogue to inform the large language model to modify the current text that does not meet the requirements, and finally generate paragraph text that meets the regulations until all rounds of dialogue are completed.

10. The event argument extraction task data synthesis system based on a large language model according to claim 9, characterized in that: The specific implementation steps of the sentence or paragraph generation module are: Step 4.1: Construct a prompt word C, including the identity information of the large language model, generation content requirements, argument type groups 1, ..., and N to be synthesized, and generation format requirements. The generation format requirements include generating single sentences and paragraphs based on the same event arguments. Step 4.2: Input the prompt word C and two or more groups of data to be synthesized into the large language model to obtain a single sentence or paragraph text; The specific implementation steps of the trigger word adding module are: Step 5.1: Use a regularization method to detect whether there are formal issues in the results obtained in Steps 3 and 4. If so, filter them out; otherwise, retain them. Formal issues include missing event arguments, XML tags, and each event argument only appears once in the text generated by the large language model in Steps 3 or 4. Step 5.2: Construct a prompt word D and input it into the large language model. The large language model is asked to find the most suitable and meaningful verb that best represents the event type trigger word in the filtered generated data. If it exists, the selected trigger word is XML-tagged in the generated data. Otherwise, all event types and their sub-event types of the original data to be synthesized are input into the large language model, and it is asked to generate a group of meaningful verbs suitable for the corresponding event types as event trigger words. Then, this group of event trigger words and the filtered generated data are input into the large language model, and the large language model is asked to select a meaningful verb that represents the event type trigger word from the group of event trigger words as the trigger word, and insert the meaningful verb into the appropriate position in the generated data in the correct tense. The prompt word D includes the identity information of the large language model, the requirements for marking trigger words, and the results obtained after filtering in step 5.

1. When reconstructing with different event arguments as the main body in step 6, the reconstruction methods adopted include out-of-order description and description according to time or event development.

Images