Battle document generation method and system based on large language model

Abstract

The invention belongs to the technical field of combat document generation, and particularly relates to a combat document generation method and system based on a large language model, which is used for generating a supervised fine-tuning text of a large model, and comprises the following steps: randomly selecting a combat scene and description information thereof from a pre-established combat scene library, and generating a fine-tuning text based on the selected combat scene and description information; generating known key information; based on the known key information, the provided cue words are used, and key information conforming to logic is generated by utilizing a large language model; verifying the generated information to determine whether the information accords with the logic of the whole scene; inputting the key information passing the verification into a preset cue word template, and generating a complete combat document by using a large language model; the generated combat document is verified, and content consistency and language smoothness are checked; the generated key information is extracted and constructed into a character string in a JSON format; and outputting the combat document text and the corresponding JSON format character string. According to the method, the generation speed of combat document training corpora and the reasonability of internal logic are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of combat document generation, and specifically relates to a combat document generation method and system based on a large language model. Background Art

[0002] As an important information carrier in the process of military command and decision-making, combat documents contain a large amount of key information about strategic deployment, force allocation, combat plans, etc. Accurately and efficiently extracting entity information from such documents automatically is of great significance to improving battlefield situation awareness and accelerating command decision-making.

[0003] The current automatic generation scheme for combat documents includes a template-driven generation method, which presets the document structure as a template, in which the static components are constant elements and the expected changing parts are variable factors. In the process of generating combat documents, staff officers need to fill in the reserved blanks according to the actual situation. This method can significantly reduce the workload, but it has obvious defects. Excessive structural rigidity limits the flexibility of the text, lacks adaptability to diverse scenarios, and may lead to loss of information integrity.

[0004] Another strategy for generating operational documents adopts the Schema method. This method is based on the principle of rhetorical predicates in linguistics, and uses rhetorical predicates and several operational symbols to characterize the logical pattern of text construction. For example, attribute predicates are used to detail the specific characteristics of entities or time, analogy predicates are intended to build comparative relationships in the text, and composition predicates describe the constituent elements of an object. The advantage of Schema generation technology is that it is easy to maintain and upgrade, and it can generate high-quality documents. However, this technology is limited to text styles with fixed structures.

[0005] In addition, the above-mentioned combat document generation technology has the following shortcomings: 1. Over-reliance on manual labor. In the field of automated content creation, although there are a variety of technical means, such as template-based generation technology and Schema-based generation technology, each of which has shown certain effectiveness and flexibility in specific scenarios, these methods have not been able to completely get rid of the high dependence on manual intervention. Specifically, whether it is filling content with a pre-set template or automatically generating information based on a predefined framework (Schema), it is essentially necessary for the creator to carefully design the template structure, write basic content or detailed specifications, and this process undoubtedly consumes considerable human resources. When faced with large-scale content generation needs, this method of content injection that relies on manual labor will undoubtedly significantly increase the time cost and economic cost of the project, limiting the efficiency and scale of content production. The present invention completely abandons the reliance on manual intervention and realizes the full automation from content conception to writing. This means that unlike the past, which required a large amount of manpower to manually fill in templates or adjust content, the new method significantly reduces labor costs, releases valuable intellectual resources, and enables military commanders and planners to focus on higher-level strategic decision-making and tactical analysis rather than cumbersome paperwork.

[0006] 2. The generated text lacks diversity. Human thinking has certain limitations. Even the most creative individuals find it difficult to maintain the diversity and freshness of their thinking when they are engaged in long-term and intensive document content creation. This means that if a single creator or a team of a few people is responsible for generating a large amount of text content, it is likely to result in a single text style, repeated views, and a lack of necessary perspective changes and innovative elements. Over time, this homogeneous content caused by the inertia of individual thinking will be difficult to meet the needs of information richness, diversity, and personalization. The large language model ensures the high quality and diversity of the generated text with its excellent contextual understanding and generation capabilities. It can not only automatically generate coherent, logical, and information-rich combat documents based on the specific requirements and background information of the task, but also incorporate a variety of expressions and strategic thinking in different scenario settings, avoiding the problem of text monotony caused by personal mindset. This ability enables each combat document to accurately reflect the complexity of the battlefield while maintaining a high degree of innovation and adaptability.

[0007] 3. Document format is limited. Text content generated based on fixed templates or schemas is often limited to a preset structural framework, which to a certain extent restricts the natural fluency and flexibility of the text. Summary of the invention

[0008] Purpose of the invention: The purpose of the present invention is to address the deficiencies of the prior art and provide a method and system for generating combat documents based on a large language model, making full use of the learning and adaptation advantages of the large model so that it can flexibly adjust the structure and style of the generated text according to the different sample cases or specific guidelines provided. Whether it is a concise instruction in an emergency or a detailed planning report for a complex campaign, the system can respond quickly and generate combat documents that are highly matched to the needs. This flexibility improves the practicality and pertinence of the documents.

[0009] Technical solution: The combat document generation method based on a large language model of the present invention comprises:

[0010] Step 1: randomly select a combat scenario and its description information from a pre-established combat scenario library, and generate known key information based on the selected combat scenario and description information;

[0011] Step 2: Input the known key information obtained in step 1 into the preset prompt word template Prompt to guide the large language model LLM to generate the key information to be tested;

[0012] Step 3: Input the key information to be checked generated in step 2 into the preset prompt word template check_message_Prompt, and guide the large language model LLM to perform logic verification. The verification process is as follows:

[0013]

[0014] When the verification result is True, the information m is retained and the next information is generated. When the verification result is False, the process returns to step 2 and regenerates the information until a complete valid information set Full_M = {m i};

[0015] Step 4: Input the valid information set Full_M obtained in step 3 into the preset prompt word template generate_text_Prompt, and use the large language model to generate a complete combat document text. The process is: OutputText = LLM (generate_text_Prompt (M));

[0016] Step 5: Input the combat document text generated in step 4 into the preset prompt word template check_text_Prompt, and guide the large prediction model LLM to perform content verification. The verification process is as follows:

[0017]

[0018] Among them, when the verification result is True, the text OutputText is retained, and when the verification result is False, it returns to step 4 to regenerate;

[0019] Step 6: Extract the valid information set generated in step 3 and convert it into a string in JSON format according to the predefined logical relationship and hierarchical relationship, Json = Transform (M);

[0020] Step 7. Output the combat document text OutputText obtained in step 5 and the corresponding JSON format string as a key-value pair, Output = {OutputText, Json}.

[0021] To further improve the above technical solution, the combat scenario library adopts S={s i , d i}, where S is the combat scenario set, s i For specific combat scenarios, i is the description information of the specific combat scene; through m=RandomChoice(S), the selected combat scene s and the description information d of the scene are obtained as the known key information m.

[0022] Furthermore, the process of generating the key information to be checked by using the large language model LLM in step 2 includes:

[0023] Get the key information provided in step 1 M = {m i}, M is the set of known key information, m i It is each specific key information;

[0024] The key information set M = {m i} Input the preset prompt word template Prompt, which includes the following contents: definition rules and output format requirements for key military information, information generation constraints based on combat scenarios, military terminology specifications and usage requirements, and restrictions on prohibiting the output of irrelevant auxiliary descriptions;

[0025] The large language model LLM is used to generate the key information to be checked based on the prompt word template Prompt, OutputMessage=LLM(Prompt(M).

[0026] Furthermore, step three includes: inputting the key information to be checked generated in step two and the known key information obtained in step one into a preset prompt word template check_message_Prompt, guiding the large language model LLM to perform logical verification; the prompt word template check_message_Prompt includes the following contents: logical relationship verification rules between military information, quantitative relationship restriction conditions, military information compliance inspection rules, tactical feasibility assessment standards, and resource coordination and rational judgment criteria.

[0027] Furthermore, the valid information set Full_M obtained in step three is input into the preset prompt word template generate_text_Prompt, and the prompt word template generate_text_Prompt includes the following contents: standardized format specifications of combat documents, rules for the use of military professional terms, document chapter structure and level requirements, organization and connection rules of different types of information, and document language expression specifications.

[0028] Furthermore, the document text generated in step 4 is input into the preset prompt word template check_text_Prompt for content verification, and the prompt word template check_text_Prompt includes the following contents: consistency check rules between the document content and the original information, military logic rigor review standards, document language fluency assessment criteria, and professional terminology use accuracy verification rules.

[0029] Furthermore, the scenarios in the combat scenario library s include: land combat scenarios, sea combat scenarios, air combat scenarios, network combat scenarios and space combat scenarios; the combat scenario description d includes: geographical conditions, climate conditions, enemy and friendly situations, and battlefield environment.

[0030] Furthermore, the JSON string includes: combat scenario information, enemy and friendly situation information, tactical deployment information, and resource allocation information.

[0031] The system for implementing the above-mentioned combat document generation method based on the large language model includes: a scenario library module, a prompt word template module, a large language model processing module and a text generation module;

[0032] The scenario library module stores a pre-established combat scenario library S={s i , d i}, where S is the combat scenario set, s i For specific combat scenarios, i is the description information of the specific combat scenario, and obtains the specific combat scenario s and the scenario description d through m=RandomChoice(S);

[0033] The prompt word template module includes:

[0034] The information generation prompt word unit is used to store and call the prompt template, which contains the definition rules of key military information, output format requirements, information generation constraints and military terminology specifications;

[0035] The information verification prompt unit is used to store and call the check_message_Prompt template, which contains the logic relationship verification rules, quantity relationship restriction conditions and tactical feasibility evaluation criteria between military information;

[0036] The document generation prompt unit is used to store and call the generate_text_Prompt template, which contains the standardized format specifications of combat documents, the rules for the use of military professional terms, and the document structure requirements;

[0037] The document verification prompt unit is used to store and call the check_text_Prompt template, which contains the document content verification rules and quality assessment standards;

[0038] The large language model processing module is used to process information according to the prompt word template module, including:

[0039] According to the prompt template, the key information set M = {m i} and generate the information to be checked OutputMessage = LLM (Prompt (M);

[0040] Get the generated key information to be checked and key information set M = {m i}, according to the check_message_Prompt template, perform the following verification process:

[0041]

[0042] The valid information that has passed the verification is combined into a set Full_M = {m i};

[0043] Get the valid information set Full_M, and generate the combat document text OutputText = LLM (generate_text_Prompt (M)) according to the generate_text_Promptt template;

[0044] Get the combat document text and perform the following combat document text verification process according to the check_text_Prompt template:

[0045]

[0046] The text generation module integrates the verified valid information set Full_M and converts it into the standard JSON format, outputs the combat document text OutputText and the corresponding JSON format string as a key-value pair, and generates the final combat document text Output = {OutputText, Json}.

[0047] Beneficial effect: Compared with the prior art, the advantages of the present invention are: the present invention introduces the capability of a large language model into the generation of combat documents, and guides the large prediction model to output by designing different prompt word templates, including the prompt template: focusing on the generation rules and constraints of key information, the check_message_Prompt template: focusing on the logical verification and compliance check of information, the generate_text_Prompt template: standardizing the generation format and structure of documents, and the check_text_Prompt template: ensuring the quality and accuracy of the document content. There is a progressive relationship between the various prompt word templates. First, by inputting a combat scenario and the corresponding prompt words, the key information of the next part is output, and the reflection mechanism is introduced to continuously improve the quality of the output information; and the obtained information is applied to the acquisition of the next part of information; after completing the generation of all information, the information is output as a fluent text through the large model with a reflection mechanism, and a standard format output is obtained from it. The combat documents obtained by this method significantly improve the reliability of the documents, enhance the accuracy and interpretability of causal reasoning, greatly reduce the investment of human resources, optimize resource utilization and speed up the generation speed, and improve the adaptability and generalization ability of cross-type combat documents. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of the combat document generation method of the large language model provided by the present invention;

[0049] Figure 2 It is a flowchart of the specific application of the combat document generation method of the large language model provided by the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.

[0051] Example 1: Figure 1 The combat document generation method based on the large language model shown in the figure comprises the following steps:

[0052] 1. Randomly select a combat scenario from the pre-established combat scenario library. Based on widely public military knowledge and actual combat cases, carefully summarize and extract a diverse combat scenario library, S = {s i , d i}, S represents the combat scenario set, s i Indicates a specific combat scenario, d i Represents the description of a specific combat scenario. This scenario library is a highly abstract and generalized summary of historical battles, and also integrates the development of modern war theories and tactics, covering multiple dimensions from traditional land, sea and air operations to new network and space warfare. Every time the system is ready to generate a new combat document, it will randomly select a scenario from the combat scenario library as the basic framework. A specific combat scenario s and a description of the scenario d obtained by m = RandomChoice(S) become known key information m after acquisition. RandomChoice is an existing function, that is, a method of randomly selecting a scenario from an existing scenario library list. This process is designed to simulate the uncertainty of the real world, ensure that each document can cope with challenges in different strategic environments, and improve the diversity of combat documents.

[0053] 2. According to the information obtained in step 1, M = {m i}, M is the set of known key information, m i That is, each specific key information. The key information obtained in step one will be systematically integrated into a pre-designed prompt word template Prompt, guiding the large language model to generate logical key information based on the existing key information. The prompt word template Prompt not only defines the display structure of the information, but also clarifies the display format of the content, ensuring that all information data can be organized in an easy-to-understand way. The large model LLM inputs the prompt word to generate the key information to be tested, OutputMessage = LLM (Prompt (M)). The logical relationship before and after obtaining the key information has been repeatedly verified and confirmed. Figure 2 The text generated by this logic chain can shorten the text generation time as much as possible while ensuring the high logic of the text.

[0054] 3. Use the big language model to reflect on the information generated in step 2 to ensure that the newly generated information is logically coherent and self-consistent with the provided information, and seamlessly connected to the overall combat context. The generated information is embedded in the customized prompt word template check_message_Prompt, which clarifies the logical relationship, quantitative relationship and other restrictions that must be met between each key information, and guides the big model to judge the input key information. The integrated prompt words are input into the big language model. The model has a high level of contextual understanding and reasoning ability, which can evaluate the logical consistency and rationality of the information and the relevance to the initial input combat scenario data (combat scenario s and description of the scenario d). For example, the model can check whether the description of force deployment is consistent with the known military force comparison, whether the tactical application information takes into account the actual limitations of the geographical environment, and whether the resource allocation is reasonable and feasible. If it is confirmed that the logic of the newly generated information is correct and perfectly integrated with all previous information, it will enter the next step of information generation. If the information generation is completed, it will enter the combat document language generation stage. If the large language model detects logical inconsistencies or contradictions, such as confusing timelines, unreasonable resource allocation, or tactical choices that go against strategic goals, it needs to regenerate information. This process repeats until all the information that needs to be generated can withstand logical scrutiny, forming a logical closed loop.

[0055]

[0056] The valid information set is continuously added in each loop, and finally all the valid information sets are Full_M = {m i}.

[0057] 4. Input the acquired valid information set into the prompt word template and generate a complete combat document. Through the above process, a series of complete combat document information is obtained. This type of information is designed in multiple dimensions such as strategic deployment, enemy and friendly situation, environmental factors, etc. These key information are embedded in the customized prompt word template generate_text_Prompt. As a transformation link from multiple single information to complete combat documents, the prompt word template needs to carefully set the information framework to ensure that all kinds of information can be presented in an orderly and clear manner, and set strict format specifications and language styles, aiming to make the final generated document both professional and rigorous and easy to read quickly. Input the filled template into the large prediction model to obtain a coherent, complete and tactically guiding combat document OutputText = LLM (generate_text_Prompt (M)).

[0058] 5. Reflect on some of the key documents obtained to check whether the content is consistent with the content provided, whether the sentences are fluent, etc. After completing the generation of preliminary operational documents, it is necessary to reflect on and review the documents obtained. The generated information is embedded in the customized prompt word template check_text_Prompt. The prompt word template clarifies the text logic, sequence and other requirements, and guides the large model to judge the key information input. The large language model needs to verify whether the real-time description and data reference in the document are consistent based on the key information provided. This process emphasizes the accuracy and consistency of the information, and ensures the authenticity and reliability of the content of the document. Secondly, it is necessary to check the language expression and logical coherence of the document to determine whether it is clear, the sentences are fluent, and the logic is rigorous.

[0059]

[0060] 6. Extract all the key information of the document obtained in step 4, and construct a JSON string according to the logical relationship and hierarchical relationship of each data as the standard output of document entity information extraction, Json = Transform (M).

[0061] 7. The output content includes the combat document text and the corresponding JSON string:

[0062] Output = {OutputText, Json}.

[0063] Example 2: Figure 2 As shown, the method provided in Example 1 is used for specific application, including:

[0064] 1. Randomly select a combat scenario from the pre-established combat scenario library. Randomly select any combat scenario from the combat scenario library summarized and improved in the public data, such as a hilly combat scenario. The description of the combat scenario obtained at the same time will provide detailed information on the geographical conditions, climate conditions, etc. in the combat scenario, such as "the local average temperature is 10 degrees Celsius, and the mountain altitude is 200 meters". These descriptions will be included in the prompt words to provide reference for the generation of some subsequent key information.

[0065] 2. Based on the information provided in the previous step, use the large language model to generate logical key information. For example, the previous step provides information such as combat scenarios and descriptions. In the prompt word, put this type of information into the preset prompt word template Prompt, which clearly indicates the definition of the next key information to be generated, such as generating the enemy's military number. You need to define the generation rules of the enemy's military number. And tell the large language model that the output cannot be fabricated, only key information needs to be output, and other auxiliary languages ​​cannot be output.

[0066] 3. Further reflect on the generated information using the large language model. Put the information obtained in the previous step into the large model together with the information obtained before, and embed the generated information into the customized prompt word template check_message_Prompt. The prompt word template clarifies the logical relationship, quantitative relationship and other restrictions that must be met between each key information, and guides the large model to judge the input key information. For example, whether the military branch represented by the enemy's military number is suitable for activities in this scenario, whether the naming of the enemy's military number is in compliance with the specifications, etc. If the model output result is correct, it will enter the next step of key information generation. If an error occurs, it is necessary to re-run the previous step.

[0067] 4. After obtaining all the key information through the above steps, all the information can be written into a JSON string with a fixed structure as the standard output structure for information extraction. For example, the enemy number "*** Brigade" obtained in the above information needs to be put into {emermyName:*** Brigade}.

[0068] 5. Put all the above information into the pre-set prompt word template check_text_Prompt. All kinds of information need to be put in according to a certain structure so that the large language model can better understand the information and make certain requirements, such as: the output text must be fluent, the content of the output article must be generated completely according to the given information, and content that does not match the sample content is not allowed.

[0069]

[0070] 6. Input the generated text into the reflection template. Some questions are set in the template, such as whether the text is fluent, whether the text is consistent with the provided information, etc. If it meets the requirements, construct a JSON string as the standard output of document entity information extraction, Json = Transform (M). If it does not meet the requirements, regenerate it.

[0071] 7. Output the obtained combat documents and key information JSON strings as key-value pairs.

[0072] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for generating combat documents based on a large language model, characterized in that: include: Step 1: randomly select a combat scenario and its description information from a pre-established combat scenario library, and generate known key information based on the selected combat scenario and description information; Step 2: Input the known key information obtained in step 1 into the preset prompt word template Prompt to guide the large language model LLM to generate the key information to be tested; Step 3: Input the key information to be checked generated in step 2 into the preset prompt word template check_message_Prompt, and guide the large language model LLM to perform logic verification. The verification process is as follows: When the verification result is True, the information m is retained and the next information is generated. When the verification result is False, the process returns to step 2 and regenerates the information until a complete valid information set Full_M = {m i }; Step 4: Input the valid information set Full_M obtained in step 3 into the preset prompt word template generate_text_Prompt, and use the large language model to generate a complete combat document text. The process is: OutputText = LLM (generate_text_Prompt (M)); Step 5: Input the combat document text generated in step 4 into the preset prompt word template check_text_Prompt, and guide the large prediction model LLM to perform content verification. The verification process is as follows: Among them, when the verification result is True, the text OutputText is retained, and when the verification result is False, it returns to step 4 to regenerate; Step 6: Extract the valid information set generated in step 3 and convert it into a string in JSON format according to the predefined logical relationship and hierarchical relationship, Json = Transform (M); Step 7. Output the combat document text OutputText obtained in step 5 and the corresponding JSON format string as a key-value pair, Output = {OutputText, Json}.

2. The method for generating combat documents based on a large language model according to claim 1, characterized in that: The combat scenario library adopts S={s i , d i }, where S is the combat scenario set, s i For specific combat scenarios, i is the description information of the specific combat scene; through m=RandomChoice(S), the selected combat scene s and the description information d of the scene are obtained as the known key information m.

3. The method for generating combat documents based on a large language model according to claim 2, characterized in that: The process of using the large language model LLM to generate key information to be tested in step 2 includes: Get the key information provided in step 1 M = {m i }, M is the set of known key information, m i It is each specific key information; The key information set M = {m i } Input the preset prompt word template Prompt, which includes the following contents: definition rules and output format requirements for key military information, information generation constraints based on combat scenarios, military terminology specifications and usage requirements, and restrictions on prohibiting the output of irrelevant auxiliary descriptions; The large language model LLM is used to generate the key information to be checked based on the prompt word template Prompt, OutputMessage=LLM(Prompt(M).

4. The method for generating combat documents based on a large language model according to claim 3 is characterized in that: Step three includes: inputting the key information to be checked generated in step two and the known key information obtained in step one into the preset prompt word template check_message_Prompt, guiding the large language model LLM to perform logical verification; the prompt word template check_message_Prompt includes the following contents: logical relationship verification rules between military information, quantitative relationship restriction conditions, military information compliance inspection rules, tactical feasibility assessment standards, and resource coordination and rational judgment criteria.

5. The method for generating combat documents based on a large language model according to claim 4 is characterized in that: The valid information set Full_M obtained in step three is input into the preset prompt word template generate_text_Prompt, which contains the following contents: standardized format specifications of combat documents, rules for the use of military professional terms, document chapter structure and hierarchical requirements, organization and connection rules of different types of information, and document language expression specifications.

6. The method for generating combat documents based on a large language model according to claim 5 is characterized in that: The document text generated in step 4 is input into the preset prompt word template check_text_Prompt for content verification. The prompt word template check_text_Prompt includes the following contents: consistency check rules between document content and original information, military logic rigor review standards, document language fluency assessment criteria, and professional terminology accuracy verification rules.

7. The method for generating combat documents based on a large language model according to claim 1, characterized in that: The scenarios in the combat scenario library s include: land combat scenarios, sea combat scenarios, air combat scenarios, network combat scenarios and space combat scenarios; the combat scenario description d includes: geographical conditions, climate conditions, enemy and friendly situations, and battlefield environment.

8. The method for generating combat documents based on a large language model according to claim 1, characterized in that: The JSON string includes: combat scenario information, enemy and friendly situation information, tactical deployment information, and resource allocation information.

9. A combat document generation system based on a large language model, used to implement the combat document generation method based on a large language model as claimed in claim 1, characterized in that: include: Scenario library module, prompt word template module, large language model processing module and text generation module; The scenario library module stores a pre-established combat scenario library S={s i , d i }, where S is the combat scenario set, s i For specific combat scenarios, i is the description information of the specific combat scenario, and obtains the specific combat scenario s and the scenario description d through m=RandomChoice(S); The prompt word template module includes: The information generation prompt word unit is used to store and call the prompt template, which contains the definition rules of key military information, output format requirements, information generation constraints and military terminology specifications; The information verification prompt unit is used to store and call the check_message_Prompt template, which contains the logic relationship verification rules, quantity relationship restriction conditions and tactical feasibility evaluation criteria between military information; The document generation prompt unit is used to store and call the generate_text_Prompt template, which contains the standardized format specifications of combat documents, the rules for the use of military professional terms, and the document structure requirements; The document verification prompt unit is used to store and call the check_text_Prompt template, which contains the document content verification rules and quality assessment standards; The large language model processing module is used to process information according to the prompt word template module, including: According to the prompt template, the key information set M = {m i } and generate the information to be checked OutputMessage = LLM (Prompt (M); Get the generated key information to be checked and key information set M = {m i }, according to the check_message_Prompt template, perform the following verification process: The valid information that has passed the verification is combined into a set Full_M = {m i }; Get the valid information set Full_M, and generate the combat document text OutputText = LLM (generate_text_Prompt (M)) according to the generate_text_Promptt template; Get the combat document text and perform the following combat document text verification process according to the check_text_Prompt template: The text generation module integrates the verified valid information set Full_M and converts it into the standard JSON format, outputs the combat document text OutputText and the corresponding JSON format string as a key-value pair, and generates the final combat document text Output = {OutputText, Json}.

Images