Text generation method and device based on confrontation mechanism, equipment and storage medium

Abstract

The invention discloses a text generation method and device based on an adversarial mechanism, equipment and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the fine tuning of an initial generator model, processing a to-be-processed prompt word generated based on a historical summary dialogue text pair through an obtained to-be-trained generator model, and obtaining an abstract generation result; scoring the abstract generation result by using the to-be-trained scoring model, and performing gradient updating on the to-be-trained generator model by using a target optimization function determined based on the obtained target scoring result to obtain a target generator model; performing comparative learning on the abstract generation result and a historical summary dialogue text, and adjusting a to-be-trained scoring model based on an obtained comparison result to obtain a target scoring model; and processing the to-be-recognized text by utilizing the obtained target scoring model and the target generator model, and if an obtained text processing result meets a preset processing stop condition, setting the text processing result as a target abstract text. Therefore, the text generation efficiency can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly relates to a text generation method, device, equipment and storage medium based on an adversarial mechanism. Background Art

[0002] Currently, current summary systems (systems for summarizing medical consultations, meetings, conversations, customer service conversations, etc.) mainly rely on large language models to generate summaries through prompt words. However, such systems face many challenges in practical applications: they often have difficulty recognizing the logical coherence of conversation content and have limited understanding of specific contexts; when dealing with professional fields, they are unable to accurately extract core decision points and professional terms; at the same time, these systems lack a mechanism to effectively calibrate the generated content, resulting in a significant gap between the generated content and high-quality summaries written by humans, and the quality of the generated summaries is far inferior to that of human-written ones.

[0003] As can be seen from the above, how to improve the efficiency of text generation in the process of text generation based on an adversarial mechanism is an urgent problem to be solved at present. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a text generation method, device, equipment and storage medium based on an adversarial mechanism, which can improve the efficiency of text generation in the process of text generation based on an adversarial mechanism, and further improve the speed of the production process. The specific solutions are as follows:

[0005] In the first aspect, the present application provides a text generation method based on an adversarial mechanism, including:

[0006] Performing domain knowledge fine-tuning operation on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generating a prompt word to be processed based on the corresponding conversation content features and domain attributes of the historical minutes conversation text, and then using the generator model to be trained to process the prompt word to be processed to obtain a summary generation result;

[0007] Using the generator model to be trained and based on a preset multi-dimensional scoring rule to perform dimensional quality scoring on the summary generation result to obtain a target scoring result, and determining a target optimization function based on the target scoring result, so as to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain a target generator model;

[0008] Using a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text of the historical minutes conversation text to obtain a comparison result, and adjusting the model parameters of the generator model to be trained based on the comparison result to obtain a target generator model;

[0009] Process the text to be recognized using the target generator model and the target scoring model to obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target summary text.

[0010] Optionally, the initial generator model is fine-tuned with domain knowledge based on a preset domain corpus to obtain a generator model to be trained, and a prompt to be processed is generated based on the corresponding dialogue content features and domain attributes of the historical summary dialogue text pair, including:

[0011] Construct a preset domain corpus containing specific domain professional terms and concepts, and fine-tune and train the initial generator model based on a preset gradient accumulation rule and a preset low learning rate strategy to obtain a generator model to be trained; the initial generator model is a large language model;

[0012] Collect historical summary dialogue text pairs, and use a preset chain of thought to insert a triple training structure including dialogue text, reasoning process, and final summary into the historical summary dialogue text pairs. Then, based on the triple training structure and the historical summary dialogue text pairs, determine the corresponding dialogue content features, domain attributes, and historical experience of the historical summary dialogue text pairs, so as to determine the prompt to be processed based on the dialogue content features, the domain attributes, and the historical experience.

[0013] Optionally, the generator model to be trained is used to process the prompt to be processed to obtain a summary generation result, including:

[0014] Use the generator model to be trained and process the prompt to be processed based on a temperature parameter index; the temperature parameter index is an index determined based on the creativity and accuracy of the prompt to be processed;

[0015] Use the generator model to be trained and perform structured prompt marking on the first prompt to be processed based on a structural integrity index; the structural integrity index includes a background information index, a main idea index, a decision conclusion index, and an action item index;

[0016] Use the generator model to be trained and perform professional term interpretation on the second prompt to be processed based on a professional depth index; the professional depth index is the usage density information and interpretation depth information set based on the target audience;

[0017] Using the to-be-trained generator model and based on the abstract length metric, perform a professional term explanation on the third to-be-processed prompt word to obtain an abstract generation result; the abstract length metric is a metric determined based on the information density and length balance of the to-be-processed prompt word.

[0018] Optionally, the using the to-be-trained scoring model and based on a preset multi-dimensional scoring rule to perform a dimensional quality scoring on the abstract generation result to obtain a target scoring result includes:

[0019] Using the to-be-trained scoring model to perform an information integrity scoring on the abstract generation result to obtain an information integrity scoring result; the numerical size of the information integrity scoring result is positively correlated with the size of the core content coverage rate in the abstract generation result;

[0020] Using the to-be-trained scoring model to perform a logical structure scoring on the abstract generation result to obtain a logical structure scoring result; the logical structure scoring result is used to characterize the content organization and coherence in the abstract generation result;

[0021] Using the to-be-trained scoring model to determine whether there are features of artificial writing in the abstract generation result to obtain a authenticity scoring result;

[0022] Using the to-be-trained scoring model and based on the normativity and accuracy of specific domain terms to process the abstract generation result to obtain a professional term accuracy rate;

[0023] Using the to-be-trained scoring model to perform a structural integrity scoring on the abstract generation result to obtain a structural integrity scoring result;

[0024] Based on the information integrity scoring result, the logical structure scoring result, the authenticity scoring result, the professional term accuracy rate, and the structural integrity scoring result, determine the target scoring result.

[0025] Optionally, the using a preset contrastive learning algorithm to compare and learn the abstract generation result with the corresponding minutes text in the historical minutes dialogue text pair to obtain a comparison result, and based on the comparison result, adjust the model parameters of the to-be-trained scoring model to obtain a target scoring model includes:

[0026] Compare the abstract generation result with the corresponding minutes text in the historical minutes dialogue text pair to obtain a text comparison result, and then based on the text comparison result, the abstract generation result, and the minutes text, determine a difference report;

[0027] Determine a comparison result by using a preset contrastive learning algorithm and based on the difference report, and use the comparison result to adjust the model parameters of the to-be-trained scoring model to obtain a to-be-processed scoring model; the model parameters are parameters for distinguishing the discrimination ability of the abstract generation result and the summary text.

[0028] Adjust the to-be-processed scoring model by using a preset dynamic learning rate adjustment mechanism to obtain a target scoring model.

[0029] Optionally, the processing of the to-be-identified text by using the target generator model and the target scoring model to obtain a text processing result includes:

[0030] Process the to-be-identified text by using preset data preprocessing rules to obtain a standardized text, and process the standardized text by using a preset prompt word generation module to obtain a target prompt word.

[0031] Perform abstract generation processing on the target prompt word by using the target generator model to obtain a to-be-processed abstract generation result, and score the to-be-processed abstract generation result by using the target scoring model to obtain a target scoring result.

[0032] Determine a feature prompt template based on the dialogue features corresponding to the to-be-identified text and historical optimization experience, and then determine a to-be-processed prompt word template library based on the feature prompt template, a preset task instruction, and a preset example summary.

[0033] Adjust the to-be-processed prompt word template library based on the target scoring result to obtain a target prompt word template library, and use the target prompt word template library to adjust the to-be-processed abstract generation result to obtain a target abstract generation result.

[0034] Optionally, the judging whether the text processing result meets a preset processing stop condition, and if so, setting the text processing result as the target abstract text includes:

[0035] Judge whether the scoring fluctuation corresponding to the text processing result is less than a preset scoring fluctuation threshold. If the scoring fluctuation corresponding to the text processing result is not less than the preset scoring fluctuation threshold, jump to the step of processing the to-be-identified text by using the target generator model and the target scoring model to obtain a text processing result until the scoring fluctuation corresponding to the text processing result is less than the preset scoring fluctuation threshold, and set the text processing result as the target abstract text.

[0036] Alternatively, determine whether the current iteration count corresponding to the text processing result is less than a preset iteration count. If the current iteration count corresponding to the text processing result is less than the preset iteration count, jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, until the current iteration count corresponding to the text processing result is not less than the preset iteration count, and set the text processing result as the target summary text.

[0037] In a second aspect, the present application provides a text generation device based on an adversarial mechanism, including:

[0038] A summary generation result determination module, configured to perform domain knowledge fine-tuning operations on an initial generator model based on a preset domain corpus to obtain a generator model to be trained, generate a prompt to be processed for corresponding dialogue content features and domain attributes based on a historical minutes dialogue text, and then use the generator model to be trained to process the prompt to be processed to obtain a summary generation result;

[0039] A generator model determination module, configured to use a scoring model to be trained and perform dimensional quality scoring on the summary generation result based on a preset multi-dimensional scoring rule to obtain a target scoring result, and determine a target optimization function based on the target scoring result, so as to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain a target generator model;

[0040] A scoring model determination module, configured to use a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text in the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the scoring model to be trained based on the comparison result to obtain a target scoring model;

[0041] A summary text determination module, configured to use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target summary text.

[0042] In a third aspect, the present application provides an electronic device, including:

[0043] A memory, configured to store a computer program;

[0044] A processor, configured to execute the computer program to implement the foregoing text generation method based on an adversarial mechanism.

[0045] Fourthly, the present application provides a computer-readable storage medium for storing a computer program, wherein when the computer program is executed by a processor, the foregoing text generation method based on an adversarial mechanism is implemented.

[0046] As can be seen from the above, before performing text generation based on an adversarial mechanism in the present application, it is necessary to perform domain knowledge fine-tuning on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generate a prompt word to be processed based on the corresponding dialogue content features and domain attributes of the historical minutes dialogue text. Then, use the generator model to be trained to process the prompt word to be processed to obtain a summary generation result; use the scoring model to be trained and perform dimensional quality scoring on the summary generation result based on a preset multi-dimensional scoring rule to obtain a target scoring result, and determine a target optimization function based on the target scoring result to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain a target generator model; use a preset contrast learning algorithm to compare and learn the summary generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the scoring model to be trained based on the comparison result to obtain a target scoring model; use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target summary text.

[0047] Thus, it can be seen that the present application first needs to perform domain knowledge fine-tuning on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generate a prompt word to be processed based on the corresponding dialogue content features and domain attributes of the historical minutes dialogue text. Then, use the generator model to be trained to process the prompt word to be processed to obtain a summary generation result; subsequently, use the scoring model to be trained and perform dimensional quality scoring on the summary generation result based on a preset multi-dimensional scoring rule to obtain a target scoring result, and determine a target optimization function based on the target scoring result to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain a target generator model; furthermore, use a preset contrast learning algorithm to compare and learn the summary generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the scoring model to be trained based on the comparison result to obtain a target scoring model; finally, use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target summary text. In this way, the efficiency of text generation is improved during the text generation process based on an adversarial mechanism, thereby enhancing the speed of the production process. Description of the Drawings

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0049] Figure 1 Flowchart of a text generation method based on an adversarial mechanism disclosed in the present application;

[0050] Figure 2 Schematic structural diagram of a text generation device based on an adversarial mechanism disclosed in the present application;

[0051] Figure 3 Structural diagram of an electronic device disclosed in the present application. Detailed implementation manners

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] Currently, the current summary system mainly relies on large language models to generate summaries through prompt words. However, such systems face many challenges in practical applications: they often have difficulty in identifying the logical coherence of the conversation content and have limited understanding of specific contexts; when dealing with professional fields, they are unable to accurately extract the core decision points and professional terms; at the same time, these systems lack a mechanism to effectively calibrate the generated content, resulting in an obvious gap compared with high-quality manually written minutes, and the quality of the generated summaries is far inferior to that of manually written ones. Therefore, the present application provides a text generation method based on an adversarial mechanism, which can improve the efficiency of text generation and thus enhance the speed of the production process.

[0054] See Figure 1 As shown, the embodiments of the present invention disclose a text generation method based on an adversarial mechanism, including:

[0055] Step S11: Perform domain knowledge fine-tuning operations on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, generate a prompt word to be processed based on the corresponding conversation content features and domain attributes of the historical minutes conversation text, and then use the generator model to be trained to process the prompt word to be processed to obtain a summary generation result.

[0056] In this embodiment, first, the construction operation of the domain corpus needs to be carried out. A professional summary corpus is constructed for a specific application domain (such as medical, financial, legal, etc.). Subsequently, the domain knowledge of the generator model is fine-tuned using the domain-specific data in the professional summary corpus to strengthen the model's understanding of professional terms and domain concepts. Among them, the fine-tuning technology adopts the strategies of gradient accumulation and low learning rate to prevent catastrophic forgetting. Subsequently, a prompt template suitable for the current dialogue scenario is automatically generated based on a preset optimization strategy. The above prompt module is used to maintain the prompt template library. Among them, the prompt template library contains domain-specific prompts, task instructions, and example summaries, and adjusts and optimizes the prompt generation strategy according to the feedback of the subsequent discriminator model and the reinforcement learning algorithm to form a closed-loop training system.

[0057] It is worth mentioning that in the embodiment of this application, the generator model and the scoring model are selected for adversarial training, and the process of adversarial training of the generator model and the scoring model is as follows: First, a large number of real summary-dialogue text pairs covering multiple application domains are collected as training data. Among them, in order to improve the quality of the training data, the embodiment of this application introduces a chain of thought to add intermediate reasoning steps to the training data to form a triple training data including "dialogue → reasoning steps → final summary". Subsequently, an innovative adversarial training framework is used to enable the generator model and the discriminator model to promote and optimize each other during the training process. Among them, the specific process is as follows:

[0058] First, the differential sampling generation operation needs to be carried out, that is, a set of prompt words to be processed is generated using the generator model based on different parameter configurations. , . Specifically, the domain knowledge of the initial generator model is fine-tuned based on a preset domain corpus to obtain a generator model to be trained, and the prompt words to be processed are generated based on the dialogue content features and domain attributes corresponding to the historical summary dialogue text pair, which may include: constructing a preset domain corpus containing specific domain professional terms and concepts to fine-tune the initial generator model based on the preset gradient accumulation rule and preset low learning rate strategy to obtain the generator model to be trained; the initial generator model is a large language model; collecting historical summary dialogue text pairs, and using the preset chain of thought to insert a triple training structure including dialogue text, reasoning process, and final summary into the historical summary dialogue text pair, and then determining the dialogue content features, domain attributes, and historical experience corresponding to the historical summary dialogue text pair based on the triple training structure and the historical summary dialogue text pair, so as to determine the prompt words to be processed based on the dialogue content features, domain attributes, and historical experience.

[0059] In a specific embodiment, the generator model is DeepSeek-R1-Distill-Qwen-32B (a distilled large language model with 32B (32 billion parameters) scale developed by DeepSeek) obtained by distilling the large language model DeepSeek R1 (i.e., DeepSeek Robot R1), and is fine-tuned with domain knowledge to receive preprocessed dialogue text and prompt words, and generate a summary under the control of multiple parameters. Among them, the multiple parameters include but are not limited to temperature parameter, structural integrity, professional depth, and summary length. Among them, the temperature parameter is adjusted to dynamically set the generation temperature value (within the range of 0.1-0.9), to find the best balance between creativity and accuracy; the structural integrity is controlled by marking the structured prompt words to control the completeness of the generated content in different structural elements (background information, main ideas, decision conclusions, action items); the professional depth is controlled by setting the usage density and explanation depth of professional terms according to the target audience; the summary length is regulated by setting the target summary length range to ensure the balance between information density and length. That is, in the embodiment of the present application, 3-5 candidate versions of summaries with different parameter configurations will be generated in parallel for subsequent evaluation and screening operations.

[0060] Specifically, using the generator model to be trained to process the prompt words to be processed to obtain the summary generation result may include: using the generator model to be trained and processing the prompt words to be processed based on the temperature parameter index to obtain the first prompt word to be processed; the temperature parameter index is an index determined based on the creativity and accuracy of the prompt words to be processed; using the generator model to be trained and performing structured prompt word marking on the first prompt word to be processed based on the structural integrity index to obtain the second prompt word to be processed; the structural integrity index includes background information index, main idea index, decision conclusion index, and action item index; using the generator model to be trained and performing professional term explanation on the second prompt word to be processed based on the professional depth index to obtain the third prompt word to be processed; the professional depth index is the usage density information and explanation depth information set based on the target audience; using the generator model to be trained and performing professional term explanation on the third prompt word to be processed based on the summary length index to obtain the summary generation result; the summary length index is an index determined based on the balance between the information density and length of the prompt words to be processed.

[0061] Step S12: Use the scoring model to be trained and perform dimensional quality scoring on the summary generation result based on a preset multi-dimensional scoring rule to obtain the target scoring result, and determine the target optimization function based on the target scoring result, so as to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain the target generator model.

[0062] In this embodiment, after obtaining the abstract generation result, it is necessary to use a scoring model to evaluate the quality of the abstract generation result, so as to determine whether it is manually written or machine-generated, and provide detailed differential feedback information. It is worth mentioning that both the scoring model and the generator model are DeepSeek-R1-Distill-Qwen-32B after being distilled by DeepSeek R1, and have also undergone domain knowledge fine-tuning operations, and will be optimized and updated together with the generator model during the adversarial training process.

[0063] It is worth mentioning that when the scoring model scores the abstract generation result, it scores the abstract generation result based on multiple scoring dimensions, including but not limited to information integrity scoring, logical structure scoring, authenticity scoring, professional term accuracy, and structural integrity. Among them, the authenticity score is the authenticity score of the scoring model for each generated record, evaluating whether it has the characteristics of being manually written; the information integrity is to analyze the coverage ratio of the key information points of the original conversation in the generated record; the logical coherence is to evaluate the integrity of the internal logical structure of the record and the naturalness of the transition between paragraphs; the professional term accuracy is the standardization and accuracy of the use of specific domain terms; the structural integrity is to evaluate whether the record contains all the expected structural elements (such as background, decision, action items, etc.).

[0064] In a specific implementation manner, the scoring model comprehensively evaluates each sample to generate a score and a detailed differential report , and after obtaining the scoring result, the embodiment of the present application needs to construct an optimization objective function based on the scoring result to perform gradient update on the generator parameters using the objective function. In addition, the embodiment of the present application needs to dynamically adjust the prompt word template library according to the evaluation result output by the scoring model to optimize the distribution of control parameters.

[0065] In this embodiment, after obtaining the abstract generation result, the embodiment of the present application needs to continuously improve the discrimination ability of the scoring model through the comparative learning of the real minutes and the generated minutes, forming a benign competition with the generator model. It is worth mentioning that the optimization objective of the discriminator is to maximize the discrimination between the real minutes and the generated minutes, while providing more accurate difference feedback information. Specifically, the dimension quality score of the abstract generation result is obtained by using the scoring model to be trained and based on the preset multi-dimensional scoring rules. The target scoring result can include: using the scoring model to be trained to perform an information integrity score on the abstract generation result to obtain an information integrity score result; the numerical size of the information integrity score result is positively correlated with the coverage rate of the core content in the abstract generation result; using the scoring model to be trained to perform a logical structure score on the abstract generation result to obtain a logical structure score result; the logical structure score result is used to characterize the content organization and coherence in the abstract generation result; using the scoring model to be trained to determine whether there are characteristics of manual writing in the abstract generation result to obtain a authenticity score result; using the scoring model to be trained and based on the standardization and accuracy of specific domain terms to process the abstract generation result to obtain the accuracy of professional terms; using the scoring model to be trained to perform a structural integrity score on the abstract generation result to obtain a structural integrity score result; the target scoring result is determined based on the information integrity score result, the logical structure score result, the authenticity score result, the accuracy of professional terms, and the structural integrity score result.

[0066] In a specific implementation manner, the embodiment of the present application is based on GRPO (Group Relative Policy Optimization, that is, group relative policy optimization) and calculates the advantage value of each generated sample through intra-group relative comparison. At the same time, the scoring model also continuously optimizes its discrimination ability and difference feedback quality during this process, thus forming a dual-model co-evolution mechanism, without training an additional value model, thereby significantly reducing the computational cost. And the expression for the scoring model to perform quality scoring is as follows:

[0067] ;

[0068] Among them, is the final quality score, is the comprehensive reward score of the th generated minutes, and are respectively the reward mean and standard deviation of all minutes samples in the current batch, is a small constant for stable calculation.

[0069] It is worth mentioning that the scoring model is not only used to provide scoring and feedback functions, but also continuously improves the scoring accuracy through gradient updates. As the capabilities of the generator model improve, the scoring model also needs to correspondingly raise the discrimination criteria to ensure adversarial balance and form a virtuous cycle improvement mechanism. That is, the embodiments of the present application ensure an appropriate competitive tension during the training process by dynamically adjusting the learning rate ratio of the two models, thereby avoiding training imbalance caused by either party overly dominating.

[0070] Step S13: Use a preset contrastive learning algorithm to compare and learn the abstract generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the to-be-trained scoring model based on the comparison result to obtain a target scoring model.

[0071] In this embodiment, after using the target optimization function to perform gradient update on the parameters corresponding to the to-be-trained generator model to obtain the target generator model, the embodiments of the present application need to construct an information feedback loop between the scoring model and the generator to achieve a spiral improvement in the quality of minutes generation. Specifically, using a preset contrastive learning algorithm to compare and learn the abstract generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjusting the model parameters of the to-be-trained scoring model based on the comparison result to obtain a target scoring model may include: comparing the abstract generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a text comparison result, and then determining a difference report based on the text comparison result, the abstract generation result, and the minutes text; using a preset contrastive learning algorithm and based on the difference report to determine a comparison result, so as to use the comparison result to adjust the model parameters of the to-be-trained scoring model to obtain a to-be-processed scoring model; the model parameters are parameters for the discrimination ability to distinguish the abstract generation result and the minutes text; using a preset dynamic learning rate adjustment mechanism to adjust the to-be-processed scoring model to obtain a target scoring model.

[0072] Step S14: Use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target abstract text.

[0073] In this embodiment, after the generator model and the scoring model are adversarially trained to obtain the target generator model and the target scoring model, the embodiments of the present application can use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result. Specifically, using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result may include: processing the text to be recognized using a preset data preprocessing rule to obtain a standardized text, and processing the standardized text using a preset prompt word generation module to obtain a target prompt word; performing a summary generation process on the target prompt word using the target generator model to obtain a to-be-processed summary generation result, and scoring the to-be-processed summary generation result using the target scoring model to obtain a target scoring result; determining a feature prompt template based on the dialogue features corresponding to the text to be recognized and historical optimization experience, and then determining a to-be-processed prompt word template library based on the feature prompt template, a preset task instruction, and a preset example summary; adjusting the to-be-processed prompt word template library based on the target scoring result to obtain a target prompt word template library, so as to adjust the to-be-processed summary generation result using the target prompt word template library to obtain a target summary generation result.

[0074] It is worth mentioning that after the scoring model outputs a scoring result and a difference report is obtained based on the scoring result, the embodiments of the present application need to determine a prompt word generation strategy based on the scoring result and the difference report to regenerate the summary generation result using the prompt word generation strategy until a preset quality threshold is reached or a specified number of iterations is completed. Specifically, determining whether the text processing result meets a preset processing stop condition, and if so, setting the text processing result as the target summary text may include: determining whether the scoring fluctuation corresponding to the text processing result is less than a preset scoring fluctuation threshold. If the scoring fluctuation corresponding to the text processing result is not less than the preset scoring fluctuation threshold, then jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result until the scoring fluctuation corresponding to the text processing result is less than the preset scoring fluctuation threshold, and setting the text processing result as the target summary text; or, determining whether the current number of iterations corresponding to the text processing result is less than a preset number of iterations. If the current number of iterations corresponding to the text processing result is less than the preset number of iterations, then jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result until the current number of iterations corresponding to the text processing result is not less than the preset number of iterations, and setting the text processing result as the target summary text.

[0075] It is worth mentioning that the embodiments of the present application provide a system interaction interface for displaying the summary generation result and collecting user feedback information to optimize the model based on the user feedback information.

[0076] As can be seen from the above, before performing text generation based on the adversarial mechanism in the embodiments of the present application, it is first necessary to perform domain knowledge fine-tuning operations on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generate a to-be-processed prompt based on the corresponding dialogue content features and domain attributes of the historical minutes dialogue text. Then, use the generator model to be trained to process the to-be-processed prompt to obtain a summary generation result; subsequently, use the scoring model to be trained and based on a preset multi-dimensional scoring rule to perform dimensional quality scoring on the summary generation result to obtain a target scoring result, and determine a target optimization function based on the target scoring result to use the target optimization function to update the parameters corresponding to the generator model to be trained to obtain a target generator model; furthermore, use a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the scoring model to be trained based on the comparison result to obtain a target scoring model; finally, use the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determine whether the text processing result meets the preset processing stop condition. If it meets, set the text processing result as the target summary text. In this way, the efficiency of text generation is improved during the text generation process based on the adversarial mechanism, thereby enhancing the speed of the production process.

[0077] Correspondingly, referring to Figure 2 as shown, the present application further provides a text generation device based on an adversarial mechanism, including:

[0078] A summary generation result determination module 11, configured to perform domain knowledge fine-tuning operations on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generate a to-be-processed prompt based on the corresponding dialogue content features and domain attributes of the historical minutes dialogue text, and then use the generator model to be trained to process the to-be-processed prompt to obtain a summary generation result;

[0079] A generator model determination module 12, configured to use the scoring model to be trained and based on a preset multi-dimensional scoring rule to perform dimensional quality scoring on the summary generation result to obtain a target scoring result, and determine a target optimization function based on the target scoring result to use the target optimization function to update the parameters corresponding to the generator model to be trained to obtain a target generator model;

[0080] A scoring model determination module 13, configured to use a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a comparison result, and adjust the model parameters of the scoring model to be trained based on the comparison result to obtain a target scoring model;

[0081] The abstract text determination module 14 is configured to process the text to be recognized by using the target generator model and the target scoring model, obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets the condition, the text processing result is set as the target abstract text.

[0082] As can be seen from the above, before performing text generation based on the adversarial mechanism in the embodiments of the present application, it is first necessary to perform domain knowledge fine-tuning on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generate a prompt word to be processed based on the corresponding dialogue content features and domain attributes with the historical minutes dialogue text. Then, the generator model to be trained is used to process the prompt word to be processed to obtain an abstract generation result. Subsequently, the target scoring result is obtained by using the generator model to be trained and performing dimensional quality scoring on the abstract generation result based on a preset multi-dimensional scoring rule, and a target optimization function is determined based on the target scoring result to update the parameters corresponding to the generator model to be trained by using the target optimization function to obtain a target generator model. Furthermore, a comparison result is obtained by using a preset contrastive learning algorithm to compare and learn the abstract generation result with the corresponding minutes text in the historical minutes dialogue text, and the model parameters of the generator model to be trained are adjusted based on the comparison result to obtain a target scoring model. Finally, the text to be recognized is processed by using the target generator model and the target scoring model to obtain a text processing result, and it is determined whether the text processing result meets a preset processing stop condition. If it meets the condition, the text processing result is set as the target abstract text. In this way, the efficiency of text generation is improved in the process of text generation based on the adversarial mechanism, thereby enhancing the speed of the production process.

[0083] In some specific embodiments, the abstract generation result determination module 11 may specifically include:

[0084] The model fine-tuning unit is configured to construct a preset domain corpus containing specific domain professional terms and concepts, and perform fine-tuning training on the initial generator model based on a preset gradient accumulation rule and a preset low learning rate strategy to obtain a generator model to be trained; the initial generator model is a large language model;

[0085] The prompt word determination unit is configured to collect historical minutes dialogue text pairs, insert a triple training structure including dialogue text, reasoning process, and final minutes into the historical minutes dialogue text pairs by using a preset chain of thought, and then determine the corresponding dialogue content features, domain attributes, and historical experience of the historical minutes dialogue text pairs based on the triple training structure and the historical minutes dialogue text pairs, so as to determine the prompt word to be processed based on the dialogue content features, the domain attributes, and the historical experience.

[0086] In some specific embodiments, the abstract generation result determination module 11 may specifically include:

[0087] A first to-be-processed prompt word determination unit, configured to process the to-be-processed prompt word by using the to-be-trained generator model and based on a temperature parameter index, so as to obtain a first to-be-processed prompt word; the temperature parameter index is an index determined based on the creativity and accuracy of the to-be-processed prompt word;

[0088] A second to-be-processed prompt word determination unit, configured to perform structured prompt word marking on the first to-be-processed prompt word by using the to-be-trained generator model and based on a structural integrity index, so as to obtain a second to-be-processed prompt word; the structural integrity index includes a background information index, a main idea index, a decision conclusion index, and an action item index;

[0089] A third to-be-processed prompt word determination unit, configured to perform professional term interpretation on the second to-be-processed prompt word by using the to-be-trained generator model and based on a professional depth index, so as to obtain a third to-be-processed prompt word; the professional depth index is an index set based on the usage density information and interpretation depth information of the target audience;

[0090] An abstract generation result determination subunit, configured to perform professional term interpretation on the third to-be-processed prompt word by using the to-be-trained generator model and based on an abstract length index, so as to obtain an abstract generation result; the abstract length index is an index determined based on the information density and length balance of the to-be-processed prompt word.

[0091] In some specific embodiments, the generator model determination module 12 may specifically include:

[0092] A first scoring result determination unit, configured to perform information integrity scoring on the abstract generation result by using a to-be-trained scoring model, so as to obtain an information integrity scoring result; the numerical value of the information integrity scoring result is positively correlated with the core content coverage rate in the abstract generation result;

[0093] A second scoring result determination unit, configured to perform logical structure scoring on the abstract generation result by using the to-be-trained scoring model, so as to obtain a logical structure scoring result; the logical structure scoring result is used to characterize the content organization and coherence in the abstract generation result;

[0094] A third scoring result determination unit, configured to use the to-be-trained scoring model to determine whether there are characteristics of artificial writing in the abstract generation result, so as to obtain a authenticity scoring result;

[0095] A fourth scoring result determination unit, configured to use the to-be-trained scoring model and process the summary generation result based on the normativeness and accuracy of specific domain terms to obtain the accuracy rate of professional terms;

[0096] A fifth scoring result determination unit, configured to use the to-be-trained scoring model to perform a structural integrity scoring on the summary generation result to obtain a structural integrity scoring result;

[0097] A sixth scoring result determination unit, configured to determine a target scoring result based on the information integrity scoring result, the logical structure scoring result, the authenticity scoring result, the accuracy rate of professional terms, and the structural integrity scoring result.

[0098] In some specific embodiments, the scoring model determination module 13 may specifically include:

[0099] A difference report generation unit, configured to compare the summary generation result with the corresponding minutes text of the historical minutes dialogue text to obtain a text comparison result, and then determine a difference report based on the text comparison result, the summary generation result, and the minutes text;

[0100] A first scoring model determination subunit, configured to use a preset contrast learning algorithm and determine a comparison result based on the difference report, so as to adjust the model parameters of the to-be-trained scoring model by using the comparison result to obtain a to-be-processed scoring model; the model parameters are parameters for distinguishing the discrimination ability of the summary generation result and the minutes text;

[0101] A second scoring model determination subunit, configured to adjust the to-be-processed scoring model by using a preset dynamic learning rate adjustment mechanism to obtain a target scoring model.

[0102] In some specific embodiments, the summary text determination module 14 may specifically include:

[0103] A text processing unit, configured to process the to-be-recognized text by using preset data preprocessing rules to obtain a standardized text, and process the standardized text by using a preset prompt word generation module to obtain a target prompt word;

[0104] A summary generation result scoring unit, configured to perform a summary generation process on the target prompt word by using the target generator model to obtain a to-be-processed summary generation result, and score the to-be-processed summary generation result by using the target scoring model to obtain a target scoring result;

[0105] A prompt template library determination unit, configured to determine a feature prompt template based on the dialogue features corresponding to the text to be recognized and historical optimization experience, and then determine a to-be-processed prompt template library based on the feature prompt template, a preset task instruction, and a preset example summary;

[0106] An abstract generation result adjustment unit, configured to adjust the to-be-processed prompt template library based on the target scoring result to obtain a target prompt template library, so as to use the target prompt template library to adjust the to-be-processed abstract generation result to obtain a target abstract generation result.

[0107] In some specific embodiments, the abstract text determination module 14 may specifically include:

[0108] A first abstract text determination subunit, configured to determine whether the scoring fluctuation corresponding to the text processing result is less than a preset scoring fluctuation threshold. If the scoring fluctuation corresponding to the text processing result is not less than the preset scoring fluctuation threshold, then jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, until the scoring fluctuation corresponding to the text processing result is less than the preset scoring fluctuation threshold, and set the text processing result as the target abstract text;

[0109] A second abstract text determination subunit, configured to determine whether the current iteration number corresponding to the text processing result is less than a preset iteration number. If the current iteration number corresponding to the text processing result is less than the preset iteration number, then jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, until the current iteration number corresponding to the text processing result is not less than the preset iteration number, and set the text processing result as the target abstract text.

[0110] Furthermore, an embodiment of the present application also discloses an electronic device, Figure 3 which is a structural diagram of an electronic device 20 shown according to an exemplary embodiment. The content in the figure should not be considered as any limitation to the scope of use of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. Among them, the memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to implement the relevant steps in the text generation method based on the adversarial mechanism disclosed in any of the foregoing embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0111] In this embodiment, the power supply 23 is used to provide operating voltages for the various hardware devices on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and specific limitations thereof are not provided herein; the input / output interface 25 is used to obtain external input data or output data to the outside, and the specific interface type can be selected according to specific application requirements, and specific limitations are not provided herein.

[0112] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, a random access memory, a magnetic disk, an optical disk, etc., and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0113] Among them, the operating system 221 is used to manage and control the various hardware devices and the computer program 222 on the electronic device 20, and it can be Windows Server, Netware, Unix, Linux, etc. The computer program 222 can further include a computer program capable of performing other specific tasks in addition to the computer program capable of implementing the adversarial mechanism-based text generation method executed by the electronic device 20 disclosed in any of the foregoing embodiments.

[0114] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the adversarial mechanism-based text generation method disclosed above is implemented. For the specific steps of this method, reference can be made to the corresponding content disclosed in the foregoing embodiments, and details are not repeated herein.

[0115] In this specification, the various embodiments are described in a progressive manner, and the key points of each embodiment are the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0116] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0117] The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be implemented directly in hardware, in software modules executed by a processor, or in a combination thereof. The software modules may be disposed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0118] Finally, it should also be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0119] The technical solutions provided in this application have been introduced in detail above. Specific examples are used herein to illustrate the principles and implementation manners of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A text generation method based on an adversarial mechanism, characterized in that Including: Performing domain knowledge fine-tuning operation on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, generating a prompt to be processed for the corresponding dialogue content features and domain attributes based on the historical minutes dialogue text, and then using the generator model to be trained to process the prompt to be processed to obtain a summary generation result; Using the generator model to be trained and based on a preset multi-dimensional scoring rule to perform dimensional quality scoring on the summary generation result to obtain a target scoring result, and determining a target optimization function based on the target scoring result to use the target optimization function to perform gradient update on the parameters corresponding to the generator model to be trained to obtain a target generator model; Using a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text in the historical minutes dialogue text to obtain a comparison result, and adjusting the model parameters of the generator model to be trained based on the comparison result to obtain a target generator model; Using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, and determining whether the text processing result meets a preset processing stop condition. If it meets, setting the text processing result as the target summary text.

2. The text generation method based on an adversarial mechanism according to claim 1, wherein The performing domain knowledge fine-tuning operation on the initial generator model based on a preset domain corpus to obtain a generator model to be trained, and generating a prompt to be processed for the corresponding dialogue content features and domain attributes based on the historical minutes dialogue text, includes: Constructing a preset domain corpus containing specific domain professional terms and concepts, and performing fine-tuning training on the initial generator model based on a preset gradient accumulation rule and a preset low learning rate strategy to obtain a generator model to be trained; the initial generator model is a large language model; Collecting historical minutes dialogue text pairs, and using a preset chain of thought to insert a triple training structure including dialogue text, reasoning process, and final minutes in the historical minutes dialogue text pairs, and then determining the corresponding dialogue content features, domain attributes, and historical experience based on the triple training structure and the historical minutes dialogue text pairs, and determining the prompt to be processed based on the dialogue content features, the domain attributes, and the historical experience.

3. The text generation method based on an adversarial mechanism according to claim 1, wherein The using the generator model to be trained to process the prompt to be processed to obtain a summary generation result, includes: Using the generator model to be trained and based on a temperature parameter index to process the prompt to be processed to obtain a first prompt to be processed; the temperature parameter index is an index determined based on the creativity and accuracy of the prompt to be processed; Using the generator model to be trained and based on a structural integrity index to perform structured prompt marking on the first prompt to be processed to obtain a second prompt to be processed; the structural integrity index includes a background information index, a main idea index, a decision conclusion index, and an action item index. Using the to-be-trained generator model and based on the professionalism depth index, conduct professional term explanations on the second to-be-processed prompt word to obtain a third to-be-processed prompt word; the professionalism depth index is the usage density information and explanation depth information set based on the target audience; Using the to-be-trained generator model and based on the abstract length index, conduct professional term explanations on the third to-be-processed prompt word to obtain an abstract generation result; the abstract length index is an index determined based on the information density and length balance of the to-be-processed prompt word.

4. The text generation method based on an adversarial mechanism according to claim 1, wherein The using the to-be-trained scoring model and based on a preset multi-dimensional scoring rule to conduct dimensional quality scoring on the abstract generation result to obtain a target scoring result, including: Using the to-be-trained scoring model to conduct information integrity scoring on the abstract generation result to obtain an information integrity scoring result; the numerical size of the information integrity scoring result is positively correlated with the size of the core content coverage rate in the abstract generation result; Using the to-be-trained scoring model to conduct logical structure scoring on the abstract generation result to obtain a logical structure scoring result; the logical structure scoring result is used to characterize the content organization and coherence in the abstract generation result; Using the to-be-trained scoring model to determine whether there are features of artificial writing in the abstract generation result to obtain a authenticity scoring result; Using the to-be-trained scoring model and based on the normativity and accuracy of terms in a specific field to process the abstract generation result to obtain a professional term accuracy rate; Using the to-be-trained scoring model to conduct structural integrity scoring on the abstract generation result to obtain a structural integrity scoring result; Determine the target scoring result based on the information integrity scoring result, the logical structure scoring result, the authenticity scoring result, the professional term accuracy rate, and the structural integrity scoring result.

5. The text generation method based on an adversarial mechanism according to claim 1, characterized in that, The using a preset contrastive learning algorithm to compare and learn the abstract generation result with the corresponding minutes text in the historical minutes dialogue text pair to obtain a comparison result, and based on the comparison result, adjust the model parameters of the to-be-trained scoring model to obtain a target scoring model, including: Compare the abstract generation result with the corresponding minutes text in the historical minutes dialogue text pair to obtain a text comparison result, and then determine a difference report based on the text comparison result, the abstract generation result, and the minutes text; Using a preset contrastive learning algorithm and based on the difference report to determine a comparison result, so as to use the comparison result to adjust the model parameters of the to-be-trained scoring model to obtain a to-be-processed scoring model; the model parameters are parameters for distinguishing the discriminative ability of the abstract generation result and the minutes text; Using a preset dynamic learning rate adjustment mechanism to adjust the to-be-processed scoring model to obtain a target scoring model.

6. The text generation method based on an adversarial mechanism according to claim 1, wherein The using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, including: Process the text to be recognized using preset data preprocessing rules to obtain a standardized text, and process the standardized text using a preset prompt word generation module to obtain a target prompt word; Use the target generator model to perform abstract generation processing on the target prompt word to obtain a to-be-processed abstract generation result, and use the target scoring model to score the to-be-processed abstract generation result to obtain a target scoring result; Determine a feature prompt template based on the dialogue features corresponding to the text to be recognized and historical optimization experience, and then determine a to-be-processed prompt word template library based on the feature prompt template, preset task instructions, and preset example minutes; Adjust the to-be-processed prompt word template library based on the target scoring result to obtain a target prompt word template library, so as to adjust the to-be-processed abstract generation result using the target prompt word template library to obtain a target abstract generation result.

7. The text generation method based on an adversarial mechanism according to any one of claims 1 to 6, characterized in that Judging whether the text processing result meets a preset processing stop condition, if so, setting the text processing result as the target abstract text, including: Judging whether the score fluctuation corresponding to the text processing result is less than a preset score fluctuation threshold. If the score fluctuation corresponding to the text processing result is not less than the preset score fluctuation threshold, jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, until the score fluctuation corresponding to the text processing result is less than the preset score fluctuation threshold, and setting the text processing result as the target abstract text; Or, judging whether the current iteration number corresponding to the text processing result is less than a preset iteration number. If the current iteration number corresponding to the text processing result is less than the preset iteration number, jump to the step of using the target generator model and the target scoring model to process the text to be recognized to obtain a text processing result, until the current iteration number corresponding to the text processing result is not less than the preset iteration number, and setting the text processing result as the target abstract text.

8. A text generation device based on an adversarial mechanism, characterized in that Including: An abstract generation result determination module, configured to perform domain knowledge fine-tuning operations on an initial generator model based on a preset domain corpus to obtain a to-be-trained generator model, generate a to-be-processed prompt word based on the dialogue content features and domain attributes corresponding to the historical minutes dialogue text, and then use the to-be-trained generator model to process the to-be-processed prompt word to obtain an abstract generation result; A generator model determination module, configured to use a to-be-trained scoring model and perform dimensional quality scoring on the abstract generation result based on a preset multi-dimensional scoring rule to obtain a target scoring result, and determine a target optimization function based on the target scoring result, so as to use the target optimization function to perform gradient update on the parameters corresponding to the to-be-trained generator model to obtain a target generator model; A scoring model determination module, configured to use a preset contrastive learning algorithm to compare and learn the summary generation result with the corresponding minutes text in the historical minutes dialogue text pair, obtain a comparison result, and adjust the model parameters of the to-be-trained scoring model based on the comparison result to obtain a target scoring model; A summary text determination module, configured to use the target generator model and the target scoring model to process the text to be recognized, obtain a text processing result, and determine whether the text processing result meets a preset processing stop condition. If it meets, set the text processing result as the target summary text.

9. An electronic device, characterized in that, including: A memory, configured to store a computer program; A processor, configured to execute the computer program to implement the adversarial mechanism-based text generation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, For storing a computer program, wherein when the computer program is executed by the processor, the adversarial mechanism-based text generation method according to any one of claims 1 to 7 is implemented.

Images