Lyric model training method and system based on model self-iteration

Abstract

The invention discloses a lyric model training method and system based on model self-iteration, and relates to the field of digital music. The method comprises the steps that lyrics serve as training samples to train a lyric generation model; according to the lyrics output by the lyric generation model, generating reflection cue words, enabling the lyric generation module to perform self-evaluation and optimization on the lyrics according to the cue words, and outputting the optimized lyrics; and expanding a training sample by using the optimized lyrics, and iteratively training the lyrics to generate a model. According to the method, a model self-evaluation and correction mechanism is introduced in the self-iteration process, self-generated output is continuously optimized, and the optimized output is easier for model learning compared with real lyrics, so that the training process of the model is more stable, an over-fitting phenomenon is effectively prevented, the generated lyrics have a correct grammar structure, and the accuracy of the lyrics is improved. The content is rich, the form is diversified, and the application requirements in different scenes can be met.

Description

Technical Field

[0001] This application relates to the field of digital music technology, and particularly to a method and system for training a lyrics model based on model self-iteration. Background Art

[0002] With the rise of digital music platforms and the development of artificial intelligence technology, the field of music creation and processing is undergoing unprecedented changes. In recent years, the successful application of deep learning technology in the field of natural language processing has inspired researchers' interest in music text (such as lyrics) generation. Through training with large-scale datasets, neural network models have been able to simulate human creativity to a certain extent and generate lyrics with certain artistic value. However, most existing lyrics generation models rely on static datasets for one-time training, lacking the dynamics and interactivity that match the actual creative process, which limits the innovation ability and practicality of the models.

[0003] To this end, the industry has proposed a training method based on a customized model and a fine-tuning method for a pre-trained language model. The training method based on a customized model means building a model specifically for lyrics generation from scratch and directly using a lyrics dataset for training after random initialization. The model generates lyrics that meet the requirements by learning the vocabulary, sentence patterns, rhymes, and other characteristics in the data. This method is highly flexible and can be customized for the task, but it has high requirements for data quality, quantity, and computing resources. The fine-tuning method for a pre-trained language model is an efficient lyrics generation strategy that has emerged with the development of deep learning technology in recent years. This method first pre-trains a general language model using a large-scale text corpus to enable it to master a wide range of language rules and expression methods. Subsequently, it is fine-tuned for the specific task of lyrics generation. In this stage, a specially labeled lyrics dataset is used to further train the model to make it better adapt to the characteristics and styles of lyrics.

[0004] However, the training method based on a customized model relies on a large amount of high-quality data. When the data is insufficient, it is easy to cause overfitting to the training data or insufficient training of the model's capabilities. At the same time, the training process is time-consuming and requires high computing resources, and the initial generation quality is low, requiring repeated debugging and optimization. The fine-tuning method for a pre-trained language model has the problem of unstable generation quality. Although pre-trained models can usually generate smooth and somewhat creative text, in some cases, the generated lyrics may lack coherence or logic, or even have grammar errors. In addition, the model may sometimes overfit the training data, resulting in overly similar or uncreative generated content. Summary of the Invention

[0005] Based on this, it is necessary to provide a method and system for training a lyric model based on model self-iteration to address the above technical problems, which can continuously improve the lyric generation ability through the continuous learning and optimization of the model itself and improve the problem of unstable lyric content quality.

[0006] In a first aspect, the present application provides a method for training a lyric model based on model self-iteration. The method includes:

[0007] Using lyrics as training samples to train a lyric generation model;

[0008] Generating reflection prompts for the lyrics output by the lyric generation model, and enabling the lyric generation module to self-evaluate and optimize the lyrics according to the prompts, and output the optimized lyrics;

[0009] Using the optimized lyrics to expand the training samples and iteratively train the lyric generation model.

[0010] In one embodiment, self-evaluating and optimizing the lyrics includes:

[0011] Self-evaluating the lyrics. If the self-evaluation does not meet the expectation, iteratively optimize the lyrics according to the reflection prompts.

[0012] In one embodiment, the method further includes:

[0013] When the self-evaluation meets the expectation or the number of iterations reaches the preset number of times, stop the iteration and output the lyrics at the time of iteration stop as the optimized lyrics.

[0014] In one embodiment, self-evaluating the lyrics includes:

[0015] Determine several evaluation dimensions, including the content level, literary level, adaptability level, and innovation level, and assign corresponding weights to each evaluation dimension;

[0016] Give scores to each evaluation dimension respectively, and use the weights for weighted summation to obtain the quantified value of the self-evaluation;

[0017] Compare the quantified value with the preset threshold to evaluate whether the score meets the expectation.

[0018] In one embodiment, the indicators at the content level include: subject clarity, emotional sincerity, and content integrity;

[0019] The indicators at the literary level include: word accuracy, rhetorical expressiveness, and rhythm beauty;

[0020] The indicators at the adaptability level include: music style adaptability and rhythm coordination;

[0021] The innovation level includes creative uniqueness and style integration.

[0022] In one embodiment, generating reflection prompting words includes:

[0023] Integrating expert knowledge to construct a knowledge graph related to the lyrics;

[0024] Generating reflection prompting words in combination with the knowledge graph.

[0025] In a second aspect, the present application also provides a lyrics model training system based on model self-iteration. The system includes:

[0026] A model generation module for training a lyrics generation model using the lyrics as training samples;

[0027] A lyrics optimization module for generating reflection prompting words for the lyrics output by the lyrics generation model, enabling the lyrics generation module to self-evaluate and optimize the lyrics according to the prompting words, and outputting optimized lyrics;

[0028] An iterative update module for using the optimized lyrics to expand the training samples and iteratively training the lyrics generation model.

[0029] In a third aspect, the present application also provides a lyrics generation model, which is obtained by training using the above-mentioned lyrics model training method based on model self-iteration.

[0030] In a fourth aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps in the above-mentioned lyrics model training method based on model self-iteration.

[0031] In a fifth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps in the above-mentioned lyrics model training method based on model self-iteration.

[0032] The above-mentioned lyrics model training method and system based on model self-iteration use the lyrics as training samples to train the lyrics generation model; for the lyrics output by the lyrics generation model, generate reflection prompting words, enable the lyrics generation module to self-evaluate and optimize the lyrics according to the prompting words, and output optimized lyrics; use the optimized lyrics to expand the training samples and iteratively train the lyrics generation model. The present invention introduces a mechanism of model self-evaluation and correction in the self-iteration process, continuously optimizes the self-generated output. The optimized output is more conducive to the model's learning compared to the real lyrics, thereby making the model training process more stable, effectively preventing the occurrence of overfitting phenomena, enabling the generated lyrics to have correct grammar structures, and being rich in content and diverse in form, capable of meeting the application requirements in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the system architecture diagram of the lyric model training method based on model self-iteration in an embodiment. Specific implementation manner

[0034] In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0035] The embodiment of the present application provides a lyric model training method based on model self-iteration, as Figure 1 shown, including the following steps:

[0036] Step 102, using the lyrics as training samples to train the lyric generation model.

[0037] Specifically, format the high-quality lyric data screened manually and convert it into the alpaca data format required for model training as the training samples.

[0038] On the pre-trained large language model, use the above training samples to fine-tune the large language model to obtain a model with basic lyric generation ability, and complete the initial training of the lyric generation model.

[0039] Step 104, generate reflection prompt words for the lyrics output by the lyric generation model, and make the lyric generation module self-evaluate and optimize the lyrics according to the prompt words, and output the optimized lyrics.

[0040] Send an instruction to the lyric generation model obtained in step 102. The instruction can be the expectation for the lyric theme, style, etc., such as "write a lyric about a new beginning". The lyric generation model will output corresponding lyrics according to the instruction.

[0041] For the above output lyrics, the lyric generation model compares with the preset lyric evaluation system to self-reflect on the problems and optimization directions of the lyrics, and generates reflection prompt words according to the self-reflection. The lyric generation module self-evaluates and optimizes the above output lyrics according to the reflection prompt words to obtain the optimized lyrics.

[0042] For example, for the initial lyrics output by the lyric generation model, self-reflect on the problem that the content is slightly ordinary and lacks artistic appeal, and generate reflection prompt words according to this problem. The lyric generation module re-outputs the optimized lyrics according to the reflection prompt words and self-evaluates the optimized lyrics according to the preset evaluation system.

[0043] In one embodiment, the self-evaluation of lyrics includes: determining several evaluation dimensions, including the content level, literary level, adaptability level, and innovation level, and assigning corresponding weights to each evaluation dimension; giving scores to each evaluation dimension respectively, and using the weights for weighted summation to obtain the quantitative value of the self-evaluation.

[0044] To more precisely conduct self-evaluation of lyrics, this embodiment designs an evaluation system including several dimensions. The specific dimensions include: the content level, literary level, adaptability level, and innovation level. To fully consider the importance differences of each dimension and balance the influence of each dimension, corresponding weights are assigned to each evaluation dimension. In different evaluation scenarios and purposes, the weights can be adjusted according to specific needs. For example, when promoting a song in the mass market, the adaptability and innovation of the lyrics may be more important because well-adapted lyrics can better blend with the music and are easy to sing; highly innovative lyrics are more likely to attract the attention of listeners. At this time, the weights of the adaptability and innovation levels can be appropriately increased. Another example is that when conducting academic research on lyric creation, more attention may be paid to the content depth and literary value of the lyrics. Therefore, the weights of the content level and literary level can be increased to deeply analyze the creation techniques and cultural connotations of the lyrics. Since the basic model of the lyric generation model is a large language model, semantic analysis technology, sentiment analysis algorithms and other technologies of the large language model can be used to score the lyrics from multiple evaluation dimensions, and the quantitative value of the self-evaluation is obtained after weighted summation using the corresponding weights.

[0045] In one embodiment, the indicators at the content level include: subject clarity, emotional sincerity, and content integrity; the indicators at the literary level include: word precision, rhetorical expressiveness, and rhythm beauty; the indicators at the adaptability level include: music style adaptability and rhythm coordination; the innovation level includes creative uniqueness and style integration.

[0046] For example, set the weight of the content level to 30%, the weight of the literary level to 30%, the weight of the adaptability level to 25%, and the weight of the innovation level to 15%.

[0047] For the lyrics generated by the lyric generation model, semantic analysis techniques can be used to determine the frequency and coherence of the core ideas within the lyrics. If the lyrics frequently and coherently revolve around a core theme, a score of 8-10 is assigned. If a theme is present but the coherence is average, a score of 5-7 is assigned. If the theme is fragmented and difficult to discern, a score of 1-4 is assigned. Sentiment analysis algorithms are used to identify the emotional tendency and intensity within the lyrics. Strong and nuanced emotional expression, with a variety of emotional details, is assigned a score of 8-10. Clear but not rich emotional tendencies are assigned a score of 5-7. Ambiguous or false emotional expression is assigned a score of 1-4. The structure and logic of the lyrics are analyzed. A complete narrative structure or emotional development, along with rich content, is assigned a score of 8-10. A generally complete structure with minor flaws is assigned a score of 5-7. A disorganized structure with incomplete content is assigned a score of 1-4.

[0048] On the literary level, use lexical and grammatical analysis to check whether the wording is appropriate and accurate. If the wording is precise and consistent with the style of the lyrics, score 8-10; if there are a few inaccuracies, score 5-7; if there are many word errors, score 1-4. Use pattern recognition technology to identify the use of rhetorical devices in the lyrics. If the rhetorical devices are used naturally and enhance the expressiveness, score 8-10; if some rhetorical devices are used but the effect is average, score 5-7; if the rhetorical devices are used awkwardly or rarely, score 1-4. Analyze the rhyme and rhythm patterns of the lyrics. If the rhymes are harmonious and the rhythm is well-controlled, score 8-10; if the rhymes are basically reasonable and the rhythm is relatively coordinated, score 5-7; if the rhymes are confusing and the rhythm is mismatched, score 1-4.

[0049] Regarding adaptability, once the large model has been trained to match musical style with lyric features, it can use its own knowledge to determine the degree of fit between the lyric style and the preset musical style. A perfect match is assigned a score of 10-12; a relatively good match with minor issues is assigned a score of 6-9; a mismatch, seriously affecting the overall effect, is assigned a score of 1-5. The lyrics are evaluated for their fluency and singability by analyzing pronunciation and rhythmic patterns. A score of 10-12 is assigned for fluency and coordination with the beat; a score of 6-9 is assigned for fluency with some awkwardness; and a score of 1-5 is assigned for awkwardness and difficulty.

[0050] Regarding innovation, compare the lyrics with a large number of existing lyrics to determine the novelty of the ideas. If the ideas are novel and unique, with a unique perspective and fresh insights, score 6-8 points; if they are relatively novel but not outstanding, score 3-5 points; if they lack creativity and are similar to common lyrics, score 1-2 points. Analyze the integration of different elements in the lyrics. If the integration is natural, smooth, and harmonious, score 6-8 points; if it is relatively natural but has some flaws, score 3-5 points; if the integration is abrupt and inconsistent, score 1-2 points.

[0051] After the lyrics generation model gives specific scores based on the scoring criteria of each dimension mentioned above, it performs weighted calculations according to the set weights to obtain the quantitative total score of the lyrics self-evaluation.

[0052] In one embodiment, the lyrics that may be generated after step 104 undergoes a self-reflection and optimization step still do not achieve the desired effect. Therefore, an iterative optimization method is adopted to obtain the expected lyrics.

[0053] Specifically, the lyrics are self-evaluated. If the self-evaluation does not meet the expectation, the lyrics are iteratively optimized according to the reflection prompt words.

[0054] In this embodiment, by comparing the quantified value of the self-evaluation with a preset threshold, it is evaluated whether the score meets the expectation.

[0055] When the self-evaluation meets the expectation, the iteration can be stopped, and the lyrics at the time of iteration stop are output as the finally optimized lyrics.

[0056] In one embodiment, to avoid an infinite loop, a reasonable upper limit of the number of iterations is set. When this upper limit is reached, even if the scoring criteria are not met, the iteration needs to be stopped, and the lyrics at the time of iteration stop are output as the finally optimized lyrics.

[0057] Step 106, use the optimized lyrics to expand the training samples and iteratively train the lyrics generation model.

[0058] Collect all the optimized lyrics data generated during the self-iteration stage of the lyrics generation model, add them as new training samples to the original training data, return to step 102, and use the expanded training data to iteratively fine-tune the lyrics generation model again.

[0059] Evaluate the performance indicators of the lyrics generation model to ensure that the lyrics generation model is stable and reliable, and it can be applied to actual lyrics creation scenarios.

[0060] In one embodiment, the generated reflection prompt words include: integrating expert knowledge, constructing a knowledge graph related to the lyrics; generating reflection prompt words in combination with the knowledge graph.

[0061] By integrating expert knowledge, a domain knowledge graph is constructed. The knowledge graph represents the relationships and attributes between entities in a graphical way, and can clearly show the structure and hierarchy of expert knowledge. The expert knowledge that can be integrated includes knowledge in literature (such as intention expression, part-of-speech collocation, rhetorical devices, etc.), knowledge in music (such as the fit between pronunciation and melody), knowledge in the market (such as popular trends, audience psychology research, etc.). Using the reasoning function of the knowledge graph, new knowledge is derived from the existing expert knowledge. The content of the prompt words is further improved according to the reasoning results.

[0062] This embodiment integrates expert knowledge to design reflection prompt words for the lyrics generation model, thereby providing a more accurate professional reference framework for the lyrics generation model and exploring deeper lyrics features to obtain more reliable and accurately expressed lyrics output.

[0063] In one embodiment, the lyrics model training method based on model self-iteration includes the following steps:

[0064] S1. Data preprocessing:

[0065] S11. Formatting lyrics data: Converting lyrics data into the unified alpaca data format and performing word segmentation on the data.

[0066] S12. Build a dataset: Divide the formatted lyrics data into a training set and a validation set according to a certain ratio to ensure that the data in each set is evenly distributed.

[0067] S2. Preliminary model training:

[0068] S21. Loading a pre-trained model: Select a large language model that has been pre-trained on a large corpus as the base model and load its parameter configuration and weights.

[0069] S22. Fine-tune the model: Use the training set constructed above to fine-tune the basic model, adjusting hyperparameters such as learning rate and batch size until the model's performance on the validation set meets the expected standards.

[0070] S23. Test model performance: Use the test set to evaluate the lyrics generation ability of the fine-tuned model to ensure that the model has basic lyrics generation capabilities.

[0071] S3. Model self-iteration:

[0072] S31. Design reflection prompts: Refer to professional knowledge of lyric writing and design a series of prompts that can guide the model to self-reflect and optimize, such as "Please check whether the lyrics rhyme" and "Try to increase emotional expression".

[0073] S32. Generate and optimize lyrics: The model generates an initial version of the lyrics based on the preliminary training results, and then self-evaluates and optimizes the lyrics based on the designed prompt words to form a higher quality version of the lyrics.

[0074] S33. Self-evaluation and iteration: The model self-scores the optimized lyrics. If the score does not reach the preset threshold, it continues to optimize the lyrics using reflection prompts. If the score reaches or exceeds the threshold, it stops iteration and outputs the final version of the lyrics.

[0075] S34. Set the iteration upper limit: To avoid infinite loops, set a reasonable upper limit for the number of iterations. When this limit is reached, even if the scoring criteria are not met, the iteration must stop, and the true lyric data corresponding to this piece of data should be output.

[0076] S4. Model retraining stage:

[0077] S41. Organize and optimize data: Collect all the optimized lyric data generated during the model self-iteration stage as new training samples.

[0078] S42. Expand the training set: Add the newly collected optimized data to the original training set to construct a larger-scale and higher-quality training data set.

[0079] S43. Fine-tune the model again: Use the expanded training set to retune the model to further improve the model's lyric generation ability and creative level.

[0080] S44. Final evaluation and release: After completing the retraining, re-evaluate the performance metrics of the model. After ensuring that the model is stable and reliable, it can be applied to actual lyric creation scenarios.

[0081] The present invention proposes a lyric model training method and system based on model self-iteration, which can improve the problem of unstable quality of lyric content generated by existing technical solutions.

[0082] Although the training method based on a customized model and the fine-tuning method of a pre-trained language model can generate acceptable lyric texts to a certain extent, they still have deficiencies in terms of content coherence, logic, and innovation. By integrating expert knowledge to guide the model for self-iteration, the present invention can not only ensure the basic quality of the generated lyrics but also stimulate the creativity of the model to a certain extent, generating more unique and personalized lyric works.

[0083] In summary, compared with the existing lyric model training methods, the present invention for the first time proposes a lyric model training method and system based on model self-iteration. The core lies in continuously learning and optimizing the model itself to continuously improve the lyric generation ability. Aiming at the problem of unstable generation quality in the fine-tuning method of the pre-trained language model, the present invention optimizes the model training process through the technology of model self-iteration. While improving the fluency of the generated lyrics, it also enhances its logic and creativity. Specifically, the present invention introduces a mechanism for model self-evaluation and correction during the self-iteration process, continuously optimizing the self-generated output. The optimized output is more conducive to the model's learning compared to the true lyrics, thus making the model training process more stable, effectively preventing the occurrence of overfitting phenomena, enabling the generated lyrics to have correct grammar structures, rich content, and diverse forms, and meeting the application requirements in different scenarios.

[0084] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the indications of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0085] Based on the same inventive concept, an embodiment of the present application also provides a model self-iterative lyric model training system for implementing the above-mentioned model self-iterative lyric model training method. The implementation solutions provided by this system for solving problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the following model self-iterative lyric model training system can refer to the limitations on the model self-iterative lyric model training method in the above text, and will not be repeated here.

[0086] In one embodiment, a model self-iterative lyric model training system is provided, including:

[0087] A model generation module, configured to use lyrics as training samples to train a lyric generation model;

[0088] A lyric optimization module, configured to generate reflection prompt words for the lyrics output by the lyric generation model, and cause the lyric generation module to self-evaluate and optimize the lyrics according to the prompt words, and output optimized lyrics;

[0089] An iterative update module, configured to use the optimized lyrics to expand the training samples and iteratively train the lyric generation model.

[0090] Each module in the above model self-iterative lyric model training system can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0091] In one embodiment, a lyric generation model is provided, and the lyric generation model is obtained by training using the above model self-iterative lyric model training method.

[0092] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in all the above method embodiments are implemented.

[0093] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0094] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0095] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. And the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0096] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0097] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0098] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for training a lyric model based on model self-iteration, characterized in that, The method includes: Using the lyrics as training samples to train a lyrics generation model; Generating reflection prompt words for the lyrics output by the lyrics generation model, and causing the lyrics generation module to self-evaluate and optimize the lyrics according to the prompt words, and output the optimized lyrics; Using the optimized lyrics to expand the training samples and iteratively train the lyrics generation model.

2. The method according to claim 1, wherein The self-evaluation and optimization of the lyrics include: Conducting self-evaluation on the lyrics. If the self-evaluation does not meet the expectation, iteratively optimize the lyrics according to the reflection prompt words.

3. The method according to claim 2, characterized in that, The method further includes: When the self-evaluation meets the expectation or the number of iterations reaches the preset number, stop the iteration and output the lyrics at the time of iteration stop as the optimized lyrics.

4. The method according to claim 2, wherein The self-evaluation of the lyrics includes: Determining several evaluation dimensions, including the content level, literary level, adaptability level, and innovation level, and respectively assigning corresponding weights to each of the evaluation dimensions; Giving scores to each of the evaluation dimensions respectively, and performing weighted summation using the weights to obtain a quantitative value of the self-evaluation; Comparing the quantitative value with a preset threshold to evaluate whether the score meets the expectation.

5. The method according to claim 4, wherein: The indicators at the content level include: subject clarity, emotional sincerity, and content integrity; The indicators at the literary level include: word precision, rhetorical expressiveness, and rhythm beauty; The indicators at the adaptability level include: music style adaptability and rhythm coordination; The innovation level includes creative uniqueness and style integration.

6. The method according to claim 1, wherein The generation of the reflection prompt words includes: Integrating expert knowledge to construct a knowledge graph related to the lyrics; Generating the reflection prompt words in combination with the knowledge graph.

7. A lyric model training system based on model self-iteration, characterized in that The system includes: A model generation module for using the lyrics as training samples to train a lyrics generation model; A lyrics optimization module for generating reflection prompt words for the lyrics output by the lyrics generation model, and causing the lyrics generation module to self-evaluate and optimize the lyrics according to the prompt words, and output the optimized lyrics; An iterative update module for using the optimized lyrics to expand the training samples and iteratively train the lyrics generation model.

8. A lyric generation model, characterized in that, The lyrics generation model is obtained by training using the steps of the method according to any one of claims 1 to 6.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images