Cboth modification method and device, terminal and storage medium

Abstract

The invention discloses a copywriting modification method and device, a terminal and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps of generating a second evaluation standard for an initial copywriting and a first evaluation standard through a copywriting evaluation model; evaluating the initial copywriting according to the second evaluation standard to obtain a target evaluation opinion; and generating a target modification opinion based on the target evaluation opinion. According to the document evaluation model, the initial first evaluation standard is supplemented and perfected, and a more comprehensive second evaluation standard is generated. The initial copywriting is evaluated based on the second evaluation standard, accurate target evaluation opinions can be output to clarify specific problems existing in the copywriting, targeted target modification opinions are generated according to the target evaluation opinions, and automatic modification of the copywriting can be achieved. The problems that the manual modification level is difficult to keep consistent and the cost is high are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a script modification method and device, a terminal and a storage medium. BACKGROUND

[0002] The existing script generation is usually completed by using a large language model, such as a short video voiceover script and a public number article. In order to make the generated script more in line with the requirements, a prompt engineering method is generally used to guide the model to generate more suitable content. In addition, in order to solve the illusion problem of the large model and the problem of insufficient real-time information, a retrieval-augmented generation (RAG) method is also used to obtain a reference basis for the model to generate a script by retrieving an external knowledge base, so as to enhance the accuracy and timeliness of the content. However, the script generated based on the two methods still has some problems in the perspective of professionals, such as insufficient information, not attractive enough at the beginning, and incoherent logic.

[0003] At present, after the script is generated by using the large language model, professional modification is still needed according to the scene and experience, and the artificial modification varies from person to person, and the modification level is not consistent, and in addition, the artificial modification also increases the cost of script generation.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a script modification method, device, terminal and storage medium to solve the problem that the script generated by using the large language model still needs to be manually modified, the artificial modification level is difficult to maintain consistency and the cost is high.

[0006] The technical scheme adopted by the present application to solve the problem is as follows: In a first aspect, the present application provides a script modification method, which comprises: generating a second evaluation standard for an initial script and a first evaluation standard by using a script evaluation model; evaluating the initial script according to the second evaluation standard to obtain target evaluation opinions; generating target modification opinions based on the target evaluation opinions.

[0007] In an embodiment, the method for obtaining the initial script comprises: obtaining a topic and a script type; generating the initial script according to the topic and the script type by using a preset script generation model.

[0008] In an implementation, a second evaluation standard is generated for the initial script and the first evaluation standard, including: The first evaluation standard is perfected based on at least one of the script type, the topic, and the script content corresponding to the initial script, to obtain the second evaluation standard.

[0009] In an implementation, the initial script is evaluated according to the second evaluation standard to obtain a target evaluation opinion, including: The initial script is evaluated according to each evaluation dimension in the second evaluation standard to obtain an evaluation grade and an evaluation reason corresponding to each evaluation dimension, respectively; The target evaluation opinion is generated according to the evaluation grade and the evaluation reason of each evaluation dimension.

[0010] In an implementation, a target modification opinion is generated based on the target evaluation opinion, including: A number of target evaluation dimensions to be modified are determined based on the target evaluation opinion; For each target evaluation dimension, a modification opinion is determined according to the evaluation grade and the evaluation reason corresponding to the target evaluation dimension; The target modification opinion is generated according to the modification opinions of all the target evaluation dimensions.

[0011] In an implementation, the target modification opinion includes all specific problem points of the evaluation dimensions whose evaluation grades meet a preset grade, and corresponding specific modification instructions; the specific modification instructions include one or more instructions of a specific operation mode of modification, content of modification, and specific location of modification.

[0012] In an implementation, the script evaluation model is pre-trained, and a training method includes: A number of training scripts corresponding to each script type are obtained; For each training script, a second training evaluation standard is generated for the training script and a first training evaluation standard by a preset large language model; a training evaluation opinion is obtained by evaluating the training script according to the second training evaluation standard; a training modification opinion is generated based on the training evaluation opinion; and training data is generated according to the training script, the training evaluation opinion, and the training modification opinion; The initial script evaluation model is fine-tuned according to all the training data, and the fine-tuned initial script evaluation model is used as the script evaluation model.

[0013] In an implementation, a second training evaluation standard is generated for the training script and a first training evaluation standard, including: The first training evaluation standard is perfected based on at least one of the data of the type of the training script corresponding to the training script, the topic and the content of the training script, to obtain the second training evaluation standard.

[0014] In an implementation, the training script is evaluated according to the second training evaluation standard to obtain a training evaluation opinion, including: The training script is evaluated according to each training evaluation dimension in the second training evaluation standard to obtain an evaluation grade and an evaluation reason corresponding to each training evaluation dimension, respectively; The training evaluation opinion is generated according to the evaluation grade and the evaluation reason of each training evaluation dimension.

[0015] In an implementation, the training modification opinion is generated based on the training evaluation opinion, including: A plurality of target training evaluation dimensions to be modified are determined based on the training evaluation opinion; For each target training evaluation dimension, a modification opinion is determined according to the evaluation grade and the evaluation reason corresponding to the target training evaluation dimension; The training modification opinion is generated according to the modification opinions of all the target training evaluation dimensions.

[0016] In an implementation, the training data is generated according to the training script, the training evaluation opinion and the training modification opinion, including: The training evaluation opinion and the training modification opinion are checked; If all the checks are qualified, the training evaluation opinion and the training modification opinion are taken as label data, and the training data is generated according to the training script and the label data.

[0017] In an implementation, the training evaluation opinion and the training modification opinion are checked, including: The consistency of the inference path of the large language model with the training evaluation opinion is checked by a preset checking model; And / or, the accuracy of the evaluation result of each evaluation dimension in the training evaluation opinion is checked by a preset checking model; And / or, the consistency of the evaluation reason of each evaluation dimension in the training evaluation opinion with the training modification opinion is checked by a preset checking model; And / or, the specificity of the training modification opinion is checked by a preset checking model.

[0018] In an implementation, after the target modification opinion is generated, it further includes: The initial script is modified according to several pieces of modification opinions in the target modification opinions based on a hierarchical modification principle through a preset script modification model, and a target script is obtained.

[0019] In an implementation, the initial script is modified according to several pieces of modification opinions in the target modification opinions based on a hierarchical modification principle, including: The several pieces of modification opinions in the target modification opinions are taken as execution modification opinions; For each execution modification opinion, a modification priority is determined according to an evaluation grade corresponding to the execution modification opinion; The initial script is modified according to modification priorities of the execution modification opinions based on a hierarchical modification principle.

[0020] In an implementation, the initial script is modified according to modification priorities of the execution modification opinions based on a hierarchical modification principle, including: The initial script is modified according to modification priorities of the execution modification opinions in sequence; For each execution modification opinion, if data of a key evaluation dimension corresponding to the initial script changes more than a preset range when the initial script is modified according to the execution modification opinion, a modification manner and / or a modification range of the execution modification opinion are adjusted; the key evaluation dimension is an evaluation dimension whose evaluation grade meets a preset requirement.

[0021] In an implementation, after the initial script is modified according to several pieces of modification opinions in the target modification opinions based on a hierarchical modification principle through a preset script modification model, and a target script is obtained, the method further includes: The target script is input into the script evaluation model to obtain updated evaluation opinions and updated modification opinions; It is judged whether the updated evaluation opinions meet a termination condition, if not, the target script is taken as the initial script, and the updated modification opinions are taken as the target modification opinions; The step of modifying the initial script according to several pieces of modification opinions in the target modification opinions based on a hierarchical modification principle through a preset script modification model is continuously performed until the updated evaluation opinions meet the termination condition.

[0022] In a second aspect, the embodiments of the present application further provide a script modification device, and the device includes: A script evaluation model is configured to generate second evaluation standards for an initial script and first evaluation standards; The initial script is evaluated according to the second evaluation standards to obtain target evaluation opinions; Target modification opinions are generated based on the target evaluation opinions.

[0023] In a third aspect, an embodiment of the present application further provides a terminal, comprising a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the script modification method according to any one of the above; and the processors are configured to execute the programs.

[0024] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, which stores a plurality of instructions, the instructions being adapted to be loaded and executed by a processor to implement the steps of the script modification method according to any one of the above.

[0025] The script modification method provided by the embodiment of the present application has the following beneficial effects: the script evaluation model is used to generate a second evaluation standard from an initial script and a first evaluation standard; the initial script is evaluated according to the second evaluation standard to obtain target evaluation opinions; and target modification opinions are generated based on the target evaluation opinions. The script evaluation model can supplement and improve the initial first evaluation standard to generate a more comprehensive second evaluation standard. The initial script is evaluated based on the second evaluation standard, accurate target evaluation opinions can be output to clearly indicate specific problems of the script, and target modification opinions can be generated according to the target evaluation opinions, so that the script can be automatically modified. The problem that the manual modification level is difficult to keep consistent and the cost is high is solved. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0027] Figure 1 is a flowchart of the script modification method provided by the embodiment of the present application.

[0028] Figure 2 is a flowchart of the full-automatic script processing process provided by the embodiment of the present application.

[0029] Figure 3 is a module diagram of the script modification device provided by the embodiment of the present application.

[0030] Figure 4 is a principle block diagram of the terminal provided by the embodiment of the present application. DETAILED DESCRIPTION

[0031] The application discloses a copywriting modification method and device, a terminal and a storage medium. To make the purpose, technical solutions and effects of the application clearer and more explicit, the application is further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0032] Those skilled in the art can understand that the singular forms "a", "an" and "the" used herein include plural forms unless specifically stated otherwise. It should be further understood that the use of the phrase "comprises" in the specification of the application means that the features, integers, steps, operations, elements and / or components exist, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. The phrase "and / or" used herein includes all or any single unit and all combinations of the associated listed items.

[0033] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as that generally understood by those skilled in the art to which the application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have meanings consistent with those in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as such.

[0034] To overcome the above-mentioned defects of the prior art, the application provides a copywriting modification method, relating to the field of artificial intelligence. The method comprises: generating a second evaluation standard for an initial copywriting and a first evaluation standard through a copywriting evaluation model; evaluating the initial copywriting according to the second evaluation standard to obtain target evaluation opinions; and generating target modification opinions based on the target evaluation opinions. The copywriting evaluation model of the application can supplement and improve the initial first evaluation standard to generate a more comprehensive second evaluation standard. Based on the second evaluation standard, the initial copywriting can be evaluated to output accurate target evaluation opinions to clearly indicate specific problems existing in the copywriting, and target modification opinions can be generated according to the target evaluation opinions to realize automatic modification of the copywriting. The problem that the level of manual modification is difficult to keep consistent and the cost is high is solved.

[0035] As shown in Figure 1 The method specifically comprises the following steps: Step S100, generating a second evaluation standard for an initial copywriting and a first evaluation standard through a copywriting evaluation model.

[0036] Specifically, the initial script refers to an original generated script, which can be generated by a large language model or other tools. The first evaluation standard can be a known evaluation standard input into the script evaluation model, which is artificially formulated based on industry experience. However, the artificial evaluation standard is subjective, and it is difficult to objectively evaluate the quality of the script, and the cost of artificial evaluation is high. The first evaluation standard can also be an evaluation standard defined by the script evaluation model based on the initial script. The first evaluation standard includes one or more indicators for measuring the quality of the initial script. The present embodiment pre-establishes a script evaluation model. The script evaluation model is pre-trained through data training and can supplement and improve the first evaluation standard to obtain a more comprehensive second evaluation standard. It should be noted that the second evaluation standard is improved based on the first evaluation standard, but is not limited to updating only the first evaluation standard.

[0037] In an implementation manner, the method for obtaining the initial script comprises: obtaining a topic and a script type; generating the initial script according to the topic and the script type through a preset script generation model.

[0038] Specifically, the topic and the script type need to be determined before the script is generated, and then the corresponding script draft is output through the script generation model, that is, the initial script is obtained. The topic is used to determine the content of the script; and the script type is used to determine the style or expression form of the script. In actual application scenarios, in order to make the generated script more in line with the requirements, a prompt engineering method can be used to construct script generation prompt data according to the topic and the script type, so as to guide the script generation model to generate a more suitable initial script.

[0039] For example, a general open source model or a closed source model (such as ChatGPT) can be used to establish a generation model for generating an initial script. Different script types are pre-designed to correspond to prompt data (prompt) templates. In actual application scenarios, users can call corresponding prompt data templates according to script types to generate appropriate script generation prompt data; or customize prompt data templates to obtain customized prompt data. The present embodiment does not need excessive prompt engineering (prompt engineering), and only needs to use relatively simple prompt data to obtain an initial script.

[0040] Taking the short video voiceover script as an example, the customized prompt data template is as follows: Please write an {article_type} around the topic "A country's cuisine" that can attract users' attention on short video platforms and return the preset data format (such as JSON format): {{"title": "xxx", "content": "xxx"}}; Topic: A country's cuisine; Script type: short video voiceover script.

[0041] The custom prompt data template can guide the model to write a short video voiceover script based on the topic, thereby obtaining an initial script. It can be understood that xxx in the above custom prompt data template refers to example text content.

[0042] Step S200, evaluating the initial script according to the second evaluation standard to obtain target evaluation opinions.

[0043] Step S300, generating target modification opinions based on the target evaluation opinions.

[0044] Specifically, the script evaluation model will comprehensively evaluate the initial script according to the second evaluation standard to generate reasonable target evaluation opinions. And generate accurate, detailed and specific target modification opinions based on the target evaluation opinions. Because the target modification opinions are clear and operable, the user can manually modify the initial script according to the target modification opinions, or use the large model to accept part or all of the modification opinions and modify the initial script.

[0045] In an implementation manner, the script evaluation model generates a second evaluation standard for the initial script and the first evaluation standard, including: Based on at least one of the script type, topic, and script content corresponding to the initial script, the first evaluation standard is perfected to obtain the second evaluation standard.

[0046] The first evaluation standard is prone to imperfections. Therefore, the script evaluation model in this embodiment will supplement and perfect the first evaluation standard, and then evaluate the initial script according to the second evaluation standard. Specifically, the evaluation standard usually includes multiple evaluation dimensions, and the evaluation dimensions of different types of scripts are different, and different topics and script content may also derive different evaluation dimensions. Therefore, one or more of the script type, topic, and script content of the initial script can be referred to to supplement the missing evaluation dimensions in the first evaluation standard, and / or to perfect the existing evaluation dimensions of the first evaluation standard. After perfecting, the second evaluation standard is obtained. In actual application scenarios, the script evaluation prompt data can be constructed according to at least one of the script type, topic, and script content to guide the script evaluation model to evaluate the initial script in detail and reasonably.

[0047] In an implementation manner, the initial script is evaluated according to the second evaluation standard to obtain a target evaluation opinion, including: The initial script is evaluated according to each evaluation dimension in the second evaluation standard to obtain an evaluation grade and an evaluation reason corresponding to each evaluation dimension respectively; The target evaluation opinion is generated according to the evaluation grade and the evaluation reason of each evaluation dimension.

[0048] Specifically, the second evaluation standard contains specific multiple evaluation dimensions, and the script evaluation model will finely evaluate the initial script according to the second evaluation standard by dimensions, and finally output the evaluation grade and the evaluation reason of each evaluation dimension to form a comprehensive and targeted target evaluation opinion. The evaluation grade is a quality grading of the evaluation dimension, which is used to intuitively reflect the performance of the initial script in the evaluation dimension, and can adopt a grading system such as good, medium and poor, or a score system. The evaluation reason is an explanation of the evaluation grade, which needs to specifically explain the problems of the initial script in the evaluation dimension or the reasons for excellent performance, avoiding formal evaluation. In this embodiment, each evaluation dimension has a clear evaluation grade and evaluation reason, which can clearly and intuitively reflect the advantages and disadvantages of the initial script in each evaluation dimension, and provide a data basis for subsequent generation of modification opinions.

[0049] In an implementation manner, the target modification opinion is generated based on the target evaluation opinion, including: A plurality of target evaluation dimensions to be modified are determined based on the target evaluation opinion; For each target evaluation dimension, a modification opinion is determined according to the evaluation grade and the evaluation reason corresponding to the target evaluation dimension; The target modification opinion is generated according to the modification opinions of all the target evaluation dimensions.

[0050] Specifically, the initial script is evaluated by the trained script evaluation model, and target evaluation opinions are output. The target evaluation opinions have clearly indicated the performance of the initial script in each evaluation dimension, and one or more evaluation dimensions need to be further selected as target evaluation dimensions. For each target evaluation dimension, the evaluation level of the dimension can reflect the severity of the problem, and the evaluation reason can clearly indicate the specific performance of the problem. The modification opinions need to be generated in combination with the evaluation level and the evaluation reason to avoid generating vague modification opinions. The modification opinions of all target evaluation dimensions are summarized to obtain target modification opinions. In actual application scenarios, the target modification opinions can include all specific problem points of the evaluation dimensions whose evaluation levels meet the preset levels, and corresponding specific modification instructions. The specific modification instructions include one or more instructions of the specific operation mode of the modification, the content of the modification, and the specific position of the modification. Thus, the target modification opinions are clear and operable.

[0051] For example, script evaluation prompt data can be generated based on the script itself corresponding to the initial script, the script type, and the evaluation standard. The script evaluation model is guided by the script evaluation prompt data to first perfect the evaluation standard to obtain a second evaluation standard, and then performs script evaluation according to the second evaluation standard. The evaluation standard includes multiple evaluation dimensions, for example, one of the evaluation dimensions can be logical clarity. When the script evaluation is completed, the evaluation level and the evaluation reason need to be output for each evaluation dimension respectively, and are summarized into target evaluation opinions. The evaluation dimensions whose evaluation levels do not meet the requirements in the target evaluation opinions can be selected as target evaluation dimensions in priority. For example, the evaluation levels include three levels of good, medium, and poor, and the evaluation dimensions whose evaluation levels do not meet the good level are selected as target evaluation dimensions. Alternatively, one or more evaluation dimensions can be selected from the target evaluation opinions as target evaluation dimensions in combination with one or more indicators in the script type, the topic, and the application scenario. In other words, even if the evaluation level of a certain evaluation dimension is not low, but it deviates from the script type, the topic, and the application scenario, it can also be selected as a target evaluation dimension. Finally, specific modification opinions are generated for all target evaluation dimensions, and target modification opinions are obtained by summarizing all the modification opinions. The target modification opinions cover all specific problem points of the evaluation dimensions whose evaluation levels are medium and poor, and corresponding specific modification instructions. The specific modification instructions include but are not limited to the specific operation mode of the modification (such as addition, deletion, modification, replacement, and replacement), the content of the modification, and the specific position of the modification.

[0052] In an implementation manner, the script evaluation model is pre-trained, and the training method includes: obtaining a plurality of training scripts corresponding to each script type respectively; For each of the training scripts, a second training evaluation standard is generated for the training script and the first training evaluation standard by a preset large language model; the training script is evaluated according to the second training evaluation standard to obtain training evaluation opinions; training modification opinions are generated based on the training evaluation opinions; and training data is generated according to the training script, the training evaluation opinions and the training modification opinions; The initial script evaluation model is fine-tuned according to all the training data, and the fine-tuned initial script evaluation model is used as the script evaluation model.

[0053] The initial script evaluation model refers to an untrained model, and a small or medium-sized basic model can be used as the initial script evaluation model to reduce the training cost. In this embodiment, the initial script evaluation model is fine-tuned by constructing training data that fits the actual scene, so that it has the ability to perfect the evaluation standard and the ability to generate accurate and specific evaluation opinions and modification opinions. Specifically, the evaluation logic of different script types is different. In order to ensure that the script evaluation model can adapt to the evaluation needs of different scenes, different types of training scripts need to be collected according to the script type. Taking a collected training script as an example, a preset large language model will update the first training evaluation standard according to the relevant information (such as the application scene) of the training script. The first training evaluation standard can be a known evaluation standard input into the large language model and artificially formulated based on industry experience. The first training evaluation standard can also be an evaluation standard defined by the large language model according to the training script. The first training evaluation standard includes one or more indicators for measuring the quality of the training script. The large language model can supplement and perfect the first training evaluation standard to obtain a more comprehensive second training evaluation standard. Further, the large language model will also evaluate the training script according to the second training evaluation standard and generate targeted modification opinions. In this embodiment, the evaluation opinions generated by the large language model in the training stage are defined as training evaluation opinions, and the modification opinions generated by the large language model in the training stage are defined as training modification opinions. The training script, the corresponding training evaluation opinions and the training modification opinions are integrated into a training data. Multiple training data are generated by collecting various training scripts to fine-tune the initial script evaluation model.

[0054] For example, a certain number of script data is collected as training scripts for different script types. A certain number of script data can be collected around a certain topic under a specific script type. For example, short video voiceover script, a variety of script data can be collected in different ways. For example, collect the voiceover content script of the real short video published on the network, or use the script data generated by the large model, or collect the script data obtained by imitating the real script.

[0055] An evaluation standard for different types of scripts is formulated. The evaluation standard can be formulated before training, for example, by a professional, or can be formulated at the beginning of training. The embodiment does not specifically limit the timing of formulating the evaluation standard. Some evaluation standards are general standards, which refer to evaluation dimensions suitable for all script types, including but not limited to: attractive beginning, clear logic, compliance, etc. Some evaluation standards are non-general standards, which refer to evaluation dimensions suitable for specific script types, including but not limited to: script-specific structure, information content, creativity, etc. For example, short video voice-over scripts require colloquial language and interactive endings.

[0056] For example, a general large language model (such as DeepSeek) is used as training data to build the model, and a small-scale model is used as an initial script evaluation model. The large language model first refines the evaluation standard for the corresponding script type based on the input training script, and then evaluates the training script based on the refined evaluation standard. The evaluation and modification opinions output by the large language model are generated based on prior knowledge, and therefore can be defined as distilled script evaluation data. The distilled script evaluation data is used to fine-tune the small-scale model. After fine-tuning, the script evaluation model is obtained. The script evaluation model obtained by fine-tuning will not use a unified evaluation standard when evaluating different types of training scripts, so the evaluation results will be more accurate. In actual application scenarios, the initial script evaluation model can use a small-scale model, such as Qwen2.5-7B-Instruct (model name). The small-scale model has high inference efficiency and lower training cost. Fine-tuning the small-scale model with distilled script evaluation data output by the large language model enables the small-scale model to have script evaluation capabilities comparable to or even exceeding those of the large model.

[0057] In one implementation, the second training evaluation standard is generated based on the training script and the first training evaluation standard, including: The first training evaluation standard is refined based on at least one of the script type, the topic, and the script content corresponding to the training script, to obtain the second training evaluation standard.

[0058] Specifically, to avoid the problem of incomplete first training evaluation standards, the embodiment relies on the prior knowledge of the large language model to refine the first training evaluation standard based on one or more of the script type, the topic, and the script content, to obtain the second training evaluation standard. The prior knowledge of the large language model contains more evaluation dimensions for different script types, so that the evaluation standard can have more evaluation dimensions, thereby injecting more understanding and insight into the evaluation system, and significantly improving the professionalism and accuracy of the subsequent script evaluation model in multiple evaluation dimensions.

[0059] In an implementation manner, the training script is evaluated according to the second training evaluation standard to obtain a training evaluation opinion, which comprises: The training script is evaluated according to each training evaluation dimension in the second training evaluation standard to obtain an evaluation grade and an evaluation reason corresponding to each training evaluation dimension, respectively; The training evaluation opinion is generated according to the evaluation grade and the evaluation reason of each training evaluation dimension.

[0060] Specifically, the second training evaluation standard contains specific multiple evaluation dimensions, and the evaluation dimensions contained in the second training evaluation standard are defined as training evaluation dimensions in this embodiment. The large language model will finely evaluate the training script according to the second training evaluation standard, and finally output the evaluation grade and the evaluation reason of each training evaluation dimension to form a comprehensive and targeted training evaluation opinion. The evaluation grade is a quality classification of the training evaluation dimension, which is used to intuitively reflect the performance of the training script in the evaluation dimension. A grading system of good, medium and poor can be used, or a score system can be used. The evaluation reason is an explanation of the evaluation grade, which needs to specifically explain the problems of the training script in the training evaluation dimension or the reasons for excellent performance, avoiding formal evaluation. The large language model will generate clear evaluation grade and evaluation reason for each training evaluation dimension, which can clearly and intuitively reflect the advantages and disadvantages of the training script in each training evaluation dimension, and provide a data basis for subsequent generation of modification opinions.

[0061] In an implementation manner, the training modification opinion is generated based on the training evaluation opinion, which comprises: A plurality of target training evaluation dimensions to be modified are determined based on the training evaluation opinion; For each target training evaluation dimension, a modification opinion is determined according to the evaluation grade and the evaluation reason corresponding to the target training evaluation dimension; The training modification opinion is generated according to the modification opinions of all the target training evaluation dimensions.

[0062] Specifically, the present embodiment uses a large language model to evaluate the training script and output training evaluation opinions. The training evaluation opinions have clearly indicated the performance of the training script in each training evaluation dimension, and one or more training evaluation dimensions need to be further selected as the target training evaluation dimension to be modified. For each target training evaluation dimension, the evaluation level of this dimension can reflect the severity of the problem, and the evaluation reason can clearly indicate the specific performance of the problem. The modification opinions need to be generated in combination with the evaluation level and the evaluation reason of this dimension to avoid generating vague modification opinions. The modification opinions of all target training evaluation dimensions are summarized, i.e., the training modification opinions are obtained. In actual training scenarios, the training modification opinions can include all specific problem points of the evaluation dimensions whose evaluation levels meet the preset levels, as well as corresponding specific modification instructions. The specific modification instructions include one or more instructions of the specific operation mode, the modified content, and the specific location of the modification. Thus, the purpose of the training modification opinions being clear and operable is achieved.

[0063] For example, the training evaluation opinions output by the large language model include the evaluation level and the evaluation reason of each training evaluation dimension in the second training evaluation standard. The evaluation level can be divided into three levels: good, medium, and poor. If the evaluation level of a training evaluation dimension is medium or poor, the large language model will give the corresponding modification opinions for this training evaluation dimension to modify the training script. The final output training modification opinions cover all specific problem points of the evaluation dimensions with medium or poor evaluation levels, and output specific modification instructions, including but not limited to: the specific operation mode of the modification (such as addition, deletion, modification, replacement, replacement, etc.), the modified content, and the specific location of the modification.

[0064] Specifically, the training script, the script type, and the evaluation standard corresponding to the script type (i.e., the first training evaluation standard) are used to construct the script evaluation prompt data (prompt) used by the large language model. Taking a short video voice-over script as an example, the large language model can use the script evaluation prompt data shown below for reasoning: Please act as a fair script evaluation professional and supplement and improve the evaluation standard of the script according to the application scenario of the script. Then, compare each item in the improved evaluation standard to carefully analyze whether the script generated by the topic selection meets the standard and explain the reasons. If there are any non-standard places, point them out and provide modification opinions.

[0065] You need to provide detailed analysis content in the thinking process, and the final output evaluation opinions (answer_content) should be consistent with the analysis content (reasoning_content) in the thinking process (think).

[0066] Script type: short video voiceover script; evaluation criteria; topic title; script content; Output format: output the final evaluation opinion in json list format, which must include each item in the improved evaluation criteria: {{"Evaluation criteria name": "an evaluation dimension in the Evaluation Criterias", "Evaluation level": "one of good, medium, and poor", "Reason for grading": "explain the reasons for grading the script according to each evaluation dimension", "Modification suggestions": "if the evaluation level is not good, please provide specific modification suggestions, using Chinese quotation marks \"\" to quote the content that needs to be modified"}}

[0067] As can be seen, the script evaluation prompt data guides the large language model to supplement and improve the evaluation criteria, and evaluates the training script according to the improved evaluation criteria. The evaluation criteria include multiple evaluation dimensions, and the large language model outputs the corresponding evaluation level and evaluation reason for each evaluation dimension, and further outputs modification suggestions for evaluation dimensions with medium and poor evaluation levels. Modification suggestions will cover all specific problem points of evaluation dimensions with medium and poor evaluation levels, and output specific modification instructions, including but not limited to: specific operation methods (such as adding, deleting, modifying, replacing, changing, etc.), modified content, and specific location. In actual application scenarios, the trained script evaluation model can use the same script evaluation prompt data as the large language model to guide the script evaluation model to output accurate and specific evaluation opinions and modification suggestions.

[0068] In one implementation, the training data is generated according to the training script, the training evaluation opinion, and the training modification opinion, including: several checks are performed on the training evaluation opinion and the training modification opinion; If all checks are passed, the training evaluation opinion and the training modification opinion are used as label data, and the training data is generated according to the training script and the label data.

[0069] Specifically, in order to ensure the quality of the training data, the training evaluation opinion and the training modification opinion output by the large language model need to pass one or more checks before they can be used as label data to evaluate the evaluation opinion and modification opinion output by the small and medium-sized models in the fine-tuning process. Each training script and its corresponding label data constitute a training data. It can be understood that in actual application scenarios, one or more checks can be selected according to task requirements, and the specific type and number of checks are not limited in this embodiment.

[0070] In an implementation manner, the training evaluation opinions and the training modification opinions are subjected to several checks, including: checking, by a preset check model, consistency of an inference path of the large language model with the training evaluation opinions; and / or, checking, by a preset check model, accuracy of an evaluation result of each evaluation dimension in the training evaluation opinions; and / or, checking, by a preset check model, consistency of an evaluation reason of each evaluation dimension in the training evaluation opinions with the training modification opinions; and / or, checking, by a preset check model, specificity of the training modification opinions.

[0071] Specifically, in order to improve the checking efficiency and accuracy, the check model can be used to complete one or more of the following checks: checking consistency of an inference path of the large language model with its output training evaluation opinions, i.e., checking whether the thinking process and the evaluation conclusion of the large language model are consistent; checking accuracy, consistency and specificity of each evaluation dimension. Among them, the accuracy refers to whether the evaluation level is accurate and reasonable; the consistency refers to whether the rating reason and the output modification opinions are consistent; and the specificity refers to whether the modification opinions are specific.

[0072] By checking one or more of the training evaluation opinions and the training modification opinions output by the large language model, the quality of the training data can be improved, so that the evaluation ability of the fine-tuned small and medium-sized model can be higher than that of the large language model, thereby obtaining a professional and advanced copywriting evaluation model that can provide accurate, detailed and specific evaluation opinions and modification opinions.

[0073] For example, the check model can use the ChatGPT model. When checking, the first check prompt data shown below can be used to guide the check model to check the consistency of the thinking process of the large language model with the training evaluation opinions: Below is the result of a large model evaluating a training copy generated by a topic according to evaluation standards, including the thinking process and the evaluation opinions. Please determine whether the thinking process and the evaluation opinions are consistent.

[0074] Topic {topic_title}; Copy title: {title}; Copy content: {content}; Thinking process {reasoning_content}; Final conclusion {answer_content}; Output format: Please output the judgment results and reasons in the following format, do not output any other irrelevant information; Is reasonable: output yes or no; Reason: elaborate the reasons.

[0075] When performing the check, the second check prompt data shown below can be used to guide the check model to check the accuracy, consistency, and specificity of the evaluation results for each evaluation dimension: Below is the result of a large model evaluating a certain dimension of a short video oral script generated by a topic according to evaluation criteria. Please judge the evaluation quality from three aspects.

[0076] Topic {topic_title}; Script: Title: {title}; Script content: {content}; Evaluation dimension {dimension}; Evaluation criteria {Evaluation Criterias (perfected)}; Evaluation opinion {opinion}; Modification suggestion {suggestion}; Evaluation requirements: Reasonableness: Is the evaluation level accurate? Consistency: Are the grading reasons and modification suggestions consistent? Specificity: Are the modification suggestions specific and operable? Output format: Reasonableness: output yes or no; Consistency: output yes or no; Specificity: output yes or no; Reason: Reasonableness: Detailed analysis of the reasonableness of the evaluation opinion; Consistency: Detailed analysis of the consistency of the modification suggestions and the evaluation opinion; Specificity: Detailed analysis of the specificity and operability of the modification suggestions.

[0077] In actual application scenarios, the key requirements for whether the modification suggestions are specific can include but are not limited to: must cover all specific problem points based on the evaluation opinion, and specify the specific modification content and modification location, so that the modification suggestions have higher operability and are convenient for users to read the modification suggestions.

[0078] In one implementation, after generating the target modification opinion, the method further includes: modifying the initial script based on the target modification opinion and the hierarchical modification principle to obtain a target script.

[0079] Specifically, the script modification model refers to a model that optimizes scripts according to modification opinions. This model can implement an automated script modification process according to modification opinions. In actual application scenarios, the script modification model will execute each modification opinion on the initial script according to one or more modification opinions and input hierarchical modification principles to achieve script modification, and finally generate a target script with higher quality.

[0080] For example, one or more improvement suggestions can be selected to adjust the initial script according to user preferences. A general large language model such as ChatGPT model can also be used as a script modification model. Using the script modification model, the initial script is modified according to the selected one or more improvement suggestions and the input hierarchical modification principle. The modified script data is defined as the target script.

[0081] The modification prompt data shown below can be used to guide the script modification model to modify the script: You are a professional scriptwriter. Please improve the following script according to the modification suggestions in the evaluation results.

[0082] Topic {topic_title}; Original script: title: {title}; content: {content}; Modification suggestions {improvement_suggestions}; The output format is: title: [title]; content: [modified content]; Please output the modified script directly without adding any explanations or instructions.

[0083] In an implementation, the initial script is modified based on the hierarchical modification principle according to several modification suggestions in the target modification suggestions, including: Several modification suggestions in the target modification suggestions are used as execution modification suggestions; For each execution modification suggestion, the modification priority is determined according to the evaluation level corresponding to the execution modification suggestion; Based on the hierarchical modification principle, the initial script is modified according to the modification priority of each execution modification suggestion.

[0084] Specifically, the target modification suggestions may contain multiple modification suggestions for different dimensions, but not all modification suggestions need to be executed, i.e. all modification suggestions can be executed, or part of the modification suggestions can be executed. Execution modification suggestion refers to the selected modification suggestion that must be executed from the target modification suggestion. These modification suggestions may play a key role in improving the quality of the script. By screening the execution modification suggestion, it can avoid low efficiency caused by too large modification range, and ensure that the modification work can be concentrated on the modification suggestions that can significantly improve the quality of the script. The evaluation level of the execution modification suggestion can usually reflect the severity of the problem, so the modification priority of each execution modification suggestion can be divided based on the evaluation level, and the initial script is modified by the modification priority of each execution modification suggestion, so that the modification work can be carried out in an orderly manner.

[0085] In an implementation manner, the initial script is modified according to the modification priority of each execution modification suggestion based on the hierarchical modification principle, comprising: The initial script is modified in sequence according to the modification priority of each execution modification suggestion; For each execution modification suggestion, if the data change of the key evaluation dimension exceeds the preset range when the initial script is modified according to the execution modification suggestion, the modification method and / or modification range of the execution modification suggestion is adjusted; the key evaluation dimension is an evaluation dimension whose evaluation level meets the preset requirement.

[0086] Specifically, the modification priority can be used to indicate the modification order. The execution modification suggestion with high modification priority is processed first, and the execution modification suggestion with low modification priority is delayed for processing, so that the modification work can be carried out in order. However, unlike the simple modification in sequence, the embodiment also needs to monitor the data change of other evaluation dimensions, especially the key evaluation dimension whose evaluation level meets the preset requirement, when each execution modification suggestion is executed. The key evaluation dimension can correspond to the advantageous part in the initial script. In view of the quality of the script, it is necessary to avoid large data changes in the key evaluation dimension. Therefore, if the implementation process of a certain execution modification suggestion causes a large data change in the key evaluation dimension, the modification method needs to be adjusted or the modification range needs to be reduced to avoid affecting the performance of the key evaluation dimension.

[0087] For example, when there are multiple execution modification suggestions, the script modification model will use the hierarchical modification method to complete the modification work. The hierarchical modification principle includes but is not limited to the following points: Advantage protection principle: taking the evaluation level containing three levels of good, medium and poor as an example, the evaluation dimension with the evaluation level of good is the existing advantage of the script, and the structure framework and order of the initial script need to be maintained during modification.

[0088] Hierarchical processing principle: different evaluation levels of evaluation dimensions have different modification priorities, corresponding to different modification orders. For example, the evaluation dimension with the evaluation level of poor has the highest modification priority and needs to be processed first; the evaluation dimension with the evaluation level of medium has the second modification priority and needs to be processed later; the evaluation dimension with the evaluation level of good has the lowest modification priority and needs to be processed last. In addition, different evaluation levels of evaluation dimensions have different modification ranges. For example, the evaluation dimension with the evaluation level of poor needs to be improved, the evaluation dimension with the evaluation level of medium needs to be improved moderately, and the evaluation dimension with the evaluation level of good does not need to be modified, but needs to be maintained.

[0089] Conflict handling principle: if a certain modification suggestion may damage the evaluation dimension with a good evaluation level, a milder modification method and / or a smaller modification range are selected to avoid damaging the evaluation dimension with a good evaluation level. If the evaluation dimension with a good evaluation level includes the dimension of logical clarity, the modification suggestion to be performed includes adding new content, but the new content may damage the original logical structure, and the modification is completed by integrating the new content into the structural framework of the initial script.

[0090] In an implementation manner, after the initial script is modified according to the several modification suggestions in the target modification suggestions based on the hierarchical modification principle through the preset script modification model to obtain the target script, the method further includes: inputting the target script into the script evaluation model to obtain updated evaluation suggestions and updated modification suggestions; judging whether the updated evaluation suggestions meet a termination condition, and if not, taking the target script as the initial script and taking the updated modification suggestions as the target modification suggestions; continuing to perform the step of modifying the initial script according to the several modification suggestions in the target modification suggestions based on the hierarchical modification principle through the preset script modification model until the updated evaluation suggestions meet the termination condition.

[0091] Specifically, the target script is a version modified by the script modification model, but it may not meet the user's requirements. In this embodiment, the target script is input into the script evaluation model again for secondary evaluation. Through the cyclic process from script evaluation to script modification and then to script evaluation, the script quality is continuously improved until the script data meeting the evaluation standard is obtained. In the secondary evaluation, the updated evaluation suggestions refer to the evaluation suggestions output by the script evaluation model based on the target script, and the performance of the target script in different evaluation dimensions can be known through the updated evaluation suggestions. The cyclic termination condition from script evaluation to script modification and then to script evaluation is a preset script quality standard condition. The updated evaluation suggestions are compared with the termination condition to judge whether the quality of the target script meets the standard. If the script evaluation does not meet the standard, a new round of script modification process is triggered; if the standard is met, the iteration is ended.

[0092] For example, the embodiment can simulate the process of writing a script by a person by using the script evaluation model and the script modification model: first, write a draft of the script, and then modify the part that is not good enough. The script evaluation model evaluates the initial script and outputs evaluation opinions and modification opinions. Specifically, the modification opinions are generated for the evaluation dimensions whose evaluation levels do not meet the requirements. The training data of the script evaluation model is constructed by distilling the script evaluation data, and the script evaluation model is obtained by training a small and medium-sized basic reasoning model. The script evaluation model can output evaluation opinions and modification opinions that are more suitable for different types of scripts, avoiding the use of a uniform evaluation standard to evaluate and modify the script. The script modification model modifies the initial script based on one or more selected modification opinions and directly outputs the modified script data, i.e., the target script. The script evaluation and script modification process can be repeated until the user's requirements for the script are met or the preset proportion of the evaluation dimensions meets the requirements, for example, when the evaluation levels of more than 90% of the evaluation dimensions are good.

[0093] In actual application scenarios, the script evaluation model can be used alone to evaluate or score an existing script to determine the quality of the script. Since the script evaluation model outputs specific modification opinions, the initial script can be improved according to the specific modification opinions. The script evaluation model and the script modification model can also be used simultaneously to automatically improve the script according to the specific modification opinions by using the script modification model, and a high-quality script can be obtained.

[0094] In an implementation manner, as shown in Figure 2 the script generation model, the script evaluation model, and the script modification model can be integrated into one model. This model can efficiently and high-quality realize the full-automatic script processing process from generating a draft of the script to script evaluation and script modification, thereby improving the efficiency and collaboration of script processing.

[0095] Based on the above embodiment, the application further provides a script modification device, as shown in Figure 3 the device includes: a script evaluation model 01, configured to generate a second evaluation standard for an initial script and a first evaluation standard; evaluate the initial script according to the second evaluation standard to obtain target evaluation opinions; generate target modification opinions based on the target evaluation opinions.

[0096] Based on the above embodiment, the application further provides a terminal, and the principle block diagram thereof can be as shown in Figure 4The terminal includes a processor, a memory, a network interface, and a display screen connected through a system bus. The processor of the terminal is configured to provide computing and control capabilities. The memory of the terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the terminal is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement the method for modifying a script. The display screen of the terminal can be a liquid crystal display screen or an electronic ink display screen.

[0097] Those skilled in the art can understand that Figure 4 The principle block diagram shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the terminal to which the scheme of the present application is applied. The specific terminal can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0098] In an implementation manner, the memory of the terminal stores more than one program, and is configured to execute the more than one program by more than one processor to include instructions for performing the method for modifying a script.

[0099] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. 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 above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0100] To sum up, the application discloses a copywriting modification method and device, a terminal and a storage medium, and relates to the technical field of artificial intelligence.The method comprises the following steps: generating a second evaluation standard for an initial copywriting and a first evaluation standard through a copywriting evaluation model; evaluating the initial copywriting according to the second evaluation standard to obtain target evaluation opinions; and generating target modification opinions based on the target evaluation opinions.The copywriting evaluation model of the application can supplement and improve the initial first evaluation standard to generate a more comprehensive second evaluation standard. Based on the second evaluation standard, the initial copywriting can be evaluated to output accurate target evaluation opinions to clearly indicate specific problems existing in the copywriting, and the target modification opinions can be generated according to the target evaluation opinions, so that the automatic modification of the copywriting can be realized. The problem that the manual modification level is difficult to keep consistent and the cost is high is solved.

[0101] It should be understood that the application of the present application is not limited to the above examples, and can be improved or changed according to the above description for those skilled in the art, and all these improvements and changes shall belong to the protection scope of the appended claims of the present application.

Claims

1. A copywriting modification method, characterized in that: The method comprises: Generate a second evaluation criterion based on the initial copy and the first evaluation criterion through the copy evaluation model; Evaluate the initial text according to the second evaluation criteria to obtain a target evaluation opinion; Generate target modification opinions based on the target evaluation opinions.

2. The method for modifying a copywriting according to claim 1, wherein: The method for obtaining the initial copy includes: Get the topic and copy type; The initial copy is generated according to the topic and the copy type through a preset copy generation model.

3. The method for modifying a copywriting according to claim 1, wherein: Generate a second evaluation criterion based on the initial copy and the first evaluation criterion, including: Based on at least one data of the copy type, the topic, and the copy content corresponding to the initial copy, the first evaluation standard is improved to obtain the second evaluation standard.

4. The method for modifying a copywriting according to claim 1, wherein: Evaluate the initial text according to the second evaluation criteria to obtain a target evaluation opinion, including: Evaluate the initial document according to each evaluation dimension in the second evaluation standard to obtain an evaluation grade and evaluation reason corresponding to each evaluation dimension; The target evaluation opinion is generated based on the evaluation level and evaluation reasons of each evaluation dimension.

5. The method for modifying a copywriting according to claim 4, wherein: Generating target modification opinions based on the target evaluation opinions, including: Determining several target evaluation dimensions to be modified based on the target evaluation opinions; For each target evaluation dimension, determine the modification opinion according to the evaluation level and evaluation reason corresponding to the target evaluation dimension; The target modification opinions are generated based on the modification opinions of all the target evaluation dimensions.

6. The method for modifying a copywriting according to claim 5, wherein: The target modification opinions include all specific problem points of the evaluation dimensions whose evaluation levels meet the preset levels, and the corresponding specific modification instructions; the specific modification instructions include: one or more instructions on the specific operation method of the modification, the content of the modification, and the specific location of the modification.

7. The method for modifying a copywriting according to claim 1, wherein: The copywriting evaluation model is pre-trained, and the training method includes: Obtain several training documents corresponding to each document type; For each of the training documents, generating a second training evaluation standard based on the training document and the first training evaluation standard using a preset large language model; evaluating the training document according to the second training evaluation standard to obtain a training evaluation opinion; generating a training modification opinion based on the training evaluation opinion; and generating training data based on the training document, the training evaluation opinion, and the training modification opinion; The initial copywriting evaluation model is fine-tuned according to all the training data, and the fine-tuned initial copywriting evaluation model is used as the copywriting evaluation model.

8. The method for modifying a copywriting according to claim 7, wherein: Generating a second training evaluation standard based on the training text and the first training evaluation standard includes: Based on at least one data of the copy type, topic and copy content corresponding to the training copy, the first training evaluation standard is improved to obtain the second training evaluation standard.

9. The method for modifying a copywriting according to claim 7, wherein: Evaluate the training document according to the second training evaluation standard to obtain a training evaluation opinion, including: Evaluate the training document according to each training evaluation dimension in the second training evaluation standard to obtain an evaluation grade and evaluation reason corresponding to each training evaluation dimension; The training evaluation opinion is generated according to the evaluation level and evaluation reason of each training evaluation dimension.

10. The method for modifying a document according to claim 9, wherein: Generating training modification opinions based on the training evaluation opinions includes: Determining a number of target training evaluation dimensions to be modified based on the training evaluation opinions; For each target training evaluation dimension, determining a modification suggestion according to the evaluation level and evaluation reason corresponding to the target training evaluation dimension; The training modification opinions are generated based on the modification opinions of all the target training evaluation dimensions.

11. The method for modifying a document according to claim 7, wherein: Generating training data according to the training document, the training evaluation opinion, and the training modification opinion includes: Conducting several checks on the training evaluation opinions and the training modification opinions; If all items are checked as qualified, the training evaluation opinion and the training modification opinion are used as label data, and the training data is generated according to the training text and the label data.

12. The method for modifying a copywriting according to claim 11, wherein: Conduct several checks on the training evaluation opinions and the training modification opinions, including: Checking the consistency between the inference path of the large language model and the training evaluation opinion through a preset checking model; and / or, checking the accuracy of the evaluation results of each evaluation dimension in the training evaluation opinion through a preset checking model; and / or, checking the consistency between the evaluation reasons for each evaluation dimension in the training evaluation opinion and the training modification opinion through a preset checking model; And / or, the specificity of the training modification suggestions is checked through a preset checking model.

13. The method for modifying a copywriting according to claim 1, wherein: After generating the target modification suggestion, the following steps are also included: Through a preset copy modification model, based on a hierarchical modification principle, the initial copy is modified according to several modification opinions in the target modification opinions to obtain a target copy.

14. The method for modifying a copywriting according to claim 13, wherein: Based on the principle of graded revision, the initial draft is revised according to several revision suggestions in the target revision suggestions, including: Taking some of the target revision opinions as implementation revision opinions; For each of the execution modification opinions, determining the modification priority according to the evaluation level corresponding to the execution modification opinion; Based on the hierarchical modification principle and the modification priority of each of the execution modification opinions, the initial text is modified.

15. The method for modifying a document according to claim 14, wherein: Based on the hierarchical modification principle and the modification priority of each of the execution modification opinions, the initial draft is modified, including: Modify the initial document in order according to the modification priority of each execution modification suggestion; For each of the execution modification opinions, if the data changes corresponding to the key evaluation dimensions exceed the preset range when the initial document is modified according to the execution modification opinions, the modification method and / or modification range of the execution modification opinions shall be adjusted; the key evaluation dimensions are evaluation dimensions whose evaluation levels meet the preset requirements.

16. The method for modifying a copywriting according to claim 13, wherein: The initial copy is modified according to several modification opinions in the target modification opinions based on the principle of hierarchical modification by using a preset copy modification model. After obtaining the target copy, the following steps are further included: Inputting the target copy into the copy evaluation model to obtain updated evaluation opinions and updated modification opinions; Determine whether the updated evaluation opinion meets the termination condition; if not, use the target document as the initial document and the updated modification opinion as the target modification opinion; Continue to execute the step of modifying the initial copy according to several modification opinions in the target modification opinions based on the hierarchical modification principle through the preset copy modification model until the updated evaluation opinion meets the termination condition.

17. A copywriting modification device, characterized in that: The device comprises: A copywriting evaluation model, used to generate a second evaluation criterion based on the initial copywriting and the first evaluation criterion; Evaluate the initial text according to the second evaluation criteria to obtain a target evaluation opinion; Generate target modification opinions based on the target evaluation opinions.

18. A terminal, characterized in that: The terminal includes a memory and one or more processors; the memory stores one or more programs; the programs include instructions for executing the document modification method according to any one of claims 1 to 16; and the processor is used to execute the programs.

19. A computer-readable storage medium having a plurality of instructions stored thereon, characterized in that: The instructions are suitable for being loaded and executed by a processor to implement the steps of the document modification method as described in any one of claims 1 to 16.

Images