Expert system-based ai writing plan aided decision-making platform construction method

Abstract

The application relates to the technical field of decision platform construction, in particular to an AI writing scheme auxiliary decision platform construction method based on an expert system, which comprises the following steps: obtaining knowledge features by preprocessing collected knowledge data, injecting the knowledge features into an initial model for training to obtain a scheme writing model; inputting input text and a predetermined text style into the scheme writing model to generate a writing scheme, optimizing the scheme writing model according to the writing scheme, the text style of the writing scheme and the knowledge features to obtain a writing scheme auxiliary decision platform; determining whether the attenuation coefficient of redundant information needs to be increased; determining whether the adaptive coefficient of feature extraction granularity needs to be increased; and determining a conflict feature fusing threshold. The application improves the construction accuracy of the auxiliary decision platform.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of decision platform construction, and in particular to an AI writing scheme auxiliary decision platform construction method based on an expert system. BACKGROUND

[0002] In the current era of deep integration of artificial intelligence and knowledge engineering and gradual development of intelligent decision technology as a key support capability, writing auxiliary decision demand is increasingly prominent in high-value text production scenarios such as scientific research papers, important documents and enterprise reports. The core value of the AI writing scheme auxiliary decision platform based on the expert system lies in that knowledge data can be injected into the large model training process in the form of a knowledge graph and logical rules, realizing the organic combination of rule-driven and data-driven, so as to generate a writing scheme that conforms to professional specifications and has dynamic adaptability, to solve the deficiencies of traditional writing large models in terms of professionalism, explainability and stability, promote intelligent writing from 'content generation' to 'knowledge decision', and have important practical significance and application prospect. However, the existing platform still has problems such as insufficient coordination between knowledge injection and model training, uneven granularity of writing feature extraction, and limited style consistency control, which cannot fully meet the demand for high accuracy and high reliability of writing results.

[0003] Chinese Patent Publication No. CN116611424A discloses a writing auxiliary model construction method, a writing auxiliary method and a storage medium, which comprises providing a first model and a second model, the first model comprising a decoder and an encoder, and adding a Cross-attention mechanism for learning controllable writing ability in the decoder; the second model has the same structure as the first model, and the parameter amount of the second model is less than that of the first model; the first model is trained by multi-task learning, which comprises a repeated generation punishment training method based on Self-Training, a style control training method combining Prompt and controllable Cross-attention mechanism, a sentence combination function training method combining Prompt and controllable Cross-attention mechanism, and a continuation training method based on the above content; based on the trained first model, the knowledge of the first model is distilled to the second model by using the technology of knowledge distillation based on curriculum learning to obtain a lightweight writing auxiliary model. As can be seen, the writing auxiliary model construction method, the writing auxiliary method and the storage medium have the problem that the combination of generation logic and knowledge data decision system is not close enough between knowledge injection and writing feature extraction, which leads to the problem of insufficient construction accuracy of the auxiliary decision platform in complex writing tasks. SUMMARY

[0004] To this end, the application provides an AI writing scheme auxiliary decision platform construction method based on an expert system, to overcome the problem that the combination of generation logic and knowledge data decision system is not close enough between knowledge injection and writing feature extraction in the prior art, thereby causing the problem of insufficient construction accuracy of the auxiliary decision platform in complex writing tasks.

[0005] To achieve the above-mentioned object, the application provides an AI writing scheme auxiliary decision platform construction method based on an expert system, comprising:

[0006] The collected knowledge data is preprocessed to obtain knowledge features, and the knowledge features are injected into an initial model for training to obtain a scheme writing model;

[0007] The input text and the predetermined text style are input into the scheme writing model to generate a writing scheme, and the scheme writing model is optimized according to the writing scheme, the text style of the writing scheme and the knowledge features to obtain a writing scheme auxiliary decision platform;

[0008] The generation error rate of the writing scheme is obtained, and whether the construction accuracy of the writing scheme auxiliary decision platform meets the requirements is determined based on the generation error rate of the writing scheme;

[0009] If the construction accuracy of the writing scheme auxiliary decision platform does not meet the requirements, it is determined whether the attenuation coefficient of the redundant information needs to be increased;

[0010] If the attenuation coefficient of the redundant information does not need to be increased, the deep feature coverage rate in the writing scheme is obtained to determine whether the collaborative adaptability of the knowledge data and the knowledge features meets the requirements;

[0011] If the collaborative adaptability of the knowledge data and the knowledge features does not meet the requirements, it is determined whether the adaptive coefficient of the feature extraction granularity needs to be increased;

[0012] If the adaptive coefficient of the feature extraction granularity does not need to be increased, the conflict feature blowout threshold is determined based on the style confusion times of the text style of the writing scheme.

[0013] Further, whether the construction accuracy of the writing scheme auxiliary decision platform meets the requirements is determined based on the generation error rate of the writing scheme, comprising:

[0014] The generation error rate of the writing scheme is compared with a preset first error rate;

[0015] If the generation error rate of the writing scheme is less than or equal to the preset first error rate, it is determined that the construction accuracy of the writing scheme auxiliary decision platform meets the requirements;

[0016] If the generation error rate of the writing scheme is greater than the preset first error rate, it is determined that the construction accuracy of the writing scheme auxiliary decision platform does not meet the requirements.

[0017] Further, it is determined whether the attenuation coefficient of the redundant information needs to be increased, including:

[0018] The generation error rate of the writing scheme is compared with the preset first error rate and the preset second error rate respectively;

[0019] If the generation error rate of the writing scheme is greater than the preset first error rate and less than or equal to the preset second error rate, it is determined that the attenuation coefficient of the redundant information does not need to be increased;

[0020] If the generation error rate of the writing scheme is greater than the preset second error rate, it is determined that the attenuation coefficient of the redundant information needs to be increased.

[0021] Further, the increase range of the attenuation coefficient of the redundant information is determined by the difference between the generation error rate of the writing scheme and the preset second error rate.

[0022] Further, based on the deep feature coverage rate in the writing scheme, it is determined whether the collaborative adaptability of the knowledge data and the knowledge feature meets the requirements, including:

[0023] The deep feature coverage rate in the writing scheme is compared with the preset second coverage rate;

[0024] If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, it is determined that the collaborative adaptability of the knowledge data and the knowledge feature meets the requirements, and it is determined whether the attenuation coefficient of the redundant information meets the requirements;

[0025] If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, it is determined that the collaborative adaptability of the knowledge data and the knowledge feature does not meet the requirements.

[0026] Further, it is determined whether the adaptive coefficient of the feature extraction granularity needs to be increased, including:

[0027] The deep feature coverage rate in the writing scheme is compared with the preset first coverage rate and the preset second coverage rate respectively;

[0028] If the deep feature coverage rate in the writing scheme is greater than the preset first coverage rate and less than or equal to the preset second coverage rate, it is determined that the adaptive coefficient of the feature extraction granularity needs to be increased, and the adaptive coefficient of the feature extraction granularity is increased;

[0029] If the deep feature coverage rate in the writing scheme is less than or equal to the preset first coverage rate, it is determined that the adaptive coefficient of the feature extraction granularity does not need to be increased.

[0030] Further, the increasing range of the adaptive coefficient of the feature extraction granularity is determined by the difference between the deep feature coverage in the writing scheme and the preset first coverage.

[0031] Further, the conflict feature fusing threshold is determined based on the style confusion times of the text style of the writing scheme, comprising:

[0032] Comparing the style confusion times of the text style of the writing scheme with a preset confusion times;

[0033] If the style confusion times of the text style of the writing scheme is less than or equal to the preset confusion times, it is determined that the collaborative dynamics of knowledge injection and model training meets the requirements, and there is no need to reduce the conflict feature fusing threshold, and whether the adaptive coefficient of the feature extraction granularity meets the requirements is determined;

[0034] If the style confusion times of the text style of the writing scheme is greater than the preset confusion times, it is determined that the collaborative dynamics of knowledge injection and model training does not meet the requirements, and the conflict feature fusing threshold needs to be reduced.

[0035] Further, the confusion times of the text style of the writing scheme is the ratio of the number of different text styles between the writing scheme and the predetermined text style to the total number of the writing scheme.

[0036] Further, the reducing range of the conflict feature fusing threshold is determined by the difference between the style confusion times of the text style of the writing scheme and the preset confusion times.

[0037] Compared with the prior art, the method has the beneficial effects that the attenuation coefficient of the redundant information is adjusted according to the generation error rate of the writing scheme, the data required for platform training and reasoning is often from multiple channels, the data from different sources has uneven quality and may contain irrelevant redundant content, thereby causing inaccurate platform construction, by increasing the attenuation coefficient of the redundant information, the weight weakening of the content determined as redundant is enhanced, that is, the redundant information is given a lower contribution in subsequent training or reasoning, thereby reducing the interference of the redundant information on platform training and writing scheme generation, and finally improving the accuracy of platform construction, the adaptive coefficient of the feature extraction granularity is adjusted according to the deep feature coverage rate in the writing scheme, since writing is a multi-dimensional complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, and ignores deep features, thereby causing the model to be under-fitted due to insufficient input features to describe the writing rules, by increasing the adaptive coefficient of the feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill the original feature blank, make the input features more complete to match the writing rules, and make the model preferentially capture high-correlation features, reduce the under-fitting performance such as logical break and scene misplacement caused by relying only on word frequency, and the conflict feature fusion threshold is adjusted according to the style confusion times of the text style in the writing scheme, since the knowledge data may be injected only once in the model initialization stage, and is not dynamically adjusted according to the data feedback in the training process, the model cannot reconcile the contradiction between knowledge and data, cannot learn formal expression or colloquial expression, and is under-fitted, by reducing the conflict feature fusion threshold, the conflict features exceeding the threshold can be determined as unacceptable, the training samples containing such high-conflict features are filtered, the model is prevented from learning the wrong style association from the data, and reasonable features with low conflict are preferentially retained, while unreasonable features with high conflict are fused, thereby helping the model to distinguish compatible general features from incompatible style conflict features, and improving the construction accuracy of the auxiliary decision-making platform.

[0038] Further, the method adjusts the attenuation coefficient of the redundant information by setting the preset first error rate and the preset second error rate, the data required for platform training and reasoning is often from multiple channels, the data from different sources has uneven quality and may contain irrelevant redundant content, thereby causing inaccurate platform construction, by increasing the attenuation coefficient of the redundant information, the weight weakening of the content determined as redundant is enhanced, that is, the redundant information is given a lower contribution in subsequent training or reasoning, thereby reducing the interference of the redundant information on platform training and writing scheme generation, and finally improving the accuracy of platform construction, and the construction accuracy of the auxiliary decision-making platform is further improved.

[0039] Further, the method sets the preset first coverage rate and the preset second coverage rate to adjust the adaptive coefficient of the feature extraction granularity, since writing is a multi-dimensional complex task, the feature extraction process may only focus on word frequency statistics or simple semantic similarity and other surface features, ignoring deep features, resulting in underfitting of the model due to insufficient input features to describe the writing rules, by increasing the adaptive coefficient of the feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill the original feature blank, make the input features more complete to match the writing rules, let the model preferentially capture high correlation features, reduce the underfitting performance of logical breakage, scene mispositioning and other underfitting performances caused by relying only on word frequency, and further improve the construction accuracy of the auxiliary decision-making platform.

[0040] Further, the method sets the preset confusion times to adjust the conflict feature fuse threshold, since the knowledge data may be injected only once in the model initialization stage, and not dynamically adjusted according to the data feedback in the training process, resulting in the model being unable to reconcile the contradiction between knowledge and data, neither learning formal expression nor learning colloquial expression, and falling into underfitting, by reducing the conflict feature fuse threshold, the conflict features exceeding the threshold can be determined as unacceptable, and the training samples containing such high conflict features are filtered, avoiding the model learning from the data. The wrong style association, and preferentially retaining low conflict reasonable features, while fusing high conflict unreasonable features, helping the model to distinguish compatible general features and incompatible style conflict features, and further improving the construction accuracy of the auxiliary decision-making platform. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The overall flowchart of the AI writing scheme auxiliary decision-making platform construction method based on the expert system of the embodiment of the application is shown in the figure.

[0042] Figure 2 The logic flowchart of the process of determining whether the adaptive coefficient of the feature extraction granularity needs to be increased in the AI writing scheme auxiliary decision-making platform construction method based on the expert system of the embodiment of the application is shown in the figure.

[0043] Figure 3 The logic flowchart of the process of determining whether the adaptive coefficient of the feature extraction granularity needs to be increased in the AI writing scheme auxiliary decision-making platform construction method based on the expert system of the embodiment of the application is shown in the figure.

[0044] Figure 4 The logic flowchart of the process of determining the conflict feature fuse threshold in the AI writing scheme auxiliary decision-making platform construction method based on the expert system of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0045] In order to make the objects and advantages of the present application clearer, the following further describes the present application with reference to examples; it should be understood that the specific examples described herein are merely intended to explain the present application, and are not intended to limit the present application.

[0046] The preferred embodiments of the present application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely intended to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0047] Please refer to Figure 1 , Figure 2 , Figure 3 and Figure 4 , which are respectively the overall flowchart of the AI writing scheme auxiliary decision platform construction method based on the expert system of the embodiments of the present application, the logic flowchart of the process of determining whether the attenuation coefficient of redundant information needs to be increased, the logic flowchart of the process of determining whether the adaptive coefficient of feature extraction granularity needs to be increased, and the logic flowchart of the process of determining the conflict feature fusing threshold. The AI writing scheme auxiliary decision platform construction method based on the expert system of the present application comprises:

[0048] Step S1, pre-processing the collected knowledge data to obtain knowledge features, and injecting the knowledge features into an initial model for training to obtain a scheme writing model;

[0049] Step S2, inputting an input text and a predetermined text style into the scheme writing model to generate a writing scheme, and optimizing the scheme writing model according to the writing scheme, the text style of the writing scheme, and the knowledge features to obtain a writing scheme auxiliary decision platform;

[0050] Step S3, obtaining the generation error rate of the writing scheme, and determining whether the construction accuracy of the writing scheme auxiliary decision platform meets the requirements based on the generation error rate of the writing scheme;

[0051] Step S4, if the construction accuracy of the writing scheme auxiliary decision platform does not meet the requirements, determining whether the attenuation coefficient of redundant information needs to be increased;

[0052] Step S5, if the attenuation coefficient of redundant information does not need to be increased, obtaining the deep feature coverage rate in the writing scheme to determine whether the collaborative adaptability of the knowledge data and the knowledge features meets the requirements;

[0053] Step S6, if the collaborative adaptability of the knowledge data and the knowledge features does not meet the requirements, determining whether the adaptive coefficient of feature extraction granularity needs to be increased;

[0054] In step S7, if the adaptive coefficient of increasing the feature extraction granularity is not needed, the conflict feature fusing threshold is determined based on the style confusion times of the text style of the writing scheme.

[0055] Specifically, the knowledge data is syntax type, industry term, and writing type rule.

[0056] Specifically, the preprocessing includes cleaning, deduplication, unified formatting, and feature extraction.

[0057] Specifically, the knowledge features include lexical style features, logical connection features, and domain concept features.

[0058] Specifically, the initial model is a pre-trained language model with basic language understanding and generation capabilities.

[0059] Specifically, the process of injecting knowledge features into the initial model for training to obtain a scheme writing model is to extract and induce corresponding logical rules by identifying the semantics of entities, relationships, and attributes in knowledge features, and then map these logical rules and knowledge features to the initial model as constraints during training, so that the model has writing ability and logical reasoning ability during iterative learning to obtain a scheme writing model.

[0060] Specifically, the decision model is a BERT model, a Seq2Seq model, and a Grammarly style adjustment model.

[0061] Specifically, the input text includes hosting scripts, learning plans, and drug names.

[0062] Specifically, the writing scheme is a hosting speech script, a learning content schedule, and a drug usage instruction.

[0063] Specifically, the predetermined text style is determined in advance according to the writing scene and purpose, and the style type required for writing scheme generation, including: formal academic style, narrative speech style, and legal document style.

[0064] Specifically, the text style of the writing scheme is the style type used in the generated writing scheme.

[0065] Specifically, the process of optimizing the scheme writing model according to the generated writing scheme, text style, and knowledge features is to analyze the generated writing scheme through logical rules to obtain the text style of the writing scheme, and to adjust the model parameters inversely according to the difference between the text style in the writing scheme and the predetermined text style, and to iteratively optimize the decision logic.

[0066] Specifically, the decay coefficient of redundant information is a parameter that measures the use of repeated and useless information in the input text by the writing model, that is, the larger the decay coefficient, the greater the decay degree of redundant information.

[0067] In particular, the adaptive coefficient of feature extraction granularity is a parameter used by the scheme-writing model to dynamically adjust the weight distribution between shallow features and deep features when invoking the text style.

[0068] In particular, the shallow features are formalized features that can be extracted from the text through basic statistical methods, including: word frequency statistics, sentence length, and lexical richness.

[0069] In particular, the deep features are abstract features that require a deep understanding of the semantics, logic, and overall structure of the text to extract, including: implied meaning, grammatical structure, and text coherence.

[0070] In particular, the conflict feature blowout threshold is the maximum value used to determine whether the conflict between knowledge data (such as formal expression rules, colloquial taboos) and knowledge features (such as style features in actual text) exceeds an acceptable range.

[0071] In implementation, the method adjusts the decay coefficient of redundant information according to the generation error rate of the writing scheme. Since the data required for platform training and reasoning often comes from multiple channels, the quality of data from different sources is uneven, and may contain irrelevant redundant content, resulting in inaccurate platform construction. By increasing the decay coefficient of redundant information, the weight of the content determined to be redundant can be weakened, that is, the redundant information is given a lower contribution in subsequent training or reasoning, thereby reducing the interference of redundant information on platform training and writing scheme generation, and finally improving the accuracy of platform construction. According to the adaptive coefficient of feature extraction granularity according to the deep feature coverage rate in the writing scheme. Since writing is a multi-dimensional complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, resulting in underfitting of the model due to insufficient input features to describe the writing rules. By increasing the adaptive coefficient of feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill the original feature gap, and let the input features more completely match the writing rules, so that the model can preferentially capture high-correlation features and reduce the underfitting performance caused by relying only on word frequency, such as logical breakage and scene misplacement. According to the conflict feature fusion threshold adjusted by the style confusion times of the text style of the writing scheme. Since the knowledge data may be injected only once in the model initialization stage, it is not dynamically adjusted according to data feedback in the training process, resulting in the model being unable to reconcile the contradiction between knowledge and data, failing to learn both formal expression and colloquial expression, and falling into underfitting. By reducing the conflict feature fusion threshold, conflict features exceeding the threshold can be determined as unacceptable, and training samples containing such high-conflict features can be filtered to avoid the model learning incorrect style associations from the data, and to preferentially retain low-conflict reasonable features while fusing high-conflict unreasonable features, helping the model to distinguish between compatible general features and incompatible style conflict features, and improving the construction accuracy of the auxiliary decision-making platform.

[0072] Specifically, whether the construction accuracy of the writing scheme auxiliary decision-making platform meets the requirements is determined based on the generation error rate of the writing scheme, comprising:

[0073] Comparing the generation error rate of the writing scheme with a preset first error rate;

[0074] If the generation error rate of the writing scheme is less than or equal to the preset first error rate, it is determined that the construction accuracy of the writing scheme auxiliary decision-making platform meets the requirements;

[0075] If the generation error rate of the writing scheme is greater than the preset first error rate, it is determined that the construction accuracy of the writing scheme auxiliary decision-making platform does not meet the requirements.

[0076] The reason that the construction accuracy of the writing scheme auxiliary decision-making platform does not meet the requirements can be that the collaborative adaptability of the knowledge data and the knowledge characteristics does not meet the requirements, or that the attenuation coefficient of the redundant information does not meet the requirements. Next, it is necessary to determine the specific reason, and the process of determining the specific reason is also the process of determining whether to increase the attenuation coefficient of the redundant information.

[0077] Specifically, it is determined whether the attenuation coefficient of the redundant information needs to be increased. This includes:

[0078] The generation error rate of the writing scheme is compared with the preset first error rate and the preset second error rate, respectively;

[0079] If the generation error rate of the writing scheme is greater than the preset first error rate and less than or equal to the preset second error rate, it is determined that the attenuation coefficient of the redundant information does not need to be increased;

[0080] If the generation error rate of the writing scheme is greater than the preset second error rate, it is determined that the attenuation coefficient of the redundant information needs to be increased.

[0081] When the generation error rate of the writing scheme is greater than the preset second error rate, it is determined that the reason that the construction accuracy of the writing scheme auxiliary decision-making platform does not meet the requirements is that the attenuation coefficient of the redundant information does not meet the requirements, and therefore the attenuation coefficient of the redundant information needs to be increased. When the generation error rate of the writing scheme is greater than the preset first error rate and less than or equal to the preset second error rate, it can be preliminarily determined that the collaborative adaptability of the knowledge data and the knowledge characteristics does not meet the requirements, and next it is necessary to finally determine whether the collaborative adaptability of the knowledge data and the knowledge characteristics meets the requirements according to the deep feature coverage rate in the writing scheme, that is, to determine whether the reason that the construction accuracy of the writing scheme auxiliary decision-making platform does not meet the requirements is that the collaborative adaptability of the knowledge data and the knowledge characteristics does not meet the requirements.

[0082] It can be understood that the preset first error rate is less than the preset second error rate, and the three intervals divided by the preset first error rate and the preset second error rate correspond to three situations, respectively:

[0083] The first interval is that the generation error rate of the writing scheme is less than or equal to the preset first error rate, and the corresponding situation is that the construction accuracy of the writing scheme auxiliary decision-making platform meets the requirements, and at this time no adjustment is needed;

[0084] The second interval is that the generation error rate of the writing scheme is greater than the preset first error rate and less than or equal to the preset second error rate, and the corresponding case is that, due to the multi-dimensional complexity of writing, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, resulting in underfitting of the model due to insufficient input features to describe the writing rules, at this time, it is necessary to further judge whether the collaborative adaptability of knowledge data and knowledge features meets the requirements;

[0085] The third interval is that the generation error rate of the writing scheme is greater than the preset second error rate, and the corresponding case is that, due to the data required for platform training and reasoning often coming from multiple channels, the data from different sources have uneven quality and may contain irrelevant redundant content, resulting in inaccurate platform construction, at this time, it is necessary to adjust the decay coefficient of redundant information.

[0086] It can be understood that, in the process of generating the writing scheme, the preset first error rate and the preset second error rate are used to represent the accuracy of the scheme generation, which is essentially a hierarchical control logic based on scheme effectiveness and knowledge fit degree, which avoids the one-sidedness of a single error threshold for scheme quality evaluation, and realizes the linkage of qualitative judgment of writing guidance rationality and dynamic optimization of rule engine, finally meets the dual needs of content accuracy and stability in the writing environment. The core role of the first error rate is to provide quantitative basis for the construction accuracy judgment of the writing scheme auxiliary decision-making platform; the core role of the second error rate is to determine whether the decay coefficient of redundant information needs to be adjusted. The preset first error rate and the preset second error rate can be set according to the actual working conditions. The setting of the preset first error rate and the preset second error rate aims to ensure the construction accuracy and practicality of the writing scheme auxiliary decision-making platform. Optionally, the preset first error rate and the second error rate are determined through a limited number of tests by evaluating the construction effect of different generation error rates on the writing scheme auxiliary decision-making platform. The determined preset first error rate and preset second error rate should meet the requirements that they cannot be too small and cannot cause too much interference to the construction process of the writing scheme auxiliary decision-making platform. Illustratively, the preset first error rate is generally selected in the range of [2%, 4%], and the preset second error rate is generally selected in the range of [7%, 9%].

[0087] Preferably, the preferred embodiment of the preset first error rate is 3%, and the preferred embodiment of the preset second error rate is 8%.

[0088] Specifically, the generation error rate of the writing scheme is the ratio of the number of generated writing scheme errors to the total number of writing schemes.

[0089] Specifically, the writing scheme generation error is that the generated writing scheme has problems such as inconsistency with the predetermined text style, format disorder, and logical contradiction.

[0090] Specifically, the increasing range of the attenuation coefficient of the redundant information is determined by the difference between the generation error rate of the writing scheme and the preset second error rate.

[0091] Specifically, when the difference between the generation error rate of the writing scheme and the preset second error rate is within 3%, the attenuation coefficient of the redundant information is increased to 1.2 times of the original value, and when the difference between the generation error rate of the writing scheme and the preset second error rate exceeds 3%, the attenuation coefficient of the redundant information is increased by 0.05 for every 0.5% increase on the basis of 1.2 times of the original value, for example, when the difference between the generation error rate of the writing scheme and the preset second error rate is 4% and the attenuation coefficient of the redundant information is 0.6, the increased attenuation coefficient of the redundant information is 0.6*1.2+0.05*2=0.82.

[0092] In implementation, the method of the present application adjusts the attenuation coefficient of the redundant information by setting the preset first error rate and the preset second error rate. Since the data required for platform training and reasoning often comes from multiple channels, the data from different sources has uneven quality and may contain irrelevant redundant content, which leads to inaccurate platform construction. By increasing the attenuation coefficient of the redundant information, the weight weakening of the content determined as redundant can be enhanced, that is, the redundant information is given a lower contribution in subsequent training or reasoning, thereby reducing the interference of redundant information on platform training and writing scheme generation, and ultimately improving the accuracy of platform construction, further improving the construction accuracy of the auxiliary decision-making platform.

[0093] Specifically, whether the collaborative adaptability of the knowledge data and the knowledge feature meets the requirements is determined based on the deep feature coverage rate in the writing scheme, including;

[0094] The deep feature coverage rate in the writing scheme is compared with the preset second coverage rate;

[0095] If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, it is determined that the collaborative adaptability of the knowledge data and the knowledge feature meets the requirements, and whether the attenuation coefficient of the redundant information meets the requirements is determined;

[0096] If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, it is determined that the collaborative adaptability of the knowledge data and the knowledge feature does not meet the requirements.

[0097] When the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, it is determined that the collaborative adaptability of the knowledge data and the knowledge feature meets the requirements, and the construction accuracy of the writing scheme auxiliary decision-making platform has been determined to not meet the requirements, so it is necessary to further determine whether the attenuation coefficient of the redundant information meets the requirements.

[0098] In the implementation, the decay coefficient of the redundancy information is determined to meet the requirements based on the comparison of the actual decay coefficient of the redundancy information and the predetermined decay coefficient threshold value, and if the actual decay coefficient of the redundancy information is less than the predetermined decay coefficient threshold value, it is determined that the decay coefficient of the redundancy information does not meet the requirements, wherein the predetermined decay coefficient threshold value is the average of the decay coefficient of the redundancy information monitored in the last three months of the historical period.

[0099] If the decay coefficient of the redundancy information does not meet the requirements, the decay coefficient of the redundancy information is increased; if the decay coefficient of the redundancy information meets the requirements, the generation error rate of the writing scheme is re-collected, and whether the construction accuracy of the writing scheme auxiliary decision platform meets the requirements is re-determined.

[0100] When the deep feature coverage in the writing scheme is less than or equal to the preset second coverage, it can be determined that the reason for the construction accuracy of the writing scheme auxiliary decision platform not meeting the requirements is that the collaborative adaptability of the knowledge data and the knowledge features does not meet the requirements, and the reason for the collaborative adaptability of the knowledge data and the knowledge features not meeting the requirements can be that the adaptive coefficient of the feature extraction granularity does not meet the requirements, or the collaborative dynamics of the knowledge injection and the model training does not meet the requirements. Next, it is necessary to determine which specific reason is, and the process of determining the specific reason is also the process of whether to increase the adaptive coefficient of the feature extraction granularity.

[0101] Specifically, determining whether the adaptive coefficient of the feature extraction granularity needs to be increased includes:

[0102] Comparing the deep feature coverage in the writing scheme with a preset first coverage and the preset second coverage, respectively;

[0103] If the deep feature coverage in the writing scheme is greater than the preset first coverage and less than or equal to the preset second coverage, it is determined that the adaptive coefficient of the feature extraction granularity needs to be increased, and the adaptive coefficient of the feature extraction granularity is increased;

[0104] If the deep feature coverage in the writing scheme is less than or equal to the preset first coverage, it is determined that the adaptive coefficient of the feature extraction granularity does not need to be increased.

[0105] When the deep feature coverage in the writing scheme is greater than the preset first coverage and less than or equal to the preset second coverage, it is determined that the reason why the collaborative adaptability of the knowledge data and the knowledge feature does not meet the requirement is that the adaptive coefficient of the feature extraction granularity does not meet the requirement, and therefore the adaptive coefficient of the feature extraction granularity needs to be increased. When the deep feature coverage in the writing scheme is less than or equal to the preset first coverage, it is preliminarily determined that the collaborative dynamics of the knowledge injection and the model training does not meet the requirement, and then it is necessary to finally determine whether the collaborative dynamics of the knowledge injection and the model training meets the requirement according to the style confusion times of the text style in the writing scheme, that is, whether the reason why the collaborative adaptability of the knowledge data and the knowledge feature does not meet the requirement is that the collaborative dynamics of the knowledge injection and the model training does not meet the requirement.

[0106] It can be understood that the preset first coverage is less than the preset second coverage, and the preset first coverage and the preset second coverage divide three intervals, which correspond to three situations respectively.

[0107] The first interval is that the deep feature coverage in the writing scheme is less than or equal to the preset first coverage, and the corresponding situation is that because the knowledge data may be injected only once in the model initialization stage and is not dynamically adjusted according to the data feedback in the training process, the model cannot reconcile the contradiction between knowledge and data, and cannot learn to express formally or colloquially, and falls into underfitting. At this time, it is necessary to further determine whether the collaborative dynamics of the knowledge injection and the model training meets the requirement.

[0108] The second interval is that the deep feature coverage in the writing scheme is greater than the preset first coverage and less than or equal to the preset second coverage, and the corresponding situation is that because writing is a multi-dimensional complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, resulting in underfitting of the model due to insufficient input features to describe the writing rules. At this time, it is necessary to adjust the adaptive coefficient of the feature extraction granularity.

[0109] The third interval is that the deep feature coverage in the writing scheme is greater than the preset second coverage, and the corresponding situation is that the collaborative adaptability of the knowledge data and the knowledge feature meets the requirement, and it is necessary to further determine whether the decay coefficient of the redundant information meets the requirement.

[0110] It can be understood that in the process of extracting writing features, the preset first coverage and the preset second coverage are introduced to represent the integrity and effectiveness of deep feature extraction. The essence is based on the hierarchical control logic of feature dimension and writing task demand, which avoids one-sided evaluation of model fitting by single threshold, realizes the linkage of qualitative judgment and self-adaptive adjustment of feature coverage sufficiency, and finally adapts to the dual needs of multi-dimensional features and model stability in writing environment. The core role of the first coverage is to provide quantitative basis for the self-adaptive coefficient of adjusting the feature extraction granularity in the feature extraction stage; the core role of the second coverage is to distinguish whether the collaborative adaptability of knowledge data and knowledge features meets the requirements. The preset first coverage and the preset second coverage can be set according to the actual working conditions. The setting of the preset first coverage and the preset second coverage aims to ensure the construction accuracy and practicability of the writing scheme auxiliary decision-making platform. Optionally, the preset first coverage and the preset second coverage are determined through limited experiments by evaluating the construction effect of different deep feature coverage on the writing scheme auxiliary decision-making platform. The determined preset first coverage and the preset second coverage should meet the requirements that they cannot be too small and cannot cause too much interference to the construction process of the writing scheme auxiliary decision-making platform. Illustratively, the preset first coverage is generally selected in the range of [76%, 84%], and the preset second coverage is generally selected in the range of [86%, 94%].

[0111] Preferably, the preferred embodiment of the preset first coverage is 80%, and the preferred embodiment of the preset second coverage is 90%.

[0112] Specifically, the deep feature coverage in the writing scheme is the ratio of the number of deep features to the total number of core deep features required by the writing task.

[0113] Specifically, the increase range of the self-adaptive coefficient of the feature extraction granularity is determined by the difference between the deep feature coverage in the writing scheme and the preset first coverage.

[0114] Specifically, when the difference between the deep feature coverage in the writing scheme and the preset first coverage is within 5%, the self-adaptive coefficient of the feature extraction granularity is increased to 1.1 times of the original value. When the difference between the deep feature coverage in the writing scheme and the preset first coverage exceeds 5%, the self-adaptive coefficient of the feature extraction granularity is increased by 0.03 for every 1% exceeding 5% on the basis of being increased to 1.1 times of the original value. For example, when the difference between the deep feature coverage in the writing scheme and the preset first coverage is 7%, and the current self-adaptive coefficient of the feature extraction granularity is 0.7, the increased self-adaptive coefficient of the feature extraction granularity is 0.7x1.1+0.03x2=0.83.

[0115] In implementation, the method of the present application adjusts the adaptive coefficient of feature extraction granularity by setting preset first coverage and preset second coverage. Since writing is a multi-dimensional complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, resulting in underfitting of the model due to insufficient input features to describe writing rules. By increasing the adaptive coefficient of feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill in the original feature blank, and make the input features more complete to match the writing rules, so that the model can preferentially capture high-correlation features, reduce the underfitting performance caused by relying only on word frequency, and further improve the construction accuracy of the auxiliary decision-making platform.

[0116] Specifically, the conflict feature blowout threshold is determined based on the style confusion times of the writing scheme Chinese text style, including:

[0117] The style confusion times of the writing scheme Chinese text style are compared with the preset confusion times;

[0118] If the style confusion times of the writing scheme Chinese text style are less than or equal to the preset confusion times, it is determined that the collaborative dynamics of knowledge injection and model training meet the requirements, and there is no need to reduce the conflict feature blowout threshold, and whether the adaptive coefficient of feature extraction granularity meets the requirements is determined;

[0119] If the style confusion times of the writing scheme Chinese text style are greater than the preset confusion times, it is determined that the collaborative dynamics of knowledge injection and model training do not meet the requirements, and the conflict feature blowout threshold needs to be reduced.

[0120] Wherein, when the style confusion times of the writing scheme Chinese text style are less than or equal to the preset confusion times, it is determined that the collaborative dynamics of knowledge injection and model training meet the requirements, and the collaborative adaptability of knowledge data and knowledge features has been determined not to meet the requirements, so it is necessary to further determine whether the adaptive coefficient of feature extraction granularity meets the requirements.

[0121] In implementation, whether the adaptive coefficient of feature extraction granularity meets the requirements is determined based on the comparison between the actual adaptive coefficient of feature extraction granularity and the predetermined adaptive coefficient threshold. If the actual adaptive coefficient of feature extraction granularity is less than the predetermined adaptive coefficient threshold, it is determined that the adaptive coefficient of feature extraction granularity does not meet the requirements, wherein the predetermined adaptive coefficient threshold is the average of the adaptive coefficient of feature extraction granularity monitored in the last three months before the method.

[0122] If the adaptive coefficient of the actual feature extraction granularity does not meet the requirement, the adaptive coefficient of the actual feature extraction granularity is increased; if the adaptive coefficient of the actual feature extraction granularity meets the requirement, the deep feature coverage in the writing scheme is re-collected, and whether the collaborative adaptability of the knowledge data and the knowledge feature meets the requirement is re-determined.

[0123] When the style confusion times of the writing scheme text style are greater than the preset confusion times, it can be determined that the reason why the collaborative adaptability of the knowledge data and the knowledge feature does not meet the requirement is that the collaborative dynamics of the knowledge injection and the model training does not meet the requirement, so it is necessary to reduce the conflict feature blowout threshold.

[0124] It can be understood that the two intervals divided by the preset confusion times correspond to two situations respectively.

[0125] The first interval is that the style confusion times of the writing scheme text style are less than or equal to the preset confusion times, and the corresponding situation is that the collaborative dynamics of the knowledge injection and the model training meets the requirement, at this time it is necessary to further determine whether the adaptive coefficient of the feature extraction granularity meets the requirement;

[0126] The second interval is that the style confusion times of the writing scheme text style are greater than the preset confusion times, and the corresponding situation is that since the knowledge data may be injected only once in the model initialization stage, it is not dynamically adjusted according to the data feedback in the training process, which leads to that the model cannot reconcile the contradiction between knowledge and data, neither can it learn formal expression nor can it learn colloquial expression, and it falls into underfitting, at this time it is necessary to adjust the conflict feature blowout threshold.

[0127] It can be understood that in the model training process, the preset confusion times are used to represent the tolerance and judgment standard of the text style conflict, which is essentially based on the hierarchical management and control logic of style consistency and knowledge-data collaborative dynamics, which not only avoids the one-sidedness of a single threshold for style deviation evaluation, but also realizes the qualitative judgment of model training stability and the dynamic linkage of knowledge injection strategy, and finally adapts to the dual needs of style uniformity and generation diversity in the writing scene. The core role of the preset confusion times is to serve as a quantitative basis for the model to distinguish between acceptable and unacceptable style deviations. The preset confusion times can be set according to the actual working conditions. The setting of the preset confusion times aims to ensure the construction accuracy and practicality of the writing scheme auxiliary decision-making platform. Optionally, the preset confusion times are determined through limited experiments by evaluating the construction effect of the writing scheme auxiliary decision-making platform with different text style confusion times. The determined preset confusion times should meet the requirements that they cannot be too small and cannot cause too much interference to the construction process of the writing scheme auxiliary decision-making platform. Exemplarily, the preset confusion times are generally selected in the range of [2 times, 4 times].

[0128] Preferably, the preferred embodiment of the preset confusion times is 3 times.

[0129] Specifically, the style confusion times of the writing scheme text style is the ratio of the number of times that the writing scheme text style is different from the predetermined text style to the total number of writing schemes.

[0130] Specifically, the reduction range of the conflict feature fusion threshold is determined by the difference between the style confusion times of the writing scheme text style and the preset confusion times.

[0131] Specifically, when the difference between the style confusion times of the text style and the preset confusion times is within 2 times, the conflict feature fusion threshold is reduced to 0.9 times of the original, and when the difference between the style confusion times of the text style and the preset confusion times exceeds 2 times, the conflict feature fusion threshold is reduced by 0.05 for each excess of 1 time on the basis of being reduced to 0.9 times of the original, for example, when the difference between the style confusion times of the text style and the preset confusion times is 4 times, the current conflict feature fusion threshold is 0.85, and the reduced conflict feature fusion threshold is 0.85*0.9-0.05*2=0.665.

[0132] In implementation, the method of the present application adjusts the conflict feature fusion threshold by setting the preset confusion times. Since the knowledge data may be injected only once in the model initialization stage and not dynamically adjusted according to data feedback in the training process, the model cannot reconcile the contradiction between knowledge and data, cannot learn formal expression or colloquial expression, and falls into underfitting. By reducing the conflict feature fusion threshold, the conflict features exceeding the threshold can be determined as unacceptable, the training samples containing such high-conflict features can be filtered, the model can be prevented from learning incorrect style associations from the data, and the low-conflict reasonable features can be preferentially retained while the high-conflict unreasonable features are fused, helping the model to distinguish between compatible general features and incompatible style conflict features, and further improving the construction accuracy of the auxiliary decision-making platform.

[0133] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the accompanying drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

Claims

1. A method for constructing an expert system-based AI writing plan aided decision platform, characterized in that, The method comprises the following steps: obtaining knowledge features by preprocessing the collected knowledge data, and injecting the knowledge features into an initial model to train the initial model to obtain a scheme writing model; inputting input text and a predetermined text style into the scheme writing model to generate a writing scheme, and optimizing the scheme writing model according to the writing scheme, a text style of the writing scheme, and the knowledge features to obtain a writing scheme assisted decision-making platform; obtaining a generation error rate of the writing scheme, and determining whether a construction accuracy of the writing scheme assisted decision-making platform meets a requirement based on the generation error rate of the writing scheme; if the construction accuracy of the writing scheme assisted decision-making platform does not meet the requirement, determining whether an attenuation coefficient of redundant information needs to be increased; if the attenuation coefficient of the redundant information does not need to be increased, obtaining a deep feature coverage rate in the writing scheme to determine whether a collaborative adaptability of the knowledge data and the knowledge features meets the requirement; if the collaborative adaptability of the knowledge data and the knowledge features does not meet the requirement, determining whether an adaptive coefficient of a feature extraction granularity needs to be increased; if the adaptive coefficient of the feature extraction granularity does not need to be increased, determining a conflict feature blowout threshold value based on a style confusion number of a text style of the writing scheme; the attenuation coefficient of the redundant information is a parameter for measuring the use of repeated useless information in the input text by the writing model, that is, the greater the attenuation coefficient, the greater the attenuation degree of the redundant information; the adaptive coefficient of the feature extraction granularity is a parameter for dynamically adjusting the weight distribution between the shallow features and the deep features when the text style is called by the scheme writing model; the shallow features are formalized features that can be extracted by basic statistical methods, including word frequency statistics, sentence length, and vocabulary richness; the deep features are abstract features that can be extracted by deeply understanding the semantics, logic, and overall structure of the text, including implied meaning, grammatical structure, and text coherence; the conflict feature blowout threshold value is a maximum value for determining whether the conflict degree between the knowledge data and the knowledge features exceeds an acceptable range; the generation error rate of the writing scheme is a ratio of the number of errors in the generated writing scheme to the total number of the writing schemes; the writing scheme generation error is a problem of inconsistency with the predetermined text style, format disorder, and logical contradiction in the generated writing scheme; the style confusion number of the text style of the writing scheme is a ratio of the number of different text styles of the writing scheme to the total number of the writing schemes; the reduction range of the conflict feature blowout threshold value is determined by the difference between the style confusion number of the text style of the writing scheme and a preset confusion number.

2. The expert system based AI writing protocol aided decision platform building method according to claim 1, characterized in that, The method for determining whether the construction accuracy of the writing scheme assisted decision-making platform meets the requirement based on the generation error rate of the writing scheme comprises the following steps: comparing the generation error rate of the writing scheme with a preset first error rate; if the generation error rate of the writing scheme is less than or equal to the preset first error rate, it is determined that the construction accuracy of the writing scheme assisted decision-making platform meets the requirement; if the generation error rate of the writing scheme is greater than the preset first error rate, it is determined that the construction accuracy of the writing scheme assisted decision-making platform does not meet the requirement.

3. The expert system based AI writing protocol aided decision platform building method according to claim 2, characterized in that, The method for determining whether the attenuation coefficient of the redundant information needs to be increased comprises the following steps: Compare the generation error rate of the writing scheme with the preset first error rate and the preset second error rate, respectively. If the generation error rate of the writing scheme is greater than the preset second error rate, it is determined that the attenuation coefficient of the redundant information needs to be increased. If the generation error rate of the writing scheme is greater than the preset first error rate and less than or equal to the preset second error rate, it is determined that the attenuation coefficient of the redundant information does not need to be increased.

4. The expert system based AI writing protocol aided decision platform building method according to claim 3, characterized in that, The increase range of the attenuation coefficient of the redundant information is determined by the difference between the generation error rate of the writing scheme and the preset second error rate.

5. The expert system based AI writing protocol aided decision platform building method according to claim 4, characterized in that, Determine whether the collaborative adaptability of knowledge data and knowledge features meets the requirements based on the deep feature coverage rate in the writing scheme, including: Compare the deep feature coverage rate in the writing scheme with the preset second coverage rate. If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, it is determined that the collaborative adaptability of knowledge data and knowledge features meets the requirements, and whether the attenuation coefficient of the redundant information meets the requirements is determined. If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, it is determined that the collaborative adaptability of knowledge data and knowledge features does not meet the requirements.

6. The expert system based AI writing protocol aided decision platform building method according to claim 5, characterized in that, Determine whether the adaptive coefficient of feature extraction granularity needs to be increased, including: Compare the deep feature coverage rate in the writing scheme with the preset first coverage rate and the preset second coverage rate, respectively. If the deep feature coverage rate in the writing scheme is greater than the preset first coverage rate and less than or equal to the preset second coverage rate, it is determined that the adaptive coefficient of feature extraction granularity needs to be increased. If the deep feature coverage rate in the writing scheme is less than or equal to the preset first coverage rate, it is determined that the adaptive coefficient of feature extraction granularity does not need to be increased.

7. The expert system based AI writing protocol aided decision platform building method according to claim 6, characterized in that, The increase range of the adaptive coefficient of feature extraction granularity is determined by the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate.

8. The expert system based AI writing protocol aided decision platform building method according to claim 7, characterized in that, Determine the conflict feature blowout threshold based on the style confusion number of the text style in the writing scheme, including: Compare the style confusion number of the text style in the writing scheme with the preset confusion number. If the style confusion number of the text style in the writing scheme is less than or equal to the preset confusion number, it is determined that the collaborative dynamics of knowledge feature injection and model training meets the requirements, the conflict feature blowout threshold does not need to be reduced, and whether the adaptive coefficient of feature extraction granularity meets the requirements is determined. If the style confusion number of the text style in the writing scheme is greater than the preset confusion number, it is determined that the collaborative dynamics of knowledge injection and model training does not meet the requirements, and the conflict feature blowout threshold needs to be reduced.

Images