AI writing scheme auxiliary decision-making platform construction method based on expert system

Abstract

The invention relates to the technical field of decision-making platform construction, in particular to an AI writing scheme auxiliary decision-making platform construction method based on an expert system, and the method comprises the steps: carrying out the preprocessing of collected knowledge data, obtaining knowledge features, injecting the knowledge features into an initial model, and carrying out the training, so as to obtain a scheme writing model; 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-making platform; determining whether the attenuation coefficient of the redundant information needs to be increased; determining whether the adaptive coefficient of the feature extraction granularity needs to be increased or not; and determining a conflict feature fusing threshold. According to the invention, the construction accuracy of the auxiliary decision-making platform is improved.

Description

Technical Field

[0001] This invention relates to the field of decision-making platform construction technology, and in particular to a method for constructing an AI-assisted decision-making platform based on an expert system-based writing scheme. Background Technology

[0002] In an era where artificial intelligence and knowledge engineering are deeply integrated, and intelligent decision-making technology is gradually becoming a key supporting capability, the need for writing-assisted decision-making is increasingly prominent in high-value text production scenarios such as scientific research papers, important documents, and corporate reports. The core value of AI-assisted decision-making platforms for writing based on expert systems lies in their ability to inject knowledge data into the training process of large models in the form of knowledge graphs and logical rules. This achieves an organic combination of rule-driven and data-driven approaches, generating writing solutions that are both professionally compliant and dynamically adaptable. This addresses the shortcomings of traditional large-scale writing models in terms of professionalism, interpretability, and stability, driving the upgrade of intelligent writing from "content generation" to "knowledge-based decision-making," and has significant practical implications and application prospects. However, existing platforms still generally suffer from insufficient coordination between knowledge injection and model training, uneven granularity of writing feature extraction, and limited control over style consistency, making it difficult to fully meet the demands for high accuracy and reliability in writing results.

[0003] Chinese Patent Publication No. CN116611424A discloses a method for constructing a writing assistance model, a writing assistance method, and a storage medium. The method includes providing a first model and a second model. The first model includes a decoder and an encoder, and a Cross-attention mechanism for learning controllable writing capabilities is added to the decoder. The second model has the same structure as the first model, but with fewer parameters. The first model is trained using multi-task learning, which includes a self-training-based repetition generation penalty training method, a style control training method combining Prompt and a controllable Cross-attention mechanism, a sentence-building function training method combining Prompt and a controllable Cross-attention mechanism, and a continuation-writing training method based on the preceding text. Based on the trained first model, knowledge distillation technology based on curriculum learning is used to distill the knowledge of the first model into the second model to obtain a lightweight writing assistance model. Therefore, it can be seen that the writing assistance model construction method, writing assistance method and storage medium mentioned above have the problem of insufficient accuracy in constructing the auxiliary decision platform in complex writing tasks due to the lack of generation logic between knowledge injection and writing feature extraction and the loose integration with the knowledge data decision system. Summary of the Invention

[0004] To address this, the present invention provides a method for constructing an AI-assisted decision-making platform for writing based on an expert system. This method overcomes the problem in existing technologies where the lack of generation logic between knowledge injection and writing feature extraction, coupled with the loose integration of knowledge data decision-making systems, leads to insufficient accuracy in constructing the auxiliary decision-making platform for complex writing tasks.

[0005] To achieve the above objectives, this invention provides a method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems, comprising: The collected knowledge data is preprocessed to obtain knowledge features, which are then injected into the initial model for training to obtain the solution writing model. Input text and a predetermined text style are input into the solution writing model to generate a writing solution. The solution writing model is then optimized based on the writing solution, the text style of the writing solution, and the knowledge features to obtain a writing solution auxiliary decision-making platform. Obtain the error rate of the generated writing plan, and determine whether the accuracy of the construction of the writing plan auxiliary decision-making platform meets the requirements based on the error rate of the generated writing plan; If the accuracy of the writing scheme-assisted decision-making platform does not meet the requirements, then it is necessary to determine whether the attenuation coefficient of redundant information needs to be increased. If it is not necessary to increase the attenuation coefficient of redundant information, then obtain the deep feature coverage in the writing scheme to determine whether the synergistic adaptability of knowledge data and knowledge features meets the requirements. If the collaborative adaptability between the knowledge data and knowledge features does not meet the requirements, then determine whether it is necessary to increase the adaptive coefficient of the feature extraction granularity. If there is no need to increase the adaptive coefficient of feature extraction granularity, the conflict feature circuit breaker threshold is determined based on the number of style confusions in the text style in the writing scheme.

[0006] Furthermore, based on the error rate of the generated writing scheme, it is determined whether the accuracy of the construction of the writing scheme auxiliary decision-making platform meets the requirements, including: Compare the error rate of the generated writing plan with the preset first error rate; If the error rate of the writing scheme generation is less than or equal to the preset first error rate, then the accuracy of the construction of the writing scheme auxiliary decision-making platform is determined to meet the requirements. If the error rate of the generated writing scheme is greater than the preset first error rate, then it is determined that the accuracy of the construction of the writing scheme auxiliary decision-making platform does not meet the requirements.

[0007] Further, determine whether it is necessary to increase the attenuation coefficient of redundant information. This includes: The error rate of the writing scheme is compared with the preset first error rate and the preset second error rate, respectively; If the 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, then it is determined that there is no need to increase the attenuation coefficient of redundant information. If the error rate of the writing scheme is greater than the preset second error rate, then it is determined that the attenuation coefficient of redundant information needs to be increased.

[0008] Furthermore, the increase in 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.

[0009] Furthermore, based on the deep feature coverage in the writing scheme, it is determined whether the collaborative adaptability of knowledge data and knowledge features meets the requirements, including; The deep feature coverage rate in the writing scheme is compared with the preset second coverage rate; If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, then it is determined that the collaborative adaptability of knowledge data and knowledge features meets the requirements, and it is determined whether the attenuation coefficient of redundant information meets the requirements. If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, then the collaborative adaptability between knowledge data and knowledge features is determined to be unsatisfactory.

[0010] Further, determine whether the adaptive coefficient for increasing the granularity of feature extraction needs to be adjusted, including: The deep feature coverage rate in the writing scheme is compared 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, then it is determined that the adaptive coefficient of feature extraction granularity needs to be increased, and 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, then it is determined that there is no need to increase the adaptive coefficient of feature extraction granularity.

[0011] Furthermore, the increase in the adaptive coefficient of the feature extraction granularity is determined by the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate.

[0012] Furthermore, the conflict feature circuit breaker threshold is determined based on the number of style confusions in the writing scheme, including: Compare the number of style confusions in the text style of the writing scheme with the preset number of confusions; If the number of style confusions in the writing scheme is less than or equal to the preset number of confusions, then the collaborative dynamics of knowledge injection and model training meet the requirements, there is no need to reduce the conflict feature circuit breaker threshold, and it is determined whether the adaptive coefficient of the feature extraction granularity meets the requirements. If the number of style confusions in the writing scheme is greater than the preset number of confusions, then the collaborative dynamics of knowledge injection and model training are not met, and the conflict feature circuit breaker threshold needs to be reduced.

[0013] Furthermore, the number of text style confusions in the writing scheme is the ratio of the number of different text styles in the writing scheme to the total number of writing schemes.

[0014] Furthermore, the reduction in the conflict feature circuit breaker threshold is determined by the difference between the number of style confusions in the text style in the writing scheme and the preset number of confusions.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: The method adjusts the attenuation coefficient of redundant information based on the error rate of the writing scheme generation. Since the data required for platform training and inference often comes from multiple channels, and the quality of data from different sources varies, it may contain irrelevant redundant content, leading to inaccurate platform construction. By increasing the attenuation coefficient of redundant information, the weight reduction of content judged as redundant is enhanced, meaning that redundant information is assigned a lower contribution in subsequent training or inference, thereby reducing the interference of redundant information on platform training and writing scheme generation, ultimately improving the accuracy of platform construction. Furthermore, the method adjusts the adaptive coefficient of feature extraction granularity based on the deep feature coverage in the writing scheme. Since writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, causing the model to underfit due to insufficient input features to describe the writing pattern. By increasing the adaptive coefficient of feature extraction granularity, the attenuation coefficient can be reduced. This approach enables the model to automatically expand the dimensions of feature extraction, filling in existing feature gaps and allowing input features to more completely match writing patterns. It prioritizes capturing highly relevant features, reducing underfitting issues such as logical breaks and scene misalignment caused by relying solely on word frequency. The conflict feature fusing threshold is adjusted based on the number of style confusions in the writing scheme. Since knowledge data may only be injected once during model initialization and not dynamically adjusted based on data feedback during training, the model cannot reconcile the contradictions between knowledge and data, failing to learn either formal or colloquial expressions, leading to underfitting. By reducing the conflict feature fusing threshold, conflict features exceeding the threshold are deemed unacceptable, filtering training samples containing such high-conflict features. This prevents the model from learning incorrect style associations from the data and prioritizes retaining reasonable features with low conflict while fusing unreasonable features with high conflict. This helps the model distinguish between compatible general features and incompatible style conflict features, improving the accuracy of the auxiliary decision-making platform.

[0016] Furthermore, the method described in this invention adjusts the attenuation coefficient of redundant information by setting a preset first error rate and a preset second error rate. Since the data required for platform training and inference often comes from multiple channels, and the quality of data from different sources varies, it may contain irrelevant redundant content, which may lead to inaccurate platform construction. By increasing the attenuation coefficient of redundant information, the weight reduction of content judged as redundant can be enhanced. That is, redundant information is given a lower contribution in subsequent training or inference, 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.

[0017] Furthermore, the method described in this invention adjusts the adaptive coefficient of feature extraction granularity by setting a preset first coverage rate and a preset second coverage rate. Since writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features. This can lead to underfitting of the model because the input features are insufficient to describe the writing pattern. By increasing the adaptive coefficient of feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill in the original feature gaps, and allow the input features to more completely match the writing pattern. This allows the model to prioritize capturing highly relevant features, reducing underfitting performance such as logical breaks and scene misalignment caused by relying solely on word frequency, and further improving the accuracy of constructing the auxiliary decision-making platform.

[0018] Furthermore, the method described in this invention adjusts the conflict feature circuit breaker threshold by setting a preset number of confusions. Since knowledge data may only be injected once during the model initialization stage and not dynamically adjusted according to data feedback during training, the model cannot reconcile the contradiction between knowledge and data, and thus cannot learn either formal or colloquial expressions, resulting in underfitting. By reducing the conflict feature circuit breaker threshold, conflict features exceeding the threshold can be judged as unacceptable, filtering training samples containing such high-conflict features, preventing the model from learning incorrect style associations from the data, and prioritizing the retention of reasonable features with low conflict. At the same time, it circuit breaks down unreasonable features with high conflict, helping the model distinguish between compatible general features and incompatible style conflict features, further improving the accuracy of constructing the auxiliary decision-making platform. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the method for constructing an AI-assisted decision-making platform for writing based on an expert system, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of determining whether to increase the attenuation coefficient of redundant information in the construction method of the AI ​​writing scheme auxiliary decision-making platform based on the expert system according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of determining whether to increase the adaptive coefficient of feature extraction granularity in the construction method of the AI ​​writing scheme auxiliary decision-making platform based on the expert system according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the process of determining the conflict feature circuit breaker threshold in the method for constructing an AI writing scheme-assisted decision-making platform based on an expert system, as described in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] Please see Figure 1 , Figure 2 , Figure 3 as well as Figure 4 The diagrams shown are, respectively, the overall flowchart of the AI ​​writing scheme-assisted decision-making platform construction method based on an expert system according to an embodiment of the present invention, the logical flowchart of the process of determining whether to increase the attenuation coefficient of redundant information, the logical flowchart of the process of determining whether to increase the adaptive coefficient of feature extraction granularity, and the logical flowchart of the process of determining the conflict feature circuit breaker threshold. The present invention provides a method for constructing an AI writing scheme-assisted decision-making platform based on an expert system, comprising: Step S1: The collected knowledge data is preprocessed to obtain knowledge features, and the knowledge features are injected into the initial model for training to obtain the scheme writing model; Step S2: Input the input text and the predetermined text style into the solution writing model to generate a writing solution. Optimize the solution writing model based on the writing solution, the text style of the writing solution, and the knowledge features to obtain a writing solution auxiliary decision-making platform. Step S3: Obtain the error rate of the writing scheme generation, and determine whether the accuracy of the writing scheme auxiliary decision-making platform construction meets the requirements based on the error rate of the writing scheme generation. Step S4: If the accuracy of the construction of the writing scheme-assisted decision-making platform does not meet the requirements, determine whether it is necessary to increase the attenuation coefficient of redundant information. Step S5: If it is not necessary to increase the attenuation coefficient of redundant information, then obtain the deep feature coverage in the writing scheme to determine whether the collaborative adaptability of knowledge data and knowledge features meets the requirements. Step S6: If the collaborative adaptability between the knowledge data and knowledge features does not meet the requirements, determine whether it is necessary to increase the adaptive coefficient of the feature extraction granularity. Step S7: If it is not necessary to increase the adaptive coefficient of feature extraction granularity, then determine the conflict feature circuit breaker threshold based on the number of style confusions in the text style in the writing scheme.

[0023] Specifically, the knowledge data includes grammar types, industry terminology, and writing style rules.

[0024] Specifically, preprocessing includes cleaning, deduplication, uniform formatting, and feature extraction.

[0025] Specifically, knowledge features include lexical style features, logical connection features, and domain concept features.

[0026] Specifically, the initial model is a pre-trained language model with basic language understanding and generation capabilities.

[0027] Specifically, the process of injecting knowledge features into the initial model for training to obtain the solution writing model involves identifying the semantics of entities, relationships, and attributes in the knowledge features, extracting and summarizing the corresponding logical rules, and then mapping these logical rules and knowledge features into the initial model as constraints during training. This enables the model to possess writing ability and logical reasoning ability during iterative learning, thus obtaining the solution writing model.

[0028] Specifically, the decision-making models are the BERT model, the Seq2Seq model, and Grammarly's style adjustment model.

[0029] Specifically, the input text includes the host's script, study plan, and drug name.

[0030] Specifically, the writing plan includes a speech for hosting a meeting, a schedule of learning content, and instructions for use of medicines.

[0031] Specifically, the pre-determined text style is the style type that needs to be generated by the writing plan, based on the writing scenario and purpose. This includes formal academic style, narrative speech style, and legal document style.

[0032] Specifically, the text style of the writing scheme is the style type used in the already generated writing scheme.

[0033] Specifically, the process of optimizing the writing model based on the generated writing scheme, text style, and knowledge features involves analyzing the generated writing scheme through logical rules to obtain the text style of the writing scheme, adjusting the model parameters in reverse based on the difference between the text style in the writing scheme and the predetermined text style, and iteratively optimizing the decision logic.

[0034] Specifically, the redundancy decay coefficient is a parameter that measures the amount of redundant and useless information used by the writing model in the input text. That is, the larger the decay coefficient, the greater the decay of redundant information.

[0035] Specifically, the adaptive coefficient of feature extraction granularity is a parameter used by the scheme writing model to dynamically adjust the weight distribution between shallow and deep features when invoking text style.

[0036] Specifically, shallow features are formal features in the text that can be extracted using basic statistical methods, including word frequency statistics, sentence length, and lexical richness.

[0037] Specifically, deep features are abstract features that require a deep understanding of the semantics, logic, and overall structure of the text to extract, including: implicit meaning, grammatical structure, and textual coherence.

[0038] Specifically, the conflict feature circuit breaker threshold is the maximum value used to determine whether the degree of conflict between knowledge data (such as formal expression rules and colloquial taboos) and knowledge features (such as stylistic features in actual texts) exceeds the acceptable range.

[0039] In implementation, the method of this invention adjusts the attenuation coefficient of redundant information based on the error rate of writing scheme generation. Since the data required for platform training and inference often comes from multiple channels, and the quality of data from different sources varies, it may contain irrelevant redundant content, leading to inaccurate platform construction. By increasing the attenuation coefficient of redundant information, the weight reduction of content judged as redundant is strengthened, meaning that redundant information is given a lower contribution in subsequent training or inference, thereby reducing the interference of redundant information on platform training and writing scheme generation, ultimately improving the accuracy of platform construction. The method also adjusts the adaptive coefficient of feature extraction granularity based on the deep feature coverage in the writing scheme. Because writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features. This can cause the model to underfit due to insufficient input features to describe the writing pattern. By increasing the adaptive coefficient of feature extraction granularity, the model can automatically expand... The feature extraction dimension fills in the original feature gaps, allowing input features to more completely match writing patterns. This enables the model to prioritize capturing highly relevant features, reducing underfitting issues such as logical breaks and scene misalignment caused by relying solely on word frequency. The conflict feature fusing threshold is adjusted based on the number of style confusions in the writing scheme. Since knowledge data may only be injected once during model initialization and not dynamically adjusted based on data feedback during training, the model cannot reconcile the contradictions between knowledge and data, failing to learn both formal and colloquial expressions, thus falling into underfitting. By reducing the conflict feature fusing threshold, conflict features exceeding the threshold can be judged as unacceptable, filtering training samples containing such high-conflict features. This prevents the model from learning incorrect style associations from the data and prioritizes retaining reasonable features with low conflict while fusing unreasonable features with high conflict. This helps the model distinguish between compatible general features and incompatible style conflict features, improving the accuracy of the auxiliary decision-making platform construction.

[0040] Specifically, determining whether the accuracy of the writing scheme auxiliary decision-making platform meets the requirements based on the error rate of the writing scheme generation includes: Compare the error rate of the generated writing plan with the preset first error rate; If the error rate of the writing scheme generation is less than or equal to the preset first error rate, then the accuracy of the construction of the writing scheme auxiliary decision-making platform is determined to meet the requirements. If the error rate of the generated writing scheme is greater than the preset first error rate, then it is determined that the accuracy of the construction of the writing scheme auxiliary decision-making platform does not meet the requirements.

[0041] The reasons why the accuracy of the writing scheme auxiliary decision-making platform construction may not meet the requirements include inadequate synergy and adaptability between knowledge data and knowledge features, or insufficient attenuation coefficient of redundant information. The next step is to determine which specific cause it is, which is essentially the process of determining whether to increase the attenuation coefficient of redundant information.

[0042] Specifically, this involves determining whether the attenuation coefficient for redundant information needs to be increased. This includes: The error rate of the writing scheme is compared with the preset first error rate and the preset second error rate, respectively; If the 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, then it is determined that there is no need to increase the attenuation coefficient of redundant information. If the error rate of the writing scheme is greater than the preset second error rate, then it is determined that the attenuation coefficient of redundant information needs to be increased.

[0043] Specifically, when the error rate of generating the writing plan exceeds the preset second error rate, it is determined that the reason for the inaccuracy of the writing plan-assisted decision-making platform is that the attenuation coefficient of redundant information does not meet the requirements, thus requiring an increase in the attenuation coefficient of redundant information. When the error rate of generating the writing plan exceeds the preset first error rate but is less than or equal to the preset second error rate, it can be preliminarily determined that the synergistic adaptability between knowledge data and knowledge features does not meet the requirements. The next step is to determine, based on the deep feature coverage rate in the writing plan, whether the synergistic adaptability between knowledge data and knowledge features meets the requirements, i.e., to determine whether the reason for the inaccuracy of the writing plan-assisted decision-making platform is due to the inadequate synergistic adaptability between knowledge data and knowledge features.

[0044] It is understandable that the preset first error rate is lower than the preset second error rate, and the three intervals divided by the preset first and second error rates correspond to three different scenarios: The first interval is when the error rate of the generated writing plan is less than or equal to the preset first error rate. The corresponding situation is that the accuracy of the construction of the writing plan auxiliary decision-making platform meets the requirements, and no adjustment is needed. The second interval is when the error rate of the writing scheme generation is greater than the preset first error rate and less than or equal to the preset second error rate. The corresponding situation is: Since writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, which leads to the model underfitting because the input features are insufficient to describe the writing rules. At this time, it is necessary to further judge whether the synergistic adaptability of knowledge data and knowledge features meets the requirements. The third interval is when the error rate of the generated writing scheme is greater than the preset second error rate. The corresponding situation is: since the data required for platform training and inference often comes from multiple channels, and the quality of data from different sources varies, it may contain irrelevant and redundant content, which leads to inaccurate platform construction. In this case, it is necessary to adjust the attenuation coefficient of redundant information.

[0045] Understandably, using preset first and second error rates to characterize the accuracy of writing plan generation during the process is essentially a hierarchical control logic based on plan effectiveness and knowledge fit. This avoids the one-sidedness of evaluating plan quality with a single error threshold, while also achieving a linkage between qualitative judgment of the rationality of writing guidance and dynamic optimization of the rule engine, ultimately meeting the dual requirements of content accuracy and stability in the writing environment. The core function of the first error rate is to provide a quantitative basis for judging the accuracy of the writing plan auxiliary decision-making platform; the core function of the second error rate is to serve as a criterion for determining whether to adjust the attenuation coefficient of redundant information. The preset first and second error rates can be set according to actual working conditions. The setting of the preset first and second error rates aims to ensure the accuracy and practicality of the writing plan auxiliary decision-making platform. Optionally, the preset first and second error rates are determined through a limited number of experiments by evaluating the effect of different generation error rates on the construction of the writing plan auxiliary decision-making platform. The determined preset first and second error rates should be neither too small nor cause excessive interference to the construction process of the writing plan auxiliary decision-making platform. For example, 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%].

[0046] Preferably, the first error rate is 3% in a preferred embodiment, and the second error rate is 8% in a preferred embodiment.

[0047] Specifically, the error rate in generating writing schemes is the ratio of the number of incorrect writing schemes to the total number of writing schemes.

[0048] Specifically, errors in the generated writing scheme include inconsistencies with the intended text style, formatting issues, and logical contradictions.

[0049] Specifically, the increase in the attenuation coefficient of redundant information is determined by the difference between the generation error rate of the writing scheme and the preset second error rate.

[0050] Specifically, when the difference between the error rate of the generated writing scheme and the preset second error rate is within 3%, the attenuation coefficient of redundant information increases to 1.2 times the original value. When the difference between the error rate of the generated writing scheme and the preset second error rate exceeds 3%, the attenuation coefficient of redundant information increases by 0.05 for every 0.5% increase beyond the original value, in addition to increasing to 1.2 times the original value. For example, when the difference between the error rate of the generated writing scheme and the preset second error rate is 4%, the current attenuation coefficient of redundant information is 0.6, and the increased attenuation coefficient of redundant information is 0.6×1.2+0.05×2=0.82.

[0051] In practice, the method of this invention adjusts the attenuation coefficient of redundant information by setting a preset first error rate and a preset second error rate. Since the data required for platform training and inference often comes from multiple channels, and the quality of data from different sources varies, it may contain irrelevant redundant content, which may lead to inaccurate platform construction. By increasing the attenuation coefficient of redundant information, the weight reduction of content judged as redundant can be strengthened. That is, redundant information is given a lower contribution in subsequent training or inference, 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.

[0052] Specifically, the synergistic adaptability of knowledge data and knowledge features is determined based on the deep feature coverage rate in the writing scheme, including: The deep feature coverage rate in the writing scheme is compared with the preset second coverage rate; If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, then it is determined that the collaborative adaptability of knowledge data and knowledge features meets the requirements, and it is determined whether the attenuation coefficient of redundant information meets the requirements. If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, then the collaborative adaptability between knowledge data and knowledge features is determined to be unsatisfactory.

[0053] Specifically, 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 between knowledge data and knowledge features meets the requirements. However, if it has been previously determined that the accuracy of the construction of the writing scheme's auxiliary decision-making platform does not meet the requirements, then it is necessary to further determine whether the attenuation coefficient of redundant information meets the requirements.

[0054] In implementation, the attenuation coefficient of the actual redundant information is compared with the predetermined attenuation coefficient threshold to determine whether the attenuation coefficient of the redundant information meets the requirements. If the attenuation coefficient of the actual redundant information is less than the predetermined attenuation coefficient threshold, the attenuation coefficient of the redundant information is determined to be unacceptable. The predetermined attenuation coefficient threshold is the average value of the attenuation coefficient of the redundant information monitored in the previous three months of the historical period.

[0055] If the attenuation coefficient of redundant information does not meet the requirements, the attenuation coefficient of redundant information is increased; if the attenuation coefficient of redundant information meets the requirements, the error rate of the writing scheme generation is re-collected, and the accuracy of the construction of the writing scheme auxiliary decision-making platform is re-evaluated.

[0056] When the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, it can be determined that the reason for the inaccuracy of the writing scheme's auxiliary decision-making platform is that the synergistic adaptability between knowledge data and knowledge features is not up to standard. The reasons for this inaccuracy may include: the adaptive coefficient of feature extraction granularity not meeting requirements, or the synergistic dynamism between knowledge injection and model training not meeting requirements. The next step is to determine which specific reason it is, which is essentially the process of deciding whether to increase the adaptive coefficient of feature extraction granularity.

[0057] Specifically, determining whether an adaptive coefficient is needed to increase the granularity of feature extraction includes: The deep feature coverage rate in the writing scheme is compared 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, then it is determined that the adaptive coefficient of feature extraction granularity needs to be increased, and 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, then it is determined that there is no need to increase the adaptive coefficient of feature extraction granularity.

[0058] Specifically, when the deep feature coverage rate in the writing scheme is greater than the preset first coverage rate but less than or equal to the preset second coverage rate, it is determined that the reason for the non-compliance of the collaborative adaptation between knowledge data and knowledge features is that the adaptive coefficient of the feature extraction granularity does not meet the requirements. Therefore, it is necessary to increase the adaptive coefficient of the feature extraction granularity. When the deep feature coverage rate in the writing scheme is less than or equal to the preset first coverage rate, it can be preliminarily determined that the collaborative dynamics between knowledge injection and model training do not meet the requirements. Next, it is necessary to make a final determination on whether the collaborative dynamics between knowledge injection and model training meet the requirements based on the number of style confusions in the text style in the writing scheme. That is, to determine whether the reason for the non-compliance of the collaborative adaptation between knowledge data and knowledge features is the non-compliance of the collaborative dynamics between knowledge injection and model training.

[0059] It is understandable that the preset first coverage rate is less than the preset second coverage rate. The three intervals divided by the preset first and preset second coverage rates correspond to three different scenarios: The first interval is when the deep feature coverage rate in the writing scheme is less than or equal to the preset first coverage rate. The corresponding situation is: since knowledge data may only be injected once during the model initialization stage and not dynamically adjusted according to data feedback during the training process, the model cannot reconcile the contradiction between knowledge and data. It cannot learn formal expression or colloquial expression and falls into underfitting. At this time, it is necessary to further judge whether the synergistic dynamics between knowledge injection and model training meet the requirements. The second interval is when 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. The corresponding situation is: since writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features, which causes the model to underfit because the input features are insufficient to describe the writing pattern. At this time, it is necessary to adjust the adaptive coefficient of feature extraction granularity. The third interval is when the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate. The corresponding situation is that the synergistic adaptability of knowledge data and knowledge features meets the requirements. At this time, it is necessary to further determine whether the attenuation coefficient of redundant information meets the requirements.

[0060] Understandably, introducing preset first and second coverage rates to characterize the completeness and effectiveness of deep feature extraction during the process of extracting writing features is essentially a hierarchical control logic based on feature dimensions and writing task requirements. This avoids a one-sided evaluation of model fit using a single threshold, while simultaneously achieving a linkage between qualitative judgment and adaptive adjustment of feature coverage sufficiency, ultimately adapting to the dual requirements of multi-dimensional features and model stability in the writing environment. The core function of the first coverage rate is to provide a quantitative basis for the adaptive coefficient of feature extraction granularity adjustment during the feature extraction stage; the core function of the second coverage rate is to serve as a direct standard for distinguishing whether the synergistic adaptability of knowledge data and knowledge features meets the requirements. The preset first and second coverage rates can be set according to actual working conditions, aiming to ensure the accuracy and practicality of the writing scheme auxiliary decision-making platform. Optionally, the preset first and second coverage rates are determined through a limited number of experiments by evaluating the construction effect of different deep feature coverage rates on the writing scheme auxiliary decision-making platform. The determined preset first and second coverage rates should be neither too small nor cause excessive interference to the construction process of the writing scheme auxiliary decision-making platform. For example, the preset first coverage rate is generally selected in the range of [76%, 84%], and the preset second coverage rate is generally selected in the range of [86%, 94%].

[0061] Preferably, the first coverage rate is 80% in the preferred embodiment, and the second coverage rate is 90% in the preferred embodiment.

[0062] Specifically, the deep feature coverage rate in the writing scheme is the ratio of the number of deep features to the total number of core deep features required for the writing task.

[0063] Specifically, the increase in the adaptive coefficient of the feature extraction granularity is determined by the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate.

[0064] Specifically, when the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate is within 5%, the adaptive coefficient of the feature extraction granularity increases to 1.1 times the original value. When the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate exceeds 5%, the adaptive coefficient of the feature extraction granularity increases by 0.03 for every 1% increase beyond the original 1.1 times. For example, when the difference between the deep feature coverage rate in the writing scheme and the preset first coverage rate is 7%, the current adaptive coefficient of the feature extraction granularity is 0.7, and the increased adaptive coefficient of the feature extraction granularity is 0.7×1.1+0.03×2=0.83.

[0065] In practice, the method of this invention adjusts the adaptive coefficient of feature extraction granularity by setting a preset first coverage rate and a preset second coverage rate. Since writing is a multi-dimensional and complex task, the feature extraction process may only focus on surface features such as word frequency statistics or simple semantic similarity, ignoring deep features. This can lead to underfitting of the model because the input features are insufficient to describe the writing pattern. By increasing the adaptive coefficient of feature extraction granularity, the model can automatically expand the dimension of feature extraction, fill in the original feature gaps, make the input features more completely match the writing pattern, and allow the model to prioritize capturing highly relevant features. This reduces underfitting performance such as logical breaks and scene misalignment caused by relying solely on word frequency, and further improves the accuracy of constructing the auxiliary decision-making platform.

[0066] Specifically, the conflict feature circuit breaker threshold is determined based on the number of style confusions in the writing scheme, including: Compare the number of style confusions in the text style of the writing scheme with the preset number of confusions; If the number of style confusions in the writing scheme is less than or equal to the preset number of confusions, then the collaborative dynamics of knowledge injection and model training meet the requirements, there is no need to reduce the conflict feature circuit breaker threshold, and it is determined whether the adaptive coefficient of the feature extraction granularity meets the requirements. If the number of style confusions in the writing scheme is greater than the preset number of confusions, then the collaborative dynamics of knowledge injection and model training are not met, and the conflict feature circuit breaker threshold needs to be reduced.

[0067] Specifically, when the number of style confusions in the writing scheme is less than or equal to the preset number of confusions, it is determined that the synergistic dynamics of knowledge injection and model training meet the requirements. However, if the synergistic adaptability of knowledge data and knowledge features has been determined to be unacceptable, then it is necessary to further determine whether the adaptive coefficient of feature extraction granularity meets the requirements.

[0068] In implementation, the adaptive coefficient of the actual feature extraction granularity is compared with the predetermined adaptive coefficient threshold to determine whether the adaptive coefficient of the feature extraction granularity meets the requirements. If the adaptive coefficient of the actual feature extraction granularity is less than the predetermined adaptive coefficient threshold, the adaptive coefficient of the feature extraction granularity is determined to be unacceptable. The predetermined adaptive coefficient threshold is the average value of the adaptive coefficients of the feature extraction granularity monitored in the previous three months of the historical period.

[0069] If the adaptive coefficient of the actual feature extraction granularity does not meet the requirements, then increase the adaptive coefficient of the actual feature extraction granularity; if the adaptive coefficient of the actual feature extraction granularity meets the requirements, then re-collect the deep feature coverage in the writing scheme, and re-determine whether the synergistic adaptability of knowledge data and knowledge features meets the requirements.

[0070] When the number of style confusions in the text style of the writing scheme is greater than the preset number of confusions, it can be determined that the reason why the synergistic adaptability of knowledge data and knowledge features does not meet the requirements is that the synergistic dynamics of knowledge injection and model training do not meet the requirements. Therefore, it is necessary to reduce the conflict feature circuit breaker threshold.

[0071] It is understandable that the two intervals for the preset number of confusions correspond to two different scenarios: The first interval is when the number of style confusions in the text style of the writing scheme is less than or equal to the preset number of confusions. The corresponding situation is that the collaborative dynamics of knowledge injection and model training meet the requirements. At this time, it is necessary to further determine whether the adaptive coefficient of the feature extraction granularity meets the requirements. The second interval is when the number of style confusions in the writing scheme is greater than the preset number of confusions. The corresponding situation is that the knowledge data may only be injected once during the model initialization stage and not dynamically adjusted according to the data feedback during the training process. As a result, the model cannot reconcile the contradiction between knowledge and data, and can neither learn formal expressions nor colloquial expressions, thus falling into underfitting. At this time, it is necessary to adjust the conflict feature circuit breaker threshold.

[0072] Understandably, using a preset number of confusions to characterize the tolerance and judgment criteria for text style conflicts during model training is essentially a hierarchical management logic based on style consistency and the dynamic synergy of knowledge and data. This avoids the one-sidedness of evaluating style deviation with a single threshold, while also achieving qualitative judgment of model training stability and dynamic linkage of knowledge injection strategies, ultimately adapting to the dual needs of style uniformity and generation diversity in writing scenarios. The core function of the preset number of confusions is to serve as a quantitative basis for the model to distinguish between acceptable and unacceptable style deviations. The preset number of confusions can be set according to actual working conditions. The setting of the preset number of confusions aims to ensure the accuracy and practicality of the construction of the writing scheme auxiliary decision-making platform. Optionally, the preset number of confusions is determined through a limited number of experiments by evaluating the effect of different text style confusions on the construction of the writing scheme auxiliary decision-making platform. The determined preset number of confusions should be neither too small nor cause excessive interference to the construction process of the writing scheme auxiliary decision-making platform. For example, the preset number of confusions is generally selected in the range of [2 times, 4 times].

[0073] Preferably, the preferred embodiment of the preset number of confusions is 3 times.

[0074] Specifically, the number of style confusions in the writing scheme is the ratio of the number of different text styles in the writing scheme to the total number of writing schemes.

[0075] Specifically, the reduction in the conflict feature circuit breaker threshold is determined by the difference between the number of style confusions in the text style in the writing scheme and the preset number of confusions.

[0076] Specifically, when the difference between the style confusion of the text style and the preset number of confusions is within 2, the conflict feature circuit breaker threshold is reduced to 0.9 times the original value. When the difference between the style confusion of the text style and the preset number of confusions exceeds 2, the conflict feature circuit breaker threshold is reduced by 0.05 for each additional difference beyond the original value of 0.9. For example, when the difference between the style confusion of the text style and the preset number of confusions is 4, the current conflict feature circuit breaker threshold is 0.85, and the reduced conflict feature circuit breaker threshold is 0.85×0.9-0.05×2=0.665.

[0077] In practice, the method described in this invention adjusts the conflict feature circuit breaker threshold by setting a preset number of confusions. Since knowledge data may only be injected once during the model initialization stage and not dynamically adjusted according to data feedback during training, the model cannot reconcile the contradiction between knowledge and data, and thus cannot learn either formal or colloquial expressions, resulting in underfitting. By reducing the conflict feature circuit breaker threshold, conflict features exceeding the threshold can be judged as unacceptable, filtering training samples containing such high-conflict features, preventing the model from learning incorrect style associations from the data, and prioritizing the retention of reasonable features with low conflict. At the same time, it circuit breaks down unreasonable features with high conflict, helping the model distinguish between compatible general features and incompatible style conflict features, further improving the accuracy of constructing the auxiliary decision-making platform.

[0078] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems, characterized in that, include: The collected knowledge data is preprocessed to obtain knowledge features, which are then injected into the initial model for training to obtain the solution writing model. Input text and a predetermined text style are input into the solution writing model to generate a writing solution. The solution writing model is then optimized based on the writing solution, the text style of the writing solution, and the knowledge features to obtain a writing solution auxiliary decision-making platform. Obtain the error rate of the generated writing plan, and determine whether the accuracy of the construction of the writing plan auxiliary decision-making platform meets the requirements based on the error rate of the generated writing plan; If the accuracy of the writing scheme-assisted decision-making platform does not meet the requirements, then it is necessary to determine whether the attenuation coefficient of redundant information needs to be increased. If it is not necessary to increase the attenuation coefficient of redundant information, then obtain the deep feature coverage in the writing scheme to determine whether the synergistic adaptability of knowledge data and knowledge features meets the requirements. If the collaborative adaptability between the knowledge data and knowledge features does not meet the requirements, then determine whether it is necessary to increase the adaptive coefficient of the feature extraction granularity. If there is no need to increase the adaptive coefficient of feature extraction granularity, the conflict feature circuit breaker threshold is determined based on the number of style confusions in the text style in the writing scheme.

2. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 1, characterized in that, Determining whether the accuracy of the writing scheme auxiliary decision-making platform meets the requirements based on the error rate of the writing scheme generation includes: Compare the error rate of the generated writing plan with the preset first error rate; If the error rate of the writing scheme generation is less than or equal to the preset first error rate, then the accuracy of the construction of the writing scheme auxiliary decision-making platform is determined to meet the requirements. If the error rate of the generated writing scheme is greater than the preset first error rate, then it is determined that the accuracy of the construction of the writing scheme auxiliary decision-making platform does not meet the requirements.

3. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 2, characterized in that, Determine whether the attenuation coefficient of redundant information needs to be increased, including: The error rate of the writing scheme is compared with the preset first error rate and the preset second error rate, respectively; If the error rate of the writing scheme is greater than the preset second error rate, then it is determined that the attenuation coefficient of 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, then it is determined that there is no need to increase the attenuation coefficient of redundant information.

4. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 3, characterized in that, The increase in 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 method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 4, characterized in that, Based on the deep feature coverage in the writing scheme, determine whether the collaborative adaptability between knowledge data and knowledge features meets the requirements, including: The deep feature coverage rate in the writing scheme is compared with the preset second coverage rate; If the deep feature coverage rate in the writing scheme is greater than the preset second coverage rate, then it is determined that the collaborative adaptability of knowledge data and knowledge features meets the requirements, and it is determined whether the attenuation coefficient of redundant information meets the requirements. If the deep feature coverage rate in the writing scheme is less than or equal to the preset second coverage rate, then the collaborative adaptability between knowledge data and knowledge features is determined to be unsatisfactory.

6. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 5, characterized in that, Determine whether the adaptive coefficient for increasing the granularity of feature extraction needs to be adjusted, including: The deep feature coverage rate in the writing scheme is compared 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, then it is determined that the adaptive coefficient of feature extraction granularity needs to be increased, and 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, then it is determined that there is no need to increase the adaptive coefficient of feature extraction granularity.

7. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 6, characterized in that, The increase in the adaptive coefficient of the 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 method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 7, characterized in that, The conflict feature circuit breaker threshold is determined based on the number of style confusions in the text style of the writing scheme, including: Compare the number of style confusions in the text style of the writing scheme with the preset number of confusions; If the number of style confusions in the writing scheme is less than or equal to the preset number of confusions, then it is determined that the collaborative dynamics of knowledge feature injection and model training meet the requirements, and it is not necessary to reduce the conflict feature circuit breaker threshold. It is also determined whether the adaptive coefficient of the feature extraction granularity meets the requirements. If the number of style confusions in the writing scheme is greater than the preset number of confusions, then the collaborative dynamics of knowledge injection and model training are not met, and the conflict feature circuit breaker threshold needs to be reduced.

9. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 8, characterized in that, The number of style confusions in the writing scheme is the ratio of the number of different text styles in the writing scheme to the total number of writing schemes.

10. The method for constructing an AI-assisted decision-making platform for writing solutions based on expert systems according to claim 9, characterized in that, The reduction in the conflict feature circuit breaker threshold is determined by the difference between the number of style confusions in the writing scheme and the preset number of confusions.

Images